content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
It is well known \cite{B}, \cite{Kalman}, that the number of feedback equivalence classes of reachable control systems over a $n$-dimensional $\mathbb{K}$-vector space equals the number $p_{\mathbb{N}}(n)$ of partitions of integer $n$. This number equals the number, $\operatorname{sol}_{\mathbb{N}}$, of solutions in $\mathbb{N}$ of linear diophantine equation
\begin{equation}\label{ne+}
n=z_1+2z_2+\cdots+nz_n
\end{equation}
Above result is generalized in \cite{miguel2013} to the general framework of regular (locally Brunovsky) linear systems over a commutative ring. In fact the number of feedback equivalence classes of regular systems with (finitely generated projective) state space $X$ equals the number $\operatorname{sol}_{\mathbf{P}(R)}$ of solutions of the linear equation
\begin{equation}
X=Z_1\oplus Z_2^2\oplus\cdots\oplus Z_n^n
\end{equation}
in monoid $(\mathbf{P}(R),\oplus)$ of finitely generated projective $R$-modules. This number equals the number $p_{\mathbf{P}(R)}(X)$ of partitions(direct sum decompositions) of projective module $X$ into direct summands if monoid $\mathbf{P}(R)$ happens to be cancellative.
Our goal in this paper is, applying results of \cite{miguel2013}, to give a complete account and obtain formulae relating the number of classes of feedback isomorphisms of regular systems with state space $X$ over different commutative rings with unit element using partitions.
First, we compute that number when $R$ is a finite product of rings $R\simeq R_{1} \times \ldots \times R_{t}$ in terms of each direct factor $R_{i}$. This case is a generalization of $R=\mathbb{Z}/l\mathbb{Z}$, the ring of modular integers. Finally, we compute the case of $R$ being a Dedekind domain.
The paper is organized as follows: In Section 2 we have some details about monoid $\mathbf{P}(R)$ of isomorphism classes of finitely generated projective $R$-module.
In Section 3 we obtain the number of classes of feedback isomorphisms of regular systems with state space $X$ when the ring is projectively trivial and when the ring splits as a finite product of rings. In Section 4, we obtain the formula over Dedekind domains. Finally, we give our conclusions.
\section{The equation}
The paper deals with the solutions of equation
\begin{equation}
\label{Xeoplus}
X=Z_1\oplus Z_2^2\oplus\cdots\oplus Z_n^n
\end{equation}
in the monoid $(\mathbf{P}(R),\oplus)$ of isomorphism classes of finitely generated projective $R$-modules.
Let us review some elementary properties of $\mathbf{P}(R)$ which will be applied in the sequel. The reader is referred to \cite{Kbook} for more details.
\begin{proposition}
\label{preli}
Let $R$ be a commutative ring and let $\mathbf{P}(R)$ be the set of isomorphism classes of finitely generated projective $R$-modules. Then the following properties hold:
\begin{itemize}
\item[(i)] $\mathbf{P}(R)$ is a monoid under operation $[P]\oplus [Q] =[P\oplus Q]$. Identity element is the zero $R$-module.
\end{itemize}
In the sequel we denote by $P$ the finitely generated projective $R$-module and its isomorphism class.
\begin{itemize}
\item[(ii)] $\mathbf{P}(R)$ is a commutative monoid (i.e. $P \oplus Q=Q \oplus P$)
\item[(iii)] $\mathbf{P}(R)$ is a zero-sum-free monoid (i.e. $P\oplus Q = 0 \Rightarrow P=Q=0$)
\item[(iv)] The mapping $\varphi:(\mathbb{N},+)\rightarrow (\mathbf{P}(R),\oplus)$ sending $0\mapsto 0$ and $n\mapsto R^n$ is an injective morphism of monoids.
\item[(v)] If every finitely generated projective $R$-module is free (i.e. $R$ is projectively trivial) then above morphism $\varphi$ is an isomorphism.
\end{itemize}
If $R$ is a domain with field of fractions $\mathbb{K}_R$ then every finitely generated $R$-module $P$ has constant rank $\operatorname{rk}(P)=\dim (P\otimes_{R} \mathbb{K}_R)$.
\begin{itemize}
\item[(vi)] Mapping $\operatorname{rk}:(\mathbf{P}(R),\oplus)\rightarrow (\mathbb{N},+)$ is a monoid morphism
\item[(vii)] $\operatorname{rk}$ is left inverse of $\varphi$; that is, $\operatorname{rk}\circ\varphi=Id_{\mathbb{N}}$
\end{itemize}
\end{proposition}
We also need some notation:
\begin{notation}
\label{notation}
Let $\Sigma$ be a regular linear system.
We denote
\begin{enumerate}
\item By $fe_{R}(X)$ as the number of feedback classes of regular systems over $R$ with state space $X$.
\item By $sol_{\mathbf{P(R)}}(X)$ as the number of solutions of equation (\ref{Xeoplus}) in the $\mathbf{P}(R)$ where $X$ is a finitely generated projective $R$-module.
\end{enumerate}
\end{notation}
Let us remark that above two numbers are equal \cite{miguel2013}. This gives the pass from systems theory to combinatorial issues.
\section{The equation in monoids $\mathbb{N}$ and $\mathbb{N}^t$. Projectively trivial rings and finite product of rings.}
Now we study the equation (\ref{Xeoplus}) in monoid of nonnegative integers and its finite products. From the systems theory point of view, this is to solving the case of regular systems over projectively trivial rings and over finite product of projectively trivial rings.
Examples of projectively trivial rings are: Fields $\mathbb{K}$; local rings like $\mathbb{K}[[x_1,...,x_s]]$ or $\mathbb{Z}/p^r\mathbb{Z}$, $p$ prime; principal ideal domains like $\mathbb{Z}$ or $\mathbb{K}[x]$; and polynomial rings like $\mathbb{K}[x_1,...,x_s]$ or $\mathbb{Z}[x_1,...,x_s]$. Example of finite products of projectively trivial rings are modular integers rings and their rings of polinomials $(\mathbb{Z}/m\mathbb{Z})[x_{1},\ldots,x_{s}]$.
\subsection{Projectively trivial rings}
If $R$ is projectively trivial, then the number $\operatorname{fe}_R(R^n)$ of feedback classes of isomorphisms of regular systems over $R^n$ (via the isomorphism $\varphi:\mathbf{P}(R)\cong\mathbb{N}$) equals the number, $\operatorname{sol}_{\mathbf{P}(R)}(R^n)$, of solutions of the linear equation in $\mathbb{N}$
\begin{equation}
n=z_1+2z_2+\cdots +nz_n
\end{equation}
This number is the number of partitions $p_{\mathbb{N}}(n)$ of integer $n$. Thus we have the result:
\begin{theorem}[cf. Corollary 8.1 \cite{miguel2013}]\label{teoprincipal}
Let $R$ be a projectively trivial ring. Then
\begin{equation}
\operatorname{fe}_R(R^n)=\operatorname{sol}_{\mathbf{P}(R)}(R^n)=\operatorname{sol}_{\mathbb{N}}(n)=p_{\mathbb{N}}(n)
\end{equation}
\end{theorem}
\subsection{Product rings}
Now suppose that $R=R_1\times\cdots\times R_t$ is a finite product of rings. Our goal in this section is to prove the formula
\begin{equation}
\operatorname{sol}_{\mathbf{P}(R)}(R^n)=\operatorname{sol}_{\mathbf{P}(R_1)}(R_1^n)\cdot \ldots \cdot \operatorname{sol}_{\mathbf{P}(R_t)}(R_t^n)
\end{equation}
and thus, from the systems theory point of view we will have the account
\begin{equation}
\operatorname{fe}_R(R^n)=\operatorname{fe}_{R_1}(R_1^n) \cdot \ldots \cdot\ fe_{R_t}(R_t^n)
\end{equation}
In order to prove above formulae we need to describe the structure of finitely generated projective $R$-modules when $R$ is a direct product of rings.
\begin{lemma}
\label{teo}
If $R\simeq R_{1}\times\cdots\times R_{t}$ is a finite product of rings.Then the following holds
\begin{center}
$\mathbf{P}(R) \cong \mathbf{P}(R_{1})\times...\times \mathbf{P}(R_{t})$.
\end{center}
\end{lemma}
\begin{proof}
The set $M(n,R)$ of $n\times n$ matrices over $R$ is embedded in $M(n+1,R)$ by
\begin{center}
$(a) \mapsto \left(\begin{array}{cc}
a & 0 \\
0& 0
\end{array}
\right)$
\end{center}
and $\displaystyle{M(R)=\bigcup_{n\geq 1}M(n,R)}$. Note that every matrix in $M(R)$ has finite size. The set of idempotent matrices in $M(R)$ is denoted by $Idem(R)$.
On the other hand $GL(n,R)$ is embedded in $GL(n+1,R)$ by
\begin{center}
$a \mapsto \left(\begin{array}{cc}
a & 0 \\
0& 1
\end{array}
\right)$
\end{center}
and $\displaystyle{GL(R)=\bigcup_{n\geq 1}GL(n,R)}$. Every matrix in $GL(R)$ is invertible having finite size.
$\mathbf{P}(R)$ may be identified \cite[Th. 1.2.3]{Rosenberg} with the set of conjugation orbits of group $GL(R)$ on set $Idem(R)$. Since $GL(R)\cong GL(R_1)\times\cdots\times GL(R_t)$ acts on $Idem(R)=Idem(R_1)\times\cdots\times Idem(R_t)$ componentwise it follows the result.
\end{proof}
\begin{theorem}
Let $R\simeq R_{1}\times \ldots \times R_{t}$ be a finite product of rings. Then:
\begin{itemize}
\item[(i)] $\operatorname{sol}_{\mathbf{P}(R)}(R^n)=\operatorname{sol}_{\mathbf{P}(R_1)}(R_1^n)\cdot \ldots \cdot \operatorname{sol}_{\mathbf{P}(R_t)}(R_t^n)$
\item[(ii)] $fe_{R}(R^n)=fe_{R_{1}}(R_1^n) \cdot \ldots \cdot fe_{R_{t}}(R_t^n)$
\end{itemize}
\end{theorem}
\begin{proof}
By Lemma \ref{teo}, $\mathbf{P}(R) \cong \mathbf{P}(R_{1})\times...\times \mathbf{P}(R_{t})$ we can solve the equation $R^n \simeq Z_1\oplus Z_2^2\oplus\cdots\oplus Z_n^n$ componentwise on every factor ring. Therefore the conditions hold.
\end{proof}
\begin{corollary}
\label{coro}
Let $R \simeq R_{1} \times \ldots \times R_{t}$ be a finite product of projectively trivial rings $R_{i}$ for $i=1, \ldots t$. Then $\operatorname{fe}_{R}(R^n)=\left(\operatorname{sol}_{\mathbb{N}}(n)\right)^{t}=\left(p_{\mathbb{N}}(n)\right)^t$
\end{corollary}
\begin{proof}
Since the rings $R_{i}$ are projectively trivial rings, then all finitely generated projective modules over each $R_{i}$ are free, that is, $\mathbf{P}(R_{i})\simeq \mathbb{N}$ for $i=1,...,t$.
By Theorem \ref{teo} we deduce that
\begin{equation}
\label{bu}
\mathbf{P}(R)\simeq \mathbf{P}(R_{1})\times ... \times \mathbf{P}(R_{t})\simeq \mathbb{N} \times\overset{t)}{...} \times \mathbb{N}.
\end{equation}
So, the number of classes of feedback isomorphisms of locally Brunovsky linear systems over $R$ is the number of solutions of the equation
\begin{center}
$m = x_{1}+2x_{2}+3x_{3}+...+sx_{s}$ where $m$ and $x_{i}$ are t-uples of natural numbers.
\end{center}
Now, $m=(m_{1}, \ldots, m_{t}) \in \mathbf{P}(X)$ and $X\simeq R^{n}$ then $m=(n, \overset{t)}{\ldots},n)$. So, the equation to solve is
\begin{center}
$(n,n,\overset{t)}{...},n)=(a_{1},b_{1},...,t_{1})+2(a_{2},b_{2},...,t_{2})+3(a_{3},b_{3},...,t_{3})+...+s(a_{s},b_{s},...,t_{s})$,
\end{center}
Then, we look for the number of solutions of the system
\begin{equation}
\left\{\begin{array}{l}
n= a_{1}+2a_{2}+3a_{3}+...+sa_{s} \\
n=b_{1}+2b_{2}+3b_{3}+...+sb_{s} \\
\vdots \\
n=t_{1}+2t_{2}+3t_{3}+...+st_{s}\\
\end{array}\right.
\end{equation}
where $n, a_{i}, b_{i},...,t_{i} \in \mathbb{N}$ for $i=1,...,t$.
Now, since $\mathbb{N}$ is cancellative, the number of solutions of i-th equation is equal to $p(n)$, the partitions of $n$ for each $i=1,...t$. Thus, the number of solutions of the system is equal to
\begin{center}
$fe_{R}(R^{n})=p_{\mathbb{N}}(n)\cdot p_{\mathbb{N}}(n) \cdot \overset{t)}{\ldots} \cdot p_{\mathbb{N}}(n)= (p_{\mathbb{N}}(n))^{t}$
\end{center}
\end{proof}
\begin{corollary}
For the ring $R= \mathbb{Z}_{l} \simeq \mathbb{Z}_{p_{1}^{r_{1}}} \times \ldots \times \mathbb{Z}_{p_{t}^{r{t}}}$ we have
\begin{center}
$fe_{\mathbb{Z}_{l}}(R^{n})= (p_{\mathbb{N}}(n))^{t}$
\end{center}
\end{corollary}
\begin{proof}
Since $\mathbf{P}(\mathbb{Z}/{p_{i}^{r_{i}}}\mathbb{Z})\simeq \mathbb{N}$ for each $i$, we conclude the proof.
\end{proof}
\section{The equation in $\mathbb{N}\times G$, G abelian group. Dedekind domains.}
Let $R$ be a commutative ring. Let $\mathrm{Pic}(R)$ be the set of isomorphism classes of line bundles over $R$ (finitely generated projective $R$-modules of $rank$ one). Then \cite[I\S3]{Kbook} $(\mathrm{Pic}(R),\otimes_R)$ is an abelian group where $R=1_{\mathrm{Pic}(R)}$ and $P^{-1}=\mathrm{Hom_R(P,R)}$.
In the sequel $R$ denotes a Dedekind domain with field of fractions $\mathbb{K}(R)$; that's to say, a commutative domain (no nonzero zero-divisors) which is noetherian, integrally closed and $1$-dimensional. Then \cite[I.3.4]{Kbook} finitely generated projective $R$-module $P$ is completely classifed by its rank $\operatorname{rk}(P)=\dim (P \otimes \mathbb{K}(R)$) and its determinant $\displaystyle{\wedge^{\mathrm{rk}(P)}P}$. To be precise, $P$ is isomorphic to $R^{\mathrm{rk} (P)-1}\oplus \displaystyle{\wedge^{\mathrm{rk}(P)} P}$.
Thus, $\mathbf{P}(R)$ equals $[\mathbb{N}^{+} \times Pic(R)]\cup \{0\}$ as a set. Arithmetic in $(\mathbf{P}(R), \oplus)$ is given by
\begin{equation}
P\oplus Q \cong R^{\operatorname{rk}(P)+\operatorname{rk}(Q)-1}\oplus \left( \displaystyle{\wedge^{\mathrm{rk}(P)} P}\otimes_{R} \displaystyle{\wedge^{\mathrm{rk}(Q)} Q}\right)
\end{equation}
where zero module $0$ is the identity and internal law $\oplus$ is commutative by Proposition \ref{preli}.
\begin{notation}
The determinant line bundle $\wedge^{\mathrm{rk}(P)}$ of a finitely generated $R$- module $P$ is denoted by $\operatorname{det}(P)$.
\end{notation}
\begin{remark}
\label{remark}
Let $X$ be a finitely generated projective $R$-module. Looking for the number of classes of feedback isomorphisms of regular systems over $X$ is equivalent to computing the number of solutions of equation (\ref{Xeoplus})
\begin{equation}
X=Z_1\oplus Z_2^2\oplus\cdots\oplus Z_n^n
\end{equation}
in $\mathbf{P}(R)\cong [\mathbb{N}^{+}\times Pic(R)]\cup \{0\}$. These solutions are determined by solutions of
\begin{equation}
\label{rank}
\operatorname{rk}(X)=\operatorname{rk}(Z_1)+2\operatorname{rk}(Z_2)+\cdots+n\operatorname{rk}(Z_n)
\end{equation}
in $(\mathbb{N}, +)$ together with a solution of
\begin{equation}
\label{det}
\operatorname{det}(X)=\operatorname{det}(Z_1)\otimes\operatorname{det}(Z_2)^{\otimes 2}\otimes\cdots\otimes\operatorname{det}(Z_n)^{\otimes n}
\end{equation}
in $Pic(R)$.
Note that if we only fix the rank of the state space, $rk(X)=n$, then $X\cong R^{n-1}\oplus L$ and
\begin{equation*}
\operatorname{det}(X)= \operatorname{det}(R^{n-1}\oplus L)= \displaystyle{\wedge^{n} (R^{n-1}\oplus L)}=\displaystyle{\bigoplus^{n}_{i=0}[(\wedge^{i} R^{n-1})\otimes (\wedge^{n-i} L)]}=
\end{equation*}
\begin{equation*}
=\displaystyle{\bigoplus^{n-2}_{i=0} [R^{\binom{n-1}{i}}\otimes 0] \oplus[(\wedge^{n-1} R^{n-1})\otimes (\wedge^{1} L)] \oplus [\wedge^{n} R^{n-1} \otimes \wedge^{0} L]}= R\otimes L = L
\end{equation*}
and equation (\ref{det}) turns to be
\begin{equation}
L=\operatorname{det}(Z_1)\otimes\operatorname{det}(Z_2)^{\otimes 2}\otimes\cdots\otimes\operatorname{det}(Z_n)^{\otimes n}
\end{equation}
\end{remark}
\begin{remark}
\label{detcero}
Let $P$ a finitely generated projective module over a Dedekind domain $R$. In particular note that if $\operatorname{rk}(P)=0$, then $\operatorname{det}(P)=R$.
Then, note that the solutions of equation (\ref{det}) in $Pic(R)$ are entangled with solutions of ranks equation in $\mathbb{N}$.
\end{remark}
\begin{remark}
\label{alfa}
A classical result by Claborn \cite{Claborn} shows that given any abelian group $G$, there exists a Dedekind domain $R$ such that $Pic(R) \cong G$.
Let $\mid Pic(R) \mid=p$ be the order of $Pic(R)$ with $p$ prime. We will use the isomorphism
\begin{center}
$(Pic(R)$, $\otimes) \overset{\overset{\alpha}{\cong}}{\longrightarrow}$ $(\mathbb{Z}/p\mathbb{Z}, +)$
\end{center}
\begin{center}
$L$ \hspace{0.5cm} $ \mapsto$ \hspace{0.3cm} $ \alpha(L)$
\end{center}
where in particular $\alpha(R)=\overline{0}$ in $\mathbb{Z}/p\mathbb{Z}$
\end{remark}
Finally, we introduce some notation about the account of regular systems over a Dedekind domain, $R$.
\begin{notation}
\begin{enumerate}
\item $fe_{R}(X)$ denotes the number of feedback classes of regular systems over $X$ (\ref{notation}).
\item $fe_{R}(n)$ will be the number of classes of feedback isomorphisms of regular systems over $R$ with a state space a finitely projective $R$-module of rank $n$.
\item In the case of Dedekind domains $X\cong R^{n-1}\oplus L$ and therefore $fe_{R}(X)= fe_{R}(R^{n-1}\oplus L )$.
\end{enumerate}
\end{notation}
In order to solve above equations (\ref{rank}) and (\ref{det}) we introduce the following combinatorial number
\begin{definition}
Let $n$ be a positive integer and $1\leq k \leq n$. We denote by $\nu(n,k)$ the set of partitions of integer $n$ into $k$ different summands. We also denote by $\nu(n,k)$ its cardinal.
\end{definition}
As matter of example $\nu(6,2)$ is the number of partitions of integer $6$ into $2$ different summands and hence contains exactly partitions $$(51),(42),(411),(3111),(2211),(21111)$$ and therefore $\nu(6,2)=6$.
\begin{definition}
\label{nukp}
Let $p$ be a prime number. We denote by $\nu(n,k,p)$ the set of partitions in $\nu(n,k)$ where all coefficientes of the summands are multiples of $p$. We also denote by $\nu(n,k,p)$ its cardinal.
For convenience let's denote by $\nu'(n,k,p)= \nu(n,k)-\nu(n,k,p)$.
\end{definition}
As matter of example, $\nu(6,2,2)=1$ because the only partition in $\nu(6,2)$ with the property that all summands are multiple of $2$ is $(42)$ .
\begin{remark}
Combinatorial number $\nu(n,k)$ needs further study. We only point out two straightforward properties:
\begin{itemize}
\item[(i)] $\nu(n,1)=\text{div}(n)$; that is $\nu(n,1)$ equals the number of divisors (including both $1$ and $n$) of integer $n$
\item[(ii)] If $n<k(k+1)/2$ then $\nu(n,k)=0$ because the least partition one can form with $k$ different summands is $(k,k-1,...,2,1)$ and therefore an $n\geq 1+2+\cdots +k=\frac{k(k+1)}{2}$ is needed.
\end{itemize}
\end{remark}
We state our main result:
\begin{theorem}
Let $R$ be a Dedekind domain and let $Pic(R)$ be its Picard Group. Then, the number of feedback classes of regular systems is as follows:
\begin{itemize}
\item[(i)] $fe_{R}(n)$ is the number of solutions $(Z_{1},Z_{2},\ldots,Z_{n})$ of equation
\begin{equation}
\label{rankn}
n=rk(Z_{1})+2rk(Z_{2})+\ldots+nrk(Z_{n}) \textit{ in } (\mathbb{N}, +)
\end{equation}
\item[(ii)] If $\mid Pic(R) \mid=\infty$ then $\operatorname{fe}_R(n)=\infty$.
\item[(iii)] If $\mid Pic(R) \mid=d<\infty$ then $fe_R(n)=\displaystyle{\sum_{k=1}^{n}\nu(n,k)\cdot d^k}$
\item[(iv)] $fe_{R}(X)$ is the number of solutions $(Z_{1},Z_{2},\ldots,Z_{n})$ of the system of equations (see Remarks \ref{remark} and \ref{detcero}).
\begin{center}
$\left\{\begin{array}{l}
\operatorname{rk}(X)=\operatorname{rk}(Z_1)+2\operatorname{rk}(Z_2)+\cdots+n\operatorname{rk}(Z_n)$ in $(\mathbb{N}, +) \\
\\
\operatorname{det}(X)=L=\operatorname{det}(Z_1)\otimes\operatorname{det}(Z_2)^{\otimes 2}\otimes\cdots\otimes\operatorname{det}(Z_n)^{\otimes n}$ in $(Pic(R),\otimes).\\
\end{array}\right.$
\end{center}
\item[(v)] If $\mid Pic(R) \mid=p$ is prime then $fe_R(X\simeq R^{n})=\displaystyle{\sum_{k=1}^{n}\left[\nu(n,k,p)\cdot p^{k}+\nu'(n,k,p)\cdot p^{k-1}\right]}$.
\item[(vi)]) If $\mid Pic(R) \mid=p$ is prime then $fe_{R}(R^{n-1}\oplus L)=\displaystyle{\sum_{k=1}^{n}\nu'(n,k,p)\cdot p^{k-1}}$
\end{itemize}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item[(i)] Is clear by Remark \ref{remark}.
\item[(ii)] Suppose that $Pic(R)$ is of infinite order and $L$ varies in $Pic(R)$. Then
\begin{center}
$(Z_1=R^{n-1}\oplus L,Z_2=0,...)$
\end{center}
are infinitely many different solutions of equation (\ref{rankn}).
\item[(iii)] $\nu(n,k)$ is the set of solutions $(\operatorname{rk}(Z_{1}),\operatorname{rk}(Z_{2}),\ldots,\operatorname{rk}(Z_{n}))$ of the equation (\ref{rankn}) where $k$ of the entries of above tuple are non zero. Thus
\begin{center}
$(\operatorname{rk}(Z_{1}),\operatorname{rk}(Z_{2}),\ldots,\operatorname{rk}(Z_{n}))=(0,\ldots,\operatorname{rk}(Z_{i_{1}}),\ldots,\operatorname{rk}( Z_{i_{k}}),\ldots,0,\ldots)$
\end{center}
In order to realize solutions $(Z_{1}, \ldots, Z_{n})$ we are free to choose $L_{1},\ldots, L_{k}$ in $Pic(R)$ to obatin solutions
\begin{center}
$(0,\ldots,0,R^{rk(Z_{i_{1}})-1}\oplus L_{1},0,\ldots,0, R^{rk(Z_{i_{k}})-1}\oplus L_{k},0,\ldots)$
\end{center}
Since $L_{i}$ varies in $Pic(R)$, then there are exactly $d^{k}$ different choices and therefore
\begin{center}
$fe_R(n)=\displaystyle{\sum_{k=1}^{n}\nu(n,k)\cdot d^k}$
\end{center}
\item[(iv)] Is clear from the Remark \ref{remark}.
\item[(v)] If $\mid Pic(R) \mid=p$ is prime, then by Remark \ref{alfa}, $\operatorname{Pic}(R) \cong \mathbb{Z}/p\mathbb{Z}$, and equations giving $fe_{R}(R^{n})$ are:
\begin{center}
$\left\{\begin{array}{l}
n=\operatorname{rk}(Z_1)+2\operatorname{rk}(Z_2)+\cdots+n\operatorname{rk}(Z_n)$ in $(\mathbb{N}, +) \\
\\
0=a_{1} + 2a_{2} + \cdots + na_{n}$ in $(\mathbb{Z}/p\mathbb{Z},+).\\
\end{array}\right.$
\end{center}
where $a_{i}=\alpha(\operatorname{det}(Z_{i}))$
There are exactly $\nu(n,k)$ different solutions for the ranks equation with exactly $k$ non zero $\operatorname{rk}(Z_{i})'s$. Every solution of ranks equation gives some choices for the second equation. But is crucial to know how many coefficients are non zero modulo $p$.
The equation over determinants by $\alpha$ is on the form
\begin{equation}
0= a_{1}+2a_{2}+\ldots+pa_{p}+...+(2p)a_{2p}+\ldots+ na_{n} \textit{ in } \mathbb{Z}/p\mathbb{Z}
\end{equation}
Let us reorder the summands such that we have $l= \lfloor \frac{n}{p}\rfloor$ summands which coefficientes are multiple of $p$, and $n-l$ summands whose coefficientes are prime with $p$.
\begin{equation*}
\label{ecup}
0= \overbrace{pa_{p}+2pa_{2p}+\ldots+lpa_{lp}}^{l \text{ summands}}+
\overbrace{a_{1}+\ldots+(p-1)a_{p-1}+(p+1)a_{p+1}+...}^{(n-l) \text{ summands }}
\end{equation*}
Since the group of $l$ summands vanishes module $p$, then above equation in $\mathbb{Z}/p\mathbb{Z}$ is in fact
\begin{center}
$0= a_{1}+\ldots+(p-1)a_{p-1}+(p+1)a_{p+1}+...$
\end{center}
or even
\begin{center}
$0= 0$ if all non zero $(z_{i})'s $ are on the form $i=\lambda p$.
\end{center}
In the former case, corresponding to $\nu(n,k,p)$ in Definition \ref{nukp}, we have exactly $p^{(k-l-1)}$ choices of $a_{1},...,a_{p-1},a_{p+1},...$ and $p^{l}$ choices for $a_{p},a_{2p},...,a_{lp}$. So, there are $p^{k-l-1} \cdot p^{l}=p^{k-1} $ different choices for every solution in $\nu(n,k)$.
In the latter case, corresponding to $\nu'(n,k,p)$ in Definition \ref{nukp}, $p^{k}$ different solutions are freely chosen for $a_{p},\ldots, a_{kp}$.
Therefore
\begin{center}
$fe_{R}(R^{n})=\displaystyle{\sum_{k=1}^{n}\left(\nu(n,k,p)\cdot p^{k}+\nu'(n,k,p)\cdot p^{k-1}\right)}$
\end{center}
\item[(vi)] If $rk(X)=n$ but $X$ is not free, then $X\cong R^{n-1}\oplus L$ and $\alpha(L)\neq 0$ in $\mathbb{Z}/p\mathbb{Z}$. The equations to compute $fe_{R}(R^{n-1}\oplus L)$ are
\begin{center}
$\left\{\begin{array}{l}
n=\operatorname{rk}(Z_1)+2\operatorname{rk}(Z_2)+\cdots+n\operatorname{rk}(Z_n)$ in $(\mathbb{N}, +) \\
\\
0\neq \alpha(L) =a_{1} + 2a_{2} + \cdots + na_{n}$ in $(\mathbb{Z}/p\mathbb{Z},+) \textit{ where } a_{i}=\alpha(\operatorname{det}(z_{i}))\\
\end{array}\right.$
\end{center}
Analogous reasoning of $(v)$ gives us to
\begin{center}
$0\neq \alpha(L) =a_{1} + 2a_{2} + \cdots + (p-1)a_{p-1}+(p+1)a_{p+1}+\ldots$
\end{center}
or
\begin{center}
$0\neq \alpha(K)=0$ having no solution.
\end{center}
Therefore if $X$ is not free of rank $n$ we have
\begin{center}
$fe_{R}(R^{n-1}\oplus L)=\displaystyle{\sum_{k=1}^{n}\nu'(n,k,p)\cdot p^{k-1}}$
\end{center}
\end{enumerate}
\end{proof}
\begin{remark}
Note that if we perform the sum of all computations over elements of $Pic(R)$, then we obatin coherent relationship between our formulae.
\begin{equation*}
\displaystyle{\sum_{L \in Pic(R)} fe_{R}(R^{n-1}\oplus L)}= \displaystyle{\sum_{k=1}^{n}\left(\nu(n,k,p)\cdot p^{k}+\nu'(n,k,p)\cdot p^{k-1}\right)}+ (p-1)\cdot \displaystyle{\sum_{k=1}^{n} \nu'(n,k,p)\cdot p^{k-1}} =
\end{equation*}
\begin{equation*}
=\displaystyle{\sum_{k=1}^{n} \left(\nu(n,k,p)\cdot p^{k} + p\cdot \nu'(n,k,p)\cdot p^{k-1} \right)}=\displaystyle{\sum_{k=1}^{n} \left(\nu(n,k,p)\cdot p^{k} + \nu'(n,k,p)\cdot p^{k} \right)}=
\end{equation*}
\begin{equation*}
=\displaystyle{\sum_{k=1}^{n} \left(\nu(n,k,p)+ \nu'(n,k,p)\right) \cdot p^{k}}= \displaystyle{\sum_{k=1}^{n} \nu(n,k)\cdot p^{k}}=fe_{R}(n).
\end{equation*}
\end{remark}
\section{Conclusions}
This paper gives a combinatorial approach to a well known problem in systems theory. New (as far as we now) combinatorial numbers $\nu(n,k)$ is introduced. Further study of these combinatorial numbers would be interesting.
A motivation for the study of partitions in monoids is introduced. In particular feedback equivalence problems over product rings translate to partitions over product monoids and feedback equivalence problems over Dedekind domains $R$ translate to partitions and linear equations in $Pic(R)$.
\section*{References}
|
\section{Introduction}
The measurement and control of light produced by quantum systems have been the focus of interest of cavity quantum electrodynamics \cite{Walther:2006,Kavokin:2007}. Specially, the emission of light powered by solid-state devices coupled to nanocavities is an extensive area of research due to its promising technological applications, such as infrared and low-threshold lasers \cite{Altug:2006,Mu:1992}, single and entangled photon sources \cite{Stevenson:2006,Stace:2003}, as well as various applications in quantum cryptography \cite{Gisin:2002}, and quantum information \cite{Monroe:2002}. Experiments with semiconductor quantum dots (QDs) embedded in microcavities have revealed a plethora of quantum effects and offer desirable properties for harnessing coherent quantum phenomena at the single photon level. For example, the Purcell enhancement \cite{Todorov:2007}, photon anti-bunching \cite{Wiersig:2009}, vacuum Rabi splitting \cite{Khitrova:2006} and strong light matter coupling \cite{Reithmaier:2004}. These and
many others quantum phenomena are being confirmed experimentally by observing the power spectral density of the light (PSD) emitted by quantum-dot-cavity systems (QD-Cavity). Thus, the PSD or so-called emission spectrum is the only relevant information about the system which allows to study the properties of light via measurements on correlations functions as stated by the Wiener-Khintchine theorem \cite{Mandel:1997}. In order to compute the absorption or emission spectrum in open quantum systems, more precisely, in QD-Cavity systems different approaches have been developed from theoretical point of view. For example, the method of the thermodynamic Green functions which is applied to the determination of the susceptibilities and absorption spectra of atomic systems embbeded in nanocavities \cite{Jedrkiewicz:2000}, and the time-resolved photoluminescence approach whose application allows to determine the emission spectrum by consideration of an additional subsystem called the photon reservoir \cite{Hieu:2010}. However, these methods have their own approximations and restrictions and therefore are not widely used. Frequently, the emission spectrum in QD-Cavity systems is computed through the Quantum Regression Theorem (QRT) \cite{Walls:1994,Lax:1966,Swain:1981}, since it relates the evolution of mean values of observables and the two-time correlation functions. It is worth mentioning that this approach can be difficult to
implement in a computer program, it due to that computational complexity of QRT approach increases significantly as the number of
QDs or modes inside the cavity, and the dimensionalities of the Hilbert spaces are large. In general, this approach is time-consuming due to that it requires to solve a large system of coupled differential equations, and numerical instabilities can arise. Moreover, theoretical complications can appears related to dynamics of the operators involved, as we will point out in the next section. In spite of this, the QRT approach is widely used for theoretical works, for example, in studies of the luminescence spectra of coupled light-matter systems in microcavities in the presence of a continuous and incoherent pumping \cite{delValle:2009,Quesada:2011}, and the relation between dynamical regimes and entanglement in QD-Cavity systems \cite{Vera:2009,Ishida:2013}. In the past, the Green's functions technique (GFT) was successfully applied for calculation of the micromaser spectrum \cite{Quang:1993}, as a methodology in which the two-time correlation function is treated as a Green's function that decays as the off-diagonal elements of the reduced density matrix of the system for a very specific initial conditions. Nevertheless, this approach has not been widely noticed in many significant situations in open quantum systems. Possibly, it is due to their work having a limitation of implementation.
The purpose of this work is to present a simple, but efficient numerical method based on QRT formalism which overcome the inherent difficulties associated to the direct application of the QRT, by solving the dynamics of the system in the frequency domain directly. This paper is organized as follows: Section \ref{sec:QRT} review the theoretical background of quantum regression theorem and its relationship with the Green's functions technique. Section \ref{sec:appl} deal with a concrete application of our proposed method for computing the emission spectra of a dissipative QD-Cavity system. In section \ref{results} we show the numerical calculations of the emission of spectrum for the cavity and the quantum dot from both GFT and QRT approaches. Finally, we conclude in the last section.
\section{Theoretical background}\label{sec:QRT}
One of the most important measurements when the light excites resonantly a QD-Cavity system is the emission spectrum of the system. From theoretical point of view, it is assumed that corresponds to a stationary and ergodic process which can be calculated as a PSD of light using the well-known Wiener-Khintchine theorem \cite{Mandel:1997}. It states that the emission spectrum is given by the Fourier Transform of the correlation function (two-time expectation value) of the operator field $\hat{c}$,
\begin{equation}
S(\omega)=\frac{1}{\pi n_c}\mathfrak{Re}\lim_{t\to\infty}\int_{0}^\infty K(\tau) e^{i\omega \tau} d\tau,
\end{equation}
where $K(\tau)=\langle \hat{c}^{\dagger}(t+\tau)\hat{c}(t)\rangle$ and the normalizing factor is the population $n_c$ at the steady-state. In order to calculate the two-time expectation value is frequently used the QRT which states that if a set of operators $\{\hat{O}_{j}(t+\tau)\}$ satisfy
the dynamical equations $\frac{d}{d\tau}\prom{\hat{O}_{i}(t+\tau)}=\sum_{j}L_{ij}\prom{\hat{O}_j(t+\tau)}$
then $\frac{d}{d\tau}\prom{\hat{O}_{i}(t+\tau)\hat{O}(t)}=\sum_{j}L_{ij}\prom{\hat{O}_j(t+\tau)\hat{O}(t)}$ is valid for any operator $\hat{O}(t)$ at arbitrary time $t$. It is worth mentioning that vality of this theorem holds whenever a closed set of operators are involved in the dynamics. In general, to obtain the closed set of operators can be difficult or an impossible task, since it must be added as many operators as necessary in order to close the dynamics of the system. For example, in order to compute the emission of spectrum in a simple model of QD-Cavity system \cite{Quesada:2011,Vera:2009} two new operadors are required due to that the field operators in the interaction picture does not lead to a complete set. Before we consider the Green's functions technique, we will briefly describe the calculation of the QRT in an alternate form which will be the starting point in the following section. Lets consider a system operador $\hat{A}$ which does not operate on the reservoir, then its single-time expectation value in the Heisenberg picture is given by
\begin{equation}\label{sec:QRT:01}
\prom{\hat{A}(t+\tau)}=Tr_{S\otimes R}[\hat{A}(t+\tau)\hat{\rho}_{S\otimes R}(t)].
\end{equation}
The operator $\hat{\rho}_{S\otimes R}(t)=\hat{\rho}_{S}(t)\otimes\hat{\rho}_{R}(t)$ depics the composite density operador of the system and reservoir. It is worth pointing out that the dynamics of the system depends directly on $\hat{\rho}_{S\otimes R}(t)$ for all times, but the validity of the Markovian approximation requires that the state of the system is sufficiently well described by $\hat{\rho}_{S}(t)=Tr_R(\hat{\rho}_{S\otimes R}(t))$, therefore it is sufficient to write $\hat{\rho}_{S\otimes R}(t)=\hat{\rho}_{S}(t)\otimes\hat{\rho}_{R}(t)$. In what follows, we change to the Schr\"odinger representation using $\hat{A}(t+\tau)=\hat{U}^{\dagger}(t+\tau,t)\hat{A}(t)\hat{U}(t+\tau,t)$ with $\hat{U}(t+\tau,t)$ being the unitary time-evolution operator, and after tracing over degrees of freedom of the reservoirs, we have
\begin{equation}\label{sec:QRT:02}
\prom{\hat{A}(t+\tau)}=Tr_{S}[\hat{A}(t)\hat{\rho}_{S}(t+\tau)],
\end{equation}
where the reduced density operador for the system is given by $\hat{\rho}_S(t+\tau)=Tr_{R}[\hat{U}(t+\tau,t)\hat{\rho}_{S\otimes R}(t)\hat{U}^{\dagger}(t+\tau,t)]$. Then, if the $\hat{\rho}_S(t+\tau)$ satisfies the Markovian master equation $d\hat{\rho}_S(t+\tau)/d\tau=\mathcal{L}\hat{\rho}_S(t+\tau)$ with $\mathcal{L}$ the Liouvillian superoperator, the evolution of $\prom{\hat{A}(t+\tau)}$ can be computed by solving the dynamics of the master equation. To calculate the two-time correlation function $\prom{\hat{A}(t+\tau)\hat{B}(t)}$ where $\hat{A}(t+\tau)$ and $\hat{B}(t)$ are arbitrary Heisenberg operators, we proceed in a similar manner, it is,
\begin{eqnarray}\label{sec:QRT:03}
\prom{\hat{A}(t+\tau)\hat{B}(t)}&=&Tr_{S\otimes R}[\hat{A}(t+\tau)\hat{B}(t)\hat{\rho}_{S\otimes R}(t)],\notag \\
&=&Tr_S[\hat{A}(t)\hat{G}(t+\tau)],
\end{eqnarray}
where we have used the well-known composition and inversion properties of the evolution operator. Then, the two-time operator is given by
\begin{equation}\label{sec:QRT:04}
\hat{G}(t+\tau)=Tr_{R}[\hat{U}(t+\tau,t)\hat{B}(t)\hat{\rho}_{S\otimes R}(t)\hat{U}^{\dagger}(t+\tau,t)].
\end{equation}
By comparison of the Eq.~(\ref{sec:QRT:02}) and Eq.~(\ref{sec:QRT:03}), we find that $\hat{G}(t+\tau)$ is an operator that obeys the same dynamical equations as $\hat{\rho}_S(t+\tau)$, but as function of $\tau$. It is, $d\hat{G}(t+\tau)/d\tau=\mathcal{L}\hat{G}(t+\tau)$ with the boundary condition $\hat{G}(t)=\hat{B}(t)\hat{\rho}_S(t)$ at arbitrary time $t$. Hence, in the long-time limit the QRT reads,
\begin{equation}\label{eq-final}
\lim_{t\to\infty}\prom{\hat{A}(t+\tau)\hat{B}(t)}=Tr_S[\hat{A}\hat{G}(\tau)]
\end{equation}
where $\hat{G}(\tau)=Tr_{R}[\hat{U}(\tau)\hat{B}\hat{\rho}^{(ss)}_{S\otimes R}\hat{U}^{\dagger}(\tau)]$ is the Green's functions operator, and the operators $\hat{A}$, $\hat{B}$ and $\hat{\rho}^{(ss)}_{S\otimes R}$ are written in the Schr\"odinger representation. The superscript "(ss)" refers to the steady state of the reduced density operator of the system. After taking the Laplace transform on Eq.~(\ref{eq-final}), we obtain an expression for the emission of spectrum in terms of the Green's functions operator, it is,
\begin{equation}
S(\omega)=\frac{1}{\pi n_c}\mathfrak{Re}\,Tr_S[\hat{A}\hat{\tilde{G}}(i\omega)].
\end{equation}
Prior to leaving this section, we mention that this equation will be considered for computing the emission spectrum due to the cavity as well as the quantum dot, e.g. by considering the photon and fermionic operators in a separated way. Therefore, we will describe in the next subsetion a general approach that can be applied for both cases.
\subsection{Algorithm for the Green's functions technique}
Before to describe a simple algorithm for calculating the emission spectrum, we take into account that the dynamics for both opertors $\hat{G}(\tau)$ and $\hat{\rho}_S(\tau)$ are governed by the same Master equation, i.e., $d\hat{G}(\tau)/d\tau=\mathcal{L}\hat{G}(\tau)$ with $\mathcal{L}$ the Liouvillian superoperator, that efectivelly has a larger tensor rank than the reduced density operator of the system. So, we can write the dynamical equations for the Green's functions operator in a component form:
\begin{equation}\label{components}
\frac{dG_{\tilde{\alpha}}(\tau)}{d\tau}=\sum_{\tilde{\beta}}\mathcal{L}_{\tilde{\alpha}\tilde{\beta}}G_{\tilde{\beta}}(\tau),
\end{equation}
together with the initial condition $G_{\tilde{\beta}}(0)$. The symbol $\tilde{\alpha}$ is a composite index for labeling the states of the reduced density operator of the system, e.g. for indexing both matter and photon states in the QD-Cavity system, see section \ref{sec:appl} for example. Hence, $G_{\tilde{\beta}}$ and $\mathcal{L}_{\tilde{\alpha}\tilde{\beta}}$ acts as a column vector and a matrix in this notation. (ii) To obtain the solution to the Eq.~(\ref{components}) in frequency domain via the Laplace Transform, it is $-\tilde{G}_{\tilde{\alpha}}(0)=\sum_{\tilde{\beta}}(\mathcal{L}_{\tilde{\alpha}\tilde{\beta}}-i\omega\delta_{\tilde{\alpha}\tilde{\beta}})\tilde{G}_{\tilde{\beta}}(i\omega),$ (iii) We perform the invertion of the matrix
$\mathcal{M}_{\tilde{\alpha}\tilde{\beta}}=(i\omega\delta_{\tilde{\alpha}\tilde{\beta}}-\mathcal{L}_{\tilde{\alpha}\tilde{\beta}}$) and finally, the spectrum of emission is computed in terms of the initial conditions given by,
\begin{equation}
\tilde{G}_{\tilde{\beta}}(i\omega)=\sum_{\tilde{\alpha}}\mathcal{M}^{-1}_{\tilde{\beta}\tilde{\alpha}}\tilde{G}_{\tilde{\alpha}}(0).
\end{equation}
These initial conditions are easily obtained by evaluating the Green's function operator at $\tau=0$.
\section{Application to the quantum dot-cavity system}\label{sec:appl}
\subsection{Model}\label{sec:model}
In order to apply our proposed method for calculating the emission spectrum in QD-Cavity system, we will consider a simple but illustrative system composed of a quantum dot interacting with a confined mode of the electromagnetic field inside a semiconductor cavity. This quantum system is well described by the Jaynes-Cummings Hamiltonian \cite{Cummings:1963}
\begin{equation}\label{eq:Hsystem}
\hat{H}_{S}=\omega_X\hat{\sigma}^{\dagger}\hat{\sigma}+(\omega_X-\Delta)\hat{a}^{\dagger}\hat{a}+g(\hat{\sigma}\hat{a}^{\dagger}+\hat{a}\hat{\sigma}^{\dagger}),
\end{equation}
where the quantum dot is described as a fermionic system with only two possible states, $\ket{G}$ and $\ket{X}$ are the ground and excited state, respectively. $\hat{\sigma}=|G\rangle\langle X|$ and $\hat{a}$ ($\hat{\sigma}^{\dagger}=|X\rangle\langle G|$ and $\hat{a}^\dagger$) are the annihilation (creation) operators for the fermionic system and the cavity mode. $g$ is the light-matter coupling constant, and we have set $\hbar=1$. We also define the detuning between frequencies of the quantum dot and the cavity mode as $\Delta=\omega_X-\omega_a$, moreover $\omega_X$ is the energy to create an exciton and $\omega_a$ is the energy associated to the photons inside de cavity, respectively. This Hamiltonian system is far away for describing any real physical situation since it is completely integrable \cite{Scully:1996} and no measurements could be done since the light remains always inside the cavity.\\
In order to include the effects of environment on the dynamics of the system, we consider the usual approach to model an open quantum system by considering a whole system-reservoir hamiltonian which is frequently splitted in three parts. The first part corresponds to the system of quantum dot-microcavity. The second part is the hamiltonian of the reservoirs and finally, the third part which is a bilinear coupling between the system and the reservoirs \cite{Perea:2004}. After tracing out the degrees of freedom of all the reservoirs and assuming the validity of the Born-Markov approximation, one arrives to a master equation for the reduced density matrix of the system,
\begin{eqnarray}\label{eq:master
\frac{d\hat{\rho}_{S}}{dt}&=&i\left[\hat{\rho}_{S},\hat{H}_{S}\right]+\frac{\kappa}{2}(2 \hat{a} \hat{\rho}_{S} \hat{a}^{\dagger}-\hat{a}^{\dagger} \hat{a} \hat{\rho}_{S}-\hat{\rho}_{S} \hat{a}^{\dagger}\hat{a})\notag\\&+&\frac{\gamma}{2}(2 \hat{\sigma} \hat{\rho}_{S} \hat{\sigma}^{\dagger}-\hat{\sigma}^{\dagger} \hat{\sigma} \hat{\rho}_{S}-\hat{\rho}_{S} \hat{\sigma}^{\dagger}\hat{\sigma})\notag \\
&+&\frac{P}{2}(2 \hat{\sigma}^{\dagger} \hat{\rho}_{S} \hat{\sigma}-\hat{\sigma} \hat{\sigma}^{\dagger} \hat{\rho}_{S}-\hat{\rho}_{S} \hat{\sigma}\hat{\sigma}^{\dagger}).
\end{eqnarray}
Where $\gamma$ is the decay rate due to the spontaneous emission, $\kappa$ is the decay rate of the cavity photons across the cavity mirrors, and $P$ is the rate at which the excitons are being pumped. Fig.~\ref{systemscheme} shows a scheme of the simplified model of the QD-cavity system showing the processes of continuous pumping $P$ and cavity loses $\kappa$. The physical process begin when the light from the pumping laser enters into the cavity and excites one of the quantum dots in the QD layer. Thus, light from this source couples to the cavity and a fraction of photons escapes through the partly transparent mirror from the cavity and goes to the spectrometer for measurements of the emission of spectrum.\\ A general approach for solving the dynamics of the coupled system, consist of writting the Bloch equations for the reduced density matrix of the system in the bared basis. It is, an extended Hilbert space formed by taking the tensor product of the state vectors for each of the system components, ${\left\lbrace|G\rangle,|X\rangle\right\rbrace}\otimes \{{|n\rangle}\}^\infty_{n=0}$. In this basis, the reduced density matrix $\hat{\rho}_{S}$ can be written in terms of its matrix elements as $\rho_{S \alpha n,\beta m}=\langle\alpha n|\hat{\rho}_{S} |\beta m\rangle$. Hence, the Eq.~\eqref{eq:master} explicitly reads,
\begin{eqnarray}\label{eq:maestra}
\frac{d\rho_{S \alpha n,\beta m}}{d\tau}&=&i\Big[(\omega_X-\Delta)(m-n)\rho_{S \alpha n,\beta m}\notag\\&+&\omega_X(\delta_{\beta X}\rho_{S \alpha n,Xm}-\delta_{\alpha X}\rho_{S Xn,\beta m})\Big]\notag\\&+&ig\Big[\Big(\sqrt{m+1}\delta_{\beta X}\rho_{S \alpha n,G m+1}\notag\\&+&\sqrt{m}\delta_{\beta G}\rho_{S \alpha n,X m-1}\Big)\notag\\&-&\Big(\sqrt{n}\delta_{\alpha G}\rho_{S X n-1,\beta m}\notag\\&+&\sqrt{n+1}\delta_{\alpha X}\rho_{S G n+1,\beta m}\Big)\Big]
\notag\\&+&\frac{\kappa}{2}\Big(2\sqrt{(m+1)(n+1)}\rho_{S \alpha n+1,\beta m+1}\notag \\ &-&(n+m)\rho_{S \alpha n,\beta m}\Big)
-\frac{\gamma}{2}\Big(\delta_{\alpha X}\rho_{S X n,\beta m}\notag \\&-&2\delta_{\alpha G}\delta_{\beta G}\rho_{S X n,X m}+\delta_{\beta X}\rho_{S \alpha n,X m}\Big)\notag\\&+&\frac{P}{2}\Big(2\delta_{\alpha X}\delta_{\beta X}\rho_{S Gn,Gm}-\delta_{\alpha G}\rho_{S Gn,\beta m}\notag\\&-&\delta_{\beta G}\rho_{S \alpha n,Gm}\Big).
\end{eqnarray}
Note that we use the convention that all indices written in greek alphabet are used for matter states and take values $\ket{G}$, $\ket{X}$, and the indices written in latin alphabet are used for Fock states and take values $0,1,2,3\dots$. Additionally, it is worth mentioning that our proposed method does not require to solve a system of coupled differential equations, instead of it, we solve a reduced set of algebraic equations that speed up the numerical solution. \\Prior to leaving this section, we point out that the number of excitations of the system is defined by the operator $\hat{N}=\hat{a}^{\dagger}\hat{a}+\hat{\sigma}^{\dagger}\hat{\sigma}$. The closed system and the number of excitations of the system is conserved, i.e., $[\hat{H}_S, \hat{N}]=0$. It allows us to organize the states of the system through the number of excitations criterion such that the density matrix elements $\rho_{Gn,Gn}$, $\rho_{Xn-1,Xn-1}$ $\rho_{Gn,Xn-1}$ and $\rho_{Xn-1,Gn}$ are related by having the same number of quanta. It
is, subspaces of a fixed number of excitation evolve independently from each other. The Fig.~\ref{qs} shows a schematic representation of the action of the dissipative processes involved in the dynamics of the system according to the excitation number ($N_{exc}$).
\begin{figure}[h]
\centering
\includegraphics[scale=.19]{system5.png}
\caption{Scheme of the simplified model of the QD-cavity system showing the processes of continuous pumping $P$ and cavity loses $\kappa$.}\label{systemscheme}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[scale=.5]{escalera3.pdf}
\caption{Ladder of bared states for a two level quantum dot coupled to a single cavity mode. The double headed green arrow depics the matter coupling constant $g$, dashed red lines the emission of the cavity mode $\kappa$, solid black lines the exciton pumping rate $P$ and solid blue lines the spontaneous emission rate $\gamma$.}\label{qs}
\end{figure}
\subsection{Emission spectrum of the cavity}\label{cavityspectrum}
\noindent
In order to compute the emission spectrum of the cavity, we will consider the two-time correlation function according to the Eq.~(\ref{eq-final}) for the photon operator as follows:
\begin{equation}
K(\tau)=\lim_{t\to\infty}\langle \hat{a}^{\dagger}(t+\tau)\hat{a}(t)\rangle.
\end{equation}
After performing the partial trace over the degrees of freedom of the system we have that
\begin{eqnarray}\label{ka}
K(\tau)&=&\sum_{\alpha,\beta,\gamma,l,m,n}\sqrt{(l+1)(m+1)}Tr_{R}[U_{\alpha l,\beta m}(\tau)\notag \\ &\times&\langle\beta m+1|\hat{\rho}^{(ss)}_{S\otimes R}|\gamma n\rangle U^{\dagger}_{\gamma n,\alpha l+1}(\tau)],
\end{eqnarray}
where the matrix elements for the time evolution operator are given by $U_{\alpha l,\beta m}(\tau)=\langle \alpha l|\hat{U}(\tau)|\beta m\rangle$ and $U^{\dagger}_{\gamma n,\alpha l+1}(\tau)=\langle \gamma n|\hat{U}^{\dagger}(\tau)|\alpha l+1\rangle$. In what follows, we assume the validity of the Markovian approximation, it means that the correlations between the system and the reservoir must be unimportant even at the steady state. Thus, the density operator system-reservoir can written as $\hat{\rho}^{(ss)}_{S\otimes R}=\hat{\rho}^{(ss)}_{S}\otimes\hat{\rho}^{(ss)}_{R}$ which implies that
\begin{equation}\label{densitySR}
\langle\beta m+1|\hat{\rho}^{(ss)}_{S\otimes R}|\gamma n\rangle = \hat{\rho}^{(ss)}_{R}\langle\beta m+1|\hat{\rho}^{(ss)}_{S}|\gamma n\rangle.
\end{equation}
Replacing the previous expression in Eq.~(\ref{ka}), it is straightforward to shows that the two-time correlation function reads
\begin{equation}\label{corr}
K(\tau)=\sum_{\alpha l}\sqrt{l+1}\langle\alpha l|\hat{G}(\tau)|\alpha l+1\rangle,
\end{equation}
where the Green's functions operator $\hat{G}(\tau)$ is given by
\begin{eqnarray}\label{op.green}
\hat{G}(\tau)&=&Tr_{R}\Big[\hat{U}(\tau)\hat{\rho}^{(ss)}_R\sum_{\beta \gamma m n}\Big(\sqrt{m+1}\ket{\beta m}\bra{\gamma n}\notag\\
&\times&\bra{\beta m+1}\hat{\rho}^{(ss)}_S\ket{\gamma n}\Big)\hat{U}^{\dagger}(\tau)\Big].
\end{eqnarray}
As we pointed out in section \ref{sec:QRT}, this operator must obey the same master equation as the reduced density operator of the system. In fact, the terms that only contribute in the Eq.~(\ref{corr}) are given by the matrix elements $G_{\beta m, \gamma n}(\tau)\equiv\bra{\beta m}\hat{G}(\tau)\ket{\gamma n}$ of the Green's functions operator. This is due to the fact that the projection operator $\ket{\beta m}\bra{\gamma n}$ enter into $\hat{G}(\tau)$ in the same way as into the reduced density operator of the system.\\
In order to identify these matrix elements, it should consider that for the QD-Cavity system the dynamics of the all coherences asymptotically vanish and remains only the reduced density matrix elements which are ruled by the number of excitations criterion, i.e. $\rho_{Gn,Gn}$, $\rho_{Xn-1,Xn-1}$, $\rho_{Gn,Xn-1}$, $\rho_{Xn-1,Gn}$. Then, the Eq.~(\ref{densitySR}) can be written as follows,
\begin{eqnarray} \label{diagonalrho}
\langle\beta m+1|\hat{\rho}^{(ss)}_{S\otimes R}|\gamma n\rangle &=& \hat{\rho}^{(ss)}_{R}\Big(\delta_{\beta G}\delta_{\gamma G}\delta_{m+1,n}\notag \\
&+& \delta_{\beta X}\delta_{\gamma X}\delta_{m,n-1}+\delta_{\beta G}\delta_{\gamma X}\delta_{m,n}\notag \\
&+& \delta_{\beta X}\delta_{\gamma G}\delta_{m+1,n-1}\Big)\rho^{(ss)}_{S\beta m+1,\gamma n}.\notag\\
\end{eqnarray}
By replacing the Eq.~(\ref{diagonalrho}) into Eq.~(\ref{op.green}) we find that the Green's functions operator explicitly reads
\begin{eqnarray}
\hat{G}(\tau)&=&Tr_R\Big[\hat{U}(\tau)\hat{\rho}^{(ss)}_R\sum_{m}\sqrt{m+1}\Big(\ket{G m}\bra{Gm+1}\notag \\
&\times& \rho^{(ss)}_{S Gm+1,Gm+1}+\ket{X m}\bra{X m+1}\rho^{(ss)}_{S X m+1,X m+1}\notag \\&+&\ket{G m}\bra{X m}\rho^{(ss)}_{S Gm+1,Xm}+\ket{X m}\bra{G m+2}\notag \\&\times&\rho^{(ss)}_{S X m+1,Gm+2}\Big)\hat{U}^{\dagger}(\tau)\Big].
\end{eqnarray}
Note that from this expression is easy to identify the nonzero matrix elements of the Green's functions operator that contribute to the emission spectrum. Finally, after performing the Laplace transform we have that the emission spectrum of the cavity is given by
\begin{eqnarray}
S(\omega)&=&\frac{1}{\pi n_c}\sum_l\sqrt{l+1}\Big(\tilde{G}_{Gl,Gl+1}(i\omega)\notag\\
&+&\tilde{G}_{Xl,Xl+1}(i\omega)+\tilde{G}_{Gl,Xl}(i\omega)\notag \\
&+&\tilde{G}_{Xl,Gl+2}(i\omega)\Big).
\end{eqnarray}
It is worth mentioning that the initial conditions may be obtained by evaluating the Green's function operator at $\tau=0$, then using the fact that the time evolution operators become the identity and $Tr_R[\hat{\rho}^{(ss)}_R]=1$, we obtain a set of initial conditions given by
\begin{eqnarray}\label{greencondini}
\tilde{G}_{Gl,Gl+1}(0)&=&\sqrt{l+1}\rho^{(ss)}_{S Gl+1,Gl+1},\notag \\
\tilde{G}_{Xl,Xl+1}(0)&=&\sqrt{l+1}\rho^{(ss)}_{S Xl+1,Xl+1},\notag \\
\tilde{G}_{Gl,Xl}(0)&=&\sqrt{l+1}\rho^{(ss)}_{S Gl+1,Xl},\notag \\
\tilde{G}_{Xl,Gl+2}(0)&=&\sqrt{l+1}\rho^{(ss)}_{S Xl+1,Gl+2}.
\end{eqnarray}
Note that this set of initial conditions corresponds to the asymptotic solution of the Bloch equations for the reduced density matrix of the system.
\subsection{Emission spectrum of the quantum dot}
\noindent
In order to compute the emission spectrum of the quantum dot, we will consider the two-time correlation function given by Eq.~(\ref{eq-final}), but for the case of the matter operator:
\begin{equation}
K(\tau)=\lim_{t\to\infty}\langle \hat{\sigma}^{\dagger}(t+\tau)\hat{\sigma}(t)\rangle.\notag
\end{equation}
It is straightforward to show after performing the partial trace over the degrees of freedom of the system that the two-time correlation function reads
\begin{equation}\label{corr2}
K(\tau)=\sum_{\alpha l}\delta_{\alpha X}\langle G l|\hat{G}(\tau)|\alpha l\rangle,
\end{equation}
where the Green's functions operator $\hat{G}(\tau)$ is given by
\begin{eqnarray}\label{op.green2}
\hat{G}(\tau)&=&Tr_{R}\Big[\hat{U}(\tau)\sum_{\beta \gamma m n}\Big(\delta_{\beta X}\ket{G m}\bra{\gamma m}\notag\\
&\times&\bra{\beta m}\hat{\rho}^{(ss)}_{S\otimes R}\ket{\gamma n}\Big)\hat{U}^{\dagger}(\tau)\Big].
\end{eqnarray}
Assuming again the validity of the Markovian approximation and taking into account the number of excitations criterion, we have that the density operator system-reservoir can be written as:
\begin{eqnarray} \label{diagonalrho2}
\langle\beta m|\hat{\rho}^{(ss)}_{S\otimes R}|\gamma n\rangle &=& \hat{\rho}^{(ss)}_{R}\Big(\delta_{\beta G}\delta_{\gamma G}\delta_{m,n}\notag \\
&+& \delta_{\beta X}\delta_{\gamma X}\delta_{m,n}+\delta_{\beta G}\delta_{\gamma X}\delta_{m,n+1}\notag \\
&+& \delta_{\beta X}\delta_{\gamma G}\delta_{m,n-1}\Big)\rho^{(ss)}_{S\beta m,\gamma n}.\notag\\
\end{eqnarray}
By inserting the Eq.~(\ref{diagonalrho2}) into Eq.~(\ref{op.green2}) we find that the Green's functions operator explicitly reads
\begin{eqnarray}
\hat{G}(\tau)&=&Tr_R\Big[\hat{U}(\tau)\hat{\rho}^{(ss)}_R\sum_{m}\Big(\ket{G m}\bra{Xm}\rho^{(ss)}_{S Xm,Xm}\notag \\
&+&\ket{G m}\bra{G m+1}\rho^{(ss)}_{S X m,G m+1}\Big)\hat{U}^{\dagger}(\tau)\Big].
\end{eqnarray}
Analogously as in section~\ref{cavityspectrum}, we identify the nonzero matrix elements of the Green's functions operator that contribute to the emission spectrum and after performing the Laplace transform the emission spectrum of the quantum dot is given by
\begin{equation}
S(\omega)=\frac{1}{\pi n_{\sigma}}\sum_l\Big(\tilde{G}_{Gl,Xl}(i \omega)+\tilde{G}_{Gl,Gl+1}(i\omega)\Big).
\end{equation}
where $n_{\sigma}=\langle\hat{\sigma}^{\dagger}\hat{\sigma} \rangle$ is the normalizing factor at the steady-state.
Taking into account that the initial conditions are obtained by evaluating the Green's function operator at $\tau=0$, we have the time evolution operators become the identity and $Tr_R[\hat{\rho}^{(ss)}_R]=1$, thus, we obtain a set of initial conditions given by
\begin{eqnarray}
\tilde{G}_{Gl,Xl}(0)&=&\rho^{(ss)}_{S Xl,Xl},\notag \\
\tilde{G}_{Gl,Gl+1}(0)&=&\rho^{(ss)}_{S Xl,Gl+1},\notag \\
\tilde{G}_{Xl,Xl+1}(0)&=&0,\notag \\
\tilde{G}_{Gl+2,Xl}(0)&=&0.
\end{eqnarray}
\section{Results and Discussion}\label{results}
In this section, we compare the numerical calculations based on GFT and QRT approach for the emission spectrum of the cavity as well as the quantum dot. Due to that the QD-Cavity system can display two different dynamical regimes by changing the values of the free parameters of the system and transitions between these two regimes can be achieved when the loss and pump rates are modified. Particularly, in the strong coupling regime the relation $P/\kappa\ll g$ holds and the relation $P/\kappa\gg g$ remains valid for the weak coupling regime. Fig.~\ref{panel2} shows the numerical calculations of the emission spectrum due to the cavity in the weak coupling regime, the parameters values are $g=1\,meV$, $\gamma=0.005\,meV$, $\kappa=0.2\,meV$, $P=0.3\,meV$, $\Delta=2\,meV$, $\omega_a=1000\,meV$. Panel (a) shows the emission spectrum for the GFT compared to the QRT approach. Panel (b) shows the quantity $\vert S(\omega)_{GFT}-S(\omega)_{QRT}\vert$ as a measurement of the error between the numerical calculations in the emission spectrum. For this set of parameters values, we can easily to identify two peaks associated to the modes of the cavity and the quantum dot, it is $\omega_a\approx998.3meV$ and $\omega_X\approx1000.3meV$, respectively. Fig.~\ref{panel1} shows the same calculations as in Fig.~\ref{panel2}, but in the strong coupling regime and the parameters values are $g=1\,meV$, $\gamma=0.005\,meV$, $\kappa=2\,meV$, $P=0.005\,meV$, $\Delta=0.0\,meV$, $\omega_a=1000\,meV$. In the case of resonance, the modes associated to the cavity and the quantum dot do not match, but
repel each other, resulting in a structure of two separate peaks a distance $2g\approx 2meV$. Fig.~\ref{panel3} shows the numerical calculations of the emission of spectrum due to the quantum dot with a high value of the rate $\kappa=5\,meV$ and a smaller, although non negligible, pumping $P=1\,meV$. The rest of parameters values are $g=1\,meV$, $\gamma=0.1\,meV$, $\Delta=5\,meV$, $\omega_a=1000\,meV$.\\
We observed that our numerical method based on GFT is in full agreement with the QRT approach and reproduces very well the spectrum of emission associated with this system. The quantity $\vert S(\omega)_{GFT}-S(\omega)_{QRT}\vert$ shows the discrepancy between both methods which is the order of $10^{-3}-10^{-2}$ as it is seen in Fig.~\ref{panel1} and Fig.~\ref{panel2}. Mainly, the discrepancy between both methods is due to the numerical errors accumulated in the numerical integration of the Bloch equations in the QRT approach, it causes some differences in the spectrum with respect to the results computed by the GFT. Note that, there is no integration of any equations in the GFT, therefore, we expect a more accurate spectrum of emission. We mention that for the numerical calculations based on QRT approach, we have followed the Ref.~\cite{Vera:2009}.
In order to test the performance of the numerical method, we regard four calculation times for computing the emission spectrum of the cavity based on GFT and QRT approach at different excitation numbers. Table~\ref{table01} shows in first column the excitation number, i.e. the truncation level in the bare-state basis for the numerical calculations involved. Second and third column show the results of elapsed time in seconds for both the GFT and QRT approach, respectively. Note that for comparison purposes all numerical calculations were performed at the same truncation level i.e. $N_{exc}=10$. Additionally, we have solved numerically the Bloch equations (see Eq.~(\ref{eq:maestra})) until time $t_{max}=2^{17}$ in order to obtain a good resolution in the frequency domain for the QRT approach, i.e. $\Delta\omega\approx0.048$. Hence, we have evaluated the emission spectrum for the GFT in a grid with the same resolution in the frequency domain (we emphasize that QRT approach is time consuming due to the number of coupled differential equations to be solved, rather than the number of evaluations in the grid size used). In addition, the numerical calculations were carried out with the same parameters values as in Fig.~\ref{panel1} for both GFT and QRT approach. We found that our numerical approach based on GFT is very efficient and accurate for calculating the emission spectrum in QD-Cavity systems. Moreover, this method can easily be implemented in the numerical linear algebra packages as well as in any programming language.
\begin{table}[h]
\caption{\label{table01} Comparison of calculation times between the Green's Functions Technique (GFT) and the Quantum Regression Theorem (QRT) in the numerical calculation of the emission spectrum of the cavity. The calculations were made using a commercial Intel(R) Core(TM) $i7-4770$ processor of $3.4$\,GHz $\times8$, and $12$ GB RAM.}
\begin{center}
\begin{tabular}{cccc}
\hline\hline
Excitation & Calculation times & Calculation times & \\
number & for the GFT (s)& for the QRT (s) & \\
\hline
$5$ &$\,\,\,\,\,\,\,0.4$ & $\,\,\,\,\,\,\,92.5$ & \\
$10$ &$\,\,\,\,\,\,\,2.0$ & $\,\,\,\,273.4$ & \\
$20$ &$\,\,\,\,14.0$ & $\,\,\,\,390.2$ & \\
$40$ &$\,100.2$ & $\,\,\,\,\,673.8$ & \\
\hline \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[h!]
\centering
\includegraphics[scale=.44]{b2.pdf}
\caption{Panel (a) shows a comparison of the emission spectrum of the cavity. The numerical calculation based on the Green's functions technique (GFT) is shown as solid blue line and the corresponding numerical calculation based on quantum regression theorem (QRT) approach is shown as dashed red line. Panel (b) shows in solid blue line the quantity $\vert S(\omega)_{GFT}-S(\omega)_{QRT}\vert$ as a measure of the difference in the numerical calculations of the emission spectrum of the cavity between two methods.
}\label{panel2}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[scale=.44]{b1.pdf}
\caption{Panel (a) shows a comparison of the emission spectrum of the cavity. The numerical calculation based on the Green's functions technique (GFT) is shown as solid blue line and the corresponding numerical calculation based on quantum regression theorem (QRT) approach is shown as dashed red line. Panel (b) shows in solid blue line the quantity $\vert S(\omega)_{GFT}-S(\omega)_{QRT}\vert$ as a measure of the difference in the numerical calculations of the emission spectrum of the cavity between two methods.
}\label{panel1}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=.44]{b3.pdf}
\caption{Panel (a) shows a comparison of the emission spectrum of the quantum dot. See text and caption of Fig.~\ref{panel1} for details.}\label{panel3}
\end{figure}
\section {Conclusions}\label{sec:conclusions}
We have developed the Green's function technique as an alternative methodology to the QRT for calculating the two-time correlation functions in open quantum systems. In particular, we have shown the performance of the Green's function technique by calculating the emission spectrum in an open quantum system composed by a quantum dot embedded in a microcavity. This theoretical approach is rather general and allows to overcome the inherent theoretical difficulties presented in the direct application of the QRT, i.e., to find a closure condition on the set of operators involved in the dynamics equations, by considering that all coherences asymptotically vanish, and remains only the reduced density matrix elements which are ruled by the number of excitations criterion. We have shown that the Green's function technique offers several computational advantages, namely, the speeding up numerical computations via a transformation of the dynamics of the master equation in a set of linear algebraic equations, which are efficiently solvable by a numerical linear algebra routine, a faster convergence and significant reduction of computational time since the emission spectrum is calculated as a sum of terms of non-diagonal matrix elements of the reduced density operator of the system. We mention that our methodology can be extended for calculating the emission spectrum in significant situations of quantum dots in biexcitonic regime or involving coupled photonic cavities.
\section{Acknowledgements}
\noindent
This work was financed by Vicerrector\'ia de investigaciones of the Universidad del Quind\'io within the project with code $659$, and by Colciencias within the project with code $110156933525$, contract number $026-2013$, and HERMES code $17432$.
|
\section{Introduction} \label{introduction}
Random key graphs, also known as homogeneous random intersection
graphs, have been investigated widely in the literature
\cite{r1,ryb3,zz,yagan_onoff,yagan,ZhaoYaganGligor,ISIT,r10,r4}.
The notion of random key graph results from the seminal
Eschenauer--Gligor (EG) random key predistribution scheme \cite{virgil},
which is the most recognized solution to secure communication using cryptographic keys in wireless sensor networks \cite{yagan}.
The definition of a random key graph can also be generalized beyond cryptographic keys. Consider
a random key graph
$G(n, X_n, Y_n)$ that is constructed on a set of $n$ nodes as follows.
Each node is independently assigned a set of $X_n$ distinct objects,
selected {\em uniformly at random} from a pool of $Y_n$ objects,
where $X_n$ and $Y_n$ are both functions of $n$. An undirected edge exists
between two nodes if and only if they possess at least one common
object. An object is a cryptographic key in the application of random key graphs to
the Eschenauer--Gligor random key predistribution scheme. In addition to the area of secure sensor networks,
random key graphs have
also been used in various applications including cryptanalysis \cite{r10}, social networks \cite{ZhaoYaganGligor}, and recommender systems \cite{r4}.
($k$-)Connectivity of a random key graph has received much interest
\cite{r1,ryb3,zz,yagan_onoff,yagan,ZhaoYaganGligor,ISIT}.
A graph is said to be $k$-connected if it remains connected despite
the deletion of at most $(k-1)$ nodes or
edges\footnote{$k$-connectivity given here is equivalent to $k$-{\em
vertex}-connectivity, which can also be defined when \emph{only} node
failure is considered; i.e., the ability of the graph remaining
connected in spite of the removal of at most $(k-1)$ nodes. $k$-{\em
edge}-connectivity is defined similarly for graphs that are still
connected despite the failure of any $(k-1)$ edges.
It is plain to prove that $k$-vertex-connectivity implies $k$-edge-connectivity
\cite{erdos61conn}.}; an equivalent definition of $k$-connectivity
is that for each pair of nodes there exist at least $k$ mutually
disjoint paths connecting them \cite{erdos61conn}. In the case of
$k$ being 1, $k$-connectivity becomes connectivity, meaning that
each node in the graph can find at least one path to any other node,
either directly or with the help of other relaying nodes. A graph
property related to and implied by $k$-connectivity
is
that the minimum degree of the the graph is at least $k$ (i.e.,
each node is directly connected to no less than $k$ other
nodes), where the minimum degree refers to the minimum among the
numbers of neighbors that nodes have.
We investigate $k$-connectivity of random key graphs. Our
contribution is, for a random key graph, to derive the
asymptotically exact probabilities for $k$-connectivity and the
property that the {minimum degree} is at least $k$.
The rest of the paper is organized as follows. Section
\ref{sec:main:res} presents the results. We elaborate the proof of
Theorem \ref{THM_RKGk} in Section \ref{sec_est}. Section
\ref{sec:expe} provides numerical findings to support the
theoretical results. Section \ref{related} surveys related work; and
Section \ref{sec:Conclusion} concludes the paper.
%
\section{The Results} \label{sec:main:res}
For a random key graph $G(n,X_n,Y_n)$, Theorem \ref{THM_RKGk} and
Corollary \ref{COR_RKGk} below present the asymptotically exact
probabilities for $k$-connectivity and the
property of the {minimum degree} being at least $k$, where $k$ is a
positive integer and does not scale with $n$. The term $\ln$ stands for
the natural logarithm function, and $e$ is its base. We use the standard asymptotic notation $O(\cdot), o(\cdot), \Omega(\cdot),
\omega(\cdot), \Theta(\cdot), \sim$; in particular, for two
positive sequences $x_n$ and $y_n$, the relation $x_n \sim y_n$
means $\lim_{n \to
\infty} (x_n/y_n)=1$. All asymptotic statements are understood with $n \to \infty$. Also, $\mathbb{P}[\mathcal {E}]$
denotes the probability that event $\mathcal {E}$ occurs.
\begin{thm}\label{THM_RKGk} For a random key graph $G(n,X_n,Y_n)$,
let $q_n$ be the probability that there exists an edge between two
nodes. With a sequence $\alpha_n$ defined by
\begin{align}
q_n &
= \frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n},
\label{thm_eq_ps}
\end{align}
then under
$ X_n \geq 2$, it follows that
\begin{align}
& \lim_{n \to \infty}\mathbb{P} \big[\hspace{2pt}G(n,X_n,Y_n)\text{
is $k$-connected}.\hspace{2pt}\big] \nonumber \\
&= \lim_{n \to \infty} \mathbb{P}
\left[\hspace{2pt}G(n,X_n,Y_n)\text{ has a minimum degree at least
}k.\hspace{2pt}\right]
\nonumber \\
&
\quad =
\begin{cases} 0, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=-\infty$}, \\ 1, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=\infty$,} \\ e^{- \frac{e^{-\alpha ^*}}{(k-1)!}},
&\text{ if $\lim_{n \to \infty}{\alpha_n}
=\alpha ^* \in (-\infty, \infty)$.} \end{cases}
\label{doublexp}
\end{align}
\end{thm}
We have the following corollary by replacing the condition (\ref{thm_eq_ps}) on the edge probability $q_n$ with a condition on \vspace{1.7pt} the asymptotics $\frac{{X_n}^2}{Y_n} $ of $q_n$ (formally, $q_n \sim \frac{{X_n}^2}{Y_n}$ holds \vspace{1pt} under $\frac{{X_n}^2}{Y_n} = o(1)$; see \cite[Lemma 8]{ZhaoYaganGligor}.)
\begin{cor}\label{COR_RKGk} For a random key graph $G(n,X_n,Y_n)$,
with a sequence $\beta_n$ defined by
\begin{align}
\frac{{X_n}^2}{Y_n} &
= \frac{\ln n + {(k-1)} \ln \ln n + {\beta_n}}{n},
\label{thm_eq_psXnPn}
\end{align}
then under
$ X_n \geq 2$, it follows that
\begin{align}
& \lim_{n \to \infty}\mathbb{P} \big[\hspace{2pt}G(n,X_n,Y_n)\text{
is $k$-connected}.\hspace{2pt}\big] \nonumber \\
& = \lim_{n \to \infty} \mathbb{P}
\left[\hspace{2pt}G(n,X_n,Y_n)\text{ has a minimum degree at least
}k.\hspace{2pt}\right]
\nonumber \\
&
\quad =
\begin{cases} 0, &\text{ if $\lim_{n \to \infty}{\beta_n}
=-\infty$}, \\ 1, &\text{ if $\lim_{n \to \infty}{\beta_n}
=\infty$,} \\ e^{- \frac{e^{-\beta ^*}}{(k-1)!}},
&\text{ if $\lim_{n \to \infty}{\beta_n}
=\beta ^* \in (-\infty, \infty)$.} \end{cases}
\nonumber
\end{align}
\end{cor}
\begin{rem} \label{rem1}
From Lemma \ref{lem-only-prove-lnlnn-1} (resp., Lemma \ref{lem-only-prove-lnlnn-2}) in the Appendix, we can introduce an
extra condition $\alpha_n = \pm O(\ln \ln n) = \pm o(\ln n)$ (resp., $\beta_n = \pm O(\ln \ln n) = \pm
o(\ln n)$) in proving Theorem
\ref{THM_RKGk} (resp., Corollary \ref{COR_RKGk}).
\end{rem}
\begin{rem}
In Theorem
\ref{THM_RKGk} and Corollary \ref{COR_RKGk}, since the results are in the asymptotic sense,
the conditions only need to hold for all $n$ sufficiently
large.
\end{rem}
Establishing Corollary \ref{COR_RKGk} given Theorem
\ref{THM_RKGk} is straightforward and is given in the Appendix. Below we explain how to obtain Theorem
\ref{THM_RKGk}. Since a necessary condition for a graph to be $k$-connected is that the minimum degree is at least $k$, the proof of Theorem \ref{THM_RKGk} will be completed
once we have the following two lemmas. Lemma \ref{lemma-1} is from our prior work \cite{ZhaoYaganGligor}. Lemma \ref{lemma-2} simply reproduces the result on the minimum degree in Theorem \ref{THM_RKGk}.
\begin{lem}[\hspace{0pt}{Our work \cite[Lemma 5]{ZhaoYaganGligor}}\hspace{0pt}] \label{lemma-1}
For a random key graph $G(n,X_n,Y_n)$ under (\ref{thm_eq_ps}) and $ X_n \geq 2$, it follows that
\begin{align}
\hspace{-2pt}\lim_{n \to \infty} \hspace{-2pt} \mathbb{P}\bigg[\hspace{-4pt}\begin{array}{l}G(n,X_n,Y_n)\text{ has a minimum degree at least
}k,\\\text{but is not $k$-connected}.\end{array}\hspace{-4pt}\bigg] & \hspace{-2pt}=\hspace{-2pt} 0.
\end{align}
\end{lem}
\begin{lem} \label{lemma-2}
For a random key graph $G(n,X_n,Y_n)$ under (\ref{thm_eq_ps}) and $ X_n \geq 2$, it follows that
\begin{align}
& \lim_{n \to \infty} \mathbb{P}
\left[\hspace{2pt}G(n,X_n,Y_n)\text{ has a minimum degree at least
}k.\hspace{2pt}\right] \nonumber \\
& \quad = \text{right hand side of (\ref{doublexp})}. \nonumber
\end{align}
\end{lem}
By \cite[Lemma 2]{mobihocQ1}, Lemma \ref{lemma-2} will follow once we show Lemma \ref{lemma-3} below, where we let $\mathcal {V} = \{v_1, v_2,
\ldots, v_n \}$ be the set of nodes in a random key graph $G(n, X_n, Y_n)$.
\begin{lem} \label{lemma-3}
For a random key graph $G(n,X_n,Y_n)$ under (\ref{thm_eq_ps}) and $ X_n \geq 2$, it follows for integers $m\geq 1$ and $h \geq 0$ that
\begin{align}
& \mathbb{P} [\text{Nodes }v_{1}, v_{2}, \ldots, v_{m}\text{ have
degree }h] \nonumber \\
& \quad \sim (h!)^{-m} (n q_n)^{hm} e^{-m n q_n}.
\label{eqn_node_v12n}
\end{align}
\end{lem}
We detail the proof of Lemma \ref{lemma-3} in the next section.
\section{The Proof of Lemma \ref{lemma-3}} \label{sec_est}
In a random key graph $G(n, X_n, Y_n)$, recalling that
$\mathcal {V} = \{v_1, v_2,
\ldots, v_n \}$ is the set of nodes, we let $S_i$ be the set of $X_n$
distinct objects assigned to node $v_i \in \mathcal {V}$.
We further define $\mathcal {V}_m$ as $ \{v_1, v_2, \ldots, v_m\}$ and
$\overline{\mathcal {V}_m} $ as $ \mathcal {V} \setminus \mathcal {V}_m
$. Among nodes in $\overline{\mathcal {V}_m}$, we denote by $N_i$
the set of nodes neighboring to $v_i$ for $i=1,2,\ldots,m$. We
denote $N_i \bcap N_j$ by $N_{ij}$, and $S_i \bcap S_j$ by $S_{ij}$.
We have the following two observations:
\begin{itemize}
\item [i)] If node $v_i$ has degree $h$,
then $|N_{i}| \leq h$, where the equal sign holds if and only if
$v_i$ is directly connected to none of nodes in $V_m \setminus
\{v_i\}$; i.e., if and only if event $\bigcap_{j \in \{1,2,\ldots,m\}\setminus\{i\}}
(S_{ij}=\emptyset)$ happens. \vspace{2pt}
\item [ii)] If $|N_{i}| \leq h$ for any $i=1,2,\ldots,m$, then
\begin{align}
& \bigg|\bigcup_{1\leq i \leq m} N_{i}\bigg| \leq \sum_{1\leq i \leq
m}N_{i} \leq hm , \label{Nileq}
\end{align}
where the two equal signs in (\ref{Nileq}) \emph{both} hold if and only if \vspace{-1pt}
\begin{align}
\bigg(\bigcap_{1\leq i <j \leq m} (N_{ij}=\emptyset)\bigg) \bcap
\bigg(\bigcap_{1\leq i \leq m}(|N_{i}| = h)\bigg). \vspace{-1pt}\label{Nij}
\end{align}
\end{itemize}
From i) and ii) above, if nodes $v_{1}, v_{2}, \ldots, v_{m}$ have degree $h$,
we have either of the following two cases:
\begin{itemize}
\item [(a)] Any two of $v_{1}, v_{2}, \ldots, v_{m}$ have no edge in between (namely, $\bigcap_{1\leq i <j \leq m}
(S_{ij}=\emptyset)$); and event (\ref{Nij}) happens. \vspace{1pt}
\item [(b)] $\big|\bigcup_{1\leq i \leq m} N_{i}\big|
\leq hm -1$.
\end{itemize}
In addition, if case (a) happens, then nodes $v_{1}, v_{2}, \ldots,
v_{m}$ have degree $h$. However, if case (b) occurs, there is no
such conclusion. With $P_a$ (resp., $P_b$) denoting the probability
of case (a) (resp., case (b)), we obtain \vspace{-1pt}
\begin{align}
& P_a \leq \mathbb{P} [\text{Nodes }v_{1}, v_{2}, \ldots, v_{m}\text{ have
degree }h] \leq P_a + P_b, \vspace{-1pt}\nonumber
\end{align}
where \vspace{-1pt}
\begin{align}
P_a = \mathbb{P}\bigg[ \bigg(\bigcap_{1\leq i <j \leq m}
(S_{ij}=\emptyset)\bigg) & \bcap \bigg(\bigcap_{1\leq i <j \leq m}
(N_{ij}=\emptyset)\bigg) \nonumber \\
& \bcap \bigg(\bigcap_{1\leq i \leq
m}(|N_{i}| = h)\bigg)\bigg], \vspace{-1pt} \nonumber
\end{align}
and \vspace{-1pt}
\begin{align}
P_b & = \mathbb{P}\bigg[\bigg|\bigcup_{1\leq i \leq m} N_{i}\bigg|
\leq hm -1\bigg]. \vspace{-1pt} \nonumber
\end{align}
Hence, (\ref{eqn_node_v12n}) holds after we prove the
following (\ref{prop2}) and (\ref{prop1}):
\begin{align}
P_b & = o \left((nq_n)^{hm} e^{-m n q_n}\right). \label{prop2}
\end{align}
and
\begin{align}
P_a & \sim (h!)^{-m} (n q_n)^{hm} e^{-m n q_n} \cdot
[1+o(1)], \label{prop1}
\end{align}
We will prove (\ref{prop2}) and (\ref{prop1}) below.
We let $\mathbb{S}_m$ denote the tuple $(S_1,S_2,\ldots,S_m)$. The
expression ``$\given \mathbb{S}_m = \mathbb{S}_m^*$'' means ``given
$S_1=S_1^*,S_2=S_2^*,\ldots,S_m=S_m^*$'', where $\mathbb{S}_m^* =
(S_1^*,S_2^*,\ldots,S_m^*)$ with $S_1^*,S_2^*,\ldots,S_m^*$ being
arbitrary $X_n$-size subsets of the object pool. Note that
$S_{ij}^{*} : = S_{i}^{*} \cap S_{j}^{*}$. For two different nodes $v$ and $w$ in the graph $G(n, X_n, Y_n)$, we use $v\leftrightarrow w$ to denote
the event that there is an edge between $v$ and $w$; i.e., the symbol ``$\leftrightarrow $'' means
``is directly connected to''.
\subsection{The Proof of (\ref{prop2})}
Let $w$ be an arbitrary node in $\overline{V_m}$. We have
\begin{align}
& \mathbb{P}\bigg[\bigg|
\bigcup_{1\leq i \leq m} N_{i}\bigg| = t \bgiven
\mathbb{S}_m= \mathbb{S}_m^*\bigg] \label{t} \\
& = \frac{(n-m)!}{t!(n-m-t)!} \nonumber \\
& \times \big\{\mathbb{P}[w\leftrightarrow \text{ at least one of nodes in }V_m \given
\mathbb{S}_m= \mathbb{S}_m^*]\big\}^t \nonumber \\
& \times \big\{\mathbb{P}[w\leftrightarrow \text{none of
nodes in }V_m \given \mathbb{S}_m= \mathbb{S}_m^*]\big\}^{n-m-t}.
\label{x}
\end{align}
By the union bound, it holds that
\begin{align}
& \mathbb{P}[w\leftrightarrow \text{at least one of
nodes in }V_m \given \mathbb{S}_m= \mathbb{S}_m^*] \nonumber \\
& \leq \sum_{1\leq i \leq m}\mathbb{P}
[w\leftrightarrow v_i \given \mathbb{S}_m= \mathbb{S}_m^*] = m q_n,\label{lll}
\end{align}
which yields
\begin{align}
& \mathbb{P}[w\leftrightarrow \text{none of nodes in }V_m \given
\mathbb{S}_m= \mathbb{S}_m^*] \geq 1 - m q_n. \label{sstar3}
\end{align}
In addition,
\begin{align}
& \mathbb{P}[w\leftrightarrow \text{none of nodes in }V_m \given
\mathbb{S}_m= \mathbb{S}_m^*] \nonumber \\ & = \frac{\binom{Y_n -
|\bigcup_{1\leq i \leq m} S_i^*|}{X_n}}{\binom{Y_n}{X_n}} \nonumber \\
& \leq (1-q_n)^{{X_n}^{-1}{|\bigcup_{1\leq i \leq m} S_i^*|}} \quad
\text{(by \cite[Lemma 5.1]{yagan_onoff})} \nonumber \\
& \leq e^{-{X_n}^{-1}q_n{|\bigcup_{1\leq i \leq m} S_i^*|}} \quad
\text{(by $1+x \leq e^x$ for any real $x$)}.
\label{sstar4}
\end{align}
We will prove
\begin{align}
& \sum_{\mathbb{S}_m^*} \Big\{ \mathbb{P}[\mathbb{S}_m =
\mathbb{S}_m^*] \nonumber \\ & \quad \times \big\{ \mathbb{P}[w\leftrightarrow \text{none of nodes in }V_m \given \mathbb{S}_m= \mathbb{S}_m^*]
\big\}^{n-m-hm}
\Big\} \label{sstar2} \\
& \quad \leq e^{-m
n q_n} \cdot
[1+o(1)] . \label{sstar}
\end{align}
From (\ref{x}) (\ref{lll}) and (\ref{sstar}), we derive
\begin{align}
P_b & = \mathbb{P}\bigg[\bigg|\bigcup_{1\leq i \leq m} N_{i}\bigg|
\leq hm -1\bigg] \nonumber \\ &= \sum_{t=0}^{hm -1}
\sum_{\mathbb{S}_m^*} \Big\{ \mathbb{P}[\mathbb{S}_m = \mathbb{S}_m^*]
\cdot (\ref{t})
\Big\}
\nonumber \\
& \leq \sum_{t=0}^{hm -1} \Big[ n^t \cdot (m q_n)^t \cdot (\ref{sstar2})\Big]
\nonumber \\ & \leq (nq_n)^{hm} e^{-m n q_n} \cdot
[1+o(1)] \cdot m^{hm}
\sum_{t=0}^{hm -1} (mnq_n)^{t-hm}
. \label{bp}
\end{align}
As noted in Remark \ref{rem1}, we can introduce an extra condition
$\alpha_n = \pm O(\ln \ln n)= \pm o(\ln n)$ in establishing Theorem \ref{THM_RKGk}.
From $\alpha_n = \pm o(\ln n)$ and (\ref{thm_eq_ps}), we obtain
\begin{align}
q_n & \sim \frac{\ln n}{n}.\label{eq_pe_lnnn}
\end{align}
Applying (\ref{eq_pe_lnnn}) to (\ref{bp}), we obtain (\ref{prop2}). Hence,
we complete the proof of (\ref{prop2}) once showing (\ref{sstar}),
whose proof is detailed below.
From (\ref{sstar3}) (\ref{sstar4}) and (\ref{eq_pe_lnnn}), we have
\begin{align}
(\ref{sstar2}) \hspace{-1pt}& \hspace{-1pt}\leq\hspace{-1pt} (1 - m q_n)^{-m-hm} \nonumber \\
\hspace{-1pt}& \hspace{-1pt}\quad \times \sum_{\mathbb{S}_m^*}
\Big\{ \mathbb{P}[\mathbb{S}_m = \mathbb{S}_m^*] \cdot
e^{-{X_n}^{-1}nq_n{|\bigcup_{1\leq i \leq m} S_i^*|}}
\Big\} \nonumber \\
\hspace{-1pt}& \hspace{-1pt}\leq\hspace{-1pt}[1+o(1)]\hspace{-1pt}\cdot\hspace{-1pt}\sum_{\mathbb{S}_m^*}
\Big\{ \mathbb{P}[\mathbb{S}_m \hspace{-1pt}=\hspace{-1pt} \mathbb{S}_m^*] \hspace{-1pt}\cdot\hspace{-1pt}
e^{-{X_n}^{-1}nq_n{|\bigcup_{1\leq i \leq m} S_i^*|}} \Big\},
\end{align}
so (\ref{sstar}) holds once we demonstrate
\begin{align}
& \sum_{\mathbb{S}_m^*} \Big\{ \mathbb{P}[\mathbb{S}_m =
\mathbb{S}_m^*] \cdot e^{-{X_n}^{-1}nq_n{|\bigcup_{1\leq i \leq m}
S_i^*|}} \Big\} \nonumber \\
& \quad \leq e^{-m n q_n} \cdot
[1+o(1)] . \label{ms}
\end{align}
We denote the left hand side of (\ref{ms}) by $Z_{m,n}$. Dividing $\mathbb{S}_{m}^*$ into two parts $\mathbb{S}_{m-1}^*$
and $S_m^*$,
we derive
\begin{align}
Z_{m,n}
&= \sum_{\begin{subarray}{c}\mathbb{S}_{m-1}^*
\\S_m^*\end{subarray}} \Big\{ \mathbb{P}[(\mathbb{S}_{m-1} = \mathbb{S}_{m-1}^*)
\bcap(S_m = S_m^*)] \nonumber \\
& \quad\quad\quad\quad \times e^{-{X_n}^{-1}nq_n{|\bigcup_{1\leq i \leq m}
S_i^*|}} \Big\}\nonumber \\ &= \sum_{\mathbb{S}_{m-1}^*}
\mathbb{P}[\mathbb{S}_{m-1} = \mathbb{S}_{m-1}^*] \bigg\{
e^{-{X_n}^{-1}nq_n{|\bigcup_{1\leq i \leq m-1} S_i^*|}} \nonumber \\
& \quad\quad \times \sum_{S_m^* }
\mathbb{P}[ S_m = S_m^* ] e^{-{X_n}^{-1}nq_n{|S_m^* \setminus
\bigcup_{1\leq i \leq m-1} S_i^*|}}\bigg\} ,\label{HnmHnm1}
\end{align}
where
\begin{align}
& \sum_{S_m^* } \mathbb{P}[ S_m = S_m^* ] e^{-{X_n}^{-1}nq_n{|S_m^*
\setminus \bigcup_{1\leq i \leq m-1} S_i^*|}} \nonumber \\ & \leq
e^{-n q_n}\sum_{S_m^* } \mathbb{P}[ S_m = S_m^* ] e^{{X_n}^{-1}{ n
q_n}\big|S_m^* \cap
\big(\bigcup_{i =1}^{m-1} S_{i }^* \big) \big|} \nonumber \\ &
= e^{-n q_n} \sum_{r=0}^{X_n}
\mathbb{P}\bigg[\bigg|S_m\bcap \bigg(\bigcup_{i =1}^{m-1}S_{i
}^*\bigg)\bigg| = r \bigg] e^{{X_n}^{-1}{n q_nr} } . \label{SS}
\end{align}
For $r$ satisfying
\begin{align}
0 &\leq r \leq |S_m|=X_n \nonumber
\end{align}
and
\begin{align}
r & = |S_m| + \bigg|\bigcup_{i =1}^{m-1}S_{i
}^*\bigg| - \bigg|S_m\bcup \bigg(\bigcup_{i =1}^{m-1}S_{i
}^*\bigg)\bigg| \nonumber \\
& \geq X_n + \bigg|\bigcup_{i =1}^{m-1}S_{i }^*\bigg|
- Y_n , \nonumber
\end{align}
as given in \cite[Eq. (36)]{mobihocQ1},
we have
\begin{align}
\mathbb{P}\bigg[\bigg|S_m\bcap \bigg(\bigcup_{i =1}^{m-1}S_{i
}^*\bigg)\bigg| = r \bigg] & \leq \frac{1}{r!} \bigg( \frac{m
{X_n}^2}{Y_n - X_n}\bigg)^r. \label{probsm2}
\end{align}
Applying (\ref{probsm2}) to (\ref{SS}), we establish
\begin{align}
& \sum_{S_m^* } \mathbb{P}[ S_m = S_m^* ] e^{-{X_n}^{-1}nq_n{|S_m^*
\setminus \bigcup_{1\leq i \leq m-1} S_i^*|}} \nonumber \\ & \leq
e^{-n q_n} \sum_{r=0}^{X_n} \frac{1}{r!} \bigg( \frac{m {X_n}^2}{Y_n - X_n}\bigg)^r
\cdot e^{{X_n}^{-1}{n q_nr} } \nonumber \\ &
\leq e^{-n q_n} \cdot e^{\frac{m {X_n}^2}{Y_n - X_n}
\cdot e^{{X_n}^{-1}{n q_n}}} . \label{psnm}
\end{align}
From (\ref{eq_pe_lnnn}) and (\ref{PnKK}), it holds that
$\frac{{X_n}^2}{Y_n}\sim \frac{\ln n}{n}$, resulting in
\begin{align}
\frac{m {X_n}^2}{Y_n - X_n} & \sim \frac{m{X_n}^2}{Y_n} \sim
\frac{m\ln n}{n}. \label{e1}
\end{align}
For an arbitrary $\epsilon > 0$, from (\ref{eq_pe_lnnn}), we obtain
$q_n \leq (1+\epsilon)\frac{\ln n}{n}$ for all $n$ sufficiently
large, which with condition $X_n \geq 2$ yields that for all $n$ sufficiently large,
\begin{align}
e^{{X_n}^{-1}{n q_n}} & \leq e^{\frac{1}{2}(1+\epsilon)\ln n} =
n^{\frac{1}{2}(1+\epsilon)}. \label{e2}
\end{align}
From (\ref{e1}) and (\ref{e2}), we get
\begin{align}
\frac{m {X_n}^2}{Y_n - X_n} \cdot e^{{X_n}^{-1}{n q_n}} & \leq
\frac{m\ln n}{n} \cdot [1+o(1)] \cdot n^{\frac{1}{2}(1+\epsilon)}
\nonumber
\\ & \leq m\ln n \cdot n^{\frac{1}{2}(\epsilon-1)} \cdot [1+o(1)] .
\label{e3}
\end{align}
Since $\epsilon > 0$ is arbitrary, it follows from (\ref{e3}) that
for arbitrary $0<c<\frac{1}{2}$, then for all $n$ sufficiently
large, it is clear that
\begin{align}
\frac{m {X_n}^2}{Y_n - X_n} \cdot e^{{X_n}^{-1}{n q_n}} & \leq
n^{-c}. \label{e4}
\end{align}
Using (\ref{e4}) in (\ref{psnm}), for all $n$ sufficiently large, it follows that
\begin{align}
\sum_{S_m^* } \mathbb{P}[ S_m = S_m^* ] e^{-{X_n}^{-1}nq_n{|S_m^*
\setminus \bigcup_{1\leq i \leq m-1} S_i^*|}} & \leq e^{-n q_n}
\cdot e^{n^{-c}} . \label{e5}
\end{align}
Substituting (\ref{e5}) into (\ref{HnmHnm1}), for all $n$
sufficiently large, we obtain
\begin{align}
&\hspace{-2pt} Z_{m,n} \nonumber \\
& \hspace{-3pt}\leq\hspace{-1.5pt} e^{-n q_n} \hspace{-1.5pt}\cdot\hspace{-1.5pt} e^{n^{-c}}
\hspace{-1.5pt}\cdot \hspace{-1.5pt}\sum_{\mathbb{S}_{m-1}^*} \hspace{-1pt}\mathbb{P}[\mathbb{S}_{m-1} \hspace{-1.5pt}=\hspace{-1.5pt}
\mathbb{S}_{m-1}^*]
e^{-{X_n}^{-1}nq_n{|\bigcup_{1\leq i \leq m-1} S_i^*|}} \nonumber \\
& \hspace{-3pt}\leq\hspace{-1pt} e^{-n q_n} \hspace{-1pt}\cdot \hspace{-1pt}e^{n^{-c}} \hspace{-1pt}\cdot\hspace{-1pt} Z_{m-1,n}.
\end{align}
We then evaluate $Z_{2,n}$. By (\ref{ms}), it holds that
\begin{align}
& \hspace{-2pt}Z_{2,n} \nonumber \\
& \hspace{-3pt}=\hspace{-1.5pt}\sum_{S_1^*}\hspace{-1pt}\sum_{S_2^*} \hspace{-1pt}\Big\{ \hspace{-1pt}\mathbb{P}[(S_1 \hspace{-1pt}=\hspace{-1pt}
S_1^*)\hspace{-1pt}\bcap \hspace{-1pt}(S_2 \hspace{-1pt}=\hspace{-1pt} S_2^*)] \hspace{-1.5pt}\cdot\hspace{-1.5pt}e^{-{X_n}^{-1}nq_n{|S_1^* \bcup
S_2^*|}} \Big\}\nonumber \\ & = \sum_{S_1^*} \mathbb{P}[ S_1 =
S_1^* ] \sum_{S_2^*} \mathbb{P}[ S_2 = S_2^* ]
e^{-{X_n}^{-1}nq_n{|S_1^* \bcup S_2^*|}}. \label{mm1}
\end{align}
Setting $m=2$ in (\ref{e5}), for all $n$ sufficiently large, we derive
\begin{align}
\sum_{S_2^*} \mathbb{P}[ S_2 = S_2^* ] e^{-{X_n}^{-1}nq_n{| S_2^*
\setminus S_1^*|}} & \leq e^{-n q_n} \cdot e^{n^{-c}} . \nonumber
\end{align}
Then for all $n$ sufficiently large, it follows that
\begin{align}
& \sum_{S_2^*} \mathbb{P}[ S_2 = S_2^* ] e^{-{X_n}^{-1}nq_n{|S_1^*
\bcup S_2^*|}} \nonumber \\
& = e^{-n q_n} \sum_{S_2^*} \mathbb{P}[ S_2 =
S_2^* ] e^{-{X_n}^{-1}nq_n{| S_2^* \setminus S_1^*|}} \nonumber \\
& \leq e^{-2n q_n} \cdot e^{n^{-c}} . \label{e6}
\end{align}
From (\ref{mm1}) and (\ref{e6}), for all $n$ sufficiently large, we
obtain
\begin{align}
Z_{m,n} & \leq \big(e^{-n q_n} \cdot e^{n^{-c}}\big)^{m-2}
\cdot Z_{2,n} \nonumber \\
& \leq \big(e^{-n q_n} \cdot e^{n^{-c}}\big)^{m-2} \cdot e^{-2n
q_n} \cdot e^{n^{-c}}
\nonumber \\
& \leq e^{-mn q_n} \cdot e^{(m-1)n^{-c}} .
\end{align}
Letting $n \to \infty$, we finally establish
\begin{align}
Z_{m,n} & \leq e^{-m n q_n} \cdot
[1+o(1)] ; \nonumber
\end{align}
i.e., (\ref{ms}) is proved. Then as explained above, (\ref{sstar})
holds; and then (\ref{prop2}) follows.
\subsection{The Proof of (\ref{prop1})}
Again let $w$ be an arbitrary node in $\overline{V_m}$. We have
\begin{align}
& \mathbb{P}\bigg[\bigg(\bigcap_{1\leq i <j \leq m}
(N_{ij}=\emptyset)\bigg) \bcap \bigg(\bigcap_{1\leq i \leq
m}(|N_{i}| = h)\bigg) \bgiven \mathbb{S}_m= \mathbb{S}_m^*\bigg]
\label{h} \\ & = \frac{(n-m)!}{(h!)^m(n-m-hm)!} \nonumber \\
& \quad \times \prod_{1\leq i \leq m}\left(\left\{\mathbb{P}
\left[\begin{array}{l}w\leftrightarrow v_i,\\\text{but }w\leftrightarrow\text{none of}\\\text{nodes
in }V_m \setminus \{v_i\}\end{array}\Bigg|\hspace{3pt} \mathbb{S}_m= \mathbb{S}_m^*\right]\right\}^h\right) \nonumber \\
& \quad \times \big\{\mathbb{P}[w\leftrightarrow \text{none of
nodes in }V_m \given \mathbb{S}_m= \mathbb{S}_m^*]\big\}^{n-m-hm}
\label{y}
\end{align}
and
\begin{align}
P_a & = \sum_{\mathbb{S}_m^*:\hspace{2pt}\bigcap_{1\leq i <j \leq
m} (S_{ij}^*=\emptyset)} \Big\{ \mathbb{P}[\mathbb{S}_m =
\mathbb{S}_m^*] \cdot (\ref{h})
\Big\} \label{pra} ,
\end{align}
where $S_{ij}^{*} : = S_{i}^{*} \cap S_{j}^{*}$.
For $i=1,2,\ldots,m$, under
$\mathbb{S}_m^*:\hspace{2pt}\bigcap_{1\leq i <j \leq m}
(S_{ij}^*=\emptyset)$, we have
\begin{align}
& \mathbb{P}
[w\leftrightarrow v_i,\text{ but none of nodes
in }V_m \setminus \{v_i\}\given \mathbb{S}_m= \mathbb{S}_m^*] \nonumber \\
& \geq \mathbb{P}
[w\leftrightarrow v_i \given \mathbb{S}_m= \mathbb{S}_m^*]
\nonumber \\
& \quad - \sum_{\begin{subarray}{c} 1 \leq j \leq m \\ j\neq
i\end{subarray}} \mathbb{P}
[w\leftrightarrow \text{both }v_i\text{ and }v_j
\given \mathbb{S}_m= \mathbb{S}_m^*] \nonumber \\
&\geq q_n - (m-1) \cdot 2{q_n} ^2 \quad \text{(by \cite[Lemma
3]{mobihocQ1})} . \label{ps2}
\end{align}
Substituting (\ref{sstar3}) and (\ref{ps2}) to (\ref{y}), and then
from (\ref{pra}), we obtain
\begin{align}
P_a & \geq \frac{(n-m-hm)^{hm}}{(h!)^m} \cdot [q_n - 2(m-1) q_n
^2]^{hm} \nonumber \\
& \quad \times (1-mq_n)^{n-m-hm} \sum_{\mathbb{S}_m^*:\hspace{2pt}\bigcap_{1\leq i
<j \leq m} (S_{ij}^*=\emptyset)} \mathbb{P}[\mathbb{S}_m =
\mathbb{S}_m^*]. \nonumber
\end{align}
Then from (\ref{eq_pe_lnnn}), it further hold that
\begin{align}
P_a & \geq \frac{n^{hm}}{(h!)^m} \cdot (q_n)^{hm}
\cdot e^{- m n q_n} \nonumber \\
& \quad \times [1-o(1)] \cdot
\mathbb{P}\bigg[\bigcap_{1\leq i <j \leq m} (S_{ij}=\emptyset)\bigg]
. \label{poa1}
\end{align}
From (\ref{sstar4}), under
$\mathbb{S}_m^*:\hspace{2pt}\bigcap_{1\leq i <j \leq m}
(S_{ij}^*=\emptyset)$, it holds that
\begin{align}
\mathbb{P}[w\leftrightarrow \text{none of nodes in }V_m \given
\mathbb{S}_m= \mathbb{S}_m^*] & \leq e^{- m q_n} . \label{ps3a}
\end{align}
For each $i=1,2,\ldots,m$, we have
\begin{align}
& \mathbb{P}
[w\leftrightarrow v_i,\text{ but }w\leftrightarrow \text{none of nodes
in }V_m \setminus \{v_i\}\given \mathbb{S}_m= \mathbb{S}_m^*] \nonumber \\
& \leq \mathbb{P}
[w\leftrightarrow v_i \given \mathbb{S}_m= \mathbb{S}_m^*] =
q_n. \label{ps}
\end{align}
Substituting (\ref{ps}) and (\ref{ps3a}) to (\ref{y}), and then from
(\ref{pra}), we obtain
\begin{align}
P_a & \hspace{-1.5pt}\leq\hspace{-1.5pt} \frac{n^{hm}}{(h!)^m} \hspace{-1.5pt}\cdot\hspace{-1.5pt} (q_n)^{hm}
\hspace{-1.5pt}\cdot\hspace{-1.5pt} e^{- m n q_n} \hspace{-1.5pt}\cdot\hspace{-1.5pt} \sum_{\mathbb{S}_m^*:\hspace{2pt}\hspace{-1.5pt}\bigcap_{1\leq i
<j \leq m} (S_{ij}^*=\emptyset)} \hspace{-1.5pt}\mathbb{P}[\mathbb{S}_m \hspace{-1.5pt}=\hspace{-1.5pt}
\mathbb{S}_m^*] \nonumber \\
& = \frac{n^{hm}}{(h!)^m} \cdot (q_n)^{hm}
\cdot e^{- m n q_n} \cdot
\mathbb{P}\bigg[\bigcap_{1\leq i <j \leq m} (S_{ij}=\emptyset)\bigg]
. \label{poa2}
\end{align}
From (\ref{poa1}) and (\ref{poa2}), we obtain\vspace{-1pt}
\begin{align}
P_a & \sim \frac{n^{hm}}{(h!)^m} \cdot (q_n)^{hm}
\cdot e^{- m n q_n} \cdot
\mathbb{P}\bigg[\bigcap_{1\leq i <j \leq m} (S_{ij}=\emptyset)\bigg].
\label{y2} \vspace{-1pt}
\end{align}
By the union bound, it is clear that\vspace{-1pt}
\begin{align}
& \mathbb{P}\bigg[\bigcap_{1\leq i <j \leq m} (S_{ij}=\emptyset)\bigg]\vspace{-1pt}
\nonumber \\ & \quad = 1 - \mathbb{P}\bigg[\bigcup_{1\leq i <j \leq m} (S_{ij}\neq
\emptyset)\bigg]\vspace{-1pt} \nonumber \\ & \quad \geq 1 - \sum_{1\leq i <j \leq m}
\mathbb{P}[S_{ij}\neq \emptyset] = 1 - \binom{m}{2}q_n. \label{m2ps}\vspace{-1pt}
\end{align}
From (\ref{eq_pe_lnnn}) and (\ref{m2ps}), since a
probability is at most $1$, we get\vspace{-1pt}
\begin{align}
\lim_{n \to \infty}\mathbb{P}\bigg[\bigcap_{1\leq i <j \leq m}
(S_{ij}=\emptyset)\bigg] & = 1 .\vspace{-1pt} \label{m2ps2}
\end{align}
Using (\ref{m2ps2}) in (\ref{y2}), we establish (\ref{prop1}).
\section{Numerical Experiments} \label{sec:expe}
We present numerical experiments to back up our
theoretical results. %
Figure \ref{fig} depicts the probability that graph
$G(n,X,Y)$ is $k$-connected. We let $X$ vary, with other parameters
fixed at $n=3,000$, $Y=30,000$ and $k=3 ,7$. The empirical
probabilities corresponding to the experimental curves are obtained
as follows: we count the times of $k$-connectivity out of $500$
independent samples of $G(n,X,Y)$, and derive the empirical
probability through dividing its corresponding count by $500$. For
the theoretical curves, we first compute $\alpha$ by setting \vspace{1pt} the edge probability $1-
\frac{\binom{Y- X}{X}}{\binom{Y}{X}}$ (viz., (\ref{qnp}) in the Appendix) \vspace{-1pt}
as $ \frac{\ln n + {(k-1)} \ln \ln n + {\alpha}}{n}$ and then use
$e^{- \frac{e^{-\alpha}}{(k-1)!}}$ as the theoretical
value for the probability of $k$-connectivity. Figure \ref{fig}
confirms our analytical results as the experimental and theoretical
curves are close.
\begin{figure}[!t]
\centering
\includegraphics[width=0.45\textwidth]{figure.eps}
\caption{A plot for the probability of $k$-connectivity in graph
$G(n,X,Y)$ with $k = 3,7$ under $n=3,000$ and
$Y=20,000$.\vspace{-5pt}
}
\label{fig}
\end{figure}
\section{Related Work} \label{related}
For a random key graph $G(n,X_n,Y_n)$, Rybarczyk \cite{ryb3} derives
the asymptotically exact probabilities of connectivity and of the
property that the minimum node degree is no less than $1$, covering
a weaker form of the results -- the zero-one laws which are also
obtained in \cite{r1,yagan}. As demonstrated in \cite{ryb3}, in
$G(n,X_n,Y_n)$ with $X_n \geq 2$, $\frac{{X_n}^2}{Y_n}=\frac{\ln n +
{\alpha_n}}{n}$ and $\lim_{n \to \infty} \alpha_n = \alpha ^* $, the
probability of connectivity and that of the minimum degree being at
least $1$ both approach to $e^{- e^{-\alpha ^*}}$ as $n\to \infty$.
Rybarczyk \cite{zz} implicitly obtains zero-one laws (but not the
asymptotically exact probabilities) for $k$-connectivity and for the
property that the minimum degree is at least $k$. The implicit
result is that if $X_n \Theta(n
^{\beta})$ for some $\beta
>0$ and $\frac{{X_n}^2}{Y_n}=\frac{\ln n + (k-1) \ln \ln n +
{\alpha_n}}{n}$, graph $G(n,X_n,Y_n)$ has (resp., does not have) the
two properties with probability approaching to $1$, given that $
{\alpha_n} $ tends to $\infty$ (resp., $-\infty$) as $n\to \infty$.
Our Corollary \ref{COR_RKGk} significantly improves her result
\cite{zz} in the following two aspects: (i) we cover the wide range of $ X_n \geq 2$
all $n$ sufficiently large, instead of the much stronger condition
$X_n = \Omega\big((\ln n)^3\big)$ in \cite{zz} (note that the analysis under $X_n=1$ is trivial), and
(ii) we establish not only zero--one laws for $k$-connectivity
and the minimum degree, but also the asymptotically exact
probabilities. The latter results are not given by Rybarczyk \cite{zz}. Recently, we \cite{JZCDC} give
the asymptotically exact
probability of $k$-connectivity in graph $G(n,X_n,Y_n)$ under $X_n = \Omega(\sqrt{\ln n}\hspace{1.5pt})$ through a rather involved proof. We improve this result
to cover $ X_n \geq 2$ through a simpler proof and fill the gap where $X_n$ is at least $2$, but is not $ \Omega(\sqrt{\ln n}\hspace{1.5pt})$. This improvement is of technical interest as well as of practical importance since random key graphs have been used in diverse applications including modeling the Eschenauer--Gligor random key predistribution scheme (the most recognized solution to secure communication in wireless sensor networks).
\section{Conclusion}
\label{sec:Conclusion}
In this paper, for a random key graph, we derive the asymptotically
exact probabilities for two properties with an arbitrary $k$: (i)
the graph is $k$-connected; and (ii) each node has at least $k$
neighboring nodes. Numerical experiments are in accordance with our
analytical results.
\renewcommand\baselinestretch{.95}
\small
\bibliographystyle{abbrv}
|
\section{Introduction}
The demonstration of single photon emission from a semiconductor quantum dot (QD) \cite{Michler-Science00} triggered 15 years of prospering research devoted towards the application of semiconductor quantum light emitters. In particular the most commonly studied InGaAs/GaAs based QDs can emit highly non-classical light, both under optical and electrical excitation \cite{Yuan-Science02, Michler-Science00, Heindel-APL10, Michler-1}. However, for many applications beyond simple quantum cryptography schemes, it is vitally important that the emitted single photons are comprising highest degrees of indistinguishability, which means that they have to be identical in all spectral characteristics: Their color, polarization, and furthermore the extension of the wave packet (i.e. their coherence) should be Fourier limited \cite{Michler-2, Santori-Nature02}. Such photons are at the heart of applications in quantum communication \cite{Pan-RMP12}, quantum networks \cite{Kok-RMP07} and linear optical quantum computing \cite{Obrien-Science07}. Most quantum teleportation schemes strongly rely on photon indistinguishability, and in particular the route towards quantum repeater networks for future long distance quantum communications highly relies on this property \cite{Briegel-PRL98, Hofmann-Science12, Cirac-PRL97}. First important experiments relying on quantum interference and teleportation of photons emitted from single QDs have recently been carried out \cite{Nilsson-NatPhot13, Gao-NatCom13}. In contrast to isolated quantum emitters such as atoms in dilute vapors, QDs are embedded in a solid state environment which imposes limitations on
\begin{itemize}
\item the brightness of the source since the photons have to be extracted out of a high refractive index material
\item the interference properties of photons emitted from these sources, as coherence and color of the emission can be affected by coupling to the environment of the emitter.
\end{itemize}
While the source efficiency can be boosted to very high values by embedding quantum emitters in photonic micro- and nanostructures \cite{Heindel-APL10, Gazzano-NatCom13, Claudon-NatPhot10, Reimer-NatCom12}, increasing the degree of indistinguishability at least partly requires to decouple the emitter from its environment. In particular frequency shifts induced by charges in the QD's vicinity, or effects of phonon induced emitter dephasing strongly and detrimentally affect the interference properties of single QDs.
This chapter is structured as follows: In section~\ref{section:theory} we will briefly address the fundamental mechanisms of two photon interference and its experimental implementation. In section~\ref{section:SPS} we will describe the experimental realization of a bright single photon source as the basis of the following studies. Finally, in sections~\ref{section:experimentI} and \ref{section:experimentII} we will describe two photon interference experiments carried out on single QDs. We will assess in detail the influence of the excitation scheme on the interference properties of the quantum light emitted from the QD. While carrier refilling effects are identified to strongly detrimentally influence the photon interference in non-resonant excitation schemes, resonance fluorescence conditions can lead to almost perfect interference visibilities of photons. We can furthermore demonstrate significant interference of single photons emitted from separate sources and compare our experimental findings with an analytical model.
\section{A pedestrian's guide to two photon interference}
\label{section:theory}
\subsection{Quantum dot single photon source}
In a very simplified picture, single QDs can be considered as two level systems embedded in a solid state environment: Electrons and holes can be captured and localized in QDs if the band configuration provides an potential well in the conduction and valence band, and the small size of the QDs leads to strong Coulomb and exchange interactions. As a result, the energetic ground state of a single dot can only be occupied by a single electron-hole pair. This excitonic carrier complex can decay spontaneously and its energy is transferred to a single photon. Since the maximum occupation number of the excited state cannot exceed unity, not more than one photon of the corresponding energy can be emitted at a time interval on the order of the exciton lifetime. The distribution function of the photon stream is hence sub-poissonian (or antibunched), which is usually characterized by the second order correlation function which we write here in terms of the emitted photon intensities and the delay time $\tau$ between two detection events: $g^{(2)}(\tau)=\frac{<I(t)I(t+\tau)>}{<I(t)>^2}$. For an ideal single photon source, the value of this function approaches 0 at $\tau=0$. The antibunched nature of the emission from a single QD makes these quantum emitters highly interesting for quantum cryptography schemes relying on encoding the information of a quantum key into the polarization of a single light particle (such as the famous BB84 protocol \cite{Bennett-84}). First successful experimental demonstrations of quantum key distribution with QD single photon sources, both under optical \cite{Waks-Nature02} and electrical excitation \cite{Heindel-NJP12} have been realized. However, for more advanced schemes such as the remote entanglement of stationary quantum bits (Qubits), the interference properties of these photons play a dominant role, which we will detail in the following subchapter.
\subsection{Photon interference with quantum light}
The interference properties of single photons and their indistinguishability are very closely related properties. Indistinguishable photons share all relevant properties, such as color, polarization, and extension of the wave-packet in time.
Directly probing the indistinguishability of single photons is usually carried out in the interference experiment pioneered by Hong, Ou and Mandel \cite{HOM-PRL87}. For further details see also the chapter (by Kuhn, Zhao). It is manifested as a quantum interference effect when two photons arrive from different sides on a beam splitter: If these photons are indistinguishable and if they overlap in space and time (spatio-temporal overlap) on the beam splitter, quantum interference will force them always to exit through a common output port, which creates a path entangled (N=2) NOON-state. This is schematically sketched in Fig.~\ref{Abb:HOM}a and d). The fact that the photons leave the beam splitter in bunches with a suppression of the scenarios sketched in ~\ref{Abb:HOM}b and c) as a result of destructive quantum interference reflects their bosonic nature, and cannot be explained by classical electrodynamics.
\begin{figure}[h]
\centering
\includegraphics[width=12cm]{Schneider_Fig1.pdf}
\caption{\label{Abb:HOM} a-d) Possible configurations of two photons entering a beam splitter from different sides. Quantum interference determines the output paths of the photons. If the two input photons were indistinguishable, both photons can only leave the beam splitter in pairs through a common arm (a and d), rather than separate arms (b and c). e) Experimental implementation of the two photon interference experiment: The photons are emitted either from the same source (QD1), or from two separate sources (QD1 and QD2) mounted in different cryostats. In the latter case, the photons are brought together on a polarizing beam splitter (PBS). After spectral filtering, the light is fed into an asymmetric Mach-Zehnder interferometer via a polarizing fiber beam splitter (PFBS), and the two photon interference effect occurs on the last 50/50 beam splitter. Single photon counters (SPCs) are connected to each exit port of the beam splitter to record quantum correlations.
}
\end{figure}
The Hong-Ou-Mandel effect can be experimentally probed by utilizing a configuration of single photon detectors similarly to a Hanbury Brown and Twiss setup which is routinely used to probe the quantum nature of the emitted light \cite{Michler-1}. The setup is schematically sketched in Fig.~\ref{Abb:HOM}e). The photons can be emitted from a single source, or from separate, distant sources mounted in separate cryostats. In the latter case, the photon beams are brought together on a polarizing beam splitter and fed into an unbalanced Mach-Zehnder fiber interferometer after spectral filtering. One arm of the interferometer has a tunable length to adjust the arrival time of the photons on the second beam splitter, where the two photon interference is probed. Single photon detectors are connected to each output port of the beam splitter. The second order photon correlation function is recorded by measuring the time delays between the measurement events of the individual detectors. In case that the photons always take a common exit on the beam splitter, no coincidence detections between both APDs occur, and the second order correlation function approaches 0 at $\tau=0$. Here, we have to remember that the single photons emitted from our QD source are photon wave packets, which interfere on the beam splitter. Ideally, these wave packets are Fourier-transform limited, with a Lorentzian spectral broadening $\Delta \omega$ being solely determined by the emitter decay time $\tau_r$, and the temporal extension of the wave packet is given by $\tau_c=2 \times \tau_r$. If additional dephasing channels with a characteristic time $\tau_{deph}$, such as coupling to phonons start to play a role, the coherence time $\tau_c$ is reduced to $\frac{1}{\tau_c}=\frac{1}{2\tau_r}+\frac{1}{\tau_{deph}}$. For Fourier-transform limited wave packets, the second order correlation function should reach a value of 0 at $\tau=0$, which converts into a two photon indistinguishability of 100$\%$. To understand the shape of the second order correlation function around $\tau=0$ in the presence of dephasing, one has to calculate the overlap integral of two photons incident on the beam splitter. The correlation function of the $\tau=0$-peak for two photons with identical frequency and a time delay $\delta\tau$ between them is given by \cite{Bylander-EPJD03}
\begin{equation}\label{eq:g2hom}
g^{(2)}(\tau, \delta\tau)= \frac{1}{4} e^{-\frac{\left|\tau-\delta\tau\right|}{\tau_r}}+\frac{1}{4} e^{-\frac{\left|\tau+\delta\tau\right|}{\tau_r}}-\frac{1}{2}e^{-\left(\frac{2}{\tau_c}-\frac{1}{\tau_r}\right)\left|\tau\right|-\frac{\left|\tau-\delta\tau\right|}{2\tau_r}-\frac{\left|\tau+\delta\tau\right|}{2\tau_r}}.
\end{equation}
In equation (\ref{eq:g2hom}) a homogeneous broadening is assumed by an exponential decay of the coherent amplitude due to dephasing and the existence of (gaussian) spectral jitter is ignored.
\begin{figure}
\centering
\includegraphics[width=11cm]{Schneider_Fig2.pdf}
\caption{Calculated correlation function $g^{(2)}(\tau)$ of the $\tau=0$-Peak for the interference of indistinguishable photons with a radiative decay time of $\tau_r=1$ ns for different visibilities $v=\tau_c/2\tau_r$.}
\label{fig:Fig2}
\end{figure}
The correlation function $g^{(2)}(\tau)$ of the $\tau=0$-peak for the interference of photons with the same energy and polarization from a pulsed source is shown in Fig.~\ref{fig:Fig2} for different ratios of $v=\tau_c/2\tau_r$. We calculated $g^{(2)}(\tau)$ for a time delay of $\delta\tau=0$ and a constant radiative decay time of $\tau_r=1$ ns. The variation of $v$ is achieved by varying the coherence time $\tau_c$ between 2 ns and 0 ns. In the case of Fourier-tranform limited photons the visibility is 100\% and the $\tau=0$-peak disappears completely. For a homogeneous broadening of the emission, represented by a coherence time $\tau_c<2\tau_r$, the wavepackets do not overlapp perfectly which results in a non-zero contribution to the correlation funtion around $\tau=0$. In the limit of very short coherence times $\tau_c\rightarrow0$, the photons leave the beam splitter independently and randomly resulting from the reduced coalescence probability. The outcome of this is a peak in the correlation histogram with a $g^{(2)}(0)$ value of 0.5, which equals the $g^{(2)}(0)$-value for a two photon source $1-1/n$ with $n=2$.
\section{A bright quasi-planar single photon source}
\label{section:SPS}
\begin{figure}[h]
\centering
\includegraphics[width=12cm]{Schneider_Fig3.pdf}
\caption{\label{Abb:SPS1} a) Sample structure and quantum dots: A single layer of In(Ga)As QDs is integrated in a single sided low-Q cavity realized by AlAs/GaAs distributed Bragg reflectors sandwiching a GaAs cavity layer. b) The very low QD density facilitates the identification of QD emission spots straight forwardly by white light imaging. c) Nanohills on the sample surface are formed during the growth process acting as natural lenses and significantly improve light outcoupling of the structure. The figure is reproduced from Maier et al. \cite{Maier-OE14}.
}
\end{figure}
The single photon source which is at the heart of this study is based on a low density layer of single In(Ga)As QDs integrated in an asymmetric AlAs/GaAs optical microcavity. The lower distributed Bragg reflector (DBR) consists of 18 AlAs/GaAs mirrorpairs, providing a reflectivity near unity in the spectral range of the QD emission, which directs the light towards the top surface of the structure. A single layer of ultra low density InAs QDs is vertically centered in a GaAs $\lambda$-thick cavity layer, which is covered by 5 AlAs/GaAs distributed Bragg reflector (DBR) segments. The microcavity has a low quality factor (Q-factor) of $\approx$ 200. A combination of very low growth rates ($<$ 0.01 nm/s), and the partial capping and annealing growth technique \cite{Garcia-APL97} allows us to realize sufficiently low QD densities to spectroscopically isolate single QDs in the wavelength range between 900 nm and 940 nm. It is interesting to note that under such growth conditions, which allow for very long migration lengths of the supplied material, the QDs tend to nucleate at crystal steps, defects or nanoholes \cite{Schmidt-Buch, Michler-1}. This peculiar nucleation behavior can be directly exploited in QD positioning schemes \cite{Schmidt-Buch, Schneider-Nanotechnology09}, where nanodefects are intentionally generated on a surface via lithography and etching. In our case, the natural formation of oval crystal defects, which was most likely induced by Gallium droplets during the growth of the bottom DBR similar to observations in \cite{Zajac-PRB12}, serves as such nucleation sites for the QDs in the cavity layer. These defects propagate through the top DBR and are well detectable as nanohills on the the surface (Fig.~\ref{Abb:SPS1}c) via atomic force microscopy, with a height on the order of 10 nm.
Fig.~\ref{Abb:SPS1}b) depicts a CCD image of the sample surface under illumination with white light at a sample temperature of 4K. We used a long pass filter (750 nm) to monitor the emission from the QDs in the infrared range. The image is characterized by bright spots which we attribute to the emission of clusters of QDs, whereas no detectably signals occur between these sites.
Comparing the position of bright photoluminescence spots recorded via spatially resolved sample imaging with the nanohill position reveals a coalescence between the position of these hills and the location of QDs in the cavity. Furthermore, the oval shape of the hills provides a gentle optical lateral confinement \cite{Zajac-PRB12} which serves to guide the emitted light out of the semiconductor structure and enhance the photon outcoupling efficiency of the device. For a perfectly two dimensional microcavity structure of our geometry, this efficiency can hardly exceed theoretical values of $\approx 30\%$ when the light is collected in the normal direction with a 0.7 NA microscope objective \cite{Royo-JAP01}. In contrast, due to the waveguiding effect provided by the nanodefect, this efficiency can be theoretically increased to $\approx 50\%$, as described in \cite{Maier-OE14}. It is worth noting, that more carefully designed shapes of buried Gaussian nanohills are predicted to facilitate strong mode confinement to the sub-micrometer range without strongly reducing cavity Q-factor via lateral scattering losses \cite{Ding-PRB13}, which makes them very appealing for cavity quantum electrodynamic experiments. In our structure, we could experimentally determine single photon outcoupling efficiencies up to 42 $\%$ in a calibrated photoluminescence setup, being in good agreement with the numerical estimations based on realistic sample parameters, and exceeds the theoretical maximum of a perfectly planar two dimensional microcavity ($\approx 30\%$) \cite{Maier-OE14}. Most other strategies to achieve bright single photon emission from QDs embedded in a semiconductor matrix are based on the integration in nano- and microphotonic structures, such as pillar microcavities, photonic crystal membranes, nano-waveguides and antennas \cite{Barnes-EPJD02}. In all those approaches, the lithographic definition of the photonic structure creates open surfaces in the close vicinity to the quantum emitter, which can lead to significant dephasing and spectral wondering of the QD emission line. This is partly reflected in emitter line broadening \cite{Claudon-NatPhot10, Reimer-NatCom12}, limitations of the QD single photon interference properties \cite{Gazzano-NatCom13} and fast spin dephasing \cite{Press-Naturephotonics10}.
\section{Emission of single and indistinguishable photons from single quantum dots}
\label{section:experimentI}
\subsection{Single photon emission from single QDs}
\begin{figure}[h]
\centering
\includegraphics[width=12cm]{Schneider_Fig4.pdf}
\caption{\label{Abb:pumping-sketch} Excitation configurations of a single QD: a) Non-resonant wetting layer excitation creates carriers in the surrounding of the QD, which relax into the ground state by scattering. b) Quasi-resonant excitation leads to a direct excitation of the QD without generating a carrier reservoir. c) Resonance fluorescence selectively excites the QD ground state. For suppression of back scattered laser light in resonance fluorescence, the pathways for excitation and photon collection can be separated in the experimental implementation by an orthogonal arrangement for planar waveguides (d) and micropillars (e). f) Sketch for the cross-polarization configuration facilitating resonance fluorescence measurements under normal incidence. d) and (e) are reproduced from Muller et al. \cite{Muller-PRL07} and Ates et al. \cite{Ates-PRL09}. }
\end{figure}
There have been three different optical excitation method used for single-photon generation from QDs, as sketched in Fig. ~\ref{Abb:pumping-sketch}. The most conventional method is non-resonant excitation by pump laser with an energy above the band gap (in the barrier or the wetting layer surrounding the QDs, sketched in Fig.~\ref{Abb:pumping-sketch} a). These carriers can then be captured by the QD and relax to the ground state via phonon scattering, from where they can decay radiatively. In order to facilitate local generation of excitons in the QD and to reduce possibly detrimental effects from the surrounding, carriers can as well be generated quasi-resonantly in the excited states of the QD (Fig.~\ref{Abb:pumping-sketch} b). These states are typically located 20-50 meV on the high-energy side of the exciton ground state in the QD, facilitating spectral filtering of the excitation laser from the collected signal in most cased. The relaxation from the p-shell to the ground state of the QD typically occurs on the ten picoseconds scale \cite{Flagg-PRL12, Santori-Nature02}, which leads to a strong reduction of time jitter in the emission. The third pump-configuration is strictly resonant excitation (resonance fluorescence) a more controlled method widely used in standard atomic physics experiments (Fig.~\ref{Abb:pumping-sketch}c). Here, the excitation laser tuned on resonance with the QD transition coherently excites the QD. This excitation condition is by far the hardest to implement, since it requires a careful distinction between QD signal and laser stray light. Spectral filtering of the pump-laser is no longer feasible in strictly resonant excitation conditions, hence other methods have to be applied to isolate the QD emission signal: spatial filtering is one option, when the pump-laser is exciting the QD in the perpendicular direction to the collection beam-path \ref{Abb:pumping-sketch}d). This technique was first employed in Bragg waveguides and led to the first successful demonstration of resonance fluorescence from a single QD \cite{Muller-PRL07, Flagg-NatPhys2009}. Strictly resonant excitation perpendicular to the photon collection direction has also been carried out on single micropillar cavities with sufficiently strong suppression of scattered laser light facilitating two photon interference studies \cite{Ates-PRL09}. Another technique is polarization filtering: Here, the excitation laser is linearly polarized, and in case the reflected laser beam preserves this polarization, the QD emission can be detected in the perpendicular polarization basis \cite{Vamivakas-NatPhys09}. In our confocale microscope setup, the polarization extinction of the pump laser can reach values in excess of $10^7$ which allows us to carry out correlation measurements in the resonance fluorescence configuration \cite{He-NatNano13}. \\
Fig.~\ref{Abb:SPS2}a) depicts a photoluminescence spectrum recorded under non-resonant excitation condition. The spectrum is characterized by a pronounced emission feature which is attributed to the neutral exciton emission from a single QD via its spectral properties (such as polarization and power dependency). In the close spectral vicinity we detect a number of other lines, possibly stemming from different charge configuration of the same QD or from neighboring QDs. In the following, we will focus on the characteristics from the brightest transition: The corresponding second order autocorrelation function recorded from this line is shown in Fig.~\ref{Abb:SPS2}b), recorded under pulsed excitation with a repetition frequency of 82 MHz. During each excitation pulse, carriers are excited non-resonantly in the wetting layer of the QDs, which is reflected in the distance between the peaks in the correlation histogram. As expected from a single quantum emitter, at $\tau=0$ the coalescence probability is strongly suppressed and a dip occurs in the histogram. The corresponding $g^{(2)}(\tau=0)$ value of 0.05 is a clear signature of single photon emission.
\begin{figure}[H]
\includegraphics[width=12cm]{Schneider_Fig5.pdf}
\caption{\label{Abb:SPS2} a) Photoluminescence (Inset: Time resolved PL) signal of a QD under non-resonant excitation; c) quasi-resonant excitation; e) resonance fluorescence. The inset in e) depicts the power dependency of the emission under pulsed resonance fluorescence, which is characterized by the occurrence of distinct Rabi-oscillations. Corresponding second order correlation histograms: b) non-resonant pumping, d) p-shell excitation; f) resonance fluorescence.).
}
\end{figure}
However, as a result of the non-resonant excitation, a large number of carriers is generated in the surrounding of the QD during one excitation pulse. If the lifetime of this reservoir exceeds the average recombination time of the QD trion, carriers can re-excite the QD after the first recombination event, which leads to a strong broadening of the peaks in the correlation histogram. \\
This effect is largely suppressed when the QD is quasi-resonantly excited into an excited state: In Fig \ref{Abb:SPS2}c) we plot the photoluminescence spectrum of a QD which is excited with a laser detuning of 29 meV on the high energy side of the recombination line. Stray-light from the excitation laser is spectrally suppressed by a combination of bandpass filters. Due to the quasi-resonant nature of the excitation, the spectrum is almost background free and a single bright emission line dominates the spectrum over a wide range. More importantly, the effects of strong time jitter and carrier recapturing are suppressed in the corresponding correlation histogram (Fig \ref{Abb:SPS2}d) and the width of the coincident peaks in the histogram are now determined by the lifetime of the excitonic transition with a characteristic time of 700 ps, which is in good agreement with the time resolved PL trace (inset of Fig \ref{Abb:SPS2}c)). This effective reduction of the measured lifetime (compared to inset of Fig \ref{Abb:SPS2}a) is again consequence of the absence of carrier recapturing under quasi-resonant excitation.
The purity of the source is even improved, as characterized by a value of $g^2(\tau=0)=0.023$ which was directly extracted from the raw data without any background correction. \\
A truly coherent, time-jitter free excitation method made use of an ultrafast (3 ps) pulsed laser with its central frequency resonant with the QD transition.
In Fig.~\ref{Abb:SPS2}e) we plot a spectrum of the QD-emission signal under such resonance flourescence (RF) conditions. The narrow emission line stemming from the driven QD resonance sits on top of a broad, yet dim background from the pump laser (plotted in log-scale). The much broader laser background can be further filtered using a narrow-band etalon, resulting in a signal to background ratio exceeding $300:1$ \cite{He-NatNano13}. The QD emission intensity is plotted as a function of the square root of the pump power of the excitation laser in the inset of Fig.~\ref{Abb:SPS2}e): The observed Rabi-oscillations are characteristic for a resonantly driven two-level system and reflect the coherent nature of the excitation process. Single RF photons are deterministically generated at the peak maximum, which corresponds to a $\pi$-pulse in the picture of driven Rabi-oscillations.
The according second order correlation measurement under these conditions is shown in Fig.~\ref{Abb:SPS2}f). In order to further filter out the broad background from the excitation laser, the emission is fed through a high finesse Fabry-Perot ethalon. Under such configurations, the second order autocorrelation function can reach values as low as $g^2(\tau=0)=0.003$, pointing out the character of our single QD as a almost perfect single photon source \cite{Wei-arxiv14}.
\subsection{Two photon interference with single photons}
The combination of high brightness and high purity of the single photon emission allows us to carry out photon interference measurements with the experimental configuration briefly described in section \ref{section:theory} and chapter (by Lanco i Senellart).
First, we study the two-photon interference (TPI) of consecutive photons emitted from a single QD, to scrutinize the dependency of the interference visibility on the excitation condition. Therefore, we adjust the path length difference of the two arms in our Mach-Zehnder interferometer to the laser repetition period of 12.2 ns (see Fig. \ref{Abb:HOM}d), so that two consecutively excited photons can coincide at the same time on the beam-splitter. Additionally, there is the possibility to change the time delay between the two arms of the interferometer via a variable optical fiber delay.
\begin{figure}[t]
\centering
\includegraphics[width=12cm]{Schneider_Fig6.pdf}
\caption{\label{Abb:ID1} a) Two photon interference histogram for a QD. The emitter is non-resonantly excited into the wetting layer. b) TPI under p-shell excitation. c) The Hong-Ou-Mandel dip evolves when the TPI visibility is plotted against the delay time $\Delta t$.
}
\end{figure}
We first study the TPI of consecutive emitted photons under wetting layer excitation. As we have discussed above, the very long diffusion lengths of our sample lead to a recapturing of charge carriers after a first recombination. This recapturing results in a background in the autocorrelation histogram at $|\tau|>0$. The corresponding correlation histogram for the TPI of consecutive emitted photons at an interferometer path length difference of $\Delta t \approx 0$ is shown in Fig. \ref{Abb:ID1}a). Fitting the data with a model based on Ref. \cite{Bylander-EPJD03} yields a visibility of only 12$\%$. This small value is a result of the very large time uncertainty induced by the long emission time induced from the carrier recapturing, which makes a simultaneous collision of two photons on the beam splitter very unlikely.
As a direct comparison, in Fig. \ref{Abb:ID1}b) the second order correlation function for TPI is shown for zero path length difference for a QD under p-shell excitation. Here, the peak at $\tau=0$ is strongly suppressed below a value of 0.5. The probability for two photons that coincide at the beamsplitter and exit in opposite directions $g_{indist}^{(2)} (\tau=0)$ is given by the area under the peak at $\tau=0$ divided by the averaged area of four peaks for $|\tau|>\pm12.2$ ns. From the raw data we extract values of $g_{indist}^{(2)}=0.16 <0.5$, verifying the indistinguishability of the photons generated under quasi resonant pumping.
In order to accurately extract the visibility of the two photon interference, we investigate $g_{indist}^{(2)} (\tau=0)$ in dependence of interferometer path length offset $\Delta t$. In this manner, we observe the characteristic Hong-Ou-Mandel dip for $\Delta t =0$.
Via fitting the data with a two sided exponential $g_{indist}^{(2)} (\Delta t)=0.5[1- v e^{(-|\Delta t|/\tau_m )} ]$ we can extract a non-postselected value of TPI visibility $v$ of 69$\%$. This value is comparable to the values of QDs embedded in micropillar cavities \cite{Santori-Nature02, Gazzano-NatCom13}, where the Purcell effect is employed to reduce the radiative decay time $\tau_r$ and hence improve $\frac{\tau_c}{2\tau_r}$.
The visibility under quasi resonant excitation is strongly increased compared to the non-resonant excitation scheme resulting from the direct excitation conditions which lead to a reduced uncertainty in the emission time, a lack of carrier re-capturing processes and reduced charge carriers in the wetting layer. For an ideal spontaneous-emission source, with instantaneous initial excitation and no decoherence, $\tau_m$ which characterizes the arrival time of photons on the beam splitter would be equal to the spontaneous emission lifetime of the quantum emitter. From the fit we get values of $\tau_m = 630$ ps which indeed is close to the spontaneous emission lifetime of this QD (see inset Fig. \ref{Abb:SPS2}c)).
\begin{figure}[h]
\centering
\includegraphics[width=12cm]{Schneider_Fig7.pdf}
\caption{\label{Abb:T2} a) Coherence time and b) decay time of the QD characterized in Fig~\ref{Abb:ID1}.
}
\end{figure}
Ideally, in order to obtain maximum degrees of indistinguishability, Fourier-transform limited sources are required. As we have assessed in \ref{section:theory}, the relation between coherence time $\tau_c$ and lifetime of the QD emission $\tau_r$, that describes the visibility of TPI is $\nu_{max}=\frac{\tau_c}{2\tau_{r}}=1$. In order to directly assess the coherence time of the QD emission under p-shell excitation, we use a Michelson interferometer, and measure the photon interference signal as a function of the variable time delay (see \cite{Michler-2} for details on the method). The fringe contrast in dependence of interferometer path length difference is shown in Fig. \ref{Abb:T2} a). The fine structure splitting of the neutral exciton line leads to oscillations in the interference fringe contrast. From a fit (red solid line) to the experimental data with the Fourier transform of two Lorentzians we can extract the coherence times of the fine structure split lines $\tau_{c1}=330$ ps and $\tau_{c2}=180$ ps. Compared with the extracted decay time $\tau_r=670$ ps we can extract a maximum visibility for TPI in this case of 25$\%$.
This seeming discrepancy to the extracted value of 69 $\%$ has been observed before \cite{Santori-Nature02} and can be explained by a primary inhomogeneous broadening of the emission lines, for example by charge fluctuations in the vicinity of the QD. These fluctuations can take place on a timescale much longer than the laser repetition frequency, and are hence not affecting the TPI measurement, since only the interference from consecutive photons is measured. On the other hand, in the time averaged Michelson experiment, the interference of photons emitted at much larger time delays contribute, which explains the reduced coherence times. As we will show in the following chapter, however, this argument is no longer true if photons from independent sources are interfered, which asks for the capability to generate photons close to the Fourier limit.
To generate such photons and increase the visibility of the TPI, we study a QD under pulsed resonance fluorescence conditions. In this experiment, each excitation pulse of the pumplaser is split into two pulses with a delay of 2 ns, generating two single photons each 12.5 ns (see Fig~\ref{Abb:RF} a). The according correlation histogram from the TPI is depicted in Fig.~\ref{Abb:RF}b) and c). If we combine photons with opposite polarizations on the last beam splitter, we observe the correlation histogram of two perfectly distinguishable photons depicted in Fig~\ref{Abb:RF}b), featuring a central peak at $\tau=0$ with the same magnitude as the neighboring peaks stemming from photons with a time difference of 2 ns. If photons with the same polarization are combined on the beam splitter (Fig.~\ref{Abb:RF}c)) the strong suppression of the central peak indicates the high degree of indistinguishability of the photons generated under these conditions. By evaluating the areas under the coincidence peaks, we can directly extract a raw TPI visibility as high as 91 $\%$, clearly exceeding the value for quasi-resonant excitation. A more recent experiment used adiabatic rapid passage to deterministically and more robustly generate single photons and demonstrated a new record of two-photon interference raw visibility of about 98$\%$ \cite{Wei-arxiv14}.
The strong increase of the TPI for resonance fluorescence clearly underlines the superiority of this excitation scheme, and points towards the possibility to deterministically generate single photons close to the Fourier-limit.
\begin{figure}[h]
\centering
\includegraphics[width=12cm]{Schneider_Fig8.pdf}
\caption{\label{Abb:RF} Photon interference spectroscopy in the resonance fluorescence configuration: a) Sketch of the optical setup: Each pulse of the excitation laser is split into two pulses with a time delay of 2 ns, resulting in a two-fold excitation of the same QD. The fluorescence is fed into a Hong-Ou-Mandel setup, and the RF-photons are recombined on the second beamsplitter. b) Interference histogram with photons of perpendicular polarizations and c) photons of the same polarization. The absence of the peak at $\tau=0$ demonstrates highly indistinguishable photons.
}
\end{figure}
\section{Two photon interference from remote, single quantum dots}
\label{section:experimentII}
\begin{figure}[h]
\centering
\includegraphics[width=12cm]{Schneider_Fig9.pdf}
\caption{\label{Abb:TPI} Two photon interference of QDs emitted from separate sources. The QD samples are mounted in separate cryostats with an overall spatial separation of 0.5 m. a) Interference histogram under spectral resonance. The QDs are tuned into resonance via adjusting the sample temperature. b) Dependency of the two photon interference visibility on the spectral detuning of the QDs. c) Theoretical maximum of the indistinguishability as a function of the QD coherence time (in relation to the decay time $\tau_r$), under consideration of homogeneous (dashed line) and inhomogeneous (solid line) broadening mechanisms.
}
\end{figure}
The very high TPI visibilities which we have discussed in the previous section puts the observation of pronounced TPI effects from photons emitted from separate sources clearly within reach. Such interference effects have previously been observed under non-resonant exictation conditions \cite{Flagg-PRL10} \cite{Patel-NatPhot10}, or under CW resonant fluorescence excitation resulting in time post selecting \cite{Konth-PRL12}. For realistic applications, the non-postselected value of the TPI is however of greater importance, which can only be probed under pulsed excitation conditions. In order to carry out this experiment, we have installed two QD samples in separate cryostats and identified individual QDs with similar emission frequencies and coherence properties as the QD characterized in Fig.~\ref{Abb:ID1}c). Both QDs were excited quasi-resonantly with the same pulsed excitation laser, and the emission was combined on a polarizing beam splitter before it was fed into the Mach-Zehnder interferometer (see. Fig. \ref{Abb:HOM}a).
In order to probe the two photon interference from separate QDs, the emitters have to be tuned to spectral resonance. A number of tuning mechanisms are possible, including electric fields \cite{Patel-NatPhot10}, strain \cite{Flagg-PRL10} or magnetic fields \cite{Reitzenstein-PRL09}. In our experiments, we utilize the sample temperature to spectrally tune the QD energies to resonance. Since the temperature in the cryostats cannot be varied in a wide range without detrimentally affecting the QD emission properties, a pair of QDs was selected with an energy difference as small as 3 $\mu$eV. The emission energies could then be equalized by changing the temperature of one sample by only 1.8 K, which should only minorly affect coherence properties of the emission.
The second order correlation function for TPI from separate sources is shown in Fig. \ref{Abb:TPI}a) for QD1 at 5.00 K and QD2 at 6.75 K. We determine the opposite output probability $g_{indist}^{(2)} (\tau=0)$ from the raw data by the area of the peak at $\tau=0$ divided by the averaged area of 6 peaks for $|\tau|>0$. From the data we extract $g_{indist}^{(2)} (\tau=0)=(0.31 \pm 0.01)$, which verifies that the photons from the two QDs have a nonzero coalescence probability. The according TPI visibility amount to $v=(39\pm2)\%$, which is the highest value of non-postselected two photon interference from separate QDs observed so far under quasi- or nonresonant excitation conditions. By varying the sample temperature in one cryostat, we can tune the QDs out of resonance, which is directly reflected in a reduced two photon interference visibility, as depiced in Fig.\ref{Abb:TPI}b).
We will now compare these experimentally observed values with a theory which only takes into account pure dephasing as a decoherence mechanism. As we have described in section \ref{section:theory}, in the presence of pure dephasing limiting the coherence time $\frac{1}{\tau_c}=\frac{1}{2\tau_r}-\frac{1}{\tau_{deph}}$ the maximum visibility of the TPI is obtained for $v = \frac{\tau_c}{2\tau_r}$. Taking into account the experimentally extracted radiative decay time of 670 ps (Fig. \ref{Abb:T2}b) and coherence time of 330 ps (Fig. \ref{Abb:T2}a), we can infer a maximum interference visibility of 25 $\%$, which is clearly exceeded in our experiment. For two photons emitted from the same QD, we argued that frequency jitter on a time scale beyond the repetition time of the pulsed excitation laser led to a reduction of the coherence time which was however only weakly affecting the interference visibility from consecutive photons. For photons emitted from independent sources, clearly this argumentation is not valid. In order to take account for the effects of inhomogeneous broadening in the evaluation of the TPI, various frequency components have to be taken into account. We follow the analysis by Legero et al. (see chapter (Kuhn, Zhao) of this book and \cite{Legero-AAMOP06}) to derive an expression for the visibility of TPI in the presence of inhomogeneous broadening represented by a Gaussian frequency distribution. We assume that the two photons which interfere at the beam splitter originate from independent ensembles of Fourier transform limited photons. The Fourier limited single-photon wave packets for QD1 and QD2 are one sided exponential functions:
\begin{equation}
\xi_1(t)=\begin{cases} \sqrt[4]{\frac{1}{\pi\tau_r}}e^{-\frac{t-\frac{\delta\tau}{2}}{2\tau_r}-i\left(\omega-\frac{\Delta}{2}\right)t} & \text{if }t-\frac{\delta\tau}{2}>0\\
0 &\text{otherwise}
\end{cases}
\end{equation}
and
\begin{equation}
\xi_2(t)=\begin{cases} \sqrt[4]{\frac{1}{\pi\tau_r}}e^{-\frac{t+\frac{\delta\tau}{2}}{2\tau_r}-i\left(\omega+\frac{\Delta}{2}\right)t} & \text{if }t+\frac{\delta\tau}{2}>0\\
0 &\text{otherwise}
\end{cases},
\end{equation}
where $\delta\tau$ is the time delay and $\Delta$ the frequency difference between them. An inhomogeneous broadening of the emission lines can be considered by a Gaussian frequency distribution $f(\omega)$ with $\sigma^2$ being the variance:
\begin{equation}
f_{i=1,2}(\omega_i)=\frac{1}{\sqrt{\pi} \, 2 \sigma_i}e^{-\frac{\left(\omega_{0i}-\omega_i\right)^2}{2 \sigma_i^2}}.
\end{equation}
With $\omega_1=\omega$ and $\omega_2=\omega+\Delta$ we get the frequency distribution as a function of the frequency difference $\Delta$
\begin{equation}
f(\Delta)=\int\mathrm{d}\omega \, f_1(\omega)f_2(\omega,\,\Delta)=\frac{1}{\sqrt{\pi} \, 2\sigma_g}e^{-\frac{\left(\Delta-\Delta_0\right)^2}{4\sigma^2_g}},
\end{equation}
with $\sigma_g=\sqrt{\sigma_1^2+\sigma_2^2}$ and $\Delta_0=\omega_{02}-\omega_{01}$. The correlation function is then given by
\begin{equation}
G_{\mathrm{inhom}}^{(2)}(t_0,\,t_0+\tau)=\int\mathrm{d}\Delta \, f(\Delta)\,\mathrm{tr}\!\left(\hat{\rho}(\xi_1,\xi_2) \, \hat{A}(t_0,\,t_0+\tau)\right),
\end{equation}
where $\mathrm{tr}\!\left(\hat{\rho}(\xi_1,\xi_2) \, \hat{A}(t_0,\,t_0+\tau)\right)$ is the correlation function for two Fourier transform limited photons with the same polarization:
\begin{equation}
G_{\mathrm{TL}}^{2}(t_0,t_0+\tau)=\frac{|\xi_1(t_0)\xi_2(t_0+\tau)-\xi_2(t_0)\xi_1(t_0+\tau)|^2}{4}
\end{equation}
Finally, the probability for detecting a photon at time $t_0 + \tau$ in one output of the beam splitter while a photon is detected at time $t_0$ in the other one for inhomogeneous
broadened ensembles of photons is given by
\begin{align}
P_{\mathrm{inhom}}&=\int^{\infty}_{-\infty}\!\mathrm{d}t_0 \, G_{\mathrm{inhom}}^{(2)}(t_0,\,t_0+\tau)= \nonumber\\
&=\frac{1}{8\tau_r}\left(e^{-\frac{\left|\delta\tau-\tau\right|}{\tau_r}}+e^{-\frac{\left|\delta\tau+\tau\right|}{\tau_r}}-2\cos\!\left(\Delta_0\,\tau\right)e^{-\frac{\left|\delta\tau\right|+\left|\tau\right|}{\tau_r}}\,e^{-\sigma_g^2\tau^2}\right)
\end{align}
From this expression, we yield an expression for the two photon interference visibility for $\delta \tau=0$ and $\Delta_0=0$, only depending on the radiative decay time $\tau_r$ and the geometric average of the broadening of the two photon ensembles $\sigma_g=\sqrt{\sigma_1^2+\sigma_2^2}$:
\begin{equation}
v_{\mathrm{inhom}}=1-\frac{1}{\tau_r \sigma_g} \left( 2 \tau_r \sigma_g - e^{\frac{1}{4\tau_r^2 \sigma_g^2}} \sqrt{\pi} \mathrm{erfc}\left(\frac{1}{2 \tau_r \sigma_g}\right)\right)
\label{eq:vis}
\end{equation}
In Fig. \ref{Abb:TPI}c) we plot the maximum TPI of such inhomogeneously broadened wave packets as a function of the coherence time (in multiples of the lifetime $\tau_r$), to visualize the strong influence broadening's origin on the maximum interference visibilty. From this analysis, we can estimate a maximum visibility of $\nu_{inhom}=(36.4\pm 1.5)\%$ which is in good agreement with the experiment.
This underlines the importance of effects as spectral wandering, time jitter and other inhomogeneous broadening channels in particular for photon interference experiments from independent sources. It is worth noting, that major improvements have recently been accomplished by utilizing resonance fluorescence conditions in such an experiment. Due to the suppression of inhomogeneous broadening effects under strict resonant excitation, two photon interference visibilities beyond $80\%$ \cite{Gao-NatCom13, He-PRL13} could be obtained.
\subsection{Conclusion}
Single semiconductor quantum dots have been established as compact single photon sources on a solid state platform. Towards the implementation of these quantum emitters as sources of highly indistinguishable photons, which is key to realize quantum teleportation schemes and highly desired quantum repeaters, the degree of indistinguishability of the photon emission is a key parameter. As we have reviewed in this chapter, besides utilizing the effects of cavity quantum electrodynamics to modify the radiative decay time of the photon emission, the appropriate excitation scheme plays a crucial role to realize high degrees of indistinguishabilities. In particular resonance fluorescence conditions can be considered as a reliable technique to generate single photons near unity indistinguishability. We highly anticipate that a combination of such sophisticated pumping schemes and the exploitation of light matter coupling effects can lead to even further simultaneous improvements of the photon coupling efficiencies and degrees of indistinguishability, which makes single QDs an truly appealing alternative to cold atoms and ions towards the realization of quantum repeaters.
\
\section{Acknowledgements} %
\begin{acknowledgement}
The authors acknowledge the great support of the following persons throughout the last years:
S. Maier, A. Thoma, Y. He, Y.-M. He, N. Gregersen, J. Mork, J. Schary, M. Lermer, M. Wagenbrenner, L. Worschech, S. Reitzenstein and A. Forchel. We acknowledge financial support by the BMBF (Projects QuaHLRep and Q.com-H) as well as the state of Bavaria.
\end{acknowledgement}
\section{References}
\bibliographystyle{SpringerPhysMWM}
|
\subsection*{Extended abstract}
We introduce and study a certain class of quadratic algebras,
which are nonhomogeneous in general,
together with the distinguish set of mutually commuting elements inside of
each, the so-called {\it Dunkl elements}. We describe relations among the
Dunkl elements in the case of a family of quadratic algebras corresponding to
a certain splitting of the {\it universal classical Yang--Baxter relations}
into two {\it three term relations}. This result is a further extension
and generalization of analogous results obtained in~\cite{FK,P} and~\cite{KM}. As an application we describe explicitly the set of relations among
the Gaudin elements in the group ring of the symmetric group, cf.~\cite{MTV}.
We also study relations among the Dunkl elements in the case of
(nonhomogeneous) quadratic algebras related with the {\it universal dynamical
classical Yang--Baxter relations}. Some relations of results obtained in
papers~\cite{FK,K,KM3} with those obtained in~\cite{GRTV}
are pointed out. We also identify a subalgebra generated by the generators
corresponding to the {\it simple roots} in the extended Fomin--Kirillov
algebra with the $\rm DAHA$, see Section~\ref{section4.3}.
The set of generators of algebras in question, naturally corresponds to the
set of edges of the {\it complete graph} $K_n$ (to the set of edges and loops
of the complete graph with (simple) loops~${\widetilde{K}}_n$ in
{\it dynamical} and {equivariant} cases). More generally, starting from any
subgraph $\Gamma$ of the complete graph with simple loops~${\widetilde{K}}_{n}$ we def\/ine a (graded) subalgebra $3T_n^{(0)}(\Gamma)$ of
the (graded) algebra $3T_n^{(0)}({\widetilde{K}}_n)$~\cite{K2}. In the case
of loop-less graphs $\Gamma \subset K_n$ we state {\it conjecture},
Conjecture~\ref{conjecture4.2} in the main text, which relates the {\it Hilbert polynomial}
of the abelian quotient~$3T_n^{(0)}(\Gamma)^{ab}$ of the algebra $3T_n^{(0)}(\Gamma)$ and the {\it chromatic polynomial} of the graph~$\Gamma$ we are
started with\footnote{We {\it expect} that a similar conjecture is true for any f\/inite
(oriented) matroid~$\cal{M}$. Namely, one (A.K.) can def\/ine an analogue of the
three term relations algebra $3T^{(0)}({\cal{M}})$ for any (oriented) matroid~$\cal{M}$. We {\it expect} that the abelian quotient
$3T^{(0)}({\cal{M}})^{ab}$ of the algebra $3T^{(0)}({\cal{M}})$ is
isomorphic to the {\it Orlik--Terao}
algebra~\cite{OTe}, denoted by ${\rm OT}({\cal{M}})$ (known also as {\it even}
version of the Orlik--Solomon algebra, denoted by ${\rm OS}^{+}({\cal{M}})$ )
associated with matroid~$\cal{M}$~\cite{Cor}.
Moreover, the anticommutative quotient of the {\it odd} version of the
algebra $3T^{(0)}({\cal{M}})$, as we {\it expect}, is isomorphic to the Orlik--Solomon algebra ${\rm OS}({\cal{M}})$ associated with matroid~${\cal{M}}$, see, e.g.,
\cite{B, GR}. In particular,
\begin{gather*}
{\rm Hilb}(3T^{(0)}\big({\cal{M}})^{ab},t\big)= t^{r({\cal{M}})} {\rm Tutte}\big({\cal{M}}; 1+t^{-1},0\big).
\end{gather*}
We {\it expect} that the Tutte polynomial of a matroid, ${\rm Tutte}({\cal{M}},x,y)$,
is related with the Betti polynomial of a~matroid $\cal{M}$. Replacing
relations $u_{ij}^2=0$, $\forall\, i, j$, in the def\/inition of the algebra
$3T^{(0)}(\Gamma)$ by relations $u_{ij}^2=q_{ij}$, $\forall\, i, j$, $(i,j) \in E(\Gamma)$, where $\{q_{ij}\}_{(i,j) \in E(\Gamma)}$, $q_{ij}=q_{ji}$, is a~collection of {\it central} elements, give rise to a~{\it quantization} of the Orlik--Terao algebra ${\rm OT}(\Gamma)$. It seems an interesting {\it task} to clarify
combinatorial/geometric signif\/icance of {\it noncommutative} versions of
Orlik--Terao algebras (as well as Orlik--Solomon ones) def\/ined as follows:
${\cal{OT}}(\Gamma) := 3T^{(0)}(\Gamma)$, its ``quantization''~$3T^{({\boldsymbol{q}})}(\Gamma)^{ab}$ and $K$-theoretic analogue $3T^{({\boldsymbol{q}})}(\Gamma, \beta)^{ab}$, cf.\ Def\/inition~\ref{definition3.1}, in the theory of
hyperplane arrangements.
{\it Note} that a small modif\/ication of arguments in~\cite{Li} as were used
for the proof of our Conjecture~\ref{conjecture4.2}, gives rise to a theorem that the
algebra $3T_n(\Gamma)^{ab}$ is isomorphic to the Orlik--Terao algebra~${\rm OT}(\Gamma)$ studied in~\cite{ScT}.\label{footnote1}}\footnote{In the case of simple graphs our Conjecture~\ref{conjecture4.2} has been proved in~\cite{Li}.}.
We check our conjecture for the complete graphs~$K_n$ and the complete bipartite graphs~$K_{n,m}$. Besides, in the case of
{\it complete multipartite graph $K_{n_1,\ldots,n_r}$}, we identify the
commutative subalgebra in the algebra $3T_{N}^{(0)}(K_{n_1,\ldots,n_r})$,
$N=n_1+\cdots+n_r$, generated by the elements
\begin{gather*}
\theta_{j,k_j}^{(N)}:=e_{k_j}\big(\theta_{N_{j-1}+1}^{(N)},\ldots,
\theta_{N_{j}}^{(N)}\big), \\ 1 \le j \le r, \quad 1 \le k_j \le n_j, \quad N_j:=n_1+\cdots+
n_j, \quad N_0=0,
\end{gather*}
with the cohomology ring $H^{*}({\cal{F}}l_{n_1,\ldots,n_r},\mathbb{Z})$ of the
partial f\/lag variety ${\cal{F}}l_{n_1,\dots,n_r}$. In other words, the set of
(additive) Dunkl elements $\big\{ \theta_{N_{j-1}+1}^{(N)},\ldots,
\theta_{N_j}^{(N)}\big\}$ plays a role of the {\it Chern roots} of the
tautological vector bundles~$\xi_j$, $j=1,\ldots,r$, over the partial f\/lag
variety ${{\cal{F}}l}_{n_1,\ldots,n_r}$, see Section~\ref{section4.1.2} for details.
In a similar fashion, the set of {\it multiplicative} Dunkl elements
$\big\{\Theta_{N_{j-1}+1}^{(N)},\ldots,\Theta_{N_{j}}^{(N)} \big\}$ plays a role of
the {\it $K$-theoretic version of Chern roots} of the tautological vector
bundle~$\xi_j$ over the~partial f\/lag variety ${\cal{F}}l_{n_1,\ldots,n_r}$.
As a byproduct for a given set of weights ${\boldsymbol{\ell}} = \{ \ell_{ij} \}_{1 \le i < j \le r}$ we compute the {\it Tutte polynomial}
$T(K_{n_1,\ldots,n_k}^{({\boldsymbol{\ell}})},x,y)$ of the ${\boldsymbol{\ell}}$-weighted
complete multipartite graph $K_{n_1,\ldots,n_k}^{({\boldsymbol{\ell}})}$, see
Section~\ref{section4}, Def\/inition~\ref{section4.4} and Theorem~\ref{section4.3}.
More generally, we introduce {\it universal Tutte polynomial}
\begin{gather*}
T_n(\{q_{ij}\},x,y) \in \mathbb{Z}[\{q_{ij}\}] [x,y]
\end{gather*}
in such a way that for any collection of non-negative integers ${\boldsymbol{m}}=
\{ m_{ij} \}_{1 \le i < j \le n}$
and a subgraph $\Gamma \subset K_n^{({\boldsymbol{m}})}$ of the weighted complete
graph on $n$ labeled vertices with each edge $(i,j) \in K_n^{({\boldsymbol{m}})}$
appears with multiplicity~$m_{ij}$, the specialization
\begin{gather*}
q_{ij} \longrightarrow 0 \quad \text{if edge} \ \ (i,j) \notin \Gamma,\qquad
q_{ij} \longrightarrow [m_{ij}]_y := \frac{y^{m_{ij}}-1}{y-1} \quad \text{if edge} \ \ (i,j)
\in \Gamma
\end{gather*}
of the universal Tutte polynomial is equal to the Tutte polynomial of graph
$\Gamma$ multiplied by $(x-1)^{\kappa(\Gamma)}$, see Section~\ref{section4.1.2},
Theorem~\ref{theorem4.3}, and {\it comments and examples}, for details.
We also introduce and study a family of {\it $($super$)$ $6$-term relations}
algebras, and suggest a~def\/inition of ``multiparameter quantum
deformation'' of the algebra of the curvature of $2$-forms of the Hermitian
linear bundles over the complete f\/lag variety ${\cal{F}}l_n$. This algebra
can be treated as a natural generalization of the (multiparameter) quantum
cohomology ring $QH^{*}({\cal{F}}l_n)$, see Section~\ref{section4.2}. In a similar fashion as in the case of three term relations algebras, for any subgraph~$\Gamma \subset K_n$, one~(A.K.) can also def\/ine an algebra
$6T^{(0)}(\Gamma)$ and projection\footnote{We treat this map as an algebraic version of the homomorphism which
sends the curvature of a Hermitian vector bundle over a smooth algebraic variety to its cohomology class, as well as a splitting of classical Yang--Baxter
relations (that is six term relations) in a couple of three term relations.}
\begin{gather*}
\text{Ch}\colon \ 6T^{(0)}(\Gamma) \longrightarrow 3T^{(0)}(\Gamma).
\end{gather*}
Note that subalgebra ${\cal{A}}(\Gamma):= {\mathbb{Q}}[\theta_1,\ldots,\theta_n] \subset 6T^{(0)}(\Gamma)^{ab}$
generated by additive Dunkl elements
\begin{gather*}
\theta_i = \sum_{j \atop (ij) \in E(\Gamma)} u_{ij}
\end{gather*}
is closely related with problems have been studied in \cite{PSS, SS}, \dots, and \cite{T} in the case $\Gamma= K_n$, see Section~\ref{section4.2.2}. We want to draw
attention of the reader to the following {\it problems} related with {\it arithmetic Schubert\footnote{See for example~\cite{T} and the literature quoted therein.}
and Grothendieck calculi}:
\begin{enumerate}\itemsep=0pt
\item[(i)] Describe (natural) quotient $6T^{\dagger}(\Gamma)$ of the algebra
$6T^{(0)}(\Gamma)$ such that the natural epimorphism
${\rm pr}\colon {\mathbb{A}}(\Gamma) \longrightarrow {\cal{A}}(\Gamma)$ turns out to be {\it isomorphism}, where we denote by
${\mathbb{A}}(\Gamma)$ a~subalgebra of~$6T^{\dagger}(\Gamma)$ generated over~${\mathbb{Q}}$ by additive Dunkl elements.
\item[(ii)] It is not dif\/f\/icult to see \cite{K} that {\it multiplicative} Dunkl
elements $\{ \Theta_i \}_{1 \le i \le n}$ also mutually commute in the algebra~$6T^{(0)}$, cf.\ Section~\ref{section3.2}. {\it
Problem} we are interested in is to describe commutative subalgebras generated by {\it multiplicative} Dunkl elements in the algebras~$6T^{\dagger}(\Gamma)$ and~$6T^{(0)}(\Gamma)^{ab}$. In the latter case one will come to the $K$-theoretic version of algebras studied in~\cite{PSS}, \dots.
\end{enumerate}
Yet another objective of our paper\footnote{This part of our paper had its origin in the study/computation of
relations among the additive and multiplicative Dunkl elements in the quadratic
algebras we are interested in, as well as the author's attempts to construct
a monomial basis in the algebra~$3T_n^{(0)}$ and f\/ind its Hilbert series for~$n \ge 6$. As far as I'm aware these {\it problems} are still widely open.}
is to describe several combinatorial
properties of some special elements in the associative quasi-classical
Yang--Baxter algebras~\cite{K}, including among others, the so-called
{\it Coxeter element} and the {\it longest element}. In the
{\it case} of {\it Coxeter element} we relate the corresponding reduced
polynomials introduced in~\cite[Exercise~6.C5(c)]{ST3}, and independently in
\cite{K}, cf.~\cite{K2}, with the $\beta$-Grothendieck polynomials~\cite{FK1} for some special permutations~$\pi_{k}^{(n)}$. More generally, we
identify the $\beta$-Grothendieck polynomial~$\mathfrak{G}_{\pi_{k}^{(n)}}^{(\beta)}(X_n)$ with a~certain weighted sum running over the set of $k$-dissections of a convex $(n+k+1)$-gon. In particular we show that the specialization
$\mathfrak{G}_{\pi_{k}^{(n)}}^{(\beta)}(1)$ of
the $\beta$-Grothendieck polynomial $\mathfrak{G}_{\pi_{k}^{(n)}}^{(\beta)}(X_n)$ counts the number of {\it $k$-dissections} of a convex $(n+k+1)$-gon
according to the number of diago\-nals involved. When the number of diagonals in
a $k$-dissection is the maximal possible (equals to $n(2 k-1)-1$), we
recover the well-known fact that the number of $k$-triangulations of a convex
$(n+k+1)$-gon is equal to the value of a certain Catalan--Hankel determinant,
see, e.g.,~\cite{SS}. In Section~\ref{section5.4.2} we study multiparameter generalizations
of reduced polynomials associated with Coxeter elements.
We also show that for a certain $5$-parameters family of vexillary permutations, the specia\-li\-za\-tion $x_i=1$, $\forall\, i \ge 1$, of the
corresponding {\it $\beta$-Schubert} polynomials
${{\mathfrak{S}}}_{w}^{(\beta)}(X_n)$ turns out to be coincide either with the
Fuss--Narayana polynomials and their generalizations, or with a
$(q,\beta)$-deformation of $\rm VSASM$ or that of $\rm CSTCPP$ numbers, see
Corollary~\ref{corollary5.2}{\rm B}.
As examples we show that
\begin{enumerate}\itemsep=0pt
\item[(a)] the reduced polynomial corresponding to
a monomial $x_{12}^{n} x_{23}^{m}$ counts the number of {\it $(n,m)$-Delannoy
paths} according to the number of $NE$-steps, see Lemma~\ref{lemma5.3};
\item[(b)]
if $\beta=0$, the reduced polynomial corresponding to monomial
$(x_{12} x_{23})^n x_{34}^{k}$, $n \ge k$, counts the number of~$n$
up, $n$ down permutations in the symmetric group ${\mathbb{S}}_{2n+k+1}$, see
Proposition~\ref{proposition5.10}; see also Conjecture~\ref{conjecture5.10}.
\end{enumerate}
We also point out on a conjectural connection between the sets of
{\it maximal} compatible sequences for the permutation $\sigma_{n,2n,2,0}$
and that $\sigma_{n,2n+1,2,0}$ from one side, and the set of ${\rm VSASM}(n)$ and
that of ${\rm CSTCPP}(n)$ correspondingly, from the other, see Comments~\ref{comments5.7} for
details. Finally, in Sections~\ref{section5.1.1} and~\ref{section5.4.1} we introduce and study a~multiparameter generalization of reduced polynomials considered
in~\cite[Exercise~6.C5(c)]{ST3}, as well as that of the Catalan, Narayana and
(small) Schr\"oder numbers.
In the {\it case} of the {\it longest element} we relate the
corresponding reduced polynomial with the Ehrhart polynomial of the
Chan--Robbins--Yuen polytope, see Section~\ref{section5.3}. More generally, we relate the
$(t,\beta)$-reduced polynomial corresponding to monomial
\begin{gather*}
\prod_{j=1}^{n-1} x_{j,j+1}^{a_{j}} \prod_{j=2}^{n-2}
\left(\prod_{k=j+2}^{n} x_{jk} \right), \qquad a_j \in \mathbb{Z}_{\ge 0}, \qquad \forall\, j,
\end{gather*}
with positive $t$-deformations of the Kostant partition function and that of
the Ehrhart polynomial of some f\/low polytopes, see Section~\ref{section5.3}.
In Section~\ref{section5.4} we investigate reduced polynomials associated with certain
monomials in the algebra $({\widehat{{\rm ACYB}}})^{ab}_n(\beta)$, known also as
Gelfand--Varchenko algebra~\cite{K3,K}, and study its combinatorial properties. Our main objective in Section~\ref{section5.4.2} is to study reduced polynomials for Coxeter element treated in a certain multiparameter deformation of the
(noncommutative) quadratic algebra ${\widehat{{\rm ACYB}}}_n(\alpha,\beta)$.
Namely, to each dissection of a convex $(n+2)$-gon we associate a certain
weight and consider the generating function of all dissections of~$(n+2)$-gon selected taken with that weight. One can show that the reduced polynomial
corresponding to the Coxeter element in the deformed algebra is equal to that
generating function. We show that certain specializations of that reduced
polynomial coincide, among others, with the Grothendieck polynomials corresponding to the permutation $1 \times w_{0}^{(n-1)} \in \mathbb{S}_n$, the Lagrange inversion formula, as well as give rise to combinatorial (i.e., positive
expressions) multiparameters deformations of Catalan and Fuss--Catalan, Motzkin, Riordan and Fine numbers, Schr\"{o}der numbers and Schr\"{o}der trees. We
{\it expect} (work in progress) a~similar connections between Schubert and Grothendieck polynomials associated with the Richardson permutations $1^{k} \times
w_{0}^{(n-k)}$, $k$-dissections of a convex $(n+k+1)$-gon investigated in the
present paper, and $k$-dimensional Lagrange--Good inversion formula studied
from combinatorial point of view, e.g., in~\cite{CL, Ge}.
\section{Introduction}\label{section1}
The Dunkl operators have been introduced in the later part of 80's of the
last century by Charles Dunkl \cite{Du,Du1} as a powerful mean to
study of harmonic and orthogonal polynomials related with f\/inite Coxeter
groups. In the present paper we don't need the def\/inition of Dunkl
operators for arbitrary (f\/inite) Coxeter groups, see, e.g.,~\cite{Du}, but only for the special case of the symmetric group ${\mathbb S}_n$.
\begin{Definition}\label{definition1.1} Let $P_n= \mathbb{C}[x_1,\ldots,x_n]$ be the ring of polynomials in
variables $x_1,\ldots,x_n$. The type~$A_{n-1}$ (additive) rational Dunkl
operators $D_1, \ldots, D_n$ are the dif\/ferential-dif\/ference operators of
the following form
\begin{gather}\label{equation1.1}
D_i= \lambda {\partial \over \partial x_i} + \sum_{j \not= i}
{1-s_{ij} \over x_i-x_j},
\end{gather}
Here $s_{ij}$, $1 \le i < j \le n$, denotes the exchange (or permutation)
operator, namely,
\begin{gather*}
s_{ij}(f)(x_1,\ldots,x_i,\ldots,x_j,\ldots, x_n)=f(x_1,\ldots,x_j,\ldots,
x_i,\ldots, x_n),
\end{gather*}
${\partial \over \partial x_i}$ stands for the derivative w.r.t.\ the variable
$x_i$, $\lambda \in \mathbb{C}$ is a parameter.
\end{Definition}
The key property of the Dunkl operators is the following result.
\begin{Theorem}[C.~Dunkl~\cite{Du}]\label{theorem1.1} For any finite Coxeter group $(W,S)$,
where $S=\{s_1,\ldots,s_l\}$ denotes the set of simple reflections, the
Dunkl operators $D_i:=D_{s_{i}}$ and $D_j:=D_{s_{j}}$ pairwise
commute: $D_i D_j=
D_jD_i$, $1 \le i ,j \le l$.
\end{Theorem}
Another fundamental property of the Dunkl operators which f\/inds a wide variety
of applications in the theory of integrable systems, see, e.g.,~\cite{HW}, is
the following statement:
the operator
\begin{gather*}
\sum_{i=1}^{l} (D_i)^2
\end{gather*}
``essentially'' coincides with the Hamiltonian of the rational Calogero--Moser
model related to the f\/inite Coxeter group~$(W,S)$.
\begin{Definition}\label{definition1.2}
Truncated (additive) Dunkl operator (or the Dunkl operator at
critical level), denoted by ${\cal D}_i$, $i=1,\ldots,l$,~is an operator of
the form~\eqref{equation1.1} with parameter $\lambda=0$.
\end{Definition}
For example, the type $A_{n-1}$ rational truncated Dunkl operator has the
following form
\begin{gather*}
{\cal D}_i = \sum_{j \not= i} {1-s_{ij} \over x_i-x_j }.
\end{gather*}
Clearly the truncated Dunkl operators generate a commutative algebra.
The important property of the truncated Dunkl operators is the following
result discovered and proved by C.~Dunkl~\cite{Du1}; see also~\cite{Ba} for a~more recent proof.
\begin{Theorem}[C.~Dunkl~\cite{Du1}, Yu.~Bazlov~\cite{Ba}] \label{theorem1.2} For any finite Coxeter
group $(W,S)$ the algebra over $\mathbb{Q}$ generated by the truncated Dunkl
operators ${\cal D}_1,\ldots,{\cal D}_l$ is canonically isomorphic to the
coinvariant algebra ${\cal{A}}_W$ of the Coxeter group~$(W,S)$.
\end{Theorem}
Recall that for a f\/inite {\it crystallographic} Coxeter group~$(W,S)$ the
coinvariant algebra~${\cal{A}}_W$ is isomorphic to the cohomology ring
$H^{*}(G/B, \mathbb{Q})$ of the f\/lag variety~$G/B$, where $G$ stands for the Lie group
corresponding to the crystallographic Coxeter group~$(W,S)$ we started with.
\begin{Example} \label{example1.1}
In the case when $W= {\mathbb S}_n$ is the symmetric group,
Theorem~\ref{theorem1.2} states that the algebra over~$\mathbb{Q}$ generated by the truncated
Dunkl operators ${\cal D}_i=\sum\limits_{j \not= i} {1-s_{ij} \over x_i-x_j}$, $i=1,
\ldots,n$, is canonically isomorphic to the cohomology ring of the full f\/lag
variety ${\cal F}l_n$ of type~$A_{n-1}$
\begin{gather}\label{equation1.2}
\mathbb{Q}[{\cal D}_1,\ldots,{\cal D}_n] \cong \mathbb{Q}[x_1,\ldots,x_n] / J_n,
\end{gather}
where $J_n$ denotes the ideal generated by the elementary symmetric
polynomials $ \{e_k(X_n), \, 1 \le k \le n\}$.
Recall that the elementary symmetric polynomials~$e_i(X_n)$,
$i=1,\ldots,n$, are def\/ined through the generating function
\begin{gather*}
1+\sum_{i=1}^{n} e_i(X_n) t^{i}=\prod_{i=1}^{n} (1+t x_i),
\end{gather*}
where we set $X_n:=(x_1,\ldots,x_n)$. It is well-known that in the case
$W=\mathbb{S}_n$, the isomorphism~\eqref{equation1.2} can be def\/ined over the ring of
integers $\mathbb{Z}$.
\end{Example}
Theorem~\ref{theorem1.2} by C.~Dunkl has raised a number of natural questions:
\begin{enumerate}\itemsep=0pt
\item[(A)] What is the algebra generated by the {\it truncated}
\begin{itemize}\itemsep=0pt
\item trigonometric,
\item elliptic,
\item super, matrix, \dots,
\begin{enumerate}\itemsep=0pt
\item[(a)] additive Dunkl operators?
\item[(b)] Ruijsenaars--Schneider--Macdonald operators?
\item[(c)] Gaudin operators?
\end{enumerate}
\end{itemize}
\item[(B)] Describe commutative subalgebra generated by the Jucys--Murphy
elements in
\begin{itemize}\itemsep=0pt
\item the group ring of the symmetric group;
\item the Hecke algebra;
\item the Brauer algebra, ${\rm BMW}$ algebra, \dots.
\end{itemize}
\item[(C)] Does there exist an analogue of Theorem~\ref{theorem1.2} for
\begin{itemize} \itemsep=0pt
\item classical and quantum equivariant cohomology and equivariant
$K$-theory rings of the partial f\/lag varieties?
\item chomology and $K$-theory rings of af\/f\/ine f\/lag varieties?
\item diagonal coinvariant algebras of f\/inite Coxeter groups?
\item complex ref\/lection groups?
\end{itemize}
\end{enumerate}
The present paper is an extended introduction to a few items from Section~5 of~\cite{K}.
The main purpose of my paper ``On some quadratic algebras,~II'' is to give
some partial answers on the above questions basically in the case of the
symmetric group~${\mathbb S}_n$.
The purpose of the {\it present paper} is to draw attention to an
interesting class of nonhomogeneous quadratic algebras closely connected
(still mysteriously!) with dif\/ferent branches of Mathematics such as
classical and quantum Schubert and Grothendieck calculi,
low-dimensional topology,
classical, basic and elliptic hypergeometric functions,
algebraic combinatorics and graph theory,
integrable systems, etc.
What we try to explain in~\cite{K} is that upon passing to {\it a suitable
representation} of the quadratic algebra in question, the subjects mentioned
above, are a manifestation of certain general properties of that quadratic
algebra.
From this point of view, we treat the commutative subalgebra generated (over a~universal {\it Lazard} ring~${\mathbb{L}}_n$~\cite{Lazard1955})
by
the additive (resp.\ multiplicative) truncated Dunkl elements in the algebra
$3T_n(\beta)$, see Def\/inition~\ref{definition3.1}, as {\it universal cohomology}
(resp.\ {\it universal $K$-theory})~ring of the complete f\/lag variety
${\cal F}l_n$. The classical or quantum
cohomology (resp.\ the classical or quantum $K$-theory) rings of the f\/lag
variety~${\cal F}l_n$ are certain quotients of that {\it universal ring.}
For example, in~\cite{KM2} we have computed relations among the (truncated)
Dunkl elements $\{ \theta_i, \, i=1,\ldots,n \}$ in the {\it elliptic
representation} of the algebra $3T_n(\beta=0)$. We {\it expect} that the
commutative subalgebra obtained is isomorphic to {\it elliptic cohomology
ring} (not def\/ined yet, but see~\cite{GRa, GKV}) of the f\/lag variety~${\cal F}l_n$.
Another example from \cite{K}. Consider the algebra $3T_n(\beta=0)$.
One can prove~\cite{K} the following {\it identities} in the algebra
$3T_n(\beta=0)$:
\begin{enumerate}\itemsep=0pt
\item[(A)] {\it summation formula}
\begin{gather*}
\sum_{j=1}^{n-1}
\left( \prod_{b=j+1}^{n-1} u_{b,b+1} \right) u_{1,n}
\left(\prod_{b=1}^{j-1} u_{b,b+1} \right)=\prod_{a=1}^{n-1} u_{a,a+1};
\end{gather*}
\item[(B)] {\it duality transformation formula}, let $m \le n$, then
\begin{gather*}
\sum_{j=m}^{n-1} \left( \prod_{b=j+1}^{n-1} u_{b,b+1} \right)
\left[\prod_{a=1}^{m-1} u_{a,a+n-1} u_{a,a+n} \right] u_{m,m+n-1}
\left( \prod_{b=m}^{j-1} u_{b,b+1} \right) \\
\qquad\quad{} +
\sum_{j=2}^{m}\left[\prod_{a=j}^{m-1}u_{a,a+n-1} u_{a,a+n} \right] u_{m,n+m-1} \left(\prod_{b=m}^{n-1} u_{b,b+1} \right) u_{1,n}\\
\qquad {} = \sum_{j=1}^{m} \left[ \prod_{a=1}^{m-j} u_{a,a+n} u_{a+1,a+n} \right] \left(
\prod_{b=m}^{n-1}u_{b,b+1} \right) \left[ \prod_{a=1}^{j-1}
u_{a,a+n-1} u_{a,a+n} \right].
\end{gather*}
\end{enumerate}
One can check that upon passing to the {\it elliptic representation} of the
algebra $3T_n(\beta=0)$, see Section~\ref{section3.1} or~\cite{KM2}, for the def\/inition of {\it elliptic representation}, the above
identities~(A) and~(B) f\/inally end up correspondingly, to be
the {\it summation formula} and the $N=1$ case of the {\it duality
transformation formula} for multiple elliptic hypergeometric series (of type
$A_{n-1})$, see, e.g.,~\cite{NK} or Appendix~\ref{appendixA.6} for the explicit forms of the latter. After passing to the so-called {\it Fay representation}~\cite{K}, the
identities~(A) and~(B) become correspondingly to be the
summation formula and duality transformation formula for the Riemann theta
functions of genus $g > 0$~\cite{K}. These formulas in the case~$g \ge 2$
seems to be new.
Worthy to mention that the relation (A) above can be treated as
a~``non-commutative analogue'' of the well-known recurrence relation among
the {\it Catalan numbers}. The study of ``descendent relations'' in the
quadratic algebras in question was originally motivated by the author
attempts to construct a~{\it monomial basis} in the algebra $3T_n^{(0)}$,
and compute ${\rm Hilb}(3T_n^{(0)},t)$ for $n \ge 6$.
These problems are still widely open, but gives rise the author to discovery
of several interesting connections with
\begin{itemize}\itemsep=0pt
\item classical and quantum Schubert and Grothendieck calculi,
\item combinatorics of reduced decomposition of some special elements in
the symmetric group,
\item combinatorics of generalized {\it Chan--Robbins--Yuen} polytopes,
\item relations among the Dunkl and Gaudin elements,
\item computation of Tutte and chromatic polynomials of the weighted complete
multipartite graphs, etc.
\end{itemize}
A few words about the content of the {\it present paper}.
Example~\ref{example1.1} can be viewed as an illustration of the main problems we are
treated in Sections~\ref{section2} and~\ref{section3} of the present paper, namely the following ones.
\begin{itemize}\itemsep=0pt
\item Let $\{u_{ij},\, 1 \le i,j \le n\}$ be a set of generators of a certain algebra over a commutative ring~$K$. The f\/irst {\it problem} we are
interested in is to describe
``a natural set of relations'' among the generators
$\{u_{ij}\}_{1 \le i,j \le n}$ which
implies the pair-wise commutativity of {\it dynamical Dunkl elements}
\begin{gather*}
\theta_{i}= \theta_{i}^{(n)}=:\sum\limits_{j=1}^{n} u_{ij}, \qquad 1 \le i \le n.
\end{gather*}
\item Should this be the case then we are interested in to describe the
algebra generated by ``the integrals of motions'', i.e., to describe the
quotient of the algebra of polynomials $K[y_1,\ldots,y_n]$ by the two-sided ideal~${\cal{J}}_n$ generated by non-zero polynomials
$F(y_1,\ldots,y_n)$ such that $F(\theta_1,\ldots,\theta_n)=0$ in the algebra
over ring~$K$ generated by the elements \linebreak $\{u_{ij} \}_{1 \le i,j \le n}$.
\item We are looking for a set of additional relations which imply that
the elementary symmetric polynomials $e_k(Y_n)$, $1 \le k \le n$, belong to the
set of integrals of motions. In other words, the value of elementary symmetric
polynomials $e_k(y_1,\ldots, y_n)$, $1 \le k \le n$, on the Dunkl elements
$\theta_1^{(n)},\ldots,\theta_n^{(n)}$ {\it do not depend} on the
variables $\{u_{ij}, \, 1 \le i \not= j \le n\}$. If so,
one can def\/ined {\it deformation} of elementary symmetric polynomials, and
make use of it and the Jacobi--Trudi formula, to def\/ine deformed Schur
functions, for example. We try to realize this program in Sections~\ref{section2} and~\ref{section3}.
\end{itemize}
In Section~\ref{section2}, see Def\/inition~\ref{definition2.2}, we introduce the
so-called {\it dynamical classical Yang--Baxter algebra} as ``a natural
quadratic algebra'' in which the Dunkl elements form a pair-wise commuting
family. It is the study of the algebra generated by the (truncated) Dunkl
elements that is the main objective of our investigation in \cite{K} and the
present paper. In Section~\ref{section2.1} we describe few representations of
the dynamical classical Yang--Baxter algebra~${\rm DCYB}_n$ related with
\begin{itemize}\itemsep=0pt
\item quantum cohomology $QH^{*}({\cal{F}}l_n)$ of the
complete f\/lag variety ${\cal{F}}l_n$, cf.~\cite{FGP};
\item quantum equivariant
cohomology $QH^{*}_{T^n \times C^{*}}(T^{*}{\cal{F}}l_n)$ of the cotangent
bundle $T^{*} {\cal{F}}l_n$ to the complete f\/lag variety, cf.~\cite{GRTV};
\item Dunkl--Gaudin and Dunkl--Uglov representations, cf.~\cite{MTV,U}.
\end{itemize}
In Section~\ref{section3}, see Def\/inition~\ref{section3.1}, we introduce the algebra
$3HT_n (\beta)$, which seems to be the most
general (noncommutative) deformation of the (even) Orlik--Solomon algebra of
type~$A_{n-1}$,
such that it's still possible to describe relations among the Dunkl elements,
see Theorem~\ref{theorem3.1}. As an application we describe explicitly a set of relations
among the (additive) Gaudin/Dunkl elements, cf.~\cite{MTV}.
It should be stressed at this
place that we treat the Gaudin elements/operators (either additive or
multiplicative) as {\it images} of the
{\it universal} Dunkl elements/operators (additive or multiplicative)
in the {\it Gaudin representation} of the algebra~$3HT_n(0)$. There are
se\-ve\-ral other important representations of that algebra, for example, the
Calogero--Moser, Bruhat, Buchstaber--Felder--Veselov (elliptic), Fay trisecant
($\tau$-functions), adjoint, and so on, considered (among others) in~\cite{K}.
Specif\/ic properties of a representation chosen\footnote{For example, in the cases of either {\it Calogero--Moser} or
{\it Bruhat} representations one has an additional constraint, namely,
$u_{ij}^2=0$ for all $i \not= j$. In the case of {\it Gaudin} representation
one has an additional constraint $u_{ij}^2 = p_{ij}^2$, where the (quantum)
parameters $ \{p_{ij}={1 \over x_i-x_j}, \, i \not=j \}$, satisfy
{\it simultaneously} the {\it Arnold} and {\it Pl\"{u}cker} relations, see
Section~\ref{section2},~{\bf II}. Therefore, the (small) quantum cohomology ring of the
type $A_{n-1}$ full f\/lag variety ${{\cal F}l}_{n}$ and the Bethe subalgebra(s)
(i.e., the subalgebra generated by Gaudin elements in the algebra
$3HT_n(0)$) correspond to {\it different specializations} of ``{\it quantum
parameters}'' $\{q_{ij} :=u_{ij}^2 \}$ of the {\it universal cohomology
ring} (i.e., the subalgebra/ring in $3HT_n(0)$ generated by (universal) Dunkl
elements). For more details and examples, see Section~\ref{section2.1} and~\cite{K}.}
(e.g., {\it Gaudin representation}) imply some
additional relations among the images of the universal Dunkl elements (e.g., {\it Gaudin elements}) should to be unveiled.
We start Section \ref{section3} with def\/inition of algebra~$3T_n(\beta)$ and
its ``Hecke'' $3HT_n(\beta)$ and ``elliptic'' $3MT_n(\beta)$ quotients.
In particular we def\/ine an elliptic representation of the algebra~$3T_n(0)$~\cite{KM2}, and show how the well-known elliptic solutions of the quantum
Yang--Baxter equation due to A.~Belavin and V.~Drinfeld, see, e.g.,~\cite{BD}, S.~Shibukawa and K.~Ueno~\cite{SU+}, and G.~Felder and V.~Pasquier~\cite{Fe},
can be plug in to our construction, see Section~\ref{section3.1}. At the end of
Section~\ref{section3.1.1} we point out on a {\it mysterious} (at least for the author)
appearance of the Euler numbers and ``traces'' of the Brauer algebra in the
equivariant Pieri rules hold for the algebra $3TM_n(\beta, {\boldsymbol{q}}, \psi)$ stated in
Theorem~\ref{theorem3.1}.
In Section \ref{section3.2} we introduce a {\it multiplicative} analogue of
the Dunkl elements $\{ \Theta_j \in 3T_n(\beta)$, $1 \le j \le n\}$ and
describe the commutative subalgebra in the algebra $3T_n(\beta)$ generated
by multiplicative Dunkl elements~\cite{KM}. The latter commutative subalgebra
turns out to be isomorphic to the quantum equivariant $K$-theory of the
complete f\/lag variety ${\cal{F}}l_n$~\cite{KM}.
In Section~\ref{section3.3} we describe relations among the truncated
Dunkl--Gaudin elements. In this case the quantum parameters $q_{ij}=
p_{ij}^2$, where parameters $\{ p_{ij}= (z_i-z_j)^{-1} ,\, 1 \le i < j \le n \}$ satisfy the both Arnold and Pl\"{u}cker relations. This observation
has made it possible to describe a set of additional {\it rational
relations} among the Dunkl--Gaudin elements, cf.~\cite{MTV}.
In Section~\ref{section3.4} we introduce an equivariant version of
multiplicative Dunkl elements, called {\it shifted Dunkl elements} in our
paper, and describe (some) relations among the latter. This result is a
generalization of that obtained in Section~\ref{section3.1} and~\cite{KM}. However we
don't know any geometric interpretation of the commutative subalgebra
generated by shifted Dunkl elements.
In Section \ref{section4.1} for any subgraph $\Gamma \subset K_n$ of the
complete graph~$K_n$ we introduce\footnote{Independently the algebra $3T_n^{(0)}(\Gamma)$ has been studied in~\cite{BLM}, where the reader can f\/ind some examples and conjectures.}
\cite{K2,K}, algebras $3T_n(\Gamma)$ and $3T_n^{(0)}(\Gamma)$ which
can be seen as analogues of algebras~$3T_n$ and~$3T_n^{(0)}$
correspondingly\footnote{To avoid confusions, it must be emphasized that the def\/ining
relations for algebras $3T_n(\Gamma)$~and $3T_n(\Gamma)^{(0)}$ may have more
then three terms.}.
We want to point out in the Introduction, cf.\ footnote~\ref{footnote1}, that an analog of the algebras~$3T_n$ and~$3T_n^{(\beta)}$, $3HT_n$, etc.\ treated in the present paper, can be def\/ined for any (oriented or not)
{\it matroid}~$\cal{M}$. We denote these algebras as $3T({\cal{M}})$ and
$3T^{(\beta)}({\cal{M}})$. One can show~(A.K.) that the {\it abelianization}
of the algebra $3T^{(\beta)}({\cal{M}})$, denoted by
${3T^{(\beta)}({\cal{M}})}^{ab}$, is isomorphic to the {\it Gelfand--Varchenko}
algebra corresponding to a matroid~$\cal{M}$, whereas the algebra ${3T^{(\beta=
0)}({\cal{M}})}^{ab}$ is isomorphic to the (even) {\it Orlik--Solomon} algebra
${\rm OS}^{+}({\cal{M}})$ of a matroid $\cal{M}$.\footnote{For a def\/inition and basic properties of the Orlik--Solomon algebra corresponding to a matroid, see, e.g., \cite{GR,Kawahara2004}.} We consider and treat the
algebras $3T({\cal{M}})$, $3HT({\cal{M}})$, \dots, as
{\it equivariant noncommutative}~(or {\it quantum}) versions of the (even)
Orlik--Solomon algebras associated with matroid (including hyperplane,
graphic, \dots\ arrangements). However a meaning of a quantum deformation of the
(even or odd) Orlik--Solomon algebra suggested in the present paper, is missing, even for the braid arrangement of type~$A_{n}$. Generalizations of the
Gelfand--Varchenko algebra has been suggested and studied in~\cite{K3,K} and in the present paper under the name quasi-associative Yang--Baxter
algebra, see Section~\ref{section5}.
In the present paper we basically study the {\it abelian quotient} of the
algebra $3T_n^{(0)}(\Gamma)$, where graph $\Gamma$ has no loops and multiple
edges, since we expect some applications of our approach to the
theory of {\it chromatic polynomials} of planar graphs, in particular to
the complete multipartite graphs $K_{n_1,\ldots,n_{r}}$ and the grid graphs~$G_{m,n}$.\footnote{See \url{http://reference.wolfram.com/language/ref/GridGraph.html} for a def\/inition of {\it grid
graph} $G_{m,n}$.} Our main results hold for
the complete multipartite, cyclic and line graphs. In particular we compute
their {\it chromatic} and {\it Tutte} polynomials, see Proposition~\ref{proposition4.2} and
Theorem~\ref{theorem4.3}. As a~byproduct we compute the Tutte polynomial of the
${\boldsymbol{\ell}}$-weighted complete multipartite graph
$K_{n_1,\ldots,n_r}^{({\boldsymbol{\ell}})}$ where ${\boldsymbol{\ell}}=\{\ell_{ij}\}_{1 \le i < j \le r}$, is a collection of weights, i.e., a set of non-negative integers.
More generally, for a set of variables $ \{ \{q_{ij} \}_{1 \le i < j \le n},
x,y \}$ we def\/ine {\it universal Tutte polynomial}
$T_n(\{q_{ij}\},x,y) \in \mathbb{Z}[q_{ij}] [x,y]$ such that for any collection of non-negative integers $\{m_{ij} \}_{1 \le i < j \le n}$ and a subgraph $\Gamma
\subset K_n^{({\boldsymbol{m}})}$ of the complete graph~$K_n$ with each edge~$(i,j)$
comes with multiplicity~$m_{ij}$, the specialization
\begin{gather*}
q_{ij} \longrightarrow 0 \quad \text{if~edge} \ \ (i,j) \notin \Gamma, \qquad q_{ij}
\longrightarrow [m_{ij}]_{y}:= \frac{y^{m_{ij}}-1}{y-1} \quad \text{if~edge} \ \ (i,j) \in
\Gamma
\end{gather*}
of the universal Tutte polynomial $T_n(\{q_{ij} \},x,y)$ is equal to the Tutte
polynomial of graph~$\Gamma$ multiplied by the factor $(t-1)^{\kappa(\Gamma)}$:
\begin{gather*}
(x-1)^{\kappa(\Gamma)} {\rm Tutte} (\Gamma,x,y):= T_n(\{q_{ij}\}, x,y) \bigg |_{q_{ij}=0 \, \text{if} \, (i,j) \notin \Gamma \atop
q_{ij}={[m_{ij}]}_{y} \, \text{if} \, (i,j) \in \Gamma}.
\end{gather*}
Here and after $\kappa(\Gamma)$ demotes the number of connected components
of a~graph~$\Gamma$. In other words, one can treat the universal Tutte
polynomial $T_n(\{q_{ij} \},x,y)$ as a ``reproducing kernel'' for the Tutte
polynomials of all (loop-less) graphs with the number of vertices not exceeded~$n$.
We also state Conjecture~\ref{conjecture4.2} that for any loopless graph~$\Gamma$ (possibly
with multiple edges) the algebra ${3T_{|\Gamma|}^{(0)}(\Gamma)}^{ab}$ is
isomorphic to the even Orlik--Solomon algebra ${\rm OS}^{+}({\cal{A}}_{\Gamma})$
of the {\it graphic arrangement} associated with graph~$\Gamma$ in question\footnote{For simple graphs, i.e., without loops and multiple edges, this conjecture has been proved in~\cite{Li}.}.
At the end we emphasize that the case of the complete graph $\Gamma=K_n$
reproduces the results of the present paper and those of~\cite{K}, i.e.,
the case of the full f\/lag variety ${\cal F}l_n$. The case of the
{\it complete multipartite graph} $\Gamma=K_{n_{1},\ldots, n_{r}}$ reproduces
the analogue of results stated in the present paper for the case of full
f\/lag variety ${\cal F}l_n$, to the case of the {\it partial flag}
variety~${\cal F}_{n_{1},\ldots,n_{r}}$, see~\cite{K} for details.
In Section~\ref{section4.1.3} we sketch how to generalize our constructions
and some of our results to the case of the Lie algebras of {\it classical
types}\footnote{One can def\/ine an analogue of the algebra~$3T_n^{(0)}$ for the root
system of~$BC_n$ and $C_n^{\vee}C_n$-types as well, but we are omitted these
cases in the present paper.}.
In Section~\ref{section4.2} we brief\/ly overview our results concerning yet
another interesting family of quadratic algebras, namely the {\it six-term
relations algebras}~$6T_n$, $6T_n^{(0)}$ and related ones. These algebras
also contain a distinguished set of mutually commuting elements called
{\it Dunkl elements} $\{\theta_i,\,i=1,\ldots, n \}$ given by
$\theta_i= \sum\limits_{j \not= i} r_{ij}$, see Def\/inition~\ref{definition4.10}.
In Section~\ref{section4.2.2} we introduce and study the algebra
$6T_n^{\bigstar}$ in greater detail. In particular we introduce a ``quantum
deformation'' of the algebra generated by the curvature of $2$-forms of of
the Hermitian linear bundles over the f\/lag variety~${\cal{F}}l_n$,
cf.~\cite{PSS}.
In Section~\ref{section4.2.3} we state our results concerning the {\it
classical Yang--Baxter algebra} ${\rm CYB}_n$ and the $6$-term relation algebra~$6T_n$. In particular we give formulas for the Hilbert series of these
algebras. These formulas have been obtained independently in~\cite{BEE} The
paper just mentioned, contains a description of a basis in the algebra~$6T_n$, and much more.
In Section~\ref{section4.2.4} we introduce a {\it super analog} of the
algebra~$6T_n$, denoted by~$6T_{n,m}$, and compute its Hilbert series.
Finally, in Section~\ref{section4.3} we introduce {\it extended nil-three
term relations} algebra~${3 \mathfrak{T}}_n$ and describe a subalgebra inside
of it which is isomorphic to the double af\/f\/ine Hecke algebra of type~$A_{n-1}$,
cf.~\cite{Ch}.
In Section~\ref{section5} we describe several combinatorial properties of some
special elements in the associative quasi-classical Yang--Baxter algebra\footnote{The algebra $\widehat{{\rm ACYB}}_n$ can be treated as ``one-half''
of the algebra~$3T_n(\beta)$. It appears that the basic relations
among the Dunkl elements, which do {\it not} mutually commute anymore, are
still {\it valid}, see Lemma~\ref{lemma5.1}.}, denoted by $\widehat{{\rm ACYB}}_n$. The main results in that direction were
motivated and obtained as a by-product, in the process of the study of the
{\it the structure} of the algebra~$3HT_n(\beta)$. More specif\/ically, the
main results of Section~\ref{section5} were obtained in the course of
``hunting for descendant relations'' in the algebra mentioned, which is an
important problem to be solved to construct {\it a basis} in the
nil-quotient algebra $3T_n^{(0)}$. This {\it problem} is still widely-open.
The results of Section~\ref{section5.1}, see Proposition~\ref{proposition5.1}, items~(1)--(5), are more or less well-known among the specialists
in the subject, while those of the item~(6) seem to be new. Namely,
we show that the polynomial $Q_n(x_{ij}=t_i)$
from~\cite[Exercise~6.C8(c)]{ST3}, essentially coincides with the
$\beta$-deformation \cite{FK1} of the Lascoux--Sch\"utzenberger
Grothendieck polynomial \cite{LS} for some particular permutation.
The results of Proposition~\ref{proposition5.1}(6), point out on a deep connection
between reduced forms of monomials in the algebra~$\widehat{{\rm ACYB}}_n$ and the
Schubert and Grothendieck calculi. This observation was the starting point
for the study of some combinatorial properties of certain specializations of
the Schubert, the $\beta$-Grothendieck~\cite{FK2} and the double $\beta$-Grothendieck polynomials in Section~\ref{section5.2}. One of the main
results of Section~\ref{section5.2} can be stated as follows.
\begin{Theorem}\label{theorem1.3} \quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] Let $w \in \mathbb{S}_n$ be a permutation, consider the
specialization $x_1:=q$, $x_i=1$, $\forall\, i \ge 2$, of the $\beta$-Grothendieck
polynomial $\mathfrak{G}_{w}^{(\beta)}(X_n)$. Then
\begin{gather*} {\cal{R}}_{w}(q,\beta +1) := \mathfrak{G}_{w}^{(\beta)}(x_1=q,\, x_i=1,
\, \forall\, i \ge 2) \in \mathbb{N} [q,1+\beta].
\end{gather*}
In other words, the polynomial
${\cal{R}}_{w}(q,\beta)$ has non-negative integer coefficients\footnote{For a more general result see Appendix~\ref{appendixA.1}, Corollary~\ref{corollary6.2}.}.
For late use we define polynomials
\begin{gather*}
\mathfrak{R}_{w}(q,\beta) := q^{1-w(1)} {\cal {R}}_{w}(q,\beta).
\end{gather*}
\item[$(2)$] Let $w \in \mathbb{S}_n$ be a permutation, consider the
specialization $x_i:=q$, $y_i=t$, $\forall\, i \ge 1$, of the double
$\beta$-Grothendieck
polynomial $\mathfrak{G}_{w}^{(\beta)}(X_n,Y_n)$. Then
\begin{gather*}
\mathfrak{G}_{w}^{(\beta -1)}(x_i:=q,\, y_i :=t, \, \forall \, i \ge 1) \in \mathbb{N} [q,t,
\beta].
\end{gather*}
\item[$(3)$] Let $w$ be a permutation, then
\begin{gather*}
\mathfrak{R}_{w}(1,\beta)=\mathfrak{R}_{1 \times w}(0,\beta).
\end{gather*}
Note that ${\cal{R}}_{w}(1,\beta) ={\cal {R}}_{w^{-1}}(1,\beta)$,
but ${\cal{R}}_{w}(t,\beta) \not={\cal {R}}_{w^{-1}}(t,\beta)$, in general.
\end{enumerate}
\end{Theorem}
For the reader convenience we collect some basic def\/initions and results
concerning the $\beta$-Grothendieck polynomials in Appendix~\ref{appendixA.1}.
Let us observe that $\mathfrak{R}_{w}(1,1)= \mathfrak{S}_{w}(1)$, where
$\mathfrak{S}_{w}(1)$ denotes the specialization $x_i:=1$, $\forall \, i \ge 1$,
of the Schubert polynomial $\mathfrak{S}_{w}(X_n)$ corresponding to
permutation~$w$. Therefore,
$\mathfrak{R}_{w}(1,1)$ is equal to the number of {\it compatible sequences}~\cite{BJS} (or {\it pipe dreams}, see, e.g.,~\cite{SS}) corresponding to
permutation~$w$.
\begin{Problem} \label{problem1.1}
Let $w \in \mathbb{S}_n$ be a permutation and $l:= \ell(w)$ be its length.
Denote by ${\rm CS}(w) = \{ {\boldsymbol{a}}=(a_1 \le a_2 \le \cdots \le a_l) \in \mathbb{N}^{l}\}$
the set of
compatible sequences~{\rm \cite{BJS}} corresponding to permutation~$w$.
\begin{itemize} \itemsep=0pt
\item Define statistics $r({\boldsymbol{a}})$ on the set of all
compatible sequences ${\rm CS}_n := \coprod\limits_{{w \in \mathbb{S}_n}} {\rm CS}(w)$
in a such way that
\begin{gather*}
\sum _{{\boldsymbol{a}} \in {\rm CS}(w)} q^{a_1} \beta ^{r({\boldsymbol{a}})} =
{\cal {R}}_{w}(q,\beta).
\end{gather*}
\item Find a geometric interpretation, and
investigate combinatorial and algebra-geometric pro\-per\-ties of
polynomials $\mathfrak{S}_{w}^{(\beta)}(X_n)$,
where for a permutation $w \in \mathbb{S}_n$ we denoted by
$\mathfrak{S}_{w}^{(\beta)}(X_n)$ the $\beta$-Schubert polynomial
defined as follows
\begin{gather*}
\mathfrak{S}_{w}^{(\beta)}(X_n) = \sum_{{\boldsymbol{a}} \in {\rm CS}(w)} \beta^{r({\boldsymbol{a}})} \prod_{i=1}^{l:=\ell(w)} x_{a_{i}}.
\end{gather*}
\end{itemize}
\end{Problem}
We {\it expect} that polynomial $\mathfrak{S}_{w}^{(\beta)}(1)$ coincides
with the Hilbert polynomial of a certain graded commutative ring naturally
associated to permutation~$w$.
\begin{Remark}
It should be mentioned that, in general, the principal specialization
\begin{gather*}
\mathfrak{G}_{w}^{(\beta-1)}\big(x_i:=q^{i-1}, \,\forall \, i \ge 1\big)
\end{gather*}
of the $(\beta-1)$-Grothendieck polynomial may have {\it negative}
coef\/f\/icients.
\end{Remark}
Our main objective in Section~\ref{section5.2} is to study the polynomials
$\mathfrak{R}_{w}(q,\beta)$ for a special class of permutations in the
symmetric group $\mathbb{S}_{\infty}$.
Namely, in Section~\ref{section5.2} we study some combinatorial properties of
polynomials $\mathfrak{R}_{ \varpi_{\lambda,\phi}}(q,\beta)$ for the f\/ive
parameters family of {\it vexillary}
permutations $ \{\varpi_{\lambda,\phi} \}$ which have the {\it shape}
$\lambda:= \lambda_{n,p,b}=(p(n-i+1)+b$, $i=1,\ldots,n+1)$
and {\it flag}
$\phi:= \phi_{k,r} = (k+r (i-1),~i=1, \ldots,n+1)$.
This class of permutations is notable for many reasons, including that the
specialized value of the Schubert polynomial $\mathfrak{S}_{\varpi_{\lambda,
\phi}}(1)$ admits a nice product formula\footnote{One can prove a product formula for the principal
specialization $\mathfrak{S}_{\varpi_{\lambda,\phi}}(x_i := q^{i-1},\,
\forall\, i \ge 1)$ of the correspon\-ding Schubert polynomial. We don't need a~such formula in the present paper.}, see Theorem~\ref{theorem5.6}. Moreover, we describe also some interesting connections of
polynomials $\mathfrak{R}_{\varpi_{\lambda,\phi}}(q,\beta)$ with plane
partitions, the Fuss--Catalan numbers\footnote{We def\/ine the (generalized)
Fuss--Catalan numbers to be ${\rm FC}_{n}^{(p)}(b):={1+b \over 1+b+(n-1)p} {n p + b \choose n}$. Connection of
the Fuss--Catalan numbers with the {\it $p$-ballot numbers} ${\rm Bal}_{p}(m,n):=
{n-mp+1 \over n+m+1}~{n+m+1 \choose m}$ and the {\it Rothe numbers}
$R_{n}(a,b):= {a \over a+b n} {a+b n \choose n}$ can be described as follows
\begin{gather*}
{\rm FC}_{n}^{(p)}(b)=R_{n}(b+1,p)={\rm Bal}_{p-1}(n,(n-1) p+b).
\end{gather*}}
and Fuss--Narayana polynomials,
$k$-triangulations and $k$-dissections of a~convex polygon, as well as a connection with two families of ${\rm ASM}$. For
example, let $\lambda=(b^{n})$ and $\phi=(k^n)$ be rectangular shape
partitions, then the polynomial
$\mathfrak{R}_{ \varpi_{\lambda,\phi}}(q,\beta)$ def\/ines a~$(q,\beta)$-deformation of the number of (ordinary) plane partitions\footnote{Let $\lambda$ be a partition. An ordinary plane partition (plane
partition for short)bounded by~$d$ and shape $\lambda$ is a~f\/illing of the
shape $\lambda$ by the numbers from the set $\{0,1,\ldots,d \}$ in such a
way that the numbers along columns and rows are weakly {\it decreasing}.
A {\it reverse} plane partition bounded by
$d$ and shape $\lambda$ is a~f\/illing of the shape $\lambda$ by the numbers from the set $\{0,1,\ldots,d \}$ in such a way that the numbers along columns
and rows are weakly {\it increasing}.}
sitting in the box~$b \times k \times n$. It seems an interesting
{\it problem} to f\/ind an algebra-geometric interpretation of polynomials
$\mathfrak{R}_{w}(q,\beta)$ in the general case.
\begin{Question}\label{question1.1}
Let $a$ and $b$ be mutually prime positive integers. Does
there exist a family of permutations $w_{a,b} \in {\mathbb{S}}_{ab(a+b)}$ such
that the specialization $x_i =1$, $\forall\, i$ of the Schubert polyno\-mial~${\mathfrak{S}}_{w_{a,b}}$ is equal to the rational Catalan number $C_{a/b}$? That is
\begin{gather*}
{\mathfrak{S}}_{w_{a,b}}(1)={1 \over a+b} {a+b \choose a}.
\end{gather*}
\end{Question}
Many of the computations in Section~\ref{section5.2} are based on the following
determinantal formula for $\beta$-Grothendieck polynomials corresponding to
grassmannian permutations, cf.~\cite{L}.
\begin{Theorem}[see Comments~\ref{comments5.5}(b)]\label{theorem1.4}
If $w=\sigma_{\lambda}$ is the grassmannian permutation with
shape $\lambda =(\lambda_,\ldots, \lambda_n)$ and a unique
descent at position~$n$, then\footnote{The equality
\begin{gather*}
\mathfrak{G}_{\sigma_{\lambda}}^{(\beta)}(X_n)=
{\operatorname{DET}\big| x_i^{\lambda_j +n-j}
(1+\beta x_i)^{j-1} \big|_{1 \le i,j \le n} \over \prod\limits_{1 \le i < j \le n}
(x_i-x_j) },
\end{gather*}
has been proved independently in \cite{MS}.}
\begin{gather*}
({\rm A}) \quad \mathfrak{G}_{\sigma_{\lambda}}^{(\beta)}(X_n)=
\operatorname{DET}\big|h_{\lambda_j +i,j}^{(\beta)}(X_n)\big|_{1 \le i,j \le n} =
{\operatorname{DET} \big| x_i^{\lambda_j +n-j}
(1+\beta x_i)^{j-1} \big|_{1 \le i,j \le n} \over \prod\limits_{1 \le i < j \le n}
(x_i-x_j) },
\end{gather*}
where $X_{n}=(x_i,x_{1},\ldots,x_n)$, and for any set of variables~$X$,
\begin{gather*}
h_{n,k}^{(\beta)}(X) = \sum_{a=0}^{k-1}~{k-1 \choose a} h_{n-k+a}(X)
\beta^{a},
\end{gather*}
and $h_k(X)$ denotes the complete symmetric polynomial of degree~$k$ in the variables from the set~$X$.
\begin{gather*}
({\rm B}) \quad {\mathfrak G}_{\sigma_{\lambda}}(X,Y) =
{ \operatorname{DET}\Big| \prod\limits_{a=1}^{\lambda_{j}+n-j} (x_i +y_a + \beta x_i y_a) (1+\beta x_i)^{j-1} \Big|_{1 \le i,j \le n} \over \prod\limits_{1 \le i < j \le n} (x_i-x_j) }.
\end{gather*}
\end{Theorem}
In Sections~\ref{section5.2.2} and~\ref{section5.4.2} we study connections of Grothendieck polynomial
associated with the Richardson permutation~$ w_{k}^{(n)}= 1^{k} \times w_{0}^{(n-k)}$, $k$-dissections of a convex $(n+k+1)$-gon, generalized reduced
polynomial corresponding to a certain monomial in the algebra~${\widehat{{\rm ACYB}}}_{n}$ and the Lagrange inversion formula. In the case of generalized Richardson
permutation~$w_{n,p}^{(k)}$ corresponding to the
$k$-shifted dominant permutations~$w^{(p,n)}$ associated with the Young diagram
$ \lambda_{p,n}:= p(n-1,n-2,\ldots,1)$, namely, $w_{n,p}^{(k)} = 1^{k} \times
w^{(p,n)}$, we treat only the case $k=1$, see also~\cite{EM}. In the case
$k \ge 2$ one comes to a {\it task} to count and f\/ind a lattice path type
interpretation for the number of {\it $k$-{\bf p}gulations} of a convex
$n$-gon that is the number of partitioning of a convex $n$-gon on parts which
are all equal to a convex $(p+2)$-gon, by a (maximal) family of diagonals such
that each diagonal has at most $k$ internal intersections with the members of a family of diagonals selected.
In Section~\ref{section5.3} we give a partial answer on Question~{6.C8}(d) by R.~Stanley~\cite{ST3}. In particular, we relate the reduced
polynomial corresponding to monomial
\begin{gather*}
\bigl( x_{12}^{a_{2}} \cdots {x_{n-1,n}}^{a_{n}} \bigr) \prod_{j=2}^{n-2}
\prod_{k=j+2}^{n} x_{jk},\qquad a_j \in \mathbb{Z}_{\ge 0}, \qquad \forall\, j,
\end{gather*}
with the Ehrhart polynomial of the generalized Chan--Robbins--Yuen polytope,
if $a_2=\cdots=a_n=m+1$, cf.~\cite{Me2}, with a $t$-deformation of the
Kostant partition function of type $A_{n-1}$ and the Ehrhart polynomials of
some f\/low polytopes, cf.~\cite{MM}.
In Section~\ref{section5.4} we investigate certain specializations of the
reduced polynomials corresponding to monomials of the form
\begin{gather*}
x_{12}^{m_1} \cdots x_{n-1,n}^{m_n}, \qquad m_j \in \mathbb{Z}_{\ge 0}, \qquad \forall\, j.
\end{gather*}
First of all we observe that the corresponding specialized reduced polynomial
appears to be a {\it piece-wise polynomial function} of parameters
${\boldsymbol{m}}=(m_1,\ldots, m_n) \in (\mathbb{R}_{\ge 0})^{n}$, denoted by $P_{{\boldsymbol{m}}}$.
It is an interesting {\it problem} to compute the {\it Laplace transform} of
that piece-wise polynomial function. In the present paper we compute the value
of the function $P_{{\boldsymbol{m}}}$ in the dominant chamber
${\cal{C}}_n = (m_1 \ge m_2 \ge \cdots \ge m_n \ge 0)$, and give a~combinatorial interpretation of the values of that function in points
$(n,m)$ and $(n,m,k)$, $n \ge m \ge k$.
For the reader convenience, in Appendices~\ref{appendixA.1}--\ref{appendixA.6} we collect some
useful auxiliary information about the subjects we are treated in the present
paper.
Almost all results in Section~\ref{section5} state that some two specif\/ic sets have the
same number of elements. Our proofs of these results are pure algebraic. It is
an interesting {\it problem} to f\/ind {\it bijective proofs} of results from
Section~\ref{section5} which generalize and extend remarkable bijective proofs presented
in \cite{MM,SS,St, W} to the {\it cases} of
\begin{itemize}\itemsep=0pt
\item the $\beta$-Grothendieck polynomials,
\item the (small) Schr\"oder numbers,
\item $k$-dissections of a convex $(n+k+1)$-gon,
\item special values of reduced polynomials.
\end{itemize}
We are planning to treat and present these bijections in separate
publication(s).
We {\it expect} that the reduced polynomials corresponding to the higher-order
powers of the Coxeter elements also admit an interesting combinatorial
interpretation(s). Some preliminary results in this direction are discussed in
Comments~\ref{comments5.8}.
At the end of introduction I want to add a few remarks.
(a)~After a suitable modif\/ication of the algebra $3HT_n$, see~\cite{KM3}, and the case $\beta \not= 0$ in~\cite{K}, one can compute the set
of relations among the (additive) Dunkl elements (def\/ined in Section~\ref{section2},
equation~\eqref{equation2.1}). In the case $\beta=0$ and $q_{ij}=q_i \delta_{j-i,1}$,
$1 \le i < j \le n$, where $\delta_{a,b}$ is the Kronecker delta symbol, the
commutative algebra generated by additive Dunkl elements~\eqref{equation2.3} appears to be
``almost'' isomorphic to the equivariant quantum
cohomology ring of the f\/lag variety ${\cal F}l_n$, see~\cite{KM3} for
details. Using the {\it multiplicative} version of Dunkl elements, see
Section~\ref{section3.2}, one can extend the results from~\cite{KM3} to the case of equivariant quantum
$K$-theory of the f\/lag variety~${\cal F}l_n$, see~\cite{K}.
(b)~As it was pointed out previously, one can def\/ine an analogue of the algebra
$3T_n^{(0)}$ for any (oriented) matroid ${\cal{M}}_n$, and state a conjecture
which connects the Hilbert polynomial of the algebra $3T_n^{(0)}(({\cal{M}}_n)^{ab},t)$ and the chromatic polynomial of matroid ${\cal{M}}_n$. We {\it
expect} that algebra $3T_n^{(\beta=1)}({\cal{M}}_n)^{ab}$ is isomorphic to
the {\it Gelfand--Varchenko} algebra associated with matroid~$\cal{M}$. It
is an interesting {\it problem} to f\/ind a combinatorial meaning of the algebra
$3T_n^{(\beta)}({\cal{M}}_n)$ for $\beta =0$ and $\beta \not= 0$.
(c)~Let $R$ be a (graded) ring (to be specif\/ied later) and
${\mathfrak{F}}_{n^{2}}$ be the free associative algebra over~$R$ with the set
of generators $\{ u_{ij},\, 1 \le i,j \le n \}$. In the subsequent text we
will distinguish the set of generators $\{ u_{ii} \}_{1 \le i \le n}$ from
that $\{u_{ij}\}_{1 \le i \not= j \le n}$, and set
\begin{gather*}
x_i:=u_{ii}, \qquad i=1,\ldots, n.
\end{gather*}
A guiding idea to choose def\/initions and perform constructions in the present
paper is to impose a set of relations~${\cal{R}}_n$ among the generators
$\{x_{i} \}_{1 \le i \le n}$ and that $\{u_{ij} \}_{1 \le i \not= j \le n}$
which ensure the mutual commutativity of the following elements
\begin{gather*}
\theta_{i}^{(n)}:=\theta_i = x_i + \sum_{j \not= i}^{n} u_{ij}, \qquad i=1,\ldots,n,
\end{gather*}
in the algebra ${\cal{F}}_{n^{2}} / {\cal{R}}_{n}$, as well as to have a good chance to describe/compute
$\bullet$~``Integral of motions'', that is f\/inding a big enough set of
algebraically independent polynomials (quite possibly that polynomials are
trigonometric or elliptic ones) $I_{\alpha}^{(n)}(y_1,\ldots,y_n) \in R [Y_n]$ such that
\begin{gather*}
I_{\alpha}^{(n)}\big(\theta_1^{(n)}, \ldots, \theta_n^{(n)}\big) \in R [X_n],
\qquad \forall\, \alpha,
\end{gather*}
in other words, the latter specialization of any integral of motion has to be
independent of the all generators $\{ u_{ij} \}_{1 \le i \not=j \le n}$.
$\bullet$~Give a presentation of the algebra ${\cal{I}}_{n}$ generated by the
integral of motions that is to f\/ind a set of def\/ining relations among the
elements $\theta_1,\ldots, \theta_n$, and describe a~$R$-basis
$ \big\{m_{\alpha}^{(n)}\big\}$ in the algebra~${\cal{I}}_{n}$.
$\bullet$~Generalized Littlewood--Richardson and Murnaghan--Nakayama problems.
Given an integral of motion $I_{\beta}^{(m)}(Y_m)$ and an integer $n \ge m$,
f\/ind an explicit positive (if possible) expression in the quotient algebra
${\cal{F}}_{n^{2}} / {\cal{R}}_{n}$ of the element
\begin{gather*}
I_{\beta}^{(m)}\big(\theta_{1}^{(n)}, \ldots , \theta_{m}^{(n)}\big).
\end{gather*}
For example in the case of the 3-term relations algebra $3T_n^{(0)}$ (as
well as its equivariant, quantum, etc.\ versions) the generalized
Littlewood--Richardson problem is to f\/ind a positive expression in the algebra~$3T_n^{(0)}$ for the element ${\mathfrak{S}}_{w}\big(\theta_1^{(n)},\ldots,\theta_{m}^{(n)}\big)$, where ${\mathfrak{S}}_{w}(Y_n)$ stands for the Schubert polynomial corresponding to a~permutation $w \in \mathbb{S}_n$.
Generalized Murnaghan--Nakayama problem consists in f\/inding a combinatorial
expression in the algebra $3T_n^{(0)}$ for the element $ \sum\limits_{i=1}^{m}
(\theta_{i}^{(n)})^{k}$.
Partial results concerning these problems have been obtained as far as we
aware in \cite{FK,K2, K, KMa,MPP, P}.
$\bullet$~``Partition functions''. Assume that the (graded) algebra ${\cal{I}}_{n}$ generated over $R$ by the
elements $\theta_1,\ldots, \theta_n$ has f\/inite dimension/rank, and the (non
zero) maximal degree component ${\cal{I}}_{\max}^{(n)}$ of that algebra has
dimension/rank one and generated by an element $\omega$.~For any element
$ g \in {\cal{F}}_{n^{2}}$ let us denote by $\operatorname{Res}_{\omega}(g)$ an element in~$R$ such that
\begin{gather*}
\overline{g} = \operatorname{Res}_{\omega}(g) \omega,
\end{gather*}
where we denote by $\overline{g}$ the image of element $g$ in the component
${\cal{I}}_{\max}^{(n)}$.
We def\/ine {\it partition function} associated with the algebra ${\cal{I}}_{n}$ as follows
\begin{gather*}
{\cal{Z}}({\cal{I}}_{n}) = \operatorname{Res}_{w} \bigg( \exp \bigg(\sum_{\alpha} q_{\alpha} m_{\alpha}^{(n)} \bigg) \bigg),
\end{gather*}
where $ \{q_{\alpha}\}$ is a set of parameters which is consistent in
one-to-one correspondence with a~basis $\big\{m_{\alpha}^{(n)}\big\}$ chosen.
We are interesting in to f\/ind a closed formula for the partition function
${\cal{Z}}({\cal{I}}_{n})$ as well as that for a {\it small} partition function
\begin{gather*}
{\cal{Z}}^{(0)}({\cal{I}}_{n}):= \operatorname{Res}_{\omega} \bigg(\exp \bigg( \sum_{1 \le i, j \le n} \lambda_{ij} u_{ij} \bigg) \bigg),
\end{gather*}
where $\{\lambda_{ij} \}_{1 \le i,j \le n}$ stands for a set of parameters.
One can show~\cite{K4} that the partition function ${\cal{Z}}({\cal{I}}_{n})$
associated with algebra $3T_n^{{\boldsymbol{q}}}$ satisf\/ies the famous Witten--Dijkraaf--Verlinde--Verlinde equations.
As a preliminary steps to perform our guiding idea we
\begin{enumerate}\itemsep=0pt
\item[(i)] investigate properties of the abelianization of the algebra
${\cal{F}}_{n^{2}} / {\cal{R}}_n$. Some unexpected connections with
the theory of hyperplane arrangements and graph theory are discovered;
\item[(ii)] investigate a variety of descendent relations coming from the
def\/ining relations. Some polynomials with interesting combinatorial properties
are naturally appear.
\end{enumerate}
To keep the size of the present paper reasonable, several new results are
presented as exercises.
We conclude Introduction by a short historical remark. As far as we aware, the
commutative version of $3$-term relations which provided the framework for a
def\/inition of the FK algebra~${\cal{E}}_n$~\cite{FK} and a plethora of its
generalizations, have been frequently used implicitly in the theory of elliptic
functions and related topics, starting at least from the middle of the 19th century, see, e.g.,~\cite{WW} for references, and up to now, and for sure will be used for ever.
The key point is that the Kronecker sigma function
\begin{gather*}
\sigma_{z}(w):= {\frac{ \sigma(z-w) \theta' (0)}{\sigma(z) \sigma(-w)}},
\end{gather*}
where $\sigma(z)$ denotes the Weierstrass sigma function, satisf\/ies the
quadratic three terms {\it addition formula} or {\it functional equation}
discovered, as far as we aware, by
K.~Weierstrass. In fact this functional equation~is really equivalent\footnote{We refer the reader to a nice written paper by Tom H.~Koornwinder \cite{Koornwinder2014} for more historical information.}
to the famous Jacobi--Riemann three term relation of degree four between
the Riemann theta functions~$\theta(x)$. In the rational degeneration of theta
functions, the three term relation between Kronecker sigma functions turns to the
famous three term Jacobi identity which can be treated as an associative
analogue of the Jacobi identity in the theory of Lie algebras.
To our best knowledge, in an abstract form that is as a set of def\/ining
relations in a certain algebra, an anticommutative version of three term
relations had been appeared in a~remarkable paper by V.I.~Arnold~\cite{Arnold1969}. Nowadays these relations are known as {\it Arnold relations}.
These relations and its various generalizations play fundamental role in the
theory of arrangements, see, e.g.,~\cite{OT}, in topology, combinatorics and
many other branches of Mathematics.
In commutative set up abstract form of $3$-term relations has been invented by
O.~Mathieu~\cite{Mathieu1995}. In the context of the braided Hopf algebras (of type~A)
$3$-term relations like algebras (as some examples of the Nichols algebras)
have appeared in papers by A.~Milinski and H.-J.~Schneider (2000), N.~Andruskiewitsch (2002), S.~Madjid (2004), I.~Heckenberger (2005) and many others\footnote{We refer the reader to the site
\url{https://en.wikipedia.org/wiki/Nichols_algebra}
for basic def\/initions and results concerning Nichols' algebras and references on vast literature treated dif\/ferent aspects of the theory of Nichols' algebras and braided Hopf algebras.}.
It is well-known that the Nichols algebra associated with the symmetric group~${\mathbb{S}}_n$ and trivial conjugacy class is a quotient of the algebra~$FK_n$. It is still an open {\it problem} to prove (or disprove) that these two algebras are isomorphic.
\section{Dunkl elements}\label{section2}
Having in mind to fulf\/ill conditions suggested by our guiding line mentioned
in Introduction as far as it could be done till now, we are led to introduce the following algebras\footnote{Surprisingly enough, in many cases to f\/ind relations among the
elements $\theta_1, \ldots, \theta_n$ there is no need to require that the
elements $\{\theta_i \}_{1 \le i \le n}$ are pairwise commute.}.
\begin{Definition}[additive Dunkl elements] \label{definition2.1}
The (additive) Dunkl elements $\theta_i$, $i=1,\dots,n$, in the
algebra ${\cal F}_n$ are def\/ined to be
\begin{gather}\label{equation2.1}
\theta_i=x_i+\sum_{j=1 \atop j \not=i}^{n} u_{ij}.
\end{gather}
\end{Definition}
We are interested in to f\/ind ``natural relations'' among the generators
$\{u_{ij} \}_{1 \le i,j \le n}$ such that the Dunkl elements~\eqref{equation2.1} are
pair-wise {\it commute}. One of the natural conditions which is the
commonly accepted in the theory of integrable systems, is
\begin{itemize}\itemsep=0pt
\item locality conditions:
\begin{gather}
(a) \quad [x_i,x_j]=0 \qquad \text{if} \ \ i \not= j,\nonumber\\
(b) \quad u_{ij} u_{kl}=u_{kl} u_{ij} \qquad \text{if} \ \ i \not=j, \ \ k \not= l \ \ \text{and} \ \ \{i,j \} \cap \{k,l \}=\varnothing.\label{equation2.2}
\end{gather}
\end{itemize}
\begin{Lemma}\label{lemma2.1} Assume that elements $\{u_{ij} \}$ satisfy the locality condition~\eqref{equation2.1}. If $i \not= j$,
then
\begin{gather*}
[\theta_i,\theta_j]= \biggl[ x_i+ \sum_{k \not= i,j} u_{ik}, u_{ij}+u_{ji}
\biggr] +\biggl[u_{ij},\sum_{k=1}^{n} x_k \biggr]+ \sum_{k \not= i,j} w_{ijk},
\end{gather*}
where
\begin{gather}
w_{ijk}=[u_{ij},u_{ik}+u_{jk}]+[u_{ik},u_{jk}]+[x_i,u_{jk}]+[u_{ik},x_j]+
[x_k,u_{ij}].\label{equation2.3}
\end{gather}
\end{Lemma}
Therefore in order to ensure that the Dunkl elements form a pair-wise
{\it commuting} family, it's natural to assume that the following
conditions hold
\begin{itemize}
\item unitarity:
\begin{gather}
[u_{ij}+u_{ji},u_{kl}]=0=[u_{ij}+u_{ji},x_k] \qquad \text{for all distinct} \ \ i,\,j,\,k,\, l,\label{equation2.4}
\end{gather}
i.e., the elements $u_{ij}+u_{ji}$ are {\it central}.
\item ``conservation laws'':
\begin{gather}
\left[\sum_{k=1}^{n} x_k,u_{ij}\right] =0 \qquad \text{for all} \ \ i,\, j,
\label{equation2.5}
\end{gather}
i.e., the element $E:=\sum\limits_{k=1}^{n} x_k$ is {\it central},
\item unitary dynamical classical Yang--Baxter relations:
\begin{gather}\label{equation2.6}
[u_{ij},u_{ik}+u_{jk}]+[u_{ik},u_{jk}]+[x_i,u_{jk}]+[u_{ik},x_j]+
[x_k,u_{ij}]=0,
\end{gather}
if $i$, $j$, $k$ are pair-wise distinct.
\end{itemize}
\begin{Definition}[dynamical six term relations algebra $6DT_n$] \label{definition2.2}
We denote by $6DT_n$ (and frequently will use also notation ${\rm DCYB}_n$) the
quotient of the algebra ${\cal F}_n$ by the two-sided ideal generated by relations~\eqref{equation2.2}--\eqref{equation2.6}.
\end{Definition}
Clearly, the Dunkl elements~\eqref{equation2.1} generate a commutative subalgebra inside of
the algebra~$6DT_n$, and the sum $\sum\limits_{i=1}^{n} \theta_i =\sum\limits_{i=1}^{n} x_i$ belongs to the center of the algebra~$6DT_n$.
\begin{Remark}\label{remark2.0}
Occasionally we will call the Dunkl elements of the form \eqref{equation2.1} by {\it
dynamical Dunkl elements} to distinguish the latter from {\it truncated
Dunkl elements}, corresponding to the case $x_i=0$, $\forall\, i$.
\end{Remark}
\subsection[Some representations of the algebra $6DT_n$]{Some representations of the algebra $\boldsymbol{6DT_n}$} \label{section2.1}
\subsubsection{Dynamical Dunkl elements and equivariant quantum cohomology}\label{section2.1.1}
{\bf (I)} ( cf.~\cite{FGP}). Given a set $q_1,\ldots,q_{n-1} $ of mutually
commuting parameters, def\/ine
\begin{gather*}
q_{ij}=\prod_{a=i}^{j-1} q_a \qquad \text{if} \quad i < j,
\end{gather*}
and set $q_{ij}=q_{ji}$ in the
case $i > j$. Clearly, that if $i < j < k$, then $q_{ij}q_{jk}=q_{ik}$.
Let $z_1, \ldots,z_n$ be a set of (mutually commuting) variables. Denote by
$P_n:=\mathbb{Z}[z_1,\ldots,z_n]$ the corresponding ring of polynomials. We consider
the variable~$z_i$, $i=1,\ldots,n$, also as the operator acting on the ring
of polynomials $P_n$ by multiplication on the variable~$z_i $.
Let $s_{ij} \in \mathbb{S}_n$ be the transposition that swaps the letters~$i$
and $j$ and f\/ixes the all other letters $k \not= i,j$. We consider the
transposition~$s_{ij}$ also as the operator which acts on the ring~$P_n$ by
interchanging $z_i$ and $z_j$, and f\/ixes all other variables. We denote by
\begin{gather*}
\partial_{ij}={1-s_{ij} \over z_i-z_j },\qquad \partial_{i} :=
\partial_{i,i+1},
\end{gather*}
the divided dif\/ference operators corresponding to the transposition
$s_{ij}$ and the simple transposition $s_{i}:= s_{i,i+1}$ correspondingly.
Finally we
def\/ine operator (cf.~\cite{FGP})
\begin{gather*}
\partial_{(ij)}:= \partial_{i}\cdots \partial_{j-1}\partial_{j}\partial_{j-1} \cdots \partial_{i} \qquad \text{if} \ \ i < j.
\end{gather*}
The operators $\partial_{(ij)}$, $1 \le i < j \le n$, satisfy (among other
things) the following set of relations (cf.~\cite{FGP})
\begin{itemize}\itemsep=0pt
\item $[z_j,\partial_{(ik)}]=0$ if $j \notin [i,k]$, $\Big[\partial_{(ij)},\sum\limits_{a=i}^{j} z_a\Big]=0$,
\item $[\partial_{(ij)},\partial_{(kl)}] =
\delta_{jk} [z_j,\partial_{(il)}]+\delta_{il} [\partial_{(kj)},z_i]$ if $i < j$, $k < l$.
\end{itemize}
Therefore, if we set $u_{ij}=q_{ij} \partial_{(ij)}$ if $i < j$, and
$u_{ij}=-u_{ji}$ if $i > j$, then for a triple $i < j < k$ we will have
\begin{gather*}
[u_{ij},u_{ik}+u_{jk}]+[u_{ik},u_{jk}]+[z_i,u_{jk}]+[u_{ik},z_j]+[z_k,u_{jk}]\\
\qquad{} =q_{ij}q_{jk}[\partial_{(ij)},\partial_{(jk)}]+q_{ik}[\partial_{(ik)},z_j]=0.
\end{gather*}
Thus the elements $\{ z_i,\, i=1,\ldots,n \}$ and
$\{u_{ij},\, 1 \le i < j \le n \}$ def\/ine a representation of the algebra
${\rm DCYB}_n$, and therefore the Dunkl elements
\begin{gather*}
\theta_i:= z_i+ \sum_{j \not= i} u_{ij}=z_i-\sum_{j < i}q_{ji}\partial_{(ji)}+\sum_{j > i} q_{ij}\partial_{(ij)}
\end{gather*}
form a pairwise commuting family of operators acting on the ring of
polynomials
\begin{gather*}
\mathbb{Z}[q_1,\ldots,q_{n-1}][z_1,\ldots,z_n],
\end{gather*}
cf.~\cite{FGP}. This representation
has been used in~\cite{FGP} to construct the small quantum cohomology ring
of the complete f\/lag variety of type~$A_{n-1}$.
{\bf (II)} Consider degenerate af\/f\/ine Hecke algebra ${\mathfrak {H}}_n$
generated by the central element~$h$, the elements of the symmetric group
${\mathbb{S}}_n$, and the mutually commuting elements $y_1,\ldots,y_n$,
subject to relations
\begin{gather*}
s_i y_i - y_{i+1} s_i = h, \quad 1 \le i < n, \qquad s_i y_j =y_j s_i, \quad j \not= i,i+1,
\end{gather*}
where $s_i$ stand for the simple transposition that swaps only indices~$i$
and~$i+1$. For $i < j$, let $s_{ij}=s_i\cdots s_{j-1}s_{j}s_{j-1}\cdots
s_{i}$ denotes the permutation that swaps only indices~$i$ and~$j$. It is
an easy exercise to show that
\begin{itemize}\itemsep=0pt
\item $[y_j,s_{ik}] = h [s_{ij},s_{jk}]$ if $i < j < k$,
\item $y_{i} s_{ik} -s_{ik} y_{k} = h + h s_{ik} \sum\limits_{ i < j < k} s_{jk}$ if $ i < k$.
\end{itemize}
Finally, consider a set of mutually commuting parameters $\{p_{ij}, \, 1 \le i
\not= j \le n, \, p_{ij}+p_{ji}=0 \}$, subject to the constraints
\begin{gather*}
p_{ij} p_{jk} = p_{ik} p_{ij} + p_{jk} p_{ik} + h p_{ik}, \qquad i < j < k.
\end{gather*}
\begin{Comments} \label{comments2.1} If parameters $\{p_{ij} \}$ are {\it invertible}, and
satisfy relations
\begin{gather*}
p_{ij} p_{jk} = p_{ik} p_{ij} + p_{jk} p_{ik} + \beta p_{ik}, \qquad i < j < k,
\end{gather*}
then one can rewrite the above displayed relations in the following form
\begin{gather*}
1+ {\beta \over p_{ik}} =\left(1+{\beta \over p_{ij}} \right)
\left(1 +{\beta \over p_{jk}}\right), \qquad 1 \le i < j < k \le n .
\end{gather*}
Therefore there exist parameters $\{q_1,\ldots,q_n \}$ such that
$1+\beta /p_{ij}=q_i/q_j$, $1 \le i < j \le n$. In other words, $p_{ij} =
{ \beta q_j \over q_j -q_j}$, $1 \le i < j \le n$. However in general,
there are many other types of solutions, for example, solutions related to
the Heaviside function\footnote{See \url{https://en.wikipedia.org/wiki/Heaviside_step_function}.}
$H(x)$, namely, $p_{ij} = H(x_i-x_j)$, $x_i \in \mathbb{R}$, $\forall\, i$, and its
discrete analogue, see Example~{\bf (III)} below. In the both cases $\beta= -1$; see also Comments~\ref{comments2.3} for other examples.
\end{Comments}
To continue presentation of Example~{\bf (II)}, def\/ine elements
$u_{ij}= p_{ij} s_{ij}$, $1 \le i \not= j \le n$.
\begin{Lemma}[dynamical classical Yang--Baxter relations]\label{lemma2.2}
\begin{gather}\label{equation2.7}
[u_{ij}, u_{ik}+u_{jk}]+[u_{ik},u_{jk}]+[u_{ik},y_j] = 0, \qquad 1 < i < j < k \le n.
\end{gather}
\end{Lemma}
Indeed,
\begin{gather*}
u_{ij}u_{jk}=u_{ik}u_{ij}+u_{jk}u_{ik}+h p_{ik}s_{ij}s_{jk}, \qquad
u_{jk}u_{ij}=u_{ij}u_{ik}+u_{ik}u_{jk}+h p_{ik}s_{jk}s_{ij},
\end{gather*}
and moreover, $[y_j,u_{ik}] = h p_{ik} [s_{ij},s_{jk}]$.
Therefore, the elements
\begin{gather*}
\theta_i = y_i - h \sum_{j < i} u_{ij} + h \sum_{i < j} u_{ij}, \qquad i=1,\ldots,n,
\end{gather*}
form a mutually commuting set of elements in the algebra
$\mathbb{Z}[\{p_{ij} \}] \otimes_{\mathbb{Z}} {\mathfrak{H}}_n$.
\begin{Theorem} \label{theorem2.1} Define matrix $M_n =(m_{i,j})_{1 \le i,j \le n}$ as follows
\begin{gather*}
m_{i,j}(u;z_1,\ldots,z_n) = \begin{cases}
u-z_i& \text{if $i=j$},\\
-h-p_{ij} & \text{if $i < j$}, \\
p_{ij} & \text{if $i > j$}.
\end{cases}
\end{gather*}
Then
\begin{gather*}
\operatorname{DET} \big | M_n(u;\theta_1,\ldots,\theta_n) \big |
= \prod_{j=1}^{n} (u-y_j).
\end{gather*}
Moreover, let us set $q_{ij}:=h^{2}( p_{ij}+p_{ij}^2) =
h^2 q_{i}q_{j}(q_{i}-q_{j})^{-2}$, $i < j$, then
\begin{gather*}
e_k(\theta_1,\ldots,\theta_n) = e_k^{({\boldsymbol{q}})}(y_1,\ldots,y_n), \qquad
1 \le k \le n,
\end{gather*}
where $e_k(x_1,\ldots,x_n)$ and $e_k^{({\boldsymbol{q}})}(x_1,\ldots,x_n)$ denote
correspondingly the classical and multiparameter quantum~{\rm \cite{FK}}
elementary polynomials\footnote{For the reader convenience we remind~\cite{FK} a~def\/inition of
the quantum elementary polynomial $e_{k}^{\boldsymbol{q}}(x_1,\ldots, x_n)$. Let
${\boldsymbol{q}}:= \{q_{ij} \}_{1 \le i <j \le n}$ be a collection of ``quantum
parameters'', then
\begin{gather*}
e_{k}^{\boldsymbol{q}}(x_1,\ldots,x_n)=\sum_{\ell} \sum_{1 \le i_{1} < \cdots < i_{\ell} \le n \atop j_{1} > i_{1}, \ldots ,j_{\ell} > i_{\ell}} e_{k-2\ell}(X_{
\overline{I \cup J}}) \prod_{a=1}^{\ell} q_{i_{a},j_{a}},
\end{gather*}
where $I=(i_1, \ldots, i_{\ell})$, $J=(j_1,\ldots,j_{\ell})$ should be
distinct elements of the set $\{1, \ldots, n \}$, and
$X_{\overline{I \cup J}}$ denotes set of variables~$x_a$ for which the
subscript~$a$ is neither one of~$i_m$ nor one of the~$j_m$.}.
\end{Theorem}
Let's stress that the elements $y_i$ and $\theta_j$ {\it do not} commute
in the algebra ${\mathfrak{H}}_n$, but the symmetric functions of
$y_1,\ldots,y_n$, i.e., the center of the algebra ${\mathfrak{H}}_n$, do.
A few remarks in order. First of all, $u_{ij}^2=p_{ij}^{2}$ are central
elements. Secondly, in the case $ h=0 $ and $y_i=0$, $\forall\, i$, the equality
\begin{gather*}
\operatorname{DET} \big | M_n(u;x_1,\ldots,x_n) \big | = u^{n}
\end{gather*}
describes the set of {\it polynomial} relations among the Dunkl--Gaudin
elements (with the following choice of parameters $p_{ij}=(q_i-q_j)^{-1}$ are
taken). And our f\/inal remark is that according to~\cite[Section~8]{GRTV}, the
quotient ring
\begin{gather*}
{\cal {H}}_n^{{\boldsymbol{q}}} : = \mathbb{Q}[y_1,\ldots,y_n]^{{\mathbb {S}}_n} \otimes
\mathbb{Q}[\theta_1, \dots, \theta_n] \otimes \mathbb{Q}[h]\Big/ \bigg\langle M_n(u;\theta_1,
\ldots,\theta_n)=\prod_{j=1}^{n}(u-y_j) \bigg\rangle
\end{gather*}
is isomorphic to the quantum equivariant cohomology ring of the cotangent
bundle $T^{*}{\cal {F}}l_n$ of the complete f\/lag variety of type $A_{n-1}$,
namely,
\begin{gather*}
{\cal {H}}_n^{{\boldsymbol{q}}} \cong QH^{*}_{T^{n} \times \mathbb{C}^{*}} (T^{*} {\cal{F}}l_n)
\end{gather*}
with the following choice of quantum parameters: $Q_i:=h q_{i+1}/ q_{i}$,
$i=1,\ldots,n-1$.
On the other hand, in~\cite{KM3} we computed the so-called {\it multiparameter
deformation} of the equivariant cohomology ring of the complete f\/lag variety
of type~$A_{n-1}$.
\begin{center}
\framebox{\parbox[t]{6in}{A deformation def\/ined in~\cite{KM3} depends on parameters
$\{q_{ij},\,1 \le i < j \le n \}$ without any constraints are imposed. For the
special choice of parameters
\begin{gather*}
q_{ij}:= h^2 {q_i~q_j \over (q_i-q_j)^2}
\end{gather*}
the multiparameter deformation of the equivariant cohomology ring of the type
$A_{n-1}$ complete f\/lag variety~${\cal{F}}l_n$ constructed in~\cite{KM3}, is
isomorphic to the ring~${\cal {H}}_n^{{\boldsymbol{q}}}$. }}
\end{center}
\begin{Comments} \label{comments2.2} Let us f\/ix a set of independent parameters $\{q_1,\ldots,
q_n \}$ and def\/ine new parameters
\begin{gather*}
\left\{ q_{ij}:= h p_{ij}(p_{ij}+h)= h^2
{q_i q_j \over (q_i-q_j)^2 } \right\}, \quad 1 \le i < j \le n,\! \qquad \text{where} \quad p_{ij}= {q_{j}
\over q_{i}-q_{j}}, \quad i < j.
\end{gather*}
We set $\deg(q_{ij})= 2$, $\deg(p_{ij})= 1$, $\deg(h)=1$.
The new parameters $\{q_{ij} \}_{1 \le i < j \le n}$,
do not free anymore, but satisfy rather complicated algebraic relations. We
display some of these relations soon, having in mind a question:
is there some
intrinsic meaning of the algebraic variety def\/ined by the set of def\/ining
relations among the ``quantum parameters''~$\{q_{ij} \}$?
Let us denote by
${{\cal {A}}}_{n,h}$ the quotient ring of the ring of polynomials
$\mathbb{Q}[h][x_{ij},\, 1 \le i < j \le n]$ modulo the ideal generating by
polynomials $f(x_{ij})$ such that the specialization $x_{ij}=q_{ij}$ of a~polynomial $f(x_{ij})$, namely $f(q_{ij})$, is equal to zero. The algebra
${\cal{A}}_{n,h}$ has a natural f\/iltration, and we denote by
${\cal{A}}_n =\operatorname{gr}{\cal{A}}_{n,h}$ the corresponding associated graded
algebra.
To describe (a part of) relations among the parameters $\{q_{ij}\}$ let us
observe that parame\-ters~$\{p_{ij}\}$ and~$\{q_{ij} \}$ are related by the
following identity
\begin{gather*}
q_{ij} q_{jk} -q_{ik}( q_{ij} +q_{jk})+h^2 q_{ik} =
2 p_{ij} p_{ik} p_{jk}(p_{ik}+h) \qquad \text{if} \quad i < j < k.
\end{gather*}
Using this identity we can f\/ind the following relations among parameters in
question
\begin{gather}
q_{ij}^2 q_{jk}^2 +q_{ij}^2 q_{ik}^2 + h^4q_{ik}^2 q_{jk}^2 -
2q_{ij} q_{ik} q_{jk}(q_{ij}+q_{jk} +q_{ik}) \nonumber\\
\qquad {} -2 h^2 q_{ik}(q_{ij} q_{jk}+q_{ij} q_{ik}+q_{jk} q_{ik})
= 8 h q_{ij}q_{ik} q_{jk} p_{ik},
\label{eq:xdef}
\end{gather}
if $ 1 \le i < j < k \le n$.
Finally, we come to a relation of degree $8$
among the ``quantum parameters'' $\{ q_{ij} \}$
\begin{gather*}
\big(\text{l.h.s.\ of \eqref{eq:xdef}}\big)^2 = 64 h^2 q_{ij}^2 q_{ik}^3 q_{jk}^2, \qquad
1 \le i <j < k \le n.
\end{gather*}
There are also higher degree relations among the parameters $\{q_{ij} \}$
some of whose in degree $16$ follow from the deformed Pl\"{u}cker
relation between parameters $\{p_{ij} \}$:
\begin{gather*}
{1 \over p_{ik} p_{jl}}={1 \over p_{ij} p_{kl}} + {1 \over p_{il} p_{jk}} +
{h \over p_{ij} p_{jk} p_{kl}}, \qquad i < j < k < l.
\end{gather*}
However, we don't know how to describe the algebra ${{\cal{A}}}_{n,h}$
generated by quantum parameters $\{ q_{ij} \}_{1 \le i < j \le n}$ even for
$n=4$.
The algebra ${\cal{A}}_n = \operatorname{gr}({\cal{A}}_{n,h})$ is isomorphic to the quotient
algebra of $\mathbb{Q}[x_{ij}, \, 1 \le i < j \le n]$ modulo the ideal generated by the
set of relations between ``quantum parameters''
\begin{gather*}
\left\{\overline{q}_{ij}:=
\left({1 \over z_i-z_j} \right)^2\right\}_{1 \le i < j \le n},
\end{gather*}
which correspond to the Dunkl--Gaudin elements $\{\theta_i \}_{1 \le i \le n}$, see Section~\ref{section3.2} below for details. In this case the parameters
$\{ \overline{q}_{ij} \}$ satisfy the following relations
\begin{gather*}
\overline{q}_{ij}^2 \overline{q}_{jk}^2+\overline{q}_{ij}^2
\overline{q}_{ik}^2+
\overline{q}_{jk}^2 \overline{q}_{ik}^2 = 2 \overline{q}_{ij} \overline{q}_{ik} \overline{q}_{jk}(\overline{q}_{ij} + \overline{q}_{jk}+\overline{q}_{jk})
\end{gather*}
which correspond to the relations~\eqref{eq:xdef} in the special case $h=0$. One can
f\/ind a set of relations in degrees~$6$,~$7$ and~$8$, namely for a given pair-wise distinct integers $1 \le i,j,k,l \le n$, one has
\begin{itemize}\itemsep=0pt
\item one relation in degree $6$
\begin{gather*}
\overline{q}_{ij}^2\overline{q}_{ik}^2\overline{q}_{il}^2+
\overline{q}_{ij}^2\overline{q}_{jk}^2\overline{q}_{jl}^2+
\overline{q}_{ik}^2\overline{q}_{jk}^2\overline{q}_{kl}^2+
\overline{q}_{il}^2\overline{q}_{jl}^2\overline{q}_{kl}^2\\
\qquad {}-2 \overline{q}_{ij} \overline{q}_{ik} \overline{q}_{il} \overline{q}_{jk}
\overline{q}_{jl} \overline{q}_{kl} \left(
{{{\overline{q}_{ij}} \over {\overline{q}_{kl}}}}+
{{{\overline{q}_{kl}} \over {\overline{q}_{ij}}}}+
{{{\overline{q}_{ik}} \over {\overline{q}_{jl}}}}+
{{{\overline{q}_{jl}} \over {\overline{q}_{ik}}}}+
{{{\overline{q}_{il}} \over {\overline{q}_{jk}}}}+
{{{\overline{q}_{jk}} \over {\overline{q}_{il}}}} \right)\\
\qquad{} +
8 \overline{q}_{ij} \overline{q}_{ik} \overline{q}_{il}\overline{q}_{jk}
\overline{q}_{jl}\overline{q}_{kl} = 0;
\end{gather*}
\item three relations in degree $7$
\begin{gather*}
\overline{q}_{ik}
\bigl(\overline{q}_{ij} \overline{q}_{il} \overline{q}_{kl} -
\overline{q}_{ij} \overline{q}_{il} \overline{q}_{jk} +
\overline{q}_{ij} \overline{q}_{jk} \overline{q}_{kl} -
\overline{q}_{il} \overline{q}_{jk} \overline{q}_{kl} \bigr)^2 \\
\qquad {} =
8 \overline{q}_{ij}^2 \overline{q}_{ik}^2 \overline{q}_{jk} \overline{q}_{kl}
\bigl(\overline{q}_{jk}+\overline{q}_{jl}+ \overline{q}_{kl} \bigr) -
4 \overline{q}_{ij}^2 \overline{q}_{il}^2\overline{q}_{jl}
\bigl( \overline{q}_{jk}^2+ \overline{q}_{kl}^2 \bigr) ,
\end{gather*}
\item one relation in degree $8$
\begin{gather*}
\overline{q}_{ij}^2 \overline{q}_{il}^2\overline{q}_{jk}^2\overline{q}_{kl}^2+
\overline{q}_{ij}^2 \overline{q}_{ik}^2 \overline{q}_{jl}^2 \overline{q}_{kl}^2+
\overline{q}_{ik}^2 \overline{q}_{il}^2 \overline{q}_{jk}^2 \overline{q}_{jl}^2 =
2 \overline{q}_{ij} \overline{q}_{ik} \overline{q}_{il}\overline{q}_{jk}
\overline{q}_{jl}\overline{q}_{kl} \bigl(\overline{q}_{ij}\overline{q}_{kl}+
\overline{q}_{ik}\overline{q}_{jl}+\overline{q}_{il}\overline{q}_{jk} \bigr),
\end{gather*}
\end{itemize}
However we don't know does the list of relations displayed above, contains
the all independent relations among the elements
$\{\overline{q}_{ij} \}_{1 \le i < j \le n}$ in degrees~$6$, $7$ and~$8$,
even for $n=4$. In degrees $\ge 9$ and $n \ge 5$ some independent relations
should appear.
Notice that the parameters $\big\{p_{ij}= { hq_j \over q_i-q_j}, \, i < j \big\}$
satisfy the so-called {\it Gelfand--Varchenko} relations, see, e.g.,~\cite{K3}
\begin{gather*}
p_{ij} p_{jk}=p_{ik} p_{ij} + p_{jk} p_{ik}+ h p_{ik}, \qquad i < j < k,
\end{gather*}
whereas parameters $\big\{ {\overline{p}}_{ij}= {1 \over q_i - q_j}, \,i <j \big\}$
satisfy the so-called {\it Arnold} relations
\begin{gather*}
{\overline{p}}_{ij}{\overline{p}}_{jk}=
{\overline{p}}_{ik}{\overline{p}}_{ij}+{\overline{p}}_{jk}{\overline{p}}_{ik},
\qquad i < j < k.
\end{gather*}
\begin{Project}\label{project2.1}
Find Hilbert series ${\rm Hilb}({\cal{A}}_n,t)$ for $n \ge 4$.\footnote{This is a particular case of more general problem we are
interested in. Namely, let $\{f_{\alpha} \in \mathbb{R}[x_1,\ldots,x_n] \}_{1 \le
\alpha \le N}$ be a collection of linear forms, and $k \ge 2$ be an integer.
Denote by $I(\{ f_{\alpha} \})$ the ideal in the ring of polynomials $\mathbb{R}[z_1,
\ldots,z_N]$ generated by polynomials $\Phi(z_1,\ldots,z_N)$ such that
\begin{gather*}
\Phi\big(f_1^{-k},\ldots,f_{N}^{-k}\big) =0.
\end{gather*}
{\it Compute} the Hilbert series (polynomial?) of the quotient algebra
$\mathbb{R}[z_1,\ldots,z_N] / I(\{f_{\alpha} \})$.}
\end{Project}
For example, ${\rm Hilb}({\cal{A}}_3,t)={(1+t)(1+t^2) \over (1-t)^2}$.
Finally, if we set $q_i:= \exp(h z_i)$ and take the limit
$ \lim\limits_{h \to 0} \frac{h^2 q_i q_j}{(q_i-q_j)^2}$,
as a result we obtain the Dunkl--Gaudin parameter ${\overline{q}}_{ij}=
\frac{1}{(z_i-z_j)^2}$.
\end{Comments}
{\bf (III)} Consider the following representation of the degenerate
af\/f\/ine Hecke algebra ${\mathfrak{H}}_n$ on the ring of polynomials $P_n =
\mathbb{Q}[x_1,\ldots,x_n]$:
\begin{itemize}\itemsep=0pt
\item the symmetric group ${\mathbb{S}}_n$ acts on $P_n$ by means of
operators
\begin{gather*}
\overline{s}_i =1+ (x_{i+1}-x_{i}-h) \partial_i, \qquad i=1,\ldots,n-1,
\end{gather*}
\item $y_i$ acts on the ring $P_n$ by multiplication on the variable
$x_i$: $y_i(f(x)) = x_i f(x)$, $f \in P_n$. Clearly,
\begin{gather*}
y_i \overline{s_i}-y_{i+1} \overline{s_i} = h \qquad \text{and} \qquad
y_i({\overline{s}}_i-1)=({\overline{s}}_i-1) y_{i+1} +x_{i+1} - x_i- h.
\end{gather*}
\end{itemize}
In the subsequent discussion we will identify the operator of multiplication
by the variable~$x_i$, namely the operator~$y_i$, with~$x_i$.
This time def\/ine $u_{ij}= p_{ij} (\overline{s}_i-1)$, if $i < j$ and set
$u_{ij} = -u_{ji}$ if $i > j$, where parameters
$\{p_{ij} \}$ satisfy the same conditions as in the previous example.
\begin{Lemma}\label{lemma2.3} The elements $\{ u_{ij},\, 1 \le i < j \le n \}$, satisfy the
dynamical classical Yang--Baxter relations displayed in Lemma~{\rm \ref{lemma2.2}}, equation~\eqref{equation2.7}.
\end{Lemma}
Therefore, the Dunkl elements
\begin{gather*}
\overline{\theta}_i := x_i+\sum_{j \atop j \not= i} u_{ij}, \qquad i=1,\ldots,n,
\end{gather*}
form a commutative set of elements.
\begin{Theorem}[\cite{GRTV}] \label{theorem2.2} Define matrix $\overline{M}_n =(\overline{m}_{ij})_{1 \le i,j \le n}$ as follows
\begin{gather*}
\overline{m}_{i,j}(u;z_1,\ldots,z_n) = \begin{cases}
u-z_i + \sum\limits_{j \not= i} h p_{ij}& \text{if $i=j$},\\
-h-p_{ij} & \text{if $i < j$}, \\
p_{ij} & \text{if $i > j$}.
\end{cases}
\end{gather*}
Then
\begin{gather*}
\operatorname{DET} \big | \overline{M}_n(u;\overline{\theta}_1,\ldots,\overline{\theta}_n)
\big | = \prod_{j=1}^{n}(u-x_j).
\end{gather*}
\end{Theorem}
\begin{Comments}\label{comments2.3}
Let us list a few more representations of the {\it dynamical}
classical Yang--Baxter relations.
\begin{itemize}\itemsep=0pt
\item Trigonometric Calogero--Moser representation. Let $i < j$, def\/ine
\begin{gather*}
u_{ij}={x_j \over x_i-x_j } (s_{ij}-\epsilon), \qquad \epsilon=0 \ \text{or} \ 1, \\
s_{ij}(x_i)
=x_j, \qquad s_{ij}(x_{j})=x_i, \qquad s_{ij}(x_k)=x_k, \qquad \forall \, k \not= i,j.
\end{gather*}
\item Mixed representation:
\begin{gather*}
u_{ij} = \left({\lambda_j \over \lambda_i-\lambda_j} - {x_j \over x_i-x_j}\right)
(s_{ij}-\epsilon), \qquad \epsilon = 0 \ \text{or} \ 1, \qquad s_{ij}(\lambda_k) =\lambda_k, \qquad \forall\,
k.
\end{gather*}
\end{itemize}
We set $u_{ij}=-u_{ji}$, if $i > j$. In all cases we def\/ine Dunkl elements to
be $\theta_i=\sum\limits_{j \not= i} u_{ij}$.
Note that operators
\begin{gather*}
r_{ij} =\left({\lambda_i+\lambda_j \over \lambda_i-\lambda_j} - {x_i+x_j \over
x_i-x_j}\right) s_{ij}
\end{gather*}
satisfy the three term relations: $r_{ij}r_{jk}=r_{ik}r_{ij}+r_{jk}r_{ik}$,
and $r_{jk}r_{ij}=r_{ij}r_{jk}+r_{ik}r_{jk}$, and thus satisfy the
{\it classical} Yang--Baxter relations.
\end{Comments}
\subsubsection[Step functions and the Dunkl--Uglov representations of the degenerate
af\/f\/ine Hecke algebras~\cite{U}]{Step functions and the Dunkl--Uglov representations\\ of the degenerate
af\/f\/ine Hecke algebras~\cite{U}}\label{section2.1.2}
Con\-sider step functions $\eta^{\pm}\colon \mathbb{R} \longrightarrow \{0,1 \}$
\begin{gather*}
\text{(Heaviside function)} \qquad
\eta^{+}(x) =\begin{cases}
1 & \text{if $x \ge 0$},\\
0 & \text{if $x < 0$},
\end{cases} \qquad
\eta^{-}(x) =\begin{cases}
1 & \text{if $x > 0$},\\
0 & \text{if $x \le 0$}.
\end{cases}
\end{gather*}
For any two real numbers $x_i$ and $x_j$ set $\eta_{ij}^{\pm}=\eta^{\pm}(x_i-x_j)$.
\begin{Lemma} \label{lemma2.4} The functions $\eta_{ij}$ satisfy the following relations
\begin{gather*}
\eta_{ij}^{\pm}+\eta_{ji}^{\pm}=1 + \delta_{x_i,x_j}, \qquad
(\eta_{ij}^{\pm})^{2}= \eta_{ij}^{\pm},\\
\eta_{ij}^{\pm} \eta_{jk}^{\pm} = \eta_{ik}^{\pm} \eta_{ij}^{\pm}+
\eta_{jk}^{\pm} \eta_{ik}^{\pm} - \eta_{ik}^{\pm},
\end{gather*}
where $\delta_{x,y}$ denotes the Kronecker delta function.
\end{Lemma}
To introduce the Dunkl--Uglov operators~\cite{U} we need a few more
def\/initions and notation. To start with, denote by $\Delta_i^{\pm}$ the
f\/inite dif\/ference operators: $\Delta_{i}^{\pm}(f)(x_1,\ldots,x_n)=f(\ldots,
x_i \pm 1,\ldots)$. Let as before, $ \{s_{ij},\, 1 \le i \not= j \le n,\,
s_{ij}=s_{ji} \}$, denotes the set of transpositions in the symmetric group
${\mathbb {S}}_n$. Recall that $s_{ij}(x_i)=x_j$, $s_{ij}(x_k)=x_k$, $\forall\, k \not= i,j$.
Finally def\/ine Dunkl--Uglov operators $d_i^{\pm} \colon \mathbb{R}^n \longrightarrow \mathbb{R}^n$
to be
\begin{gather*}
d_i^{\pm} =\Delta_i^{\pm} + \sum_{j < i} \delta_{x_i,x_j}- \sum_{j < i}
\eta_{ji}^{\pm} s_{ij} + \sum_{j > i} \eta_{ij}^{\pm} s_{ij}.
\end{gather*}
To simplify notation, set $u_{ij}^{\pm}:=\eta_{ij}^{\pm} s_{ij}$ if $i < j$, and
${\widetilde{\Delta}}_i^{\pm}= \Delta_{i}^{\pm} +\sum\limits_{j < i} \delta_{x_i,x_j}$.
\begin{Lemma} \label{lemma2.5} The operators $\{ u_{ij}^{\pm},\, 1 \le i < j \le n \}$ satisfy the
following relations
\begin{gather*
\big[u_{ij}^{\pm},u_{ik}^{\pm} + u_{jk}^{\pm}\big]+\big[u_{ik}^{\pm},u_{jk}^{\pm}\big] +
\bigg[u_{ik}^{\pm}, \sum_{j < i} \delta_{x_i,x_j}\bigg] =0 \qquad \text{if} \ \ i < j < k.
\end{gather*}
\end{Lemma}
From now on we {\it assume that $ x_i \in \mathbb{Z}$, $\forall\, i$},
that is, we will work with the restriction of the all operators def\/ined at
beginning of Example~\ref{example2.1}(c), to the subset $\mathbb{Z}^n
\subset \mathbb{R}^n$. It is easy to see that under the assumptions $x_i \in \mathbb{Z}$, $\forall\, i$, we will have
\begin{gather}\label{equation2.10}
\Delta_j^{\pm} \eta_{ij}^{\pm} = (\eta_{ij}^{\pm} \mp \delta_{x_i,x_j})
\Delta_{i}^{\pm}.
\end{gather}
Moreover, using relations~\eqref{equation2.13}, \eqref{equation2.14} one can prove that
\begin{Lemma} \label{lemma2.6} \quad
\begin{itemize}\itemsep=0pt
\item $[u_{ij}^{\pm}, {\widetilde{\Delta}}_i^{\pm} +
{\widetilde{\Delta}}_j^{\pm} ] =0$,
\item $[ u_{ik}^{\pm},{\widetilde{\Delta}_j}^{\pm}]=
\big[u_{ik}^{\pm}, \sum\limits_{j < i} \delta_{x_i,x_j}\big]$, $ i < j < k$.
\end{itemize}
\end{Lemma}
\begin{Corollary} \label{corollary2.1}\quad
\begin{itemize}\itemsep=0pt
\item The operators $\{ u_{ij}^{\pm}, \, 1 \le i < j < k \le n \}$,
and ${\widetilde{\Delta}}_i^{\pm}$, $i=1,\ldots,n $ satisfy the dynamical
classical Yang--Baxter relations
\begin{gather*}
\big[u_{ij}^{\pm},u_{ik}^{\pm} + u_{jk}^{\pm}\big]+\big[u_{ik}^{\pm},u_{jk}^{\pm}\big] +
\big[u_{ik}^{\pm}, {\widetilde{\Delta}}_j\big] =0 \qquad \text{if} \ \ i < j < k.
\end{gather*}
\item The operators $\{s_i:=s_{i,i+1}, \, 1 \le i < n, \, and\,
{\widetilde{\Delta}}_j^{\pm}, \, 1 \le j \le n \}$ give rise to two
representations of the degenerate affine Hecke algebra~${\mathfrak{H}}_n$.
In particular, the Dunkl--Uglov operators are mutually commute:
$[d_i^{\pm},d_j^{\pm}]=0$~{\rm \cite{U}}.
\end{itemize}
\end{Corollary}
\subsubsection{Extended Kohno--Drinfeld algebra and Yangian Dunkl--Gaudin
elements}\label{section2.1.3}
\begin{Definition}\label{definition2.3}
Extended Kohno--Drinfeld algebra is an associative algebra over
$\mathbb{Q}[\beta]$ ge\-ne\-ra\-ted by the elements $\{z_1,\ldots,z_n \}$ and
$\{ y_{ij} \}_{1 \le i \not= j \le n}$ subject to the set of relations
\begin{enumerate}\itemsep=0pt
\item[(i)] The elements $\{ y_{ij} \{_{1 \le i \not= j \le n}$ satisfy the Kohno--Drinfeld relations
\begin{itemize}\itemsep=0pt
\item $y_{ij}=y_{ji}$, $[y_{ij},y_{kl}]=0$ if $i$, $j$, $k$, $l$ are distinct,
\item $[y_{ij},y_{ik}+y_{jk}]=0=[y_{ij}+y_{ik},y_{jk}]$ if $i < j < k$.
\end{itemize}
\item[(ii)] The elements $z_1,\ldots,z_n$ generate the free associative algebra
${\cal{F}}_n$.
\item[(iii)] Crossing relations:
\begin{itemize}\itemsep=0pt
\item $[z_i, y_{jk}] =0$ if $i \not= j,k$, $[z_i,z_{j}]= \beta
[y_{ij},z_i]$ if $i \not= j$.
\end{itemize}
\end{enumerate}
\end{Definition}
To def\/ine the (Yangian) Dunkl--Gaudin elements, cf.~\cite{GRTV}, let us consider
a set of elements $\{p_{ij} \}_{1 \le i \not= j \le n}$ subject to relations
\begin{itemize}\itemsep=0pt
\item $p_{ij}+p_{ji}= \beta$, $[p_{ij},y_{kl}]=0 =[p_{ij},z_k]$ for all
$i$, $j$, $k$,
\item $p_{ij} p_{jk}= p_{ik} ( p_{jk}-p_{ji} )$ if $i < j < k$.
\end{itemize}
Let us set $u_{ij}= p_{ij} y_{ij}$, $i \not= j$, and def\/ine the (Yangian)
Dunkl--Gaudin elements as follows
\begin{gather*}
\theta_i = z_i+ \sum_{j \not= i} u_{ij}, \qquad i=1,\ldots, n.
\end{gather*}
\begin{Proposition}[cf.~\protect{\cite[Lemma~3.5]{GRTV}}] \label{proposition2.1}
The elements $\theta_1,\ldots,\theta_n$ form a mutually commuting family.
\end{Proposition}
Indeed, let $i < j$, then
\begin{gather*}
[\theta_i,\theta_j] =
[z_i,z_j] + \beta [z_i, y_{ij}] + p_{ij} [y_{ij},z_i+z_j]\\
\hphantom{[\theta_i,\theta_j] = }{}
+ \sum_{k \not= i,j} \big(p_{ik} p_{jk} \big[y_{ij}+y_{ik}, y_{jk} \big]
+p_{ik} p_{ji} \big[ y_{ij},y_{ik}+y_{jk} \big] \big) =0.
\end{gather*}
A representation of the extended Kohno--Drinfeld algebra has been constructed
in~\cite{GRTV}, namely one can take
\begin{gather*}
y_{ij}:= T_{ij}^{(1)} T_{ji}^{(1)} -T_{jj}^{(1)}=y_{ji},\qquad z_{i}:=
\beta T_{ii}^{(2)} -\frac{\beta}{2} T_{ii}^{(1)} \big (T_{ii}^{(1)}-1\big),\\
p_{ij}:= \frac{\beta q_j}{q_i -q_j}, \qquad i \not= j,
\end{gather*}
where $q_1,\ldots,q_n$ stands for a set of mutually commuting {\it quantum}
parameters, and $ \big\{T_{ij}^{(s)} \big\}_{1 \le i,j \le n \atop s \in \mathbb{Z}_{\ge 0}}$
denotes the set of generators of the Yangian $Y({\mathfrak{gl}}_n)$, see, e.g.,~\cite{Mo}.
A proof that the elements $\{z_i \}_{1 \le i \le n}$ and $\{y_{ij} \}_{1 \le i \not= j \le n}$ satisfy the extended Kohno--Drinfeld algebra relations
is based on the following relations, see, e.g., \cite[Section~3]{GRTV},
\begin{gather*}
\big[T_{ij}^{(1)}, T_{kl}^{(s)}\big]= \delta_{il} T_{kj}^{(s)} - \delta_{jk} T_{il}^{(s)} , \qquad i,j,k,l = 1,\ldots,n, \qquad s \in \mathbb{Z}_{\ge 0}.
\end{gather*}
\subsection[``Compatible'' Dunkl elements, Manin matrices and
algebras related with weighted\\ complete graphs $r K_{n}$]{``Compatible'' Dunkl elements, Manin matrices and
algebras\\ related with weighted complete graphs $\boldsymbol{r K_{n}}$}\label{section2.2}
Let us consider a collection of generators $\{ u_{ij}^{(\alpha)},\,1 \le i,j
\le n,\, \alpha =1,\ldots,r \}$, subject to the following relations
\begin{itemize}\itemsep=0pt
\item either the unitarity
(the case of sign~``${+}$'') or the symmetry relations (the case of sign~``${-}$'')\footnote{More generally one can impose the $q$-symmetry conditions
\begin{gather*}
u_{ij}+q u_{ji} = 0, \qquad 1 \le i < j \le n
\end{gather*}
and ask about relations among the local Dunkl elements to ensure the
commutativity of the global ones. As one might expect, the matrix
$Q := \big(\theta_j^{(a)}\big)_{1 \le a \le r \atop 1 \le j \le n}$ composed from the
local Dunkl elements should be a~$q$-Manin matrix. See, e.g.,~\cite{CF}, or
\url{https://en.wikipedia.org/wiki/Manin.matrix} for a def\/inition and basic properties of the latter.}
\begin{gather}\label{equation2.11}
u_{ij}^{(\alpha)} \pm u_{ji}^{(\alpha)}=0, \qquad \forall \, \alpha, i,j,
\end{gather}
\item {\it local $3$-term relations}:
\begin{gather}\label{equation2.12}
u_{ij}^{(\alpha)} u_{jk}^{(\alpha)}+ u_{jk}^{(\alpha)} u_{ki}^{\alpha)}+
u_{ki}^{(\alpha)} u_{ij}^{(\alpha)}=0, \qquad i,j,k \ \ \text{are distinct}, \quad 1 \le
\alpha \le r.
\end{gather}
\end{itemize}
We def\/ine {\it global} 3-term relations algebra $3T_{n,r}^{(\pm)}$ as ``compatible product'' of the local 3-term relations algebras. Namely, we
require that the elements
\begin{gather*}
U_{ij}^{({\boldsymbol{\lambda}})}:= \sum_{\alpha=1}^{r} \lambda_{\alpha}
u_{ij}^{(\alpha)}, \qquad 1 \le i,j \le n,
\end{gather*}
satisfy the global 3-term relations
\begin{gather*}
U_{ij}^{(\boldsymbol{\lambda})} U_{jk}^{(\boldsymbol{\lambda})} +
U_{jk}^{(\boldsymbol{\lambda})} U_{ki}^{(\boldsymbol{\lambda})} +
U_{ki}^{(\boldsymbol{\lambda})} U_{ij}^{(\boldsymbol{\lambda})} = 0
\end{gather*}
for all values of parameters
$\{\lambda_i \in \mathbb{R} , \, 1 \le \alpha \le r \}$.
It is easy to check that our request is equivalent to a validity of the
following sets of relations among the generators $\big\{u_{ij}^{(\alpha)} \big\}$
\begin{enumerate}\itemsep=0pt
\item[(a)] {\it local $3$-term relations}: $ u_{ij}^{(\alpha)} u_{jk}^{\alpha)}+u_{jk}^{(\alpha)} u_{ki}^{(\alpha)} +u_{ki}^{\alpha)} u_{ij}^{(\alpha)} =0$,
\item[(b)] {\it $6$-term crossing relations}:
\begin{gather*}
u_{ij}^{(\alpha)} u_{jk}^{(\beta)}+u_{ij}^{(\beta)} u_{jk}^{(\alpha)}+
u_{k,i}^{(\alpha)} u_{ij}^{(\beta)} u_{ki}^{(\alpha)} + u_{jk}^{(\alpha)}
u_{ki}^{(\beta)} + u_{jk}^{(\beta)} u_{ki}^{(\alpha)}=0,
\end{gather*}
$i$, $j$, $k$ are distinct, $\alpha \not= \beta$.
\end{enumerate}
Now let us consider {\it local} Dunkl elements
\begin{gather*}
\theta_{i}^{(\alpha)}:= \sum_{j \neq i} u_{ij}^{(\alpha)}, \qquad j=1,\ldots,n, \quad
\alpha=1,\ldots,r.
\end{gather*}
It follows from the local 3-term relations \eqref{equation2.12} that for a f\/ixed
$\alpha \in [1,r]$ the local Dunkl elements
$\big\{ \theta_i^{(\alpha)} \big\}_{1 \le i \le n \atop 1 \le \alpha \le r}$
either mutually commute
(the sign~``$+$''), or pairwise anticommute (the sign~``$-$''). Similarly, the
global 3-term relations imply that the global Dunkl
elements
\begin{gather*}
\theta_i^{(\lambda)}:= \lambda_1 \theta_i^{(1)}+ \cdots + \lambda_r \theta_i^{(r)} = \sum_{j \not=i} U_{ij}^{(\lambda)}, \qquad
i=1,\ldots,n,
\end{gather*}
also either mutually commute (the case~``$+$'') or pairwise anticommute
(the case~``$-$'').
Now we are looking for a set of relations among the local Dunkl elements
which is a consequence of the commutativity (anticommutativity) of the
global Dunkl elements.
It is quite clear that if $i < j$, then
\begin{gather*}
\big[\theta_i^{(a)},\theta_j^{(b)}\big]_{\pm} =\sum_{a=1}^{r} \lambda_{a}^2 \big[\theta_i^{(a)},\theta_j^{(a)}\big]_{\pm} + \sum_{1 \le a < b \le r}
\lambda_{a} \lambda_b
\big(\big[\theta_i^{(a)},\theta_j^{(b)}\big]_{\pm}+\big[\theta_i^{(b)},
\theta_j^{(a)}\big]_{\pm} \big),
\end{gather*}
and the commutativity (or anticommutativity) of the global Dunkl elements for
all $(\lambda_1,\ldots, \lambda_r) \in \mathbb{R}^{r}$ is equivalent to the following set of relations
\begin{itemize}\itemsep=0pt
\item $[\theta_i^{(a)},\theta_j^{(a)}]_{\pm} =0$,
\item $[\theta_i^{(a)},\theta_j^{(b)}]_{\pm} + [\theta_i^{(b)},\theta_j^{(a)}]_{\pm} = 0$, $a < b$ and $i < j$,
where by def\/inition we set $[a,b]_{\pm}:=a b \mp b a$.
\end{itemize}
In other words, the matrix $\varTheta_n: = \big(\theta_i^{(a)}\big)_{1 \le a \le r
\atop 1 \le i \le n}$
should be either a {\it Manin matrix} (the case~``$+$''), or its super analogue
(the case~``$-$''). Clearly enough that a similar
construction can be applied to the algebras studied in Section~\ref{section2}, {\bf I}--{\bf III},
and thus it produces some interesting examples of the Manin matrices.
It is an interesting {\it problem} to describe
the algebra generated by the local Dunkl elements $\big\{ \theta_i^{(a)}\big\}_{1 \le
a \le r \atop 1 \le i \le n}$ and a commutative subalgebra generated by the
global Dunkl elements inside the former. It is also an interesting {\it
question} whether or not the coef\/f\/icients
$C_1,\ldots ,C_n$ of the column characteristic polynomial $\operatorname{Det}^{\rm col} |\varTheta_n - t I_n| = \sum\limits_{k=0}^{n} C_k t^{n-k}$ of the Manin matrix
$\varTheta_n$ generate a commutative subalgebra? For a def\/inition of the
column determinant of a matrix, see, e.g.,~\cite{CF}.
However a close look at this problem and the question posed needs an
additional treatment and has been omitted from the content of the present
paper.
Here we are looking for a ``natural conditions'' to be imposed on the set
of generators $\{ u_{ij}^{\alpha} \}_{1 \le \alpha
\le r \atop 1 \le i,j \le n}$ in order to ensure that the local
Dunkl elements satisfy the commutativity (or anticommutativity) relations:
\begin{gather*}
\big[\theta_i^{(\alpha)},\theta_j^{(\beta)}\big]_{\pm}=0, \qquad \text{for all} \ \ 1 \le i < j
\le n, \qquad 1 \le \alpha, \beta \le r.
\end{gather*}
The ``natural conditions'' we have in mind
are
\begin{itemize}\itemsep=0pt
\item {\it locality relations}:
\begin{gather}\label{equation2.13}
\big[u_{ij}^{(\alpha)}, u_{kl}^{(\beta)}\big]_{\pm} = 0 ,
\end{gather}
\item {\it twisted classical Yang--Baxter relations}:
\begin{gather}\label{equation2.14}
\big[u_{ij}^{(\alpha)}, u_{jk}^{(\beta)}\big]_{\pm}+ \big[u_{ik}^{(\alpha)},u_{ji}^{(\beta)}\big]_{\pm}+ \big[u_{ik}^{(\alpha)},u_{jk}^{(\beta)}\big]_{\pm}=0,
\end{gather}
if $i$, $j$, $k$, $l$ are distinct and $1 \le \alpha, \beta \le r$.
\end{itemize}
Finally we def\/ine a multiple analogue of the three term relations algebra,
denoted by \linebreak $3T^{\pm}(r K_n)$, to be the quotient of the global $3$-term
relations algebra $3T_{n,r}^{\pm}$ modulo the two-sided ideal generated by
the left hand sides of relations~\eqref{equation2.13},~\eqref{equation2.14} and that of the
following relations
\begin{itemize}\itemsep=0pt
\item $ \big(u_{ij}^{(\alpha)} \big)^2=0$,
$\big[u_{ij}^{(\alpha)},u_{ij}^{(\beta)}\big]_{\pm}=0$, for all $i \not= j$, $\alpha
\not= \beta $.
\end{itemize}
The outputs of this construction are
\begin{itemize}\itemsep=0pt
\item commutative (or anticommutative) quadratic algebra $3T^{(\pm)}(r K_n)$ generated by
the elements $\big\{u_{ij}^{(\alpha)} \big\}_{ 1 \le i < j \le n \atop \alpha =1,
\ldots,r} $,
\item a family of $nr$ either mutually commuting (the case~``$+$''), or
pair-wise anticommuting (the case~``$-$'') local Dunkl elements
$\big\{ \theta_{i}^{(\alpha)} \big\} _{i=1,\ldots,n \atop \alpha=1,\ldots,r} $.
\end{itemize}
We {\it expect} that the subalgebra generated by local Dunkl elements in the
algebra $3T^{+}( r K_n)$ is closely related (isomorphic for $r=2$) with
the coinvariant algebra of
the diagonal action of the symmetric group ${\mathbb{S}}_n$ on the ring of
polynomials $\mathbb{Q}\big[ X_{n}^{(1)},\ldots,X_{n}^{(r)}\big]$, where $X_{n}^{(j)}$ stands
for the set of variables $\big\{x_1^{(j)},\ldots,x_n^{(j)} \big\}$. The algebra $3T^{-}(2 K_n)^{\rm anti}$ has been studied in~\cite{K} and~\cite{BDK}.
In the present paper we state only our old conjecture.
\begin{Conjecture}[A.N.~Kirillov, 2000]\label{conjecture2.1}
\begin{gather*}
{\rm Hilb}\big(3T^{-}( 3 K_n)^{\rm anti},t\big) = (1+t)^n (1+n t)^{n-2} ,
\end{gather*}
where for any algebra~$A$ we denote by $A^{\rm anti}$ the quotient of algebra~$A$
by the two-sided ideal generated by the set of anticommutators $\{ a b + b a \,|\, (a,b) \in A \times A \}$.
\end{Conjecture}
According to observation of M.~Haiman~\cite{H}, the number $2^n (n+1)^{n-2}$
is thought of as being equal to the dimension of the space of triple
coinvariants of the symmetric group~$\mathbb{S}_n$.
\subsection{Miscellany}\label{section2.3}
\subsubsection[Non-unitary dynamical classical Yang--Baxter algebra
${\rm DCYB}_n$]{Non-unitary dynamical classical Yang--Baxter algebra
$\boldsymbol{{\rm DCYB}_n}$}\label{section2.3.1}
Let $\widetilde{{\cal A}_n}$ be the quotient of the algebra ${\mathfrak F}_n$
by the two-sided ideal generated by the rela\-tions~\eqref{equation2.2}, \eqref{equation2.5} and~\eqref{equation2.6}. Consider elements
\begin{gather*}
\theta_i=x_i+\sum_{a \not= i} u_{ia} \qquad \text{and} \qquad {\bar {\theta_j}}= -x_j+
\sum_{b \not= j} u_{bj}, \qquad 1\le i < j \le n.
\end{gather*}
Clearly, if $i < j$, then
\begin{gather*}
[\theta_i,{\bar \theta_j}] +[x_i,x_j] = \left[\sum_{k=1}^{n} x_k , u_{ij}\right] +
\sum_{k \not= i,j} w_{ikj},
\end{gather*}
where the elements $w_{ijk}$, $i < j$, have been def\/ined in Lemma~\ref{lemma2.1}, equation~\eqref{equation2.3}.
Therefore the elements $\theta_i$ and ${\bar \theta_j}$ {\it commute}
in the algebra ${\widetilde{A}_n}$.
In the case when $x_i=0$ for all $i=1, \ldots,n$, the relations
\begin{gather*}
w_{ijk} := [u_{ij},u_{ik}+u_{jk}]+[u_{ik},u_{jk}]=0 \qquad \text{if $i$, $j$, $k$ are all distinct},
\end{gather*}
are well-known as the {\it non-unitary classical Yang--Baxter relations}.
Note that for a given triple of pair-wise distinct~$(i,j,k)$ one has in fact
6~relations. These six relations imply that $[\theta_i,{\bar {\theta_j}}]=0$.
However, in general,
\begin{gather*}
[\theta_i,\theta_j]=\biggl[\sum_{k \not= i,j} u_{ik},
u_{ij}+u_{ji} \biggr] \not= 0.
\end{gather*}
{\bf Dynamical classical Yang--Baxter algebra ${\rm DCYB}_n$.}
In order to ensure the commutativity relations among the Dunkl
elements~\eqref{equation2.1}, i.e., $[\theta_i,\theta_j]=0$ for all $i$, $j$, let us remark
that if $i \not=j$, then
\begin{gather*}
[\theta_,\theta_j] = [x_i+u_{ij},x_j+ u_{ji}]+
[x_i+x_j,u_{ij}] + \left [u_{ij},\sum_{k=1}^{n} x_k\right] \\
\hphantom{[\theta_,\theta_j] =}{}
+
\sum_{k=1 \atop k \not= i,j}^{n} [u_{ij}+u_{ik},u_{jk}]+[u_{ik}, u_{ji}] +[x_i,u_{jk}]+[u_{ik},x_j]+ [x_k,u_{ij}].
\end{gather*}
\begin{Definition} \label{definition2.4}
Def\/ine {\it dynamical non-unitary classical Yang--Baxter algebra ${\rm DNUCYB}_n$}
to be the quotient of the free associative algebra
$\mathbb{Q} \langle \{ x_{i}, \,1 \le i \le n\}, \, \{ u_{ij} \}_{1 \le i \not= j \le n}
\rangle$
by the two-sided ideal generated by the following set of relations
\begin{itemize}\itemsep=0pt
\item zero curvature conditions:
\begin{gather}\label{equation2.15}
[x_i+u_{ij},x_j+ u_{ji}]=0, \qquad 1 \le i \not= j \le n,
\end{gather}
\item conservation laws conditions:
\begin{gather*}
\left[u_{ij},\sum_{k=1}^{n} x_k\right] =0 \qquad \text{for all} \ \ i \not= j, k.
\end{gather*}
\item crossing relations:
\begin{gather*}
[x_i+x_j,u_{ij}]=0, \qquad i \not= j.
\end{gather*}
\item twisted dynamical classical Yang--Baxter relations:
\begin{gather*
[u_{ij}+u_{ik},u_{jk}]+[u_{ik}, u_{ji}] +[x_i,u_{jk}]+[u_{ik},x_j]+
[x_k,u_{ij}]=0,
\end{gather*}
$i$, $j$, $k$ are distinct.
\end{itemize}
\end{Definition}
It is easy to see that the twisted classical Yang--Baxter relations
\begin{gather}\label{equation2.17}
[u_{ij}+u_{ik},u_{jk}]+[u_{ik},u_{ji}]= 0, \qquad i,j,k \ \ \text{are distinct},
\end{gather}
for a f\/ixed triple of distinct indices $i$, $j$, $k$ contain in fact~$3$ dif\/ferent relations whereas the non-unitary classical
Yang--Baxter relations
\begin{gather*}
[u_{ij}+u_{ik},u_{jk}]+[u_{ij}, u_{ik}], \qquad i,j,k \ \ \text{are distinct},
\end{gather*}
contain $6$ dif\/ferent relations for a f\/ixed triple of distinct indices
$i$, $j$, $k$.
\begin{Definition}\label{definition2.5} \quad
\begin{itemize}\itemsep=0pt
\item Def\/ine {\it dynamical classical Yang--Baxter algebra ${\rm DCYB}_n$} to
be the quotient of the algebra ${\rm DNUCYB}_n$ by the two-sided ideal generated by
the elements
\begin{gather*}
\sum_{k \not=i,j} [u_{ik},u_{ij}+u_{ji}] \qquad \text{for all} \ \ i \not= j.
\end{gather*}
\item Def\/ine {\it classical Yang--Baxter algebra ${\rm CYB}_n$} to be the
quotient of the dynamical classical Yang--Baxter algebra ${\rm DCYB}_n$ by the set
of relations
\begin{gather*}
x_i=0\qquad \text{for} \ \ i=1,\dots,n.
\end{gather*}
\end{itemize}
\end{Definition}
\begin{Example}\label{examples2.1}
Def\/ine
\begin{gather*}
p_{ij}(z_1,\ldots,z_n) =\begin{cases}
\dfrac{z_i}{z_i-z_j} & \text{if $1 \le i < j \le n$},\vspace{1mm}\\
- \dfrac{z_j}{z_j-z_i} & \text{if $ n \ge i > j \ge 1$}.
\end{cases}
\end{gather*}
Clearly, $p_{ij}+p_{ji}=1$. Now def\/ine operators $u_{ij} = p_{ij} s_{ij}$,
and the truncated Dunkl operators to be $\theta_i = \sum\limits_{j \not= i}
u_{ij}$, $i=1, \ldots,n$. All these operators act on the f\/ield of rational
functions $\mathbb{Q}(z_1,\ldots,z_n)$; the operator $s_{ij}=s_{ji}$ acts as the
exchange operator, namely, $s_{ij}(z_i)=z_j$, $s_{ij}(z_k)=z_k$, $\forall\, k \not=
i,j$, $s_{ij}(z_j)=z_i$.
Note that this time one has
\begin{gather*}
p_{12} p_{23} = p_{13} p_{12}+p_{23} p_{13} - p_{13}.
\end{gather*}
It is easy to see that the operators $\{ u_{ij},\, 1
\le i \not= j \le n \}$ satisfy relations~\eqref{equation3.1}, and
therefore, satisfy the twisted classical Yang--Baxter relations~\eqref{equation2.14}. As a~corollary we obtain that the truncated Dunkl operators $\{\theta_i,\,i=1,
\ldots,n \}$ are pair-wise commute. Now consider the Dunkl operator $D_i=
\partial_{{z_{i}}} + h \theta_i$, $i=1,\ldots,n$, where~$h$ is a~parameter.
Clearly that $ [ \partial_{{z_i}} +\partial_{{z_j}}, u_{ij}]=0$, and
therefore $[D_i,D_j]=0$, $\forall\, i,j$. It easy to see that
\begin{gather*}
s_{i,i+1} D_{i} - D_{i+1} s_{i,i+1}=h, \qquad [D_i, s_{j,j+1}]=0 \qquad \text{if} \ \ j \not=i, i+1.
\end{gather*}
In such a manner we come to the well-known representation of the degenerate
af\/f\/ine Hecke algebra ${\mathfrak H}_n$.
\end{Example}
\subsubsection{Dunkl and Knizhnik--Zamolodchikov elements}\label{section2.3.2}
Assume that $\forall \, i$, $x_i=0$, and generators $\{u_{ij},\,
1\le i < j \le n \}$ satisfy the locality conditions~\eqref{equation2.2} and the
classical Yang--Baxter relations
\begin{gather*}
[u_{ij},u_{ik}+u_{jk}]+[u_{ik},u_{jk}]=0 \qquad \text{if} \ \ 1 \le i < j < k \le n.
\end{gather*}
Let $y,z,t_1,\ldots,t_n $ be parameters, consider the rational function
\begin{gather*}
F_{\rm CYB}(z;{\boldsymbol{t}}):= F_{\rm CYB}(z;t_1,\ldots,t_n)= \sum_{1 \le i < j \le n} {(t_i
-t_j) u_{ij} \over (z-t_i)(z-t_j)}.
\end{gather*}
Then
\begin{gather*}
[F_{\rm CYB}(z;{\boldsymbol{t}}),F_{\rm CYB}(y;{\boldsymbol{t}})]=0 \qquad \text{and} \qquad \operatorname{Res}_{z=t_{i}}
F_{\rm CYB}(z;{\boldsymbol{t}}) =\theta_i .
\end{gather*}
Now assume that a set of generators $\{ c_{ij},\, 1 \le i \not= j
\le n\}$ satisfy the locality and symmetry (i.e., $c_{ij}=c_{ji}$) conditions,
and the Kohno--Drinfeld relations:
\begin{gather*}
[c_{ij}, c_{kl}]=0 \qquad \text{if} \ \ \{i,j\} \cap \{k,l\} = {\varnothing}, \\
[c_{ij},c_{jk}+c_{ik}]=0=[c_{ij}+c_{ik},c_{jk}], \qquad i < j < k.
\end{gather*}
Let $y,z,t_1,\ldots,t_n $ be parameters, consider the rational function
\begin{gather*}
F_{\rm KD}(z;{\boldsymbol{t}}):= F_{\rm KD}(z;t_1,\ldots,t_n)=\sum_{1 \le i \not= j \le n}
{ c_{ij} \over (z-t_i)(t_i-t_j)} = \sum_{1 \le i < j \le n} {c_{ij} \over
(z-t_i)(z-t_j)}.
\end{gather*}
Then
\begin{gather*}
[F_{\rm KD}(z;{\boldsymbol{t}}),F_{\rm KD}(y;{\boldsymbol{t}})] =0 \qquad \text{and} \qquad \operatorname{Res}_{z=t_i}
F_{\rm KD}(z;{\boldsymbol{t}})= {\rm KZ}_{i},
\end{gather*}
where
\begin{gather*}
{\rm KZ}_i=\sum_{j=1 \atop j \not=i}^{n} {c_{ij} \over t_i -t_j}
\end{gather*}
denotes the truncated {\it Knizhnik--Zamolodchikov} element.
\subsubsection{Dunkl and Gaudin operators}\label{section2.3.3}
{\bf (a)~Rational Dunkl operators.} Consider the quotient of the algebra
${\rm DCYB}_n$, see Def\/i\-ni\-tion~\ref{definition2.2}, by the
two-sided ideal generated by elements
\begin{gather*}
\{[x_i+x_j, u_{ij}] \} \qquad \text{and} \qquad \{ [x_k,u_{ij}],\, k \not= i,j \}.
\end{gather*}
Clearly the Dunkl elements~\eqref{equation2.1} mutually commute. Now let us consider
the so-called {\it Calogero--Moser} representation of the algebra ${\rm DCYB}_n$ on
the ring of polynomials $R_n:=\mathbb{R}[z_1,\ldots,z_n]$ given by
\begin{gather*}
x_i(p(z))= \lambda {\partial p(z) \over \partial z_i}, \qquad u_{ij}(p(z))=
{1 \over
z_i-z_j} (1-s_{ij}) p(z),\qquad p(z) \in R_n.
\end{gather*}
The symmetric group ${\mathbb{S}}_n$ acts on the ring $R_n$ by means of
transpositions $s_{ij} \in {\mathbb{S}}_n$:
$s_{ij}(z_i)=z_j$, $s_{ij}(z_j)=z_i$, $s_{ij}(z_k)=z_k$ if $k \not=i,j$.
In the Calogero--Moser representation the Dunkl elements~$\theta_i$ becomes the
rational Dunkl operators~\cite{Du}, see Def\/inition~\ref{definition1.1}. Moreover, one has
$[x_k,u_{ij}]=0$ if$k \not= i,j$, and
\begin{gather*}
x_i u_{ij}=u_{ij} x_j + {1 \over z_i - z_j} (x_i-x_j-u_{ij}), \qquad
x_j u_{ij}=u_{ij} x_i - {1 \over z_i - z_j} (x_i-x_j-u_{ij}).
\end{gather*}
{\bf (b)~Gaudin operators.}
The Dunkl--Gaudin representation of the algebra ${\rm DCYB}_n$ is def\/ined on the
f\/ield of rational functions $K_n:= \mathbb{R}(q_1,\ldots,q_n)$~and given by
\begin{gather*}
x_i(f(q)):= \lambda {\partial f(q) \over \partial q_i},\qquad u_{ij}= {s_{ij}
\over q_i-q_j}, \qquad f(q) \in K_n,
\end{gather*}
but this time we {\it assume} that $w(q_i)=q_i$, $\forall\, i \in [1,n]$ and
for all $w \in {\mathbb{S}}_n$. In the Dunkl--Gaudin representation the
Dunkl elements becomes the rational Gaudin operators, see, e.g.,~\cite{MTV}.
Moreover, one has $[x_k,u_{ij}]= 0$, if~$k \not=i,j$, and
\begin{gather*}
x_i u_{ij}= u_{ij} x_j - {u_{ij} \over q_i-q_j},\qquad x_j u_{ij}= u_{ij} x_i+
{u_{ij} \over q_i -q_j}.
\end{gather*}
\begin{Comments}\label{comments2.4}
It is easy to check that if $f \in \mathbb{R}[z_1,\ldots,z_n]$, and $x_i:= {\frac{\partial}{\partial z_{i}}}$, then the
following commutation relations are true
\begin{gather*}
x_i f= f x_i + \frac{\partial}{\partial_{z_{i}}}(f), \qquad u_{ij} f = s_{ij}(f) u_{ij} + \partial_{z_{i},z_{j}}(f).
\end{gather*}
Using these relations it easy to check that in the both cases~$({\boldsymbol{a}})$ and~$({\bf b})$ the elementary symmetric polynomials $e_k(x_1,\ldots,x_n)$
commute with the all generators $ \{u_{ij} \}_{1 \le i,j \le n}$, and
therefore commute with the all Dunkl elements $\{\theta_i \}_{1 \le i \le n}$.
Let us {\it stress} that $[\theta_i, x_k] \not= 0$ for all $1 \le i,k \le n$.
\end{Comments}
\begin{Project}\label{project2.2}
Describe a commutative algebra generated by the Dunkl elements
$\{\theta_i \}_{1 \le i \le n}$ and the elementary symmetric polynomials
$ \{ e_k(x_1,\ldots,x_n) \}_{1 \le k \le n}$.
\end{Project}
\subsubsection[Representation of the algebra $3T_n$ on the free algebra
$\mathbb{Z} \langle t_1,\ldots,t_n \rangle$]{Representation of the algebra $\boldsymbol{3T_n}$ on the free algebra
$\boldsymbol{\mathbb{Z} \langle t_1,\ldots,t_n \rangle}$}\label{section2.3.4}
Let ${\mathcal{F}}_n =\mathbb{Z} \langle t_1,\ldots,t_n \rangle $ be free associative
algebra over the ring of integers $\mathbb{Z}$, equipped with the action of the
symmetric group $\mathbb{S}_n$: $s_{ij}(t_i)=t_j$, $s_{ij}(t_k)=t_k$, $\forall\,
k \not=i,j$.
Def\/ine the action of $u_{ij} \in 3T_n$ on the set of generators of the algebra
$\mathcal{F}_n$ as follows
\begin{gather*
u_{ij}(t_k)= \delta_{i,k} t_i t_j - \delta_{j,k} t_j t_i.
\end{gather*}
The action of generator $u_{ij}$ on the whole algebra $\mathcal{F}_n$ is
def\/ined by linearity and the twisted Leibniz rule:
\begin{gather*}
u_{ij}(1)=0, \qquad u_{ij}(a+b)=u_{ij}(a)+u_{ij}(b),\qquad u_{ij}(a b)=u_{ij}(a) b+s_{ij}(a) u_{ij}(b).
\end{gather*}
It is easy to see from~\eqref{equation2.15} that
\begin{gather*
s_{ij} u_{jk}= u_{ik} s_{ij},\qquad s_{ij} u_{kl}=u_{kl} s_{ij} \qquad \text{if} \ \ \{i,j\}
\cap \{ k,l\}=\varnothing, \qquad u_{ij}+u_{ji}=0.
\end{gather*}
Now let us consider operator
\begin{gather*}
u_{ijk}:= u_{ij} u_{jk}-u_{jk} u_{ik} -u_{ik} u_{ij},\qquad 1 \le i < j < k \le n.
\end{gather*}
\begin{Lemma}\label{lemma2.7}
\begin{gather*}
u_{ijk}(a b)=u_{ijk}(a) b+s_{ij} s_{jk}(a) u_{ijk}(b),\qquad a,b \in \mathcal{F}_n.
\end{gather*}
\end{Lemma}
\begin{Lemma}\label{lemma2.8}
\begin{gather*}
u_{ijk}(a)=0\qquad \forall\, a \in \mathcal{F}_n .
\end{gather*}
\end{Lemma}
Indeed,
\begin{gather*}
u_{ijk}(t_i)=-u_{jk} (u_{ij}(t_i))-u_{ik}(u_{ij}(t_i)) = -t_i u_{jk}(t_k) -u_{ik}(t_i) t_j =t_{i}(t_k t_j) -(t_i t_k) t_j =0,\\
u_{ijk}(t_k)=u_{ij}(u_{jk}(t_k)) -u_{jk}(u_{ik}(t_k)) =-u_{ij}(t_kt_j)+u_{jk}(t_kt_i)=t_k(u_{ij}(t_j) + u_{jk}(t_k)t_i = 0, \\
u_{ijk}(t_j)= u_{ij}(u_{jk}(t_j)) - u_{ik}(u_{ij}(t_j))= - u_{ij}(t_j)t_k-
t_ju_{ik}(t_i) =(t_jt_i)t_k-t_j(t_it_k)=0.
\end{gather*}
Therefore Lemma~\ref{lemma2.8} follows from Lemma~\ref{lemma2.7}.
Let $\mathcal{F}_n^{\bullet}$ be the quotient of the free algebra
$\mathcal{F}_n$ by the two-sided ideal generated by elements
$t_i^2t_j-t_jt_i^2$, $1 \le i \not= j \le n$. Since
$u_{i,j}^2(t_i)=t_it_j^2-t_j^2t_i$, one can
def\/ine a representation of the algebra~$3T_n^{(0)}$ on that
$\mathcal{F}_n^{\bullet}$. One can also def\/ine a representation of the
algebra $3T_n^{(0)}$ on that~$\mathcal{F}_n^{(0)}$, where
$\mathcal{F}_n^{(0)}$ denotes the quotient of the algebra~$\mathcal{F}_n$ by
the two-sided ideal generated by elements
$\{ t_{i}^2,\, 1 \le i \le n \}$. Note that
$(u_{i,k} u_{j,k} u_{i,j})(t_k)=[t_i t_j t_i,t_k] \not= 0$ in the algebra
$\mathcal{F}_n^{(0)}$, but the elements $u_{i,j} u_{i,k} u_{j,k} u_{i,j}$,
$ 1 \le i < j < k \le n$, which belong to the kernel of the Calogero--Moser
representation \cite{K}, act trivially both on the algebras~$\mathcal{F}_n^{(0)}$ and that~$\mathcal{F}_n^{\bullet}$.
Note f\/inally that the algebra $\mathcal{F}_n^{(0)}$ is
{\it Koszul} and has Hilbert series
${\rm Hilb}\big(\mathcal{F}_n^{(0)},t\big)={1+t \over 1-(n-1) t}$, whereas
the algebra $\mathcal{F}_n^{\bullet}$ is {\it not} Koszul for
$n \ge 3$, and
\begin{gather*}
{\rm Hilb}
(\mathcal{F}_n^{\bullet},t)= {1 \over (1-t)(1-(n-1)t)(1-t^2)^{n-1}}.
\end{gather*}
In Appendix~\ref{appendixA.5} we apply the representation introduced in this section to the
study of relations in the subalgebra $Z_{n}^{(0)}$ of the algebra
$3T_{n}^{(0)}$ generated by the elements $u_{1,n}, \ldots,u_{n-1,n}$. To
distinguish the generators $\{ u_{ij} \}$ of the algebra $3T_n^{(0)}$ from
the introduced in this section opera\-tors~$u_{ij}$ acting on it, in Appendix~\ref{appendixA.5} we will use for the latter notation $\nabla_{ij}:= u_{ij}$.
\subsubsection{Kernel of Bruhat representation}\label{section2.3.5}
Bruhat representations, classical and quantum, of algebras~$3T_{n}^{(0)}$ and
$3QT_{n}$ can be seen as a~connecting link between commutative subalgebras
generating by either additive or multiplicative Dunkl elements in these
algebras, and classical and quantum Schubert and Grothendieck calculi.
$(\bf Ia)$ {\bf Bruhat representation of algebra} $3T_n^{(0)}$, cf.~\cite{FK}.
Def\/ine action of $u_{i,j} \in 3T_n^{(0)}$ on the group ring of the symmetric
group $\mathbb{Z}[{\mathbb S}_n]$ as follows: let $w \in {\mathbb S}_n$, then
\begin{gather*} u_{i,j} w=\begin{cases}
w s_{ij} & \text{if \ $l(w s_{ij})=l(w)+1 $}, \\
0 & \text{otherwise}.
\end{cases}
\end{gather*}
Let us remind that $s_{ij} \in {\mathbb S}_n$ denotes the transposition
that interchanges~$i$ and~$j$ and f\/ixes each $k \not= i,j$; for each
permutation $u \in {\mathbb S}_n$, $l(u)$ denotes its length.
$(\bf Ib)$ {\bf Quantum Bruhat representation of algebra} $3QT_n$, cf.~\cite{FK}.
Let us remind that algebra $3QT_n$ is the quotient of the 3-term relations
algebra $3T_n$ by the two-sided ideal generated by the elements
\begin{gather*}
\{u_{ij}^2, |j-i| \ge 2 \} \bigcup \{u_{i,i+1}^2=q_i,~i=1,\ldots,n-1 \}.
\end{gather*}
Def\/ine the $\mathbb{Z}[q]-$linear action of $u_{i,j} \in 3QT_n$, $i < j$, on the
extended group ring of the symmetric group $\mathbb{Z}[q] [{\mathbb S}_n]$ as follows: let $w \in {\mathbb S}_n$, and $q_{ij}=q_i q_{i+1} \cdots q_{j-1}$, $i < j$, then
\begin{gather*}
u_{i,j} w=\begin{cases}
w s_{ij} & \text{if \ $l(w s_{ij})=l(w)+1 $}, \\
q_{ij} w s_{ij} & \text{if \ $l(ws_{ij})=l(w)-l(s_{ij})$}, \\
0 & \text{otherwise}.
\end{cases}
\end{gather*}
Let us remind, see, e.g.,~\cite{M}, that in general one has
\begin{gather*}
l(w s_{ij})=\begin{cases}
l(w)-2 e_{ij}-1 & \text{if} \ \ w(i) > w(j), \\
l(w)+2~e_{ij}+1 & \text{if} \ \ w(i) < w(j).
\end{cases}
\end{gather*}
Here $e_{ij}(w)$ denotes the number of $k$ such that $i < k < j$ and $w(k)$
lies between~$w(i)$ and~$w(j)$. In particular, $l(ws_{ij})=l(w)+1$ if\/f
$e_{ij}(w)=0$ and $w(i) < w(j)$;
$l(ws_{ij})=l(w)-l(s_{ij})=l(w)-2(j-i)+1$ if\/f $w(i) > w(j)$ and $e_{ij}=j-i-1$ is the maximal possible.
$({\bf II})$ {\bf Kernel of the Bruhat representation.}
It is not dif\/f\/icult to see that the following elements of degree
three and four belong to the kernel of the Bruhat representation:
\begin{gather*}
({\bf IIa}) \quad u_{i,j}u_{i,k}u_{i,j} \qquad \text{and} \qquad u_{i,k} u_{j,k} u_{i,k} \qquad
\text{if} \ \ 1 \le i < j < k \le n;\\
({\bf IIb}) \quad u_{i,k}u_{i,l}u_{j,l} \qquad \text{and} \qquad u_{j,l}u_{i,l}u_{i,k};\\
({\bf IIc}) \quad u_{il} u_{ik} u_{jl} u_{il}, \qquad u_{il}u_{ij}u_{kl}u_{il}, \qquad
u_{ik}u_{il}u_{jk}u_{ik}, \\
\hphantom{({\bf IIc}) \quad} u_{ij}u_{ik}u_{il}u_{ij}, \qquad u_{ik}u_{il}u_{ij}u_{ik}
\qquad \text{if} \ \ 1 \le i < j < k <l \le n .
\end{gather*}
This observation motivates the following def\/inition.
\begin{Definition}\label{definition2.6} {\it The reduced 3-term relation algebra $3T_n^{\rm red}$} is
def\/ined to be the quotient of the algebra $3T_n^{(0)}$ by the two-sided ideal
generated by the elements displayed in {\bf IIa}--{\bf IIc} above.
\end{Definition}
\begin{Example}\label{example2.1}
\begin{gather*}
{\rm Hilb}\big(3T_3^{\rm red},t\big)=(1,3,4,1), \qquad \dim \big(3T_3^{\rm red}\big) =9,\\
{\rm Hilb}\big(3T_4^{\rm red},t\big)=(1,6,19,32,19,6,1),\qquad \dim \big(3T_4^{\rm red}\big)=84, \\
{\rm Hilb}\big(3T_5^{\rm red},t\big)=(1,10,55,190,383,370,227,102,34,8,1), \qquad \dim\big(3T_5^{\rm red}\big)=1374.
\end{gather*}
\end{Example}
We {\it expect} that $\dim (3T_n^{red})_{{n \choose 2}-1} = 2(n-1)$ if $n \ge 3$.
\begin{Theorem} \label{theorem2.3} \quad
\begin{enumerate}\itemsep=0pt
\item[$1.$] The algebra $3T_n^{\rm red}$ is finite-dimensional, and its Hilbert
polynomial has degree ${n \choose 2}$.
\item[$2.$] The maximal degree ${n \choose 2}$ component of the algebra
$3T_n^{\rm red}$ has dimension one and generated by any element which is
equal to the product $($in any order$)$ of all generators of the algebra
$3T_n^{\rm red}$.
\item[$3.$] The subalgebra in $3T_n^{\rm red}$ generated by the elements
$\{ u_{i,i+1},\, i=1,\ldots,n-1 \}$ is canonically isomorphic to the
nil-Coxeter algebra~${\rm NC}_n$. In particular, its Hilbert polynomial is equal
to $[n]_{t} !:= \prod\limits_{j=1}^{n} {(1-t^{j}) \over 1-t}$, and the element
$\prod\limits_{j=1}^{n-1} \prod\limits_{a=j}^{1} u_{a,a+1}$ of degree ${n \choose 2}$
generates the maximal degree component of the algebra $3T_n^{\rm red}$.
\item[$4.$] The subalgebra over $\mathbb{Z}$ generated by the Dunkl elements
$\{\theta_1, \ldots, \theta_n \}$ in the algebra $3T_n^{\rm red}$ is canonically
isomorphic to the cohomology ring $H^{*}({\cal F}l_{n}, \mathbb{Z})$ of the type~$A$ flag variety ${\cal F}l_n$.
\end{enumerate}
\end{Theorem}
A def\/inition of the nil-Coxeter algebra ${\rm NC}_n$ one can f\/ind in Section~\ref{section4.1.1}.
It is known, see~\cite{Ba} or Section~\ref{section4.1.1}, that the subalgebra generated
by the elements $\{ u_{i,i+1},\,i=1,\ldots,n-1 \}$ in the whole algebra
$3T_n^{(0)}$ is canonically isomorphic to the nil-Coxeter algebra~${\rm NC}_n$ as well.
We {\bf expect} that the kernel of the Bruhat representation of the algebra
$3T_n^{(0)}$ is generated by all {\it monomials} of the form
$u_{i_{1},j_{1}} \cdots u_{i_{k},j_{k}}$ such that the sequence of
transpositions $t_{i_{1},j_{1}},\ldots, t_{i_{k},j_{k}}$
does not correspond to {\it a~path} in the Bruhat graph of the symmetric group~${\mathbb S}_n $. For example if $1 \le i < j < k < l \le n$, the elements $u_{i,k}u_{i,l}u_{j,l}$ and $u_{j,l}u_{i,l}u_{i,k}$ do belong to the
kernel of the Bruhat representation.
\begin{Problem} \label{problem2.1}\quad
\begin{enumerate}\itemsep=0pt
\item[$1.$] The image of the Bruhat representation of the algebra
$3T_n^{(0)}$
defines a subalgebra
\begin{gather*}
\operatorname{Im}\big(3T_n^{(0)}\big) \subset \operatorname{End}_{\mathbb{Q}} (\mathbb{Q}[{\mathbb{S}}_{n}] ).
\end{gather*}
Does this image isomorphic to the algebra $3T_n^{\rm red}$? Compute Hilbert polynomials of algebras $\operatorname{Im}\big(3T_n^{(0)}\big)$ and
$3T_n^{\rm red}$.
\item[$2.$] Describe the image$($s$)$ of the affine nil-Coxeter algebra
${\widetilde{{\rm NC}}}_n$, see Section~{\rm \ref{section4.1.1}}, in the algebras $3T_n^{\rm red}$ and
$ \operatorname{End}_{\mathbb{Q}}(\mathbb{Q}[{\mathbb{S}}_{n}] )$.
\end{enumerate}
\end{Problem}
\subsubsection[The Fulton universal ring \cite{Fu}, multiparameter quantum
cohomology of f\/lag\\ varieties \cite{FK} and the full Kostant--Toda
lattice \cite{FTL+,FTL}]{The Fulton universal ring \cite{Fu}, multiparameter
quantum cohomology\\ of f\/lag varieties \cite{FK} and the full Kostant--Toda
lattice \cite{FTL+,FTL}}\label{section2.3.6}
Let $X_{n}= (x_1,\ldots, x_{n})$ be be a set of variables, and
\begin{gather*}
{\boldsymbol{g}}:= {\boldsymbol{g}}^{(n)}= \{g_a[b] \,|\, a \ge 1, \, b \ge 1, \, a+b \le n
\}
\end{gather*}
be a set of parameters; we put $\deg(x_i)=1$ and $\deg(g_a[b])=b+1$, and set
$g_k[0]:=x_k$, $k=1,\ldots,n$. For a subset $S \subset [1,n]$ we denote by $X_{S}$ the set of variables $\{ x_i\,|\, i \in S \}$.
Let $t$ be an auxiliary variable, denote by $M=(m_{ij})_{1 \le i,j \le n}$
the matrix of size~$ n$ by~$n$ with the following elements:
\begin{gather*}
m_{i,j}=\begin{cases}
x_i+t & \text{if \ $ i=j$}, \\
g_i[j-i]& \text{if \ $j > i$}, \\
-1& \text{if \ $i-j=1$}, \\
0 & \text{if \ $i-j > 1$}.
\end{cases}
\end{gather*}
Let $P_n(X_n,t) =\det |M |$.
\begin{Definition} \label{definition2.7}
The Fulton universal ring ${\cal R}_{n-1}$ is def\/ined to be the
quotient\footnote{If $P(t,X_n)=\sum\limits_{k \ge 1} f_k(X_n) t^k$, $f_k(X_n) \in \mathbb{Q}[Xn]$ is a polynomial, we denote by
$\langle P(t,X_n) \rangle$
the ideal in the polynomial ring $\mathbb{Q}[X_n]$ generated by the coef\/f\/icients $\{f_1,f_2,\ldots \}$.\label{footnote26}}
\begin{gather*}
{\cal R}_{n-1} = \mathbb{Z}\big[{\boldsymbol{g}}^{(n)}\big][x_1,\ldots,x_{n}]/
\langle P_n(X_{n},t)-t^{n} \rangle.
\end{gather*}
\end{Definition}
\begin{Lemma} \label{lemma2.9}
Let $P_n(X_n,t)=\sum\limits_{k=0}^{n}c_k(n)t^{n-k}$, $c_0(n)=1$. Then
\begin{gather}\label{equation2.20}
c_k(n):=c_k\big(n;X_n,{\boldsymbol{g}}^{(n)}\big)=
\sum_{{1 \le i_1 < i_2 < \cdots < i_s < n
\atop j_1 \ge 1,\ldots, j_s \ge 1}
\atop m:=\sum(j_a+1) \le n} \prod_{a=1}^{s}g_{i_a}[j_a]
e_{k-m}\big(X_{[1,n] {\setminus} \bigcup\limits_{a=1}^{s} [i_a,i_a+j_a ]} \big),
\end{gather}
where in the summation we assume additionally that the sets
$[i_a,i_a+j_a]:= \{i_a,i_{a}+1,\ldots,i_a+j_a \}$, $a=1,\ldots, s$, are
pair-wise disjoint.
\end{Lemma}
It is clear that ${\cal R}_{n-1} = \mathbb{Z}[{\boldsymbol{g}}^{(n)}][x_1,\ldots,x_{n}]/
\langle c_{n}(1),\ldots,c_{n}(n) \rangle $.
One can easily see that the coef\/f\/icients $c_k(n)$ and $g_m[k]$ satisfy the
following recurrence relations~\cite{Fu}:
\begin{gather*}\label{equation2.21}
c_k(n)=c_k(n-1)+\sum_{a=0}^{k-1}g_{n-a}[a]c_{k-a-1}(n-a-1),\qquad c_0(n)=1,
\\
g_m[k]=c_{k+1}(m+k)-c_{k+1}(m+k-1)-
\sum_{a=0}^{k-1}g_{m+k-a}[a]c_{k-a}(m+k-a), \\
g_{m}[0]:=x_m.
\end{gather*}
On the other hand, let $\{ q_{ij} \}_{1 \le i < j \le n}$ be a set of
(quantum) parameters, and $e_k^{({\boldsymbol{q}})}(X_n)$ be the multiparameter
quantum elementary polynomial introduced in~\cite{FK}. We are interested in
to describe a set of relations between the parameters $ \{ g_i[j] \}_{i \ge 1, j \ge 1 \atop i+j \le n}$ and the quantum parameters $\{ q_{ij} \}_{1 \le i < j \le n}$ which implies that
\begin{gather*}
c_k(n)= e_{k}^{({\boldsymbol{q}})}(X_n) \qquad \text{for} \quad k= 1,\ldots,n.
\end{gather*}
To start with, let us recall the recurrence relations among the quantum
elementary polynomials, cf.~\cite{P}. To do so, consider the generating
function
\begin{gather*}
E_n\big( X_n; \{q_{ij} \}_{1 \le i < j \le n}\big) = \sum_{k=0}^{n}
e_k^{({\boldsymbol{q}})}(X_n) t^{n-k}.
\end{gather*}
\begin{Lemma}[\cite{FGP,P}]\label{lemma2.10}
One has
\begin{gather*}
E_n\big( X_n; \{q_{ij} \}_{1 \le i < j \le n}\big) = (t+x_n)
E_{n-1}\big( X_{n-1}; \{q_{ij} \}_{1 \le i < j \le n-1}\big) \\
\hphantom{E_n\big( X_n; \{q_{ij} \}_{1 \le i < j \le n}\big) =}{}+
\sum_{j=1}^{n-1} q_{jn} E_{n-2}\big( X_{[1,n-1] {\setminus} \{j \}};
\{q_{a,b } \}_{1 \le a < b \le n-1 \atop a\not=j, b \not= j} \big).
\end{gather*}
\end{Lemma}
\begin{Proposition}\label{proposition2.2}
Parameters $\{g_a[b] \}$ can be expressed polynomially in
terms of quantum parameters $\{q_{ij} \}$ and variables $x_1,\ldots, x_n$,
in a such way that
\begin{gather*}
c_k(n) = e_k^{({\boldsymbol{q}})}(X_n), \qquad \forall \, k,n.
\end{gather*}
Moreover,
\begin{itemize}\itemsep=0pt
\item $g_a[b]= \sum\limits_{k=1}^{a} q_{k,a+b} \prod\limits_{j=a+1}^{a+b-1} (x_j-x_k)
+ \text{lower degree polynomials in $x_1,\ldots,x_n$}$,
\item the quantum parameters $\{ q_{ij} \}$ can be presented as rational functions in terms of variables $x_1,\ldots,x_n$ and polynomially in terms
of parameters $ \{g_a[b] \}$ such that the equality $c_k(n) = e_k^{({\boldsymbol{q}})}(X_n)$ holds for all~$k$,~$n$.
\end{itemize}
\end{Proposition}
In other words, the transformation
\begin{gather*}
\{q_{ij}\}_{1 \le i < j \le n}
\longleftrightarrow \{g_{a}[b] \}_{a+b \le n \atop a \ge 1, \, b \ge 1}
\end{gather*}
def\/ines a ``birational transformation'' between the algebra
$\mathbb{Z}[{\boldsymbol{g}}^{(n)}] [X_n] / \langle P_n(X_n,t)-t^n \rangle $ and
multiparameter quantum deformation of the algebra $H^{*}({\cal{F}}l_n,\mathbb{Z})$.
\begin{Example}\label{example2.2}
Clearly,
\begin{gather*}
g_{n-1}[1]=\sum_{j=1}^{n-1} q_{j,n}, \quad n \ge 2 \qquad \text{and} \qquad g_{n-2}[2]= \sum_{j=1}^{n-2} q_{jn} (x_{n-1}-x_{j}),
\quad n \ge 3.
\end{gather*}
Moreover
\begin{gather*}
g_1[3]= q_{14} \big((x_2-x_1)(x_3-x_1)+q_{23}-q_{12} \big)+ q_{24}
\big(q_{13}-q_{12} \big), \\
g_2[3]= q_{15} \big((x_3-x_1)(x_4-x_1)+q_{24}+q_{34}-q_{12}-q_{13} \big)\\
\hphantom{g_1[3]=}{} +
q_{25} \big((x_3-x_2)(x_4-x_2) +q_{14}+q_{34}-q_{12}-q_{23} \big) +
q_{35} \big(q_{14}+q_{24}-q_{13}-q_{23} \big).
\end{gather*}
\end{Example}
\begin{Comments}\label{comments2.5}
The full Kostant--Toda lattice (FKTL for short) has been introduced in the
end of $70's$ of the last century by B.~Kostant and since that
time has been extensively studied both in Mathematical and Physical literature.
We refer the reader to the original paper by B.~Kostant~\cite{FTL+,FTL} for the def\/inition of the ${\rm FKTL}$ and its basic
properties. In the present paper we just want to point out on a connection of
the Fulton universal ring and hence the multiparameter deformation of the
cohomology ring of complete f\/lag varieties, and polynomial integral of motion
of the FKTL. Namely,
\begin{center}
\framebox{\parbox[t]{4in}{Polynomials $c_k(n;X_n,{\boldsymbol{g}}^{(n)})$ def\/ined by
\eqref{equation2.20} coincide with the polynomial integrals of motion of the FKTL.}}
\end{center}
It seems an interesting task to clarify a meaning of the ${\rm FKTL}$ rational
integrals of motion in the context of the universal Schubert calculus~\cite{Fu} and the algebra $3HT_n(0)$, as well as any meaning of universal
Schubert or Grothendieck polynomials in the context of the Toda or full
Kostant--Toda lattices.
\end{Comments}
\section[Algebra $3HT_n$]{Algebra $\boldsymbol{3HT_n}$}\label{section3}
Consider the twisted classical Yang--Baxter relation
\begin{gather*}
[u_{ij}+u_{ik},u_{jk}]+[u_{ik},u_{ji}]=0,
\end{gather*}
where $i$, $j$, $k$ are distinct.
Having in mind applications of the Dunkl elements to combinatorics and
algebraic geometry, we split the above relation into two relations
\begin{gather}\label{equation3.1}
u_{ij}u_{jk}=u_{jk}u_{ik}-u_{ik}u_{ji} \qquad \text{and} \qquad
u_{jk}u_{ij}=u_{ik}u_{jk}-u_{ji}u_{ik}
\end{gather}
and impose the following {\it unitarity} constraints
\begin{gather*}
u_{ij}+u_{ji}=\beta,
\end{gather*}
where $\beta$ is a central element.
Summarizing, we come to the following def\/inition.
\begin{Definition} \label{definition3.1}
Def\/ine algebra $3T_n(\beta)$ to be the quotient of the free
associative algebra
\begin{gather*}
\mathbb{Z}[\beta] \langle u_{ij},\, 1 \le i < j \le n
\rangle
\end{gather*}
by the set of relations
\begin{itemize}\itemsep=0pt
\item {\it locality}: $u_{ij} u_{kl}=u_{kl}u_{ij}$ if $\{i,j\}
\cap \{k,l\}=\varnothing$,
\item {\it $3$-term relations}:
$u_{ij} u_{jk}=u_{ik} u_{ij}+u_{jk} u_{ik}- \beta u_{ik}$, and
$u_{jk} u_{ij}=u_{ij} u_{ik}+u_{ik} u_{jk}-\beta u_{ik}$
if~$1 \le i < j < k \le n$.
\end{itemize}
\end{Definition}
It is clear that the elements $\{ u_{ij},\, u_{jk},\, u_{ik},\, 1 \le i < j <k
\le n \}$ satisfy the classical Yang--Baxter relations, and therefore, the
elements $\big\{\theta_{i}:=\sum\limits_{j \not= i} u_{ij}, \,1=1,\ldots,n \big\}$ form a~mutually commuting set of elements in the algebra $3T_n(\beta)$.
\begin{Definition}\label{definition3.2}
We will call $ \theta_1,\ldots,\theta_n$ by the
(universal) additive Dunkl elements.
\end{Definition}
For each pair of indices $i < j$, we def\/ine element $q_{ij}:=u_{ij}^{2} - \beta u_{ij} \in 3T_n(\beta)$.
\begin{Lemma} \label{lemma3.1} \quad
\begin{enumerate}\itemsep=0pt
\item[$1.$] The elements $\{ q_{ij}, \, 1 \le i < j \le n \}$ satisfy the Kohno--Drinfeld relations
$($known also as the horizontal four term relations$)$
\begin{gather*}
q_{ij} q_{kl}=q_{kl} q_{ij} \qquad \text{if} \ \ \{i,j\} \cap \{k,l\} = \varnothing, \\
[q_{ij},q_{ik}+q_{jk}]=0,\qquad [q_{ij}+q_{ik},q_{jk}]=0 \qquad \text{if} \ \ i < j < k.
\end{gather*}
\item[$2.$] For a triple $(i < j < k)$ define $u_{ijk}:= u_{ij}-u_{ik}+u_{jk}$.
Then
\begin{gather*}
u_{ijk}^2= \beta u_{ijk}+q_{ij}+q_{ik}+q_{jk}.
\end{gather*}
\item[$3.$] Deviation from the Yang--Baxter and Coxeter relations:
\begin{gather*}
u_{ij} u_{ik} u_{jk}-u_{jk} u_{ik} u_{ij}=[u_{ik},q_{ij}]=
[q_{jk},u_{ik}],\\
u_{ij}u_{jk}u_{ij}-u_{jk}u_{ij}u_{jk}= q_{ij}u_{ik}-u_{ik}q_{jk}.
\end{gather*}
\end{enumerate}
\end{Lemma}
\begin{Comments} \label{comments3.1}
It is easy to see that the horizontal 4-term relations
listed in Lemma~\ref{lemma3.1}(1), are consequences of the locality conditions among
the generators $\{q_{ij}\}$, together with the commutativity conditions
among the Jucys--Murphy elements
\begin{gather*}
d_i:= \sum_{j=i+1}^{n} q_{ij},\qquad i=2,\ldots,n,
\end{gather*}
namely, $[d_i,d_j]=0$. In \cite{K} we describe some properties of a
commutative subalgebra generated by the Jucys--Murphy elements in the (nil\footnote{That is the quotient of the Kohno--Drinfeld algebra generated by the elements $\{q_{ij}\}$ by the two-sided ideal generated by the elements $\{q_{ij}^2\}_{1 \le i,j \le n}$.})
Kohno--Drinfeld algebra. It is well-known that the Jucys--Murphy elements
generate a maximal commutative subalgebra in the group ring of the symmetric
group ${\mathbb S}_n$. It is an open {\it problem}
\begin{center}
\framebox{\parbox[t]{4 in}{describe def\/ining
relations among the Jucys--Murphy elements in the group ring
$\mathbb{Z}[{\mathbb S}_n]$. }}
\end{center}
\end{Comments}
Finally we introduce the ``Hecke quotient'' of the algebra $3T_n(\beta)$,
denoted by $3HT_n(\beta)$.
\begin{Definition} \label{definition3.3}
Def\/ine algebra $3HT_n(\beta)$ to be the quotient of the algebra
$3T_n(\beta)$ by the set of relations
\begin{gather*}
q_{ij} q_{kl}=q_{kl} q_{ij} \qquad \text{for all} \ \ i,\,j,\,k,\,l.
\end{gather*}
\end{Definition}
In other words we assume that the all elements $\{q_{ij},\, 1 \le i < j \le n \}$ are {\it central} in the algebra $3T_n(\beta)$.
From Lemma~\ref{lemma3.1} follows immediately that in the algebra $3HT_n(\beta)$ the
elements $\{u_{ij} \}$ satisfy the multiplicative (or quantum) Yang--Baxter
relations
\begin{gather}\label{equation3.2}
u_{ij} u_{ik} u_{jk}=u_{jk} u_{ik} u_{ij} \qquad \text{if} \ \ i < j < k.
\end{gather}
To underline the dependence of the algebra $3HT_n(\beta)$ on the central
elements ${\boldsymbol{q}}:=\{q_{ij}\}$, we will use for the former the notation
$3T_n^{({\boldsymbol{q}})}(\beta)$ as well.
\begin{Exercises}[some relations in the algebra $3T_n^{({\boldsymbol{q}})}(\beta)$] \label{Exercises3.1} \quad
\begin{enumerate}\itemsep=0pt
\item[1.] Noncommutative analogue of recurrence relation among the Catalan
numbers \cite{K2, K}, cf.\ Section~\ref{section5.1}.
Let $k$, $n$ be positive integers, $ k < n$ and $i_1,\ldots,i_k$,
$1 \le i_{k} < n $, be a collection of pairwise distinct integers. {\it Prove} the following identity in the algebra $3T_n^{({\boldsymbol{q}})}(\beta)$\footnote{{\it Hint}: denote the r.h.s.\ of of the identity stated in item~(1)
by~$R_{I}$. One possible proof is based on induction and examination of the
element $R_{ I \cup \{i_{k+2} \}}:= u_{i_{a_{1},i_{a+2}}} R_{I} -R_{I} u_{i_{a+1},i_{a_{i+2}}}$.}
\begin{gather*}
\prod_{a=1}^{k} u_{i_{a},i_{a+1}} +\sum_{r=2}^{k+1} \left( \prod_{a=r}^{n}
\beta (u_{i_{a},i_{a+1}}) u_{i_{a_{k+1}},i_{a_{1}}} \left( \prod_{a=1}^{r-2}
u_{i_{a},i_{a+1}}\right)\right)\\
\qquad{} = \beta \prod_{a=1}^{k} u_{i_{a},i_{a+1}} -
\beta \big( u_{i_{1},i_{k+1}} R_{ I{\setminus} \{i_{k+1\}}} - R_{I{\setminus}\{i_{k+1}\}} u_{i_{k},i_{k+1}} \big),
\end{gather*}
where $R_{I}$ denotes the r.h.s.\ of the above identity.
For example,
\begin{gather*}
12\,23 +23\,31 + 31\,12 = \beta (12-13+23),\\
12\,23\,34 + 23\,34\,41+ 34\,41\,12 + 41\,12\,23\\
\qquad {} = \beta ( 12\,23 - 14(12 -13+23) + (12-13 +23) 34 ),
\end{gather*}
where we use short notation $ij:=u_{ij}$.
See Introduction, {\it summation formula},~{\bf A}, for an interpretation of
the above formula in the case $\beta=0$, $q_{ij}=0$, $\forall\, i,j$.
{\it Note} that the above formula does not depend on deformation
(or quantum) parameters $\{q_{ij}\}$, in particular it also true for the
algebras $3T(\Gamma)$ associated with a simple graph~$\Gamma$, and gives rise to quantum as well as $K$-theoretic deformations of the Orlik--Terao algebra
of a~simple graph, cf.~\cite{Li}.
\item[2.] Cyclic relations, cf.~\cite{FK}.
Let $i_1,i_2,\ldots,i_k$, $1 \le i_a \le n$ be a collection of pairwise
distinct integers.
{\it Show} that
\begin{gather*}
\sum_{r=1}^{k-1} \left(\prod_{a=r+1}^{k} u_{i_{1},i_{a}} \right) \left(\prod_{a=2}^{r} u_{i_{1},i_{a}} \right) u_{i_{r+1},i_{1}}
= - \left( \sum_{a=2}^{k}
q_{i_{1},i_{a}} \left(\prod_{b =a+1}^{k } u_{i_{a},i_{b}} \right) \left( \prod_{b=2}^{a-1} u_{i_{a},i_{b}} \right) \right).
\end{gather*}
For example, $12\,13\,14\,21 + 13\,14\,12\,31+14\,12\,13\,41= -q_{12}\,23\,24 -q_{13}\,34\,32 - q_{14}\,42\,43$.
\end{enumerate}
Note that the r.h.s.\ does not depend on parameter $\beta$.
\end{Exercises}
\subsection[Modif\/ied three term relations algebra $3MT_n(\beta,\psi)$]{Modif\/ied three term relations algebra $\boldsymbol{3MT_n(\beta,\psi)}$}\label{section3.1}
Let $\beta$, $\{q_{ij}=q_{ji},\,\psi_{ij}=\psi_{ji},\, 1 \le i, j \le n \}$,
be a set of mutually commuting elements.
\begin{Definition}\label{definition3.7}
Modif\/ied $3$-term relation algebra $3MT_n(\beta, {\boldsymbol{q}}, \psi)$
is an associative algebra over the ring of polynomials
$\mathbb{Z}[ \beta, q_{ij},\psi_{ij}]$ with the set of generators
$\{ u_{ij},\, 1 \le i , j \le n \}$ subject to the set of relations
\begin{itemize}\itemsep=0pt
\item $u_{ij}+u_{ji}= \beta$, $u_{ij}u_{kl}=u_{kl}u_{ij}$
if~$\{i,j \}
\cap \{k,l \} =\varnothing$,
\item {\it three term relations}:
\begin{gather*}
u_{ij} u_{jk}+u_{ki} u_{ij}+u_{jk} u_{ki}= \beta (u_{ij}+u_{ik}+u_{jk}) \qquad \text{if} \ \ i,\,j,\,k \ \ \text{are distinct},
\end{gather*}
\item $u_{ij}^{2} =\beta u_{uj}+ q_{ij}+\psi_{ij} $ if $i \not= j$,
\item $u_{ij} \psi_{kl}= \psi_{kl} u_{ij}$ if $\{i,j\} \cap \{k,l \} =
\varnothing$,
\item {\it exchange relations}: $u_{ij} \psi_{jk}= \psi_{ik} u_{ij}$
if $i$, $j$, $k$ are distinct,
\item elements $\beta$, $\{ q_{ij}, \, 1 \le i,j \le n \}$ are
{\it central}.
\end{itemize}
\end{Definition}
It is easy to see that in the algebra $3MT_n(\beta, {\boldsymbol{q}},{\boldsymbol{\psi}})$ the
generators $\{u_{ij} \}$
satisfy the modif\/ied {\it Coxeter} and modif\/ied {\it quantum Yang--Baxter
relations}, namely
\begin{itemize}\itemsep=0pt
\item {\it modified Coxeter relations}: $u_{ij}u_{jk}u_{ij}-u_{jk}u_{ij}u_{jk}= (q_{ij}-q_{jk})u_{ik}$,
\item {\it modified quantum Yang--Baxter relations}:
\begin{gather*}
u_{ij} u_{ik} u_{jk}-u_{jk} u_{ik} u_{ij} = (\psi_{jk}- \psi_{ij}) u_{ik}
\end{gather*}
if $i$, $j$, $k$ are distinct.
\end{itemize}
Clearly the additive Dunkl elements $\big\{\theta_i:=\sum\limits_{j \not=i} u_{ij},\, i=1,
\ldots,n \big\}$ generate a commutative subalgebra in $3MT_n(\beta,\psi)$.
It is still possible to describe relations among the additive Dunkl elements~\cite{K}, cf.~\cite{KM2}. However we don't know any geometric interpretation
of the commutative algebra obtained. It is not unlikely that this commutative
subalgebra is a common generalization of the small quantum cohomology and
elliptic cohomology (remains to be def\/ined!) of complete f\/lag varieties.
The algebra $3MT_n(\beta=0,{\boldsymbol{q}}={\boldsymbol{0}},\psi)$ has an elliptic representation \cite{K,KM2}. Namely,
\begin{gather*}
u_{ij}:= \sigma_{\lambda_i -\lambda_j}(z_i-z_j) s_{ij}, \qquad q_{ij} =
\wp(\lambda_i -\lambda_j), \qquad \psi_{ij}= - \wp(z_i-z_j),
\end{gather*}
where $\{ \lambda_i, \, i=1,\ldots,n \}$ is a set of parameters (e.g., complex
numbers), and $\{z_1,\ldots,z_n \}$ is a set of variables;
$s_{ij}$, $i < j$, denotes the transposition that swaps~$i$ on~$j$ and f\/ixes
all other variables;
\begin{gather*}
\sigma_{\lambda}(z):= \frac{\theta(z-\lambda) \theta'(0)}{\theta(z)\theta(\lambda)}
\end{gather*}
denotes the {\it Kronecker sigma function}; $\wp(z)$ denotes the {\it
Weierstrass $P$-function}.
{\bf ``Multiplicative'' version of the elliptic representation.}
Let $q$ be parameter. In this place we will use the same symbol $\theta(x)$ to denote the ``multiplicative'' version of the Riemann theta function
\begin{gather*}
\theta(x):=\theta(x;q)= (x;q)_{\infty} (q/x;q)_{\infty},
\end{gather*}
where by def\/inition $ (x;q)_{\infty}= (x)_{\infty}=\prod\limits_{k \ge 0}(1-x~q^{k})$.
Let us state some well-known properties of the Riemann theta function:
\begin{itemize}\itemsep=0pt
\item $\theta(qx;q)=\theta(1/x;q)=-x^{-1} \theta(x;q)$,
\item functional equation:
\begin{gather*}
x/y \theta\big(u x^{\pm 1}\big) \theta\big(y v^{\pm 1}\big) + \theta\big(u v^{\pm 1}\big)
\theta\big(x y^{\pm 1}\big)=\theta\big(u y^{\pm 1}\big) \theta\big(x v^{\pm 1}\big),
\end{gather*}
where by def\/inition $\theta(x y^{\pm 1}):= \theta(x y) \theta(x y^{-1})$.
\item Jacobi triple product identity:
\begin{gather*}
(q;q)_{\infty} \theta(x;q) = \sum_{n \in \mathbb{Z}} (-x)^{n} q^{{n \choose 2}}.
\end{gather*}
\end{itemize}
One can easily check that after the change of variables
\begin{gather*}
x:= \left(\frac{z^2}{\lambda w}\right)^{1/2}, \qquad y:=\left(\frac{w}{\lambda}\right)^{1/2}, \qquad
u:=\left(\frac{w}{\lambda \mu^2}\right)^{1/2}, \qquad v:=(w \lambda)^{1/2},
\end{gather*}
the functional equation for the Riemann theta function $\theta(x)$ takes the following form
\begin{gather*}
\sigma_{\lambda}(z)~\sigma{\mu}(w) = \sigma_{\lambda \mu}(z)\sigma_{\mu}(w/z)
+ \sigma_{\lambda \mu}(w) \sigma_{\lambda}(z/w),
\end{gather*}
where
\begin{gather*}
\sigma_{\lambda}(z):=\frac{\theta(z/ \lambda)}{\theta(z) \theta\big(\lambda^{-1}\big)}
\end{gather*}
denotes the (multiplicative) {\it Kronecker sigma function}.
Therefore, the operators
\begin{gather*}
u_{ij}(f):= \sigma_{\lambda_i/\lambda_j}(z_i/z_j) s_{ij}(f),
\end{gather*}
where $s_{ij}$ denotes the exchange operator which swaps the variables $z_i$
and $z_{j} $, namely $s_{ij}(z_i)=z_j$, $s_{ij}(z_j)=z_i$, $s_{ij}(z_k)=z_k$,
$\forall\, k \not= i,j$, and $s_{ij}$ acts trivially on dynamical parameters
$\lambda_i$, namely, $s_{ij}(\lambda_k)=\lambda_k$, $\forall\, k$, {\it give rise} to a representation of the algebra $3MT_n(\beta=0,{\boldsymbol{q}}={\boldsymbol{0}},\psi)$.
The $3$-term relations among the elements $\{u_{ij} \}$ are consequence (in
fact equivalent) to the famous {\it Jacobi--Riemann} $3$-term relation of
degree~$4$ among the theta function $\theta(z)$, see, e.g., \cite[p.~451,
Example~5]{WW}. In several cases, see Introduction, relations~({\bf A}) and~{(\bf B}), identities among the Riemann theta functions can be rewritten in
terms of the elliptic Kronecker sigma functions and turn out to be a~consequence of certain relations in the algebra $3MT_n(\beta=0,{\boldsymbol{q}}={\boldsymbol{0}},\psi)$ for some
integer~$n$, and vice versa\footnote{It is commonly believed that any identity between the Riemann
theta functions is a consequence of the Jacobi--Riemann three term relations
among the former. However we do not expect that the all hypergeometric type
identities among the Riemann theta functions can be obtained from certain
relations in the algebra $3MT_n(\beta=0$, ${\boldsymbol{q}}={\boldsymbol{0}},\psi)$ after applying the {\it elliptic
representation} of the latter.}.
The algebra $3HT_n(\beta)$ is the quotient of algebra $3MT_n(\beta,{\boldsymbol{q}},\psi)$ by the two-sided ideal ge\-ne\-ra\-ted by the elements $\{ \psi_{ij} \}$.
Therefore the
elements $\{u_{ij} \}$ of the algebra $3HT_n(\beta)$ satisfy the quantum Yang--Baxter relations $u_{ij}u_{ik}u_{jk}=u_{jk}u_{ik}u_{ij}$, $i <j < k$,
and as a consequence, the multiplicative Dunkl elements
\begin{gather*}
\Theta_i =\prod_{a=i-1}^{1} (1+h u_{a,i})^{-1} \prod_{a=i+1}^{n} (1+h
u_{i,a}),\qquad i=1,\ldots,n, \qquad u_{0,i}=u_{i,n+1}=0
\end{gather*}
generate a commutative subalgebra in the algebra $3HT_n(\beta)$, see
Section~\ref{section3.1}. We emphasize that the Dunkl elements $\Theta_j$, $j=1,\ldots, n$,
do not pairwise commute in the algebra $3MT_n(\beta,{\boldsymbol{q}}, {\boldsymbol{\psi}})$, if
$\psi_{ij} \not= 0$ for some $i \not= j$. One way to construct a~multiplicative analog of additive Dunkl elements $\theta_i:=\sum\limits_{j \not=i}
u_{ij}$ is to add a new set of mutually commuting generators denoted by
$\{\rho_{ij}, \,\rho_{ij}+\rho_{ji}=0$, $1 \le i \not= j \le n\}$ subject
to the crossing relations
\begin{itemize}\itemsep=0pt
\item $\rho_{ij}$ commutes with $\beta$, $q_{kl}$ and $\psi_{k,l}$ for
all $i$, $j$, $k$, $l$,
\item $\rho_{ij} u_{kl}=u_{kl} \rho_{ij}$ if $\{i,j\} \cap \{k,l \} =
\varnothing$, $\rho_{ij}u_{jk}=u_{jk}\rho_{ik}$ if $i$, $j$, $k$ are distinct,
\item $\rho_{ij}^2 - \beta \rho_{ij} +\psi_{ij}= \rho_{jk}^2-\beta \rho_{jk}+ \psi_{jk}$ for all triples $1 \le i < j < k \le n$.
\end{itemize}
Under these assumptions one can check that elements
\begin{gather*}
R_{ij}:= \rho_{ij}+ u_{ij},\qquad 1 \le i < j \le n
\end{gather*}
satisfy the quantum Yang--Baxter relations
\begin{gather*}
R_{ij} R_{ik} R_{jk} =R_{jk} R_{ik} R_{ij},\qquad i < j < k.
\end{gather*}
In the case of {\it elliptic representation} def\/ined above, one can take
\begin{gather*}
\rho_{ij}:= \sigma_{\mu}(z_i-z_j),
\end{gather*}
where $\mu \in \mathbb{C}^{*}$ is a parameter. This solution to the quantum Yang--Baxter equation has been discovered in~\cite{SU+}. It can be seen as an
operator form of the famous (f\/inite-dimensional) solution to~${\rm QYBE}$ due to A.~Belavin and V.~Drinfeld~\cite{BD}. One can go to one step more and add
to the algebra in question a new set of generators corresponding to the shift
operators
$T_{i,q}\colon z_i \longrightarrow q z_i$, cf.~\cite{Fe}. In this case one
can def\/ine {\it multiplicative Dunkl elements} which are closely related
with the elliptic Ruijsenaars--Schneider--Macdonald operators.
\subsubsection{Equivariant modif\/ied three term relations algebra}\label{section3.1.1}
Let ${\boldsymbol{h}}= (h_2,\ldots,h_n)$ be a set of parameters. We def\/ine {\it equivariant modified $3$-term relations
algebra $3EMT_n(\beta,h,{\boldsymbol{q}},\psi)$} to be the extension of the algebra
$3TM_n(\beta,{\boldsymbol{q}},\psi)$ by the set of mutually commuting generators
$\{y_1,\ldots,y_n \}$ subject to the crossing relations
\begin{itemize}\itemsep=0pt
\item $y_i u_{jk}=u_{jk}y_i$ if $ i \not= j,k$, $y_i u_{ij} =u_{ij} y_j +h_{j}$, $y_j u_{ij}= u_{ij} y_i -h_{j}$, $ i < j$,
\item $ [y_k, q_{ij}] = 0 = [y_k, \psi_{ij}]$ for all $i$, $j$, $k$.
\end{itemize}
It is clear that the additive Dunkl elements $ \theta_i = y_i +\sum\limits_{j \not= i} u_{ij}$, $i=1,\ldots,n$, are pair-wise commute.
For simplicity's sake, we
shall restrict our consideration to the case $\beta =0$.
\begin{Theorem}[generalized Pieri's rule, cf.~\cite{K,KM2, P}] \label{theorem3.1}
Let $ 1 \le m \le n$, then
\begin{gather*}
e_k^{({\boldsymbol{h}}, {\boldsymbol{q}})}\big(\theta_1^{(n)},\ldots, \theta_{m}^{(n)}\big):=
\sum_{A \subset [1,m], \, |A|=2r \atop B \subset [1,m] {\setminus} A, |B|=2s}
H_{r}(A) M_B(\{q_{ij} \}) e_{k-2r-2s}(\Theta_{[1,m] {\setminus} (A \cup B)}) \\
\qquad{}=
\sum_{A \subset [1,m]} Y_{A} \sum_{B \subset [1,m] {\setminus} A \atop |B|=
2s} (-1)^{s} M_{B}(\{\psi_{ij}\}) \sum_{I \subset [1,n] {\setminus} A,\, I \cap B = \varnothing \atop |A|+|B|+|I|=k}
\prod_{{(i_{\alpha},j_{\alpha}) \in I \times I \atop 1 \le i_{\alpha} \le m < j_{\alpha} \le n, \, \forall \alpha} \atop
i_1,\ldots,i_{I} \, \text{are distinct}} u_{i_{\alpha},j_{\alpha}},
\end{gather*}
where for any subset $C \subset [1,n]$ we put $Y_{C}:=\prod\limits_{c \in C} y_{c}$, and $e_{\ell}(\Theta_C)=e_m(\{ \theta_c\}_{c \in C})$ stands for the degree~$\ell$ elementary symmetric polynomial of the elements
$\{\theta_c\}_{c \in C}$,
$e_{k}(\{\theta_c\}_{c \in C}) = \delta_{0,k}$ if $k \le 0$;
if $B \subset [1,n]$, $|B|=2s$, we set
\begin{gather*}
M_B(\{\psi_{ij}) =\prod_{{ L \subset B, \, |L|=s \atop (i_1,\ldots,i_s) \subset L} \atop (j_1,\ldots,j_s) \subset B \setminus L, \, i_{\alpha} < j_{\alpha}, j_{\alpha} \in B {\setminus} L, \, i_{\alpha} < n\, \forall \, \alpha} \psi_{i_{\alpha},j_{\alpha}};
\end{gather*}
in a similar manner one can define $M_B(\{q_{ij} \})$; finally we set
\begin{gather*}
H_{r}(A)= h_{a_{2r}} \Biggl( \sum_{(a_1,\ldots,a_{r-1}) \subset A {\setminus}
\{a_{2r} \}} \prod_{j=1}^{r-1} \max(a_j-2j+1,0)~h_{a_{j}} \Biggr).
\end{gather*}
\end{Theorem}
It is not dif\/f\/icult to {\it show} that
\begin{gather*}
H_{r}(A) \big |_{h_{a}=1, \, a \in A} = (2 r -1) !! ,
\end{gather*}
as well as the number of dif\/ferent monomials which appear in $H_r([1,2r])$ is
equal to the Catalan number ${\rm Cat}_{r}$. For example,
\begin{gather*}
H_{3}([1,6])=h_{6}
(h_{24} +2 h_{25}+ 2 h_{34}+4 h_{35}+6 h_{45}),\\
H_4([1,8])=h_{8}\big(h_{246} +2 h_{247}+ 2 h_{256}+4 h_{257}+6 h_{267} +
2 h_{346}+ 4 h_{347}+ 4 h_{356} \\
\hphantom{H_4([1,8])=}{} +8 h_{357} +12 h_{367} + 6 h_{456}+ 12 h_{457} + 18 h_{467} +24 h_{567}\big).
\end{gather*}
\begin{Exercise}\label{exercise3.1}
Write
\begin{gather*}
H_{r}([1,2r])= h_{a_{2r}} \Bigg( \sum_{A:=(a_1,\ldots,a_{r-1}) \subset [1,2r-1] \atop a_j \ge 2 j} c_{A}^{(r)} h_{A} \Bigg),
\end{gather*}
where $c_{A}^{(r)}:= \prod\limits_{j=1}^{r-1} \max(a_{j}-2j+1,0)$ and $h_{A}:= \prod\limits_{a \in A} h_{a}$. {\it Show} that
\begin{gather}
\sum_{A:=(a_1,\ldots,a_{r-1}) \subset [1,2r-1] \atop a_{j} \ge 2 j}
\big( c_{A}^{(r)} \big)^2 = E_{r},\label{****}
\end{gather}
where $E_r$ denotes the $r$-th Euler number, see, e.g., \cite[$A000364$]{SL}.
{\it Find} representation theoretic interpretation of numbers
$\{ c_{A}^{(r)} \}$ and the identity \eqref{****}.
\end{Exercise}
Clearly,
\begin{gather*}
\sum_{A:=(a_1,\ldots,a_{r-1}) \subset [1,2r-1] \atop a_{j} \ge 2 j}
c_{A}^{(r)} = (2r-1) !! .
\end{gather*}
\begin{Question} \label{question3.1} Does there exist a semisimple algebra ${\mathfrak{A}}(r)$,
$\dim ({\mathfrak{A}}(r))=E_r$ such that the all irreducible representations
$\pi_{A}^{(r)}$ of the algebra ${\mathfrak{A}}(r)$ are in one-to-one
correspondence with the set $ {\cal{P}}(r):= \{ A=(a_1,\ldots,a_{r-1}) \subset [1,2r-1], \, a_{j} \ge 2 j, \, \forall\, j\}$ and $\dim (\pi_{A}) =
c_{A}^{(r)}$, $\forall \, A \in {\cal{P}}(r)$?
\end{Question}
It is worth noting that the Dunkl element $\theta_i$, $1 \le i \le n$, doesn't
commute either with~$y_j$, $j \not=i$ or any~$\psi_{kl}$. On the other hand
one can check easily that $[e_k(y_1,\ldots,y_n), \theta_i]= 0$, $\forall\, k,i$.
\subsection{Multiplicative Dunkl elements}\label{section3.2}
Since the elements $u_{ij}$, $u_{ik}$ and $u_{jk}$, $i < j <k$, satisfy the
classical and {\it quantum} Yang--Baxter relations~\eqref{equation3.1} and~\eqref{equation3.2},
one can def\/ine a multiplicative analogue denoted by~$\Theta_i$, $1 \le i
\le n$, of the Dunkl elements~$\theta_i$. Namely, to start with, we def\/ine
elements
\begin{gather*}
h_{ij}:= h_{ij}(t)= 1+t u_{ij}, \qquad i \not= j.
\end{gather*}
We consider $h_{ij}(t)$ as an element of the algebra $\widetilde {3HT_n} :=
3HT_n(\beta) \otimes \mathbb{Z}[[ q_{ij}^{\pm 1},t,x,y, \ldots]]$, where we assume
that the all parameters $\{ q_{ij},t,x,y, \ldots \}$ are
{\it central} in the algebra $\widetilde {3HT_n}$.
\begin{Lemma} \label{lemma3.2}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1a)$] $h_{ij}(x) h_{ij}(y)=h_{ij}(x+y+\beta xy)+q_{ij} xy$,
\item[$(1b)$] $h_{ij}(x) h_{ji}(y)=h_{ij}(x-y)+\beta y-q_{ij} x y$ if $i < j$.
\end{enumerate}
It follows from $(1b)$ that $h_{ij}(t) h_{ji}(t) =1+\beta t-t^2 q_{ij}$ if~$i <j$,
and therefore the elements $ \{h_{ij} \}$ are invertible in the algebra
$\widetilde{3HT_n}$.
\begin{enumerate}\itemsep=0pt
\item[$(2)$] $h_{ij}(x) h_{jk}(y)=h_{jk}(y) h_{ik}(x)+h_{ik}(y) h_{ij}(x)-
h_{ik}(x+y+\beta xy)$,
\item[$(3)$] multiplicative Yang--Baxter relations:
\begin{gather*}
h_{ij} h_{ik} h_{jk}=h_{jk} h_{ik} h_{ij} \qquad \text{if} \ \ i < j < k,
\end{gather*}
\item[$(4)$] define multiplicative Dunkl elements $($in the algebra $\widetilde{3HT_n})$ as follows
\begin{gather*
\Theta_j:=\Theta_j(t)= \left( \prod_{a=j-1}^{1} h_{aj}^{-1} \right) \left( \prod_{a=n}^{j+1} h_{ja} \right), \qquad 1 \le j \le n.
\end{gather*}
\end{enumerate}
Then the multiplicative Dunkl elements pair-wise commute.
\end{Lemma}
Clearly
\begin{gather*}
\prod_{j=1}^{n} \Theta_j =1, \qquad \Theta_j=1+t \theta_j+t^2( \cdots) \qquad \text{and} \qquad
\Theta_{I}\prod_{i \notin I, \, j \in I \atop i < j}\big(1+t \beta -t^2 q_{ij}\big)
\in 3HT_n(\beta).
\end{gather*}
Here for a subset $I \subset [1,n]$ we use notation $\Theta_{I}= \prod\limits_{a \in I} \Theta_{a}$. Note, that the element $\Theta_{I}$ is a~product of
(exactly!) $k(n-k)$ terms of a form $h_{i_{\alpha}j_{\alpha}}$, where $k:= |I|$.
Our main result of this section is a~description of relations among the
multiplicative Dunkl elements.
\begin{Theorem}[A.N.~Kirillov and T.~Maeno~\cite{KM}]\label{theorem3.2}
In the algebra $3HT_n(\beta)$ the following relations hold true
\begin{gather*}
\sum_{I \subset [1,n] \atop |I|=k} \Theta_{I} \prod_{i \notin I, \, j \in J
\atop i < j} \big(1+t \beta-t^2 q_{ij}\big)= {n \brack k}_{1+t \beta}.
\end{gather*}
Here ${n \brack k}_{q}$ denotes the $q$-Gaussian
polynomial.
\end{Theorem}
\begin{Corollary}\label{corollary3.1}
Assume that $q_{ij} \not= 0$ for all $1 \le i < j \le n$. Then the all
elements $\{u_{ij} \}$ are invertible and $u_{ij}^{-1} =q_{ij}^{-1} (u_{ij}-
\beta)$. Now define elements $\Phi_i \in \widetilde{3HT_n}$ as follows
\begin{gather*}
\Phi_i= \left\{\prod_{a=i-1}^{1} u_{ai}^{-1} \right\} \left\{\prod_{a=n}^{i+1} u_{ia} \right\}, \qquad i=1,\ldots,n.
\end{gather*}
Then we have
\begin{enumerate}\itemsep=0pt
\item[$(1)$] relationship among $\Theta_j$~and~$\Phi_j$:
\begin{gather*}
t^{n-2j+1} \Theta_j\big(t^{-1}\big) \arrowvert_{t=0}=(-1)^{j} \Phi_j,
\end{gather*}
\item[$(2)$] the elements $\{ \Phi_i,\, 1 \le i \le n,\}$ generate a~commutative
subalgebra in the algebra ${\widetilde{3HT_n}}$,
\item[$(3)$] for each $k=1,\ldots,n$, the following relation in the algebra
$3HT_n$ among the elements $\{ \Phi_i \}$ holds
\begin{gather*}
\sum_{I \subset [1,n] \atop |I|=k} \prod_{i \notin I, \, j \in I \atop i < j}
(-q_{ij}) \Phi_{I} =\beta^{k(n-k)},
\end{gather*}
where $\Phi_{I}:= \prod\limits_{a \in I} \Phi_a$.
\end{enumerate}
\end{Corollary}
In fact the element $\Phi_i$ admits the following ``reduced expression'' (i.e., one with the minimal number of terms involved) which
is useful for proofs and applications
\begin{gather}\label{equation3.4}
\Phi_i= \Biggl\{ \overrightarrow{\prod_{j \in I}} \Biggl\{ \overrightarrow{\prod_{i \in I_{+}^{c} \atop i < j}}~
u_{ij}^{-1} \Biggr\} \Biggr\} \Biggl\{\overrightarrow{\prod_{j \in I_{+}^{c}}}
\Biggl\{ \overrightarrow{\prod_{i \in I \atop i < j }} u_{ij} \Biggr\} \Biggr\}.
\end{gather}
Let us explain notations. For any (totally) ordered set
$I=(i_1 < i_2 < \cdots < i_k)$ we denote by~$I_{+}$ the set~$I$ with the
opposite order, i.e., $I_{+}=(i_k > i_{k-1} > \cdots > i_1)$;
if $I \subset [1,n]$, then we set $I^{c} := [1,n]{\setminus}I$. For any
(totally) ordered set~$I$ we denote by $\overrightarrow{\prod\limits_{i \in I}}$ the
ordered product according to the order of the set~$I$.
Note that the total number of terms in the r.h.s.\ of \eqref{equation3.4} is equal to
$i(n-i)$.
Finally, from the ``reduced expression''~\eqref{equation3.4} for the element~$\Phi_i$ one
can see that
\begin{gather*}
\prod_{i \notin I, j \in I \atop i < j}(-q_{ij}) \Phi_{I}=
\Biggl\{\overrightarrow{\prod_{j \in I}} \Biggl\{ \overrightarrow{\prod_{i \in I_{+}^{c} \atop i < j}} (\beta-u_{ij}) \Biggr\} \Biggr\}
\Biggl\{\overrightarrow{
\prod_{j \in I_{+}^{c}}} \Biggl\{\overrightarrow{\prod_{i \in I \atop i < j}} u_{ij} \Biggr\} \Biggr\}:= \widetilde{\Phi_I} \in 3HT_n.
\end{gather*}
Therefore the identity
\begin{gather*}
\sum_{I \subset [1,n] \atop |I|=k} \widetilde{\Phi_{I}} = \beta^{k(n-k)}
\end{gather*}
is true in the algebra $3HT_n$ for any set of parameters $\{q_{ij} \}$.
\begin{Comments} \label{comments3.2}
In fact from our proof of Theorem~\ref{theorem3.1} we can deduce more
general statement, namely, consider integers~$m$ and~$k$~such that $1 \le k
\le m \le n$. Then
\begin{gather}\label{equation3.5}
\sum_{I \subset [1,m] \atop |I|=k} \Theta_{I}
\prod_{i \in [1,m] {\setminus} I, \, j \in J \atop i < j} \big(1+t \beta-t^2 q_{ij}\big)=
{m \brack k} _{1+t \beta} +\sum_{A \subset [1,n], \, B \subset
[1,n] \atop |A|=|B|=r} u_{A,B},
\end{gather}
where, by def\/inition, for two sets $A=(i_1,\ldots,i_r)$ and
$B=(j_1,\ldots,j_r)$ the symbol $u_{A,B}$ is equal to the (ordered) product
$\prod\limits_{a=1}^{r} u_{i_{a},j_{a}}$. Moreover, the elements of the sets~$A$ and~$B$ have to satisfy the following conditions:
\begin{itemize}\itemsep=0pt
\item for each $a=1,\ldots,r$ one has
$1 \le i_a \le m < j_a \le n$, and $ k \le r \le k(n-k)$.
\end{itemize}
Even more, if $r=k$, then sets $A$ and $B$ have to satisfy the following
additional conditions:
\begin{itemize}\itemsep=0pt
\item $B=(j_1 \le j_2 \le \cdots \le j_k)$, and the elements of the set~$A$ are pair-wise distinct.
\end{itemize}
In the case $\beta=0$ and $r=k$, i.e., in the case of additive (truncated) Dunkl
elements, the above statement, also known as the quantum Pieri formula, has
been stated as conjecture in~\cite{FK}, and has been proved later in~\cite{P}.
\begin{Corollary}[\cite{KM}]\label{corollary3.2}
In the case when $\beta=0$ and $q_{ij}=q_i \delta_{j-i,1}$, the algebra over
$\mathbb{Z}[q_1,\ldots,q_{n-1}]$ generated by the multiplicative Dunkl elements
$\{ \Theta_i \, \text{and}\, \Theta_i^{-1}, \, 1 \le i \le n \}$ is canonically isomorphic
to the quantum $K$-theory of the complete flag variety ${\cal F}l_n$ of type~$A_{n-1}$.
\end{Corollary}
It is still an open {\it problem} to describe explicitly the set of monomials
$\{ u_{A,B} \}$ which appear in the r.h.s.\ of~\eqref{equation3.5} when $r > k$.
\end{Comments}
\subsection{Truncated Gaudin operators}\label{section3.3}
Let $\{ p_{ij},\, 1 \le i \not= j \le n
\}$ be a set of mutually commuting parameters. We assume that parameters
$\{p_{ij}\}_{1\le i < j \le n}$ are invertible and satisfy the Arnold
relations
\begin{gather*}
{1 \over p_{ik}}=
{1 \over p_{ij}}+{1 \over p_{jk}}, \qquad i < j , k.
\end{gather*}
For example one can take $p_{ij}=(z_i-z_j)^{-1}$, where $z=(z_1,\ldots,z_n) \in
(\mathbb{C} {\setminus} 0)^{n}$.
\begin{Definition} \label{definition3.5} Truncated (rational) Gaudin operator corresponding to the set
of parame\-ters~$\{p_{ij}\}$ is def\/ined to be
\begin{gather*} G_i= \sum_{j \not= i} p_{ij}^{-1} s_{ij}, \qquad 1 \le i \le n,
\end{gather*}
where~$s_{ij}$ denotes the exchange operator which switches variables~$x_i$
and~$x_j$, and f\/ixes parame\-ters~$\{ p_{ij} \}$.
We consider the Gaudin operator $G_i$ as an element of the group ring
$\mathbb{Z}[\{p_{ij}^{\pm 1} \}][{\mathbb S}_n]$, call this element
$G_i \in \mathbb{Z}[\{p_{ij}^{\pm 1} \}][{\mathbb S}_n]$, $i=1,\ldots,n$, by {\it
Gaudin element} and denoted it by~$\theta_i^{(n)}$.
\end{Definition}
It is easy to see that the elements $u_{ij}:=p_{ij}^{-1} s_{ij}$, $1 \le i \not=
j \le n$, def\/ine a representation of the algebra $3HT_n(\beta)$ with
parameters $\beta=0$ and $q_{ij}=u_{ij}^{2}=p_{ij}^2$.
Therefore one can consider the (truncated) Gaudin elements as a special case
of the (truncated) Dunkl elements. Now one can rewrite the relations among
the Dunkl elements, as well as the quantum Pieri formula \cite{FK,P}, in terms of the Gaudin elements.
The key observation which allows to rewrite
the quantum Pieri formula as a certain relation among the Gaudin elements, is
the following one:
parameters $\big\{p_{ij}^{-1} \big\}$ satisfy the
Pl\"ucker relations
\begin{gather*}
{1 \over p_{ik} p_{jl}}={1 \over p_{ij} p_{kl}}+{1 \over p_{il} p_{jk}} \qquad
\text{if} \ \ i < j < k < l.
\end{gather*}
To describe relations among the Gaudin elements $\theta_i^{(n)}$, $i=1,\ldots,n$,
we need a bit of notation. Let $\{p_{ij}\}$ be a set of invertible
parameters as before,
$i_a < j_a$, $a=1,\ldots,r$.
Def\/ine polynomials in the variables ${\boldsymbol{h}}=(h_1,\ldots,h_n)$
\begin{gather}\label{equation3.6}
G_{m,k,r}^{(n)}({\boldsymbol{h}},\{ p_{ij} \})=\sum_{I \subset [1,n-1] \atop |I|=r}
{1 \over \prod\limits_{i \in I} p_{in} } \sum_{J \subset [1,n] \atop |I|+m=|J|+k} {n- | I \cup J | \choose n-m -|I|}
{\tilde h}_{J},
\end{gather}
where
\begin{gather*}
{\tilde h}_{J}=\sum_{K \subset J,\, L \subset J \atop |K|= |L|,
\, K \bigcap L = \varnothing} \prod_{j \in J {\setminus} (K \cup L)} h_{j}
\prod_{k_a \in K, \, l_a \in L} p_{k_a,l_a}^2,
\end{gather*}
and summation runs over subsets $K=\{k_1 < k_2 < \cdots < k_r \}$
and $L=\{l_1 < l_2 < \cdots < l_r \} \subset J \}$, such that
$k_a < l_a$, $a=1, \ldots,r$.
\begin{Theorem}[relations among the {\it Gaudin} elements~\cite{K},
cf.~\cite{MTV}] \label{theorem3.3} \quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] Under the assumption that elements $\{p_{ij},\, 1 \le i < j \le n \}$
are invertible, mutually commute and satisfy the Arnold relations, one has
\begin{gather}
G_{m,k,r}^{(n)}\big(\theta_1^{(n)},\ldots,\theta_n^{(n)}, \{ p_{ij} \}\big)=0 \qquad \text{if} \ \ m >k,\nonumber
\\
G_{0,0,r}^{(n)}\big(\theta_1^{(n)},\ldots,\theta_n^{(n)},\{p_{ij}\}\big)=
e_{r}(d_2, \ldots,d_n),\label{equation3.7}
\end{gather}
where $d_2,\ldots,d_n$ denote the Jucys--Murphy elements in the group ring
$\mathbb{Z}[\mathbb{S}_n]$~of the symmetric group $\mathbb{S}_n$, see Comments~{\rm \ref{comments3.1}} for
a definition of the Jucys--Murphy elements.
\item[$(2)$] Let $J= \{j_1 < j_2 < \cdots < j_r \} \subset [1,n]$, define matrix
$M_{J}:= (m_{a,b})_{1 \le a,b \le r}$, where
\begin{gather*}
m_{a,b}:= m_{a,b}({\boldsymbol{h}};\{p_{ij}\} ) =\begin{cases}
h_{j_{a}} & \text{if \ $ a=b$},\\
p_{{j_a},{j_b}} & \text{if \ $a < b$}, \\
- p_{{j_b},{j_a}} & \text{if \ $a > b$}.
\end{cases}
\end{gather*}
Then
\begin{gather*}
{\tilde{h}}_{J} =\operatorname{DET} \vert M_{J} \vert.
\end{gather*}
\end{enumerate}
\end{Theorem}
\begin{Examples} \label{examples3.1} \quad
(1) Let us display the polynomials $G_{m,k,r}^{(n)}({\boldsymbol{h}},\{ p_{ij} \})$
a few cases
\begin{gather*}
G_{m,0,r}^{(n)}({\boldsymbol{h}}, \{p_{ij} \})=\sum_{I \subset [1,n-1]
\atop |I| =r} \prod_{i \in I}p_{in}^{-1} \Biggl( \sum_{J \subset [1,n]
\atop {|J|=m+r,\, I \subset J}} {\tilde{h}}_{J} \Biggr),\\
G_{m,k,0}^{(n)}({\boldsymbol{h}}, \{p_{ij} \})= {n-m+k \choose k}
e_{m-k}^{{\boldsymbol{q}}}(h_1,\ldots,h_n),\\
G_{m,1,r}^{(n)}({\boldsymbol{h}}, \{p_{ij} \})=\sum_{I \subset [1,n-1] \atop
|I|=r} \prod_{ i \in I} p_{in}^{-1} \Biggl( \sum_{ J \subset [1,n] \atop {I
\subset J,\, |J|=m+r}}\!\! (n-m-r+1) {\tilde{h}}_{J}+
\!\!\sum_{{J \subset [1,n] \atop {|J|=m+r-1,\, |I \cup J| =m+r}}}\!\! {\tilde{h}}_{J}
\Biggr).
\end{gather*}
(2)~Let us list the relations~\eqref{equation3.7} among the Gaudin elements in
the case $n=3$. First of all, the Gaudin elements satisfy the ``standard''
relations among the Dunkl elements
$\theta_1 + \theta_2+ \theta_3 =0$, $\theta_1\theta_2+\theta_1\theta_3+
\theta_2\theta_3+q_{12}+q_{13}+q_{23}=0$,
$\theta_1\theta_2\theta_3+
q_{12} \theta_3+q_{13} \theta_2+q_{23} \theta_1=0$. Moreover, we have
additional relations which are specif\/ic for the Gaudin elements
\begin{gather*}
G_{2,0,1}^{(3)}= {1 \over p_{13}} (\theta_1 \theta_2 +\theta_1 \theta_3+q_{12}+
q_{13} )+{1 \over p_{23}} (\theta_1 \theta_2+ \theta_2 \theta_3 +q_{12}+
q_{23} ) =0,
\end{gather*}
the elements $p_{23} \theta_1+p_{13} \theta_2$ and $\theta_1 \theta_2$
are central.
\end{Examples}
It is well-known that the elementary symmetric polynomials
$e_r(d_2,\ldots,d_n):=C_r$, $r=1,\ldots,n-1$, generate the center of the
group ring $\mathbb{Z}[p_{ij}^{\pm 1}][\mathbb{S}_n]$, whereas the Gaudin elements
$\{ \theta_i^{(n)},\, i=1,\ldots,n \}$, generate a maximal commutative
subalgebra ${\cal B}(p_{ij})$, the so-called {\it Bethe subalgebra},
in $\mathbb{Z}[p_{ij}^{\pm 1}][{\mathbb S}_n]$. It is well-known, see, e.g.,~\cite{MTV}, that ${\cal B}(p_{ij})=
\bigoplus_{\lambda \vdash n} {\cal B}_{\lambda}(p_{ij})$, where
${\cal B}_{\lambda}(p_{ij})$ is
the $\lambda$-isotypic component of ${\cal B}(p_{ij})$. On each $\lambda$-isotypic component the value of the central element~$C_k$ is the explicitly
known constant~$c_k(\lambda)$.
It follows from~\cite{MTV} that the relations~\eqref{equation3.7} together with relations
\begin{gather*}
G_{0,0,r}\big(\theta_1^{(n)},\ldots,\theta_n^{(n)},\{p_{ij}\}\big)= c_r(\lambda)
\end{gather*}
are the def\/ining relations for the algebra ${\cal B}_{\lambda}(p_{ij})$.
Let us remark that in the def\/inition of the Gaudin elements we can use
{\it any} set of mutually commuting, invertible elements
$\{p_{ij} \}$ which
satisf\/ies the Arnold conditions. For example, we can take
\begin{gather*}
p_{ij}:= {q^{j-2}(1-q) \over 1-q^{j-i}}, \qquad 1 \le i < j \le n.
\end{gather*}
It is not dif\/f\/icult to see that in this case
\begin{gather*}
\lim_{q \rightarrow 0}{\theta_J^{(n)} \over p_{1j}}=-d_j=- \sum_{a=1}^{j-1} s_{aj},
\end{gather*}
where as before, $d_j$ denotes the Jucys--Murphy element in the group ring
$\mathbb{Z}[{\mathbb S}_n]$ of the symmetric group ${\mathbb S}_n$. Basically from
relations~\eqref{equation3.7} one can deduce the relations among the Jucys--Murphy
elements
$d_2,\ldots,d_n$ after plugging in~\eqref{equation3.6} the values
$p_{ij}:= {q^{j-2}(1-q) \over 1-q^{j-i}}$ and passing to the limit
$q \rightarrow 0$. However the real computations are rather involved.
Finally we note that the {\it multiplicative} Dunkl/Gaudin elements
$\{ \Theta_i,\, 1,\ldots,n \}$ also generate a maximal commutative subalgebra in
the group ring $\mathbb{Z}[p_{ij}^{\pm 1}][{\mathbb S}_n]$. Some relations among
the elements $\{\Theta_l\}$ follow from Theorem~\ref{theorem3.2}, but we don't know
an analogue of relations~\eqref{equation3.7} for the multiplicative Gaudin
elements, but see~\cite{MTV}.
\begin{Exercises}\label{exercises3.2}
Let $A = (a_{i,j})$ be a $2m \times 2m$ {\it skew-symmetric} matrix.
The {\it Pfaffian} and {\it Hafnian} of $A$ are def\/ined correspondingly by
the equations
\begin{gather*}
\mathrm{Pf}(A) = \frac{1}{2^m m!}\sum_{\sigma\in S_{2m}}\mathrm{sgn}(\sigma)\prod_{i=1}^{m}a_{\sigma(2i-1),\sigma(2i)},\\
\mathrm{Hf}(A)= \frac{1}{2^m m!}
\sum_{\sigma\in S_{2m}} \prod_{i=1}^{m}a_{\sigma(2i-1),\sigma(2i)},
\end{gather*}
where ${\mathbb S}_{2m}$ is the symmetric group and $\operatorname{sgn}(\sigma)$ is the
signature of a permutation $\sigma \in {\mathbb S}_{2m}$.\footnote{See, e.g.,
\url{https://en.wikipedia.org/wiki/Pfaffian}.}
Now let $n$ be a positive integer, and~$\{p_{ij}, \, 1 \le i \not= j \le n, \, p_{ij}+p_{ji}=0 \}$ be a set of
skew-symmetric, invertible and mutually commuting elements. We set $p_{ii}=0$
for all~$i$, and ${\boldsymbol{q}}:= \big\{ p_{ij}^2 \big\}_{1 \le i < j \le n}$.
Now let us {\it assume} that the elements $\{p_{ij} \}_{1 \le i < j \le n}$ satisfy the Pl\"{u}ker relations for the elements
$\big\{ p_{ij}^{-1} \big\}_{1 \le i < j \le n}$, namely,
\begin{gather*}
{1 \over p_{ik} p_{jl}}={1 \over p_{ij} p_{kl}}+{1 \over p_{il} p_{jk}} \qquad \text{for all} \quad 1 \le i <j < k < l \le n.
\end{gather*}
\begin{enumerate}\itemsep=0pt
\item[(a)] Let $n$ be an {\it even} positive integer. Let us def\/ine
$A_{n}(p_{ij}):=(p_{ij})_{1 \le i,j \le n}$ to be the $n \times n$
skew-symmetric matrix corresponding to the family $\{p_{ij} \}_{1 \le i < j
\le n}$.
Show that
\begin{gather*}
\operatorname{DET} | A_n(p_{ij})| = \operatorname{Hf}\big(A_n\big(p_{ij}^2\big)\big).
\end{gather*}
\item[(b)] Let $n$ be a positive integer, and $z_1,\ldots,z_n$ be a set of
mutually commuting variables, def\/ine polynomials $H_{i}(z_1,\ldots,z_n \,|\, \{p_{ij} \})$,
$i=1,\ldots,n$ from the equation
\begin{gather*}
\operatorname{DET} |\operatorname{diag}(t+z_1, \ldots,t+z_n)+A_n(p_{ij})| =t^n+\sum_{i=1}^{n}
H_{i}(z_1,\ldots,z_n \,|\, \{p_{ij} \}) t^{n-i},
\end{gather*}
where $ \operatorname{diag}(t+z_1, \ldots,t+z_n)$ means the diagonal matrix.
\end{enumerate}
Show that
for $k=1,\ldots,n$ the polynomial $H_k(z_1,\ldots,z_n \,|\, \{p_{ij} \})$ is
equal to the
multiparameter quantum elementary polynomial
$e_{k}^{({\boldsymbol{q}})}(z_1, \ldots,z_n)$, see, e.g.,~\cite{FK}, or Theorem~\ref{theorem2.1}.
For example, take $n=4$, then
\begin{gather*}
\operatorname{DET}| A(p_{ij})| = (p_{12} p_{34}-
p_{13} p_{24} +p_{14} p_{23})^2 = p_{12}^2 p_{34}^2 +
p_{13}^2 p_{24}^2 +p_{14}^2 p_{23}^2 \\
\hphantom{\operatorname{DET}| A(p_{ij})| =}{}
-2 p_{12} p_{13} p_{23} p_{14} p_{24} p_{34} \left( \frac{1}{p_{12} p_{34}} - \frac{1}{p_{13} p_{24}} +
\frac{1}{p_{14} p_{23}} \right) \\
\hphantom{\operatorname{DET}| A(p_{ij})|}{}
=
p_{12}^2 p_{34}^2 +p_{13}^2 p_{24}^2 +p_{14}^2 p_{23}^2 = \operatorname{Hf}\big(A_4\big(\big\{p_{ij}^2 \big\}\big)\big).
\end{gather*}
The last equality follows from the Pl\"{u}cker relations for parameters~$\{ p_{ij}^{-1}\}$.
On the other hand, if one assumes that a set of skew symmetric parameters
$\{r_{ij} \}_{1 \le i < j \le n}$,
$r_{ij}+r_{ji}=0$, satisf\/ies the
``standard'' Pl\"{u}cker relations, namely
\begin{gather*}
r_{ik} r_{jl}=r_{ij} r_{kl}+r_{il} r_{jk}, \qquad i < j < k < l,
\end{gather*}
then $\operatorname{DET}|A_n(r_{ij})| = 0 $.
\end{Exercises}
\subsection[Shifted Dunkl elements $\mathfrak{d}_i$ and $\mathfrak{D}_i$]{Shifted Dunkl elements $\boldsymbol{\mathfrak{d}_i}$ and $\boldsymbol{\mathfrak{D}_i}$}\label{section3.4}
As it was stated in Corollary~\ref{corollary3.2}, the {\it truncated} additive and
multiplicative Dunkl elements in the algebra $3HT_n(0)$ generate over the ring
of polynomials $\mathbb{Z}[q_1,\ldots,q_{n-1}]$ correspondingly the
{\it quantum cohomology} and {\it quantum $K$-theory}
rings of the full f\/lag variety ${\cal{F}}l_n$. In order to describe the
corresponding {\it equivariant} theories, we will introduce the {\it shifted} additive and multiplicative Dunkl elements. To start with we need at f\/irst to
introduce an extension of the algebra~$3HT_n(\beta)$.
Let $\{z_1,\ldots,z_n \}$ be a set of mutually commuting elements and
$\{\beta,{\boldsymbol{h}}=(h_2,\ldots,h_{n}), t, q_{ij}=q_{ji}$, $1 \le i,j \le n \}$
be a set of parameters. We set $h_{n}:=0$.
\begin{Definition}[cf.\ Def\/inition~\ref{definition3.1}]\label{definition3.20}
Def\/ine algebra
$\overline{3TH_n(\beta,{\boldsymbol{h}})}$ to be the semi-direct
product of the algebra $3TH_n(\beta)$ and the ring of polynomials
$\mathbb{Z}[{\boldsymbol{h}},t][z_1,\ldots,z_n]$ with respect to the crossing relations
\begin{enumerate}\itemsep=0pt
\item[(1)] $z_i u_{kl}=u_{kl} z_i$ if $i \notin \{k,l\}$,
\item[(2)] $z_i u_{ij}=u_{ij} z_j+\beta z_i+h_{j}$, $z_j u_{ij}=u_{ij} z_i-\beta z_i-h_{j}$ if $1 \le i < j < k \le n$.
\end{enumerate}
\end{Definition}
Now we set as before $h_{ij}:=h_{ij}(t)=1+t u_{ij}$.
\begin{Definition}\label{definition3.21}\quad
\begin{itemize}\itemsep=0pt
\item Def\/ine shifted additive Dunkl elements to be
\begin{gather*}
\mathfrak{d}_i= z_i + \sum_{ i < j} u_{ij} - \sum_{i > j} u_{ji}.
\end{gather*}
\item Def\/ine shifted multiplicative Dunkl elements to be
\begin{gather*}
\mathfrak{D}_i=\left( \prod_{a=i-1}^{1} h_{ai}^{-1} \right) (1+z_i)\left(
\prod_{a=n}^{i+1} h_{ia} \right).
\end{gather*}
\end{itemize}
\end{Definition}
\begin{Lemma}\label{lemma3.3}
\begin{gather*}
[\mathfrak{d}_i, \mathfrak{d}_j]=0, \qquad [\mathfrak{D}_i, \mathfrak{D}_j]=0 \qquad \text{for all} \quad i,\, j.
\end{gather*}
\end{Lemma}
Now we stated an analogue of Theorem~\ref{theorem3.1} for shifted multiplicative Dunkl
elements.
As a preliminary step, for any subset $I \subset [1,n]$ let us set
$\mathfrak{D}_{I} = \prod\limits_{a \in I} \mathfrak{D}_a$. It is clear that
\begin{gather*}
\mathfrak{D}_{I} \prod_{i \notin I,\, j \in I \atop i < j}
\big(1+t \beta-t^2 q_{ij}\big) \in \overline{3HT_n(\beta,{\boldsymbol{h}})}.
\end{gather*}
\begin{Theorem} \label{theorem3.4}
In the algebra $\overline{3HT_n(\beta,{\boldsymbol{h}})}$ the following relations hold
true
\begin{gather*}
\sum_{I \subset [1,n] \atop |I|=k} \mathfrak{D}_{I} \prod_{i \notin I, j \in J \atop i < j} \big(1+t \beta-t^2 q_{ij}\big)\\
\qquad {}=
\sum_{I \subset [1,n] \atop I=\{1 \le i_1 < \ldots < i_k \le n \}} \prod_{a= 1}^{k} (1+t \beta)^{n-k-i_{a}+a} \left(z_{i_{a}} (1+ t \beta)^{i_{a}-a} +1 + h_{i_{a}} \frac{(1+t \beta)^{i_{a}- a}-1}{\beta} \right).
\end{gather*}
\end{Theorem}
In particular, if $\beta=0$, we will have
\begin{Corollary}\label{corollary3.3} In the algebra $\overline{3HT_n(0,{\boldsymbol{h}})}$ the following
relations hold
\begin{gather}\label{equation3.9}
\sum_{I \subset [1,n] \atop |I|=k} \mathfrak{D}_{I} \prod_{i \notin I, j \in J \atop i < j} \big(1-t^2 q_{ij}\big)= \sum_{I \subset [1,n] \atop I=\{ 1 \le i_1,\ldots,i_k \le n \}} \prod_{a=1}^{k} \bigl(z_{i_{a}}+1 +t h_{i_{a}}
(i_{a}-a) \bigr).
\end{gather}
\end{Corollary}
\begin{Conjecture} \label{conjecture3.1} If $h_1= \cdots =h_{n-1}=1$, $t=1$ and $q_{ij}=\delta_{i,j+1}$,
then the subalgebra generated by multiplicative Dunkl elements
$\mathfrak{D}_i$, $i=1,\ldots,n$, in the algebra
$\overline{3HT_n(0,{\boldsymbol{h}} ={\boldsymbol{1}})}$ $($and $t=1)$, is isomorphic to the equivariant
quantum $K$-theory of the complete flag variety ${\cal{F}}l_n$.
\end{Conjecture}
Our proof is based on induction on $k$ and the following relations in
the algebra $\overline{3HT_n(\beta,{\boldsymbol{h}})}$
\begin{gather*}
h_{ji}\cdot ( 1+x_j) =h_{j}+(1 +\beta) x_j-x_i + (1+x_i) h_{ji}, \qquad
h_{ji} h_{jk} =h_{jk} h_{ki}+h_{ik} h_{ji} -1- \beta,
\end{gather*}
if $i < j < k$, and we set $h_{ij}:= h_{ij}(1)$. These relations allow to
reduce the left hand side of the relations listed in Theorem~\ref{theorem3.4} to the case
when $z_i=0$, $h_i =0$, $\forall\, i$. Under these assumptions one needs to proof
the following relations in the algebra $3HT_n(\beta)$, see Theorem~\ref{theorem3.2},
\begin{gather*
\sum_{I \subset [1,n] \atop |I|=k} \mathfrak{D}_{I} \prod_{i \notin I, \, j \in J \atop i < j} \big(1+t \beta-t^2 q_{ij}\big)= {n \brack k}_{1+t \beta}.
\end{gather*}
In the case $\beta =0$ the identity~\eqref{equation3.9} has been proved in~\cite{KM}.
One of the main steps in our proof of Theorem~\ref{theorem3.1} is the following explicit
formula for the elements~$\mathfrak{D}_{I}$.
\begin{Lemma}\label{lemma3.4}
One has
\begin{gather*}
\widetilde{\mathfrak{D}_I}:=\mathfrak{D}_{I} \prod_{i \notin I,\, j \in I
\atop i < j}\big(1+t \beta-t^2 q_{ij}\big) = \prod_{b \in I}^{\nearrow}
\Bigg(\prod_{a \notin I \atop a < b}^{\searrow} h_{ba} \Bigg)
\prod_{a \in I}^{\nearrow} \Bigg((1+z_a) \prod_{b \notin I \atop a < b}^{\searrow} h_{ab} \Bigg).
\end{gather*}
\end{Lemma}
Note that if $a < b$, then $h_{ba}= 1+\beta t - u_{ab}$. Here we have used
the symbol
\begin{gather*}
\prod_{b \in I}^{\nearrow}
\Bigg(\prod_{a \notin I \atop a < b}^{\searrow} h_{ba} \Bigg)
\end{gather*}
to denote the following product. At f\/irst, for a given element $b \in I$ let
us def\/ine the set $I(b):=\{ a \in [1,n] \backslash I, \, a < b\}:=
(a_1^{(b)} < \cdots < a_{p}^{(b)})$ for some $p$ (depending on~$b$). If $I=
(b_1 < b_2 < \cdots < b_k)$, i.e., $b_i=a_i^{(b)}$, then we set
\begin{gather*}
\prod_{b \in I}^{\nearrow}
\Bigg(\prod_{a \notin I \atop a < b}^{\searrow} h_{ba} \Bigg)= \prod_{j=1}^{k}
\Bigg( u_{b_{j},a_{s}} u_{b_{j},a_{s-1}} \cdots u_{b_{j},a_{1}} \Bigg).
\end{gather*}
For example, let us take $n=6$ and $I=(1,3,5)$, then
\begin{gather*}
\widetilde{\mathfrak{D}_I}=h_{32}h_{54}h_{52}(1+z_1)h_{16}h_{14}h_{12} (1+
z_3)h_{36}h_{34}(1+z_5)h_{56}.
\end{gather*}
Let us {\it stress} that the element $\widetilde{\mathfrak{D}_I} \in
\overline{3HT_n(\beta)}$ is a linear combination of {\it square free}
monomials and therefore, a computation of the left hand side of the equality
stated in Theorem~\ref{theorem3.3} can be performed in the ``classical case'' that is in
the case $q_{ij}=0$, $\forall\, i < j$. This case corresponds to the computation
of the classical equivariant cohomology of the type~$A_{n-1}$ complete f\/lag
variety~${\cal{F}}l_n$ if $h=1$.
A proof of the $\beta=0$ case given in \cite[Theorem~1]{KM}, can be
immediately extended to the case $\beta \not= 0$.
\begin{Exercises}\label{exercises3.3}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] Show that
\begin{gather*}
\sum_{1 \le i_1 < \cdots < i_k \le n } \prod_{a= 1}^{k} (1+\beta)^{n-k-i_{a}+a} = {n \brack k}_{1+t \beta}.
\end{gather*}
\item[(2)] $(\beta,h)$-Stirling polynomials of the second type.
Def\/ine polynomials $S_{n,k}(\beta,h)$ as follows
\begin{gather*}
S_{n,k}(\beta,h) = \sum_{I \subset [1,n] \atop I=\{ 1 \le i_1,\ldots,i_k \le n \}} \prod_{a=1}^{k} \left(\beta^{n-k-i_{a}+a} + h\frac{\beta^{n-k-i_{a}+a}-1}{\beta-1} \right).
\end{gather*}
Show that
\begin{gather*}
S_{n,k}(1,1) = {n+1 \brace k+1},\qquad S_{n,k}(\beta,0)= {n \brack k}_{\beta}.
\end{gather*}
\end{enumerate}
\end{Exercises}
\section[Algebra $3T_n^{(0)}(\Gamma)$ and Tutte polynomial of graphs]{Algebra $\boldsymbol{3T_n^{(0)}(\Gamma)}$ and Tutte polynomial of graphs}\label{section4}
\subsection{Graph and nil-graph subalgebras, and partial f\/lag varieties}\label{section4.1}
Let's consider the set $R_n:= \{ (i,j) \in \mathbb{Z} \times \mathbb{Z} \,|\, 1 \le i < j
\le n \}$ as the set of edges of the complete graph~$K_n$ on~$n$ labeled
vertices $v_1,\ldots,v_n$. Any subset $S \subset R_n$ is the set of edges
of a unique subgraph $\Gamma:=\Gamma_{S}$ of the complete graph~$K_n$.
\begin{Definition}[graph and nil-graph subalgebras]\label{definition4.1}
The graph subalgebra $3T_n(\Gamma)$
(resp.\ nil-graph subalgebra $3T_n^{(0)}(\Gamma)$) corresponding to a subgraph
$\Gamma \subset K_n$ of the complete graph $K_n$,
is def\/ined to be the subalgebra in the algebra $3T_n$ (resp.~$3T_n^{(0)}$)
generated by the elements $\{u_{ij}\,|\, (i,j) \in \Gamma \}$.
\end{Definition}
In subsequent Sections~\ref{section4.1.1} and~\ref{section4.1.2} we will study some examples of
graph subalgebras corresponding to the complete multipartite graphs, cycle
graphs and linear graphs.
\subsubsection[Nil-Coxeter and af\/f\/ine nil-Coxeter subalgebras in $3T_n^{(0)}$]{Nil-Coxeter and af\/f\/ine nil-Coxeter subalgebras in $\boldsymbol{3T_n^{(0)}}$}\label{section4.1.1}
Our f\/irst example is concerned with the case when the graph $\Gamma$
corresponds to either the set $S:= \{(i,i+1) \,|\, i=1,\ldots,n-1 \}$ of
{\it simple roots} of type $A_{n-1}$, or the set $S^{\rm af\/f}:= S \cup
\{(1,n) \}$ of {\it affine simple roots} of type $A_{n-1}^{(1)}$.
{\samepage \begin{Definition} \label{definition4.2}\quad
\begin{enumerate}\itemsep=0pt
\item[(a)] Denote by ${\widetilde {\rm NC}_{n}}$ subalgebra in the
algebra $3T_{n}^{(0)}$ generated by the elements~$u_{i,i+1}$, $1 \le i \le n-1$.
\item[(b)] Denote by ${\widetilde {\rm ANC}_{n}}$ subalgebra in the algebra
$3T_{n}^{(0)}$ generated by the elements $u_{i,i+1}$, $1 \le i \le n-1$ and~$- u_{1,n}$.
\end{enumerate}
\end{Definition}}
\begin{Theorem} \label{theorem4.1}\quad
\begin{enumerate}\itemsep=0pt
\item[{\rm (A)}] The subalgebra ${\widetilde {\rm NC}_{n}}$ is
canonically isomorphic to the nil-Coxeter algebra ${\rm NC}_{n}$. In particular,
${\rm Hilb}({\widetilde {\rm NC}_{n}},t)=[n]_{t}{!}$ $($cf.~{\rm \cite{Ba})}.
\item[{\rm (B)}] The subalgebra ${\widetilde {\rm ANC}_{n}}$ has finite dimension and
its Hilbert polynomial is equal to
\begin{gather*}
{\rm Hilb}({\widetilde {\rm ANC}_{n}},t)= [n]_{t} \prod_{ 1 \le j \le n-1}[j(n-j)]_{t}=
[n]_{t}{!}\prod_{1 \le j \le n-1} [j]_{t^{n-j}}.
\end{gather*}
In particular, $\dim {\widetilde {\rm ANC}_{n}}=(n-1)! n !$,
$\deg_{t} {\rm Hilb}({\widetilde {\rm ANC}_{n}},t)= {{n+1}\choose 3}$.
\item[{\rm (C)}] The kernel of the map $ \pi\colon {\widetilde {\rm ANC}_{n}}
\longrightarrow {\widetilde {\rm NC}_{n}}$, $ \pi(u_{1,n}) = 0$,
$\pi(u_{i,i+1}) = u_{i,i+1}$, $1 \le i \le n-1$, is generated by the following elements:
\begin{gather*}
f_{n}^{(k)}= \prod_{j=k}^{1} \prod_{a=j}^{n-k+j-1} u_{a,a+1}, \qquad 1 \le k \le n-1.
\end{gather*}
\end{enumerate}
\end{Theorem}
Note that $\deg~f_{n}^{(k)}=k(n-k)$.
The statement (C) of Theorem~\ref{theorem4.1} means that the element $f_{n}^{(k)}$ which
does not contain the generator $u_{1,n}$, can be written as a linear
combination of degree $k(n-k)$ monomials in the algebra ${\widetilde
{\rm ANC}_{n}}$, each contains the generator $u_{1,n}$ at least once. By this
means we obtain a~set of all extra relations (i.e., additional to those
in the algebra $\widetilde {\rm NC}_{n}$) in the algebra
${\widetilde {\rm ANC}_{n}}$. Moreover, each monomial~$M$ in all linear
combinations mentioned above,
appears with coef\/f\/icient $(-1)^{\#|u_{1,n} \in M|+1}$. For example,
\begin{gather*}
f_4^{(1)}:=u_{1,2}u_{2,3}u_{3,4}=u_{2,3}u_{3,4}u_{1,4}+u_{3,4}u_{1,4}u_{1,2}+
u_{1,4}u_{1,2}u_{2,3},\\
f_4^{(2)}:=u_{2,3}u_{3,4}u_{1,2}u_{2,3}=
u_{1,2}u_{3,4}u_{2,3}u_{1,4}+u_{1,2}u_{2,3}u_{1,4}u_{1,2}+
u_{2,3}u_{1,4}u_{1,2}u_{3,4}\\
\hphantom{f_4^{(2)}:=}{} +u_{3,4}u_{2,3}u_{1,4}u_{3,4}-
u_{1,4}u_{1,2}u_{3,4}u_{1,4}.
\end{gather*}
Worthy of mention is that $\dim(\widetilde{\rm ANC}_{n})= (n-1)! n!$ is equal
to the number of (directed) Hamiltonian cycles in the complete bipartite graph
$K_{n,n}$, see, e.g., \cite[$A010790$]{SL} for additional information.
\begin{Remark}\label{remark4.1} More generally, let $(W,S)$ be a f\/inite crystallographic
Coxeter group of rank~$l$ with the set of exponents
$1=m_1 \le m_2 \le \cdots \le m_l$.
Let ${\cal B}_{W}$ be the corresponding Nichols--Woronowicz algebra, see, e.g.,~\cite{Ba}. Follow~\cite{Ba}, denote by $\widetilde {\rm NC}_{W}$ the subalgebra in
${\cal B}_{W}$ generated by the elements $[\alpha_s] \in {\cal B}_{W}$
corresponding to simple roots $s \in S$. Denote by $\widetilde {\rm ANWC}_{W}$
the subalgebra in ${\cal B}_{W}$ generated by $\widetilde {\rm NC}_{W}$ and the
element~$[a_{\theta}]$, where $[a_{\theta}]$ stands for the element in
${\cal B}_{W}$ corresponding to the highest root $\theta$ for~$W$. In other
words, $\widetilde {\rm ANWC}_{W}$ is the image of the algebra
$\widetilde {\rm ANC}_{W}$ under the natural map
$B{\cal E}(W) \longrightarrow {\cal B}_{W}$, see, e.g., \cite{Ba,KMa}.
It follows from \cite[Section~6]{Ba},
that ${\rm Hilb}(\widetilde {\rm NC}_{W},t)=\prod\limits_{i=1}^{l}[m_{i}+1]_{t}$.
\end{Remark}
\begin{Conjecture}[Yu.~Bazlov and A.N.~Kirillov, 2002]\label{conjecture4.1}
\begin{gather*}
{\rm Hilb}\big(\widetilde {\rm ANWC}_{W},t\big)=
\prod_{i=1}^{l}{{1-t^{m_{i}+1} \over 1-t^{m_i}}} {\prod_{i=1}^{l} {1-t^{a_i} \over
1-t }}=
P_{\rm af\/f}(W,t) \prod_{i=1}^{l} (1-t^{a_i}),
\end{gather*}
where
\begin{gather*}
P_{\rm af\/f}(W,t):=\sum_{w \in W_{\rm af\/f}} t^{l(w)}=
\prod_{i=1}^{l}{(1+t+\cdots+t^{m_i}) \over 1-t^{m_i}}
\end{gather*}
denotes the Poincar\'e polynomial corresponding to the affine Weyl group
$W_{\rm af\/f}$, see {\rm \cite[p.~245]{Bo}};
$a_{i}:=( 2 \rho,\alpha_{i}^{\vee})$,
$1 \le i \le l$, denote the coefficients of the decomposition of the sum of
positive roots~$2 \rho$ in terms of the simple roots~$\alpha_i$.
\end{Conjecture}
In particular,
\begin{gather*}
\dim \widetilde {\rm ANWC}_{W}=|W|{\prod\limits_{i=1}^{l}a_i \over
\prod\limits_{i=1}^{l} m_{i}} \qquad \text{and} \qquad \deg {\rm Hilb}\big(\widetilde {ANWC}_{W},t\big)=\sum\limits_{1=1}^{l}
a_{i}.
\end{gather*}
It is well-known that the product~$\prod\limits_{i=1}^{l}{1-t^{a_{i}}
\over 1-t^{m_i}}$ is a
symmetric (and {\it unimodal}?) polynomial with non-negative integer
coef\/f\/icients.
\begin{Example}\label{example4.1}\quad
\begin{alignat*}{3}
& (a) \quad && {\rm Hilb}\big(\widetilde{\rm ANC}_{3},t\big)=[2]_{t}^2[3]_{t}, \qquad
{\rm Hilb}\big(\widetilde{ANC}_{4},t\big)=[3]_{t}^2[4]_{t}^2, &\\
&&&
{\rm Hilb}\big(\widetilde{ANC}_{5},t\big)=[4]_{t}^2[5]_{t}[6]_{t}^2,& \\
& (b)\quad && {\rm Hilb}(B{\cal E}_2,t)=(1+t)^4\big(1+t^2\big)^2, &\\
&&& {\rm Hilb}(\widetilde{\rm ANC}_{B_2},t)= (1+t)^3\big(1+t^2\big)^2=P_{\rm af\/f}(B_2,t)\big(1-t^3\big)\big(1-t^4\big).&\\
& (c) \quad && {\rm Hilb}\big(\widetilde{\rm ANC}_{B_3},t\big)=
(1+t)^3\big(1+t^2\big)^2\big(1+t^3\big)\big(1+t^4\big)\big(1+t+t^2\big)\big(1+t^3+t^6\big)& \\
&&& \hphantom{{\rm Hilb}\big(\widetilde{\rm ANC}_{B_3},t\big)}{} =
P_{\rm af\/f}(B_3,t)\big(1-t^5\big)\big(1-t^8\big)\big(1-t^9\big).
\end{alignat*}
Indeed, $m_{B_3}=(1,3,5)$, $a_{B_3}=(5,8,9)$.
\end{Example}
\begin{Definition}\label{definition4.3}
Let $\langle \widetilde{\rm ANC}_{n} \rangle$ denote the two-sided
ideal in $3T_{n}^{(0)}$ generated by the ele\-ments $ \{ u_{i,i+1} \}$,
$1 \le i \le n-1$, and $u_{1,n}$. Denote by $U_{n}$ the quotient
$U_{n}=3T_{n}^{0} / \langle \widetilde{\rm ANC}_{n} \rangle$.
\end{Definition}
\begin{Proposition}\label{proposition4.1}
\begin{gather*}
U_{4} \cong \langle u_{1,3},u_{2,4} \rangle \cong \mathbb{Z}_{2} \times \mathbb{Z}_{2},\qquad
U_{5} \cong \langle u_{1,4},u_{2,4},u_{2,5},u_{3,5},u_{1,3} \rangle
\cong \widetilde{\rm ANC}_{5}.
\end{gather*}
\end{Proposition}
In particular, ${\rm Hilb}\big(3T_5^{(0)},t\big)=\bigl[ {\rm Hilb}({\widetilde{\rm ANC}}_5,t) \bigr]^2$.
\subsubsection{Parabolic 3-term relations algebras and partial f\/lag varieties}\label{section4.1.2}
In fact one can construct an analogue of the algebra $3HT_n$
and a commutative subalgebra inside it, for {\it any} graph
$\Gamma=(V,E)$ on $n$ vertices, possibly with loops and multiple edges~\cite{K}. We denote this algebra by $3T_n(\Gamma)$, and denote by
$3T_n^{(0)}(\Gamma)$ its {\it nil-quotient}, which may be considered as
a~``classical limit of the algebra $3T_n(\Gamma)$''.
The case of the complete graph $\Gamma=K_n$ reproduces the
results of the present paper and those of~\cite{K}, i.e., the case of the full
f\/lag variety ${\cal F}l_n$. The case of the {\it complete multipartite graph}
$\Gamma=K_{n_{1},\ldots, n_{r}}$ reproduces the analogue of results stated
in the present paper for the full f\/lag variety ${\cal F}l_n$, to the case of
the {\it partial flag} variety~${\cal F}_{n_{1},\ldots,n_{r}}$, see~\cite{K} for details.
We {\it expect} that in the case of the complete graph
with all edges having the same multiplicity~$m$, denoted by either
$\Gamma = K_{n}^{(m)}$, or~$m K_{n}$ in the present paper, the
commutative subalgebra generated by the Dunkl elements in the algebra
$3T_n^{(0)}(\Gamma)$ is related to the algebra of coinvariants of the diagonal
action of the symmetric group~$\mathbb{S}_n$ on the ring of polynomials
$\mathbb{Q}\big[X_n^{(1)},\ldots,X_n^{(m)}\big]$, where we set $X_n^{(i)} =
\big\{ x_1^{(i)},\ldots,x_n^{(i)} \big\}$.
\begin{Example} \label{example4.2} Take $\Gamma=K_{2,2}$. The algebra $3T^{(0)}(\Gamma)$ is
generated by four elements $\{a=u_{13}$, $b=u_{14}, c=u_{23}, d=u_{24} \}$
subject to the following set of (def\/ining) relations
\begin{itemize}\itemsep=0pt
\item $ a^2=b^2=c^2=d^2=0$, $ c b=b c$, $a d=d a$,
\item $a b a+b a b=0=a c a+c a c$, $b d b+d b d=0=c d c+d c d$,\\
$a b d-b d c-c a b+d c a=0=a c d-b a c-c d b+d b a$,
\item $a b c a+a d b c+b a d b+b c a d+c a d c+d b c d=0$.
\end{itemize}
It is not dif\/f\/icult to see that\footnote{Hereinafter we shell use notation
$ (a_{0},a_{1},\ldots,a_{k})_{t} := a_{0}+a_{1} t +
\cdots + a_{k} t^{k}$.}
\begin{gather*}
{\rm Hilb}\big(3T^{(0)}(K_{2,2}),t\big)=[3]_{t}^{2} [4]_{t}^{2},\qquad
{\rm Hilb}\big(3T^{(0)}(K_{2,2})^{ab},t\big)=(1,4,6,3).
\end{gather*}
Here for any algebra $A$ we denote by $A^{ab}$ its {\it abelianization}\footnote{See \url{http://groupprops.subwiki.org/wiki/Abelianization}.}.
\end{Example}
The commutative subalgebra in $3T^{(0)}(K_{2,2})$, which corresponds to
the intersection
$3T^{(0)}(K_{2,2})$ $\cap \mathbb{Z}[\theta_1,\theta_2,\theta_3,\theta_4]$,
is generated by the elements $c_1:=\theta_1+\theta_2=(a+b+c+d)$ and
$c_2:= \theta_1 \theta_2=(ac+ca+bd+db+ad+bc)$. The elements $c_1$ and~$c_2$
commute and satisfy the following relations
\begin{gather*}
c_{1}^{3}-2 c_1 c_2=0,\qquad c_{2}^{2}-c_{1}^{2} c_2=0.
\end{gather*}
The ring of polynomials $\mathbb{Z}[c_1,c_2]$ is isomorphic to the cohomology ring
$H^{*}(\operatorname{Gr}(2,4),\mathbb{Z})$ of the Grassmannian variety $\operatorname{Gr}(2,4)$.
To continue exposition, let us take $m \le n$, and consider the complete
multipartite graph $K_{n,m}$ which corresponds to the Grassmannian variety
$\operatorname{Gr}(n,m+n)$. One can show
\begin{gather*}
{\rm Hilb}\big(3T_{n+m}^{(0)}(K_{n,m})^{ab},t\big)= \sum_{k=0}^{n-1} (-1)^{k}
(1+(n-k) t)^{m-1} \prod_{j=1}^{n-k} (1+j t) {n \brace n-k} \\
\hphantom{{\rm Hilb}\big(3T_{n+m}^{(0)}(K_{n,m})^{ab},t\big)}{}
=t^{n+m-1} {\rm Tutte}\big(K_{n,m},1+t^{-1},0\big),
\end{gather*}
where ${n \brace k } := S(n,k)$ denotes the
Stirling numbers of the second kind, that is the number of ways to partition
a~set of~$n$ labeled objects into~$k$ nonempty unlabeled subsets, and for
any graph $\Gamma$,
${\rm Tutte}(\Gamma,x,y)$ denotes the {\it Tutte polynomial\/}\footnote{See, e.g., \url{https://en.wikipedia.org/wiki/Tutte_polynomial}. It is
well-known that
\begin{gather*}
{\rm Tutte}(\Gamma,1+t,0)= {(-1)^{|\Gamma|}} {t^{-\kappa(\Gamma)}}
{\rm Chrom}(\Gamma,-t),
\end{gather*}
where for any graph $\Gamma$, $|\Gamma|$ is equal to the number of vertices
and $\kappa(\Gamma)$ is equal to the number of connected components of~$\Gamma$. Finally ${\rm Chrom}(\Gamma,t)$ denotes the {\it chromatic polynomial}
corresponding to graph~$\Gamma$, see, e.g.,~\cite{We}, or \url{https://en.wikipedia.org/wiki/Chromatic_polynomial}.\label{footnote32}}
corresponding to graph~$\Gamma$.
It is well-known that the Stirling numbers $S(n,k)$ satisfy the following
identities
\begin{gather*}
\sum_{k=0}^{n-1} (-1)^k~S(n,n-k) \prod_{j=1}^{n-k} (1+j t)=(1+t)^n,\qquad
\sum_{n \ge k} {n \brace k } {x^n \over n !}={(e^x-1)^k
\over k !}.
\end{gather*}
Let us observe that
\begin{gather*}
\dim\big(3T^{(0)}(K_{n,n})^{ab}\big)=
\sum_{k=0}^{n-1} (-1)^{k} (n+1-k)^{n-1} (n+1-k) ! {n \brace n-k }\\
\hphantom{\dim\big(3T^{(0)}(K_{n,n})^{ab}\big)}{} =
\sum_{k=1}^{n+1} ((k-1) !)^2 { n+1 \brace k}^2 ,
\end{gather*}
see, e.g., \cite[$A048163$]{SL}.
Moreover, if $m \ge 0$, then
\begin{gather*}
\sum_{n \ge 1} \dim\big(3T^{(0)}(K_{n,n+m})^{ab}\big) { t^{n}} = \sum_{k \ge 1}
{k^{k+m-1} (k-1) ! t^{k} \over \prod\limits_{j=1}^{k-1} (1+k j t) },\\
\sum_{n \ge 1} {\rm Hilb}\big(3T^{(0)}(K_{n,m})^{ab},t\big) z^{n-1}=
\sum_{k \ge 0} (1+k t)^{m-1} \prod_{j=1}^{k} {{z (1+j t) \over 1+j z}}.
\end{gather*}
\begin{Comments}[poly-Bernoulli numbers]\label{commenst4.1}
Based on listed above identities involving the Stirling numbers $S(n,k)$, one
can prove the following {\it combinatorial} formula
\begin{gather}\label{equation4.1}
\dim \big(3T^{(0)}(K_{n,m})^{ab}\big) = \sum_{j=0}^{\min(n,m)}(j !)^2
{n+1 \brace j+1} {m+1 \brace j+1} = {B_{n}^{(-m)}} = {B_{m}^{(-n )}},
\end{gather}
where $B_n^{(k)}$ denotes the {\it poly-Bernoulli number} introduced by M.~Kaneko~\cite{Kan}.
On the other hand, it is well-known, see, e.g., footnote~\ref{footnote32}, that for any graph
$\Gamma$ the specialization ${\rm Tutte}(\Gamma; 2,0)$ of the Tutte polynomial associated with graph~$\Gamma$, counts the number of acyclic orientations of~$\Gamma$. Therefore, the poly-Bernulli number $B_n^{(-m)}$ counts the number of acyclic orientatations of the complete bipartite graph~$K_{n,m}$.
For example, $\dim \big(3T^{(0)}(K_{3,3})^{ab}\big)= 230 =1 + 7^2 +(2 !)^2 6^2 + (3 !)^2$, cf.\ Example~\ref{examples4.16}(3).
For the reader's convenient, we recall below a def\/inition of
{\it poly-Bernoulli numbers}. To start with, let $k$ be an integer, consider
the formal power series
\begin{gather*}
\operatorname{Li}_k(z):= \sum_{n=1}^{\infty} {z^n \over n^k}.
\end{gather*}
If $k \ge 1$, $\operatorname{Li}_k(z)$ is the $k$-th {\it polylogarithm}, and if
$k \le 0$, then $\operatorname{Li}_k(z)$ is a~rational function. Clearly
$\operatorname{Li}_1(z)=-\ln(1-z)$. Now def\/ine {\it poly-Bernoulli numbers} through the
generating function
\begin{gather*}
{\operatorname{Li}_k(1-e^{-z}) \over {1-e^{-z}}} = \sum_{n=0}^{\infty} B_n^{(k)} {z^n
\over n !}.
\end{gather*}
Note that a combinatorial formula for the numbers $B_n^{(-k)}$ stated in
\eqref{equation4.1} is a consequence of the following identity~\cite{Kan}
\begin{gather*}
\sum_{n=0}^{\infty} \sum_{k=0}^{\infty} B_n^{(-k)} {{x^n \over n !}}
{{z^k \over k !}} ={e^{x+z} \over {1-(1-e^x)(1-e^z)}}.
\end{gather*}
Note that the poly-Bernoulli numbers $B_{n}^{(-m)} (=B_{m}^{(-n)})$
have the following combinatorial interpretation\footnote{See for example, \cite[$A136126$]{SL}, \cite[$A099594$]{SL} or~\cite[Theorem~3.1]{CE},
and the literature quoted therein. Recall, that the {\it excedance set} of a~permutation $\pi \in \mathbb{S}_n$ is the set of indices~$i$, $1 \le i \le n$, such that $\pi(i) > i$.}, namely, the number $B_{n}^{(-m)}$, and therefore
the dimension of the algebra $3T^{(0)}(K_{n,m})$ is
equal to
\begin{gather*}
B_{n}^{(-m)} =T(n-1,m)+T(n,m-1),
\end{gather*}
where~\cite{CE}
\begin{gather*}
T(n,m):= \sum_{j=0}^{\min(n,m)} j ! (j+1) ! {n+1 \brace j+1} {m+1 \brace j+1}
\end{gather*}
is equal to the number of permutations $w \in
\mathbb{S}_{n+m}$~having the {\it excedance set} $\{1,2,\ldots,m \}$.
\begin{Exercise} \label{exercise4.1} Show that polynomial ${\rm Hilb}(3T^{(0)}(K_{n,m}),t)$ has
degree $n+m-1$, and
\begin{gather*}
{\rm Coef\/f}_{t^{n+m-1}} \bigl({\rm Hilb}\big(3T^{(0)}(K_{n,m}),t\big)\bigr) = T(n-1,m-1).
\end{gather*}
\end{Exercise}
\begin{Problem}\label{problem4.0}
To find a bijective proof of the identity~\eqref{equation4.1}.
\end{Problem}
Final remark, the explicit expression for the chromatic polynomial of the
complete tripartite graph $K_{n,n,n}$ can be found in~\cite[$A212220$]{SL}.
\end{Comments}
Now let $\theta_i^{(n+m)} = \sum\limits_{j \not=i} u_{ij}$, $1 \le i \le n+m$, be the
Dunkl elements in the algebra $3T^{(0)}(K_{n+m})$, def\/ine the following elements
the in the algebra $3T^{(0)}(K_{n,m})$
\begin{gather*}
c_k:=e_k\big(\theta_1^{(n+m)},\ldots,\theta_{n}^{(n+m)}\big), \qquad 1 \le k \le n,
\\
{\overline{c}}_r:=e_r\big(\theta_{n+1}^{(n+m)},\ldots,\theta_{n+m}^{(n+m)}\big),\qquad
1 \le r \le m.
\end{gather*}
Clearly,
\begin{gather*}
\left(1+\sum_{k=1}^{ n} c_k t^k\right)\left(1+\sum_{r=1}^{m} {\overline{c}}_r t^r\right) =
\prod_{j=1}^{n+m} \big(1+\theta_{j}^{(n+m)}\big) = 1.
\end{gather*}
Moreover, there exist the natural isomorphisms of algebras
\begin{gather*}
H^{*}(\operatorname{Gr}(n,n+m),\mathbb{Z}) \cong \mathbb{Z}[c_1,\ldots,c_n] \bigg/ \left\langle \left(1+\sum_{k=1}^{n} c_k t^k\right)
\left(1+\sum_{r=1}^{m} {\overline{c}}_r t^r\right) -1 \right\rangle,\\
QH^{*}(\operatorname{Gr}(n,n+m)) \cong \mathbb{Z}[q][c_1,\ldots,c_n] \bigg/ \left\langle
\left(1+\sum_{k=1}^{n} c_k t ^k\right)\left(1+\sum_{r=1}^{m} {\overline{c}}_r t^r\right) -1 - q t^{n+m} \right\rangle.
\end{gather*}
Let us recall, see Section~\ref{section2}, footnote~\ref{footnote26}, that for a commutative ring~$R$
and a~polynomial $p(t)=\sum\limits_{j=1}^{s}g_j t^{j}
\in R[t]$, we denote by $\langle p(t) \rangle$ the ideal in the ring~$R$
generated by the coef\/f\/icients $g_1,\ldots,g_s$.
These examples are illustrative of the similar results valid for the
{\it general complete multipartite graphs} $K_{n_{1},\ldots,n_{r}}$, i.e.,
for the partial f\/lag varieties~\cite{K}.
To state our results for {\it partial flag varieties}
we need a bit of notation. Let $N:=n_1+\cdots+n_r$, $n_j > 0$, $\forall\, j$, be a~composition of size~$N$.
We set $N_j:=n_1+ \cdots +n_j$, $j=1,\ldots,r$, and
$N_0=0$. Now, consider the commutative subalgebra in the algebra
$3T_{N}^{(0)}(K_{N})$ generated by the set of Dunkl elements
$\big\{\theta_1^{(N)},\ldots,\theta_{N}^{(N)} \big\}$, and def\/ine elements
$\big\{ c_{k_{j}}^{(j,N)} \in 3T_N^{(0)}(K_{n_1,\ldots,n_r})\big\}$ to be the degree~$k_j$ elementary symmetric polynomials of the Dunkl
elements $\theta_{N_{j-1}+1}^{(N)},\ldots,\theta_{N_{j}}^{(N)}$, namely,
\begin{gather*}
c_k^{(j)}:= c_{k_{j}}^{(j,N)}= e_{k}\big(\theta_{N_{j-1}+1}^{(N)},\ldots,\theta_{N_{j}}^{(N)}\big),
\qquad 1 \le k_j \le n_j, \qquad j=1,\ldots,r,\\
c_{0}^{(j)}=1, \qquad \forall\, j.
\end{gather*}
Clearly
\begin{gather*}
\prod_{j=1}^{r} \left(\sum_{a=0}^{n_j} c_{a}^{(j)} t^{a}\right) = \prod_{j=1}^{N}
\big(1+\theta_j^{(N)}t^{j}\big) = 1.
\end{gather*}
\begin{Theorem}\label{theorem4.2}
The commutative subalgebra generated by the elements
$ \{ c_{k_{j}}^{(j)},\,1 \le k_j \le n_j,\,1 \le j \le r-1\}$, in the algebra
$3T_N^{(0)}(K_{n_1,\ldots,n_r})$ is isomorphic to the cohomology ring
$H^{*}({\cal{F}}l_{n_1,\ldots,n_r}, \mathbb{Z})$ of the partial flag variety
${\cal{F}}l_{n_1,\ldots,n_r}$.
\end{Theorem}
In other words, we treat the Dunkl elements $\big\{
\theta_{N_{j-1}+a}^{(N)},\,1 \le a \le n_j\big\}$, $j=1,\ldots,r$, as the
{\it Chern roots} of the vector bundles $\{ \xi_{j}:=
{\cal{F}}_j / {\cal{F}}_{j-1} \}$, $j=1,\ldots,r$,
over the partial f\/lag variety ${\cal{F}}l_{n_1,\ldots,n_r}$.
Recall that a point ${\boldsymbol{F}}$ of the partial f\/lag variety ${\cal{F}}l_{n_1,
\ldots,n_r}$, $n_1+\cdots+n_r=N$, is a sequence of embedded subspaces
\begin{gather*}
{\boldsymbol{F}}=\big\{0=F_{0} \subset F_1 \subset F_2 \subset \cdots \subset F_r=\mathbb{C}^{N}
\big\}
\end{gather*}
such that
\begin{gather*}
\dim (F_i / F_{i-1})=n_i, \qquad i=1,\ldots, r.
\end{gather*}
By def\/inition, the f\/iber of the vector bundle~$\xi_i$ over a point ${\boldsymbol{F}} \in
{\cal{F}}l_{n_1,\ldots,n_r}$ is the $n_i$-dimensional vector space
$F_{i} / F_{i-1}$.
To conclude, let us describe the set of (def\/ining) relations among the
elements $\big\{c_{a}^{(j)}\big\}$, $1 \le a \le n_{j}$, $1 \le j \le r-1$. To proceed,
let us introduce the set of variables $\big\{x_{a}^{(j)}\, |\, 1 \le a \le n_j, \, 1 \le j \le r-1 \big\}$, and def\/ine polynomials $b_0=1,b_{k}:=b_{k}\big(\big\{x_{a}^{(j)}\big\}\big)$, $k \ge 1$ by the use of generating function
\begin{gather*}
\frac{1}{\prod\limits_{j=1}^{r-1} \prod\limits_{a=1}^{n_{j}}\big(1+x_{a}^{(j)}\big) t^{a}} =
\sum_{k \ge 0} b_{k} t^{k}.
\end{gather*}
Now let us introduce matrix $M_{m}\big(\big\{x_{a}^{(j)}\big\}\big):= (m_{ij})$, where
\begin{gather*}
m_{ij}:= \begin{cases}
b_{i+j-1} & \text{if} \ \ j > i, \\
1 & \text{if} \ \ j=i-1, \ i \ge 2 , \\
0 & \text{if} \ j < i-1.
\end{cases}
\end{gather*}
\begin{Claim}\label{claim_page58} $\det M_{m}\big(\big\{c_{a}^{(i)}\big\}\big) = 0$, $N_{r-1} < m \le N$. Moreover,
\begin{gather*}
H^{*}({\cal{F}}l_{n_1,\ldots,n_r},\mathbb{Z})
\cong \mathbb{Z}[\{x_{a}^{j} \}] / \langle M_{N_{r-1}+1},\ldots, M_{N} \rangle.
\end{gather*}
\end{Claim}
A meaning of the algebra $3T_n^{(0)}(\Gamma)$ and
the corresponding commutative subalgebra inside it for a general graph
$\Gamma$ is still unclear.
\begin{Conjecture}\footnote{Part $(1)$ of this conjecture has been proved in~\cite{Li}.
In~\cite{Li} the author has used notation ${\rm OT}(\Gamma)$ for the Orlik--Terao
algebra associated with (simple) graph~$\Gamma$. In our paper we prefer to
denote the corresponding Orlik--Terao algebra by ${\rm OS}^{+}(\Gamma)$.} \label{conjecture4.2}
\begin{enumerate}\itemsep=0pt
\item[$(1)$] Let $\Gamma=(V,E)$ be a connected subgraph of the complete graph
$K_n$ on $n$ vertices. Then
\begin{gather*}
{\rm Hilb}\big(3T_n^{(0)}(\Gamma)^{ab},t\big)= t^{|V|-1} {\rm Tutte}\big(\Gamma;1+t^{-1},0\big).
\end{gather*}
\item[$(2)$] Let $\Gamma=(V,E,\{m_{ij}, (ij) \in E \})$ be a connected subgraph of
the complete graph $K^{({\boldsymbol{m})}}_n$ with multiple edges such that an edge
$(ij) \in K_n$ has the multiplicity~$m_{ij}$. Let $3T_n^{(0)}(\Gamma,{\boldsymbol{m}})$ denotes the subalgebra in the algebra $3T_n^{(0)}({\boldsymbol{m}})$ generated by
elements $\{u_{ij}^{(\alpha_{(ij)})},\, (ij) \in E,\, 1 \le \alpha_{(ij)} \le m_{ij} \}$, see Section~{\rm \ref{section2.2}}. Let ${\cal{A}}(\Gamma,\{m_{ij}\})$ denotes the
graphic arrangement cor\-res\-pon\-ding to the graph $(\Gamma,\{m_{ij}\})$,
that is the set of hyperplanes $\{ H_{(ij),a}= (x_i -x_j= a),\, 0 \le a \le
m_{ij}-1,\, (ij) \in E \}$.
Then
\begin{gather*}
3T_n^{(0)} (\Gamma,{\boldsymbol{m}})^{ab} = {\rm OS}^{+}({\cal{A}}(\Gamma,\{m_{ij} \})),
\end{gather*}
where for any arrangements of hyperplanes ${\cal{A}}$, ${\rm OS}^{+}({\cal{A}})$
denotes the even Orlik--Solomon algebra of the arrangement~$\cal{A}$~{\rm \cite{OT}}.
In the case when $m_{ij}=1$, $\forall\, 1 \le i < j \le n$,
$3T_n^{(0)}(\Gamma)^{\rm anti} = {\rm OS}({\cal{A}}(\Gamma)) $.
\end{enumerate}
\end{Conjecture}
\begin{Examples}\label{examples4.16}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] Let $G=K_{2,2}$ be complete bipartite graph of type $(2,2)$. Then
\begin{gather*}
{\rm Hilb}\big(3T_4^{0}(2,2)^{ab},t\big)=(1,4,6,3)=t^2(1+t)+t(1+t)^2+(1+t)^3,
\end{gather*}
and the Tutte polynomial for the graph $K_{2,2}$ is equal to $x+x^2+x^3+y$.
\item[(2)] Let $G=K_{3,2}$ be complete bipartite graph of type $(3,2)$. Then
\begin{gather*}
{\rm Hilb}\big(3T_5^{0}(3,2)^{ab},t\big)=(1,6,15,17,7)\\
\hphantom{{\rm Hilb}\big(3T_5^{0}(3,2)^{ab},t\big)}{}
=t^3 (1+t)+3 t^2~(1+t)^2+2t (1+t)^3+(1+t)^4,
\end{gather*}
and the Tutte polynomial for the graph $K_{3,2}$ is equal to
\begin{gather*}
x+3 x^2+2 x^3+x^4 +y+3 x y+y^2.
\end{gather*}
\item[(3)] Let $G=K_{3,3}$ be complete bipartite graph of type $(3,3)$. Then
\begin{gather*}
{\rm Hilb}\big(3T_6^{0}(3,3)^{ab},t\big)=(1,9,36,75,78,31)=(1+t)^5+4t(1+t)^4\\
\hphantom{{\rm Hilb}\big(3T_6^{0}(3,3)^{ab},t\big)=}{}+10t^{2}(1+t)^3+11t^{3}(1+t)^2+5t^{4}(1+t),
\end{gather*}
and the Tutte polynomial of the bipartite graph $K_{3,3}$ is equal to
\begin{gather*}
5x+11x^2+10x^3+4x^4+x^5+15xy+9x^2y+6xy^2+5y+9y^2+5y^3+y^4.
\end{gather*}
\item[(4)] Consider complete multipartite graph $K_{2,2,2}$. One can show
that
\begin{gather*}
{\rm Hilb}\big(3T_6^{(0)}(K_{2,2,2})^{ab},t\big)=(1,12,58,137,154,64)=
11t^4(1+t)+25t^3(1+t)^2\\
\hphantom{{\rm Hilb}\big(3T_6^{(0)}(K_{2,2,2})^{ab},t\big)=}{}+20 t^2(1+t)^3+7 t (1+t)^4+(1+t)^5,
\end{gather*}
and
\begin{gather*}
{\rm Tutte}(K_{2,2,2},x,y)= x(11,25 ,20,7,1)_{x}+ y (11,46,39,8)_{x} +
y^2(32,52,12)_{x}\\
\hphantom{{\rm Tutte}(K_{2,2,2},x,y)=}{}
+y^3(40,24)_{x}+ y^4(29,6)_{x}+15y^{5}+5 y^{6}+y^{7} .
\end{gather*}
\end{enumerate}
\end{Examples}
The above examples show that the Hilbert polynomial ${\rm Hilb}(3T_n^{0}(G)^{ab},t)$
appears to be a~certain specialization of the Tutte polynomial of the
corresponding graph~$G$.
Instead of using the Hilbert polynomial of the
algebra $3T_n^{0}(G)^{ab}$ one can consider the graded Betti numbers (over a
f\/ield $\boldsymbol{k}$) polynomial ${\rm Betti}_{\boldsymbol{k}}(3T_n^{0}(G)^{ab},x,y)$. For example,
\begin{gather*}
{\rm Betti}_{\mathbb{Q}}\big(3T_3^{0}(K_3)^{ab},x,y\big)= 1+ x y(4,2)_{x} + x^2 y^2 (3,2)_{x},\\
{\rm Betti}_{\mathbb{Q}}\big(3T_4^{0}(K_{2,2})^{ab},x,y\big)= 1+ 4 x y+ x y^2(1,9,3)_{x}+ x^2 y^3 (1,6,3)_{x},\\
{\rm Betti}_{\mathbb{Q}}\big(3T_5^{0}(K_{3,2})^{ab},x,y\big)= 1+ 6 x y+ x y^2(3,25,9)+x^2 y^3(6,45,34,7)+ x^3 y^4(3,25,25,7),\\
{\rm Betti}_{\mathbb{Q}}\big(3T_4^{0}(K_4)^{ab},x,y\big)=
1+ x y (10,10)+ x^2 y^2(25,46,26,6)+ x^3 y^3 (15,36,25,6),\\
{\rm Betti}_{\mathbb{Z}/2\mathbb{Z}}\big(3T_4^{0}(K_4)^{ab},x,y\big)=
1+ x y (10,10,{\bf 1})_{x}+ x^2 y^2(25,46,26,6)+ x^3 y^3 ({\bf 16},36,25,6), \\
{\rm Betti}_{\mathbb{Q}}\big(3T_5^{0}(K_5)^{ab},x,y\big)= 1+ x y(20,30)+ x^2 y^2(109,342,315,72)+ x^3 y^3(195,852,1470,\\
\hphantom{{\rm Betti}_{\mathbb{Q}}\big(3T_5^{0}(K_5)^{ab},x,y\big)=}{} 1232,639,190,24)+ x^4 y^4 (105,540,1155,1160,639,190,24),\\
{\rm Betti}_{\mathbb{Q}}\big(3T_5^{0}(K_5)^{ab},1,1\big)= 9304, \\
{\rm Betti}_{\mathbb{Z}/3\mathbb{Z}}\big(3T_5^{0}(K_5)^{ab},x,y\big)= 1+x y (20,30)+x^2 y^2(109,342,315,72,{\bf 1}) \\
\hphantom{{\rm Betti}_{\mathbb{Z}/3\mathbb{Z}}\big(3T_5^{0}(K_5)^{ab},x,y\big)=}{} +x^3 y^3 (195,852,1471,1232,640,190,24)\\
\hphantom{{\rm Betti}_{\mathbb{Z}/3\mathbb{Z}}\big(3T_5^{0}(K_5)^{ab},x,y\big)=}{}
+x^4 y^4 (105,540,{\bf 1156},1160,639,190,24), \\
{\rm Betti}_{\mathbb{Z}/3\mathbb{Z}}\big(3T_5^{0}(K_5)^{ab},1,1\big)= 9308,
\\
{\rm Betti}_{\mathbb{Z}/2\mathbb{Z}}\big(3T_5^{0}(K_5)^{ab},x,y\big)= 1+x y(20,30,{\bf 5})+x^2 y^2({\bf 114},342,{\bf 340},{\bf 131},{\bf 10}) \\
\hphantom{{\rm Betti}_{\mathbb{Z}/2\mathbb{Z}}\big(3T_5^{0}(K_5)^{ab},x,y\big)=}{}+
x^3 y^3({\bf 220},{\bf 911},{\bf 1500},1291,649,190,24)\\
\hphantom{{\rm Betti}_{\mathbb{Z}/2\mathbb{Z}}\big(3T_5^{0}(K_5)^{ab},x,y\big)=}{}
+x^4 y^4 ({\bf 125},{\bf 599},{\bf 1165},1160,639,190,24), \\
{\rm Betti}_{\mathbb{Z}/2\mathbb{Z}}\big(3T_5^{0}(K_5)^{ab},1,1\big)= 9680, \\
{\rm Betti}_{\mathbb{Z}/2\mathbb{Z}}\big(3T_6^{0}(K_{3,3})^{ab},x,y\big)= 1+9 xy+x y^2(9,69,27)+ x^2 y^3(40,285,257,{\bf 52}) \\
\hphantom{{\rm Betti}_{\mathbb{Z}/2\mathbb{Z}}\big(3T_6^{0}(K_{3,3})^{ab},x,y\big)=}{}+
x^3 y^4(59,526,866,563,201,31)\\
\hphantom{{\rm Betti}_{\mathbb{Z}/2\mathbb{Z}}\big(3T_6^{0}(K_{3,3})^{ab},x,y\big)=}{}
+ x^4 y^5(28,{\bf 311},636,520,201,31),\\
{\rm Betti}_{\mathbb{Z}/2\mathbb{Z}}\big(3T_6^{0}(K_{3,3})^{ab},1,1\big)= 4740,\\
{\rm Betti}_{\mathbb{Q}}\big(3T_6^{0}(K_{3,3})^{ab},x,y\big)= 1+ 9 xy + x y^2(9,69,27)+ x^2 y^3(40,285,257,43)\\
\hphantom{{\rm Betti}_{\mathbb{Q}}\big(3T_6^{0}(K_{3,3})^{ab},x,y\big)=}{}
+ x^3 y^4(59,
526,866,563,201,31)\\
\hphantom{{\rm Betti}_{\mathbb{Q}}\big(3T_6^{0}(K_{3,3})^{ab},x,y\big)=}{}
+ x^4 y^5(28,302,636,520,201,31),\\
{\rm Betti}_{\mathbb{Q}}\big(3T_6^{0}(K_{3,3})^{ab},1,1\big)= 4704.
\end{gather*}
Let us observe that in all examples displayed above, the Betti polynomials are
divisible by $1+x y$.
It should be emphasize that in the literatute one can f\/ind def\/initions of
big variety of (graded) Betti's numbers associated with a~given simple graph $\Gamma$, depending on choosing an algebra/ideal has been attached to graph~$\Gamma$. For example, to def\/ine Betti's numbers, one can start with {\it edge graph ideal/algebra} associated with a graph in question, the {\it Stanley--Reisner} ideal/ring and so on and so far. We refer the reader to carefully written
book by E.~Miller and B.~Sturmfels~\cite{MiS} for def\/initions and results
concerning combinatorial commutative algebra graded Betti's numbers. As far as
I'm aware, the graded Betti numbers we are looking for in the present paper,
are dif\/ferent from those treated in~\cite{MiS}, and more close to those studied in~\cite{B}.
It is not dif\/f\/icult to see (A.K.) that for a simple connected graph
$\Gamma$ the coef\/f\/icient just before the (unique!) monomial of the maximal degree in
${\rm Betti}_{\boldsymbol{k}}\big(3T^{0}(\Gamma)^{ab},x,y\big)$ is equal to
${\rm Tutte}(\Gamma;1,0)$. It is known \cite{BCP} that the number
${\rm Tutte}(\Gamma;1,0)$ counts that of acyclic orientations of the edges of
$\Gamma$ with a unique source at a vertex $v \in \Gamma$, or equivalently~\cite{BCP}, the number of maximum $\Gamma$-parking functions relative to vertex~$v$.
\begin{Claim}\label{claim4.0}
Let $G=(V,E)$ be a connected graph without loops. Then over any field $\textbf{k}$
\begin{gather*}
{\rm Betti}_{\textbf{k}} \big(3T_n^{0}(G)^{ab},-x,x\big)=(1-x)^{e} {\rm Hilb}\big(3T_n^{0}(G)^{ab},x\big),
\end{gather*}
where $n=|V(G)| = \text{number of vertices}$, $e=|E(G)|=\text{number of edges}$.
\end{Claim}
\begin{Question} \label{question4.1}\quad
\begin{itemize}\itemsep=0pt
\item Let $G$ be a connected subgraph of the complete graph~$K_n$. Does the graded Betti polynomial ${\rm Betti}_{\mathbb{Q}}(3T_n^{0}(G)^{ab},x,y)$ is
a~certain specialization of the Tutte polynomial $T(G,x,y)$? If not,
give example of two $($simple$)$ graphs such that their
Orlik--Terao algebras have the same Tutte polynomial, but different
Betti polynomials over $\mathbb{Q}$ , and vice versa.
\item It is clear that for any graph $\Gamma$ $($or matroid$)$ one has
${\rm Tutte}(\Gamma,x,y)= a(\Gamma) (x+y) + (\text{higher degree terms})$ for some
integer $a(\Gamma) \in \mathbb{N} $. Does the number $a(\Gamma)$ have a~simple
combinatorial interpretation?
\end{itemize}
\end{Question}
\begin{Proposition}\label{proposition4.2}
Let ${\boldsymbol{n}}=(n_1,\ldots,n_r)$ be a composition of $n \in \mathbb{Z}_{\ge 1}$, then
\begin{gather*}
{\rm Hilb}\big(3T^{(0)} (K_{n_1,\ldots,n_r})^{ab},t\big) =\sum_{{\boldsymbol{k}}=(k_1,\ldots,k_r)
\atop 0 < k_j \le n_j } (-t)^{|{\boldsymbol{n}}|-|{\boldsymbol{k}}|}
\prod_{j=1}^{r} {n_j \brace k_j }
\prod_{j=1}^{|{\boldsymbol{k}}| -1}(1+j t),
\end{gather*}
where we set~$|{\boldsymbol{k}}|:=k_1+\cdots+k_r$.
\end{Proposition}
\begin{Remark}\label{remark4.0}
This proposition is a consequence of Conjecture~\ref{conjecture4.2}(1), which
has been proved in~\cite{Li}.
\end{Remark}
\begin{Corollary} \label{corollary4.1} One has
\begin{alignat*}{3}
& (a)\quad && 1+t(t-1) \sum_{(n_1,\ldots,n_r) \in \mathbb{Z}_{\ge 0}^{r} \backslash 0^{r}}
{\rm Hilb}\big(3T^{(0)}(K_{n_1,\ldots,n_r}\big)^{ab},t) {{x_{1}^{n_1} \over n_1 !}} \cdots{{x_{r}^{n_r} \over n_r !}} &\\
&&& \qquad{} = \left(1 +t \sum_{j=1}^{r} (e^{-x_j}-1) \right)^{1-t},&\\
& (b)\quad && \sum_{(n_1,n_2,\ldots,n_r) \in \mathbb{Z}_{\ge 0} \backslash 0^{r}}
\dim \big(3T^{(0)}(K_{n_1,\ldots,n_r})^{ab} \big)
{{x^{n_1} \over n_1 !}} \cdots {{x^{n_r} \over n_r !}}= - \log \left(1-r+
\sum_{j=1}^{r} e^{-x_j} \right), & \\
& (c)\quad && {\rm Hilb}\big(3T^{(0)}(K_{n_1,\ldots,n_r})^{ab},t\big)= (-t)^{|{\boldsymbol{n}}|} {\rm Chrom} \big(K_{n_1,\ldots,n_r},-t^{-1}\big),& \\
& (d)\quad && \dim \big(3T^{(0)}(\Gamma)^{ab}\big) \ \ \text{is equal to the number of acyclic orientations of $\Gamma$}, &
\end{alignat*}
where $\Gamma$ stands for a simple graph.
\end{Corollary}
Recall that for any graph $\Gamma$ we denote by ${\rm Chrom}(\Gamma,x)$
the chromatic polynomial of that graph.
Indeed, one can show\footnote{If $r=1$, the complete unipartite graph $K_{(n)}$ consists of $n$
distinct points, and
\begin{gather*}
{\rm Chrom}(K_{(n)},x)= x^n = \sum_{k=0}^{n-1} {n \brace k } (x)_{k}.
\end{gather*}
Let us stress that to abuse of notation the complete unipartite graph~$K_{(n)}$ consists of $n$ disjoint points with the Tutte polynomial equals
to~$1$ for all $n \ge 1$, whereas the complete graph~$K_n$ is equal to the
complete multipartite graph $K_{(1^n)}$.}
\begin{Proposition}\label{proposition4.3}
If $r \in \mathbb{Z}_{\ge 1}$, then
\begin{gather*}
{\rm Chrom} (K_{n_1,\ldots,n_r},t)= \sum_{{\boldsymbol{k}}=(k_1,\ldots,k_r)}
\prod_{j=1}^{r} {n_j \brace k_j } (t)_{|{\boldsymbol{k}}|},
\end{gather*}
where by definition $(t)_{m}:= \prod\limits_{j=1}^{m-1} (t-j)$, $(t)_{0}=1$,
$(t)_m =0$ if $m< 0$.
\end{Proposition}
Finally we describe explicitly the exponential generating function for the
{\it Tutte polynomials} of the weighted complete multipartite graphs. We
refer the reader to~\cite{MRR} for a def\/inition and a~list of basic properties
of the Tutte polynomial of a~graph.
\begin{Definition} \label{definition4.4}
Let $r \ge 2 $ be a positive integer and $ \{S_1, \ldots,S_r \}$
be a~collection of sets of cardinalities $\#|S_j|= n_j$, $j=1,\ldots,r$. Let
${\boldsymbol{\ell}}:= \{\ell_{ij} \}_{1 \le i < j \le n}$ be a collection of non-negative integers.
The ${\boldsymbol{\ell}}$-weighted complete multipartite graph
$K_{n_1,\ldots,n_r}^{({\boldsymbol{\ell}})} $ is a graph with the set of vertices equals
to the disjoint union $\coprod\limits_{j=1}^{r} S_i$ of the sets $S_1,\ldots,S_r$,
and the set of edges $\{ (\alpha_i, \beta_j ), \, \alpha_i \in S_i, \, \beta_j \in
S_j \}_{1 \le i < j \le r}$ of multiplicity $\ell_{ij}$ each edge~$(\alpha_i,
\beta_j)$.
\end{Definition}
\begin{Theorem} \label{theorem4.3}
Let us fix an integer $r \ge 2$ and a collection of
non-negative integers ${\boldsymbol{\ell}}:= \{\ell_{ij} \}_{1 \le i < j \le r}$.
Then
\begin{gather*}
1 + \sum_{{\boldsymbol{n}}=(n_1,\ldots,n_r) \in \mathbb{Z}_{\ge 0}^{r} \atop {\boldsymbol{n}}
\not= {\boldsymbol{0}}} (x-1)^{\kappa({\boldsymbol{\ell}},{\boldsymbol{n}})} {\rm Tutte}\big(K_{n_1,\ldots,n_r}^{({\boldsymbol{\ell}})},x,y\big)
{{t_1^{n_1} \over n_1 !}} \cdots {{t_r^{n_r} \over n_r !}} \\
\qquad{}= \left( \sum_{{\boldsymbol{m}}=(m_1,\ldots,m_r) \in \mathbb{Z}_{\ge 0}^r}
y^{ \sum\limits_{1 \le i < j \le r} \ell_{ij} m_i m_j}
(y-1)^{- |{\boldsymbol{m}}|} {{t_1^{m_1} \over m_1 !}} \cdots {{t_r^{m_r} \over m_r !}}
\right)^{(x-1)(y-1)},
\end{gather*}
where $\kappa({\boldsymbol{\ell}},{\boldsymbol{n}})$ denotes the number of connected components
of the graph $K_{n_1,\ldots,n_r}^{({\boldsymbol{\ell}})}$.
\end{Theorem}
\begin{Comments}\quad
\begin{enumerate}\itemsep=0pt
\item[(a)] Clearly the condition $\ell_{ij}=0$ means that there are no edges
between vertices from the sets~$S_i$ and~$S_j$. Therefore Theorem~\ref{theorem4.3} allows
to compute the Tutte polynomial of {\it any} (f\/inite) graph. For
example,
\begin{gather*}
{\rm Tutte}\big(K_{2,2,2,2}^{(1^{6})},x,y\big) = \big\{(0,362,927,911,451,121,17,1)_{x},\\
\hphantom{{\rm Tutte}\big(K_{2,2,2,2}^{(1^{6})},x,y\big) = }{}
(362,2154,2928,1584,374,32)_{x},
(1589,4731,3744,1072,96)_{x},\\
\hphantom{{\rm Tutte}\big(K_{2,2,2,2}^{(1^{6})},x,y\big) = }{}
(3376,6096,2928,448,16)_{x},
(4828,5736,1764,152)_{x}, \\
\hphantom{{\rm Tutte}\big(K_{2,2,2,2}^{(1^{6})},x,y\big) =}{}
(5404,4464,900,32)_{x},
(5140,3040,380)_{x}, (4340,1840,124)_{x},\\
\hphantom{{\rm Tutte}\big(K_{2,2,2,2}^{(1^{6})},x,y\big) =}{}
(3325,984,24)_{x},
(2331,448)_{x},
(1492,168)_{x}, (868,48)_{x}, (454,8)_{x},\\
\hphantom{{\rm Tutte}\big(K_{2,2,2,2}^{(1^{6})},x,y\big) =}{}
210, 84, 28, 7, 1\big\}_{y}.
\end{gather*}
\item[(b)] One can show that a formula for the chromatic polynomials from
Proposition~\ref{proposition4.2} corresponds to the specialization $y=0$ ({\it but} not
direct substitution!) of the formula for generating function for the Tutte
polynomials stated in Theorem~\ref{theorem4.3}.
\item[(c)] The Tutte polynomial ${\rm Tutte}\big(K_{n_1,\ldots,n_{r}}^{({\boldsymbol{\ell}})},x,y\big)$ does
not symmetric with respect to parameters $\{ \ell_{ij} \}_{1 \le i < j
\le r}$.
For example, let us write ${\boldsymbol{\ell}}= (\ell_{12},\ell_{23},\ell_{13},\ell_{14},\ell_{24},\ell_{34})$, then
\begin{gather*}
{\rm Tutte}\big(K_{2,2,2,2}^{(6,3,4,5,2,4)},1,1\big)=
2^8 \cdot 3 \cdot 5 \cdot 11^3 \cdot 241= 1231760640 .
\end{gather*}
On the other hand,
\begin{gather*}
{\rm Tutte}\big(K_{2,2,2,2}^{(6,4,3,5,2,4)},1,1\big)=2^{13} \cdot 3 \cdot 7 \cdot 11^2
\cdot 61 = 1269768192.
\end{gather*}
\subsubsection[Universal Tutte polynomials]{Universal Tutte polynomials}\label{section4.1.3add}
Let ${\boldsymbol{m}}=(m_{ij}, \, 1 \le i < j \le n)$ be a collection of non-negative
integers. Def\/ine {\it generalized Tutte polynomial}
${\widetilde{T}}_n({\boldsymbol{m}},x,y)$ as
follows
\begin{gather*}
(x-1)^{\kappa(n,{\boldsymbol{m}})} {\widetilde{T}}_n({\boldsymbol{m}},x,y)\\
\qquad {} =
{\rm Coef\/f}_{[t_1 \cdots t_n]} \left[ \left( \sum_{\ell_1,\ldots,\ell_n
\atop \ell_i \in \{0,1 \}, \, \forall\, i} y^{\sum\limits_{1 \le i < j \le n} m_{ij} \ell_i \ell_j}
(y-1)^{- (\sum_{j} \ell_{j})} t_1^{\ell_{1}} \cdots t_n^{\ell_{n}} \right)^{(x-1)(y-1)} \right] ,
\end{gather*}
where as before, $\kappa(n, {\boldsymbol{m}})$ denotes the number of connected components of the graph $K_n^{({\boldsymbol{m}})}$.
\end{enumerate}
Clearly that if $\Gamma \subset K_n^{({\boldsymbol{\ell}})}$ is a subgraph of the
weighted complete graph $K_n^{({\boldsymbol{\ell}})}\stackrel{\rm def}{=}
K_{1^n}^{({\boldsymbol{\ell}})}$, then the Tutte polynomial of graph $\Gamma$
multiplied by $(x-1)^{\kappa(\Gamma)}$ is equal to the following
specialization
\begin{gather*}
m_{ij}=0 \quad \text{if edge} \ \ (i,j) \notin \Gamma,\qquad m_{ij}= \ell_{ij} \quad \text{if
edge} \ \ (i,j) \in \Gamma
\end{gather*}
of the generalized Tutte polynomial
\begin{gather*}
(x-1)^{\kappa(\Gamma)} {\rm Tutte}(\Gamma,x,y)= {\widetilde{T}}_n({\boldsymbol{m}},x,y)
\Big |_{m_{ij} =0 \ \ \text{if} \ (i,j) \notin \Gamma \atop m_{ij}=\ell_{ij} \ \text{if} \ (i,j) \in \Gamma}.
\end{gather*}
For example,
\begin{enumerate}\itemsep=0pt
\item[(a)] Take $n=6$ and $\Gamma= K_6 \backslash \{ 15,16,24,25,34,36 \}$,
then
\begin{gather*}
{\rm Tutte}(\Gamma,x,y)= \{(0,4,9,8,4,1)_{x}, (4,13,9)_{x}, (8,7)_{x}, 5, 1 \}_{y}.
\end{gather*}
\item[(b)] Take $n=6$ and $\Gamma =K_6 \backslash \{15,26,34 \}$, then
\begin{gather*}
{\rm Tutte}(\Gamma,x,y)= \{(0,11,25,20,7,1)_{x}, (11,46,39,8)_{x},
(32,52,12)_{x},\\
\hphantom{{\rm Tutte}(\Gamma,x,y)=}{}
(40,24)_{x}, (29,6)_{x}, 15, 5, 1 \}_{y}.
\end{gather*}
\item[(c)] Take $n=6$ and $\Gamma=K_6 \backslash \{12.34.56 \} = K_{2,2,2}$. As a~result one obtains an expression for the Tutte polynomial of the graph $K_{2,2,2}$ displayed in Example~\ref{examples4.16}(4).
\end{enumerate}
\end{Comments}
Now set us set
\begin{gather*}
q_{ij}:=\frac{y^{m_{ij}}-1}{y-1}.
\end{gather*}
\begin{Lemma}\label{lemma4.1}
The generalized Tutte polynomial ${\widetilde{T}}_n({\boldsymbol{m}},x,y)$
is a~polynomial in the variables $\{ q_{ij} \}_{1 \le i < j \le n}$, $x$ and~$y$.
\end{Lemma}
\begin{Definition} \label{definition4.5}
The universal Tutte polynomial $T_n(\{q_{ij}\},x,y)$ is def\/ined to be the polynomial in the variables $\{q_{ij} \}$, $x$, and~$y$ def\/ined in
Lemma~\ref{lemma4.1}.
\end{Definition}
Explicitly,
\begin{gather*}
(x-1) T_n(\{q_{ij} \},x,y)\\
{}=
{\rm Coef\/f}_{[t_1 \cdots t_n]} \! \left[\Bigg( \sum_{\ell_1,\ldots,\ell_n
\atop \ell_i \in \{0,1 \}, \, \forall\, i} \prod_{1 \le i < j \le n}
(q_{ij} (y-1)+1)^{\ell_i \ell_j}
(y-1)^{- (\sum_{j} \ell_{j})} t_1^{\ell_{1}} \cdots t_n^{\ell_{n}} \!\Bigg)^{\!(x-1)(y-1)} \right] .
\end{gather*}
\begin{Corollary}\label{corollary4.2}
Let $\{ m_{ij} \}_{1 \le i < j \le n}$ be a collection of
positive integers. Then the specialization
\begin{gather*}
q_{ij} \longrightarrow [m_{ij}]_{y}:= \frac{y^{m_{ij}}-1}{y-1}
\end{gather*}
of the universal Tutte polynomial $T_n(\{q_{ij} \},x,y)$ is equal to the
Tutte polynomial of the complete graph~$K_n$ with each edge~$(i,j)$ of the
multiplicity~$m_{ij}$.
\end{Corollary}
Further specialization $ q_{ij} \longrightarrow 0$ if edge $(i,j) \notin \Gamma$
allows to compute the Tutte polynomial for any graph
\begin{gather*}
{\rm Tutte}_{3}(\{q_{12},q_{13},q_{23} \},x,y) = (1 -
q[12])(1-q[13])(1-q[23]) +
y q[12] q[13] q[23])\\
\hphantom{{\rm Tutte}_{3}(\{q_{12},q_{13},q_{23} \},x,y) =}{} + x (q[12] + q[13] + q[23] -2 )+ x^2 .
\end{gather*}
\begin{center}
\framebox{\parbox[t]{6 in}{ It is not dif\/f\/icult to see that the
${\rm Tutte}_{n}(\{ q_{ij} \},x,y)$ is a {\it symmetric polynomial} with
respect to parameters $\{ q_{ij}\}_{1 \le i < j \le n}$. }}
\end{center}
For more compact expression, it is more
convenient to rewrite the universal chromatic polynomial in terms of parameters $p_{ij}:= 1-q_{ij},~1 \le i < j \le n$, and denote it by
$Ch_n(\{p_{ij}\},x)$.
For example,
\begin{gather*}
{\rm Ch}_{4}(\{ p_{ij}\},x)= - p_{12} p_{13} p_{14} p_{23} p_{24} p_{34} + x \bigl(2 - p_{12} - p_{13} - p_{14} - p_{23}- p_{24}- p_{34}\\
\hphantom{{\rm Ch}_{4}(\{ p_{ij}\},x)=}{} + p_{12} p_{34} + p_{14} p_{23} + p_{13} p_{24} +p_{12} p_{13} p_{23} + p_{12} p_{14} p_{24}+ p_{13} p_{14} p_{34}\\
\hphantom{{\rm Ch}_{4}(\{ p_{ij}\},x)=}{} + p_{23} p_{24} p_{34} \bigr) +
x^2 \bigl(3 - p_{12} - p_{13} - p_{14} - p_{23} - p_{24} - p_{34} \bigr) + x^3.
\end{gather*}
Note that $p_{12} p_{34} + p_{14} p_{23} + p_{13} p_{24}$
is a symmetric polynomial of the variables $ p_{12}$, $p_{34}$, $p_{13}$, $p_{24}$,
$p_{14}$,
$p_{23}$. It is important to keep in mind that parameters $\{ m_{ij}\}$ and
$\{ p_{ij}\}$ are connected by relations
\begin{gather*}
p_{ij}= {\frac{y- y^{m_{ij}}}{y-1}}, \qquad 1 \le i < j \le n.
\end{gather*}
Therefore, $p_{ij}=1$ if $(i,j) \notin {\rm Edge}(\Gamma)$, $p_{ij}=0$ if $m_{ij}=1$. We emphasize that the latter equalities are valid for arbitrary~$y$.
It is not dif\/f\/icult to see that
\begin{gather*}
{\rm Ch}_n(\{q_{ij}=0, \, \forall\, i,j\}= {\rm Tutte}(K_n;x, 0), \qquad {\rm Ch}_n(\{q_{ij}=1, \,
\forall\, i,j\} = (x-1)^{n-1}.
\end{gather*}
\begin{center}
\framebox{\parbox[t]{6 in}{Def\/ine {\it universal chromatic}
polynomial to be ${\rm Ch}_{n}(\{p_{ij}\},x)={\rm Tutte}_{n}(\{p_{ij} \},x,0)$,
where we treat $\{ p_{ij} \}_{1 \le i < j \le n}$ as a collection of a free parameters. }}
\end{center}
To state our result concerning the universal chromatic polynomial
${\rm Ch}_n(\{p_{ij} \},x)$, f\/irst we introduce a bit of notation. Let $n \ge 2$ be an integer, consider a partition ${\cal{B}}=\big\{B_{i}= \big(b_{1}^{(i)},\ldots,b_{r_{i}}^{(i)}\big)\big\}_{1 \le i \le k}$ of the set $[1,n]:=[1,2,\ldots,n]$. In other
words one has that $[1,n] = \cup_{i=1}^{k} B_{i}$ and $B_i \cap B_j = \varnothing$ if $i \not= j$. We assume that $b_{1}^{(1)} < b_{1}^{(2)} < \cdots < b_{1}^{(k)}$. We def\/ine $\kappa({\cal{B}}):=k$. To a~given partition
${\cal{B}}$ we associate a monomial $p_{\cal{B}}:= \prod\limits_{a=1}^{k}~p_{B_{a}}$, where $p_{B_{a}}=1$ if $\kappa({\cal{B}})=1$, and
\begin{gather*}
p_{B_{a}}= \prod_{i,j \in B_{a} \atop i < j} p_{ij}.
\end{gather*}
For a given partition $\lambda \vdash n$ denote by
${\cal{L}}_{\lambda}^{(\beta)}(\{p_{ij}\})$ the sum of all monomials
$p_{{\cal{B}}} \beta^{\kappa({\cal{B}}) -2 }$ such that $\lambda= \lambda({\cal{B}}) \stackrel{\rm def}{=} (|B_{1}|,\ldots,|B_{\kappa({\cal{B}})}|)^{+}$, where
for any composition $\alpha \models n$, $\alpha^{+}$ denotes a unique
partition obtained from $\alpha$ by the reordering of its parts.
Def\/ine $\beta$-universal chromatic polynomial to be
\begin{gather*}
{\rm Ch}_n^{(\beta)}(\{p_{ij}\},x) = \beta^{-1} {\cal{L}}_{(n)} + \sum_{\lambda \vdash n}
{\rm Tutte}(K_{\ell(\lambda)-1};x,0) {\cal{L}}_{\lambda}^{(\beta)},
\end{gather*}
where summation runs over all partitions $\lambda$ of $n$; we set
$K_{0}:= \varnothing$ and ${\rm Tutte}(\varnothing;x,y)=0$.
For the reader convenience we are reminded that for the complete graph~$K_n$,
$n > 0$, one has
\begin{gather*}
{\rm Tutte}(K_n,x,0)= \prod_{j=1}^{n-1} (x+j-1) = \sum_{k=0}^{n-1} s(k,n-1)x^{k},
\end{gather*}
where $s(k,n)$ denotes the Stirling number of the f\/irst kind\footnote{See, e.g., \cite[$A008275$]{SL} or
\url{https://en.wikipedia.org/wiki/Stirling_numbers_of_the_first_kind}.\label{footnote37}}.
\begin{Theorem}[formula for universal chromatic
polynomials]\label{theorem4.29}
\begin{gather*}
{\rm Ch}_{n}(\{p_{i,j}\},x) = {\rm Ch}_{n}^{(\beta = - 1)}(\{p_{ij}\},x).
\end{gather*}
\end{Theorem}
For a given partition $\lambda \vdash n$ denote by
${\cal{L}}_{\lambda}(\{p_{ij}\})$ the sum of all monomials $p_{{\cal{B}}}$
such that $\lambda= \lambda({\cal{B}}) \stackrel{\rm def}{=} (|B_{1}|,\ldots,|B_{\kappa({\cal{B}})}|)^{+}$, where for any composition
$\alpha \models n$, $\alpha^{+}$ denotes a unique partition obtained from~$\alpha$ by the reordering of its parts.
It is clear that for a graph $\Gamma \subset K_n$ and partition ${\cal{B}}$ the
value of monomial $p_{{\cal{B}}}$ under the specialization $p_{ij}= 0$ if $(ij) \in {\rm Edge}(\Gamma)$ and $p_{ij}=1$ if $(ij) \notin {\rm Edge}(\Gamma)$, is equal to~$1$ i\/f\/f the complementary graph $K_n {\setminus} \Gamma$ contains a
subgraph which is isomorphic to the disjoint union of complete graphs
$K(\lambda):= \coprod\limits_{i=1}^{k} K_{\lambda_{i}}$, where $(\lambda_1,\ldots,\lambda_k)= \lambda({\cal{B}})$. Therefore the specialization
\begin{gather*}
{\cal{L}}_{\lambda} \big\vert_{p_{ij}=0, \ (ij) \in \Gamma, \atop p_{ij}=1, \ (ij) \notin \Gamma}
\end{gather*}
is equal to the number of non isomorphic subgraphs of the complementary graph
$ K_n {\setminus} \Gamma$ which are isomorphic to the graph $K(\lambda)$.
\begin{Example} Take $n=6$, then
\begin{gather*}
{\rm Ch}_{6}^{(\beta)}= \beta^{-1} {\cal{L}}_{(6)}+x(x+1)(x+2)(x+3)(x+4) {\cal{L}}_{(1^{6})} \beta^4\\
\hphantom{{\rm Ch}_{6}^{(\beta)}=}{}
+
x(x+1)(x+2)(x+3){\cal{L}}_{(2,1^{4})} \beta^3 + x(x+1)(x+2) \bigl(
{\cal{L}}_{(2^{2},1^{2})} +{\cal{L}}_{(3,1^{3})} \bigr) \beta^2 \\
\hphantom{{\rm Ch}_{6}^{(\beta)}=}{}
+ x(x+1) \bigl( {\cal{L}}_{(2^3)} +{\cal{L}}_{(3,2,1)}+{\cal{L}}_{(4,1^2)} \bigr) \beta +
{\cal{L}}_{(3^2)}+{\cal{L}}_{(4,2)}+{\cal{L}}_{(5,1)}.
\end{gather*}
Since $p_{ij}$ is
equal to either~$1$ or $0$, one can see that ${\cal{L}}_{(n)}=0$ unless graph
$\Gamma$ is a collection of~$n$ distinct points and therefore~${\cal{L}}=1$.
\end{Example}
\begin{center}
\framebox{\parbox[t]{6 in}{The chromatic polynomial of any graph is a $\mathbb{Z}$-linear combination of the chromatic polynomials corresponding to a set of complete graphs. }}
\end{center}
\begin{Corollary}[formula for universal $\beta$-Tutte polynomials\footnote{It should be remembered that ${\rm Tutte}(K_{1};x,y)=1 $ and
${\rm Tutte}(K_{0};x,y)=0$, since the graph $K_1:=\{pt\}$ and graph $K_{0}= \varnothing$.\label{footnote38}}]
\begin{gather*}
(1-y)^{n-1}{\rm Tutte}_{n}^{(\beta)}(\{p_{i,j}\};x,y) \\
\qquad {} = \prod_{1
\le i < j \le n} p_{ij} + \sum_{\lambda \vdash n}
{\rm Tutte} (K_{\ell(\lambda)-1}; x+y+\beta x y,0) {\cal{L}}_{\lambda}^{(\beta)}(\{p_{ij}\}) .
\end{gather*}
\end{Corollary}
\begin{center}
\framebox{\parbox[t]{6in}{The polynomial $(1-y)^{|V(\Gamma)|-1}{\rm Tutte}(\Gamma;x,y)$ is a
$\mathbb{Z}[y]$-linear combination of the chromatic polynomials
${\rm Tutte}(K_m;x+y-x y,0) $ corresponding to a family of complete graphs~$\{K_m\}$.}}
\end{center}
Here $V(\Gamma)$ denotes the set of vertices of graph~$\Gamma$.
\begin{Comments}\quad
\begin{enumerate}\itemsep=0pt
\item[$(i)$] Let us write
\begin{gather*}
{\rm Ch}_{n}^{(\beta)}(\{p_{ij} \},x) = - {\cal{L}}_{(n)} \beta^{-1} + \sum_{k=1}^{n-1} a_{n}^{(k)}(\{p_{ij}) x^{k}.
\end{gather*}
It follows from Theorem~\ref{theorem4.29} that
\begin{gather*}
a_{n}^{(k)} = \sum_{\lambda \vdash n} s(\ell(\lambda)-1,k) {\cal{L}}_{\lambda}^{(\beta)},
\end{gather*}
where as before, $s(k,n)$ denote the Stirling numbers of the f\/irst kind,
see, e.g., footnote~\ref{footnote37}. For example,
\begin{gather*}
a(\Gamma) = a_{n}^{(1)}(\{p_{ij}\}) \bigg\vert_{p_{ij}=0,\, (ij) \in \Gamma,
\atop p_{ij}=1, \, (ij) \notin \Gamma} (\ell(\lambda)-2) ! {\cal{N}}_{\lambda}(\Gamma)
\beta^{\ell(\lambda)-2},
\end{gather*}
where ${\cal{N}}_{\lambda}(\Gamma)$ denotes the number non isomorphic
subgraphs in the complementary graph $K_n {\setminus}\Gamma$, which are isomorphic
to the graph~$ K_{n}(\lambda)$.
\item[$(ii)$] It is clear that for a general set of parameters $\{p_{ij}\}$ the
number of dif\/ferent monomials which appear in ${\cal{L}}_{\lambda}^{(\beta)}
(\{p_{ij}\})$, where partition $\lambda= \sum\limits_{j=1}^{n} j m_j$, $\lambda
\vdash n$, is equal to
\begin{gather*}
\frac{n !}{\prod\limits_{j \ge 1}(j !)^{m_{j}} m_{j} !}.
\end{gather*}
\item[$(iii)$] For general set of parameters $\{p_{ij}\}$ one can show
that the number of dif\/ferent monomials which appear in polynomial
$a_{n}^{(1)}(\{p_{ij}\})$ is equal to ${\rm Bell}(n) -1$, where ${\rm Bell}(n)$ denotes
the $n$-th Bell number, see, e.g., \cite[$A000110$]{SL}.
\item[$(iv)$] In the limit $y \longrightarrow 1$ one has $q_{ij}= m_{ij}$ and
$p_{ij}= 1-m_{ij}$.
\item[$(v)$] Let us introduce a modi\/f\/ied universal Tutte polynomial, namely,
\begin{gather*}
{\rm Tutte}(\{q_{ij}\};x,y,z):=
(-1)^{n-1} {\rm Coef\/f}_{[t_1 \cdots t_n]} \\
\qquad {}\times \Biggl[\Biggl( \sum_{\ell_1,\ldots,
\ell_n \atop \ell_i \in \{0,1 \}, \, \forall \,i} \prod_{1 \le i < j \le n}
(z q_{ij} y+1)^{\ell_i \ell_j}
y^{- (\sum_{j} \ell_{j})} t_1^{\ell_{1}} \cdots t_n^{\ell_{n}} \Biggr)^{ x y} x^{-1} \Bigg] .
\end{gather*}
\end{enumerate}
\end{Comments}
We set $\deg(q_{ij})=1$,
\begin{Proposition}\quad
\begin{enumerate}\itemsep=0pt
\item[$(a)$] ${\rm Tutte}(\{q_{ij}\};x,y,z) \in \mathbb{N}[\{q_{ij}] [x,y,z]$.
\item[$(b)$] Degree $n-1$ monomials of the polynomial ${\rm Tutte}(\{q_{ij}\};0,y,z)$
are in one-to-one correspondence with the set of spanning trees of the
complete graph~$K_n$. Moreover, the polynomial ${\rm Tutte}(\{q_{ij}=1,\, \forall\, i,j \};x,0,1)$ is equal to the generating function of forests on~$n$ labeled
vertices, counting according to the number of connected components,~whereas
the polynomial ${\rm Tutte}(\{q_{ij}=1,\, \forall\, i,j \};1,0,z)$ is equal to the
Hilbert polynomial of the even Orlik--Solomon algebra\footnote{Known also as Orlik--Terao algebra.}
${\rm OS}^{+}(\Gamma_n)$ associated to the type $A_{n-1}$ {\it
generic hyperplane arrangement} $\Gamma_n$, see {\rm \cite[Section~5]{PSt}} or~{\rm \cite{K}}, namely,
\begin{gather*}
{\rm Tutte}(\{q_{ij}=1,\, \forall \, i,j \};1,0,z)= {\rm Hilb}({\rm OS}^{+}(\Gamma_{n}),z)=
\sum_{{\cal{F}}} z^{|{\cal{F}} |},
\end{gather*}
where the sum runs over all forests on the vertices $\{1,\ldots,n\}$, and $|{\cal{F}}|$ denotes the number of edges of~${\cal{F}}$.
\item[$(c)$] More generally, denote by $F_{n}(x,t):= \sum\limits_{{\cal{F}}} x^{|{\cal{F}}|} t^{{\rm inv} ({\cal{F}})}$ the generating function of statistics $|{\cal{F}}|$
and ${\rm inv} ({\cal{F}})$ on the set $F(n)$ of forests on~$n$ labeled vertices.
Recall that the symbol $|{\cal{F}}|$ denotes the number of {\it edges} in a
forest ${\cal{F}} \in F(n)$ and that ${\rm inv} ({\cal{F}})$ its inversion index\footnote{For the readers convenience we recall def\/initions of statistics
${\rm inv} ({\cal{F}})$ and the major index ${\rm maj} ({\cal{F}})$.
Given a forest ${\cal{F}}$ on $n$ labeled nodes, one can
construct a~tree ${\cal{T}}$ by adding a~new vertex (root) connected with
the maximal vertices in the connected components of~${\cal{F}}$.
The inversion index ${\rm inv} ({\cal{F}})$ is equal to the number of pairs~$(i,j)$
such that $1 \le i < j \le n$, and the vertex labeled by~$j$ lies on the
shortest path in~${\cal{T}}$ from the vertex labeled by~$i$ to the root.
The major index ${\rm maj} ({\cal{F}})$ is equal to $\sum\limits_{x \in {\rm Des}({\cal{F}})}
h(x)$; here for any vertex $x \in {\cal{F}}$, $h(x)$ is the size of the
subtree rooted at~$x$; the descent set ${\rm Des}({\cal{F}})$ of~${\cal{F}}$
consists of the vertices $x \in {\cal{F}}$ which have
the labeling strictly greater than the labeling of its child's.}.
\end{enumerate}
\begin{Lemma} One can show that
\begin{gather*}
F_{n}(x,t)= (x t)^{n-1} {\rm Tutte}(K_n; 1+ (x t)^{-1},t-1),\\
{\rm Coef\/f}_{(x t)^{n-1}} [ F_{n}(x,t) ] =I_{n}(t),
\end{gather*}
where $I_{n}(t):= \sum_{{\cal{F}} \in {\rm Tree}(n)} t^{{\rm inv}({\cal{F}})}$ denotes the
tree inversion polynomial, see, e.g., {\rm \cite{GS,ST1}}.
\end{Lemma}
\begin{enumerate}\itemsep=0pt
\item[$(d)$] Set
\begin{gather*}
DU_n(x):= (z t )^{n-1} {\rm Hilb}\big(K_n; 1+(z t)^{-1},z-1\big) \Big |_{t:=-1 \atop z:= -x} = F_{n}(-x,-1).
\end{gather*}
One $($A.K.$)$ can show that $(n \ge 2)$
\begin{gather*}
DU_n(x) \in \mathbb{N} [x], \qquad DU_{n}(1) =UD_{n+1}, \qquad {\rm Coef\/f}_{x^{n-1}}[DU_n(x)] = UD_{n-1},
\end{gather*}
where $UD_n$ denote the Euler or up/down numbers associated with the
exponential ge\-ne\-ra\-ting function $\sec(x) + \tan(x)$, see\footnote{The fact that $I_n(-1)= UD_{n-1}$ is due to G.~Kreweras~\cite{Kre}.},
e.g., {\rm \cite[$A000111$]{SL}}.
\item[$(e)$] One has
\begin{gather*}
x^{{n\choose 2}} {\rm Tutte}\big(\{q_{ij}=1,\,\forall \, i,j \};x,x^{-1}-1,1\big)=
{\rm Hilb}({\cal{A}}_{n},x),
\end{gather*}
where ${\cal{A}}_n$ denotes the algebra generated by the curvature of $2$-forms of the standard Hermitian linear bundles over the flag variety ${\cal{F}}l_n$,
see {\rm \cite{K,PSS,SS}} or Section~{\rm \ref{section4.2.2}}, Theorem~{\rm \ref{theorem4.4}$(B)$}.
\item[$(f)$] Write ${\rm Tutte}(\{q_{ij}\};0,y,z)=\sum\limits_{k=0}^{n-1} a_{n}^{(k)}(y,z)$,
then monomials which appear in polynomial $a_{n}^{(k)}(y,z)$ are in
one-to-one correspondence with the set of labeled graphs with n nodes having
exactly~$k$ connected components.
\item[$(g)$] One has ${\rm Tutte}((\{q_{ij}\};x,-1,1)= {\rm Tutte}(\{q_{ij}\},x+1,0)$.
\item[$(h)$] Recurrence relations for polynomials $F_n(x,t)$, cf.~{\rm \cite{Kre}},
\begin{gather*}
F_0(x,t)=F_1(x,t)=1, \qquad F_{n+1}(x,t)= \sum_{k=0}^{n} {n \choose k} (xt)^{k} I_k(t) F_{n-k}(x,t).
\end{gather*}
\end{enumerate}
\end{Proposition}
\begin{Example}
Take $n=5$, then
\begin{gather*}
{\rm Tutte}(K_5;x,y)= (0,6,11,6,1)+(6,20, 10) y+ 15(1,1) y^2+ 5(3,1) y^3 +10 y^4+4 y^5+y^6,\\
F_5(-x,- 1)= (1,10,25,20,5).
\end{gather*}
Write $F_n(x,t)= {\tilde{F}}_{n}(u,t) \big |_{u=xt}$,
then
\begin{gather*}
{\tilde{F}}_{5}(u,t) = 1+10 u+ u^2 (35+10 t)+ u^3 (50,40,15,5)_{t}+
u^4 (24,36,30,20,10,4,1)_{t},\\
{\tilde{F}}_{n}(u,0)= \prod_{j=1}^{n-1}(1+ j~u),\\
{\rm Hilb}({\cal A}_5,t)=(1,4,10,20,35,51,64,60,35,10,1)_{t},\\
{\rm Hilb}({\rm OS}^{+}(\Gamma_5),t)= (1,10,45,110,125)_{t}.
\end{gather*}
\end{Example}
\begin{Exercises}\label{exercises4.2}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] Assume that $\ell_{ij}=\ell$ for all $1 \le i < j \le r$. Based on
the above formula for the exponential generating function for the Tutte
polynomials of the complete multipartite graphs $K_{n_1,\ldots,n_r}$,
{\it deduce} the following well-known formula
\begin{gather*}
{\rm Tutte}\big(K_{n_1,\ldots,n_r}^{(\ell)},1,1\big) = \ell^{N-1} N^{r-2} \prod_{j=1}^{r} (N-n_j)^{n_j-1},
\end{gather*}
where $N:=n_1+\cdots + n_r$. It is well-known that the number
${\rm Tutte}(\Gamma,1,1)$ is equal to the number of {\it spanning trees} of a~connected graph $\Gamma$.
\item[(2)] Take $r=3$ and let $n_1$, $n_2$, $n_3$ and $\ell_{12}$, $\ell_{13}$, $\ell_{23}$ be positive integers. Set $N:= \ell_{12} \ell_{13} n_1 + \ell_{12}
\ell_{23} n_2 +\ell_{13} \ell_{23} n_3$. Show that
\begin{gather*}
{\rm Tutte}\big(K_{n_1,n_2,n_3}^{\ell_1,\ell_2,\ell_3}, 1,1\big) =
N (\ell_{12} n_2\!+\ell_{13} n_3)^{n_{1}-1} (\ell_{12} n_1\!+\ell_{13} n_3)^{n_{2}-1} (\ell_{13} n_1\!+ \ell_{23} n_2)^{n_{3}-1}.
\end{gather*}
\item[(3)] Let $r \ge 2$, consider weighted complete multipartite graph
$K_{ \underbrace{n,\ldots,n}_{r}}^{({\boldsymbol{\ell}})}$, where ${\boldsymbol{\ell}}= (\ell_{ij})$ such that $\ell_{1,j}=\ell$, $j=1,\ldots, r$ and
$\ell_{ij}= k$, $2 \le i < j \le r$. Show that
\begin{gather*}
{\rm Tutte}\big(K_{ \underbrace{n,\ldots,n}_{r}}^{({\boldsymbol{\ell}})},1,1\big) = k^{n}
(r-1)^{n-1} ((r-1) \ell + k )^{r-2} ((r-2) \ell +k )^{(r-1)(n-1)} n^{nr-1}.
\end{gather*}
\end{enumerate}
\end{Exercises}
Let $\Gamma_{n}(*)$ be a spanning star subgraph of the complete
graph~$K_n$. For example, one can take for a graph $\Gamma_{n}(*)$ the
subgraph $K_{1,n-1}$ with the set of vertices $V:=\{1,2,\ldots,n \}$
and that of edges $E:=\{(i,n),\, i=1,\ldots,n-1 \}$. The algebra
$3T_n^{(0)}(K_{1,n-1})$ can be treated as a~``noncommutative analog''
of the projective space $\mathbb{P}^{n-1}$.
We have $\theta_1 = u_{12}+u_{13}+\cdots+u_{1n}$. It is not dif\/f\/icult
to see that
${\rm Hilb}(3T_n^{(0)}(K_{1,n-1})^{ab},t)=(1+t)^{n-1}$, and $\theta_1^n=0$.
Let us observe that ${\rm Chrom}(\Gamma_n(\star),t)= t (t-1)^{n-1}$.
\begin{Problem} \label{problem4.1} Compute the Hilbert series of the algebra
$3T_n^{(0)}(K_{n_1,\ldots,n_r})$.
\end{Problem}
The f\/irst non-trivial case is that of {\it projective space}, i.e.,
the case $r=2$, $n_1=1$, $n_2=5$.
On the other hand, if $\Gamma_n= \{(1,2) \rightarrow (2,3) \rightarrow \cdots
\rightarrow (n-1,n) \}$ is the Dynkin graph of type~$A_{n-1}$, then the
algebra $3T_n^{(0)}(\Gamma_n)$ is isomorphic to the nil-Coxeter algebra of
type $A_{n-1}$, and if $\Gamma_n^{\rm (af\/f)}= \{ (1,2) \rightarrow (2,3)
\rightarrow \cdots \rightarrow (n-1,n) \rightarrow -(1,n) \}$ is the Dynkin
graph of type~$A_{n-1}^{(1)}$, i.e., {\it a~cycle}, then the algebra
$3T_n^{(0)}(\Gamma_n^{\rm (af\/f)})$ is isomorphic to a certain quotient of the
af\/f\/ine nil-Coxeter algebra of type $A_{n-1}^{(1)}$ by the two-sided ideal
which can be described explicitly~\cite{K}. Moreover~\cite{K},
\begin{gather*}
{\rm Hilb}\big(3T_n^{0)}\big(\Gamma^{\rm (af\/f)}\big),t\big)= [n]_t \prod_{j=1}^{n-1} [j(n-j)]_t,
\end{gather*}
see Theorem~\ref{theorem4.1}. Therefore, the dimension $\dim(3T^{(0)}(\Gamma^{\rm af\/f}))$ is
equal to $n ! (n-1) !$ and is equal also, as it was pointed out in
Section~\ref{section4.1.1}, to the number of (directed)
Hamiltonian cycles in the complete bipartite graph~$K_{n,n}$, see \cite[$A010790$]{SL}.
It is not dif\/f\/icult to see that
\begin{gather*}
{\rm Hilb}\big(3T_n^{(0)}(\Gamma_n)^{ab},t\big)=(t+1)^{n-1},
\qquad
{\rm Hilb}\big(3T^{(0)}\big(\Gamma_n^{\rm af\/f}\big)^{ab},t\big)=t^{-1} \big((t+1)^{n}-t-1\big),
\end{gather*}
whereas
\begin{gather*}
{\rm Chrom}(\Gamma_n,t)= t (t-1)^{n-1}, \qquad {\rm Chrom}\big (\Gamma_n^{\rm af\/f},t\big)=(t-1)^n+(-1)^n (t-1).
\end{gather*}
\begin{Exercise}\label{Exercise4.3}
Let $K_{n_1,\ldots,n_r}$ be complete multipartite graph,
$N:= n_1+\cdots+n_r$.
Show that\footnote{It should be remembered that to abuse of notation, the complete
graph~$K_n$, by def\/inition, is equal to the complete multipartite graph
$K(\underbrace{(1,\ldots,1)}_{n})$, whereas the graph $K_{(n)}$ is a~collection of~$n$ distinct points.}
\begin{gather*}
{\rm Hilb}(3T_N(K_{n_1,\ldots,n_r}),t) = \frac{\prod\limits_{j=1}^{r} \prod\limits_{a=1}^{n_{j}-1} (1-a t)}{\prod\limits_{j=1}^{N-1}(1- jt)}.
\end{gather*}
\end{Exercise}
\subsubsection[Quasi-classical and associative classical Yang--Baxter algebras of type $B_n$]{Quasi-classical and associative classical Yang--Baxter algebras of type $\boldsymbol{B_n}$}\label{section4.1.3}
In this section we introduce an analogue of the algebra $3T_n(\beta)$ for the
classical root systems.
\begin{Definition}\label{definition4.6}\quad
\begin{enumerate}\itemsep=0pt
\item[(A)] {\it The quasi-classical Yang--Baxter algebra
$\widehat{{\rm ACYB}(B_n)}$ of type~$B_n$} is an associative algebra with the set of
generators $\{x_{ij},\,y_{ij},\,z_i,\,1 \le i \not= j \le n \}$ subject to the set
of def\/ining relations
\begin{enumerate}\itemsep=0pt
\item[(1)] $x_{ij}+x_{ij}=0$, $y_{ij}=y_{ji}$ if $i \not= j$,
\item[(2)] $z_i z_j=z_j z_i$,
\item[(3)] $x_{ij} x_{kl}=x_{kl} x_{ij}$, $x_{ij} y_{kl}=y_{kl} x_{ij}$,
$y_{ij}y_{kl}=y_{kl}y_{ij}$ if $i$, $j$, $k$, $l$ are distinct,
\item[(4)] $z_ix_{kl}=x_{kl}z_i$, $z_iy_{kl}=y_{kl}z_i$ if $i \not= k,l$,
\item[(5)] {\it three term relations}:
\begin{gather*}
x_{ij} x_{jk}=x_{ik} x_{ij}+x_{jk} x_{ik} -\beta x_{ik}, \qquad x_{ij} y_{jk}
=y_{ik} x_{ij}+y_{jk} y_{ik}-\beta y_{ik},
\\
x_{ik} y_{jk}=y_{jk} y_{ij}+y_{ij} x_{ik}+\beta y_{ij},\qquad y_{ik}x_{jk}
=x_{jk}y_{ij}+y_{ij}y_{ik}+\beta y_{ij}
\end{gather*}
if $1 \le i < j < k \le n$,
\item[(6)] {\it four term relations}:
\begin{gather*}
x_{ij} z_j=z_{i} x_{ij}+y_{ij} z_i+z_j y_{ij}-\beta z_{i}
\end{gather*}
if $i < j$.
\end{enumerate}
\item[(B)]
{\it The associative classical Yang--Baxter algebra ${\rm ACYB}(B_n)$ of type~$B_n$} is the special case $\beta=0$ of the algebra
$\widehat{{\rm ACYB}(B_n)}$.
\end{enumerate}
\end{Definition}
\begin{Comments} \label{comments4.2}\quad
\begin{itemize}\itemsep=0pt
\item In the case $\beta=0$ the algebra ${\rm ACYB}(B_n)$ has a
rational representation
\begin{gather*}
x_{ij} \longrightarrow (x_i-x_j)^{-1},\qquad
y_{ij} \longrightarrow (x_i+x_j)^{-1},\qquad
z_i \longrightarrow x_i^{-1}.
\end{gather*}
\item In the case $\beta=1$ the algebra $\widehat{{\rm ACYB}(B_n)}$ has
a ``trigonometric'' representation
\begin{gather*}
x_{ij} \longrightarrow \big(1-q^{x_i-x_j}\big)^{-1},\qquad
y_{ij} \longrightarrow \big(1-q^{x_i+x_j}\big)^{-1},\qquad
z_i \longrightarrow \big(1+q^{x_i}\big)\big(1-q^{x_i}\big)^{-1}.
\end{gather*}
\end{itemize}
\end{Comments}
\begin{Definition}\label{definition4.7} {\it The bracket algebra ${\cal E}(B_n)$ of type $B_n$}
is an associative algebra with the set of generators
$\{x_{ij},\, y_{ij},\, z_i,\, 1 \le i \not= j \le n \}$ subject to the set of
relations (1)--(6) listed in Def\/inition~\ref{definition4.6}, and the additional relations
\begin{alignat*}{3}
& (5a) \quad && x_{jk} x_{ij}=x_{ij} x_{ik}+x_{ik} x_{jk} -\beta x_{ik}, \qquad y_{jk} x_{ij}
=x_{ij} y_{ik}+y_{ik} y_{jk}-\beta y_{ik},&\\
&&& y_{jk} x_{ik}=y_{ij} y_{jk}+x_{ik} y_{ij}+\beta y_{ij}, \qquad x_{jk} y_{ik} =y_{ij} x_{jk}+y_{ik} y_{ij}+\beta y_{ij}&
\end{alignat*}
if $1 \le i < j < k \le n$,
\begin{alignat*}{3}
& (6a) \quad && z_j x_{ij}=x_{ij} z_{i}+z_i y_{ij}+y_{ij} z_j-\beta z_{i}&
\end{alignat*}
if $i < j$.
\end{Definition}
\begin{Definition}\label{definition4.8} The quasi-classical Yang--Baxter algebra
$\widehat{{\rm ACYB}(D_n)}$ of type $D_n$, as well as the algebras
${\rm ACYB}(D_n)$~and~${\cal E}(D_n)$ are def\/ined by putting
$z_i=0$, $i=1,\ldots, n$, in the corresponding $B_n$-versions of
algebras in question.
\end{Definition}
\begin{Conjecture}\label{conjecture4.3}
The both algebras ${\cal E}(B_n)$ and ${\cal E}(D_n) $ are
Koszul, and
\begin{gather*}
{\rm Hilb}({\cal E}(B_n),t)= \left(\prod_{j=1}^{n}(1-(2j-1)t) \right)^{-1};
\end{gather*}
if $n \ge 4$
\begin{gather*}
{\rm Hilb}({\cal E}(D_n),t)=
\left(\prod_{j=1}^{n-1}(1-2j t) \right)^{-1}.
\end{gather*}
\end{Conjecture}
\begin{Example}\label{example4.3}
\begin{gather*}
{\rm Hilb}({\rm ACYB}(B_2),t)=\big(1-4t+2t^2\big)^{-1}, \\
{\rm Hilb}({\rm ACYB}(B_3),t)=\big(1-9t+16t^2-4t^3\big)^{-1}, \\
{\rm Hilb}({\rm ACYB}(B_4),t)=\big(1-16t+64t^2-60t^3+9t^4\big)^{-1} , \\
{\rm Hilb}({\rm ACYB}(D_4),t)=\big(1-12t+18t^2-4t^3\big)^{-1}.
\end{gather*}
However,
\begin{gather*}
{\rm Hilb}({\rm ACYB}(B_5),t)=\big(1-25t+180t^2-400t^3+221t^4-31t^5\big)^{-1}.
\end{gather*}
\end{Example}
Let us introduce the following Coxeter type elements
\begin{gather}\label{equation4.2}
h_{B_n}:= \prod_{a=1}^{n-1} x_{a,a+1} z_n \in {\cal E}(B_n) \qquad \text{and} \qquad
h_{D_n}:= \prod_{a=1}^{n-1} x_{a,a+1} y_{n-1,n} \in {\cal E}(D_n).
\end{gather}
Let us bring the element $h_{B_n}$ (resp.~$h_{D_n}$) to the reduced form in
the algebra ${\cal E}(B_n)$ that is, let us consecutively apply the def\/ining
rela\-tions~(1)--(6), (5a), (6a) to the element~$h_{B_n}$ (resp.\ apply to~$h_{D_n}$ the def\/ining relations for algebra ${\cal E}(D_n)$) in any order
until unable to do so. Denote the resulting (noncommutative) polynomial
by $P_{B_n}(x_{ij},y_{ij},z)$ (resp.\ $P_{D_n}(x_{ij},y_{ij})$). In
principal, this polynomial itself can depend on the order in which the
rela\-tions~(1)--(6), (5a), (6a) are applied.
{\samepage
\begin{Conjecture}[cf.~\protect{\cite[Exercise~8.C5(c)]{ST3}}] \label{conjecture4.4}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] Apart from applying the commutativity relations $(1)$--$(4)$, the
polynomial $P_{B_n}(x_{ij},y_{ij},z)$ $($resp.\ $P_{D_n}(x_{ij},y_{ij}))$
does not depend on the order in which the defining relations have been applied.
\item[$(2)$] Define polynomial $P_{B_n}(s,r,t)$ $($resp.\ $P_{D_n}(s,r))$ to be the
the image of that $P_{B_n}(x_{ij},y_{ij},z)$ $($resp.\ $P_{D_n}(x_{ij},y_{ij}))$
under the specialization
\begin{gather*}
x_{ij} \longrightarrow s, \qquad y_{ij} \longrightarrow r,\qquad
z_i \longrightarrow t.
\end{gather*}
Then
$P_{B_n}(1,1,1)={1 \over 2} {2n \choose n}={1 \over 2} {\rm Cat}_{B_n}$.
\end{enumerate}
\end{Conjecture}}
Note that $P_{B_n}(1,0,1)={\rm Cat}_{A_{n-1}}$.
\begin{Problem} \label{problem4.2}
Investigate the $B_n$ and $D_n$ types reduced polynomials
corresponding to the Co\-xeter elements~\eqref{equation4.2}, and the reduced polynomials
corresponding to the longest elements
\begin{gather*}
w_{B_{n}}:= \prod_{J=1}^{n} z_j \left(\prod_{1 \le i < j \le n} x_{ij} y_{ij}\right),\qquad
w_{D_{n}}= \prod_{1 \le i < j \le n} x_{ij} y_{ij}.
\end{gather*}
\end{Problem}
\subsection{Super analogue of 6-term relations and classical
Yang--Baxter algebras}\label{section4.2}
\subsubsection[Six term relations algebra $6T_n$, its quadratic dual $(6T_n)^{!}$, and algebra $6HT_n$]{Six term relations algebra $\boldsymbol{6T_n}$, its quadratic dual $\boldsymbol{(6T_n)^{!}}$, and algebra $\boldsymbol{6HT_n}$}\label{section4.2.1}
\begin{Definition}\label{definition4.9} {\it The $6$ term relations algebra} $6T_{n}$ is an associative
algebra (say over $\mathbb{Q}$) with the set of generators
$\{ r_{i,j}, 1 \le i \not= j < n \}$, subject to the following relations:
\begin{enumerate}\itemsep=0pt
\item[1)] $r_{i,j}$ and $r_{k,l}$ commute if $\{i,j\} \cap \{k,l\}=\varnothing$,
\item[2)] {\it unitarity condition}: $r_{ij}+r_{ji}=0 $,
\item[3)] {\it classical Yang--Baxter relations}:
$[r_{ij},r_{ik}+r_{jk}]+[r_{ik},r_{jk}] =0$ if~$i$, $j$, $k$~are distinct.
\end{enumerate}
\end{Definition}
We denote by ${\rm CYB}_n$, named by {\it classical Yang--Baxter algebra}, an
associative algebra over $\mathbb{Q}$ generated by elements $\{r_{ij},\, 1 \le i \not= j \le n \}$ subject to relations~1) and~3).
Note that the algebra $6T_{n}$ is given by ${n \choose 2}$ generators
and ${n \choose 3}+3 {n \choose 4}$ quadratic relations.
\begin{Definition}\label{definition4.10} Def\/ine {\it Dunkl elements} in the algebra $6T_n$ to be
\begin{gather*}
\theta_i = \sum_{j \not=i} r_{ij}, \qquad i=1,\ldots,n .
\end{gather*}
\end{Definition}
It easy to see that the Dunkl elements $\{ \theta_i \}_{1 \le i \le n}$
generate a commutative subalgebra in the algebra~$6T_n$.
\begin{Example}[some ``rational and trigonometric'' representations
of the algebra~$6T_n$] \label{example4.4}
Let $A=U({\mathfrak{sl}}(2))$ be the universal enveloping algebra of the
Lie algebra~${\mathfrak{sl}}(2)$. Recall that the algebra ${\mathfrak{sl}}(2)$ is spanned by the
elements $e$, $f$, $h$, such that $[h,e]=2e$, $[h,f]=-2f$, $[e,f]=h$.
Let's search for solutions to the ${\rm CYBE}$ in the form
\begin{gather*}
r_{i,j}=
a(u_i,u_j) h \otimes h +b(u_i,u_j) e \otimes f+c(u_i,u_j) f \otimes e,
\end{gather*}
where $a(u,v),b(u,v) \not= 0,c(u,v) \not= 0$ are meromorphic functions
of the variables $(u,v) \in \mathbb{C}^2$, def\/ined in a neighborhood of
$(0,0)$, taking values in $A \otimes A$. Let $a_{ij}:=a(u_i,u_j)$
(resp.\ $b_{ij}:=b(u_i,u_j)$, $c_{ij}:=c(u_i,u_j)$).
\begin{Lemma}\label{lemma4.2} The elements $r_{i,j}:=a_{ij} h \otimes h
+b_{ij} e \otimes f+c_{ij} f \otimes e$ satisfy CYBE iff
\begin{gather*}
b_{ij} b_{jk} c_{ik}
=c_{ij} c_{jk} b_{ik} \qquad \text{and}\qquad 4 a_{ik}=b_{ij} b_{jk}/b_{ik}-
b_{ik} c_{jk}/b_{ij}-b_{ik} c_{ij}/b_{jk}
\end{gather*}
for $1 \le i < j < k \le n$.
\end{Lemma}
It is not hard to see that
\begin{itemize}\itemsep=0pt
\item there are three rational solutions:
\begin{gather*}
r_1(u,v)= {1/2 h \otimes h+ e \otimes f+f \otimes e \over u-v}, \\
r_2(u,v)= {u+v \over 4(u-v)} h \otimes h+{u \over u-v} e \otimes f+
{v \over u-v}f \otimes e,
\end{gather*}
and $r_3(u,v):=-r_2(v,u)$,
\item there is a trigonometric solution
\begin{gather*}
r_{\rm trig}(u,v)={1 \over 4} {q^{2u}+q^{2v} \over q^{2u}-q^{2v}} h \otimes
h+{q^{u+v} \over q^{2u}-q^{2v}} \big(e \otimes f+f \otimes e \big).
\end{gather*}
\end{itemize}
Notice that the {\it Dunkl element}
$\theta_j:= \sum\limits_{a \not= j} r_{\rm trig}(u_a,u_j)$
corresponds to the truncated (or level~$0$)
{\it trigonometric Knizhnik--Zamolodchikov operator}.
In fact, the ``${\mathfrak{sl}}_{n}$-Casimir element''
\begin{gather*}
\Omega=
{1 \over 2} \left(\sum_{i=1}^{n} E_{ii} \otimes E_{ii} \right) +
\sum_{1 \le i < j \le n} E_{ij} \otimes E_{ji}
\end{gather*}
satisf\/ies the 4-term relations
\begin{gather*}
[\Omega_{12},\Omega_{13}+\Omega_{23}]=0
=[\Omega_{12}+\Omega_{13},\Omega_{23}],
\end{gather*}
and the elements $r_{ij}:= {\Omega_{ij } \over u_i-u_j }$,~
$1 \le i < j \le n$, satisfy the classical Yang--Baxter relations.
Recall that the set
$ \{ E_{ij}:=(\delta_{ik} \delta_{jl})_{1 \le k,l \le n},
\, 1 \le i,j \le n \}$, stands for the standard basis of the
algebra $\operatorname{Mat}(n,\mathbb{R})$.
\end{Example}
\begin{Definition}\label{definition4.11}
Denote by $6T_{n}^{(0)}$ the quotient of the algebra $6T_{n}$ by the
(two-sided) ideal generated by the set of elements
$\{r_{i,j}^2,\,1 \le i < j \le n \}$.
\end{Definition}
More generally, let $\{\beta, q_{ij},\,
1 \le i < j \le n \}$ be a set of parameters.
Let $R:=\mathbb{Q}[\beta][q_{ij}^{\pm 1}]$.
\begin{Definition}\label{definition4.12} Denote by $6HT_n$ the quotient of the algebra
$6T_{n} \otimes R$ by the
(two-sided) ideal generated by the set of elements
$\{r_{i,j}^2-\beta r_{i,j}-q_{ij}, \, 1 \le i < j \le n \}$.
\end{Definition}
All these algebras are naturally graded, with $\deg (r_{i,j})=1$,
$\deg(\beta)=1$, $\deg(q_{ij})=2$.
It is clear that the algebra $6T^{(0)}_{n}$ can be considered as the
inf\/initesimal deformation $R_{i,j}:=1+\epsilon r_{i,j}$,
$\epsilon \longrightarrow 0$, of the Yang--Baxter group ${\rm YB}_{n}$.
For the reader convenience we recall the def\/inition of the {\it
Yang--Baxter} group.
\begin{Definition}\label{definition4.13} The Yang--Baxter group~${\rm YB}_n$ is a group generated by
elements $\{R_{ij}^{\pm 1}, \,1 \le i < j \le n \}$, subject to the set
of def\/ining relations
\begin{itemize}\itemsep=0pt
\item $R_{ij} R_{kl}=R_{kl} R_{ij}$ if $i$, $j$, $k$, $l$ are distinct,
\item {\it quantum Yang--Baxter relations}:
\begin{gather*}
R_{ij} R_{ik} R_{jk}=R_{jk} R_{ik} R_{ij} \qquad \text{if} \quad 1 \le i < j < k
\le n.
\end{gather*}
\end{itemize}
\end{Definition}
\begin{Corollary}\label{corollary4.3} Define $h_{ij}=1+r_{ij} \in 6HT_{n}$. Then the
following relations in the algebra $6HT_n$ are satisfied:
\begin{enumerate}\itemsep=0pt
\item[$(1)$] $r_{ij} r_{ik} r_{jk}=r_{jk} r_{ik} r_{ij}$ for all pairwise
distinct $i$, $j$ and $k$;
\item[$(2)$] Yang--Baxter relations: $h_{ij} h_{ik} h_{jk}=h_{jk} h_{ik} h_{ij}$
if $1 \le i < j < k \le n$.
\end{enumerate}
\end{Corollary}
Note, the item $(1)$ includes three relations in fact.
\begin{Proposition} \label{proposition4.4}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] The quadratic dual $(6T_n)^{!}$ of the algebra $6T_n$ is a~quadratic algebra generated by the elements $\{t_{i,j},
\, 1 \le i < j \le n \}$ subject to the set of relations
\begin{enumerate}\itemsep=0pt
\item[$(i)$] $t_{i,j}^2=0$ for all $i \not= j$;
\item[$(ii)$] anticommutativity: $t_{ij} t_{k,l}+t_{k,l} t_{i,j}=0$
for all $i \not= j$ and $k \not= l$;
\item[$(iii)$] $t_{i,j} t_{i,k}=t_{i,k} t_{j,k}=t_{i,j} t_{j,k}$ if $i$, $j$, $k$ are
distinct.
\end{enumerate}
\item[$(2)$] The quadratic dual $(6T_n^{(0)})^{!}$ of the algebra $6T_n^{(0)}$ is a quadratic algebra with generators $\{t_{i,j},\, 1 \le i < j \le n \}$
subject to the relations $(ii)$--$(iii)$ above only.
\end{enumerate}
\end{Proposition}
\subsubsection[Algebras $6T_n^{(0)}$ and $6T_n^{\bigstar}$]{Algebras $\boldsymbol{6T_n^{(0)}}$ and $\boldsymbol{6T_n^{\bigstar}}$}\label{section4.2.2}
We are reminded that the algebra $6T_n^{(0)}$ is the quotient of the six term
relation algebra $6T_n$ by the two-sided ideal generated by the elements
$\{r_{ij} \}_{1 \le i < j \le n}$. Important {\it consequence} of the
classical Yang--Baxter relations and relations $r_{ij}^2=0$, $\forall\, i \not=j$,
is that the both additive Dunkl elements $\{ \theta_{i} \}_{1 \le i \le n}$
and multiplicative ones
\begin{gather*}
\left\{ \Theta_i = \prod_{\atop a=i-1}^{1} h_{ai}^{-1}
\prod_{\atop a=i+1}^{n} h_{ia} \right\}_{1\le i \le n}
\end{gather*}
generate commutative
subalgebras in the algebra~$6T_n^{(0)}$ (and in the algebra $6T_n$ as well),
see Corollary~\ref{corollary4.3}. The problem we are interested in, is to describe
commutative subalgebras generated by additive (resp.\ multiplicative) Dunkl
elements in the algebra~$6T_{n}^{(0)}$. Notice that the subalgebra generated
by additive Dunkl elements in the abelianization\footnote{See, e.g., \url{http://mathworld.wolfram.com/Abelianization.html}.}
of the algebra $6T_n{(0)}$ has been studied in~\cite{PSS, SS}. In order
to state the result from~\cite{PSS} we need, let us introduce a bit of
notation. As before, let~${\cal{F}}l_n$ denotes the complete f\/lag variety, and
denote by~${\cal{A}}_n$ the algebra generated by the curvature of $2$-forms of
the standard Hermitian linear bundles over the f\/lag variety ${\cal{F}}l_n$,
see, e.g.,~\cite{PSS}. Finally, denote by~$I_n$ the ideal in the ring of
polynomials $\mathbb{Z}[t_1,\ldots,t_n]$ generated by the set of elements
\begin{gather*}
(t_{i_{1}}+\cdots+t_{i_{k}})^{k(n-k)+1}
\end{gather*}
for all sequences of indices $1 \le i_1 < i_2 < \cdots < i_k \le n$, $k=1,
\ldots,n$.
\begin{Theorem}[\cite{PSS,SS}]\label{theorem4.4}\quad
\begin{enumerate}\itemsep=0pt
\item[$(A)$] There exists a natural isomorphism
\begin{gather*}
{\cal{A}}_n \longrightarrow \mathbb{Z}[t_1,\ldots,t_n] /I_n,
\end{gather*}
\item[$(B)$]
\begin{gather*}
{\rm Hilb}({\cal{A}}_n,t) =t^{{n \choose 2}} {\rm Tutte}\big(K_n,1+t,t^{-1}\big).
\end{gather*}
\end{enumerate}
\end{Theorem}
Therefore the dimension of ${\cal{A}}_n$ (as a $\mathbb{Z}$-vector space) is equal to
the number ${\cal{F}}(n)$ of forests on~$n$ labeled vertices. It is well-known that
\begin{gather*}
\sum_{n \ge 1} {\cal{F}}(n) \frac{x^n}{n !} =\exp \left( \sum_{n \ge 1}
n^{n-1} \frac{x^n}{n !} \right) -1.
\end{gather*}
For example,
\begin{gather*} {\rm Hilb}({\cal A}_3,t)=(1,2,3,1), \qquad {\rm Hilb}({\cal A}_4,t)=(1,3,6,10,11,6,1),\\
{\rm Hilb}({\cal A}_5,t)=(1,4,10,20,35,51,64,60,35,10,1),\\
{\rm Hilb}({\cal A}_6,t)=(1,5,15,35,70,126,204,300,405,490,511,424,245,85,15,1).
\end{gather*}
\begin{Problem} \label{problem4.3}
Describe subalgebra in $(6T_n^{(0)})^{ab}$ generated by the
multiplicative Dunkl elements $\{\Theta_i \}_{1 \le i \le n}$.
\end{Problem}
On the other hand, the commutative subalgebra ${\cal{B}}_n$ generated by
the additive Dunkl elements in the algebra $6T_n^{(0)}$, $n \ge 3$, has
{\it infinite} dimension. For example,
\begin{gather*}
{\cal{B}}_3 \cong \mathbb{Z}[x,y]/ \langle xy(x+y) \rangle,
\end{gather*}
and the Dunkl elements $\theta_j^{(3)}$, $j=1,2,3$, have inf\/inite order.
\begin{Definition}\label{definition4.14}
Def\/ine algebra $6T_n^{\bigstar}$ to be the quotient of that
$6T_n^{(0)}$ by the two-sided ideal generated by the set of ``cyclic
relations''
\begin{gather*}
\sum_{j=2}^{m} \prod_{a=j}^{m} r_{i_{1},i_{a}} \prod_{a=2}^{j}
r_{i_{1},i_{a}} = 0
\end{gather*}
for all sequences $\{1 \le i_1,i_2,\ldots,i_m \le n \}$ of pairwise distinct
integers, and all integers $2 \le m \le n$.
\end{Definition}
For example,
\begin{itemize}\itemsep=0pt
\item ${\rm Hilb}(6T_{3}^{\bigstar},t)=(1,3,5,4,1)=(1+t)(1,2,3,1)$,
\item subalgebra (over~$\mathbb{Z}$) in the algebra $6T_{3}^{\bigstar}$
generated by Dunkl elements $\theta_{1}$ and $\theta_{2}$ has the Hilbert
polynomial equal to $(1,2,3,1)$, and the following presentation:
$\mathbb{Z} [x,y]/I_{3}$, where $I_{3}$ denotes the ideal in $\mathbb{Z}[x,y]$ generated
by $x^3$, $y^3$, and $(x+y)^3$,
\item ${\rm Hilb}(6T_{4}^{\bigstar},t) = (1,6,23,65,134,164,111,43,11,1)_{t}$.
\end{itemize}
As a consequence of the cyclic relations, one can check that for any integer
$n \ge 2$ the $n$-th power of the additive Dunkl element~$\theta_i$ is equal
to zero in the algebra $6T_{n}^{\bigstar}$ for all $i=1,\ldots,n$. Therefore,
the Dunkl elements generate a f\/inite-dimensional commutative subalgebra in
the algebra $6T_{n}^{\bigstar}$. There exist natural homomorphisms
\begin{gather}\label{equation4.3}
6T_n^{\bigstar} \longrightarrow 3T_n^{(0)}, \qquad
\begin{CD}
{\cal{B}}_n @>\tilde{\pi}>> {\cal{A}}_n \longrightarrow H^{*}({\cal{F}}l_n,\mathbb{Z})
\end{CD}
\end{gather}
The f\/irst and third arrows in~\eqref{equation4.3} are epimorphism. We expect that the
map~$\tilde{\pi}$ is also epimorphism\footnote{Contrary to the case of the map ${\rm pr}_{n}\colon \mathbb{Z}[\theta_1,\ldots,
\theta_n] \longrightarrow (3T_n{(0)})^{ab}$, where the image ${\rm Im}({\rm pr}_n)$ has
dimension equals to the number of permutations in~${\mathbb{S}}_n$ with~$(n-1)$ inversions see~\cite[$A001892$]{SL}.}, and looking for a description of the kernel $\operatorname{ker}(\tilde{\pi})$.
\begin{Comments} \label{comments4.3}\quad
\begin{itemize}\itemsep=0pt
\item Let us denote by ${\cal{B}}_n^{\rm mult}$ and ${\cal{A}}_n^{\rm mult}$ the
subalgebras generated by {\it multiplicative} Dunkl elements in the algebras
$6T_n^{(0)}$ and $\big(6T_n^{(0)}\big)^{ab}$ correspondingly. One can def\/ine a~sequence of maps
\begin{gather}\label{equation4.4}
{\cal{B}}_n^{\rm mult} \longrightarrow {\cal{A}}_n^{\rm mult}
\overset{\tilde{\phi}}{\longrightarrow}
K^{*}({\cal{F}}l_n),
\end{gather}
which is a $K$-theoretic analog of that~\eqref{equation4.3}. It is an interesting problem to f\/ind a geometric interpretation of the algebra ${\cal{A}}_n^{\rm mult}$ and the
map~$\tilde{\phi}$.
\item ``Quantization''. Let $\beta$ and $\{ q_{ij}=q_{ji}, 1 \le i,j \le n \}$ be parameters.
\end{itemize}
\begin{Definition} \label{definition4.15}
Def\/ine algebra $6HT_n$ to be the quotient of the algebra $6T_n$ by
the two sided ideal generated by the elements $\{r_{ij}^2 - \beta r_{ij} -q_{ij} \}_{1 \le i,j \le n}$.
\end{Definition}
\begin{Lemma} \label{lemma4.3}
The both additive $\{\theta_i \}_{1 \le i \le n}$ and
multiplicative $\{\Theta_{i} \}_{1 \le i \le n}$ Dunkl elements generate
commutative subalgebras in the algebra~$6HT_n$.
\end{Lemma}
Therefore one can def\/ine algebras $6{\cal{HB}}_n$ and $6{\cal{HA}}_n$ which
are a ``quantum deformation'' of algebras~${\cal{B}}_n$ and~${\cal{A}}_n$
respectively. We {\it expect} that in the case $\beta=0$ and a special choice
of ``arithmetic parameters'' $\{ q_{ij} \}$, the algebra ${\cal{HA}}_n$ is
connected with the arithmetic Schubert and Grothendieck calculi, cf.~\cite{SS, T}.
Moreover, for a ``general'' set of parameters $\{q_{ij}\}_{1 \le i,j \le n}$ and $\beta=0$, we {\it expect} an existence of a natural homomorphism
\begin{gather*}
{\cal{HA}}_n^{\rm mult} \longrightarrow {\cal{QK}}^{*}({{\cal{F}}l}_{n}),
\end{gather*}
where ${\cal{QK}}^{*}({{\cal{F}}l}_{n})$ denotes a {\it multiparameter
quantum deformation} of the $K$-theory ring $K^{*}({\cal{F}}l_n)$ \cite{K,KM}; see also Section~\ref{section3.1}. Thus, we treat the algebra
${\cal{HA}}_n^{\rm mult}$ as the $K$-theory version of a~multiparameter quantum
deformation of the algebra ${\cal{A}}_n^{\rm mult}$ which is generated by the
curvature of $2$-forms of the Hermitian linear bundles over the f\/lag variety~${\cal{F}}l_n$.
\begin{itemize}\itemsep=0pt
\item One can def\/ine an analogue of the algebras $6T_n^{(0)}$,
$6HT_n$ etc., denoted by $6T(\Gamma)$ etc., for any subgraph
$\Gamma \subset K_n$ of the complete graph~$K_n$, and in fact for any oriented
matroid. It is known that ${\rm Hilb}((6T_n(\Gamma)^{ab},t) = t^{e(\Gamma)}
{\rm Tutte}(\Gamma,1+t,t^{-1})$, see, e.g.,~\cite{B} and the literature quoted therein.
\end{itemize}
\end{Comments}
\subsubsection[Hilbert series of algebras ${\rm CYB}_n$ and $6T_n$]{Hilbert series of algebras $\boldsymbol{{\rm CYB}_n}$ and $\boldsymbol{6T_n}$\footnote{Results of this subsection have been obtained
independently in~\cite{BEE}. This paper contains, among other things, a~description of a basis in the algebra~$6T_n$, and much more.}}\label{section4.2.3}
\begin{Examples}\label{examples4.2}
\begin{gather*}
{\rm Hilb}(6T_{3},t)=\big(1-3t+t^2\big)^{-1},\qquad
{\rm Hilb}(6T_{4},t)=\big(1-6t+7t^2-t^3\big)^{-1},\\
{\rm Hilb}(6T_{5},t)=\big(1-10t+25t^2-15t^3+t^{4}\big)^{-1},\\
{\rm Hilb}(6T_{6},t)=\big(1-15t+65t^2-90t^3+31t^4-t^5\big)^{-1},\\
{\rm Hilb}\big(6T_{3}^{(0)},t\big)= [2][3](1-t)^{-1},\qquad
{\rm Hilb}\big(6T_{4}^{(0)},t\big)= [4](1-t)^{-2}\big(1-3t+t^2\big)^{-1}.
\end{gather*}
\end{Examples}
In fact, the following statements are true.
\begin{Proposition}[cf.~\cite{BEE}]\label{proposition4.5} Let $n \ge 2$, then
\begin{itemize}\itemsep=0pt
\item The algebras $6T_n$ and ${\rm CYB}_n$ are Koszul.
\item We have
\begin{gather*}
{\rm Hilb}(6T_n,t)=
\left( \sum_{k=0}^{n-1}(-1)^k {n \brace n-k } t^{k}
\right)^{-1},
\end{gather*}
where ${n \brace k } $ stands for the Stirling
numbers of the second kind, i.e., the number of ways to partition a set
of~$n$ things into~$k$ nonempty subsets.
\item
\begin{gather*} {\rm Hilb}({\rm CYB}_n,t)=
\left(\sum_{k=0}^{n-1}(-1)^k (k+1)! N(k,n) t^k \right)^{-1},
\end{gather*}
where $N(k,n)={1 \over n}{n \choose k} {n \choose k+1}$ denotes the Narayana
number, i.e., the number of Dyck $n$-paths with exactly~$k$ peaks.
\end{itemize}
\end{Proposition}
\begin{Corollary} \label{corollary4.4}\quad
\begin{enumerate}\itemsep=0pt
\item[$(A)$] The Hilbert polynomial of the quadratic dual of the algebra
$6T_n$ is equal to
\begin{gather*}
{\rm Hilb}\big(6T_n^{!},t\big)=
\sum_{k=0}^{n-1} {n \brace n-k } t^{k}.
\end{gather*}
It is well-known that
\begin{gather*}
\sum_{n \ge 0}\left( \sum_{k=0}^{n-1} {n \brace n-k } t^{k}
\right) {z^n \over n!}=\exp\left({\exp(zt)-1 \over t}\right).
\end{gather*}
Therefore,
\begin{gather*}
\dim (6T_n)^{!}={\rm Bell}_n,
\end{gather*}
where ${\rm Bell}_n$ denotes the $n$-th Bell number, i.e., the number of
ways to partition $n$ things into subsets, see~{\rm \cite{SL}}.
Recall, that
\begin{gather*}
\sum_{n \ge 0} {\rm Bell}_n {z^n \over n!}=\exp(\exp(z)-1)).
\end{gather*}
\item[$(B)$]
The Hilbert polynomial of the quadratic dual of the algebra
${\rm CYB}_n$ is equal to
\begin{gather*}
{\rm Hilb}\big(({\rm CYB}_n\big)^{!},t)=\sum_{k=0}^{n-1}(k+1)! N(k,n) t^k=
(n-1)! L_{n-1}^{(\alpha=1)}\big({-}t^{-1}\big) t^{n-1},
\end{gather*}
where
\begin{gather*}
L_n^{(\alpha)}(x)=
{x^{-\alpha} e^x \over n!}{d^n \over dx^n} \big(e^{-x} x^{n+\alpha}\big)
\end{gather*}
denotes the generalized Laguerre polynomial.
The numbers $(k+1)! N(n,k):= L(n,n-k)$ are known as Lah numbers, see,
e.g., {\rm \cite[$A008297$]{SL}}, moreover~{\rm \cite{SL}},
\begin{gather*}
\dim ({\rm CYB}_n)^{!} = A000262.
\end{gather*}
It is well-known that
\begin{gather*}
\sum_{n \ge 0}\left( \sum_{k \ge 0}^{n-1} (k+1)!
N(k,n) t^{k} \right) {z^n \over n!}=\exp\big(z(1-zt)^{-1}\big).
\end{gather*}
\end{enumerate}
\end{Corollary}
\begin{Comments} \label{comments4.4}
Let ${\cal{E}}_n(u)$, $u \not= 0,1$, be the {\it Yokonuma--Hecke} algebra, see, e.g.,~\cite{RH} and the literature quoted therein. It is
known that the dimension of the Yokonuma--Hecke algebra ${\cal{E}}_n(u)$ is
equal to $n! B_n$, where $B_n$ denotes as before the $n$-th Bell number.
Therefore, $\dim({\cal{E}}_n(u)) =
\dim ((6T_n)^{!} \rtimes {\mathbb S}_n)$, where
$(6T_n)^{!} \rtimes {\mathbb S}_n$ denotes the semi-direct product of
the algebra $(6T_n)^{!}$ and the symmetric group~${\mathbb S}_n$. It
seems an interesting task to check whether or not the algebras
$(6T_n)^{!} \rtimes {\mathbb S}_n$ and ${\cal{E}}_n(u)$ are isomorphic.
\end{Comments}
\begin{Remark}\label{remark4.2}
Denote by ${\cal M}{\rm YB}_n$ the group algebra over $\mathbb{Q}$
of the {\it monoid} corresponding to the Yang--Baxter group ${\rm YB}_n$, see, e.g.,
Def\/inition~\ref{definition4.10}. Let $P({\cal M}{\rm YB}_n,s,t)$ denotes the Poincar\'e
polynomial of the algebra ${\cal M}{\rm YB}_n$. One can show that
\begin{gather*}{\rm Hilb}(6T_n,s)=P({\cal M}{\rm YB}_n,-s,1)^{-1} .
\end{gather*}
For example,
\begin{gather*}
P({\cal M}{\rm YB}_3,s,t)= 1+3s t+s^2 t^3,\\
P({\cal M}{\rm YB}_4,s,t)=1+6s t+s^2 \big(3t^2+4t^3\big)+s^3 t^6,\\
P({\cal M}{\rm YB}_5,s,t)=1+10s t+s^2\big(15t^2+10t^3\big)+s^3\big(10t^4+5t^6\big)+s^4 t^{10}.
\end{gather*}
Note that ${\rm Hilb}({\cal M}{\rm YB}_n,t)= P({\cal M}{\rm YB}_n,-1,t)^{-1}$ and
$P({\cal M}{\rm YB}_n,1,1) = {\rm Bell}_{n}$, the $n$-th Bell number.
\end{Remark}
\begin{Conjecture}\label{conjectute4.5}
\begin{gather*}
P({\cal M}{\rm YB}_n,s,t)=\sum_{\pi} s^{\#(\pi)} t^{n(\pi)},
\end{gather*}
where the sum runs over all partitions $\pi=(I_1,\ldots,I_k)$ of the set
$[n]:=[1,\ldots,n]$ into nonempty subsets $I_1,\ldots,I_k$, and we set by
definition, $\#(\pi):=n-k$, $n(\pi):=\sum\limits_{a=1}^{k} {|I_a| \choose 2}$.
\end{Conjecture}
\begin{Remark} \label{remark4.3}
For any f\/inite Coxeter group $(W,S)$ one can def\/ine the
algebra ${\rm CYB}(W):={\rm CYB}(W,S)$ which is an analog of the algebra
${\rm CYB}_n={\rm CYB}(A_{n-1})$ for other root systems.
\end{Remark}
\begin{Conjecture}[A.N.~Kirillov, Yu.~Bazlov] \label{conjecture4.6}
Let $(W,S)$ be a finite Coxeter group
with the root system~$\Phi$. Then
\begin{itemize}\itemsep=0pt
\item the algebra ${\rm CYB}(W)$ is Koszul;
\item ${\rm Hilb}({\rm CYB}(W),t)= \left\{\sum\limits_{k=0}^{|S|}r_k(\Phi) (-t)^{k}
\right\}^{-1}$,
\end{itemize}
where $r_k(\Phi)$ is equal to the number of subsets in $\Phi^{+}$ which
constitute the positive part of a root subsystem of rank~$k$. For example,
$r_{1}(\Phi)=|\Phi^{+}|$, and $r_{2}(\Phi)$ is equal to the number of
defining relations in a representation of the algebra ${\rm CYB}(W)$.
\end{Conjecture}
\begin{Example}\label{example4.5}
\begin{gather*}
{\rm Hilb}\big({\rm CYB}(B_2)^{!},t\big)=(1,4,3), \qquad {\rm Hilb}\big({\rm CYB}(B_3)^{!},t\big)=(1,9,13,2), \\
{\rm Hilb}\big({\rm CYB}(B_4)^{!},t\big)=(1,16,46,28,5), \qquad
{\rm Hilb}\big({\rm CYB}(B_5)^{!},t\big)=(1,25,130,200,101,12),\\
{\rm Hilb}\big({\rm CYB}(D_4)^{!},t\big)=(1,12,34,24,4), \qquad
{\rm Hilb}\big({\rm CYB}(D_5)^{!},t\big)=(1,20,110,190,96,11).
\end{gather*}
\end{Example}
\begin{Definition}\label{definition4.16} The even generic Orlik--Solomon algebra
${\rm OS}^{+}(\Gamma_n)$ is def\/ined to be an associative algebra (say over~$\mathbb{Z}$)
generated by the set of {\it mutually commuting} elements~$y_{i,j}$, $1 \le i \not= j \le n$, subject to the set of cyclic relations
\begin{gather*}
y_{i,j}=y_{j,i}, \qquad y_{i_1,i_2} y_{i_2,i_3} \cdots
y_{i_{k-1},i_k} y_{i_1,i_k}=0 \qquad \text{for} \quad k=2,\ldots, n,
\end{gather*}
and all sequences of pairwise distinct integers
$1 \le i_1, \ldots, i_k \le n$.
\end{Definition}
\begin{Exercises}\label{exercises4.4}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] Show that
\begin{gather*}
\exp\big(z (1-zt)^{-q}\big)=1+\sum_{n \ge 1} \left( 1+\sum_{k=1}^{n-1} {n-1 \choose k} \prod_{a=0}^{k-1}(a+(n-k) q) t^{k} \right) {z^n \over n!}.
\end{gather*}
\item[(2)] {\it The even generic Orlik--Solomon algebra.}
Show that the number of degree $k$, $k \ge 3$,
relations in the def\/inition of the Orlik--Solomon algebra ${\rm OS}^{+}(\Gamma_n)$
is equal to ${1 \over 2} (k-1)! {n \choose k}$
and also is equal to the maximal number of $k$-cycles in the complete graph~$K_n$.
\end{enumerate}
Note that if one replaces the commutativity condition in the above def\/inition
on the condition that ${y_{i,j}}$'s pairwise {\it anticommute}, then the
resulting algebra appears to be isomorphic to the Orlik--Solomon algebra
${\rm OS}(\Gamma_n)$ corresponding to the generic hyperplane arrangement~$\Gamma_n$, see~\cite{PSt}. It is known \cite[Corollary~5.3]{PSt}, that
\begin{gather*}
{\rm Hilb}({\rm OS}(\Gamma_n),t)= \sum_{F} t^{|F|},
\end{gather*}
where the sum runs over all forests~$F$ on the vertices $1, \ldots, n$, and~$|F|$ denotes the number of edges in a~forest~$F$.
It follows from Corollary~\ref{corollary4.4}, that
\begin{gather*}
\sum_{n \ge 1} {\rm Hilb}({\rm OS}(\Gamma_n),t) {z^n \over n!}
= \exp \left( \sum_{n \ge 1} n^{n-2} t^{n-1} {z^n \over n!} \right).
\end{gather*}
It is not dif\/f\/icult to see that
${\rm Hilb}({\rm OS}^{+}(\Gamma_n),t)={\rm Hilb}({\rm OS}(\Gamma_n),t)$. In particular,
$\dim {\rm OS}^{+}(\Gamma_n)= {\cal F}(n)$. Note also that a sequence
$\{ {\rm Hilb}({\rm OS}(\Gamma_n),-1) \}_{n \ge 2}$ appears in \cite[$A057817$]{SL}.
The polynomials ${\rm Hilb}({\cal A}_n,t)$, $F_n(x,t)$ and
${\rm Hilb}({\rm OS}^{+}(\Gamma_n),t)$ can be expressed, see, e.g.,~\cite{PSS},
as certain {\it specializations} of the Tutte polynomial $T(G;x,y)$
corresponding to the complete graph $G:=K_n$. Namely,
\begin{gather*}
{\rm Hilb}({\cal A}_n,t)=t^{{n \choose 2}} T\big(K_n; 1+t, t^{-1}\big),\qquad
{\rm Hilb}\big({\rm OS}^{+}(\Gamma_n),t\big)=t^{n-1} T\big(K_n;1+t^{-1},1\big).
\end{gather*}
\end{Exercises}
\subsubsection{Super analogue of 6-term relations algebra}\label{section4.2.4}
Let $n$, $m$ be non-negative integers.
\begin{Definition}\label{definition4.17} {\it The super $6$-term relations algebra $6T_{n,m}$} is
an associative algebra over $\mathbb{Q}$ generated by the elements $\{x_{i,j}, \,
1 \le i \not= j \le n \}$ and
$\{y_{\alpha,\beta}, \, 1 \le \alpha \not= \beta \le m \}$ subject to the
set of relations
\begin{enumerate}\itemsep=0pt
\item[(0)] $x_{i,j}+x_{j,i}=0$, $ y_{\alpha, \beta}=y_{\beta, \alpha}$;
\item[(1)] $x_{i,j}x_{k,l}=x_{k,l}x_{i,j}$, $x_{i,j}y_{\alpha,\beta}=
y_{\alpha,\beta}x_{i,j}$, $y_{\alpha,\beta}y_{\gamma,\delta}+
y_{\gamma,\delta}y_{\alpha,\beta}=0$,
if tuples $(i,j,k,l)$, $(i,j,\alpha,\beta)$, as well as
$( \alpha, \beta, \gamma, \delta)$ consist of pair-wise distinct integers;
\item[(2)] classical Yang--Baxter relations and theirs super analogue:
$[x_{i,k},x_{j,i}+x_{j,k}]+[x_{i,j},x_{j,k}]=0$
if $1 \le i,j,k \le n$ are distinct,
$[x_{i,k},y_{j,i}+y_{j,k}]+[x_{i,j},y_{j,k}]=0$
if $ 1 \le i,j,k \le \min(n,m)$ are distinct,
$[y_{\alpha,\gamma},y_{\beta,\alpha}+
y_{\beta,\gamma}]_{+}+[y_{\alpha,\beta},y_{\beta,\gamma}]_{+}=0$
if $1 \le \alpha, \beta,\gamma \le m$ are distinct.
\end{enumerate}
Recall that $[a,b]_{+}:=ab+ba$ denotes
the anticommutator of elements~$a$ and~$b$.
\end{Definition}
\begin{Conjecture} \label{conjecture4.7} The algebra $6T_{n,m}$ is Koszul.
\end{Conjecture}
\begin{Theorem} \label{theorem4.5}
Let $n,m \in \mathbb{Z}_{\ge 1}$, one has
\begin{gather*}
{\rm Hilb}\big((6T_n)^{!},t\big) {\rm Hilb}\big((6T_m)^{!},t\big)=
\sum_{k=0}^{\min (n,m)-1}
{\min(n,m) \brace \min(n,m)-k}
{\rm Hilb}\big((6T_{n-k,m-k})^{!},t\big) t^{2k},
\end{gather*}
where as before ${n \brace n-k}$ denotes the Stirling numbers of the second
kind, see, e.g., {\rm \cite[$A008278$]{SL}}.
\end{Theorem}
\begin{Corollary}\label{corollary4.5} Let $n,m \in \mathbb{Z}_{ \ge 1}$. One has
\begin{enumerate}\itemsep=0pt
\item[$(a)$] Symmetry: ${\rm Hilb}(6T_{n,m},t)={\rm Hilb}(6T_{m,n},t)$.
\item[$(b)$] Let $n \le m$, then
\begin{gather*}
{\rm Hilb}\big((6T_{n,m})^{!},t\big)=
\sum_{k=0}^{n-1} s(n-1,n-k)
{\rm Hilb}\big((6T_{n-k})^{!},t\big) {\rm Hilb}\big((6T_{m-k})^{!},t\big) t^{2k},
\end{gather*}
where $s(n-1,n-k)$ denotes the Stirling numbers of the first kind, i.e.,
\begin{gather*}
\sum_{k=0}^{n-1} s(n-1,n-k) t^k= \prod_{j=1}^{n-1} (1-j t).
\end{gather*}
\item[$(c)$] $\dim (6T_{n,n})^{!}$ is equal to the number of pairs of partitions of
the set $\{1,2,\ldots,n \}$ whose meet is the partition
$\{\{1\}, \{2\},\ldots, \{n\}\}$, see, e.g., {\rm \cite[$A059849$]{SL}}.
\end{enumerate}
\end{Corollary}
\begin{Example} \label{examples4.77}
\begin{gather*}
{\rm Hilb}\big((6T_{3,2})^{!},t\big)={\rm Hilb}\big((6T_{2,3})^{!},t\big)=(1,4,3), \\
{\rm Hilb}\big((6T_{2,4})^{!},t\big)={\rm Hilb}\big((6T_{4,2})^{!},t\big)=(1,7,12,5),\qquad
{\rm Hilb}\big((6T_{3,3})^{!},t\big)=(1,6,8),\\
{\rm Hilb}\big((6T_{2,5})^{!},t\big)={\rm Hilb}\big((6T_{5,2})^{!},t\big)=(1,11,34,34,9), \\
{\rm Hilb}\big((6T_{3,4})^{!},t\big)={\rm Hilb}\big((6T_{4,3})^{!},t\big)=(1,9,23,16),\\
{\rm Hilb}\big((6T_{4,4})^{!},t\big)=(1,12,44,50,6), \\
{\rm Hilb}\big((6T_{3,5})^{!},t\big)={\rm Hilb}\big((6T_{5,3})^{!},t\big)=(1,13,53,79,34), \\
{\rm Hilb}\big((6T_{4,5})^{!},t\big)={\rm Hilb}\big((6T_{5,4})^{!},t\big)=(1,16,86,182,131,12), \\
{\rm Hilb}\big((6T_{5,5})^{!},t\big)=(1,20,140,410,462,120).
\end{gather*}
\end{Example}
Now let us def\/ine in the algebra $6T_{n,m}$ the Dunkl elements
$\theta_{i}:=\sum\limits_{j \not= i} x_{i,j}$, $1 \le i \le n$, and
${\bar \theta}_{\alpha}:=\sum\limits_{\beta \not= \alpha} y_{\alpha,\beta}$,
$1 \le \alpha \le m$.
\begin{Lemma}\label{lemma4.4} One has
\begin{itemize}\itemsep=0pt
\item $[\theta_{i},\theta_{j}]= 0$,
\item $[\theta_i, {\bar \theta}_{\alpha}]=[x_{i,\alpha},y_{i,\alpha}]$,
\item $[{\bar \theta}_{\alpha}, {\bar \theta}_{\beta}]_{+}= 2 y_{\alpha,\beta}^2
$ if $\alpha \not= \beta$.
\end{itemize}
\end{Lemma}
\begin{Remark}[``odd'' six-term relations algebra] \label{remark4.4}
In particular,
one can def\/ine an ``odd'' analog $6T_n^{(-)}=6T_{0,n}$ of the six
term relations algebra~$6T_n$. Namely, the algebra $6T_n^{(-)}$ is given by the
set of generators $\{ y_{ij},\, 1 \le i < j \le n \}$, and that of relations:
\begin{enumerate}\itemsep=0pt
\item[1)] $y_{i,j}$ and $y_{k,l}$ {\it anticommute} if $i$, $j$, $k$, $l$ are pairwise
distinct;
\item[2)] $[y_{i,j},y_{i,k}+y_{j,k}]_{+}+[y_{i,k},y_{j,k}]_{+}=0$,
if $1 \le i < j \le k \le n$, where $[x,y]_{+}=xy+yx$ denotes the
anticommutator of~$x$ and~$y$.
\end{enumerate}
The ``odd'' three term relations algebra~$3T_n^{-}$ can be obtained as the
quotient of the algebra~$6T_n^{-}$ by the two-sided ideal generated by the
three term relations
$y_{ij} y_{jk} + y_{jk} y_{ki} +y_{ki} y_{ij} = 0$ if $i$, $j$, $k$ are pairwise
distinct.
One can show that the Dunkl elements $\theta_i$ and $\theta_j$, $i \not= j$,
given by formula
\begin{gather*}
\theta_i=\sum_{j \not= i} y_{ij}, \qquad i=1,\ldots,n,
\end{gather*}
form an {\it anticommutative} family of elements in the algebra $6T_n^{(-)}$.
In a similar fashion one can def\/ine an ``odd'' analogue of the dynamical six
term relations algebra $6DT_n$, see Def\/inition~\ref{definition2.2} and Section~\ref{section2.1.1}, as well as
def\/ine an ``odd'~analogues of the algebra $3MT_n(\beta,{\boldsymbol{0}})$,
see Def\/inition~\ref{definition3.7}, the Kohno--Drinfeld algebra, the Hecke algebra and few
others considered in the present paper. Details are omitted in the present
paper.
More generally, one can ask what are natural $q$-analogues of the six term and
three term relations algebras? In other words to describe relations which
ensure the $q$-commutativity of Dunkl elements def\/ined above. First of all it
would appear natural that the ``$q$-locality and $q$-symmetry conditions''
hold among the set of generators $\{y_{ij}, \, 1 \le i \not= j \le n \}$, that is
$y_{ij}+ q y_{ji}=0$, $y_{ij} y_{kl} = q y_{kl} y_{ij}$ if $i < j$,
$k < l$, and $\{i,j\} \cap \{k,l\} =\varnothing$.
Another natural condition is the fulf\/illment of $q$-analogue of the classical
Yang--Baxter relations, namely,
$[y_{ik},y_{jk}]_{q} + [y_{ik},y_{ji}]_{q} + [y_{ij},y_{jk}]_{q} =0$ if
$i < j < k$, where $[x,y]_{q}:= x y -q y x$ denotes the $q$-commutator.
However we are not able to f\/ind the $q$-analogue of the classical Yang--Baxter relation listed above in the mathematical and physical literature yet.
Only cases $q=1$ and $q= -1$ have been extensively studied.
\end{Remark}
\subsection{Four term relations algebras / Kohno--Drinfeld algebras}\label{section4.3}
\subsubsection[Kohno--Drinfeld algebra $4T_n$ and that ${\rm CYB}_n$]{Kohno--Drinfeld algebra $\boldsymbol{4T_n}$ and algebra $\boldsymbol{{\rm CYB}_n}$}\label{section4.3.1}
\begin{Definition}\label{definition4.18} {\it The $4$-term relations algebra} (or the Kohno--Drinfeld
algebra, or inf\/initesimal pure braids algebra) $4T_{n}$ is an
associative algebra (say over $\mathbb{Q}$) with the set of generators
$y_{i,j}$, $1 \le i < j \le n$, subject to the following relations
\begin{enumerate}\itemsep=0pt
\item[1)] $y_{i,j}$ and $y_{k,l}$ are commute, if $i$, $j$, $k$, $l$ are all distinct;
\item[2)] $[y_{i,j},y_{i,k}+y_{j,k}]= 0$,
$[y_{i,j}+y_{i,k},y_{j,k}]= 0$ if $1 \le i < j \le k \le n$.
\end{enumerate}
\end{Definition}
Note that the algebra $4T_{n}$ is given by ${n \choose 2}$ generators and
$2 {n \choose 3}+3 {n \choose 4}$ quadratic relations, and the element
\begin{gather*}
c:=\sum_{1 \le i < j \le n} y_{i,j}
\end{gather*}
belongs to the center of the Kohno--Drinfeld algebra.
\begin{Definition}\label{definition4.19}
Denote by $4T_{n}^{(0)}$ the quotient of the algebra $4T_{n}$ by the
(two-sided) ideal generated by by the set of elements
$\{y_{i,j}^2,\,1 \le i < j \le n \}$.
More generally, let $\beta$,
$\{q_{ij},\, 1 \le i < j \le n \}$ be the set of parameters, denote by $4HT_n$
the quotient of the algebra $4T_n$ by the two-sided ideal generated by the
set of elements $\{y_{ij}^2- \beta y_{ij}-q_{ij}, \, 1 \le i < j \le n \}$.
\end{Definition}
These algebras are naturally graded, with $\deg( y_{i,j})=1$, $\deg(\beta)=1$,
$\deg( q_{ij})=2$, as well as each of that algebras has a natural f\/iltration by
setting $\deg( y_{i,j})=1$, $\deg( \beta )= 0$, $\deg(q_{ij})=0$, $\forall\, i
\not= j$.
It is clear that the algebra $4T_{n}$ can be considered as the
inf\/initesimal deformation $g_{i,j}:=1+\epsilon y_{i,j}$, $\epsilon \longrightarrow 0$, of the pure braid group $P_{n}$.
There is a natural action of the
symmetric group ${\mathbb{S}}_{n}$ on the algebra $4T_{n}$ (and also on
$4T_{n}^{0}$) which preserves the grading: it is def\/ined by $w \cdot y_{i,j}=
y_{w(i),w(j)}$ for $w \in {\mathbb{S}}_{n}$. The semi-direct product
$\mathbb{Q}{\mathbb{S}}_{n} \ltimes 4T_{n}$ (and also that
$\mathbb{Q}{\mathbb{S}}_{n} \ltimes 4T_{n}^{0}$) is a~Hopf algebra
denoted by ${\cal B}_{n}$ (respectively~${\cal B}_{n}^{(0)})$.
\begin{Remark}\label{remark4.5}
There exists the natural map
\begin{gather*}
{\rm CYB}_{n} \longrightarrow 4T_{n} \qquad \text{given by} \ \ y_{i,j}:=u_{i,j}+u_{j,i}.
\end{gather*}
Indeed, one can easily check that
\begin{gather*}
[y_{ij},y_{ik}+y_{jk}] = w_{ijk}+w_{jik}-w_{kij}-w_{kji} ,
\end{gather*}
see Section~\ref{section2.3.1}, Def\/inition~\ref{definition2.5} for a def\/inition of the classical
Yang--Baxter algebra $CYB_n$, and Section~\ref{section2}, equation~\eqref{equation2.3}, for a def\/inition of the
element~$w_{ijk}$.
\end{Remark}
\begin{Remark}\label{remark4.6}\quad
\begin{itemize}\itemsep=0pt
\item Much as the relations in the algebra $6T_n$ are chosen in a way to
imply (and ``essentially''\footnote{Together with locality and factorization conditions a set of
def\/ining relations in the algebra $6T_n$ is {\it equivalent} to the
commutativity property of Dunkl's elements.}
equivalent) the pair-wise commutativity of the Dunkl elements $\{\theta_{i} \}_{1 \le i \le n}$, the relations in the Kohno--Drinfeld algebra imply (and
``essentially'' equivalent) to pair-wise commutati\-vi\-ty of the Jucys--Murphy
elements (or, equivalently, {\it dual} JM-elements)
$d_j:= \sum\limits_{ 1 \le a < j} y_{aj}$, $2 \le j \le n$ (resp.\ $\overline{d}_i= \sum\limits_{1 \le a \le i} y_{n-i,n-a+1}$, $1 \le i \le n-1 $).
\item It follows from the classical 3-term identity
(``Jacobi identity'')
\begin{gather*
{1 \over (a-b)(a-c)}-{1 \over (a-b)(b-c)}+{1 \over (a-c)(b-c)}=0,
\end{gather*}
that if elements $\{ y_{i,j} \,|\, 1 \le i < j \le n \}$ satisfy the 4-term
algebra relations, see Def\/ini\-tion~\ref{definition4.18}, and $t_1,\dots,t_n, $ a set of
(pair-wise) commuting parameters, then the elements
\begin{gather*}
r_{i,j}:={y_{i,j} \over t{_i}-t{_j}}
\end{gather*}
satisfy the set of def\/in\/ing relations of the 6-term relations algebra $6T_n$,
see Section~\ref{section4.2.1},
Def\/inition~\ref{definition4.9}. In particular, the Knizhnik--Zamolodchikov
elements
\begin{gather*}
{\rm KZ}_{j}:=\sum_{i \not= j}{y_{i,j} \over t{_i}-t{_j}}, \qquad 1 \le j \le n,
\end{gather*}
form a pair-wise commuting family (by def\/inition, we put $y_{i,j}=y_{j,i}$
if $i > j$).
\end{itemize}
\end{Remark}
\begin{Example}\label{example4.7}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] {\it Yang representation of the $4T_n$}.
Let ${\mathbb S}_n$ be the symmetric group acting
identically on the set of variables $\{x_1,\ldots,x_n \}$. Clearly that the
elements $ \{y_{i,j}: = s_{ij} \}_{1 \le i < j \le n}$, $y_{i,j}:=y_{j,i}$
if $i > j$, satisfy the Kohno--Drinfeld relations listed in
Def\/inition~\ref{definition4.18}. Therefore the operators $u_{ij}$ def\/ined by
\begin{gather*}
u_{ij}=(x_i-x_j)^{-1} s_{ij}
\end{gather*}
give rise to a representation of the algebra $3T_n$ on the f\/ield of rational
functions $\mathbb{Q}(x_1,\ldots$, $x_n)$. The Dunkl--Gaudin elements
\begin{gather*}
\theta_i= \sum_{j, j\not= i} y_{ij}, \qquad i=1,\ldots,n
\end{gather*}
correspond to the truncated Gaudin operators acting in the tensor
space $({\mathbb{C}})^{\otimes n}$. Cf.\ Section~\ref{section3.3}.
\item[(2)] Let $A=U({\mathfrak{sl}}(2))$ be the universal enveloping algebra of the
Lie algebra ${\mathfrak{sl}}(2)$. Recall that the algebra ${\mathfrak{sl}}(2)$ is spanned by the
elements~$e$, $f$, $h$, so that $[h,e]=2e$, $[h,f]=-2f$, $[e,f]=h$. Consider
the element $\Omega={1 \over 2} h \otimes h+e \otimes f+f \otimes e$.
Then the map $y_{i,j} \longrightarrow \Omega_{i,j} \in A^{ \otimes n}$
def\/ines a representation of the Kohno--Drinfeld algebra~$4T_n$ on that~$A^{\otimes n}$. The element ${\rm KZ}_j$ def\/ined above, corresponds to the
truncated (or at critical level) rational Knizhnik--Zamolodchikov operator. Cf.\ Section~\ref{section4.2.1}, Example~\ref{example4.4}.
\end{enumerate}
\end{Example}
\begin{Proposition}[T.~Kohno, V.~Drinfeld]\label{proposition4.6}
\begin{gather*}
{\rm Hilb}(4T_{n},t)=\prod_{j=1}^{n-1}(1-jt)^{-1}=\sum_{k \ge 0}
{n+k-1 \brace n-1} t^{k},
\end{gather*}
where ${n \brace k }$ stands for the Stirling
numbers of the second kind, i.e., the number of ways to partition a~set of~$n$
things into~$k$ nonempty subsets.
\end{Proposition}
\begin{Remark} \label{remark4.7} It follows from~\cite{BN} that ${\rm Hilb}(4T_{n},t)$ is equal to
the generating function
\begin{gather*}
1+\sum_{ d \ge 1} v_{d}^{(n)} t^d
\end{gather*}
for the number $v_{d}^{(n)}$ of {\it Vassiliev invariants of order~$d$}
for {\it $n$-strand braids}. Therefore, one has the following equality:
\begin{gather*}
v_{d}^{(n)}= {n+d-1 \brace n-1},
\end{gather*}
i.e., the number of Vassiliev invariants of order~$d$ for $n$-strand braids is equal to the Stirling number of the second kind ${n+d-1 \brace n-1}$.
We {\it expect} that the generating function
\begin{gather*}
1+\sum_{ d \ge 1} {\widehat v}_{d}^{(n)} t^d
\end{gather*}
for the number ${\widehat v}_{d}^{(n)}$ of {\it Vassiliev invariants of
order~$d$} for {\it $n$-strand virtual braids} is equal to the
Hilbert series ${\rm Hilb}(4NT_n,t)$ of the {\it nonsymmetric} Kohno--Drinfeld
algebra~$4NT_n$, see Section~\ref{section4.3.2}.
\end{Remark}
\begin{Proposition}[cf.~\cite{BEE}] \label{proposition4.7}
The algebra $4NT_n,t)$ is Koszul, and
\begin{gather*}
{\rm Hilb} (4NT_n,t)= \left(\sum_{k=0}^{n-1} (k+1)! N(k,n) (-t)^{k} \right)^{-1}, \\
{\rm Hilb}\big((4NT_n)^{!},t\big)=(n-1) ! L_{n-1}^{(\alpha=1)}\big({-}t^{-1}\big) t^{n-1},
\end{gather*}
where $N(k,n):= \frac{1}{n} {n \choose k} {n \choose k+1}$ denotes the
Narayana number, i.e., the number of Dyck $n$-paths with exactly $k$ peaks,
\begin{gather*}
L_n^{(\alpha)}(x)=\frac{x^{\alpha} e^x}{n!} \frac{d^n}{dx^n}\big( e_x x^{n+\alpha} \big)
\end{gather*}
denotes the generalized Laguerre polynomial.
\end{Proposition}
See also Theorem~\ref{theorem4.6} below.
It is well-known that the quadratic dual $4T_n^{!}$ of the Kohno--Drinfeld
algebra $4T_n$ is isomorphic to the Orlik--Solomon algebra of type $A_{n-1}$,
as well as the algebra $3T_n^{\rm anti}$. However the algebra $4T_n^{0}$ is
{\it failed} to be Koszul.
\begin{Examples}\label{examples4.3}
\begin{gather*}
{\rm Hilb}\big(4T_{3}^{0},t\big)=[2]^{2}[3], \qquad
{\rm Hilb}\big(4T_{4}^{0},t\big)=(1,6,19,42,70,90,87,57,23,6,1),\\
{\rm Hilb}\big(\big(4T_{3}^{0}\big)^{!},t\big)(1-t)=(1,2,2,1),\qquad
{\rm Hilb}\big(\big(4T_{4}^{0}\big)^{!},t\big)(1-t)^2=(1,4,6,2,-4,-3),\\
{\rm Hilb}\big(\big(4T_{5}^{0}\big)^{!},t\big)(1-t)^2=(1,8,26,40,24,-3,-6).
\end{gather*}
\end{Examples}
We {\it expect} that ${\rm Hilb}\big((4T_{n}^{0})^{!},t\big)$ is a rational function with
the only pole at $t=1$ of order $[n/2]$, cf.\ Examples~\ref{examples4.77}.
\begin{Remark} \label{remark4.8}
One can show that if $n \ge 4$, then ${\rm Hilb}(4T_{n}^{0},t) <
{\rm Hilb}(3T_{n}^{0},t)$ {\it contrary} to the statement of Conjecture~9.6
from~\cite{K3}.
\end{Remark}
\subsubsection[Nonsymmetric Kohno--Drinfeld algebra $4NT_n$, and McCool algebras ${\cal P}\Sigma_n$ and ${\cal P}\Sigma_n^{+}$]{Nonsymmetric Kohno--Drinfeld algebra $\boldsymbol{4NT_n}$,\\ and McCool algebras $\boldsymbol{{\cal P}\Sigma_n}$ and $\boldsymbol{{\cal P}\Sigma_n^{+}}$}\label{section4.3.2}
\begin{Definition} \label{definition4.20}
{\it The nonsymmetric $4$-term relations algebra} (or the
nonsymmetric Kohno--Drinfeld algebra) $4NT_{n}$ is an
associative algebra (say over~$\mathbb{Q}$) with the set of generators
$y_{i,j}$, $1 \le i \not= j \le n$, subject to the following relations
\begin{enumerate}\itemsep=0pt
\item[1)] $y_{i,j}$ and $y_{k,l}$ are commute if $i$, $j$, $k$, $l$ are all distinct;
\item[2)] $[y_{i,j},y_{i,k}+y_{j,k}]= 0$ if $i$, $j$, $k$ are all distinct.
\end{enumerate}
\end{Definition}
We denote by $4NT_{n}^{+}$ the quotient of the algebra $4NT_{n}$ by the
two-sided ideal generated by the elements
$\{y_{ij}+y_{ji}=0, \, 1 \le i \not= j \le n \}$.
\begin{Theorem} \label{theorem4.6}
One has
\begin{gather*}
{\rm Hilb}(4NT_n,t)={\rm Hilb}({\rm CYB}_n,t), \qquad {\rm Hilb}\big(4NT_n^{+},t\big)= {\rm Hilb}(6T_n,t)
\end{gather*}
for all $n \ge 2$.
\end{Theorem}
We expect that the both algebras $4NT_n$ and $4NT_n^{+}$ are {\it Koszul}.
\begin{Definition} \label{definition4.21}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] Def\/ine {\it the McCool algebra ${\cal P}\Sigma_n$} to be the
quotient of the nonsymmetric Kohno--Drinfeld algebra $4NT_n$ by the
two-sided ideal generated by the elements
$\{ y_{ik}y_{jk}-y_{jk}y_{ik} \}$
for all pairwise distinct~$i$,~$j$ and~$k$.
\item[(2)] Def\/ine {\it the upper triangular McCool algebra
${\cal P}\Sigma_n^{+}$} to be the quotient of the McCool algebra
${\cal P}\Sigma_n$ by the two-sided ideal generated by the elements
$\{ y_{ij}+y_{ji} \}$,
$1 \le i \not= j \le n$.
\end{enumerate}
\end{Definition}
\begin{Theorem} \label{theorem4.7}
The quadratic duals of the algebras ${\cal P}\Sigma_n$ and
${\cal P}\Sigma_n^{+}$ have the following Hilbert polynomials
\begin{gather*}
{\rm Hilb}\big({\cal P}\Sigma_n ^{!},t\big)=(1+n t)^{n-1},\qquad
{\rm Hilb}\big(\big({\cal P}\Sigma_n^{+}\big)^{!},t\big)= \prod_{j=1}^{n-1} (1+j t).
\end{gather*}
\end{Theorem}
\begin{Proposition}\label{proposition4.8}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] The quadratic dual ${\cal P}\Sigma_n^{!}$ of the algebra
${\cal P}\Sigma_n$ admits the
following description. It is gene\-ra\-ted over $\mathbb{Z}$ by the set of pairwise
anticommuting elements $\{y_{ij},\, 1 \le i \not= j \le n \}$,
subject to the set of relations
\begin{enumerate}\itemsep=0pt
\item[$(a)$] $y_{ij}^2=0$, $y_{ij} y_{ji}=0$, $1 \le i \not= j \le n$,
\item[$(b)$] $y_{ik} y_{jk}=0$ for all distinct $i$, $j$, $k$,
\item[$(c)$] $y_{ij} y_{jk}+y_{ik} y_{ij}+y_{kj} y_{ik}=0$ for all distinct
$i$, $j$, $k$.
\end{enumerate}
\item[$(2)$] The quadratic dual $({\cal P}\Sigma_n^{+})^{!}$ of the algebra
${\cal P}\Sigma_n^{+}$ admits the
following description. It is generated over $\mathbb{Z}$ by the set of pairwise
anticommuting elements $\{z_{ij}, \,1 \le i < j \le n \}$, subject to
the set of relations
\begin{enumerate}\itemsep=0pt
\item[$(a)$] $z_{ij}^2=0$ for all $i < j$,
\item[$(b)$] $z_{ij} z_{jk}=z_{ij} z_{ik}$ for all $1 \le i< j < k \le n$.
\end{enumerate}
\end{enumerate}
\end{Proposition}
\begin{Comments}[the McCool groups and algebras]\label{comments4.5}
{\it The McCool group $P\Sigma_n$} is, by def\/inition, the group of pure
symmetric automorphisms of the free group $F_n$ consisting of all
automorphism that, for a~f\/ixed basis
$\{x_1,\ldots,x_n \}$, send each $x_i$ to a~conjugate of itself. This group is
generated by automorphisms $\alpha_{ij}$, $1 \le i \not= j \le n$, def\/ined by
\begin{gather*}
\alpha_{ij}(x_k)=\begin{cases}
x_j x_i x_j^{-1}, & k=i, \\
x_k, & k \not= i.
\end{cases}
\end{gather*}
McCool have proved that the relations
\begin{gather*}
\begin{cases}
[\alpha_{ij}, \alpha_{kl}]=1, & i,\,j,\, k,\, l \ \ \text{are distinct}, \\
[\alpha_{ij},\alpha_{ji}]=1, & i \not= j, \\
[\alpha_{ij},\alpha_{ik} \alpha_{jk}]=1,
& i,\,j,\,k \ \ \text{are distinct}.
\end{cases}
\end{gather*}
form the set of def\/ining relations for the group $P\Sigma_n$
The subgroup of $P\Sigma_n$ generated by the~$\alpha_{ij}$ for
$1 \le i < j \le n$ is denoted by~$P\Sigma_n^{+}$ and is called by {\it upper
triangular McCool group}. It is easy to see that the McCool algebras
${\cal P}\Sigma_n$ and ${\cal P}\Sigma_n^{+}$ are the ``inf\/initesimal
deformations'' of the McCool groups $P\Sigma_n$ and $P\Sigma_n^{+}$
respectively.
\end{Comments}
\begin{Theorem} \label{theorem4.8}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] There exists a natural isomorphism
\begin{gather*}
H^{*}(P\Sigma_n, \mathbb{Z}) \simeq {\cal P}\Sigma_n ^{!}
\end{gather*}
of the quadratic dual ${\cal P}\Sigma_n ^{!}$ of the McCool algebra
${\cal P}\Sigma_n$ and the cohomology ring $H^{*}(P\Sigma_n, \mathbb{Z})$ of
the McCool group~$P\Sigma_n$, see~{\rm \cite{JMM}}.
\item[$(2)$] There exists a natural isomorphism
\begin{gather*}
H^{*}(P\Sigma_n^{+}, \mathbb{Z}) \simeq ({\cal P}\Sigma_n^{+}) ^{!}
\end{gather*}
of the quadratic dual $({\cal P}\Sigma_n^{+}) ^{!}$ of the upper triangular
McCool algebra ${\cal P}\Sigma_n^{+}$ and the cohomology ring
$H^{*}(P\Sigma_n^{+}, \mathbb{Z})$ of the upper triangular McCool group
$P\Sigma_n^{+}$, see~{\rm \cite{CP}}.
\end{enumerate}
\end{Theorem}
\subsubsection[Algebras $4TT_n$ and $4ST_n$]{Algebras $\boldsymbol{4TT_n}$ and $\boldsymbol{4ST_n}$}\label{section4.3.3}
\begin{Definition}\label{definition4.22}\quad
\begin{enumerate}\itemsep=0pt
\item[(I)] {\it Algebra $4TT_n$} is generated over $\mathbb{Z}$ by the set of
elements $\{x_{ij}, \,1 \le i \not= j \le n \}$, subject to the set of
relations
\begin{enumerate}\itemsep=0pt
\item[(1)] $x_{ij} x_{kl}=x_{kl} x_{ij}$ if all $i$, $j$, $k$, $l$ are distinct,
\item[(2)] $[x_{ij}+x_{jk},x_{ik}]=0$, $[x_{ji}+x_{kj},x_{ki}]=0$
if $i < j < k$.
\end{enumerate}
\item[(II)] {\it Algebra $4ST_n$} is generated over~$\mathbb{Z}$ by the set of
elements $\{x_{ij},\, 1 \le i \not= j \le n \}$, subject to the set of
relations
\begin{enumerate}\itemsep=0pt
\item[(1)] $[x_{ij},x_{kl}]=0$, $[x_{ij},x_{ji}]=0$ if $i$, $j$, $k$, $l$ are distinct,
\item[(2)] $[x_{ij},x_{ik}]=[x_{ik},x_{jk}]=[x_{jk},x_{ij}]$,
$[x_{ji},x_{ki}]=[x_{ki},x_{kj}]=[x_{kj},x_{ii}]$,
\item[(3)] $[x_{ij},x_{ki}]=[x_{kj},x_{ij}]=[x_{ji},x_{ik}]=[x_{ik},x_{kj}]=
[x_{ki},x_{jk}]=[x_{jk},x_{ji}]$ if $i <j < k$.
\end{enumerate}
\end{enumerate}
\end{Definition}
\begin{Proposition}\label{proposition4.9}
One has
\begin{gather*}
t \sum_{n\ge 2} {\rm Hilb}\big((4TT_n)^{!},t\big) {z^n \over n!} =
{\exp(-t z) \over (1-z)^{2 t}}-1-t z.
\end{gather*}
Therefore, $\dim (4TT_n)^{!}$ is equal to the number of permutations of the
set $[1,\dots,n+1]$ having no substring $[k,k+1]$; also, for $n \ge 1$ equals
to the maximal permanent of a nonsingular $n \times n$
$(0,1)$-matrix, see {\rm \cite[$A000255$]{SL}}\footnote{See also a~paper by F.~Hivert, J.-C.~Novelli and J.-Y.~Thibon
\cite[Section~3.8.4]{Hivert2008} for yet another combinatorial interpretation of the dimension of the algebra~$(4TT_n)^{!}$.}. Moreover, one has
\begin{gather*}
{\rm Hilb}\big((4ST_n)^{!},t\big)=(1+t)^{n} (1+n t)^{n-2},
\end{gather*}
cf.\ Conjecture~{\rm \ref{conjecture4.9}}.
\end{Proposition}
We expect that The both algebras $4TT_n$ and $4ST_n$ are {\it Koszul}.
\begin{Problem}\label{problem-page79}
Give a combinatorial interpretation of polynomials
${\rm Hilb}((4TT_n)^{!},t)$ and construct a monomial basis in the algebras
$(4TT_n)^{!}$ and $4ST_n$.
\end{Problem}
\subsection[Subalgebra generated by Jucys--Murphy elements in $4T_n^{0}$]{Subalgebra generated by Jucys--Murphy elements in $\boldsymbol{4T_n^{0}}$}\label{section4.4}
\begin{Definition}\label{definition4.23}
{\it The Jucys--Murphy elements} $d_j$,
$2 \le j \le n$, in the quadratic algebra $4T_{n}$ are def\/ined as follows
\begin{gather*
d_j=\sum_{1 \le i < j} y_{i,j}, \qquad j=2,\dots,n.
\end{gather*}
\end{Definition}
It is clear that Jucys--Murphy's elements $d_{j}$ are the inf\/initesimal
deformation of the elements $D_{1,j} \in P_{n}$.
\begin{Theorem}\label{theorem4.9}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] The Jucys--Murphy elements $d_{j}$, $2 \le j \le n$,
commute pairwise in the algebra $4T_{n}$.
\item[$(2)$]In the algebra $4T_{n}^{0}$ the Jucys--Murphy elements~$d_{j}$,
$2 \le j \le n$, satisfy the following relations
\begin{gather*}
(d_2+\cdots+d_j) d_{j}^{2j-3}=0, \qquad 2 \le j \le n.
\end{gather*}
\item[$(3)$] Subalgebra (over $\mathbb{Z}$) in $4T_{n}^{0}$ generated by the
Jucys--Murphy elements $d_2,\dots, d_n$ has the following Hilbert polynomial
$\prod\limits_{j=1}^{n-1} [2j]$.
\item[$(4)$] There exists an $($birational$)$ isomorphism
${\mathbb{Z}} [x_{1},\ldots,x_{n-1}]/J_{n-1} \longrightarrow {\mathbb{Z}} [d_2,\ldots,d_n]$
defined by $d_{j}:= \prod\limits_{i=1}^{n-j} x_{i}$, $2 \le j \le n$,
where $J_{n-1}$ is a $($two-sided$)$ ideal generated by
$e_{i}(x_{1}^{2},\ldots,x_{n-1}^{2})$, $1 \le i \le n-1$, and
$e_{i}(x_1,\ldots,x_{n-1})$ stands for the $i$-th elementary symmetric
polynomial in the variables $x_1,\ldots,x_{n-1}$.
\end{enumerate}
\end{Theorem}
\begin{Remark}\label{remark4.9}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] It is clearly seen that the commutativity of the Jucys--Murphy elements
is equivalent to the validity of the Kohno--Drinfeld relations and the
locality relations among the generators $\{ y_{i,j} \}_{1 \le i < j \le n}$.
\item[(2)] Let's stress that $d_j^{2j-2} \not= 0$ in the algebra $4T_n^{0}$ for $j=3,
\ldots,n$. For example, $d_3^{4}=y_{13} y_{23} y_{13} y_{23}+y_{23} y_{13} y_{23} y_{13} \not= 0$
since $\dim(4T_3^{0})_{4}=1$ and it is generated by the element $d_3^{4}$.
\item[(3)] The map $\iota\colon y_{i,j} \longrightarrow y_{n+1-j,n+1-i}$
preserves the relations~$1)$ and~$2)$ in the def\/inition of the algebra $4T_n$,
and therefore def\/ines an involution of the Kohno--Drinfeld algebra. Hence the
elements
\begin{gather*} {\widehat d}_j:=\sum_{k=j+1}^{n}y_{j,k}=\iota(d_{n+1-j}),\qquad
1 \le j \le n-1
\end{gather*}
also form a pairwise commuting family.
\end{enumerate}
\end{Remark}
\begin{Problems}\label{problems4.1}\quad
\begin{enumerate}\itemsep=0pt
\item[$(a)$] Compute Hilbert series of the algebra $4T_{n}^{0}$ and
its quadratic dual algebra $(4T_{n}^{0})^{!}$.
\item[$(b)$] Describe subalgebra in the algebra $4HT_n$ generated by the
Jucys--Murphy elements $d_j$, $2 \le j \le n$.
\end{enumerate}
\end{Problems}
It is well-known that the Kohno--Drinfeld algebra $4T_n$ is {\it Koszul}, and
its quadratic dual $4T_n^{!}$ is isomorphic to the anticommutative quotient
$3T_n^{0,{\rm anti}}$ of the algebra $3T_n^{(-),0}$.
On the other hand,
if $n \ge 3$ the algebra $4T_{n}^{0}$ is not {\it Koszul}, and its quadratic
dual is isomorphic to the quotient of the ring of polynomials in the set of
anticommutative variables $\{t_{i,j} \,|\, 1 \le i < j \le n \}$, where
{\it we do not impose conditions} $t_{ij}^2=0$, modulo the ideal
generated by Arnold's relations
$\{ t_{i,j}t_{j,k}+t_{i,k}(t_{i,j}-t_{j,k})=0 \}$ for all pairwise
distinct~$i$,~$j$ and~$k$.
\subsection[Nonlocal Kohno--Drinfeld algebra $NL4T_n$]{Nonlocal Kohno--Drinfeld algebra $\boldsymbol{NL4T_n}$}\label{section4.5}
\begin{Definition}\label{definition4.24}
{\it Nonlocal Kohno--Drinfeld algebra $NL4T_n$} is an associative algebra over~$\mathbb{Z}$ with the set of generators $\{y_{ij},\, 1 \le i < j \le n \}$
subject to the set of relations
\begin{enumerate}\itemsep=0pt
\item[(1)] $y_{ij} y_{kl}=y_{kl} y_{ij}$ if $(i-k)(i-l)(j-k)(j-l) > 0$,
\item[(2)] $\big[y_{ij},\sum\limits_{a=i}^{j} y_{ak}\big]=0$ if $i < j < k$,
\item[(3)] $\big[y_{jk},\sum\limits_{a=j}^{k} y_{ia}\big]=0$ if $i < j < k$.
\end{enumerate}
\end{Definition}
It's not dif\/f\/icult to see that relations (1)--(3) imply the following
relations
\begin{enumerate}\itemsep=0pt
\item[(4)] $\big[x_{ij},\sum\limits_{a=i+1}^{j-1} (y_{ia}+y_{aj})\big]=0$ if~$i < j$.
\end{enumerate}
Let's introduce in the nonlocal Kohno--Drinfeld algebra $NL4T_n$ the
{\it Jucys--Murphy elements} (JM-elements for short) $d_{j}$
and the {\it dual JM-elements} ${\hat d}_j$ as follows
\begin{gather}\label{equation4.11}
d_{j}=\sum_{a=1}^{j-1} y_{aj}, \qquad {\hat d}_{j}=\sum_{a=n-j+2}^{n} y_{n-j+1,a}, \qquad j=2,\ldots,n.
\end{gather}
It follows from relations (1) and (2) (resp.~(1) and~(3)) that the
Jucys--Murphy elements $d_2,\ldots,d_n$ (resp.\ ${\hat d}_{2},\ldots,{\hat d}_{n}$)
form a commutative subalgebra in the algebra $NL4T_n$. Moreover, it follows from relations (1)--(3) that the element $c_1:=\sum\limits_{j=2}^{n}d_{j}= \sum\limits_{j=2}^{n} {\hat d}_{j}$ belongs to the center of the algebra $NL4T_n$.
\begin{Theorem} \label{theorem4.10}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] The algebra $NL4T_n$ is Koszul, and
\begin{gather*}
{\rm Hilb}\big((NL4T_n)^{!},t\big)=\sum_{k=0}^{n-1} C_k~{n+k-1 \choose 2k} t^{k},
\end{gather*}
where $C_k={1 \over k+1} {2k \choose k}$ stands for the $k$-th Catalan number.
\item[$(2)$] The quadratic dual $(NL4T_{n})^{!}$ of the nonlocal Kohno--Drinfeld
algebra $NL4T_n$ is an associative algebra generated by the set of
mutually anticommuting elements $\{ t_{ij},\, 1 \le i < j \le n \}$
subject to the set of relations
\begin{itemize}\itemsep=0pt
\item $t_{ij}^2=0$ if $1 \le i < j \le n$,
\item Arnold's relations: $t_{ij} t_{jk}+t_{ik} t_{ij}+t_{jk} t_{ik}=0$ if $i < j < k$,
\item disentanglement relations: $t_{ik} t_{jl}+t_{il} t_{ik}+ t_{jl} t_{il}=0$ if $i < j < k < l$.
\end{itemize}
\end{enumerate}
\end{Theorem}
Therefore the algebra $(NL4T_{n})^{!}$ is the quotient of the Orlik--Solomon algebra ${\rm OS}_n$ by the ideal generated by Disentanglement relations, and $\dim((NL4T_{n+1})^{!})$ is equal to the number of Schr\"oder paths,
i.e., paths from $(0,0)$ to $(2n,0)$ consisting of steps $U=(1,1)$,
$D=(1,-1)$, $H=(2,0)$ and never going below the $x$-axis. The Hilbert
polynomial ${\rm Hilb}((NL4T_n)^{!},t)$ is the generating function of such paths with respect to the number of $U's$, see \cite[$A088617$]{SL}.
\begin{Remark} \label{remark4.10}
Denote by $H_n(q)$ ``the normalized'' Hecke algebra of type
$A_n$, i.e., an associative algebra generated over $\mathbb{Z}[q,q^{-1}]$ by elements
$T_1,\ldots,T_{n-1}$ subject to the set of relations
\begin{enumerate}\itemsep=0pt
\item[(a)] $T_i T_j=T_j T_i$ if $|i-j| > 1$, $T_i T_j T_i=T_j T_i T_j$
if $|i-j| = 1$,
\item[(b)] $T_i^{2}=(q-q^{-1})~T_i+1$ for $i=1,\ldots,n-1$.
\end{enumerate}
\end{Remark}
If $1 \le i < j \le n-1$, let's consider elements $T_{(ij)}:=
T_i T_{i+1} \cdots T_{j-1} T_{j} T_{j-1} \cdots T_{i+1} T_i$.
\begin{Lemma}\label{lemma4.5}
The elements $\{ T_{(ij)}, \, 1 \le i < j < n-1 \}$ satisfy the
defining relations of the non-local Kohno--Drinfeld algebra $NL4T_{n-1}$, see
Definition~{\rm \ref{definition4.24}}.
Therefore the map $y_{ij} \rightarrow H_{(ij)}$ defines a epimorphism
$\iota_{n}\colon NL4T_n \longrightarrow H_{n+1}(q)$.
\end{Lemma}
\begin{Definition}\label{definition4.25}
Denote by ${\cal {NL}}4T_n$ the quotient of the non-local
Kohno--Drinfeld algebra $NL4T_n$ by the two-sided ideal ${\cal I}_n$ generated
by the following set of degree three elements:
\begin{alignat*}{3}
& (1) \quad && z_{ij}:=y_{i,j+1} y_{ij} y_{j,j+1}-y_{j,j+1} y_{ij} y_{i,j+1} \qquad \text{if} \ \ 1 \le i < j \le n,&\\
& (2) \quad && u_{i}:= y_{i,i+1} \left( \sum_{a=1}^{i-1} \sum_{b=1,\, b
\not=a}^{i-1} y_{ai} y_{b,i+1} \right) - \left( \sum_{a=1}^{i-1} \sum_{b=1,\, b \not=a}^{i-1} y_{b,i+1} y_{ai} \right) y_{i,i+1}& \\
&&& \text{if} \ \ 1 \le i \le n-1,& \\
& (3) \quad && v_{i}:=y_{i,i+1}\left( \sum_{a=i+1}^{n} \sum_{b=i+1,\, b \not= a}^{n} y_{i+1,a} y_{i,b} \right) -
\left( \sum_{a=i+1}^{n} \sum_{b=i+1,\, b \not= a}^{n}
y_{i+1,a} y_{i,b} \right) y_{i,i+1},& \\
&&& \text{if} \ \ 1 \le i \le n-1.&
\end{alignat*}
\end{Definition}
\begin{Proposition}\label{proposition4.10}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] The ideal ${\cal T}_n$ belongs to the kernel of the epimorphism
$\iota_n\colon {\cal I}_n \subset \operatorname{Ker}(\iota_n)$,
\item[$(2)$] Let $d_{2},\ldots,d_n$ $($resp.\ ${\hat d}_2, \ldots, {\hat d}_n )$
be the Jucys--Murphy elements $($resp.\ dual JM-elements$)$ in the algebra
${\cal {NL}}4T_n$ given by the formula~\eqref{equation4.11}.
\end{enumerate}
Then the all elementary
symmetric polynomials $e_k(d_2,\ldots,d_n)$ $($resp.\
$e_k({\hat d}_2,\ldots,{\hat d}_n))$ of deg\-ree~$k$, $ 1 \le k < n$,
in the Jucys--Murphy elements $d_2,\ldots,d_n$, $($resp.\ in the dual JM-elements
${\hat d}_2,\ldots,{\hat d}_n)$, commute in the algebra
${\cal {NL}}4T_n$ with the all elements $y_{i,i+1}$, $i=1,\ldots,n-1$.
\end{Proposition}
Therefore, there exists an epimorphism of algebras
${\cal {NL}}4T_n \longrightarrow H_n(q)$, and images of the elements
$e_k(d_2,\ldots,d_n)$ (resp. $e_k({\hat d}_2,\ldots,{\hat d}_n$), $1 \le k < n$, belongs to the center of the
``normalized'' Hecke algebra $H_n(q)$, and in fact {\it generate}
the center of algebra~$H_n(q)$.
Few comments in order:
\begin{enumerate}\itemsep=0pt
\item[(A)] Let $N{\ell}4T_n$ be an associative algebra over~$\mathbb{Z}$ with the set of
generators $\{ y_{ij},\, 1 \le i < j \le n \}$ subject to the set of relations
\begin{enumerate}\itemsep=0pt
\item[(1)] $y_{ij} y_{kl}=y_{kl} y_{ij}$ if $(i-k)(i-l)(j-k)(j-l) > 0$,
\item[(2)] $\big[y_{ij},\sum\limits_{a=i}^{j} y_{ak}\big]=0$ if $i < j < k$.
\end{enumerate}
\end{enumerate}
\begin{Proposition} \label{proposition4.11}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] The algebra $N{\ell}4T_n$ is Koszul and has the Hilbert
series equals to
\begin{gather*}
{\rm Hilb}(N{\ell}4T_n,t)= \left( \sum_{k=0}^{n-1} (-1)^{k} N(k,n) t^{k} \right)^{-1},
\end{gather*}
where $N(k,n):= {1 \over n} {n \choose k} {n \choose k+1}$ denotes the
Narayana number, i.e., the number of Dyck $n$-paths with exactly~$k$ peaks,
see, e.g., {\rm \cite[$A001263$]{SL}}.
Therefore, $\dim(N{\ell}4T_n)^{!}={1 \over n+1}{2 n \choose n}$, the $n$-th
Catalan number.
\item[$(2)$] Elementary symmetric polynomials $e_k(d_2,\ldots,d_n)$ of degree
$k$, $ 1 \le k < n$, in the Jucys--Murphy elements $d_2,\ldots,d_n$,
commute in the algebra ${N{\ell}}4T_n$ with the all elements $y_{i,i+1}$, $i=1,\ldots,n-1$.
\end{enumerate}
\end{Proposition}
\begin{enumerate}\itemsep=0pt
\item[(B)] The kernel of the epimorphism
${\cal {NL}}4T_n \longrightarrow H_n(q)$ contains the elements
\begin{gather*}
\{y_{i,i+1} y_{i+1,i+2} y_{i,i+1}-y_{i+1,i+2} y_{i,i+1} y_{i+1,i+2},\, i=1,
\ldots,n-2 \}, \\
\big\{T_{i,i+1}^2-\big(q-q^{-1}\big) T_{i,i+1}-1 \big\},
\end{gather*}
as well as the following set of commutators
\begin{gather*}
[y_{ij},e_{k}(d_{i},\ldots,d_j)], \qquad 1 \le k \le j-i+1.
\end{gather*}
It is an interesting {\it task} to f\/ind def\/ining relations among the
Jucys--Murphy elements $\{d_j, j=2,\ldots,n\}$ in the algebra $NL4T_n$ or that
$N{\ell}4T_n$. We {\it expect} that the Jucys--Murphy element~$d_k$ satisf\/ies
the following relation (= minimal polynomial) in the {\it Hecke algebra}
$H_n(q)$, $n \ge k$,
\begin{gather*
\prod_{a=1}^{k-1} \left(d_k-{q-q^{2a+1} \over 1-q^{2}}\right) \left(d_k+{q^{-1}-q^{-2a-1}
\over 1-q^{-2}}\right)=0.
\end{gather*}
\end{enumerate}
\subsubsection{On relations among JM-elements in Hecke algebras}\label{section4.5.1}
Let $H_n(q)$ be the ``normalized'' Hecke algebra of type $A_n$, see
Remark~\ref{remark4.10}. Let $\lambda \vdash n$ be a~partition of~$n$. For a~box
$x=(i,j) \in \lambda$ def\/ine
\begin{gather*
c_{\lambda}(x;q):=q {1-q^{2 (j-i)} \over 1-q^2}.
\end{gather*}
It is clear that if $q=1$, $c_{q=1}(x)$ is equal to the content $c(x)$ of
a box $x \in \lambda$. Denote by
\begin{gather*}
\Lambda_q^{(n)} = \mathbb{Z}\big[q,q^{-1}\big] [z_1,\ldots,z_n]^{{\mathbb S}_n}
\end{gather*}
the space of symmetric polynomials over the ring $\mathbb{Z}[q,q^{-1}]$ in variables
$\{z_1,\ldots,z_n \}$.
\begin{Definition} \label{definition4.26} Denote by $J_q^{(n)}$ the set of symmetric polynomials $f \in
\Lambda_q^{(n)}$ such that for any {\it partition} $\lambda \vdash n$
one has
\begin{gather*}
f(c_{\lambda}(x;q) \vert x \in \lambda) = 0.
\end{gather*}
\end{Definition}
For example, one can check that symmetric polynomial
\begin{gather*}
e_1^2-\big(q^2+1+q^{-2}\big) e_2 -2\big(q-q^{-1}\big) e_1 -3
\end{gather*}
belongs to the set $J_q^{(3)}$.
Finally, denote by $ {\mathbb J}_q^{(n)}$~the ideal in the ring $\mathbb{Z}[q,q^{-1}]
[z_1, \ldots,z_n]$ generated by the set $J_q^{(n)}$.
\begin{Conjecture}\label{conjecture4.9}
The algebra over $\mathbb{Z}[q,q^{-1}]$ generated by the Jucys--Murphy
elements $d_2$, $\ldots,d_n$ corresponding to the Hecke algebra $H_n(q)$ of
type $A_{n-1}$, is isomorphic to the quotient of the algebra $\mathbb{Z}[q,q^{-1}][z_1,\ldots,z_n]$ by the ideal ${\mathbb J}_q^{(n)}$.
\end{Conjecture}
It seems an interesting {\it problem} to f\/ind a minimal set of generators for
the ideal ${\mathbb J}_q^{(n)}$.
\begin{Comments}\label{comments4.6}
Denote by $JM(n)$ the algebra over $\mathbb{Z}$ generated by the
JM-elements $d_2$, $\ldots,d_n$,
$\deg(d_i)=1$, $\forall\, i$, corresponding to the symmetric group
${\mathbb S}_n$. In this case one can check Conjecture~\ref{conjecture4.9} for $n <8$, and
compute the Hilbert polynomial(s) of the associated graded algebra(s)
${\rm gr}(JM(n))$. For example\footnote{I would like to thank Dr.\ S.~Tsuchioka for computation the Hilbert
polynomials ${\rm Hilb}(JM(n),t)$, as well as the sets of def\/ining relations
among the Jucys--Murphy elements in the symmetric group $\mathbb{S}_n$ for
$n \le 7$.}
\begin{gather*}
{\rm Hilb}({\rm gr}(JM(2) ,t)=(1,1), \qquad {\rm Hilb}({\rm gr}(JM(3) ,t)=(1,2,1),\\
{\rm Hilb}({\rm gr}(JM(4) ,t)=(1,3,4,2), \qquad {\rm Hilb}({\rm gr}(JM(5) ,t)=(1,4,8,9,4),\\
{\rm Hilb}({\rm gr}(JM(6) ,t)=(1,5,13,21,21,12,3), \\
{\rm Hilb}({\rm gr}(JM(7) ,t)=(1,6,19,40,59,60,37,10).
\end{gather*}
It seems an interesting {\it task} to f\/ind a combinatorial interpretation of
the polynomials \linebreak
${\rm Hilb}({\rm gr}(JM(n)),t)$ in terms of standard Young tableaux of
size~$n$.
\end{Comments}
Let $\{ \chi^{\lambda},\, \lambda \vdash n \}$ be the characters of the
irreducible representations of the symmetric group~$\mathbb{S}_n$, which form
a basis of the center ${\cal{Z}}_n$ of the group ring $\mathbb{Z}[\mathbb{S}_n]$. The
famous result by A.~Jucys~\cite{J} states that for any symmetric polynomial
$f(z_1,\ldots,z_n)$ the character expansion of $f(d_2,\dots,d_n,0) $ $\in
{\cal{Z}}_n$ is
\begin{gather*
f(d_2,\ldots,d_n,0)= \sum_{\lambda \vdash n} \frac{f(C_{\lambda})}{H_{\lambda}} \chi^{\lambda},
\end{gather*}
where $H_{\lambda}= \prod\limits_{x \in \lambda}~h_x$ denotes the product of all
{\it hook-lengths} of $\lambda$, and $C_{\lambda}:= \{c(x) \}_{x \in
\lambda}$ denotes the set of {\it contents} of all boxes of $\lambda$.
Recall that the Jucys--Murphy elements $\big\{ d_j^{H}\big\}_{2 \le j \le n}$ in the
(normalized) Hecke algebra $H_n(q)$ are def\/ined as follows: $d_{j}^{H}:= \sum\limits_{i < j} T_{(ij)}$, where $T_{(ij)}:= T_{i} \cdots T_{j-1} T_{j} T_{j-1} \cdots T_{i}$. Finally denote by $H_{\lambda}(q)$ and $C_{\lambda}^{(q)}$ the
hook polynomial and the set $\{c_{\lambda}{x;q) \}_x \in \lambda}$. Then for
any symmetric polynomial $f(z_1,\ldots,z_n)$ one has
\begin{gather*
f\big(d_2^H,\ldots,d_n^H,0\big)= \sum_{\lambda \vdash n} \frac{f(C_{\lambda}^{(q)})}
{H_{\lambda}(q)} \chi_{q}^{\lambda},
\end{gather*}
where $\chi_{q}^{\lambda}$ denotes the $q$-character of the algebra~$H_{n(q)}$.
Therefore, if $f \in J_{q}^{(n)}$, then $f\big(d_{2}^H,\ldots,d_{n}^{H},0\big)=0$. It
is an open {\it problem} to prove/disprove that if
$f\big(d_{2}^H,\ldots,d_{n}^{H},0\big)=0$, then $f\big(C_{\lambda}^{(q)}\big) = 0$ for all
partitions of size $n$ (even in the case $q=1$).
\subsection{Extended nil-three term relations algebra and DAHA, cf.~\cite{Ch}}\label{section4.6}
Let $ A:=\{q,t, a,b,c,h,e,f, \ldots \}$ be a set of parameters.
\begin{Definition}\label{definition4.27} Extended nil-three term relations algebra
$3{\mathfrak{T}}_{n}$ is an associative algebra over ${\mathbb{Z}}[q^{\pm 1},
t^{\pm 1},a,b,c,h,e,\ldots]$ with the set of generators $\{u_{i,j},\, 1
\le i \not= j \le n, x_i,\, 1 \le i \le n, \pi \}$ subject to the set of relations
\begin{enumerate}\itemsep=0pt
\item[(0)] $u_{i,j}+u_{j,i}=0$, $u_{i,j}^2 =0$,
\item[(1)] $x_i x_j=x_j x_i$, $u_{i,j} u_{k,l}=u_{k,l} u_{i,j}$ if $i$, $j$, $k$, $l$ are
distinct,
\item[(2)] $ x_i u_{kl}= u_{k,l} x_i$ if $i \not= k,l$,
\item[(3)] $x_{i} u_{i,j}= u_{i,j} x_j +1$, $ x_{j} u_{i,j}=u_{i,j} x_{i} -1$,
\item[(4)] $u_{i,j} u_{j,k}+u_{k,i} u_{i,j}+u_{j,k} u_{k,i} =0$ if $i$, $j$, $k$ are
distinct,
\item[(5)] $\pi x_{i}= x_{i+1} \pi$ if $1 \le i < n$, $\pi x_{n} = t^{-1} x_{1}
\pi$,
\item[(6)] $\pi u_{ij}=u_{i+1,j+1}$ if $1 \le i < j < n$, $\pi^{j} u_{n-j+1,n}=
t u_{1,j} \pi^{j}$, $2 \le j \le n$.
\end{enumerate}
\end{Definition}
Note that the algebra $3\mathfrak{T}_n$ contains also the set of elements
$\{ \pi^a u_{jn}, \, 1 \le a \le n-j \}$.
\begin{Definition}[cf.\ \cite{LSZ}]\label{definition4.28} Let $1 \le i < j \le n$, def\/ine
\begin{gather*}
T_{i,j}= a+(b x_i+c x_j+h+e x_i x_j) u_{i,j}.
\end{gather*}
\end{Definition}
\begin{Lemma} \label{lemma4.6}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] $T_{i,j}^2 =(2a+b-c) T_{i,j} -a(a+b-c)$ if $a=0$, then $T_{ij}^2 =
(b-c) T_{ij}$.
\item[$(2)$] Coxeter relations
\begin{gather*}
T_{i,j}T_{j,k}T_{i,j} = T_{j,k}T_{i,j}T_{j,k},
\end{gather*}
are valid, if and only if
the following relation holds
\begin{gather}\label{equation4.16}
(a+b)(a-c)+h e =0.
\end{gather}
\item[$(3)$] Yang--Baxter relations
\begin{gather*}
T_{i,j} T_{i,k} T_{j,k}= T_{j,k} T_{i,k} T_{i,j}
\end{gather*}
are valid if and only if $b=c=e=0$, i.e., $T_{ij}=a+d u_{ij}$.
\item[$(4)$] $T_{ij}^2=1$ if and only if $a= \pm 1$, $c=b \pm 2$, $he= (b\pm 1)^2$.
\item[$(5)$] Assume that parameters $a$, $b$, $c$, $h$, $e$ satisfy the conditions~\eqref{equation4.16} and
that $b c+1=h e$. Then
\begin{gather*}
T_{ij} x_i T_{ij} =x_j+ (h+(a+b)(x_i+x_j)+e x_i x_j) T_{ij}.
\end{gather*}
\item[$(6)$] Quantum Yang--Baxterization. Assume that parameters $a$, $b$, $c$, $h$, $e$
satisfy the conditions~\eqref{equation4.16} and that $\beta:=2 a+b-c \not= 0$. Then
$($cf.\ {\rm \cite{IK, LLT}} and the literature quoted therein$)$
the elements $R_{ij}(u,v):= 1+ \frac{\lambda-\mu}
{\beta \mu} T_{ij}$ satisfy the twisted quantum Yang--Baxter relations
\begin{gather*} R_{ij}(\lambda_i,\mu_j) R_{jk}(\lambda_i,\nu_k) R_{ij}(\mu_j,\nu_k)= R_{jk}(\mu_j,\nu_k) R_{ij}(\lambda_i,\nu_k) R_{jk}(\lambda_i,\mu_j),
\qquad i < j < k,
\end{gather*}
where $\{\lambda_i,\mu_i,\nu_i \}_{1 \le i \le n}$ are parameters.
\end{enumerate}
\end{Lemma}
\begin{Corollary}\label{corollary4.6}
If $(a+b)(a-c)+h e=0$, then for any permutation
$w \in {\mathbb{S}}_n$ the element
\begin{gather*}
T_w:=T_{i_{1}} \cdots T_{i_{l}} \in 3 \mathfrak{T}_n,
\end{gather*}
where $w=s_{i_{1}} \cdots s_{i_{l}}$ is any reduced decomposition of~$w$, is well-defined.
\end{Corollary}
\begin{Example} \label{example4.8}
Each of the set of elements
\begin{gather*}
s_i^{( h)}=1+(x_{i+1}-x_i + h) u_{i,i+1}
\end{gather*}
and
\begin{gather*}
t_i^{(h)}=-1+(x_i-x_{i+1}+ h(1+x_i)(1+x_{i+1}) u_{ij}, \qquad i=1,\ldots,n-1,
\end{gather*}
by itself generate the symmetric group ${\mathbb{S}}_n$.
\end{Example}
\begin{Comments} \label{comments4.7}
Let $A=(a,b,c,h,e)$ be a sequence of integers satisfying the conditions~\eqref{equation4.16}. Denote by $\partial_i^{A}$ the divided dif\/ference operator
\begin{gather*}
\partial_{i}^{A}= (a+(b x_i+c x_{i+1}+h+e x_i x_{i+1}) \partial_i, \qquad i=1,\ldots,n-1.
\end{gather*}
It follows from Lemma~\ref{lemma4.5} that the operators $\{ \partial_{i}^{A}\}_{1 \le i
\le n}$ satisfy the Coxeter relations
\begin{gather*}
\partial_{i}^{A} \partial_{i+1}^{A} \partial_{i}^{A}=\partial_{i+1}^{A} \partial_{i}^{A} \partial_{i+1}^{A},\qquad i=1,\ldots,n-1.
\end{gather*}
\end{Comments}
\begin{Definition}\label{definition4.29}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] Let $w \in \mathbb{S}_n$ be a permutation. Def\/ine the generalized
Schubert polynomial corresponding to permutation~$w$ as follows
\begin{gather*}
\mathfrak{S}_w^{A}(X_n) = \partial_{w^{-1} w_{0}}^{A} x^{\delta_n},
\end{gather*}
where
\begin{gather*}
x^{\delta_n}:= x_1^{n-1} x_2^{n-2} \cdots x_{n-1},
\end{gather*}
and $w_0$ denotes the longest element in the symmetric group~$\mathbb{S}_{n}$.
\item[(2)] Let $\alpha$ be a composition with at most~$n$ parts, denote by
$w_{\alpha} \in \mathbb{S}_n$ the permutation such that $w_{\alpha}(\alpha)=
\overline{\alpha}$, where $\overline{\alpha}$ denotes a unique partition
corresponding to composition~$\alpha$.
\end{enumerate}
\end{Definition}
\begin{Proposition}[\cite{K5}] \label{proposition4.12} Let $w \in \mathbb{S}_n$ be a permutation.
\begin{itemize}\itemsep=0pt
\item If $A=(0,0,0,1,0)$, then $\mathfrak{S}_{w}^{A}(X_n)$ is equal to
the Schubert polynomial $\mathfrak{S}_{w}(X_n)$.
\item If $A=(-\beta,\beta,0,1,0)$, then $\mathfrak{S}_{w}^{A}(X_n)$ is
equal to the $\beta$-Grothendieck polynomial $\mathfrak{G}_w^{({\beta})}(X_n)$
introduced in~{\rm \cite{FK1}}.
\item If $A=(0,1,0,1,0)$ then $\mathfrak{S}_{w}^{A}(X_n)$ is equal to
the dual Grothendieck polynomial~{\rm \cite{K5, L}}.
\item If $A=(-1,2,0,1,1)$, then $\mathfrak{S}_{w}^{A}(X_n)$ is equal to
the Di~Francesco--Zinn-Justin polynomials and studied in~{\rm \cite{DZ0,DZ,DZ1}} and {\rm \cite{K5}}.
\end{itemize}
In all cases listed above the polynomials $\mathfrak{S}_{w}^{A}(X_n)$ have
non-negative integer coefficients.
\begin{itemize}\itemsep=0pt
\item If $A=(1,-1,1,-h,0)$, then $\mathfrak{S}_{w}^{A}(X_n)$ is equal to
the $h$-Schubert polynomials introduced in~{\rm \cite{K5}}.
\end{itemize}
Define the generalized key or Demazure polynomial corresponding to a
composition~$\alpha$ as follows
\begin{gather*}
K_{\alpha}^{A}(X_n)= \partial_{w_{\alpha}} x^{\overline \alpha}.
\end{gather*}
\begin{itemize}\itemsep=0pt
\item If $A=(1,0,1,0,0)$, then $K_{\alpha}^{A}(X_n)$ is equal to key $($or
Demazure$)$ polynomial corresponding to~$\alpha$.
\item If $A=(0,0,1,0,0)$, then $K_{\alpha}^{A}(X_n)$ is equal to the
reduced key polynomial introduced in~{\rm \cite{K5}}.
\item If $A=(1,0,1,0,\beta)$, then $K_{\alpha}^{A}(X_n)$ is equal to the
key Grothendieck polynomial $KG_{\alpha}(X_n)$ introduced in~{\rm \cite{K5}}.
\item If $A=(0,0,1,0,\beta)$, then $K_{\alpha}^{A}(X_n)$ is equal to
the reduced key Grothendieck polynomial~{\rm \cite{K5}}.
\end{itemize}
In all cases listed above the polynomials $\mathfrak{S}_{w}^{A}(X_n)$ have
non-negative integer coefficients.
\end{Proposition}
{\samepage
\begin{Exercises}\label{exercises4.5}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] Let $b$, $c$, $h$, $e$ be a collection of integers, def\/ine elements $P_{ij}:=
f_{ij} u_{ij} \in 3 \mathfrak{T}$, where $f_{ij}:= b x_i+c x_j+h+e x_i x_j$.
Show that
\begin{itemize}\itemsep=0pt
\item $P_{ij}^2= (b-c) P_{ij}$,
\item $P_{ij} P_{jk} P_{ij}=f_{ij} f_{ik} f_{jk} u_{ij} u_{jk} u_{ij}+(b c -e h) P_{ij}$,\\
$P_{jk} P_{ij} P_{jk} = f_{ij} f_{ik} f_{jk} u_{ij} u_{jk} u_{ij} -(b c- e h) P_{jk}$.
\end{itemize}
\item[$(2)$] Assume that $a=q$, $b=-q$, $c=q^{-1}$, $h=e=0$, and introduce elements
\begin{gather*}
e_{ij}:=\big(q x_i-q^{-1} x_j\big) u_{ij}, \qquad 1 \le i < j < k \le n.
\end{gather*}
\begin{enumerate}\itemsep=0pt
\item[(a)] Show that if $i$, $j$, $k$ are distinct, then
\begin{gather*}
e_{ij} e_{jk} e_{ij} = e_{ij}+\big(q x_i-q^{-1} x_j\big)\big(q x_i - q^{-1} x_k\big)\big(q x_j -
q^{-1} x_k\big) u_{ij} u_{jk} u_{ij},\\
e_{ij}^2= \big(q+q^{-1}\big)e_{ij}.
\end{gather*}
\item[(b)] Assume additionally that
\begin{gather*}
u_{ij} u_{jk} u_{ij} = 0, \qquad \text{if $i$, $j$, $k$ are distinct}.
\end{gather*}
Show that the elements $\{e_{i}:=e_{i,i+1},\, i=1,\ldots,n-1 \}$,
generate a subalgebra in $3{\mathfrak{L}}_n$ which is isomorphic to the
Temperley--Lieb algebra $TL_n(q+q^{-1})$.
\end{enumerate}
\item[(3)] Let us set $T_i:=T_{i,i+1}$, $i=1,\ldots,n-1$, and def\/ine
\begin{gather*}
T_0:= \pi T_{n-1} \pi^{-1} .
\end{gather*}
Show that if $(a+b)(a-c)+e h=0$, then
\begin{gather*}
T_1 T_0 T_1 =T_1 T_0 T_1, T_{n-1} T_0 T_{n-1}=T_0 T_{n-1} T_0,
\end{gather*}
Recall that $T_i^2= (2 a+b-c) T_i -a(a+b-c)$, $0 \le i \le n-1$.
\end{enumerate}
\end{Exercises}}
In what follows we take $a=q$, $b=-q$, $c=q^{-1}$, $h=e=0$. Therefore,
$T_{i,j}^2=(q-q^{-1})T_{i,j}+1$. We denote by ${\cal{H}}_n(q)$ a subalgebra
in~$3 {\mathfrak{T}}_n$ generated by the elements $T_i:= T_{i,i+1}$, $i=1,\ldots, n-1$.
\begin{Remark} \label{remark4.11}
Let us stress on a dif\/ference between elements $T_{ij}$ as
a part of generators of the algebra $3{\mathfrak{T}}_n$, and the elements
\begin{gather*}
T_{(ij)}:= T_i \cdots T_{j-1} T_{j} T_{j-1} \cdots T_{i} \in {\cal{H}}_n(q).
\end{gather*}
Whereas one has $[T_{ij},T_{kl}]=0$ if $i$, $j$, $k$, $l$ are distinct, the
relation $[T_{(ij)},T_{(kl)}]=0$ in the algebra ${\cal{H}}_n(q)$ holds
(for general~$q$ and $i \le k$) if and only if either one has $i <j < k < l$ or $i< k < l <j$.
\end{Remark}
\begin{Lemma}\label{lemma4.7}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] $T_{ij} T_{kl}= T_{kl} T_{ij}$ if $i$, $j$, $k$, $l$ are distinct,
\item[$(2)$] $T_{i,j} x_i T_{i,j} =x_{j}$ if $1 \le i < j \le n$,
\item[$(3)$] $\pi T_{i,j} = T_{i+1,j+1}$ if $1 \le i < j < n$, $\pi^j T_{n-j+1,n}= T_{1,j} \pi^j$.
\end{enumerate}
\end{Lemma}
\begin{Definition}\label{definition4.30}
Let $1 \le i < j \le n$, set
\begin{gather*}
Y_{i,j}= T_{i-1,j-1}^{-1} T_{i-2,j-2}^{-1} \cdots T_{1,j-i+1}^{-1} \pi^{j-i}
~T_{n-j+i,n} \cdots T_{i+1,j+1} T_{i,j}, \qquad 1 \le i < j \le n,
\end{gather*}
and $Y_n=T_{n-1,n}^{-1} \cdots T_{1,2}^{-1} \pi$.
\end{Definition}
For example,
\begin{gather*}
Y_{1,j}=\pi^{j-1} T_{n-j+1,n} \cdots T_{1,j}, \qquad j \ge 2,\\
Y_{2,j}=T_{1,j-1}^{-1}\pi^{j-2} T_{n-j+2,n}\cdots T_{2,j}, \qquad \text{and so on},\\
Y_{j-1,j} = T_{j-2,j-1}^{-1} \cdots T_{1,2}^{-1} \pi T_{n-1,n} \cdots T_{j-1,j}.
\end{gather*}
\begin{Proposition}\label{proposition4.13}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] $x_j x_j T_{ij}=T_{ij} x_i x_j$,
\item[$(2)$] $Y_{i,j}= T_{i,j} Y_{i+1,j+1} T_{i,j}$ if $1 \le i < j < n$,
\item[$(3)$] $Y_{i,j} Y_{i+k,j+k}=Y_{i+k,j+k} Y_{i,j}$ if $1 \le i < j \le n-k$,
\item[$(4)$] one has $x_{i-1}Y_{i,j}^{-1}=Y_{i,j}^{-1}x_{i-1}T_{i-1,j-1}^{2}$, $2 \le i < j \le n$,
\item[$(5)$] $Y_{i,j} x_1 x_2 \cdots x_n = t x_1 x_2 \cdots x_n Y_{i,j}$,
\item[$(6)$] $x_i Y_1 Y_2 \cdots Y_n = t^{-1} Y_1 Y_2 \cdots Y_n x_i$,
\end{enumerate}
where we set $Y_i:= Y_{i,i+1}$, $ 1 \le i < j < n$.
\end{Proposition}
\begin{Conjecture} \label{conjecture4.10}
Subalgebra of $3{\mathfrak {T}}_n$ generated by the
elements $\{T_i:= T_{i,i+1}, \, 1 \le i < n, \, Y_{1}, \ldots, Y_n$,
and $x_1,\ldots,x_n \}$, is isomorphic to the double affine Hecke algebra
${\rm DAHA}_{q,t}(n)$.
\end{Conjecture}
Note that the algebra $3{\mathfrak{T}}_n$ contains also two additional commutative subalgebras generated by {\it additive} $ \big\{ \theta_i=\sum\limits_{j \not= i} u_{ij} \big\}_{1 \le i \le n}$ and {\it multiplicative}
\begin{gather*}
\left\{ \Theta_i = \prod_{a=1}^{i-1} (1-u_{ai}) \prod_{a=i+1}^{n} (1+u_{ia})
\right\}_{1 \le i \le n}
\end{gather*}
Dunkl elements correspondingly.
Finally we introduce (cf.~\cite{Ch}) a (projective) representation of the
modular group $SL(2,\mathbb{Z})$ on the {\it extended affine Hecke algebra}
${\widehat{ \cal{H}}_n}$
over the ring $\mathbb{Z}[q^{\pm 1},t^{\pm 1}]$ generated by elements
\begin{gather*}
\{ T_{1},\ldots,T_{n-1} \}, \quad \pi , \quad \text{and}\quad \{ x_1,\ldots,x_n \}.
\end{gather*}
It is well-known that the group ${\rm SL}(2,\mathbb{Z})$ can be generated by two matrices
\begin{gather*} \tau_{+}= \left(
\begin{matrix}
1&1\\
0&1
\end{matrix}
\right),\qquad
\tau_{-}= \left(
\begin{matrix}
1&0\\
1&1
\end{matrix}
\right),
\end{gather*}
which satisfy the following relations
\begin{gather*}
\tau_{+} \tau_{-}^{-1} \tau_{+}= \tau_{-}^{-1} \tau_{+} \tau_{-}^{-1},\qquad
\big(\tau_{+} \tau_{-}^{-1} \tau_{+}\big)^6 = I_{2 \times 2}.
\end{gather*}
Let us introduce operators~$\tau_{+}$ and~$\tau_{-}$ acting on the extended
af\/f\/ine algebra ${\widehat{ \cal{H}}_n}$. Namely,
\begin{gather*}
\tau_{+}(\pi)= x_{1} \pi, \tau_{+}(T_{i}) = T_{i}, \tau_{+}(x_i) = x_{i},\qquad
\forall\, i,\\
\tau_{-}(\pi)=\pi,\qquad \tau_{-}(T_{i})=T_{i}, \qquad \tau_{-}(x_i)=
\left( \prod_{a=i-1}^{1} T_{a} \right) \pi \left( \prod_{a=n}^{i} T_{a} \right) x_{i}.
\end{gather*}
\begin{Lemma}\label{lemma4.8}\quad
\begin{gather*}
\tau_{+}(Y_i)= \left(\prod_{a=i-1}^{1} T_{a}^{-1} \right) \left(
\prod_{a=1}^{i-1}T_{a}^{-1} \right) x_{i} Y_{i},\\
\tau_{-}(x_i)= \left(\prod_{a=i-1}^{1} T_{a} \right) \left(
\prod_{a=1}^{i-1}T_{a} \right) Y_{i} x_i, \\
\big(\tau_{+} \tau_{-}^{-1} \tau_{+}\big)(x_i)=Y_i^{-1}= \big(\tau_{-}^{-1} \tau_{+} \tau_{-}^{-1}\big)(x_i),\\
\big(\tau_{+} \tau_{-}^{-1} \tau_{+}\big)(Y_i) = t x_i \left
(\prod_{a=i-1}^{1} T_{a} \right) \left( T_{1} \cdots T_{n-1}\right) \left( \prod_{a= n-1}^{i} T_{a} \right), \qquad i=1,\ldots,n.
\end{gather*}
\end{Lemma}
In the last formula we set $T_{n}=1$ for convenience.
\subsection{Braid, af\/f\/ine braid and virtual braid groups}\label{section4.7}
The main objective of this section is to describe the distinguish abelian subgroup
in the braid group $B_n$, see Proposition~\ref{proposition4.130}$\big(2^{(0)}\big)$, and that in the
Yang--Baxter groups ${\widehat{{\rm YB}}}_n$ and ${\rm YB}_n$, see Proposition~\ref{proposition4.130}$\big(5^{(0)}\big)$ and $\big(6^{(0)}\big)$ correspondingly. As far as I'm aware, these constructions go back to E.~Artin in the case of braid groups, and to C.N.~Yang in the
case of Yang--Baxter group, and nowadays are widely use in the representation
theory of Hecke's type algebras and that of integrable systems. In a~few words,
by choosing a suitable representation (f\/inite-dimensional or birational) of
either $B_n$ or ${\rm YB}_n$, or ${\widehat{{\rm YB}}}_n$, one gives rise to a family of
mutually commuting operators acting in the space of a~representation selected. In the
case of braid groups one comes to Jucys--Murphy's type operators/elements, and
in the case of Yang--Baxter groups one comes to Dunkl's type operators/elements. See, e.g., \cite{IK,IKT}, where it was used the so-called $R$-matrix
representation of the af\/f\/ine braid group of type $C_n^{(1)}$ to construct the
(two boundary) quantum Knizhnik--Zamolodchikov connections with values in the
af\/f\/ine Birman--Murakami--Wenzl algebras.
To start with, let $n \ge 2$ be an integer.
\begin{Definition}\label{definition4.40} \quad
\begin{itemize}\itemsep=0pt
\item Denote by $\mathbb{S}_{n}$ {\it the symmetric group}
on $n$ letters, and by $s_{i}$ the simple transposition
$(i,i+1)$ for $1 \le i \le n-1$.
The well-known Moore--Coxeter presentation of the symmetric group has the form
\begin{gather*}
\langle s_{1},\ldots,s_{n-1} \,|\,s_{i}^2=1,\, s_{i}s_{i+1}s_{i}=s_{i+1}s_{i}s_{i+1}
,\, s_{i}s_{j}=s_{j}s_{i}\, {\rm if}\, |i-j| \ge 2 \rangle.
\end{gather*}
Transpositions $s_{ij}:=s_i s_{i+1} \cdots s_{j-2} s_{j-1}
s_{j-2} \cdots s_{i+1} s_{i}$, $1 \le i < j < j \le n$, satisfy the following
set of (def\/ining) relations
\begin{gather*}
s_{ij}^{2}=1, \qquad s_{ij} s_{kl}=s_{kl} s_{ij} \qquad {\rm if} \quad \{i,j\} \cap \{k,l \}= \varnothing,\\
s_{ij} s_{ik}=s_{jk} s_{ij}=s_{ik} s_{jk}, \qquad s_{ik} s_{ij}=s_{ij} s_{jk}= s_{jk} s_{ik}, \qquad i < j < k.
\end{gather*}
\item {\it The Artin braid group on $n$ strands $B_{n}$} is
def\/ined by generators $\sigma_{1},\ldots,\sigma_{n-1}$ and relations
\begin{gather}\label{equation4.7new}
\sigma_{i}\sigma_{i+1}\sigma_{i}=\sigma_{i+1}\sigma_{i}\sigma_{i+1}, \qquad
1 \le i \le n-2, \qquad \sigma_{i}\sigma_{j}=\sigma_{j}\sigma_{i} \qquad {\rm if} \quad |i-j|
\ge 2.
\end{gather}
\item {\it The monoid of positive braids on $n$ strands
$B_{n}^{+}$}
is a monoid generated by the elements $\sigma_{1},\ldots,\sigma_{n-1}$ subject
to the set of relations~\eqref{equation4.7new}.
\item {\it A new representation of the braid group} \cite{BKL}.
The Birman--Ko--Lee representation of the braid group~$B_n$ has the set of
generators $\{ \sigma_{i,j} \,|\, 1 \le i < j \le n \}$ subject to
the Birman--Ko--Lee (def\/ining) relations
\begin{gather*}
\sigma_{i,j} \sigma_{k,l}=\sigma_{k,l} \sigma_{i,j} \qquad {\rm if}\quad
(j-l)(j-k)(i-l)(i-k) > 0,\\
\sigma_{i,j} \sigma_{i,k}=\sigma_{j,k} \sigma_{i,j}=
\sigma_{i,k} \sigma_{j,k} \qquad {\rm if} \quad 1 \le i < j < k \le n.
\end{gather*}
One can take $\sigma_{i,j}:=(\sigma_{j-1} \cdots \sigma_{i+1})\sigma_{i}
(\sigma_{i+1}^{-1} \cdots \sigma_{j-1}^{-1})$, see~\cite{BB} for details.
It would be well to note that as a~corollary of the Birman--Ko--Lee relations
one can deduce the {\it $2D$ Coxeter relations} among the Birman--Ko--Lee generators
\begin{gather*}
\sigma_{i,j} \sigma_{j,k} \sigma_{i,j}=\sigma_{j,k} \sigma_{i,j} \sigma_{j,k}, \qquad 1 \le i < j <k \le n.
\end{gather*}
\item {\it The Birman--Ko--Lee monoid ${\rm BKL}_n$} is a monoid generated by
the elements $\sigma_{i,j}$, $1 \le i < j \le n$, subject to the Birman--Ko--Lee
relations.
We denote by ${\rm BKL}(n)$ and called it as
{\it the Birman--Ko--Lee algebra}, the group algebra ${\mathbb{Q}}[{\rm BKL}_n]$ of the Birman--Ko--Lee monoid.
The Hilbert series of the Birman--Ko--Lee algebra ${\rm BKL}(n)$ will
be computed in Section~\ref{section4.7.3}, Theorem~\ref{theorem4.132}.
\item {\it The pure braid group} ${\rm PB}_{n}$ is def\/ined to be
the kernel of the natural (non-split) projection $p\colon B_{n} \longrightarrow \mathbb{S}_{n}$
given by $p(\sigma_{i})=s_{i}$.
It is well-known that the pure braid group ${\rm PB}_{n}$
is generated by the elements
\begin{gather*}
g_{i,j}:= \sigma_{i,j}^2=\sigma_{j-1}\sigma_{j-2}\cdots\sigma_{i+1}
\sigma_{i}^2\sigma_{i+1}^{-1}\cdots\sigma_{j-2}^{-1}\sigma_{j-1}^{-1}
\qquad {\rm for} \quad 1 \le i < j \le n,
\end{gather*}
subject to the following def\/ining relations
\begin{gather*}
g_{i,j} g_{k,l}=g_{k,l} g_{i,j} \qquad {\rm if} \quad (i-k) (i-l) (j-k) (j-l) > 0,\\
g_{i,j} g_{i,k} g_{j,k}=g_{i,k} g_{j,k} g_{i,j}=g_{j,k} g_{i,j} g_{i,k} \qquad {\rm if}\quad 1 \le i < j < k \le n,\\
g_{i,k} g_{i,l} g_{j,l} g_{k,l}=g_{i,l} g_{j,l} g_{k,l} g_{i,k} \qquad {\rm if} \quad 1 \le i < j <k < l \le n.
\end{gather*}
\end{itemize}
\begin{Comments} It is easy to see that the def\/ining relations for the pure braid group ${\rm PB}_n$ listed above are equivalent to the following list of def\/ining relations
\begin{gather*}
g_{i,j}^{-1}g_{k,l}g_{i,j}= \begin{cases}
g_{k,l} & \text{if} \ (i-k)(i-l)(j-k)(j-l) > 0, \\
g_{i,l} g_{k,l} g_{i,l}^{-1} & \text{if} \ i < k=j < l , \\
g_{i,l} g_{j,l} g_{k,l} g_{i,l}^{-1} g_{j,l}^{-1} & \text{if} \ i=k < j < l, \\
g_{i,l} g_{j,l} g_{i,l}^{-1} g_{j,l}^{-1} g_{k,l} g_{j,l} g_{i,l} g_{j,l}^{-1} g_{i,l}^{-1} & \text{if} \ i < k < j < l,
\end{cases}
\end{gather*}
commonly used in the literature to describe the def\/ining relations among the generators $\{g_{ij} \}$ of the pure braid group $P_n$, see, e.g., \cite{BB}.
\end{Comments}
\begin{itemize}\itemsep=0pt
\item {\it The affine Artin braid group $B_n^{\rm af\/f}$}, cf.~\cite{OR}, is an extension of the Artin braid group on~$n$ strands $B_n$ by the
element~$\tau$ subject to the set of crossing relations
\begin{gather*}
\sigma_1 \tau \sigma_1 \tau=\tau \sigma_1 \tau \sigma_1,
\qquad \sigma_i \tau=\tau \sigma_i \qquad {\rm for} \quad 2 \le i \le n-1.
\end{gather*}
\item {\it The virtual braid group ${\rm VB}_n$} is a group generated by
$\sigma_1,\ldots,\sigma_{n-1}$ and $s_1,\ldots,s_{n-1}$ subject to the relations:
\begin{enumerate}\itemsep=0pt
\item[$(1)$] {\it braid relations} $\sigma_1,\ldots,\sigma_{n-1}$ generate the Artin
braid group $B_n$;
\item[$(2)$] {\it Moore--Coxeter relations} $s_1,\ldots,s_{n-1}$ generate the symmetric
group ${\mathbb S}_n$;
\item[$(3)$] {\it crossing relations} $\sigma_i s_j=s_j \sigma_i$ if $|i-j| \ge 2$,
$s_is_{i+1}\sigma_i=\sigma_{i+1}s_is_{i+1}$ if $1 \le i \le n-2$.
\end{enumerate}
\item {\it The virtual pure braid group ${\rm VP}_n$} is def\/ined to
be the kernel of the natural map
\begin{gather*}
\eta\colon \ {\rm VB}_n \longrightarrow {\mathbb S}_n, \qquad \eta(\sigma_i)=\eta(s_i)=s_i,\qquad
i= 1, \ldots,n-1.
\end{gather*}
\end{itemize}
\end{Definition}
\subsubsection{Yang--Baxter groups}\label{section4.7.1}
\begin{Definition}\quad
\begin{itemize}\itemsep=0pt
\item {\it The quasitriangular Yang--Baxter group $\widehat {{\rm YB}_{n}}$},
cf.~\cite{BEE}, is a group generated by the set of elements
$\{ Q_{i,j},\, 1 \le i \not= j \le n \}$, subject to the set of def\/ining
relations
\begin{enumerate}\itemsep=0pt
\item[$(1)$] $[Q_{i,j},Q_{k,l}]=0$ if $i$, $j$, $k$ and $l$
are all distinct,
\item[$(2)$] {\it Yang--Baxter relations} $Q_{i,j}Q_{i,k}Q_{j,k}=Q_{j,k}Q_{i,k}Q_{i,j}$ if $i$, $j$, $k$ are distinct.
\end{enumerate}
According to \cite[Theorem~1]{Bar}, the
{\it quasitriangular Yang--Baxter group $\widehat {{\rm YB}_{n}}$} is isomorphic
to the {\it virtual pure braid group} ${\rm VP}_n$.
\item {\it The Yang--Baxter monoid $\widetilde {{\rm YB}_{n}}$} is a~monoid
generated by the elements $Q_{i,j}$, $1 \le i \not= j \le n$.
Important particular case corresponds to the case when $Q_{i,j}Q_{j,i}=1$
for all $1 \le i \not= j \le n$.
\item {\it The Yang--Baxter group ${\rm YB}_{n}$} is def\/ined by the set of
generators
$R_{i,j}$, $1 \le i < j \le n$, subject to the set of def\/ining relations
\begin{enumerate}\itemsep=0pt
\item[$(1)$] $R_{i,j}R_{k,l}=R_{k,l}R_{i,j}$ if $i$, $j$, $k$ and $l$ are pairwise
distinct,
\item[$(2)$] $R_{i,j}R_{i,k}R_{j,k}=R_{j,k}R_{i,k}R_{i,j}$ if
$1 \le i < j <k \le n$.
\end{enumerate}
\end{itemize}
\end{Definition}
\subsubsection{Some properties of braid and Yang--Baxter groups}\label{section4.7.2}
For the sake of convenience and future references, below we state some basic
properties of the groups $P_{n}$, ${\rm YB}_{n}$ and $B_n^{\rm af\/f}$.
\begin{Proposition}\label{proposition4.130}
Let $F_m$ denotes the free group with~$m$ generators.
\begin{enumerate}\itemsep=0pt
\item[$\big(1^{0}\big)$] The elements $g_{1,n},g_{2,n},\ldots,g_{n-1,n}$ generate
a~free normal subgroup $F_{n-1}$ in $P_{n},$ and $P_{n}=P_{n-1} \ltimes \langle
g_{1,n},g_{2,n},\dots,g_{n-1,n} \rangle$. Hence $P_{n}$ is an iterated
extension of free groups.
\item[$\big(2^{0}\big)$] Let us consider the following elements in the group
$B_n^{\rm af\/f}$:
\begin{gather*} \gamma_1=\tau,\qquad \gamma_i=
\prod_{j=i-1}^{1}\sigma_j \tau \prod_{j=1}^{i-1}\sigma_j,\qquad 2 \le i \le n.
\end{gather*}
Then
\begin{enumerate}\itemsep=0pt
\item[$(a)$] commutativity, $\gamma_i \gamma_j=\gamma_j \gamma_i$ for all $1 \le i,j \le n$;
\item[$(b)$] the elements $\gamma_1,\dots,\gamma_n$
generate a free abelian subgroup of rang $n$ in $B_n^{\rm af\/f}$.\footnote{We refer the reader to~\cite{OR} for more details about af\/f\/ine braid
groups.
Here we only remark that the type $A$ af\/f\/ine Weyl groups
$\widehat{\mathbb {S}}_n$,
the Hecke algebras~$H_{n,q}$, the af\/f\/ine Hecke
algebras $\widehat {H}_{n,q}$, the Ariki--Koike, or cyclotomic Hecke,
algebras $H_{r,1,n}$, the af\/f\/ine and cyclotomic Birman--Murakami--Wenzl
algebras ${\cal Z}_{r,1,n}$, all can be obtained as certain quotients of the
group algebra $\mathbb{C} B_n^{\rm af\/f}$ of the af\/f\/ine braid group.}
\end{enumerate}
\item[$\big(3^{0}\big)$] Let us introduce elements
\begin{gather*}
D_{i,j}:=\sigma_{j-1}\sigma_{j-2}\cdots\sigma_{i+1}~
\sigma_{i}^2~\sigma_{i+1}\cdots\sigma_{j-2}\sigma_{j-1}=
\prod_{i \le a < j}g_{a,j} \in P_{n},\\
F_{i,j}:=
\sigma_{n-j} \sigma_{n-j+1} \cdots \sigma_{n-i-1} \sigma_{n-i}^{2}
\sigma_{n-i-1} \cdots \sigma_{n-j+1} \sigma_{n-j}=
\prod_{a=j}^{i+1} g_{n-a,n-i} \in P_n,
\end{gather*}
where $1 \le i < j \le n$. For example,
\begin{gather*}
\begin{split}
& D_{i,i+1}=\sigma_{i}^2,\qquad D_{i,i+2}=
\sigma_{i+1}\sigma_{i}^2\sigma_{i+1},\qquad F_{i,i+1}=\sigma_{n-i}^{2},\\
& F_{i,i+2}=\sigma_{n-i-1} \sigma_{n-i}^2 \sigma_{n-i-1}
\end{split}
\end{gather*}
and etc.
Then
\begin{itemize}\itemsep=0pt
\item For each $j=3, \ldots,n,$ the element $D_{1,j}$ commutes with
$\sigma_{1},\ldots,\sigma_{j-2}$.
\item The elements $D_{1,2},D_{1,3},\ldots,D_{1,n}$ $($resp.\
$F_{1,2}, F_{1,3},\ldots,F_{1,n})$ generate a free
abelian subgroup in~$P_{n}$.
\item If $n \ge 3$, the element
\begin{gather*}
\prod_{2 \le j \le n}D_{1,j} = \prod_{2 \le j \le n} F_{1,j}=
(\sigma_{1}\cdots\sigma_{n-1})^n
\end{gather*}
generates the center of the braid group~$B_{n}$ and that of the pure braid
group $P_{n}$.
\item $D_{i,j}D_{i,j+1}D_{j,j+1}=D_{j,j+1}D_{i,j+1}D_{i,j}$
if $i < j$.
\item Consider the elements $s:=\sigma_1 \sigma_2 \sigma_1$,
$t:= \sigma_1 \sigma_2$ in the braid group $B_3$. Then $s^2=t^3$ and the
element $c:=s^2$ generates the center of the group~$B_3$. Moreover,
\begin{gather*}
B_3 / \langle c \rangle \cong {\rm PSL}_2(\mathbb{Z}), \qquad
B_3 / \langle c^2 \rangle \cong {\rm SL}_2(\mathbb{Z}).
\end{gather*}
\end{itemize}
\item[$\big(4^{0}\big)$] Let us introduce the following elements in the
quasitriangular Yang--Baxter group $\widehat {{\rm YB}_n}$:
\begin{gather*}
C_{i,j}= \left( \prod_{a=j-1}^{i} Q_{a,j} \right) \left( \prod_{b=i}^{j-1} Q_{j,b} \right),\qquad
f_{i,j}=\left(\prod_{a=j-1}^{i+1}Q_{a,j} \right) Q_{i,j}Q_{j,i} \left( \prod_ {b=i+1}^{j-1} Q_{b,j} \right)^{-1}.
\end{gather*}
Then
\begin{itemize}\itemsep=0pt
\item The elements $C_{1,2},C_{1,3},\ldots,C_{1,n}$ generate a free
abelian subgroup in~$\widehat {{\rm YB}_n}$.
\item The elements $ f_{i,j}$, $1 \le i < j \le n$, generate a subgroup in $\widehat {{\rm YB}_n} $, which
is isomorphic to the pure braid group~$P_n$.\footnote{It is enough to check that the elements $\{f_{i,j},\, 1 \le i < j \le n \}$ satisfy the def\/ining relations for the pure braid group $P_n$ only in the case $n=4$.
Let us prove that
\begin{gather*}
f_{1,4} f_{2,4} f_{3,4} f_{1,3}=f_{1,3} f_{1,4} f_{2,4} f_{3,4}.
\end{gather*}
Other relations are simple and can be checked in a similar fashion.
Let
\begin{gather*}
{\rm l.h.s.} =f_{1,4} f_{2,4} f_{3,4} f_{1,3}=Q_{34}Q_{24}Q_{14}Q_{41}Q_{42}Q_{43}
Q_{23}Q_{13}Q_{31}Q_{23}^{-1},\\
{\rm r.h.s.} =f_{1,3} f_{1,4} f_{2,4} f_{3,4}=Q_{23}Q_{13}Q_{31}{\boldsymbol{Q_{23}^{-1}Q_{34}Q_{24}}}Q_{14}Q_{41}Q_{42}Q_{43}.
\end{gather*}
Now we are going to apply the Yang--Baxter relations
\begin{gather*}
Q_{23}^{-1}Q_{34}Q_{24}=Q_{24}Q_{34}Q_{23}^{-1},\qquad Q_{23}^{-1}Q_{42}Q_{43}=Q_{43}Q_{42}Q_{23}^{-1},\qquad
Q_{31}Q_{34}Q_{14}=Q_{14}Q_{34}Q_{31}.
\end{gather*}
Therefore,
\begin{gather*}
{\rm r.h.s.} = Q_{23}Q_{13}Q_{31}{\boldsymbol{Q_{23}^{-1}Q_{34}Q_{24}}}Q_{14}Q_{41}Q_{42}Q_{23}^{-1}={\boldsymbol{Q_{23}Q_{24}Q_{34}}}Q_{14}Q_{13}
{\boldsymbol{Q_{31}Q_{41}Q_{43}}}Q_{42}Q_{23}^{-1}\\
\hphantom{{\rm r.h.s.}}{}
=Q_{34}Q_{24}Q_{14}Q_{23}{\boldsymbol{Q_{13}Q_{43}Q_{41}}}Q_{42}Q_{31}Q_{23}^{-1}=Q_{34}Q_{24}Q_{14}Q_{41}{Q_{23}Q_{43}Q_{42}}Q_{13}Q_{31}Q_{23}^{-1}\\
\hphantom{{\rm r.h.s.}}{}
=Q_{34}Q_{24}Q_{14}Q_{41}Q_{42}Q_{43}Q_{23}Q_{13}Q_{31}Q_{23}^{-1}={\rm l.h.s.}
\end{gather*}
}
\end{itemize}
\item[$\big(5^{0}\big)$] Assume that the following additional relations in $\widehat {{\rm YB}_n}$ are satisfied
\begin{gather*}
Q_{i,j}Q_{j,i}=Q_{j,i}Q_{i,j},\qquad Q_{k,l}Q_{i,j}Q_{j,i}=
Q_{j,i}Q_{i,j}Q_{k,l}
\end{gather*}
if $i \not= j$ and $k \not= l$. In other words, the elements $Q_{i,j}$ and
$Q_{j,i}$ commute, and the elements $Q_{i,j}Q_{j,i}=Q_{j,i} Q_{i,j}$ are central.
Under these assumptions, we have
that the elements
\begin{gather*}
\Theta_i:=\prod_{j=i-1}^{1}Q_{j,i}
\prod_{j=n}^{i+1}Q_{i,j}, \qquad \bar \Theta_{i}:= \prod_{j=i+1}^{n}Q_{j,i}
\prod_{j=1}^{i-1}Q_{i,j}, \qquad 1 \le i \le n,
\end{gather*}
satisfy the following relations
\begin{gather*}
[\Theta_{i},\Theta_{j}]=0=[\bar\Theta_{i},\bar\Theta_{j}],\\
\Theta_{i} \bar\Theta_{i}=\prod_{j \not= i}Q_{i,j}Q_{j,i}=
\bar\Theta_{i} \Theta_{i},\qquad \prod_{i=1}^{n}\Theta_{i}=
\prod_{1 \le i \not= j \le n}Q_{i,j}Q_{j,i}= \prod_{i=1}^{n}\bar\Theta_{i}.
\end{gather*}
\item[$\big(6^{0}\big)$]
In the special case $Q_{i,j}Q_{j,i}=1$ for all $i \not= j$, the following statement holds:
the elements
\begin{gather*}
\Theta_{j}=\prod_{a=j-1}^{1}R_{a,j}^{-1}
\prod_{b=n}^{j+1}R_{j,b}, \qquad 1 \le j \le n-1,
\end{gather*}
generate a subgroup
in the Yang--Baxter group ${\rm YB}_{n}$, which is isomorphic to the free abelian group of rang~$n-1$.
\end{enumerate}
\end{Proposition}
\subsubsection{Artin and Birman--Ko--Lee monoids}\label{section4.7.3}
Let $(W,S)$ be a f\/inite Coxeter group, $B(W)$ and $B^{+}(W)$ be
the corresponding braid group and monoid of positive braids.
Denote by $P_{W}(s,t)=
\sum\limits_{i \ge 0,\,j \ge 0}B_{{\mathbb{Q}}[B^{+}(W)]}(i,j)s^it^j$ the Poincar\'e
polynomial of the group algebra over $\mathbb{Q}$ of the monoid $B^{+}(W)$.
\begin{Conjecture} $P_{W}(s,1)=(s+1)^{|S|}$.
\end{Conjecture}
It is known \cite{De,Sa}
that the Hilbert series of the
group algebra of the monoid $B^{+}(W)$ is a~{\it rational} function of the form
${1 \over P(t)}$ for a some polynomial $P(t):=P_{W}(t) \in \mathbb{Z}[t]$.
\begin{Theorem} \label{theorem4.132}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] Some Betti numbers of the group algebra over $\mathbb{Q}$ of the
monoid
$B^{+}(A_{n-1})$:
\begin{gather*}
B_{{\mathbb{Q}}[B^{+}(A_{n-1})]}(k,k)={n-k \choose k}, \\
B_{{\mathbb{Q}}[B^{+}(A_{n-1})]}\left(k,{k+1 \choose 2}\right)=n-k, \qquad 1 \le k \le n-1,\\
B_{{\mathbb{Q}}[B^{+}(A_{n-1})]}(k,k+1)= (k-1) {n-k \choose k-1},\\
B_{{\mathbb{Q}}[B^{+}(A_{n-1})]}(k,k+2)= {k-2 \choose 2} {n-k \choose k-2},\\
B_{{\mathbb{Q}}[B^{+}(A_{n-1})]}(k,k+3)= (k-2) {n-k \choose k-2}+
\max(3 k-17,0) {n-k \choose k-3} \quad \text{if} \ k \ge 3.
\end{gather*}
\item[$(2)$] The Birman--Ko--Lee algebra ${\rm BKL}(n)$ is Koszul, and the Hilbert
polynomial of its quadratic dual is equal to
\begin{gather*}
{\rm Hilb}\big({\rm BKL}(n)^{!},t\big)= \sum_{k=0}^{n-1}{1 \over {k+1}} {n-1 \choose k}
{n+k-1 \choose k} t^k.
\end{gather*}
\end{enumerate}
\end{Theorem}
\begin{Conjecture}[type $A_{n-1}$ case] Let $I \subset [1,n-1]$ be a subset of
vertices in the Dynkin diagram of type $A_{n-1}$, and $R_{I}$ denotes the
root system generated by the positive roots
$\{\alpha_{ij}=\epsilon_i-\epsilon_j, \, (i,j) \in I \times I \}$. Assume that
\begin{gather*}
R_I \cong A_{n_{1}} \coprod \cdots \coprod A_{n_{k}},\qquad n_1+\cdots+n_{k}=n-1
\end{gather*}
stands for the decomposition of the root system~$R_{I}$ into the disjoint
union of irreducible root subsystems of type~$A$. The numbers $n_1, \ldots,
n_k$ are defined uniquely up to a permutation. Let us set $n(I)=
\sum\limits_{a=1}^{k} {n_a \choose 2}$. Then
\begin{gather*}
P_{A_{n-1}}(s,t)=\sum_{I} s^{|I|} t^{n(I)},
\end{gather*}
where the sum runs over the all subsets of vertices~$I$ in the Dynkin diagram
of type~$A_{n-1}$, including the empty set, and~${|I|}$ denotes the
cardinality of the set~$I$.
\end{Conjecture}
\begin{Comments} \quad
\begin{enumerate}\itemsep=0pt
\item[$(A)$] The Hilbert polynomial of the Birman--Ko--Li algebra ${\rm BKL}(n)$ has been computed also by M.~Albenque and P.~Nadeau, see~\cite{AN}.
\item[$(B)$] Let's consider the {\it truncated theta function}
$\theta^{+}(z,t)=
\sum\limits_{n \ge 0}t^{n(n+1)/2} z^n$. Then
\begin{gather*}
\sum_{n \ge 1} P_{A_{n-1}}(s,t) z^{n-1}=
\theta^{+}(t,s z)/(1-z(\theta^{+}(t,s z))).
\end{gather*}
\item[$(C)$] It is well known that the number
\begin{gather*}
T(n,k)= {1 \over {k+1}} {n \choose k}~{n+k \choose k}
\end{gather*}
counts the number of Schr\"oder paths (i.e., consisting of
steps $(1,1)$, $(1,-1)$ and $(2,0)$ and never going below $x$-axis) from
$(0,0)$ to $(2n,0)$, having exactly~$k$ $(1,1)$ steps. In particular,
\begin{gather*}
\dim({\rm BKL}(n)^{!})= {\rm Sch}(n),
\end{gather*}
is the $n$-th (large) Schr\"oder number, see \cite[$A006318$]{SL}.
It is a classical result that
\begin{gather*}
\sum_{k=0}^{n} T(n,k) x^k (1-x)^{n-k}=\sum_{k=0}^{n-1} N(n,k) x^k,
\end{gather*}
where $N(n,k):={1 \over n} {n \choose k} {n \choose k+1}$ denotes the
Narayana number.
Some explicit combinatorial interpretations of the values of
the above polynomials for $x=0,1,2,4$ can be found in~\cite[$A088617$]{SL}.
Note that ${\rm Hilb}({\rm BKL}^{!},t)= (1+t) {\rm Ass}_{n-2}(t)$, where
\begin{gather*}
{\rm Ass}_{n}(t):= \sum_{k=0}^{n} {1 \over k+1} {n \choose k} {n+k \choose k}
t^{k}
\end{gather*}
denotes the $f$-vector polynomial corresponding to the associahedron of type~$A_n$.
\item[$(D)$] The polynomials
$F(n,t):=\sum\limits_{k \ge 0} B_{{\mathbb{Q}}[B^{+}(A_{n-1})]}(k,k) t^{k}$ appear to be
equal to the so-called {\it Fibonacci} polynomials, see, e.g., \cite[$A011973$]{SL}. It is well-known that
\begin{gather*}
\sum_{n \ge 0} F(n,t) y^{n}={1+t y \over 1-y-t y^2 }.
\end{gather*}
Moreover, the coef\/f\/icient $B_{{\mathbb{Q}}[B^{+}(A_{n-1})]}(k,k)$ is equal to the
number of compositions of $n+2$ into $k+1$ parts, all $ \ge 2$, see~\cite[$A011973$]{SL}.
\item[$(E)$] {\it Monoid of positive pure braids.}
The monoid of positive pure braids ${\rm PB}_n^{+}$ (of the type $A_{n-1}$) is the
monoid generated by the set $\{g_{i,j},\, 1 \le i < j \le n \}$ of the Artin
generators of the pure braid group ${\rm PB}_n$.
\end{enumerate}
\end{Comments}
\begin{Conjecture} The following list of relations is the defining set of relations in
the monoid ${\rm PB}_n$:
\begin{gather*}
(a)\quad [g_{i,j},g_{k,l}]=0, \qquad [g_{i,l},g_{j,k}]=0, \qquad \text{if} \quad 1 \le i < j < k < l \le n,\\
(b)\quad \left[g_{\overline {j_{1}+m}, \overline{j_{k-1}+m}}, \prod_{a=1}^{k-1}g_{\overline {j_{a}+m}, \overline{j_{k}+m}} \right]=0,
\end{gather*}
for all sequences of integers $1 \le j_1 < j_2 < \cdots < j_k \le n$ of the
length $k \ge 4$ and $m= 0, \ldots, n-1$.
Here we assume that $g_{i,j}=g_{j,i}$
for all $i \not= j$, and for any non-negative integer~$a$ we denote by~$\overline {a}$ a unique integer $1 \le \overline{a} \le n$ such that
$ a \equiv \overline{a}$ $({\rm mod} \, n+1)$.
\end{Conjecture}
It is worth noting that the def\/ining relations in the pure braid group $P_n$
are that listed in~$(a)$ and the part of that listed in~$(b)$
corresponding to $k=3$, $m=0$ and~$1$, and that for $k=4$, $m=0$.
\section{Combinatorics of associative Yang--Baxter algebras}\label{section5}
Let $\alpha$~and~$\beta$ be parameters.
\begin{Definition}[\cite{Kir,K2,K}, cf.\ \cite{Ag,Pol}]\label{definition5.1}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] The associative quasi-classical Yang--Baxter algebra
of weight $(\alpha,\beta)$, denoted by\linebreak $\widehat{{\rm ACYB}}_n(\alpha,\beta)$, is
an associative algebra, over the ring of polynomials $\mathbb{Z}[\alpha,\beta]$,
generated by the set of elements $\{ x_{ij}, \, 1 \le i < j \le n\}$, subject to
the set of relations
\begin{enumerate}\itemsep=0pt
\item[(a)] $x_{ij} x_{kl}=x_{kl} x_{ij}$ if $ \{i,j\} \cap \{k,l \}=\varnothing$,
\item[(b)] $x_{ij} x_{jk}=x_{ik} x_{ij}+x_{jk} x_{ik}+\beta x_{ik} + \alpha$ if
$1 \le 1 <i < j \le n$.
\end{enumerate}
\item[(2)] Def\/ine {\it associative quasi-classical Yang--Baxter algebra of weight
$\beta$}, denoted by $\widehat{{\rm ACYB}}_n(\beta)$, to be
$\widehat{{\rm ACYB}}_n(0,\beta)$.
\end{enumerate}
\end{Definition}
\begin{Comments}\label{comments5.1}
The algebra $3T_n(\beta)$, see Def\/inition~\ref{definition3.1}, is the
quotient of the algebra \linebreak $\widehat{{\rm ACYB}}_n(- \beta)$, by the ``dual relations''
\begin{gather*}
x_{jk}x_{ij}-x_{ij} x_{ik}-x_{ik} x_{jk}+\beta x_{ik}=0, \qquad i < j < k.
\end{gather*}
The (truncated) Dunkl elements $\theta_i=\sum\limits_{j \not= i} x_{ij}$, $i=1,\ldots,n$, do not commute in the algebra $\widehat{{\rm ACYB}}_n(\beta)$. However a~certain version of noncommutative elementary polynomial of degree $k \ge 1$,
still is equal to zero after the substitution of Dunkl elements instead of
variables~\cite{K}. We state here the corresponding result only ``in classical case'', i.e., if
$\beta=0$ and $q_{ij}= 0$ for all~$i$,~$j$.
\end{Comments}
\begin{Lemma}[\cite{K}]\label{lemma5.1} Define noncommutative elementary polynomial
$L_k(x_1,\ldots,x_n)$ as follows
\begin{gather*}
L_k(x_1,\ldots,x_n)= \sum_{I=(i_1 < i_2 < \cdots < i_k) \subset [1,n]}
x_{i_{1}} x_{i_{2}} \cdots x_{i_{k}}.
\end{gather*}
Then $L_{k}(\theta_1,\theta_2,\ldots,\theta_n)=0$.
Moreover, if $1 \le k \le m \le n$, then one can show that the value of
the noncommutative polynomial $L_k(\theta_1^{(n)},\ldots,\theta_m^{(n)})$ in
the algebra $\widehat{{\rm ACYB}}_n(\beta)$ is given by the Pieri formula,
see~{\rm \cite{FK, P}}.
\end{Lemma}
\subsection{Combinatorics of Coxeter element}\label{section5.1}
Consider the ``Coxeter element'' $w \in \widehat{{\rm ACYB}}_n(\alpha,\beta)$ which
is equal to the ordered product of ``simple generators'':
\begin{gather*}
w:=w_n=\prod_{a=1}^{n-1} x_{a,a+1}.
\end{gather*}
Let us bring the element~$w$ to the reduced form in the algebra
$\widehat{{\rm ACYB}}_n(\alpha,\beta)$, that is, let us consecutively apply the
def\/ining relations~$(a)$ and~$(b)$ to the element~$w$ in any order until unable
to do so. Denote the resulting (noncommutative) polynomial by
$P_n(x_{ij};\alpha,\beta)$. In principal, the polynomial itself can depend on
the order in which the relations~$(a)$ and~$(b)$ are applied. We set
$P_n(x_{ij};\beta):=P_n(x_{ij};0,\beta)$.
\begin{Proposition}[cf.~\protect{\cite[Exercise~6.C5(c)]{ST3}}, \cite{Me,Me-b}] \label{proposition5.1}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] Apart from applying the relation $(a)$ $($commutativity$)$, the
polynomial $P_n(x_{ij};\beta)$
does not depend on the order in which
relations~$(a)$ and~$(b)$ have been applied, and can be written in a~unique way as a linear combination:
\begin{gather*}
P_{n}(x_{ij};\beta)= \sum_{s=1}^{n-1} \beta^{n-s-1}
\sum_{\{i_a \} } \prod_{a=1}^{s} x_{i_a,j_a},
\end{gather*}
where the second summation runs over all sequences of integers
$\{i_a \}_{a=1}^{s}$ such that $n-1 \ge i_1 \ge i_2 \ge \cdots \ge i_s =1$,
and $i_a \le n-a$ for $a=1,\ldots,s-1$; moreover, the corresponding
sequence $\{j_a \}_{a=1}^{n-1}$ can be defined uniquely by
that $\{i_a \}_{a=1}^{n-1}$.
\begin{itemize}\itemsep=0pt
\item It is clear that the polynomial $P(x_{ij}; \beta)$ also can be
written in a unique way as a linear combination of monomials
$\prod\limits_{a=1}^{s} x_{i_a,j_a}$ such that $j_1 \ge j_2 \cdots \ge j_s$.
\end{itemize}
\item[$(2)$] Let us set $\deg(x_{ij})=1$, $\deg(\beta)=0$. Denote by $T_{n}(k,r)$
the number of degree~$k$ monomials in the
polynomial $P(x_{ij};\beta)$ which contain exactly~$r$ factors of the form
$x_{*,n}$. $($Note that $1 \le r \le k \le n-1.)$ Then
\begin{gather*}
T_{n}(k,r)={r \over k} {n+k-r-2 \choose n-2} {n-2 \choose k-1}.
\end{gather*}
In other words,
\begin{gather*}
P_n(t,\beta)= \sum_{1 \le r \le k < n} T_n(k,r) t^{r} \beta^{n-1-k},
\end{gather*}
where $P_n(t,\beta)$ denotes the following specialization
\begin{gather*}
x_{ij} \longrightarrow 1 \qquad \text{if}\quad j < n,\qquad x_{in} \longrightarrow t, \quad \forall\, i=1,\ldots, n-1
\end{gather*}
of the polynomial $P_n(x_{ij} ; \beta)$.
In particular, $T_{n}(k,k)={n-2 \choose k-1}$ and $T_{n}(k,1)=T(n-2,k-1)$,
where
\begin{gather*}
T(n,k):={1 \over k+1} {n+k \choose k} {n \choose k}
\end{gather*}
is equal to the number of Schr\"oder paths $($i.e., consisting of steps
$U=(1,1)$, $D=(1,-1)$, $H=(2,0)$ and never going below the $x$-axis$)$ from
$(0,0)$ to $(2n,0)$, having~$k$ $U$'s, see {\rm \cite[$A088617$]{SL}}.
Moreover, $T_{n}(n-1,r)={\rm Tab}(n-2,r-1)$, where
\begin{gather*}
{\rm Tab}(n,k):={k+1 \over n+1} {2n-k \choose n} = F^{(2)}_{n-k}(k)
\end{gather*}
is equal to the number of standard Young tableaux of the shape $(n,n{-}k)$,
see {\rm \cite[$A009766$]{SL}}. Recall that $F^{(p)}_{n}(b) =
{1+b \over n} {np+b \choose n-1}$ stands for the generalized Fuss--Catalan
number.
\item[$(3)$] After the specialization $x_{ij} \longrightarrow 1$ the polynomial
$P(x_{ij})$ is transformed to the polynomial
\begin{gather*}
P_n(\beta):=\sum_{k=0}^{n-1}N(n,k) (1+\beta)^k,
\end{gather*}
where $N(n,k):={1 \over n}~{n \choose k} {n \choose k+1}$, $k=0,\ldots,n-1$,
stand for the Narayana numbers.
Furthermore,
$P_n(\beta)=\sum\limits_{d=0}^{n-1}s_n(d) \beta^d$, where
\begin{gather*}
s_n(d)={1 \over n+1}{2n-d \choose n} {n-1 \choose d}
\end{gather*}
is the number of ways to draw $n-1-d$ diagonals in a convex $(n+2)$-gon,
such that no two diagonals intersect their interior.
Therefore, the number of $($nonzero$)$ terms
in the polynomial $P(x_{ij};\beta)$ is equal to the $n$-th little
Schr\"oder number $s_n:=\sum\limits_{d=0}^{n-1}s_n(d)$, also known as the
$n$-th super-Catalan number, see, e.g., {\rm \cite[$A001003$]{SL}}.
\item[$(4)$] Upon the specialization $x_{1j} \longrightarrow t$,
$1 \le j \le n$, and that $x_{ij} \longrightarrow 1$ if
$2 \le i < j \le n$, the polynomial~$P(x_{ij};\beta)$ is transformed to the
polynomial
\begin{gather*}
P_n(\beta,t)= t \sum_{k=1}^{n} (1+\beta)^{n-k} \sum_{\pi} t^{p(\pi)},
\end{gather*}
where the second summation runs over the set of Dick paths~$\pi$ of length
$2n$ with exactly~$k$ picks $($UD-steps$)$, and~$p(\pi)$ denotes the number of
valleys $($DU-steps$)$ that touch upon the line $x=0$.
\item[$(5)$] The polynomial $P(x_{ij} ; \beta)$ is invariant under the
action of anti-involution $\phi \circ \tau$, see {\rm \cite[Section~5.1.1]{K}} for
definitions of~$\phi$ and~$\tau$.
\item[$(6)$] Follow {\rm \cite[Exercise~6.C8(c)]{ST3}} consider the
specialization
\begin{gather*}
x_{ij} \longrightarrow t_{i}, \qquad 1 \le i < j \le n,
\end{gather*}
and def\/ine $P_n(t_1,\ldots,t_{n-1};\beta)= P_n(x_{ij}=t_i;\beta)$.
One can show, cf.\ {\rm \cite{ST3}}, that
\begin{gather*
P_n(t_1,\ldots,t_{n-1}; \beta)= \sum \beta^{n-k} t_{i_{1}} \cdots t_{i_{k}},
\end{gather*}
where the sum runs over all pairs $\{(a_1,\ldots,a_k),(i_1,\ldots,i_k) \in
\mathbb{Z}_{\ge 1} \times \mathbb{Z}_{\ge 1} \}$ such that $ 1 \le a_1 < a_2 < \cdots < a_k$,
$ 1 \le i_1 \le i_2 \cdots \le i_k \le n$ and $i_{j} \le a_j$ for all~$j$.
\end{enumerate}
\end{Proposition}
Now we are ready to state our main result about polynomials
$P_n(t_1,\ldots,t_n;\beta)$.
Let $\pi:=\pi_{n} \in {\mathbb S}_n$ be the permutation
\begin{gather*}
\pi= \begin{pmatrix}1& 2& 3& \ldots& n \\
1& n& {n-1}& \cdots& 2 \end{pmatrix}.
\end{gather*}
Then
\begin{gather}\label{equation5.2}
P_n(t_1,\ldots,t_{n-1};\beta)= \left(\prod_{i= 1}^{n-1} t_i^{n-i} \right)
\mathfrak{G}_{\pi}^{(\beta)}\big(t_1^{-1},\dots,t_{n-1}^{-1}\big) = \sum_{{\cal{T}}}
wt({\cal{T}}),
\end{gather}
where $\mathfrak{G}_{w}^{(\beta)}(x_1,\ldots,x_{n-1})$ denotes the
$\beta$-Grothendieck polynomial corresponding to a permutation
$w \in {\mathbb S}_n$, see~\cite{FK1} or Appendix~\ref{appendixA.1}; summation in the right hand
side of the second equality runs over the set of all {\it dissections}
${\cal{T}}$ of a convex $(n+2)$-gon, and $wt({\cal{T}})$ denotes
{\it weight} of a~dissection ${\cal{T}}$, namely,
\begin{gather*}
wt({\cal{T}}) = \prod_{ d \in {\cal{T}}} x_{d} \beta^{n-3- |{\cal{T}}|},
\end{gather*}
where the product runs over diagonals in ${\cal{T}}$, $x_{d} = x_{ij}$, if
diagonal~$d$ connects vertices~$i$ and~$j$, $i < j$, and $|{\cal{T}}|$ denotes the number of diagonals in dissection~${\cal{T}}$.
In particular,
\begin{gather*}
\mathfrak{ G}_{\pi}^{(\beta)}(x_1=1,\ldots,x_{n-1}=1)=\sum_{k=0}^{n-1}
N(n,k) (1+\beta)^{k},
\end{gather*}
where $N(n,k)$ denotes the Narayana numbers, see item~(3) of
Proposition~\ref{proposition5.1}.
More generally, write $P_n(t,\beta)= \sum\limits_{k} P_{n}^{ (k)}(\beta) t^k$. Then
\begin{gather*}
\mathfrak{ G}_{\pi}^{(\beta)}(x_1=t,\,x_i=1,\, \forall\, i \ge 2)=
\sum_{k=0}^{n-1} P_{n-1}^{(k)}\big(\beta^{-1}\big) \beta^{k} t^{n-1-k}.
\end{gather*}
\begin{Comments}\label{comments5.2}\quad
\begin{itemize}\itemsep=0pt
\item Note that if $\beta=0$, then one has
$\mathfrak{G}_{w}^{(\beta=0)}(x_1,\ldots,x_{n-1})=
\mathfrak{S}_{w}(x_1,\ldots,x_{n-1})$, that is the $\beta$-Grothendieck
polynomial at $\beta=0$, is equal to the Schubert polynomial corresponding to
the same permutation~$w$. Therefore, if
\begin{gather*}
\pi= \begin{pmatrix} 1& 2& 3& \ldots& n \\
1& n& {n-1}& \ldots& 2 \end{pmatrix},
\end{gather*}
then
\begin{gather}\label{equation5.3}
\mathfrak{S}_{\pi}(x_1=1,\ldots,t_{n-1}=1)= C_{n-1},
\end{gather}
where $C_m$ denotes the $m$-th Catalan number. Using the formula~\eqref{equation5.3} it is not dif\/f\/icult to check that the following formula for the
principal specialization of the Schubert polynomial $\mathfrak{S}_{\pi}(X_n)$
is true
\begin{gather}\label{equation5.4}
\mathfrak{S}_{\pi}\big(1,q, \ldots,q^{n-1}\big)=q^{n-1 \choose 3} C_{n-1}(q),
\end{gather}
where $C_m(q)$ denotes the Carlitz--Riordan $q$-analogue of the Catalan
numbers, see,\linebreak e.g.,~\cite{ST1}. The formula~\eqref{equation5.4} has been proved in~\cite{FK3} using the observation that~$\pi$ is a~{\it vexillary} permutation, see~\cite{M} for the a def\/inition of the latter. A~combinatorial/bijective proof
of the formula~\eqref{equation5.3} is {\it due} to A.~Woo~\cite{W}.
\item The Grothendieck polynomials, had been def\/ined originally by
A.\,Lascoux and M.-P.\,Sch\"ut\-zen\-berger, see, e.g.,~\cite{LS}, correspond to
the case $\beta=-1$. In this case $P_n(-1)=1$ if~$n \ge 0$, and therefore the
specialization $\mathfrak{G}_{w}^{(-1)}(x_1=1,\ldots,x_{n-1}=1) =1$ for all
$ w \in \mathbb{S}_n$.
\item
In Section~\ref{section5.2.2}, Theorems~\ref{theorem5.5} and~\ref{theorem5.6}, we state a generalization of
the second equality in the formula~\eqref{equation5.2} to the case of Richardson's
permutations of the form $ 1^{k} \times w_{0}^{(n-k)}:= \pi_{k}^{(n)}$, and
relate monomials which appear in a combinatorial formula\footnote{See~\cite{BJS, FK3, KMi} for example.}
for the corresponding $\beta$-Grothendieck polynomial, and/with the
set of $k$-dissections and $k$-triangulations of a~convex $(n+k+1)$-gon,
and the Lagrange inversion formula, see Section~\ref{section5.4.2} for more details.
\end{itemize}
\end{Comments}
Clearly, the Richardson permutations $\pi_{k}^{(0)}$ are special subset of
permutations of the form $1^{k} \times w_{\lambda}:= w_{k}^{(\lambda)}$,
where $w_{\lambda}$ stands for the dominant permutation of shape $\lambda$.
An analogue and extension of the f\/irst equality in the formula~\eqref{equation5.2} for
permutations of the form $w_{1}^{(\lambda)}$ has been proved in~\cite[Theorem~5.4]{EM}. We state here a particular case of that result related with the
Fuss--Catalan numbers obtained independently by the author of the present paper
as a~generalization of \cite[Exercise~8C5(c)]{ST3} and~\cite{W} to
the case of Fuss--Catalan numbers. Namely, let $\lambda =(\lambda_1,\ldots,\lambda_k=1)$ be a~Young diagram such that $\lambda_i-\lambda_{i+1} \le 1$.
Therefore, the boundary~$\partial(\lambda)$ of~$\lambda$, that is the set of
the last boxes in each row of~$\lambda$, is a disjoint union of vertical
intervals. To the last box of the
lowermost interval we attach the generator~$x_{23}$. To the next box of that
interval, if exists, we attach the generator~$u_{24}$ and so on, up to the
top box of that interval is equipped with the generator, say~$x_{2,k_{1}}$.
It is clear that $k_1= \lambda'_1-\lambda'_2+2$. Now let us consider the
next vertical interval. To the bottom box of that interval
we attach the variable $x_{k_1,k_{1}+1}$, to the next box we attach the
variable $x_{k_1,k_1+2}$ and so on. Let the top of that vertical interval is
equipped with the generator~$x_{k_1,k_2}$; it is clear that $k_2= \lambda'_1-\lambda'_3+2$. Applying this procedure successively step by step to each
vertical interval, we attach the variable~$u_{b}$ to each box~$b$ in the boundary of Young diagram~$\lambda$. Finally we attach the monomial
\begin{gather*}
M_{\lambda} = x_{12} \prod_{ b \in \partial(\lambda)} x_{b}.
\end{gather*}
\begin{Theorem}[\cite{EM}]\label{theorem5.1}
Let $\lambda$ be a partition such that $\lambda_i-\lambda_{i+1} \le 1$, $\forall\, i \ge 1$, and set $N:=\lambda'_{1}+2$. Let
$w_{\lambda} \in {\mathbb{S}}_N$ be a unique dominant partition of shape~$\lambda$, and $M_{\lambda} \in {\widehat{{\rm ACYB}}}_{N}(\beta)$ be the monomial
associated with the boundary $\partial(\lambda)$ of partition~$\lambda$.
Then
\begin{gather*}
P_{M_{\lambda}}(x_{ij}=t_i,\, \forall \, i,j; \beta) = t^{\lambda}
{\mathfrak{G}}_{1 \times w_{\lambda}}^{(\beta)}\big(t_1^{-1}, \ldots,t_{N}^{-1}\big),
\end{gather*}
where $t^{\lambda}:=t_{1}^{\lambda'_{1}} \cdots t_{N}^{\lambda'_{N}}$. In
other words, after the specialization $x_{ij} \longrightarrow t_i^{-1}$, $\forall\, i,j$, the spe\-ciali\-zed reduced polynomial corresponding to the monomial~$M_{\lambda}$ is equal to $t^{- \lambda}$ multiplied by the $\beta$-Grothendieck polynomial associated with the permutation $1 \times w_{\lambda}$.
\end{Theorem}
Let us illustrate the above theorem by example. We take $\lambda= 43221$. In this case $N=7=5+2$ and $w:=w_{\lambda}= [1,6,5,4,7,3,2]$. The monomial
corresponding to the boundary of $\lambda$ is equal to
\begin{gather*}
x_{12} x_{23} x_{34} x_{35} x_{56} x_{67} \in {\widehat{{\rm ACYB}}}_{7}.
\end{gather*}
Since the both reduced and $\beta$-Grothendieck polynomials appearing in this
example are huge, we display only its specialized values at $x_{ij}=1$, $\forall\, i,j$ and $t_i=1$, $\forall\, i$. We set also $d:= \beta-1$. It is not dif\/f\/icult to
check that the reduced polynomial corresponding to monomial $ x_{12} x_{23} x_{34} x_{35} x_{56}$ after the specialization $x_{ij}=1$, $\forall\, 1 \le i < j \le 5$, and the identif\/ication $x_{i,6}=x_{1,6}$, $1 \le i \le 5$, is equal to
\begin{gather*}
(9,20,14,3)_{\beta} x_{16}+(9,15,6)_{\beta} x_{16}^2+(4,4)_{\beta} x_{16}^3+
x_{16}^4.
\end{gather*}
Finally after multiplication of the above expression by~$x_{67}$, applying
$3$-term relations $(b)$ in the algebra ${\widehat{{\rm ACYB}}}_7$ to the result obtained,and and taking the specialization $x_{i,7}=1$, $\forall\,i$, we will come to
the following expression
\begin{gather*}
(9,20,14,3)_{\beta} (2+\beta)+(9,15,6)_{\beta} (3+2 \beta)^2+(4,4)_{\beta} (4+3 \beta)+ (5+ 4 \beta)\\
\qquad{} =(66,144,108,32,3)_{\beta}.
\end{gather*}
One can check that the latter polynomial is equal to ${\mathfrak{G}}_{w}^{\beta}(1)$.
\begin{Corollary}[monomials and Fuss--Catalan numbers ${\rm FC}_{n}^{(p+1)}$]
Let $p$, $n$, $b$ be integers, consider diagram $\lambda=(n^{b},(n-1)^{p},(n-2)^{p},\ldots,2^{p},1^{p})$
and dominant permutation $w \in {\mathbb{S}}_{(n-1)p +b+2}$ of shape~$\lambda$. Let us define monomial
\begin{gather*}
M_{n,p,b} =x_{12} \prod_{j=0}^{n-2} \left( \prod_{a=3}^{p+2} x_{j p+2, jp+a}
\right) \prod_{a=3}^{b+2} x_{(n-1) p +2, (n-1) p +a}.
\end{gather*}
Then
\begin{gather*}
P_{M_{n,p,b}}(x_{ij}=1, \, \forall\, i,j)(\beta) = \sum_{k=1}^{n} {\frac{1}{k}}
{\binom{n-1}{k-1}}{\binom{p n -\overline{b}}{k-1}} (\beta +1)^{k-1}.
\end{gather*}
Moreover,
\begin{gather*}
P_{M_{n,p,b}}(x_{ij}=1,\, \forall\, i,j)(\beta=0)= {\frac{1}{n p - \overline{b}+1}} {\binom{n(p+1) - \overline{b}}{n}} = {\frac{1}{n}} {\binom{n(p+1)-\overline{b}}{n-1}},
\end{gather*}
where $\overline{b}:= b-\frac{1-(-1)^{b}}{2}$.
\end{Corollary}
For $b=0$ the right hand side of the above equality is equal to the Fuss--Narayana polynomial, see Theorem~\ref{theorem5.10} and Proposition~\ref{proposition5.5}; a combinatorial interpretation of the value $P_{M_{n,p,b}}(x_{ij}=1$, $\forall\, i,j)(\beta=1)$ one can f\/ind in~\cite{NT}.
Note that reduced expressions for monomial $M_{n,p,b}$ in the (noncommutative)
algebra ${\widehat{{\rm ACYB}}}_n(\beta)$ up to applying the commutativity rules~$(a)$, Def\/inition~\ref{definition5.1}, is unique.
It seems an interesting {\it problem} to construct a natural bijection between the set of monomials in the (noncommutative) reduced expression associated with monomials $M_{n,p,0}$ and the set of $(p+1)$-gulations\footnote{That is the set of dissections of a convex $pk$-gon by (maximal)
collection of non-crossing diagonals such that the all regions obtained are a
convex $(p+2)$-gons of a convex $kp$-gon.}
Finally we remark that there are certain connections of the
$\beta$-Grothendieck
polynomials corresponding to shifted dominance permutations (i.e., permutations
of the form $1^{k} \times w_{\lambda}$) and some generating functions for the set of bounded by~$k$ plane
partitions of shape~$\lambda$, see, e.g.,~\cite{FK3}. In the case
of a staircase shape partition $\lambda = (n-1,\dots, 1)$ one can envision
(cf.\ \cite{SSt, St}) a connection/bijection between the set of $k$-bounded
plane partitions of that shape and $k$-dissections of a convex $(n+k+1)$-gon.
However in the case $k \ge 2$ it is not clear does there exist a monomial
$M$ in the algebra ${\widehat{{\rm ACYB}}}_n$ such that the value of the corresponding reduced polynomial at $x_{ij}= 1$, $\forall\, i,j$ is equal to the number
of $k$-dissections ($k \ge 2$) of a convex $(n+k+1)$-gon.
\begin{Exercises}\label{Exercises5.1}\quad
$(1)$
$(a)$ Let as before,
\begin{gather*}
\pi = \begin{pmatrix} 1& 2& 3& \ldots& n \cr
1& n& {n-1}& \ldots& 2 \end{pmatrix}.
\end{gather*}
{\it Show} that
\begin{gather*} {\mathfrak{S}}_{\pi}(x_1=q,\,x_j=1, \,\forall\, j \not= i) =\sum_{a=0}^{n-2} {n-a-1
\over n-1} {n+a-2 \choose a} q^{a}.
\end{gather*}
Note that the number
\begin{gather*}
{n-k+1 \over n+1} {n+k \choose k}
\end{gather*}
is equal to the
dimension of irreducible representation of the symmetric group
${\mathbb S}_{n+k}$ that corresponds to partition $(n+k,k)$.
$(b)$ Big Schr\"{o}der numbers, paths and polynomials $\mathfrak{G}_{1 \times w_{0}^{(n-1)}}^{(\beta)}(x_1=q, \,x_i=1,\,\forall\, i \ge 2)$.
Let $n \ge 1$ and $k \ge 0$ be integers, denote by $S_{k,n}$ the number of
{\it big Schr\"{o}der paths of type $(k,n)$}, that is lattice paths from the point $(0,0)$ and
ending at point $(2n+k,k)$, using only the steps $U =(1,1)$, $H= (2 ,0)$
and $D=(1,-1)$ and never going below the line $x= 0$. The numbers $S(n):=
S_{0,n}$ commonly known as {\it big Schr\"{o}der numbers}, see, e.g.,
\cite[$A001003$]{SL}. It is well-known that
\begin{gather*}
S_{k,n} = {\frac{k+1}{n}} \sum_{a=0}^{n} {\binom{n}{a}} {\binom{n+k+a}{n-1}}.
\end{gather*}
{\it Show} that
\begin{gather*}
\mathfrak{G}_{1 \times w_{0}^{(n-1)}}^{(\beta)}(x_1=q, \, x_i=1,\, \forall\, i \ge 2) =
\sum_{k=0}^{n-2} S_{k,n-2-k}(\beta) q^{n-k-2},
\end{gather*}
where $S_{k,n}(\beta)$ is the generating functions of the big Schr\"{o}der
paths of type $(k,n)$ according to the number of horizontal steps~$H$.
$(c)$ {\it Show} that the polynomial $\mathfrak{G}_{1 \times w_{0}^{(n-1-1)}}^{(\beta)}(x_1=q,\, x_i=1,\,\forall \,i \ge 2)$ belongs to the ring $\mathbb{N} [q,\beta+1]$. For example, for $n=8$ one has
\begin{gather*}
\mathfrak{G}_{1 \times w_{0}^{(7)}}^{(\beta)}(x_1=q, \,x_i=1,\,\forall\, i \ge 2) = (0,1,15,50,50,15,1)_{\beta+1} t^6 +(0,2,24,60,40,6)_{\beta+1}t^5\\
\qquad{} + (0,3,27,45,15)_{\beta+1} t^4+ (0,4,24,20)_{\beta+1} t^3+(0,5,15)_{\beta+1} t^2+ 6(\beta+1) t+1.
\end{gather*}
{\it Show} that
\begin{gather*}
S_{k,n}(\beta) ={\frac{k+1}{n}} \sum_{a=0}^{n} {\binom{n}{a}}{\binom{n+k+a}{n-1}} \beta^{n-a}= {\frac{k+1}{k+1+n}} {\binom{2n+k}{n}} + \cdots+{\binom{n+k}{n}} \beta^{n}.
\end{gather*}
$(d)$ Write
\begin{gather*}
\mathfrak{G}_{1^{k} \times w_{0}^{(n-k)}}^{(\beta)}(x_1=q, \, x_i=1,\,\forall\, i \ge 2) =A_{k,n}(\beta) q^{n-k-1} +\cdots+ B_{n,k}(\beta).
\end{gather*}
{\it Show} that
\begin{gather*}
A_{k,n}= (1+\beta)^{k} \mathfrak{G}_{k,n-1}^{(\beta)}(x_i=1,\, \forall\, i \ge 1),\qquad
B_{k,n}= \mathfrak{G}_{k-1,n-1}^{(\beta)}(x_i=1,\, \forall\, i \ge 1).
\end{gather*}
(2)~Consider the commutative quotient ${\widetilde{{\rm ACYB}}}_n^{ab}(\alpha,
\beta)$ of the algebra ${\widetilde{{\rm ACYB}}}_n(\alpha,\beta)$,
i.e., assume that
the all generators $\{x_{ij} \,| \,1 \le i < j \le n$ are mutually commute. Denote
by ${\overline{P}}_n(x_{ij};\alpha,\beta)$ the image of polynomial the
$P_n(x_{ij};\alpha,\beta) \in {\widetilde{{\rm ACYB}}}_n(\alpha,\beta)$ in the
algebra ${\widetilde{{\rm ACYB}}}_n^{ab}(\alpha,\beta)$. Finally, def\/ine polynomials
$P_n(t,\alpha,\beta)$~to be the specialization
\begin{gather*}
x_{ij} \longrightarrow 1 \qquad \text{if}\quad j < n, \qquad x_{in} \longrightarrow t \qquad \text{if}
\quad 1 \le i < n.
\end{gather*}
{\it Show} that
$(a)$ Polynomial $P_n(t,\alpha,\beta)$ does not depend on on order in which
relations~$(a)$ and~$(b)$, see Def\/inition~\ref{definition5.1}, have been applied.
$(b)$
\begin{gather*}
P_n(1,\alpha=1,\beta=0)= \sum_{k \ge 0} {(2n-2k) ! \over k ! (n+1-k) !
(n-2k) !},
\end{gather*}
see \cite[$A052709(n)$]{SL} for combinatorial
interpretations of these numbers.
For example,
\begin{gather*}
P_7(t,\alpha,\beta)= t^7+6(1+\beta) t^6+ [(0,5,15)_{\beta+1}+
6 \alpha ] t^5 + [(0,4,24,20)_{\beta+1 }+
\alpha (5,29)_{\beta+1} ] t^4\\
\hphantom{P_7(t,\alpha,\beta)=}{} + [(0,3,27,45,15)_{\beta +1}+
\alpha (4,45,55)_{\beta +1}+14 \alpha^2 ] t^3 \\
\hphantom{P_7(t,\alpha,\beta)=}{} +
[(0,2,24,60,40,6)_{\beta +1} +\alpha (3,48,115,50)_{\beta +1}+
\alpha^2 (21,49)_{\beta +1} ] t^2 \\
\hphantom{P_7(t,\alpha,\beta)=}{} +
[(0,1,15,50,50,15,1)_{\beta +1}+\alpha (2,38,130,110,20)_{\beta +1}+
\alpha^2 (21,91,56)_{\beta +1}\\
\hphantom{P_7(t,\alpha,\beta)=}{}
+ 14 \alpha^3 ]t+\alpha (1,15,50,50,15,1)_{\beta +1} +
\alpha^2(14,70,70,14)_{\beta +1} +\alpha^3 (21,21)_{\beta + 1}.
\end{gather*}
$(c)$ {\it Show} that in fact
\begin{gather*}
P_n(1,\alpha,0)=\sum_{k \ge 0} {1 \over n+1} {2n-2k \choose n} {n+1
\choose k} \alpha^k = \sum_{k \ge 0} {T_{n+2}(n-k,k+1) \over 2n-1-2k}
\alpha^{k},
\end{gather*}
see Proposition~\ref{proposition5.1}(2), for def\/inition of numbers~$T_n(k,r)$. As for a~combinatorial interpretation of the polynomials $P_n(1,\alpha,0)$, see
\cite[$A117434$, $A085880$]{SL}.
$(3)$ Consider polynomials $P_n(t,\beta)$ as it has been def\/ined in
Proposition~\ref{proposition5.1}(2).
{\it Show} that
\begin{gather*}
P_n(t,\beta)= P_n(t,\alpha=0,\beta)= t^n+ \sum_{r=1}^{n-1} t^{r} \left( \sum_{k=0}^{n- 1-r} {r \over n} {n \choose k+r} {n-r-1 \choose k} (1+\beta)^{n-r-k} \right),
\end{gather*}
cf., e.g., \cite[$A033877$]{SL}.
A few comments in order. Several combinatorial interpretations of the
integer numbers
\begin{gather*}
U_n(r,k):= {r \over n+1} {n+1 \choose k+r} {n-r \choose k}
\end{gather*}
are well-known. For example,
if $r=1$, the numbers $U_n(1,k) = {1 \over n}
{n \choose k+1} {n \choose k}$ are equal to the Narayana numbers, see, e.g.,
\cite[$A001263$]{SL};
if $r=2$, the number $U_n(2,k)$ counts the number of Dyck $(n+1)$-paths
whose last descent has length~$2$ and which contain $n-k$ peaks, see \cite[$A108838$]{SL} for details.
Finally, it's easily seen, that $P_n(1, \beta) = A127529(n)$, and
$P_n(t,1)= A033184 (n)$, see~\cite{SL}.
$(4)$ {\it Show} that
\begin{gather*}
P_n(t,\alpha,\beta) \in \mathbb{N} [t,\alpha] [\beta+1],
\end{gather*}
that is the polynomial $P_n(t,\alpha,\beta)$ is a polynomial of $\beta +1$
with coef\/f\/icients from the ring~$\mathbb{N}[t,\alpha]$.
{\it Show} that
\begin{gather*}
P_n(0,1,\beta) \in \mathbb{N} [\beta +2].
\end{gather*}
For example,
\begin{gather*}
P_7(0,1,\beta)=(0,3,8,14,10,1)_{\beta+2}, \qquad P_8(0,1,\beta)=(1,3,11,25,35,15,1)_{\beta+2}.
\end{gather*}
{\it Show} that~\cite{SL}
\begin{gather*}
P_n(1,1,0)= A052709(n+1),
\end{gather*}
that is the number of underdiagonal lattice paths from $(0,0)$ to $(n-1,n-1)$
and such that each step is either $(1,0)$, $(0,1)$, or $(2,1)$.
For example, $P_7(1,1,0)= 1697 = A052709(8)$. Cf.\ with the next exercise.
{\it Show} that~\cite{SL}
\begin{gather*}
P_n(0,1,0) = A052705(n),
\end{gather*}
namely, the number of underdiagonal paths from (0,0) to the line $x=n-2$, using
only steps $(1,0)$, $(0,1)$ and $NE= (2,1)$. For example,
\begin{gather*}
P_7(0,1,0)= 36+106+120+64+15+1 = 342 = A052705(7).
\end{gather*}
{\it Show} that~\cite{SL}
\begin{gather*}
{\frac{\partial}{\partial a}} P_n(a,{\boldsymbol{b}} ={\boldsymbol{1}},{\boldsymbol{\beta}}={\boldsymbol{0}},{\boldsymbol{\alpha}}={\boldsymbol{1}},
{\boldsymbol{y}} = {\boldsymbol{z}} ={\boldsymbol{1}})= A005775,
\end{gather*}
that is the number number of paths in the half-plane $x \ge 0$ from $(0,0)$
to $(n-1,2)$ or $(n-1,-3)$, and consisting of steps $U=(1,1)$, $D=(1,-1)$
and $H=(1,0)$ that contain at least one $UUU$ but avoid $UUU's$ starting
above level~$0$.
\end{Exercises}
\subsubsection{Multiparameter deformation of Catalan, Narayana and Schr\"oder numbers}\label{section5.1.1}
Let $\mathfrak{b} = (\beta_1,\ldots,\beta_{n-1})$ be a set of mutually
commuting parameters. We def\/ine a multiparameter analogue of the associative
quasi-classical Yang--Baxter algebra $\widehat{{\rm MACYB}}_n(\mathfrak{b})$ as
follows.
\begin{Definition}[cf.\ Def\/inition~\ref{definition2.4}]\label{definition5.2}
The multiparameter associative quasi-
classical Yang--Baxter algebra of weight ${\mathfrak{b}}$, denoted by
$\widehat{{\rm MACYB}}_n(\mathfrak{b})$, is an associative
algebra, over the ring of polynomials $\mathbb{Z}[\beta_1,\ldots,\beta_{n-1}]$,
generated by the set of elements $\{ x_{ij}, \, 1 \le i < j \le n\}$,
subject to the set of relations
\begin{enumerate}\itemsep=0pt
\item[$(a)$] $x_{ij} x_{kl}=x_{kl} x_{ij}$ if $ \{i,j\} \cap \{k,l \}=\varnothing$,
\item[$(b)$] $x_{ij} x_{jk}=x_{ik} x_{ij}+x_{jk} x_{ik}+\beta_{i} x_{ik}$ if
$1 \le 1 <i < j \le n$.
\end{enumerate}
\end{Definition}
Consider the ``Coxeter element'' $w_n \in \widehat{{\rm MACYB}}_n(\mathfrak{b})$
which is equal to the ordered product of ``simple generators'':
\begin{gather*}
w_n:=\prod_{a=1}^{n-1} x_{a,a+1}.
\end{gather*}
Now we can use the same method as in~\cite[Exercise~8.C5(c)]{ST3}, see
Section~\ref{section5.1}, to def\/ine the {\it reduced form} of the Coxeter element~$w_n$.
Namely, let us bring the element~$w_n$ to the reduced form in the algebra~
$\widehat{{\rm MACYB}}_n({\mathfrak{b}})$, that is, let us consecutively apply the
def\/ining relations $(a)$~and~$(b)$ to the element~$w_n$ in any order until
unable to do so. Denote the resulting (noncommutative) polynomial by
$P(x_{ij};\mathfrak{b})$. In principal, the polynomial itself can depend on
the order in which the relations~$(a)$ and~$(b)$ are applied.
\begin{Proposition}[cf.~\protect{\cite[Exercise~8.C5(c)]{ST3}, \cite{Me,Me-b}}]\label{proposition5.2}
The specialized
polynomial $P(x_{ij} =1$, $\forall\, i,j, \, \mathfrak{b})$ does not depend on the
order in which relations~$(a)$ and~$(b)$ have been applied.
\end{Proposition}
To state our main result of this subsection, let us def\/ine polynomials
\begin{gather*}
Q(\beta_1,\ldots,\beta_{n-1}):=
P(x_{ij}=1, \,\forall \,i,j ; \,\beta_1-1,\beta_2-1,\ldots,\beta_{n-1}-1).
\end{gather*}
\begin{Example}\label{example5.1}
\begin{gather*}
Q(\beta_1,\beta_2)=1+2 \beta_1 +\beta_2 +\beta_1^2,\\
Q(\beta_1,\beta_2,\beta_3)=1+ 3 \beta_1+ 2 \beta_2+\beta_3 +3 \beta_1^2+ \beta_1 \beta_2 +\beta_1 \beta_3+ \beta_2^2+\beta_1^3,\\
Q(\beta_1,\beta_2,\beta_3,\beta_4)= 1+4 \beta_1+3 \beta_2+ 2 \beta_3+ \beta_4+
\beta_1 (6 \beta_1 + 3 \beta_2+3 \beta_3 +2 \beta_4) \\
\hphantom{Q(\beta_1,\beta_2,\beta_3,\beta_4)=}{} + \beta_2 (3 \beta_2+\beta_3 + \beta_4) +\beta_3^2 +
\beta_1^2 (4 \beta_1+\beta_2+\beta_3+\beta_4)\\
\hphantom{Q(\beta_1,\beta_2,\beta_3,\beta_4)=}{}
+\beta_1 (\beta_2^2+\beta_3^2)+
\beta_2^3 +\beta_1^4.
\end{gather*}
\end{Example}
\begin{Theorem}\label{theorem5.2}
Polynomial $Q(\beta_1,\ldots,\beta_{n-1})$ has non-negative
integer coefficients.
\end{Theorem}
It follows from \cite{ST3} and Proposition~\ref{proposition5.1}, that
\begin{gather*}
Q(\beta_1,\ldots,\beta_{n-1})\bigl|_{\beta_1=1,\ldots,\beta_{n-1}=1} =
{\rm Cat}_n.
\end{gather*}
Polynomials $Q(\beta_1,\ldots,\beta_{n-1})$ and
$Q(\beta_1+1,\ldots,\beta_{n-1}+1)$ can be considered as a multiparameter
deformation of the Catalan and (small) Schr\"oder numbers
correspondingly, and the homogeneous degree~$k$ part of
$Q(\beta_1,\ldots,\beta_{n-1})$ as a multiparameter analogue of Narayana
numbers.
\subsection[Grothendieck and $q$-Schr\"oder polynomials]{Grothendieck and $\boldsymbol{q}$-Schr\"oder polynomials}\label{section5.2}
\subsubsection{Schr\"oder paths and polynomials}\label{section5.2.1}
\begin{Definition} \label{definition5.3}
A Schr\"oder path of the length $n$ is an over diagonal
path from $(0,0)$ to $(n,n)$ with steps $(1,0)$, $(0,1)$ and steps
$D=(1,1)$ {\it without} steps of type $D$ on the diagonal $x=y$.
\end{Definition}
If $p$ is a Schr\"oder path, we denote by $d(p)$ the number of the
diagonal steps resting on the path~$p$, and by $a(p)$ the number of unit
squares located between the path~$p$ and the diagonal $x=y$. For each (unit)
diagonal step $D$ of a path $p$ we denote by $i(D)$ the $x$-coordinate of
the column which contains the diagonal step~$D$. Finally, def\/ine the index~$i(p)$ of a path~$p$ as the some of the numbers~$i(D)$ for all diagonal steps
of the path~$p$.
\begin{Definition}\label{definition5.4}
Def\/ine $q$-Schr\"oder polynomial $S_n(q;\beta)$ as follows
\begin{gather}\label{equation5.5}
S_n(q;\beta)= \sum_{p} q^{a(p)+i(p)} \beta^{d(p)},
\end{gather}
where the sum runs over the set of all Schr\"oder paths of length~$n$.
\end{Definition}
\begin{Example}\label{example5.2}
\begin{gather*}
S_1(q;\beta)=1, \qquad S_2(q;\beta)=1+q+\beta q,\\
S_3(q;\beta)=1+2 q+q^2+q^3+\beta \big(q+2q^2+2q^3\big )+\beta^2 q^3,\\
S_4(q;\beta)=1+3q+3q^2+3q^3+2q^4+q^5+q^6+\beta\big(q+3q^2+5q^3+6q^4+3q^5+3q^6\big)\\
\hphantom{S_4(q;\beta)=}{} + \beta^2\big(q^3+2q^4+3q^5+3q^6\big)+\beta^3 q^6.
\end{gather*}
\end{Example}
\begin{Comments}\label{comments5.3}
The $q$-Schr\"oder polynomials def\/ined by the
formula \eqref{equation5.5} are {\it different} from the $q$-analogue of
Schr\"oder polynomials which has been considered in~\cite{BK}. It seems
that there are no simple connections between the both.
\end{Comments}
\begin{Proposition}[recurrence relations for $q$-Schr\"oder polynomials]\label{proposition5.3}
The $q$-Schr\"oder polynomials satisfy the following relations
\begin{gather*
S_{n+1}(q;\beta)= \big(1+q^{n}+\beta q^{n}\big) S_n(q;\beta)+\sum_{k=1}^{k=n-1}
\big(q^k+\beta q^{n-k}\big) S_{k}(q;q^{n-k} \beta) S_{n-k}(q;\beta),
\end{gather*}
and the initial condition $S_1(q;\beta)=1$.
\end{Proposition}
Note that $P_n(\beta)=S_n(1;\beta)$ and in particular, the polynomials
$P_n(\beta)$ satisfy the following recurrence relations
\begin{gather}\label{equation5.7}
P_{n+1}(\beta)=(2+\beta)~P_n(\beta)+(1+\beta)~\sum_{k=1}^{n-1}~P_k(\beta)~
P_{n-k}(\beta).
\end{gather}
\begin{Theorem}[evaluation of the Schr\"oder--Hankel determinant]\label{theorem5.3}
Consider permutation
\begin{gather*}
\pi_{k}^{(n)}= \begin{pmatrix} 1& 2& \ldots& k& k+1& k+2&\ldots& n \\
1& 2& \ldots& k& n& {n-1}& \ldots& k+1 \end{pmatrix}.
\end{gather*}
Let as before
\begin{gather}\label{equation5.8}
P_n(\beta)=\sum_{j=0}^{n-1} N(n,j) (1+\beta)^{j},\qquad n \ge 1,
\end{gather}
be Schr\"oder polynomials. Then
\begin{gather*
(1+\beta)^{{k \choose 2}} \mathfrak{G}_{\pi_{k}^{(n)}}^{(\beta)}(x_1=1,\ldots,
x_{n-k}=1) =\operatorname{Det} \vert P_{n+k-i-j}(\beta) \vert_{1 \le i,j \le k}.
\end{gather*}
\end{Theorem}
Proof is based on an observation that the permutation $\pi_{k}^{(n)}$ is a~{\it vexillary} one and the recurrence relations~\eqref{equation5.7}.
\begin{Comments} \label{comments5.4}
{\bf (1)} In the case $\beta=0$, i.e., in the case of {\it Schubert
polynomials}, Theorem~\ref{theorem5.3} has been proved in~\cite{FK3}.
{\bf (2)} In the cases when $\beta=1$ and $ 0 \le n-k \le 2$, the value of
the determinant in the r.h.s.\ of~\eqref{equation5.8} is known\footnote{See, e.g.,~\cite{BK}, or
M.~Ichikawa talk ``Hankel determinants of Catalan, Motzkin and
Schr\"oder numbers and its $q$-analogue'',
\url{http://www.uec.tottori-u.ac.jp/~mi/talks/kyoto07.pdf}.}. One can check
that in the all cases mentioned above, the formula~\eqref{equation5.8}
gives the same results.
{\bf (3)} {\it Grothendieck and Narayana polynomials.}
It follows from the expression~\eqref{equation5.7} for the Narayana--Schr\"oder
polynomials that $P_n(\beta-1) = \mathfrak{N}_n(\beta)$,
where
\begin{gather*}
\mathfrak{N}_n(\beta):=\sum_{j=0}^{n-1} {1\over n} {n \choose j} {n \choose
j+1} \beta^{j},
\end{gather*}
denotes the $n$-th Narayana polynomial. Therefore, $P_n(\beta -1)=
\mathfrak{N}_n(\beta)$ is a
symmetric polynomial in $\beta$ with non-negative integer coef\/f\/icients.
Moreover, the value of the polynomial $P_n(\beta-1)$ at $\beta=1$ is equal to
the $n$-th Catalan number $C_n:= {1 \over n+1} {2n \choose n}$.
It is well-known, see, e.g.,~\cite{Su}, that the Narayana polynomial
$\mathfrak{N}_n(\beta)$ is equal to the ge\-ne\-ra\-ting function of the statistics
$\pi(\mathfrak{p})= (\text{number of peaks of a Dick path $\mathfrak{p}$})-1$ on the set ${\rm Dick}_n$ of Dick paths of the length~$2n$
\begin{gather*}
\mathfrak{N}_n(\beta)=\sum_{\mathfrak{p}} \beta^{\pi(\mathfrak{p})}.
\end{gather*}
Moreover, using the Lindstr\"om--Gessel--Viennot lemma\footnote{See, e.g.,
\url{https://en.wikipedia.org/wiki/Lindstrom-Gessel-Viennot_lemma}.},
one can see that
\begin{gather}\label{equation5.10}
\operatorname{DET} | \mathfrak{N}_{n+k-i-j}(\beta) |_{1 \le i,j \le k} =
\beta^{{k \choose 2}} \sum_{(\mathfrak{p}_1,\ldots,\mathfrak{p}_k)} \beta^{\pi(\mathfrak{p}_1) + \cdots+\pi(\mathfrak{p}_k)},
\end{gather}
where the sum runs over $k$-tuple of non-crossing Dick paths $(\mathfrak{p}_1,
\ldots,\mathfrak{p}_k)$ such that the path $\mathfrak{p}_i$ starts from the
point $(i-1,0)$ and has length $2(n-i+1)$, $i=1,\ldots,k$.
We {\it denote} the sum in the r.h.s.\ of~\eqref{equation5.10} by
$\mathfrak{N}_n^{(k)}(\beta)$. Note that $\mathfrak{N}_{k-1}^{(k)}(\beta)= 1$
for all $k \ge 2$.
Thus, $\mathfrak{N}_n^{(k)}(\beta)$ is a symmetric polynomial in $\beta$ with
non-negative integer coef\/f\/icients, and
\begin{gather*}
\mathfrak{N}_n^{(k)}(\beta=1)= C_n^{(k)}= \prod_{1 \le i \le j \le n-k}
{2k+i+j \over i+j} = \prod_{2 a \le n-k-1} {{2n-2a \choose 2k} \over {2k+2a+1
\choose 2k}}.
\end{gather*}
As a corollary we obtain the following statement
\begin{Proposition}\label{proposition5.4}
Let $n \ge k$, then
\begin{gather*}
\mathfrak{G}_{\pi_{k}^{(n)}}^{(\beta-1)}(x_1=1,\ldots,x_n=1)=
\mathfrak{N}_n^{(k)}(\beta).
\end{gather*}
\end{Proposition}
Summarizing, the specialization $\mathfrak{G}_{\pi_{k}^{(n)}}^{(\beta-1)}(x_1=1,
\ldots,x_n=1)$ is a symmetric polynomial in $\beta$ with non-negative
integer coef\/f\/icients, and coincides with the generating function of the
statistics $\sum\limits_{i=1}^{k} \pi(\mathfrak{p}_i)$ on the set $k$-${\rm Dick}_n$ of
$k$-tuple of non-crossing Dick paths $(\mathfrak{p}_1,\ldots,\mathfrak{p}_k)$.
\begin{Example} \label{example5.3}
Take $n=5$, $k=1$. Then $\pi_{1}^{(5)}=(15432)$ and one has
\begin{gather*}
\mathfrak{G}_{\pi_{1}^{(5)}}^{(\beta)} \big(1,q,q^2,q^3\big)=q^4(1,3,3,3,2,1,1)+q^5
(1,3,5,6,3,3) \beta+q^7(1,2,3,3) \beta^2+q^{10} \beta^3.
\end{gather*}
It is easy to compute the Carlitz--Riordan $q$-analogue of the Catalan number
$C_5$, namely,
\begin{gather*}
C_5(q)=(1,3,3,3,2,1,1).
\end{gather*}
\end{Example}
\begin{Remark} \label{remark5.1}
The value $\mathfrak{N}_n(4)$ of the Narayana polynomial at $\beta = 4$ has the
following combinatorial interpretation:
$\mathfrak{N}_{n}(4)$ is equal to the number of dif\/ferent lattice paths
from the point $(0,0)$ to that $(n,0)$ using steps from the set
$\Sigma = \{ (k,k) \, \text{or}\, (k,-k), \, k \in \mathbb{Z}_{ > 0} \}$, that never go below the
$x$-axis, see \cite[$A059231$]{SL}.
\end{Remark}
\begin{Exercises}\label{exercises5.2}\quad
\begin{enumerate}\itemsep=0pt
\item[$(a)$] Show that
\begin{gather*} \gamma_{k,n}:= { C_{n}^{(k+1)} \over C_{n}^{(k)}} = {(2n-2k) ! (2k+1) !
\over (n-k) ! (n+k+1) !}.
\end{gather*}
\item[$(b)$] Show that
$\gamma_{k,n} \le 1$ if $k \le n \le 3k+1$, and $\gamma_{k,n} \ge
2^{n-3k-1}$ if $n > 3k+1$.
\end{enumerate}
\end{Exercises}
{\bf (4)} {\it Polynomials $\mathfrak{F}_{w}(\beta)$,
$\mathfrak{H}_{w}(\beta)$, $\mathfrak{H}_{w}(q,t;\beta)$ and $\mathfrak{R}_{w}(
q;\beta)$.}
Let $w \in \mathbb{S}_n$ be a permutation,
$\mathfrak{G}^{(\beta)}_{w}(X_n)$ and~$\mathfrak{G}^{(\beta)}_{w}(X_n,Y_n)$
be the corresponding $\beta$-Grothendieck and double
$\beta$-Grothendieck polynomials. We denote by
$\mathfrak{G}^{(\beta)}_{w}(1)$ and by
$\mathfrak{G}^{(\beta)}_{w}(1;1)$ the specializations
$X_n:=(x_1=1,\ldots$, $x_n=1)$,
$Y_n:=(y_1=1,\ldots,y_n=1)$ of the $\beta$-Grothendieck polynomials introduced
above.
\begin{Theorem} \label{theorem5.4} Let $w \in \mathbb{S}_n$ be a permutation.
Then
\begin{enumerate}\itemsep=0pt
\item[$(i)$] The polynomials $\mathfrak{F}_{w}(\beta):= \mathfrak{G}_{w}^{(\beta-1)}(1)$ and $\mathfrak{H}_{w}(\beta):=\mathfrak{G}_{w}^{(\beta-1)}(1;1)$
have both non-negative integer coefficients.
\item[$(ii)$] One has
\begin{gather*}
\mathfrak{H}_{w}(\beta)=(1+\beta)^{\ell(w)} \mathfrak{F}_{w}\big(\beta^2\big).
\end{gather*}
\item[$(iii)$] Let $w \in \mathbb{S}_n$ be a permutation, define polynomials
\begin{gather*}
\mathfrak{H}_{w}(q,t;\beta):=
\mathfrak{G}_{w}^{(\beta)}(x_1=q,x_2=q, \ldots,x_n=q,y_1=t,y_2=t,\ldots,
y_n=t)
\end{gather*}
to be the specialization $\{x_i=q,\,y_i=t,\, \forall\, i \}$ of the double
$\beta$-Grothendieck polynomial $\mathfrak{G}_{w}^{(\beta)}(X_n,Y_n)$.
Then
\begin{gather*}
\mathfrak{H}_{w}(q,t;\beta) =(q+t+\beta q t)^{\ell(w)} \mathfrak{F}_{w}((1+\beta q)(1+\beta t)).
\end{gather*}
In particular, $\mathfrak{H}_{w}(1,1;\beta) = (2+\beta)^{\ell(w)}
\mathfrak{F}_{w}((1+\beta)^2)$.
\item[$(iv)$] Let $w \in \mathbb{S}_n$ be a permutation, define polynomial
\begin{gather*}
{\cal{R}}_{w}(q;\beta) := \mathfrak{G}_{w}^{(\beta-1)}(x_1=q,x_2=1,x_3=1,\ldots)
\end{gather*}
to be the specialization $\{x_1=q,\, x_i=1,\, \forall\, i \ge 2 \}$, of the
$(\beta-1)$-Grothendieck polynomial $\mathfrak{G}_{w}^{(\beta -1)}(X_n)$.
Then
\begin{gather*} {\cal{R}}_{w}(q;\beta) = q^{w(1)-1} \mathfrak{R}_{w}(q;\beta),
\end{gather*}
where $\mathfrak{R}_{w}(q;\beta)$ is a polynomial in $q$ and $\beta$ with
non-negative integer coefficients, and $\mathfrak{R}_{w}(0;\beta=0)=1$.
\item[$(v)$] Consider permutation $w_n^{(1)}:= [1,n,n-1,n-2,\dots, 3,2]
\in \mathbb{S}_n$.
Then $\mathfrak{H}_{w_{n}^{(1)}}(1,1; 1) =
3^{{n-1 \choose 2}} \mathfrak{N}_n(4)$.
\end{enumerate}
\end{Theorem}
In particular, if $w_n^{(k)}= (1,2,\ldots,k,n,n-1,\ldots,k+1) \in {\mathbb{S}}_n$, then
\begin{gather*}
{\mathfrak{S}}_{w_{n}^{(k)}}^{(\beta-1)}(1;1) =(1+\beta)^{{n-k \choose 2}} {\mathfrak{S}}_{w_{n}^{(k)}}^{(\beta-1)}\big(\beta^2\big).
\end{gather*}
See Remark~\ref{remark5.1} for a combinatorial interpretation of the number
$\mathfrak{N}_{n}(4)$.
\begin{Example}\label{example5.4}
Consider permutation $v= [2,3,5,6,8,9,1,4,7] \in
\mathbb{S}_{9}$ of the length $12$, and set
$x:=(1+\beta q)(1+\beta t)$. One can check that
\begin{gather*}
\mathfrak{H}_{v}(q,t ; \beta) = x^{12} (1+2 x)\big(1+6 x+19 x^2 + 24 x^3 + 13 x^4\big),
\end{gather*}
and $\mathfrak{F}_{v} (\beta)
= (1+2 \beta)(1 +6 \beta+ 19 \beta^2 + 24 \beta^3 + 13 \beta^4)$.
Note that $\mathfrak{F}_{v}( \beta=1) = 27 \times 7$, and $7 = {\rm AMS}(3)$,
$26={\rm CSTCTPP}(3)$, cf.\ Conjecture~\ref{conjecture5.4}, Section~\ref{section5.2.4}.
\end{Example}
\begin{Remark}\label{remark5.2}
One can show, cf.~\cite[p.~89]{M}, that if
$w \in \mathbb{S}_{n}$, then ${\cal{R}}_w(1,\beta)=
{\cal {R}}_{w^{-1}}(1,\beta)$.
However, the equality ${\mathfrak{R}}_w(q,\beta) =
{\mathfrak{R}}_{w^{-1}}(q,\beta)$ can be violated, and it seems that
in general, there are no simple connections between polynomials
${\mathfrak{R}}_w(q,\beta)$ and ${\mathfrak{R}}_{w^{-1}}(q,\beta)$, if so.
\end{Remark}
From this point we shell use the notation $(a_{0},a_{1},\ldots,a_{r})_{\beta} := \sum\limits_{j=0}^{r} a_{j} \beta^{j}$, etc.
\begin{Example}\label{example5.5}
Let us take $w = [1,3,4,6,7,9,10,2,5,8]$. Then
\begin{gather*}
\mathfrak{R}_{w}(q,\beta)= (1, 6, 21, 36, 51, 48, 26)_{\beta} +
q \beta (6, 36, 126, 216, 306, 288, 156)_{\beta}\\
\hphantom{\mathfrak{R}_{w}(q,\beta)=}{}+
q^2 \beta^3 (20, 125, 242, 403, 460, 289)_{\beta} +
q^3 \beta^5 (6, 46, 114, 204, 170)_{\beta}.
\end{gather*}
Moreover,
$\mathfrak{R}_{w}(q,1)= (189,1134,1539,540)_{q}$.
On the other hand,
$w^{-1}\! = [1,8,2,3,9,4,5,10,6,7]$, and
\begin{gather*}
\mathfrak{R}_{w^{-1}}(q,\beta) =
(1, 6, 21, 36, 51, 48, 26)_{\beta}+
q \beta (1, 6, 31, 56, 96, 110, 78)_{\beta}\\
\hphantom{\mathfrak{R}_{w^{-1}}(q,\beta) =}{} +
q^2 \beta (1, 6, 27, 58, 92, 122, 120, 78)_{\beta}+
q^3 \beta (1, 6, 24, 58, 92, 126, 132, 102, 26)_{\beta}\\
\hphantom{\mathfrak{R}_{w^{-1}}(q,\beta) =}{}
+
q^4 \beta (1, 6, 22, 57, 92, 127, 134, 105, 44)_{\beta}\\
\hphantom{\mathfrak{R}_{w^{-1}}(q,\beta) =}{}
+
q^5 \beta (1, 6, 21, 56, 91, 126, 133, 104, 50)_{\beta}\\
\hphantom{\mathfrak{R}_{w^{-1}}(q,\beta) =}{}
+
q^6 \beta (1, 6, 21, 56, 91, 126, 133, 104, 50)_{\beta}.
\end{gather*}
Moreover, $\mathfrak{R}_{w^{-1}}(q,1)=(189,378,504,567,588, 588,588)_{q}$.
Notice that $w= 1 \times u$, where $u=[2,3,5,6,8,9,1,4,7]$. One can show that
\begin{gather*}
\mathfrak{R}_{u}(q,\beta)= (1, 6, 11, 16, 11)_{\beta}+
q \beta^2 (10, 20, 35, 34)_{\beta}+
q^2 \beta^4 (5, 14, 26)_{\beta}.
\end{gather*}
On the other hand,
$u^{-1}= [7,1,2,8,3,4,9,5,6]$ and
\begin{gather*}
\mathfrak{R}_{u^{-1}}(1,\beta)= (1,6,21,36,51,48,26)_{\beta} = \mathfrak{R}_{u}(1,\beta).
\end{gather*}
\end{Example}
Recall that by our def\/inition $(a_{0},a_{1},\ldots,a_{r})_{\beta} :=
\sum\limits_{j=0}^{r} a_{j} \beta^{j}$.
\end{Comments}
\subsubsection[Grothendieck polynomials and $k$-dissections]{Grothendieck polynomials and $\boldsymbol{k}$-dissections}\label{section5.2.2}
Let $ k \in \mathbb{N}$ and $n \ge k-1$,~be a integer, def\/ine {\it a $k$-dissection} of
a convex $(n+k+1)$-gon to be a~collection $\cal{E}$ of diagonals in $(n+k+1)$-gon not containing $(k+1)$-subset
of pairwise crossing diagonals and such that at least $2(k-1)$ diagonals are
coming from each vertex of the $(n+k+1)$-gon in question. One can show that
the number of diagonals in any $k$-dissection~$\cal{E}$ of a convex
$(n+k+1)$-gon contains at least $(n+k+1)(k-1)$ and at most $n(2k-1)-1$
diagonals. We def\/ine the {\it index} of a $k$-dissection $\cal{E}$ to be
$i({\cal{E}})= n(2k-1) -1- \# | {\cal{E}}|$. Denote by
\begin{gather*}
{\cal{T}}_n^{(k)}(\beta)=\sum_{\cal{E}} \beta^{i(\cal{E})}
\end{gather*}
the generating function for the number of $k$-dissections with a f\/ixed
index, where the above sum runs over the set of all $k$-dissections of a~convex $(n+k+1)$-gon.
\begin{Theorem}\label{theorem5.5}
\begin{gather*}
\mathfrak{G}_{\pi_{k}^{(n)}}^{(\beta)}(x_1=1,\ldots,x_n=1)= {\cal{T}}_{n}^{
(k)}(\beta).
\end{gather*}
\end{Theorem}
Mopre generally, let $ n \ge k > 0$ be integers, consider a convex $(n+k+1)$-gon $P_{n+k+1}$ and a
vertex $v_0 \in P_{n+k+1}$. Let us label clockwise the vertices of
$P_{n+k+1}$ by the numbers $1,2,\ldots,n+k+1$ starting from the vertex~$v_0$.
Let $\operatorname{Dis}(P_{n+k+1})$ denotes the set of all $k$-dissections of the $(n+k+1)$-gon $P_{n+k+1}$. We denote by $D_0:=\operatorname{Dis}_{0}(P_{n+k+1})$ the ``minimal''
$k$-dissection of the $(n+k+1)$-gon $P_{n+k+1}$ in question
consisting of the set of diagonals connecting vertices~$v_{a}$ and
$v_{\overline{a+r}}$, where $ 2 \le r \le k$, $1 \le a \le n+k+1$, and for
any positive integer $a$ we denote by $\overline{a}$ a unique integer such
that $1 \le \overline{a} \le n+k+1$ and $a \equiv \overline{a} ({\rm mod}\, ( n+k+1))$.
For example, if $k=1$, then $\operatorname{Dis}_0(P_{n+2}) =\varnothing$; if $k=3$ and $n=4$,
in other words, $P_8$ is a octagon, the minimal $3$-dissection consists of~$16$ diagonals connecting vertices with the following labels
\begin{gather*}
1 \rightarrow 3 \rightarrow 5 \rightarrow 7 \rightarrow \overline{9}=1, \qquad
2 \rightarrow 4 \rightarrow 6 \rightarrow 8 \rightarrow \overline{10}=2, \\
1 \rightarrow 4 \rightarrow 7 \rightarrow \overline{10}=2 \rightarrow 5
\rightarrow 8 \rightarrow \overline{11}=3 \rightarrow 6 \rightarrow \overline{9}=1.
\end{gather*}
Now let $D \in \operatorname{Dis}(P_{n+k+1})$ be a dissection. Consider a
diagonal $d_{ij} \in (D {\setminus} D_{0})$, $i < j$ which connects vertex $v_i$
with that~$v_j$. We attach variable $x_i$ to the diagonal~$d_{ij}$ in
question and consider the following expression
\begin{gather*}
{\cal{T}}_{P_{n+k+1}}(X_{n+k+1}) = \sum_{D \in \operatorname{Dis}(P_{n+k+1})} \beta^{\#|D {\setminus} D_0|} \sum_{d_{ij} \in (D {\setminus} D_{0}) \atop i < j} \prod x_{i}.
\end{gather*}
\begin{Theorem} \label{theorem5.6}
One has
\begin{gather*}
{\cal{T}}_{P_{n+n+1}}(X_{n+k+1}) = \beta^{k(n-k)} \prod_{a=1}^{n} x_a^{\min(n-a+1,n-k)} \mathfrak{G}_{w_{k}^{n}}^{\beta^{-1}}\big(x_1^{-1},\dots,x_{n}^{-1}\big).
\end{gather*}
\end{Theorem}
\begin{Exercises} \label{exercises5.3}
It is not dif\/f\/icult to check that
\begin{gather*}
{\mathfrak{G}}_{15432}^{\beta}(X_5) = \beta^3 x_1^3 x_2^3 x_3^2 x_4
+ \beta^2 (x_1^3 x_2^3 x_3
+ 2 x_1^3 x_2^3 x_3 x_4
+ 3 x_1^3 x_2^2 x_3^2 x_4
+ 3 x_1^2 x_2^3 x_3^2 x_4) \\
\hphantom{{\mathfrak{G}}_{15432}^{\beta}(X_5) =}{}
+ \beta (x_1^3 x_2^3 x_3
+ x_1^3 x_2^3 x_4 + 2 x_1^3 x_2^2 x_3
+ 2 x_1^2 x_2^3 x_3^2 + 3 x_1^3 x_2^2 x_3 x_4
+ 3 x_1^3 x_2 x_3^2 x_4\\
\hphantom{{\mathfrak{G}}_{15432}^{\beta}(X_5) =}{}
+ 3 x_1^2 x_2^3 x_3 x_4
+ 3 x_1^2 x_2^2 x_3^2 x_4
+ 3 x_1 x_2^3 x_3^2 x_4)
+
x_1^3 x_2^2 x_3
+ x_1^3 x_2^2 x_4
+ x_1^3 x_2 x_3^2\\
\hphantom{{\mathfrak{G}}_{15432}^{\beta}(X_5) =}{}
+x_1^3 x_2 x_3 x_4
+x_1^3 x_3^2 x_4
+x_1^2 x_2^3 x_3
+x_1^2 x_2^3 x_4
+x_1^2 x_2^2 x_3^2
+x_1^2 x_2^2 x_3 x_4
+x_1^2 x_2 x_3^2 x_4\\
\hphantom{{\mathfrak{G}}_{15432}^{\beta}(X_5) =}{}
+x_1 x_2^3 x_3^2
+x_1 x_2^3 x_3 x_4
+x_1 x_2^2 x_3^2 x_4
+x_2^3 x_3^2 x_4.
\end{gather*}
Describe
bijection between dissections of hexagon $P_{6}$ (the
case $k=1$, $n=4$) and the above listed monomials involved in the $\beta$-Grothendieck polynomial ${\mathfrak{G}}_{15432}^{\beta}(x_1,x_2,x_3,x_4)$.
\end{Exercises}
A $k$-dissection of a convex $(n+k+1)$-gon with the maximal number of
diagonals (which is equal to $n(2k-1)-1$) is called {\it $k$-triangulation}.
It is well-known that the number of $k$-triangulations of a convex
$(n+k+1)$-gon is equal to the Catalan--Hankel number $C_{n-1}^{(k)}$.
Explicit bijection between the set of $k$-triangulations of a convex
$(n+k+1)$-gon and the set of $k$-tuple of non-crossing Dick paths
$(\gamma_1, \ldots,\gamma_k)$ such that the Dick path~$\gamma_i$ connects
points $(i-1,0)$ and $(2n-i-1,0)$, has been constructed in~\cite{SSt,St}.
\subsubsection[Grothendieck polynomials and $q$-Schr\"oder polynomials]{Grothendieck polynomials and $\boldsymbol{q}$-Schr\"oder polynomials}\label{section5.2.3}
Let $\pi_{k}^{(n)}=1^{k} \times w_{0}^{(n-k)} \in \mathbb{S}_n$ be the
vexillary permutation as before, see Theorem~\ref{theorem5.3}. Recall that
\begin{gather*}
\pi_{k}^{(n)}= \begin{pmatrix} 1& 2& \ldots& k& k+1& k+2&\ldots& n \\
1& 2& \ldots& k& n& {n-1}& \ldots& k+1 \end{pmatrix}.
\end{gather*}
$({\bf A})$ {\bf Principal specialization of the Schubert polynomial
$\boldsymbol{\mathfrak{S}_{\pi_{k}^{(n)}}}$.}
Note that $\pi_{k}^{(n)}$ is a~vexillary permutation of the staircase shape
$\lambda =(n-k-1,\ldots,2,1)$ and has the staircase f\/lag $\phi=(k+1,k+2,\ldots,n-1)$. It is known, see, e.g., \cite{M,Wa}, that for a vexillary
permutation $w \in \mathbb{S}_n$ of the shape $\lambda$ and f\/lag $\phi=(\phi_1,
\ldots,\phi_r)$, $r= \ell(\lambda)$, the corresponding Schubert polynomial
$\mathfrak{S}_{w}(X_n)$ is equal to the multi-Schur polynomial
$s_{\lambda}(X_{\phi})$, where $X_{\phi}$ denotes the f\/lagged set of variables, namely, $X_{\phi}=(X_{\phi_{1}}, \ldots,X_{\phi_{r}})$ and $X_m=
(x_1,\ldots,x_m)$. Therefore we can write the following determinantal formula
for the principal specialization of the Schubert polynomial corresponding to
the vexillary permutation~$\pi_{k}^{(n)}$
\begin{gather*}
\mathfrak{S}_{\pi_{k}^{(n)}}\big(1,q,q^2,\ldots\big)= \operatorname{DET} \left({n-i+j-1 \brack k+i-1 }_{q} \right)_{1 \le i,j \le n-k},
\end{gather*}
where ${n \brack k }_{q}$ denotes the $q$-binomial
coef\/f\/icient.
Let us observe that the Carlitz--Riordan $q$-analogue $C_n(q)$ of the Catalan
number $C_n$ is equal to the value of the $q$-Schr\"oder polynomial at
$\beta=0$, namely, $C_n(q)=S_n(q,0)$.
\begin{Lemma} \label{lemma5.2}
Let $k$, $n$ be integers and $n > k$, then
\begin{alignat*}{3}
& (1) \quad && \operatorname{DET} \left({n-i+j-1 \brack k+i-1}_{q} \right)_{1 \le i,j \le n-k} = q^{{n-k \choose 3}} C_{n}^{(k)}(q),&\\
& (2) \quad && \operatorname{DET} \big( C_{n+k-i-j}(q) \big)_{1 \le i,j \le k} =
q^{k(k-1)(6n-2k-5
)/6}~C_{n}^{(k)}(q).&
\end{alignat*}
\end{Lemma}
$({\bf{B}})$ {\bf Principal specialization of the Grothendieck
polynomial $\boldsymbol{\mathfrak{G}_{\pi_{k}^{(n)}}^{(\beta)}}$.}
\begin{Theorem}\label{theorem5.7}
\begin{gather*}
q^{{n-k+1 \choose 3}-(k-1) {n-k \choose 2}}
\operatorname{DET}\big|S_{n+k-i-j}\big(q; q^{i-1} \beta\big) \big|_{1 \le i,j \le k}\\
\qquad {}=
q^{k(k-1)(4k+1)/6} \prod_{a=1}^{k-1}\big(1+q^{a-1} \beta\big) \mathfrak{G}_{\pi_{k}^{(n)}}\big(1,q,q^2,\ldots\big).
\end{gather*}
\end{Theorem}
\begin{Corollary}\label{corollary5.2}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] If $k=n-1$, then
\begin{gather*}
\operatorname{DET} |S_{2n-1-i-j}\big(q; q^{i-1} \beta\big) |_{1 \le i,j \le n-1}=q^{(n-1)(n-2)(4n-3)/6} \prod_{a=1}^{n-2}\big(1+q^{a-1} \beta\big)^{n-a-1},
\end{gather*}
\item[$(2)$] If $k=n-2$, then
\begin{gather*}
q^{n-2} \operatorname{DET} \big| S_{2n-2-i-j}\big(q; q^{i-1} \beta\big) \big|_{1 \le i,j \le n-2}\\
\qquad{}=
q^{(n-2)(n-3)(4n-7)/6} \prod_{a=1}^{n-3}\big(1+q^{a-1} \beta\big)^{n-a-2}
\left\{{(1+ \beta)^{n-1}-1 \over \beta} \right\}.
\end{gather*}
\end{enumerate}
\end{Corollary}
{\bf Generalization.}
Let ${\boldsymbol{n}}=(n_1,\ldots,n_p) \in \mathbb{N}^{p}$ be a composition of $n$ so that $n=n_1+ \cdots + n_p$. We set $n^{(j)}=n_1+ \cdots+n_j$, $j=1,\ldots,p$, $n^{(0)}=0$.
Now consider the permutation $w^{({\boldsymbol{n}})}=w_{0}^{(n_1)} \times w_{0}^{(n_2)} \times \cdots \times w_{0}^{(n_p)} \in \mathbb{S}_n$,
where $w_{0}^{(m)} \in \mathbb{S}_m$ denotes the longest permutation in the
symmetric group $\mathbb{S}_m$. In other words,
\begin{gather*}
w^{({\boldsymbol{n}})} = \begin{pmatrix} 1& 2& \ldots& n_1& n^{(2)}& \ldots &
n_1+1& \ldots& n^{(p-1)}& \ldots n \\
n_{1}& {n_1-1}& \ldots& 1& {n_1+1}& \ldots & n^{(2)}& \ldots& n& \ldots
n^{(p-1)+1} \end{pmatrix}.
\end{gather*}
For the permutation $w^{({\boldsymbol{n}})}$ def\/ined above, one has the following
factorization formula for the Grothendieck polynomial corresponding to
$w^{({\boldsymbol{n}})} $~\cite{M}
\begin{gather*}
\mathfrak{G}_{w^{({\boldsymbol{n}})}}^{(\beta)} =\mathfrak{G}_{w_{0}^{(n_1)}}^{(\beta)} \times
\mathfrak{G}_{1^{n_1} \times w_{0}^{(n_2)}}^{(\beta)} \times
\mathfrak{G}_{1^{n_1+n_2} \times w_{0}^{(n_3)}}^{(\beta)} \times \cdots \times
\mathfrak{G}_{1^{n_{1}+ \ldots n_{p-1}} \times w_{0}^{(n_{p})}}^{(\beta)}.
\end{gather*}
In particular, if
\begin{gather}\label{equation5.11}
w^{({\boldsymbol{n}})}= w_{0}^{(n_1)} \times w_{0}^{(n_2)} \times \cdots \times w_{0}^{(n_p)} \in \mathbb{S}_n,
\end{gather}
then the principal specialization $\mathfrak{G}_{w^{({\boldsymbol{n}})}}^{(\beta)}$ of the Grothendieck polynomial corresponding to the permutation $w$, is the product
of $q$-Schr\"oder--Hankel polynomials. Finally, we observe that from
discussions in Section~\ref{section5.2.1}(3), {\it Grothendieck and Narayana polynomials}, one can deduce that
\begin{gather*}
\mathfrak{G}_{w^{({\boldsymbol{n}})}}^{(\beta-1)}(x_1=1,\ldots,x_n=1)= \prod_{j=1}^{p-1} \mathfrak{N}_{n^{(j+1)}}^{(n^{(j)})} (\beta).
\end{gather*}
In particular, the polynomial $\mathfrak{G}_{w^{({\boldsymbol{n}})}}^{(\beta-1)}(x_1, \ldots,x_n)$ is a symmetric polynomial in $\beta$ with non-negative integer coef\/f\/icients.
\begin{Example}\label{example5.6}\quad
$(1)$ Let us take (non vexillary) permutation $w=2143=s_1 s_3$.
One can check that
\begin{gather*}
\mathfrak{G}_{w}^{(\beta)}(1,1,1,1)=3+3 \beta+\beta^2=1+(\beta+1)+(\beta+1)^2,
\end{gather*}
and
\begin{gather*}
\mathfrak{N}_4(\beta)=(1,6,6,1) , \qquad \mathfrak{N}_3(\beta)=(1,3,1), \qquad \mathfrak{N}_2(\beta)=(1,1).
\end{gather*}
It is easy to see that
\begin{gather*}
\beta \mathfrak{G}_{w}^{(\beta)}(1,1,1,1)= \operatorname{DET} \left | \begin{matrix}
\mathfrak{N}_4(\beta) & \mathfrak{N}_3(\beta) \\
\mathfrak{N}_3(\beta) & \mathfrak{N}_2(\beta)
\end{matrix} \right |.
\end{gather*}
On the other hand,
\begin{gather*}
\operatorname{DET} \left | \begin{matrix}
P_4(\beta) & P_3(\beta) \\
P_3(\beta) & P_2(\beta)
\end{matrix} \right | = (3,6,4,1)=\big(3+3 \beta+ \beta^2\big) (1+\beta).
\end{gather*}
It is more involved to check that
\begin{gather*}
q^5 (1+\beta)~\mathfrak{G}_{w}^{(\beta)}\big(1,q,q^2,q^3\big)= \operatorname{DET} \left |
\begin{matrix}
S_4(q;\beta) & S_3(q;\beta) \\
S_3(q; q \beta) & S_2(q; q \beta)
\end{matrix} \right | .
\end{gather*}
$(2)$~Let us illustrate Theorem~\ref{theorem5.7} by a few examples. For the sake of
simplicity, we consider the case $\beta=0$, i.e., the case of {\it Schubert polynomials}. In this case $P_n(q;\beta=0)=C_n(q)$ is equal to the Carlitz--Riordan $q$-analogue of Catalan numbers. We are reminded that the $q$-Catalan--Hankel polynomials are def\/ined as follows
\begin{gather*}
C_n^{(k)}(q)= q^{k(1-k)(4k-1)/6} \operatorname{DET} | C_{n+k-i-j}(q) |_{1 \le i,j \le n} .
\end{gather*}
In the case $\beta =0$ the Theorem~\ref{theorem5.7} states that if ${\boldsymbol{n}} =(n_1, \ldots,
n_p) \in \mathbb{N}^{p}$ and the permutation $w_{({\boldsymbol{n}})} \in \mathbb{S}_n$ is def\/ined
by the use of~\eqref{equation5.10}, then
\begin{gather*}
\mathfrak{S}_{w^{({\boldsymbol{n}})}}\big(1,q,q^2,\ldots\big)
=q^{\sum {n_i \choose 3}} C_{n_{1}+n_{2}}^{(n_1)}(q) \times C_{n_{1}+n_{2}+n_{3}}^{(n_1+n_2)}(q) \times C_{n}^{(n-n_{p})}(q).
\end{gather*}
Now let us consider a few examples for $n=6$.
\begin{itemize}\itemsep=0pt
\item ${\boldsymbol{n}}=(1,5)$ $\Longrightarrow$ $\mathfrak{S}_{w^{({\boldsymbol{n}})}}(1,q,\ldots)= q^{10} C_{6}^{(1)}(q)=C_{5}(q)$.
\item ${\boldsymbol{n}}=(2,4)$ $\Longrightarrow$ $\mathfrak{S}_{w^{({\boldsymbol{n}})}}(1,q, \ldots)=q^{4} C_{6}^{(2)}(q) = \operatorname{DET} \left | \begin{matrix}
C_{6}(q) & C_{5}(q) \\
C_5(q) & C_4(q)
\end{matrix} \right |$.
\end{itemize}
Note that $\mathfrak{S}_{w^{(2,4)}}(1,q,\ldots)=\mathfrak{S}_{w^{(1,1,4)}}(1,q,\ldots)$.
\begin{itemize}\itemsep=0pt
\item ${\boldsymbol{n}}=(2,2,2)$ $\Longrightarrow$
$\mathfrak{S}_{w^{({\boldsymbol{n}})}}(1,q,\ldots)=C_{4}^{(2)}(q) C_{6}^{(4)}(q)$.
\item ${\boldsymbol{n}}=(1,1,4)$ $\Longrightarrow$ $\mathfrak{S}_{w^{({\boldsymbol{n}})}}(1,q,\ldots)= q^4 C_{2}^{(1)}(q) C_{4}^{(2)}(q)=q^4 C_{4}^{(2)}(q)$,
the last equality follows from that $C_{k+1}^{(k)}(q)=1$ for all $k \ge 1$.
\item ${\boldsymbol{n}}=(1,2,3)$ $\Longrightarrow$ $\mathfrak{S}_{w^{({\boldsymbol{n}})}}(1,q,\ldots)=q C_{3}^{(1)}(q) C_{6}^{(3)}(q)$.
\item ${\boldsymbol{n}}=(3,2,1)$ $\Longrightarrow$ $\mathfrak{S}_{w^{({\boldsymbol{n}})}}(1,q,\ldots)=q C_{5}^{(3)}(q) C_{6}^{(5)}(q)=q C_{5}^{(3)}(q)=q(1,1,1,1)$.
Note that $C_{k+2}^{(k)}(q)= {k+1 \brack 1 }_{q}$.
\end{itemize}
\end{Example}
\begin{Exercises}\label{exercises5.4} Let $1 \le k \le m \le n$ be integers, $n \ge 2k+1$. Consider permutation
\begin{gather*}
w = \begin{pmatrix} 1& 2& \ldots& k& {k+1}& \ldots& &n \\
m& {m-1}& \ldots& {m-k+1}& n& \ldots& \ldots& 1 \end{pmatrix} \in
\mathbb{S}_n.
\end{gather*}
Show that
\begin{gather*}
\mathfrak{S}_{w}(1,q,\ldots)= q^{n(D(w))} C_{n-m+k}^{(m)}(q),
\end{gather*}
where for any permutation $w$, $n(D(w))= \sum {d_i(w) \choose 2}$ and $d_i(w)$
denotes the number of boxes in the $i$-th column of the ({\it Rothe}) diagram~$D(w)$ of the permutation~$w$, see \cite[p.~8]{M}.
\end{Exercises}
({\bf C}) {\bf A determinantal formula for the Grothendieck polynomials
$\boldsymbol{\mathfrak{G}_{\pi_{k}^{(n)}}^{(\beta)}}$.}
Def\/ine polynomials
\begin{gather*}
\Phi_{n}^{(m)}(X_n)=\sum_{a=m}^{n}e_a(X_n) \beta^{a-m},\\
A_{i,j}(X_{n+k-1})={1 \over (i-j)!} \left({\partial \over \partial \beta}
\right)^{j-1} \Phi_{k+n-i}^{(n+1-i)}(X_{k+n-i}) \qquad \text{if}\quad 1 \le i \le j \le n,
\end{gather*}
and
\begin{gather*}
A_{i,j}(X_{k+n-1})=\sum_{a=0}^{i-j-1} e_{n-i-a}(X_{n+k-i})~{i-j-1 \choose a} \qquad \text{if} \quad 1 \le j < i \le n.
\end{gather*}
\begin{Theorem}\label{theorem5.8}
\begin{gather*}
\operatorname{DET} |A_{i,j}|_{1 \le i,j \le n}=\mathfrak{G}_{\pi_{k+n}^{(k)}}^{(\beta)}(X_{k+n-1}).
\end{gather*}
\end{Theorem}
\begin{Comments} \label{comments5.5}\quad
$(a)$~One can compute the Grothendieck polynomials for yet
another interesting family of permutations. namely, {\it grassmannian}
permutations
\begin{gather*}
\sigma_{k}^{(n)}= \begin{pmatrix} 1& 2& \ldots& {k-1}& k& {k+1}& {k+2}& \ldots& n+k \\
1& 2& \ldots& {k-1}& {n+k}& {k}& {k+1}& \ldots & n+k-1 \end{pmatrix}\\
\hphantom{\sigma_{k}^{(n)}}{} =
s_k s_{k+1} \cdots s_{n+k-1} \in \mathbb{S}_{n+k}.
\end{gather*}
Then
\begin{gather*}
\mathfrak{G}_{{\sigma_{k}}^{(n)}}^{(\beta)}(x_1,\ldots,x_{n+k})=
\sum_{j=0}^{k-1} s_{(n,1^{j})}(X_k) \beta^{j},
\end{gather*}
where $s_{(n,1^{j})}(X_k)$ denotes the Schur polynomial corresponding to the
hook shape partition $(n,1^{j})$ and the set of variables
$X_k:=(x_1,\ldots,x_k)$. In particular,
\begin{gather*}
\mathfrak{G}_{{\sigma_{k}}^{(n)}}^{(\beta)}(x_j=1,\, \forall\, j)=
{n+k-1 \choose k} \left(\sum_{j=0}^{k-1} {k \over n+j} {k-1 \choose j}
\beta^{j} \right) = \sum_{j=0}^{k-1} {n+j-1 \choose j} (1+\beta)^{j}.
\end{gather*}
$(b)$ {\it Grothendieck polynomials for grassmannian permutations.}
In the case of a {\it grassmannian} permutation
$w:=\sigma_{\lambda} \in {\mathbb{S}}_{\infty}$ of the shape
$\lambda=(\lambda_1 \ge \lambda_2 \ge \cdots \ge \lambda_n )$ where $n$
is a unique descent of~$w$, one can prove the following
formulas for the $\beta$-Grothendieck polynomial
\begin{gather}\label{equation5.12}
\mathfrak{G}_{\sigma_{\lambda}}^{(\beta)}(X_n)= {\operatorname{DET} \big| x_i^{\lambda_j +n-j}
(1+\beta x_i)^{j-1} \big|_{1 \le i,j \le n} \over \prod\limits_{1 \le i < j \le n}
(x_i-x_j) },
\\
\operatorname{DET}\big| h_{\lambda_j +i,j}^{(\beta)}(X_{[i,n]})\big|_{1 \le i,j \le n} =
\operatorname{DET}\big|h_{\lambda_j +i,j}^{(\beta)}(X_n)\big|_{1 \le i,j \le n},\nonumbe
\end{gather}
where $X_{[i,n]}=(x_i,x_{i+1},\ldots,x_n)$, and for any set of variables~$X$
\begin{gather*}
h_{n,k}^{(\beta)}(X) = \sum_{a=0}^{k-1} {k-1 \choose a} h_{n-k+a}(X)
\beta^{a},
\end{gather*}
and $h_k(X)$ denotes the complete symmetric polynomial of degree~$k$ in the variables from the set~$X$.
A proof is a straightforward adaptation of the proof of special case $\beta =0$ (the case of {\it Schur} polynomials) given by I.~Macdonald~\cite[Section~2, equation~(2.10) and Section~4, equation~(4.9)]{M}.
Indeed, consider $\beta$-divided dif\/ference operators $\pi_{j}^{(\beta)}$,
$j=1,\ldots,n-1$, and $\pi_{w}^{(\beta)}$, $w \in {\mathbb{S}}_n$,
introduced in~\cite{FK1}. For example,
\begin{gather*}
\pi_{j}^{(\beta)}(f) = {1 \over x_j-x_{j+1}}\big((1+\beta x_{j+1}) f(X_n)-
(1+\beta x_j) f(s_{j}(X_n) \big).
\end{gather*}
Now let $w_{0}:= w_{0}^{(n)}$ be the longest element in the symmetric group
${\mathbb{S}}_n$. The same proves of the State\-ments~2.10,~2.16 from~\cite{M} show that{\samepage
\begin{gather*}
\pi_{w_{0}}^{(\beta)}= a_{\delta}^{-1} w_{0} \left( \sum_{\sigma \in {\mathbb{S}}_n}
(-1)^{\ell(\sigma)} \prod _{j=1}^{n-1} (1+\beta x_{j})^{n-j} \sigma \right),
\end{gather*}
where $a_{\delta} = \prod\limits_{1 \le i < j \le n} (x_i - x_j)$.}
On the other hand, the same arguments as in the proof of Statement~4.8
from~\cite{M} show that
\begin{gather*}
\mathfrak{G}_{\sigma_{\lambda}}^{(\beta)}(X_n) = \pi_{w^{(0)}}^{(\beta)}
\big(x^{\lambda+\delta_{n}}\big).
\end{gather*}
Application of the formula for operator $\pi_{w_{n}^{(0)}}^{(\beta)}$
displayed above to the monomial $x^{\lambda +\delta_{n}}$ f\/inishes the proof
of the f\/irst equality in~\eqref{equation5.11}. The statement that the right hand side of
the equality~\eqref{equation5.12} coincides with determinants displayed in the identity~\eqref{equation5.12} can be checked by means of simple transformations.
\end{Comments}
\begin{Problems}\label{problems5.1}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] Give a bijective prove of Theorem~{\rm \ref{theorem5.5}}, i.e., construct a
bijection between
\begin{itemize}\itemsep=0pt
\item the set of $k$-tuple of mutually non-crossing
Schr\"oder paths $(\mathfrak{p}_1,\ldots,\mathfrak{p}_k)$ of lengths
$(n,n-1,\ldots,n-k+1)$ correspondingly, and
\item the set of pairs $(\mathfrak{m},{\cal{T}})$, where $\cal{T}$ is a $k$-dissection of a convex $(n+k+1)$-gon, and~$\mathfrak{m}$ is a upper triangle
$(0,1)$-matrix of size $(k-1) \times (k-1)$,
which is compatible with natural statistics on the both sets.
\end{itemize}
\item[$(2)$] Let $w \in \mathbb{S}_n$ be a permutation, and~${\rm CS}(w)$ be the set of
compatible sequences corresponding to~$w$, see, e.g.,~{\rm \cite{BJS}}.
Define statistics $c(\bullet)$ on the set ${\rm CS}(w)$ such that
\begin{gather*}
\mathfrak{G}_{w}^{(\beta-1)}(x_1=1,x_2=1,\ldots)= \sum_{a \in {\rm CS}(w)} \beta^{c(a)}.
\end{gather*}
\item[$(3)$] Let $w$ be a vexillary permutation.
Find a determinantal formula for the $\beta$-Grothendieck
polynomial ${\mathfrak G}_{w}^{(\beta)}(X)$.
\item[$(4)$] Let $w$ be a permutation.
Find a geometric interpretation of coefficients of the
polynomials ${\mathfrak{S}}_{w}^{(\beta)}(x_i=1)$ and ${\mathfrak{S}}_{w}^{(\beta)}(x_i=q, \,x_j=1,\,
\forall\, j \not= i)$.
\end{enumerate}
\end{Problems}
For example, let $w \in {\mathbb S}_n$ be an involution, i.e., $w^2=1$,
and $w' \in {\mathbb S}_{n+1}$ be the image of $w$ under the natural
embedding ${\mathbb S}_n \hookrightarrow {\mathbb S}_{n+1}$ given by
$ w \in {\mathbb S}_n \longrightarrow (w,n+1) \in {\mathbb S}_{n+1}$.
It is well-known, see, e.g., \cite{KMi, W}, that the multiplicity
$m_{e,w}$ of the $0$-dimensional Schubert cell $\{pt \}=Y_{w_{0}^{(n+1)}}$ in
the Schubert variety $\overline{Y}_{w'}$ is equal to the specialization
${\mathfrak{S}}_{w}(x_i=1)$ of the Schubert polynomial ${\mathfrak{S}}_{w}(X_{n})$. Therefore one can
consider the polynomial ${\mathfrak{S}}_{w}^{(\beta)}(x_i=1)$ as a $\beta$-deformation of
the multiplicity~$m_{e,w}$.
\begin{Question}\label{question-page105}
What is a geometrical meaning of the coefficients
of the polynomial \linebreak ${\mathfrak{S}}_{w}^{(\beta)}(x_i=1) \in \mathbb{N} [\beta]$?
\end{Question}
\begin{Conjecture} \label{conjecture5.1}
The polynomial ${\mathfrak{S}}_{w}^{(\beta)}(x_i=1)$ is a unimodal
polynomial for any permutation~$w$.
\end{Conjecture}
\subsubsection{Specialization of Schubert polynomials}\label{section5.2.4}
Let $n$, $k$, $r$ be positive integers and $p$, $b$ be
non-negative integers such that $r \le p+1$. It is well-known~\cite{M} that
in this case there exists a unique {\it vexillary} permutation
$\varpi:= \varpi_{\lambda,\phi} \in \mathbb{S}_{\infty}$ which has the
{\it shape} $\lambda=(\lambda_1,\ldots,\lambda_{n+1})$ and the
{\it flag} $\phi =(\phi_1,\ldots,\phi_{n+1})$, where
\begin{gather*}
\lambda_i= (n-i+1) p+b, \qquad \phi_i=k+1+r (i-1), \qquad 1 \le i \le n+1-
\delta_{b,0}.
\end{gather*}
According to a theorem by M.~Wachs~\cite{Wa}, the Schubert polynomial $\mathfrak{S}_{\varpi}(X)$ admits the following determinantal representation
\begin{gather*}
\mathfrak{S}_{\varpi}(X)= \operatorname{DET} \big( h_{\lambda_i-i+j}(X_{\phi_{i}}) \big)_{1 \le i,j \le n+1}.
\end{gather*}
Therefore we have
\begin{gather*}
\mathfrak{S}_{\varpi}(1):= \mathfrak{S}_{\varpi}(x_1=1,x_2=1,\ldots) \\
\hphantom{\mathfrak{S}_{\varpi}(1)}{} =\operatorname{DET} \left( {(n-i+1)p+b-i+j+k+(i-1)r \choose k+ (i-1)r} \right) _{1 \le i,j \le n+1}.
\end{gather*}
We denote the above determinant by $D(n,k,r,b,p)$.
\begin{Theorem}\label{theorem5.9}
\begin{gather*}
D(n,k,r,b,p)= \!\!
\prod_{(i,j) \in {\cal{A}}_{n,k,r}}\!\!\! {i+b+jp \over i}\!\!
\prod_{(i,j) \in {\cal{B}}_{n,k,r}}\!\!\!
{(k-i+1)(p+1)+(i+j-1)r+r(b+np) \over k-i+1+(i+j-1)r},
\end{gather*}
where
\begin{gather*}
{\cal{A}}_{n,k,r}= \big\{ (i,j) \in \mathbb{Z}_{\ge 0}^2 \,|\, j \le n, j < i \le k+(r-1)(n-j) \big\},\\
{\cal{B}}_{n,k,r}= \big\{ (i,j) \in \mathbb{Z}_{\ge 1}^2 \,|\, i+j \le n+1, \, i \not= k+1+r s,\, s\in \mathbb{Z}_{\ge 0} \bigr\}.
\end{gather*}
\end{Theorem}
It is convenient to re-write the above formula for $D(n,k,r,b,p)$ in the
following form
\begin{gather*}
D(n,k,r,b,p)= \prod_{j=1}^{n+1} {((n-j+1)p +b+k+(j-1)(r-1) )! (n-j+1)! \over
(k+(j-1)r )!((n-j+1)(p+1)+b )!}\\
\hphantom{D(n,k,r,b,p)=}{} \times
\prod_{1 \le i \le j \le n} ((k-i+1)(p+1)+jr+(np+b)r).
\end{gather*}
\begin{Corollary}[some special cases] \label{corollary5.3} $(A)$ The case $r=1$.
We consider below some special cases of Theorem~{\rm \ref{theorem5.9}} in the case
$r=1$. To simplify notation, we set $D(n,k,b,p):= D(n,k,r=1,b,p)$. Then we can rewrite the
above formula for $D(n,k,r,b,p)$ as follows
\begin{gather*}
D(n,k,b,p)=
\prod_{j=1}^{n+1} {((n+k-j+1)(p+1)+b ) ! ((n-j+1)p+b+k ) ! (j-1) ! \over ((n-j+1)(p+1)+b )! ((k+n-j+1)p+b+k)! (k+j-1)! }.
\end{gather*}
$(1)$ If $ k \le n+1$, then
\begin{gather*}
D(n,k,b,p)=
\prod_{j=1}^{k} {(n+k+1-j)(p+1)+b \choose n-j+1} {(k-j)p+b+k \choose j}
{j! (k-j)! (n-j+1)! \over (n+k-j+1)!}.
\end{gather*}
In particular,
\begin{itemize}\itemsep=0pt
\item if $k=1$, then
\begin{gather*}
D(n,1,b,p)={1+b \over 1+b+(n+1)p}~{(p+1)(n+1)+b \choose n+1}:=F_{n+1}^{(p+1)}(b),
\end{gather*}
where $F_{n}^{p}(b):={1+b \over 1+b+(p-1)n} {pn+b \choose n}$ denotes the generalized Fuss--Catalan number,
\item if $k=2$, then
\begin{gather*}
D(n,2,b,p)={(2+b)(2+b+p) \over (1+b)(2+b+(n+1)p)(2+b+(n+2)p)} F_{n+1}^{(p+1)}(b) F_{n+2}^{(p+1)}(b),
\end{gather*}
in particular,
\begin{gather*}
D(n,2,0,1)= \frac{6}{(n+3)(n+4)} {\rm Cat}_{n+1} {\rm Cat}_{n+2}.
\end{gather*}
See {\rm \cite[$A005700$]{SL}} for several combinatorial interpretations of these
numbers.
\end{itemize}
$(2)$ Consider the Young diagram $($see R.A.~Proctor~{\rm \cite{Pr3})}
\begin{gather*}
\lambda:=\lambda_{n,p,b}= \big\{ (i,j) \in \mathbb{Z}_{\ge 1} \times \mathbb{Z}_{\ge 1} \,|\, 1 \le i \le n+1, \,1 \le j \le (n+1-i)p+b \}.
\end{gather*}
For each box $(i,j) \in \lambda$ define the numbers $c(i,j):=n+1-i+j$, and
\begin{gather*}
l_{(i,j)}(k)= \begin{cases}\dfrac{k+c(p,j)}{c(i,j)} & \text{if} \ \ j \le (n+1-i)(p-1)+b, \vspace{1mm}\\
\dfrac{(p+1)k+c(i,j)}{c(i,j)} & \text{if} \ \ (n+1-i)(p-1) < j-b \le (n+1-i)p. \end{cases}
\end{gather*}
Then
\begin{gather}\label{equation5.14}
D(n,k,b,p)=\prod_{(i,j) \in \lambda} l_{(i,j)}(k).
\end{gather}
Therefore, $D(n,k,b,p)$ is a polynomial in $k$ with rational coefficients.
$(3)$ If $p=0$, then
\begin{gather*}
D(n,k,b,0)=\dim V_{(n+1)^{k}}^{\mathfrak{gl}(b+k)}=\prod_{j=1}^{n+k}
\left({j+b \over j}\right)^{\min(j,n+k+1-j)},
\end{gather*}
where for any partition $\mu$, $\ell(\mu) \le m$, $V_{\mu}^{\mathfrak{gl}(m)}$ denotes the irreducible $\mathfrak{gl}(m)$-module with the highest weight~$\mu$.
In particular,
\begin{gather*}
D(n,2,b,0)= {1 \over n+2+b}{n+2+b \choose b} {n+2+b \choose b+1}
\end{gather*}
is equal to the Narayana number $N(n+b+2,b)$,
\begin{gather*}
D(1,k,b,0) = {(b+k)! (b+k+1)! \over k! b! (k+1)! (b+1)!}:=
N(b+k+1,k),
\end{gather*}
and therefore the number $D(1,k,b,0)$ counts the number of pairs of
non-crossing lattice paths inside a~rectangular of size $(b+1) \times (k+1)$, which go from the point $(1,0)$
$($resp.\ from that $(0,1))$ to the point $(b+1,k)$ $($resp.\ to that $(b,k+1))$,
consisting of steps $U=(1,0)$ and $R=(0,1)$, see~{\rm \cite[$A001263$]{SL}}, for some
list of combinatorial interpretations of the Narayana numbers.
$(4)$ If $p=b=1$, then
\begin{gather*}
D(n,k,1,1)= C^{(k)}_{n+k+1}:= \prod_{1 \le i \le j \le n+1} {2k+i+j \over i+j} .
\end{gather*}
$(5)$ If $p=1$ and $b$ is odd integer, then $D(n,k,b,1)$ is
equal to the dimension of the irreducible representation of the symplectic Lie
algebra ${\rm Sp}(b+2n+1)$ with the highest weight $k\omega_{n+1}$
$($R.A.~Proctor~{\rm \cite{Pr1,Pr2})}.
$(6)$ If $p=1$ and $b=0$, then
\begin{gather*}
D(n,k,1,0) = D(n-1,k,1,1) = \prod_{ 1 \le i \le j \le n} {2k +i +j
\over i+j} =C_{n+k}^{(k)},
\end{gather*}
see section on Grothendieck and Narayana polynomials.
$(7)$ Let $\varpi_{\lambda}$ be a unique dominant
permutation of shape $\lambda:= \lambda_{n,p,b}$ and $\ell:= \ell_{n,p,b} = {1 \over 2}(n+1)(np+2b)$ be its length $($cf.~{\rm \cite{FK3})}. Then
\begin{gather*}
\sum_{{\boldsymbol{a}} \in R(\varpi_{\lambda})} \prod_{i=1}^{\ell} (x+a_i) =
\ell ! B(n,x,p,b).
\end{gather*}
Here for any permutation~$w$ of length~$l$, we denote by~$R(w)$ the set
$\{ {\boldsymbol{a}}=(a_1,\ldots,a_l) \}$ of all reduced decompositions of~$w$.
\end{Corollary}
\begin{Exercises}\label{exercises5.5}
Show that
\begin{gather*}
\operatorname{DET} \big | F_{n+i+j-2}^{(2)}(0) \big |_{1 \le i,j \le k} =
\prod_{j=1}^{k} F_{n+j-1}^{(2)}(0) \frac{{k+1 \choose 2} !}{ \prod\limits_{1 \le i
\le k-1 \atop 1 \le j \le k}(n+i+j)},\\
D(n,k,b,1)= \prod_{j=1}^{k} F_{n+j}^{(2)}(b) \frac{\prod\limits_{1 \le i \le j \le k}(b+i+j-1)}{\prod\limits_{1 \le i \le k-1 \atop 1 \le j \le k}(n+b+i+j+1)}.
\end{gather*}
Clearly that if $b=0$, then $F_{n}^{(2)}(0)=C_n$, and $D(n,k,0,1)$ is equal to
the Catalan--Hankel determinant~$C_n^{(k)}$.
Finally we recall that the generalized
Fuss--Catalan number $F_{n+1}^{(p+1)}(b)$ counts the number of lattice paths
from $(0,0)$ to $(b+np,n)$ that do not go above the line $x=py$, see,
e.g.,~\cite{Kr}.
\end{Exercises}
\begin{Comments}\label{comments5.6} It is well-known, see, e.g., \cite{Pr3} or \cite[Vol.~2, Exercise~7.101.b]{ST1}, that the number $D(n,k,b,p)$ is equal
to the total number $pp^{\lambda_{n,p,b}}(k)$ of plane
partitions\footnote{Let $\lambda$ be a partition. A plane (ordinary) partition bounded
by $d$ and shape $\lambda$ is a f\/illing of the shape~$\lambda$ by the numbers
from the set $\{0,1,\ldots,d \}$ in such a way that the numbers along
columns and rows are weakly {\it decreasing}.
A~{\it reverse} plane partition bounded by~$d$ and shape~$\lambda$ is a~f\/illing of the shape~$\lambda$ by the numbers
from the set $\{0,1,\ldots,d \}$ in such a way that the numbers along columns
and rows are weakly {\it increasing}.}
bounded by $k$ and contained in the shape $\lambda_{n,b,p}$.
More generally, see, e.g.,~\cite{FK3}, for any partition $\lambda$ denote by
$w_{\lambda} \in \mathfrak{S}_{\infty}$ a unique {\it dominant} permutation of
shape~$\lambda$, that is a unique permutation with the code $c(w)=\lambda$.
Now for any non-negative integer $k$ consider the so-called {\it shifted
dominant} permutation~$w_{\lambda}^{(k)}$ which has the shape $\lambda$ and the
f\/lag $\phi=(\phi_i=k+i-1,\, i=1,\ldots,\ell(\lambda))$. Then
\begin{gather*}
{\mathfrak{S}}_{w_{\lambda}^{(k)}}(1) = pp^{\lambda}( \le k),
\end{gather*}
where $pp^{\lambda}(\le k)$ denotes the number of all plane partitions bounded by~$k$ and contained in~$\lambda$. Moreover,
\begin{gather*}
\sum_{\pi \in PP^{\lambda}(\le k)} q^{|\pi|} = q^{n(\lambda)}
{\mathfrak{S}}_{w_{\lambda}^{(k)}}\big(1,q^{-1},q^{-2},\ldots\big),
\end{gather*}
where $PP^{\lambda}(\le k)$ denotes the set of all plane partitions bounded by~$k$ and contained in~$\lambda$.
\end{Comments}
\begin{Exercises}\label{exercises5.6}\quad
$(1)$ Show that
\begin{gather*} \lim _{k \rightarrow \infty}~{\mathfrak{S}}_{w_{\lambda}^{(k)}}\big(1,q,q^2,\ldots\big) =
\frac{q^{n(\lambda)}}{H_{\lambda}(q)},
\end{gather*}
where $H_{\lambda}(q) = \prod\limits_{x \in \lambda} (1-q^{h(x)})$
denotes the {\it hook} polynomial corresponding to a given partition~$\lambda$.
$(2)$ Let $\lambda = ((n+\ell)^{\ell}, \ell^{n})$ be a {\it fat hook}.
Show that
\begin{gather*}
\lim_{k \rightarrow \infty} q^{n(\lambda)} {\mathfrak{S}}_{w_{\lambda}^{(k)}}\big(1,q^{-1},q^{-2},\ldots\big)
= q^{s(\ell,n)} \frac{K_{\lambda}(q)}{M_{\ell}(2 n+2 \ell - 1 ;q)},
\end{gather*}
where $a(\ell,n)$ is a certain integer we don't need to specify in what
follows,
\begin{gather*}
M_{\ell}(N;q)= \prod_{j=1}^{N} \left(\frac{1}{1- q^{j}} \right)^{\min(j,N+1-j,\ell)}
\end{gather*}
denotes the MacMahon generating function for the number of plane partitions
f\/it inside the box $N \times N \times \ell$, $K_{\lambda}(q)$ is a
polynomial in~$q$ such that $K_{\lambda}(0)=1$.
$(a)$ Show that
\begin{gather*}
(1-q)^{|\lambda|} \frac{K_{\lambda}(q)}{M_{\ell}(2 n+2 \ell - 1 ;q)}
\bigg |_{q=1} = {1 \over \prod\limits_{x \in \lambda} h(x)}.
\end{gather*}
$(b)$ Show that
\begin{gather*}
K_{\lambda}(q) \in \mathbb{N} [q] \qquad \text{and}\qquad K_{\lambda}(1)= M(n,n,\ell),
\end{gather*}
where $M(a,b,c)$ denotes the number of plane partitions f\/it inside the box
$a \times b \times c$. It is well-known, see, e.g., \cite[p.~81]{M1}, that
\begin{gather*}
M(a,b,c) = \prod_{\substack{1 \le i \le a \\1 \le j \le b\\ 1 \le k \le c}}
\frac{i+j+k-1}{i+j+k-2} = \prod_{i=1}^{c}\frac{(a+b+i-1) ! (i-1) !}{(a+i-1) ! (b+1-1) !} =\dim V_{(a^c)}^{{\mathfrak{gl}}_{b+c}}.
\end{gather*}
Show that
\begin{gather*}
K_{\lambda}(q) = \sum_{\pi \in B_{n,n,\ell}} q^{wt_{\ell}(\pi)},
\end{gather*}
where the sum runs over the set of plane partitions $\pi=(\pi_{ij})_{1 \le i,j \le n}$ f\/it inside the box
$B_{n,n,\ell}:=n \times n \times \ell$, and
\begin{gather*}
wt_{\ell}(\pi)= \sum_{i,j} \pi_{ij} + \ell \sum_i \pi_{ii}.
\end{gather*}
$(c)$ Assume as before that $\lambda:=((n+\ell)^{\ell},\ell^n)$.
Show that
\begin{gather*}
\lim_{ n \rightarrow \infty} K_{\lambda}(q) = M_{\ell}(q) \sum_{\mu \atop \ell(\mu) \le \ell} q^{|\mu|} \left(\frac{q^{n(\mu)}}{\prod_{x \in \mu} (1-q^{h(x)})} \right)^2,
\end{gather*}
where the sum runs over the set of partitions $\mu$ with the number of parts
at most $\ell$, and $n(\mu)= \sum_{i} (i-1) \mu_i$,
\begin{gather*}
M_{\ell}(q) :=
\prod_{j \ge 1} \big(1-q^{j}\big)^{\min(j,\ell)}.
\end{gather*}
Therefore the generating function $PP^{(\ell,0)}(q):= \sum\limits_{\pi \in
PP^{(\ell,0)}} q^{|\pi|}$
is equal to
\begin{gather*}
\sum_{\mu \atop \ell(\mu) \le \ell} q^{|\mu|} \left(\frac{q^{n(\mu)}}{\prod\limits_{x \in \mu} (1-q^{h(x)})} \right)^2,
\end{gather*}
where $PP^{(\ell,k)}:= \{ \pi= (\pi_{ij})_{i,j \ge 1} \,|\, \pi_{ij}
\ge 0, \pi_{\ell+1,\ell+1} \le k \}$, $|\pi|= \sum\limits_{i,j} \pi_{ij}$.
$(d)$ Show that
\begin{gather*
PP^{(\ell,0)}(q) = \frac{1}{M_{\ell}(q)^2} \sum_{\mu, \atop \ell(\mu) \le \ell} (- q)^{|\mu|} q^{n(\mu)+n(\mu')} \big( \dim_{q} V_{\mu}^{{\mathfrak{gl}}(\ell)} \big)^2,
\end{gather*}
where $\mu'$ denotes the {\it conjugate} partition of $\mu$, therefore
$n(\mu') =\sum\limits_{i \ge 1} {\mu_i \choose 2}$.
The formula \eqref{equation5.14} is the special case $n=m$ of \cite[Theorem~1.2]{FM}.
In particular, if $\ell=1$ then one come to following identity
\begin{gather*}
\frac{1}{(q;q)_{\infty}^2} \sum_{k \ge 0} (-1)^{k} q^{{k+1 \choose 2}} =
\sum_{k \ge 0} q^{k} \left(\frac{1}{(q;q)_{k}} \right)^2.
\end{gather*}
$(e)$ Let $k \ge 0$, $\ell \ge 1$ be integers.
Show that the ({\it fermionic}) generating function for the
number of plane partitions $\pi = (\pi_{ij}) \in PP^{(\ell,k)}$ is equal to
\begin{gather*}
\sum_{\pi \in PP^{(\ell,k)}} q^{|\pi|} = \sum_{\mu \atop \mu_{\ell+1} \le k} q^{|\mu|} \left(\frac{q^{n(\mu)}}{\prod\limits_{x \in \mu} (1-q^{h(x)})} \right)^2.
\end{gather*}
\end{Exercises}
$(B)$ {\it The case $k=0$.}
\begin{enumerate}\itemsep=0pt
\item[(1)] $D(n,0,1,p,b) =1$ for all nonnegative $n$, $p$, $b$.
\item[(2)] $D(n,0,2,2,2) ={\rm VSASM}(n)$, i.e., the number of alternating sign
$(2n+1) \times (2n+1)$ matrices symmetric about the vertical axis, see, e.g., \cite[$A005156$]{SL}.
\item[$(3)$] $D(n,0,2,1,2)= {\rm CSTCPP}(n)$, i.e., the number of cyclically symmetric
transpose complement plane partitions, see, e.g., \cite[$A051255$]{SL}.
\end{enumerate}
\begin{Theorem} \label{theorem5.10}
Let $\varpi_{n,k,p}$ be a unique vexillary permutation
of the shape $\lambda_{n.p}:=(n,n-1, \ldots,2,1)p$ and flag
$\phi_{n,k} := (k+1,k+2,\ldots,k+n-1,k+n)$. Then
\begin{gather*}
\mathfrak{G}_{\varpi_{n,1,p}}^{(\beta-1)}(1)= \sum_{j=1}^{n+1} {1 \over n+1} {n+1 \choose j} {(n+1)p \choose j-1} \beta^{j-1}.
\end{gather*}
If $k \ge 2$, then $G_{n,k,p}(\beta) := \mathfrak{G}_{\varpi_{n,k,p}}^{(\beta -1)}(1)$ is a~polynomial of deg\-ree~$nk$ in~$\beta$, and
\begin{gather*}
{\rm Coef\/f}_{[\beta^{nk}]}(G_{n,k,p}(\beta)) = D(n,k,1,p-1,0).
\end{gather*}
\end{Theorem}
The polynomial
\begin{gather*}
\sum_{j=1}^{n} {1 \over n} {n \choose j} {p n \choose j-1} t^{j-1}:=\mathfrak{FN}_{n}(t)
\end{gather*}
is known as the Fuss--Narayana polynomial and can be considered as a $t$-deformation of the Fuss--Catalan number ${\rm FC}_{n}^{p}(0)$.
Recall that the number ${1 \over n} {n \choose j} {p n \choose j-1}$ counts
paths from $(0,0)$ to $(np,0)$ in the f\/irst quadrant, consisting of steps
$U=(1,1)$ and $D=(1,-p)$ and have~$j$ peaks (i.e., $UD$'s), cf.~\cite[$A108767$]{SL}.
For example, take $n=3$, $k=2$, $p=3$, $r=1$, $b=0$. Then
\begin{gather*}
\varpi_{3,2,3} = [1,2,12,9,6,3,4,5,7,8,10,11] \in \mathbb{S}_{12},
\\
G_{3,2,3}(\beta)= (1,18,171,747,1767,1995,1001).
\end{gather*}
Therefore,
\begin{gather*}
G_{3,2,3}(1)=5700=D(3,2,3,0) \qquad \text{and} \qquad
{\rm Coef\/f}_{[\beta^{6}]}(G_{3,2,3}(\beta))= 1001= D(3,2,2,0).
\end{gather*}
\begin{Proposition}[\cite{NT}]\label{proposition5.5}
The value of the Fuss--Catalan polynomial at $t=2$, that is the
number
\begin{gather*}
\sum_{j=1}^{n} {1 \over n} {n \choose j} {p n \choose j-1} 2^{j-1}
\end{gather*}
is equal to the number of hyperplactic classes of $p$-parking functions of
length~$n$, see~{\rm \cite{NT}} for definition of $p$-parking functions, its
properties and connections with some combinatorial Hopf algebras.
\end{Proposition}
Therefore, the value of the {\it Grothendieck polynomial}
$\mathfrak{G}_{\varpi_{n,1,p}}^{(\beta =1)}(1)$ at $\beta=1$ and $x_i =1$,
$\forall\, i$, is equal to the number of $p$-parking functions of length $n+1$.
It is an open problem to f\/ind combinatorial interpretations of the
polynomials $\mathfrak{G}_{\varpi_{n,k,p}}^{(\beta )}(1)$ in the case
$k \ge 2$. Note f\/inally, that in the case $p=2,~k=1$ the values of the Fuss--Catalan polynomials at $t=2$ one can f\/ind in \cite[$A034015$]{SL}.
\begin{Comments}\label{comments5.7} {\it $(\Longrightarrow)$ The case $r=0$.}
It follows from Theorem~\ref{theorem5.7} that in the case $r=0$ and
$k \ge n$, one has
\begin{gather*}
D(n,k,0,p,b)= \dim V_{\lambda_{n,p,b}}^{\mathfrak{gl}(k+1)}=
(1+p)^{{n+1 \choose 2}} \prod_{j=1}^{n+1} {{(n-j+1)p+b+k-j+1 \choose k-j+1}
\over {(n-j+1)(p+1)+b \choose n-j+1}} .
\end{gather*}
Now consider the conjugate $\nu:=\nu_{n,p,b}:=((n+1)^b,n^{p},(n-1)^{p},\ldots,
1^{p})$ of the partition $\lambda_{n,p,b}$, and a rectangular shape partition $\psi=(\underbrace{k,\ldots,k}_{np+b})$. If $k \ge np+b$, then there exists a
unique {\it grassmannian} permutation
$\sigma:=\sigma_{n,k,p,b}$ of the {\it shape} $\nu$ and the
{\it flag}~$\psi$~\cite{M}. It is easy to see from the above formula
for $D(n,k,0,p,b)$, that
\begin{gather*}
\mathfrak{S}_{\sigma_{n,k,p,b}}(1) = \dim V_{\nu_{n,p,b}}^{\mathfrak{gl}(k-1)}\\
\hphantom{\mathfrak{S}_{\sigma_{n,k,p,b}}(1)}{}=
(1+p)^{{n \choose 2}} {k+n-1 \choose b} \prod_{j=1}^{n}{(p+1)(n-j+1) \over (n-j+1)(p+1)+b} \prod_{j=1}^{n}{ {k+j-2 \choose (n-j+1)p+b} \over {(n-j+1)(p+1)+b-1 \choose n-j}}.
\end{gather*}
After the substitution $k:=np+b+1$ in the above formula we will have
\begin{gather*}
\mathfrak{S}_{\sigma_{n,np+b+1,p,b}}(1)= (1+p)^{{n \choose 2}} \prod_{j=1}^{n} {{{np+b+j-1 \choose (n-j+1)p }} \over {{j(p+1)-1 \choose j-1}}}.
\end{gather*}
In the case $b=0$ some simplif\/ications are happened, namely,
\begin{gather*}
\mathfrak{S}_{\sigma_{n,k,p,0}}(1) = (1+p)^{{n \choose 2}} \prod_{j=1}^{n}
{{k+j-2 \choose (n-j+1)p} \over {(n-j+1)p+n-j \choose n-j}}.
\end{gather*}
Finally we observe that if $k=np+1$, then
\begin{gather*}
\prod_{j=1}^{n} {{{ np+j-1 \choose (n-j+1)p}} \over {{(n-j+1)p+n-j \choose n-j}}}=
{\prod_{j=2}^{n} {{np+j-1 \choose (p+1)(j-1)} \over {j(p+1)-1 \choose j-1}} =
\prod_{j=1}^{n-1} {j ! (n(p+1) -j -1) ! \over ((n-j)(p+1)) ! ((n-j)(p+1)-1) !}}:=A_{n}^{(p)},
\end{gather*}
where the numbers $A_{n}^{(p)}$ are {\it integers} that generalize the
numbers of {\it alternating sign matrices} (ASM) of size $n \times n$, recovered in the case
$p=2$, see \cite{DZ, O} for details.
\end{Comments}
\begin{Examples}\label{examples5.1} \quad
$(1)$ Let us consider polynomials $\mathfrak{G}_{n}(\beta) :=
\mathfrak{G}_{{\sigma_{n,2n,2,0}}}^{(\beta-1)}(1)$.
If $n=2$, then
\begin{gather*}
\sigma_{2,4,2,0}=235614 \in \mathbb{S}_6, \qquad
\mathfrak{G}_2(\beta)= (1,2,{\bf 3}):=1+ 2 \beta+ {\bf 3} \beta^2.
\end{gather*}
Moreover,
\begin{gather*}
\mathfrak{R}_{\sigma_{2,4,2,0}}(q;\beta)= (1,{\bf 2})_{\beta}+ {\bf 3} q \beta^2.
\end{gather*}
If $n=3$, then
\begin{gather*}
\sigma_{3,6,2,0}=235689147 \in \mathbb{S}_9, \qquad
\mathfrak{G}_{3}(\beta)=(1,6,21,36,51,48,{\bf 26}).
\end{gather*}
Moreover,
\begin{gather*}
\mathfrak{R}_{\sigma_{3,6,2,0}}(q;\beta) = (1,6,11,16,{\bf 11})_{\beta} +
q \beta^2 (10,20,35,34)_{\beta}+q^2 \beta^4 (5,14,{\bf 26})_{\beta},\\
\mathfrak{R}_{\sigma_{3,6,2,0}}(q;1) = (45,99,45)_{q}.
\end{gather*}
If $n=4$, then
\begin{gather*}
\sigma_{4,8,2,0}= [2,3,5,6,8,9,11,12,1,4,7,10] \in
\mathbb{S}_{12},\\
\mathfrak{G}_{4}(\beta)= (1,12,78,308,903,2016,3528,4944,5886,5696,4320,2280,{\bf 646}).
\end{gather*}
Moreover,
\begin{gather*}
\mathfrak{R}_{\sigma_{4,8,2,0}}(q;\beta)= (1,12,57,182,392,602,763,730,493,{\bf 170})_{\beta}\\
\hphantom{\mathfrak{R}_{\sigma_{4,8,2,0}}(q;\beta)=}{} +
q \beta^2 (21,126,476,1190,1925,2626,2713,2026,804)_{\beta}\\
\hphantom{\mathfrak{R}_{\sigma_{4,8,2,0}}(q;\beta)=}{}+
q^2 \beta^4 (35,224,833,1534,2446,2974,2607,1254)_{\beta}\\
\hphantom{\mathfrak{R}_{\sigma_{4,8,2,0}}(q;\beta)=}{}
+q^3 \beta^6 (7,54,234,526,909,1026,{\bf 646})_{\beta},\\
\mathfrak{R}_{\sigma_{4,8,2,0}}(q;1) = (3402,11907,11907,3402)_{q}= 1701~(2,7,7,2)_{q}.
\end{gather*}
$\bullet$ If $n=5$, then
\begin{gather*}
\sigma_{5,10,2}=[2,3,5,6,8,9,11,12,14,15,1,4,7,10,13] \in \mathbb{S}_{15}, \\
\mathfrak{G}_{5}(\beta)= (1,20,210,1420,7085,27636,87430,230240,516375,997790,1676587,2466840,\\
\hphantom{\mathfrak{G}_{5}(\beta)= (}{}
3204065,3695650,3778095,3371612,2569795,1610910,782175,262200,{\bf 45885}).
\end{gather*}
Moreover,
\begin{gather*}
\mathfrak{R}_{\sigma_{5,10,2,0}}(q;\beta)= (1, 20, 174, 988, 4025,
12516, 31402, 64760, 111510, 162170,\\
\hphantom{\mathfrak{R}_{\sigma_{5,10,2,0}}(q;\beta)= (}{}
202957,220200, 202493, 153106, 89355,35972, {\bf 7429})_{\beta}\\
\hphantom{\mathfrak{R}_{\sigma_{5,10,2,0}}(q;\beta)=}{}
+
q \beta^2 (36, 432, 2934, 13608, 45990, 123516, 269703, 487908, 738927,
\\
\hphantom{\mathfrak{R}_{\sigma_{5,10,2,0}}(q;\beta)= (}{}
956430, 1076265,
1028808, 813177, 499374, 213597, 47538)_{\beta}\\
\hphantom{\mathfrak{R}_{\sigma_{5,10,2,0}}(q;\beta)=}{}
+
q^2 \beta^{4} (126, 1512, 9954, 40860, 127359, 314172, 627831,1029726,
1421253,\\
\hphantom{\mathfrak{R}_{\sigma_{5,10,2,0}}(q;\beta)= (}{}
1711728,
1753893, 1492974, 991809, 461322, 112860)_{\beta}\\
\hphantom{\mathfrak{R}_{\sigma_{5,10,2,0}}(q;\beta)=}{}
+
q^3 \beta^6 (84, 1104, 7794, 33408, 105840, 255492, 486324, 753984, 1019538,\\
\hphantom{\mathfrak{R}_{\sigma_{5,10,2,0}}(q;\beta)= (}{}
1169520, 1112340,
825930, 428895, 117990)_{\beta}\\
\hphantom{\mathfrak{R}_{\sigma_{5,10,2,0}}(q;\beta)=}{}
+
q^4 \beta^8 (9, 132, 1032, 4992, 17730, 48024, 102132, 173772, 244620, 276120,\\
\hphantom{\mathfrak{R}_{\sigma_{5,10,2,0}}(q;\beta)= (}{}
240420, 144210,
{\bf 45885})_{\beta},\\
\mathfrak{R}_{\sigma_{5,10,2,0}}(q;1) = (1299078,6318243,10097379,6318243,1299078)_{q}\\
\hphantom{\mathfrak{R}_{\sigma_{5,10,2,0}}(q;1)}{} =
59049 (22,107,171,107,22)_{q}.
\end{gather*}
We are reminded that over the paper we have used the notation
\begin{gather*}
(a_{0},a_{1},\ldots,a_{r})_{\beta} := \sum\limits_{j=0}^{r} a_{j} \beta^{j},
\end{gather*}
etc.
One can show that $\deg_{[\beta]} \mathfrak{G}_n(\beta)=n(n-1)$,
$\deg_{[q]} \mathfrak{R}_{\sigma_{n,2n,2,0}}(q,1) =n-1$, and looking on the
numbers $3$, $26$, $646$, $45885$ we made
\begin{Conjecture}\label{conjecture5.2}
Let $a(n):= {\rm Coef\/f}[\beta^{n(n-1)}]~
(\mathfrak{G}_n(\beta) )$. Then
\begin{gather*}
a(n)={\rm VSASM}(n)={\rm OSASM}(n)= \prod_{j=1}^{n-1} { (3j+2)(6j+3)! (2j+1)! \over (4j+2)! (4j+3)!},
\end{gather*}
where
${\rm VSASM}(n)$ is the number of alternating sign $(2n+1) \times (2n+1)$ matrices
symmetric about the vertical axis,
${\rm OSASM}(n)$ is the number of $2n \times 2n$ off-diagonal symmetric alternating
sign matrices.
See {\rm \cite[$A005156$]{SL}}, {\rm \cite{O}} and references therein, for details.
\end{Conjecture}
\begin{Conjecture}\label{conjecture5.3}
Polynomial $\mathfrak{R}_{\sigma_{n,2n,2,0}}(q;1)$ is symmetric
and
\begin{gather*}
\mathfrak{R}_{\sigma_{n,2n,2,0}}(0;1) = A20342(2n-1),
\end{gather*}
see~{\rm \cite{SL}}.
\end{Conjecture}
$(2)$ Let us consider polynomials $\mathfrak{F}_{n}(\beta) :=\mathfrak{G}_{{\sigma_{n,2n+1,2,0}}}^{(\beta-1)}(1)$.
If $n=1$, then
\begin{gather*}
\sigma_{1,3,2,0}=1342 \in \mathbb{S}_4, \qquad \mathfrak{F}_2(\beta)= (1,{\bf 2}):=1+{\bf 2} \beta.
\end{gather*}
If $n=2$, then
\begin{gather*}
\sigma_{2,5,2,0}=1346725 \in \mathbb{S}_7, \qquad \mathfrak{F}_3(\beta)= (1,6,11,16,{\bf 11}).
\end{gather*}
Moreover,
\begin{gather*}
\mathfrak{R}_{\sigma_{2,5,2,0}}(q;\beta)= (1,2, {\bf 3})_{\beta} +q \beta (4,8,12)_{\beta}+q^2 \beta^3 (4,{\bf 11})_{\beta}.
\end{gather*}
If $n=3$, then
\begin{gather*}
\sigma_{3,7,2,0}=[1,3,4,6,7,9,10,2,5,8] \in \mathbb{S}_{10}, \\
\mathfrak{F}_4(\beta)=
(1,12,57,182,392,602,763,730,493,{\bf 170}).
\end{gather*}
Moreover,
\begin{gather*}
\mathfrak{R}_{\sigma_{3,7,2,0}}(q;\beta)=
(1,6,21,36,51,48,{\bf 26})_{\beta} + q
\beta (6,36,126,216,306,288,156)_{\beta}\\
\hphantom{\mathfrak{R}_{\sigma_{3,7,2,0}}(q;\beta)=}{}+ q^2 \beta^3 (20,125,242,403,460,289)_{\beta} + q^3 \beta^5 (6,46,114,204,
{\bf 170})_{\beta},\\
\mathfrak{R}_{\sigma_{3,7,2,0}}(q;1) = (189,1134,1539,540)_{q}=27 (7,42,57,20)_{q}.
\end{gather*}
If $n=4$, then
\begin{gather*}
\sigma_{4,9,2,0} = [1,3,4,6,7,9,10,12,13,2,5,8,11] \in \mathbb{S}_{13},\\
\mathfrak{F}_5(\beta)= (1,20,174,988,4025,12516,31402,64760,111510,162170,202957,\\
\hphantom{\mathfrak{F}_5(\beta)= (}{} 220200,202493,153106,89355,35972,{\bf 7429}).
\end{gather*}
Moreover,
\begin{gather*}
\mathfrak{R}_{\sigma_{4,9,2,0}}(q;\beta)=
(1,12,78,308,903,2016,3528,4944,5886,5696,4320,2280,{\bf 646})_{\beta} \\
\hphantom{\mathfrak{R}_{\sigma_{4,9,2,0}}(q;\beta)=}{}
+ q \beta (8,96,624,2464,7224,16128,28224,39552,47088,45568,\\
\hphantom{\mathfrak{R}_{\sigma_{4,9,2,0}}(q;\beta)=(}{}
34560,18240,5168)_{\beta}\\
\hphantom{\mathfrak{R}_{\sigma_{4,9,2,0}}(q;\beta)=}{}
+q^2 \beta^3(56,658,3220,11018,27848,53135,78902,100109,103436,\\
\hphantom{\mathfrak{R}_{\sigma_{4,9,2,0}}(q;\beta)=(}{}
84201,47830,14467)_{\beta}\\
\hphantom{\mathfrak{R}_{\sigma_{4,9,2,0}}(q;\beta)=}{}
+q^3 \beta^5(56,728,3736,12820,29788,50236,72652,85444,78868,\\
\hphantom{\mathfrak{R}_{\sigma_{4,9,2,0}}(q;\beta)=(}{}
50876,17204)_{\beta}\\
\hphantom{\mathfrak{R}_{\sigma_{4,9,2,0}}(q;\beta)=}{}
+q^4 \beta^7(8,117,696,2724,7272,13962,21240,24012,18768,{\bf 7429})_{\beta},\\
\mathfrak{R}_{\sigma_{4,9,2,0}}(q;1) = (30618,244944,524880,402408,96228)_{q} = 4374 (7,56,120,92,22)_{q} .
\end{gather*}
One can show that $\mathfrak{F}_n(\beta)$ is a polynomial in $\beta$ of degree~$n^2$, and looking on the num\-bers~$2$, $11$, $170$, $7429$ we made
\begin{Conjecture}\label{conjecture5.4}
Let $b(n):= {\rm Coef\/f}_{[\beta^{(n-1)^2}]}
(\mathfrak{F}_n(\beta))$. Then
$b(n) = {\rm CSTCPP}(n)$. In other words, $b(n)$ is equal to the number of
cyclically symmetric transpose complement plane partitions in an
$2n \times 2n \times 2n$~box. This number is known to be
\begin{gather*}
\prod_{j}^{n-1} {{(3j+1)(6j)! (2j)! \over (4j+1)! (4j)! }},
\end{gather*}
see {\rm \cite[$A051255$]{SL}}, {\rm \cite[p.~199]{Br}}.
\end{Conjecture}
It ease to see that polynomial $\mathfrak{R}_{\sigma_{n,2n+1,2,0}}(q;1)$ has degree~$n$.
\begin{Conjecture}\label{conjecture5.5}
\begin{gather*}
{\rm Coef\/f}_{[\beta^n]} \big(\mathfrak{R}_{\sigma_{n,2n+1,2,0}}(q;1) \big) =
A20342(2n),
\end{gather*}
see {\rm \cite{SL}};
\begin{gather*}
\mathfrak{R}_{\sigma_{n,2n+1,2,0}}(0;1)= A_{\rm QT}^{(1)}(4n;3) = 3^{n(n-1)/2} {\rm ASM}(n),
\end{gather*}
see {\rm \cite[Theorem~5]{Ku}} or {\rm \cite[$A059491$]{SL}}.
\end{Conjecture}
\begin{Proposition}\label{proposition5.6}
One has
\begin{gather*}
\mathfrak{R}_{\sigma_{4,2n+1,2,0}}(0;\beta) = \mathfrak{G}_n( \beta) = \mathfrak{G}_{\sigma_{n,2n,2,0}}^{(\beta-1)}(1),\qquad
\mathfrak{R}_{\sigma_{n,2n,2,0}}(0,\beta)= \mathfrak{F}_n(\beta)= \mathfrak{G}_{\sigma_{n,2n+1,2,0}}^{(\beta-1)}(1).
\end{gather*}
\end{Proposition}
\end{Examples}
Finally we def\/ine $(\beta,q)$-deformations of the numbers ${\rm VSASM}(n)$ and
${\rm CSCTPP}(n)$. To accomplish these ends, let us consider permutations
\begin{gather*}
w_k^{-} = (2,4,\ldots, 2k,2k-1,2k-3,\ldots,3,1),\\
w_k^{+} = (2,4,\ldots,2k,2k+1,2k-1,\ldots,3,1).
\end{gather*}
\begin{Proposition}\label{proposition5.7} One has
\begin{gather*}
\mathfrak{S}_{{w_{k}^{-}}}(1)={\rm VSAM}(k), \qquad \mathfrak{S}_{{w_{k}^{+}}}(1) = {\rm CSTCPP}(k) .
\end{gather*}
\end{Proposition}
Therefore the polynomials $\mathfrak{G}_{{w_{k}^{-}}}^{(\beta-1)}(x_1=q,\,x_j=1,\, \forall\, j \ge 2)$ and $\mathfrak{G}_{{w_{k}^{+}}}^{(\beta-1)}(x_1=q,\,x_j=1$, $\forall \,j \ge 2)$ def\/ine $(\beta,q)$-deformations of the numbers ${\rm VSAM}(k)$ and
${\rm CSTCPP}(k)$ respectively. Note that the inverse permutations
\begin{gather*}
(w_{k}^{-})^{-1} = (\underbrace{2k,1},\ldots,\underbrace{2k+1-i,i},\ldots,
\underbrace{k+1,k}),\\
(w_{k}^{+})^{-1}= (\underbrace{2k+1,1},\ldots,\underbrace{2k+2-j,j},
\ldots, \underbrace{k+2,k},k+1)
\end{gather*}
also def\/ine a $(\beta,q)$-deformation of the
numbers considered above.
\begin{Problem}\label{problem5.1}
It is well-known, see, e.g., {\rm \cite[p.~43]{ER}}, that the set
${\cal{VSASM}}(n)$ of alternating sign $(2n+1) \times (2n+1)$ matrices
symmetric about the vertical axis has the same cardinality as the set
${\rm SYT}_{2}( \lambda(n), \le n)$ of semistandard Young tableaux of the shape
$\lambda(n):= (2n-1,2n-3,\ldots,3,1)$ filled by the numbers from the set
$\{1,2,\ldots,n \}$, and such that the entries are weakly increasing down the
anti-diagonals.
On the other hand, consider the set ${\cal{CS}}(w_{k}^{-})$
of compatible sequences, see, e.g., {\rm \cite{BJS,FK1}}, corresponding to
the permutation $w_{k}^{-} \in {\mathbb{S}}_{2k}$.
\end{Problem}
\begin{Challenge}
Construct bijections between the sets
${\cal{CS}}(w_k^{-})$, ${\rm SYT}_{2}( \lambda(k), \le k)$ and\linebreak ${\cal{VSASM}}(k)$.
\end{Challenge}
\begin{Remark}\label{remarks5.1} One can compute the principal specialization of the Schubert polynomial corresponding to the transposition $t_{k,n}:=
(k,n-k) \in \mathbb{S}_n$ that interchanges $k$ and $n-k$, and f\/ixes all
other elements of~$[1,n]$.
\end{Remark}
\begin{Proposition}\label{proposition5.8}
\begin{gather*}
q^{(n-1)(k-1)} \mathfrak{S}_{t_{k,n-k}}\big(1,q^{-1},q^{-2},q^{-3},\ldots\big)\\
\qquad{}= \sum_{j=1}^{k} (-1)^{j-1} q^{{j \choose 2}} {n-1 \brack k-j }_{q}~{n -2+j
\brack k+j-1}_{q} =
\sum_{j=1}^{n-2} q^{j} \left( {j+k-2 \brack k-1}_{q}
\right)^{2}.
\end{gather*}
\end{Proposition}
\begin{Exercises} \label{exercises5.7}\quad
(1) Show that if $k \ge 1$, then
\begin{gather*}
{\rm Coef\/f}_{[q^k \beta^{2k}]} ( {\mathfrak{R}}_{{\sigma_{n,2n,2,0}}}(q;t)
) ={2n-1 \choose 2k},\\
{\rm Coef\/f}_{[q^k \beta^{2k-1}]} ( {\mathfrak{R}}_{{\sigma_{n,2n+1,2,0}}}(q;t) ) ={2n \choose 2k-1}.
\end{gather*}
(2) Let $n \ge 1$ be a positive integer, consider ``zig-zag'' permutation
\begin{gather*}
w = \begin{pmatrix} 1& 2& 3& 4&\ldots& 2k+1& 2k+2& \ldots& {2n-1}& 2n \\
2& 1& 4& 3&\ldots& {2k+2}& {2k+1}& \ldots& 2n& {2n-1} \end{pmatrix}
\in \mathbb{S}_{2n}.
\end{gather*}
Show that
\begin{gather*}
\mathfrak{R}_{w}(q,\beta)= \prod_{k=0}^{n-1} \left({1- \beta^{2k} \over 1-\beta} +q \beta^{2k} \right).
\end{gather*}
(3) Let $\sigma_{k,n,m}$ be grassmannian permutation with shape
$\lambda =(n^m)$ and f\/lag $\phi =(k+1)^m$, i.e.,
\begin{gather*}
\sigma_{k,n,m} = \begin{pmatrix} 1& 2& \ldots& k& {k+1}& \ldots& {k+n}& {k+n+1}& \ldots& {k+n+m} \\
1& 2& \ldots& k& {k+m+1}& \ldots& {k+m+n}& {k+1}& \ldots& {k+m} \end{pmatrix}.
\end{gather*}
Clearly $\sigma_{k+1,n,m}= 1 \times \sigma_{k,n,m}$.
Show that the coef\/f\/icient
${\rm Coef\/f}_{\beta^m} ( \mathfrak{R}_{\sigma_{k,n,m}}(1,\beta) )$
is equal to the Narayana number $ N(k+n+m,k)$.
(4) Consider permutation $w:= w^{(n)}=(w_1,\ldots,w_{2n+1})$, where
$w_{2k-1}=2k+1$ for $k=1,\ldots, n$, $w_{2n+1}= 2 n$, $w_{2}=1$ and
$w_{2k}= 2 k -2$ for $k=2,\ldots,n$. For example, $w^{(3)}=(3152746)$. We
set $w^{(0)}=1$.
Show that
the polynomial ${\mathfrak{S}}_{w}^{(\beta)}(x_i=1, \forall i)$ has degree
$n(n-1)$ and the coef\/f\/icient
${\rm Coef\/f}_{\beta^{n(n-1)}} ({\mathfrak{S}}_{w}^{(\beta)}(x_i=1,\, \forall \, i) )$
is equal to the $n$-th Catalan number $C_n$.
Note that the specialization
$ {\mathfrak{S}}_{w}^{(\beta)}(x_i=1)|_{\beta=1}$ is equal to the $2n$-th Euler
(or up/down) number, see~\cite[$A000111$]{SL}.
More generally, consider permutation $w^{(n)}_{k}:= 1^k \times w^{(n)} \in
{\mathbb{S}}_{k+2n+1}$, and polynomials
\begin{gather*}
P_k(z)=\sum_{j \ge 0} (-1)^{j} {\mathfrak{S}}_{w^{(j)}_{k-2j}}(x_i=1) z^{k-2j},
\qquad k \ge 0.
\end{gather*}
Show that
\begin{gather*}
\sum_{k \ge 0} P_k(z) {t^k \over k!} = \exp (t z)\operatorname{sech}(t).
\end{gather*}
The polynomials $P_{k}(z)$ are well-known as {\it Swiss--Knife} polynomials,
see \cite[$A153641$]{SL}, where one can f\/ind an overview of some properties of
the Swiss--Knife polynomials.
(5) Assume that $n=2k+3$, $k \ge 1$, and consider permutation
$v_n=(v_1,\ldots,v_n) \in \mathbb{S}_n$, where $v_{2a+1}=2a+3$, $a=0,\ldots,n-1$, $w_2=1$ and $w_{2a}=2 a-2$, $a=2,\ldots,k+1$. For example, $v_4=[31527496,11,8,10]$ and ${\mathfrak{S}}_{v_{4}}(1)=50521 =E_{10}$.
Show that
\begin{gather*}
{\mathfrak{S}}_{v_{n}}(q,\,x_i=1,\,\forall\, i \ge 2) = (n-2) E_{n-3} q^2+ \cdots+ 2^{k-1} (k-1) ! q^{k+2}, \\
{\mathfrak{S}}_{v_{n}}(x_i=1,\,\forall\, i \ge 1) = E_{n-1}.
\end{gather*}
Set $\beta=d-1$, consider polynomials ${\cal{E}}_n(q,d)=
{\mathfrak{G}}_{v_{n}}^{(\beta)}(x_1=q, \,x_i =1,\, \forall\, i \ge 2)$.~Clearly,
see the latter formula, ${\cal{E}}_{n}(1,1)= E_{n-1}$. Give a~combinatorial prove that ${\cal{E}}_n(q,d) \in \mathbb{N} [q,d]$, that is to give combinatorial interpretation(s) of coef\/f\/icients of the polynomial
${\cal{E}}_{n}(q,d)$.
Show that $\deg_{d} {\cal{E}}_{n}(1,d) =n(n+1)$ and the
leading coef\/f\/icient is equal to the Catalan number~$C_{n+1}$.
(6) Consider permutation $u:=u_{n}=(u_1,\ldots,u_{2n}) \in
{\mathbb S}_{2n}$, $n \ge 2$, where
$u_1=2$, $u_{2k+1}=2k-1$, $k=1,\ldots,n$,
$u_{2k}=2k+2$, $k=1,\ldots,n-1$, $u_{2n}=2n-1$. For example, $u_{4}=(24163857)$.
Now consider polynomial
\begin{gather*}
R_n^{(k)}(q) ={\mathfrak{S}}_{1^{k} \times u_{n}}(x_1=q,\,x_i=1, \,\forall \,i \ge 2).
\end{gather*}
Show that
$R_{n}^{(k)}(1) ={2n+k-1 \choose k} E_{2n-1}$, where $E_{2k-1}$,
$k \ge 1$, denotes the Euler number, see \cite[$A00111$]{SL}. In particular,
$R_{n}^{(1)}(1) =2^{2n-1} G_n$, where~$G_n$ denotes the {\it unsigned Genocchi} number, see \cite[$A110501$]{SL}.
Show that $\deg_{q} R_{n}^{(k)}(q) =n$ and ${\rm Coef\/f}_{q^{n}} \big(
R_{n}^{(0)}(q) \big) =(2n-3) ! !$.
(7) Consider permutation $w_n \in \mathbb{S}_{2n+2}$, where $w_2=1$, $w_4=2$, and
\begin{gather*}
w_{2k-1}=2k+2, \quad 1 \le k \le n, \qquad w_{2k}=2k-3, \quad 3 \le k \le n,\\
w_{2n+1}=2n-3,\qquad w_{2n+2}=2n-1.
\end{gather*}
For example, $w_{5}= [4,1,6,2,8,3,10,5,12,7,9,11]$.
Show that
\begin{gather*}
{\mathfrak{S}}_{w_{n}}(x_i=1, \, \forall\, i) =(2n+1) !! \big(2^{2n} -2\big) |B_{2n} |,
\end{gather*}
where $B_{2n}$ denotes the {\it Bernoulli} numbers\footnote{See, e.g., \url{https://en.wikipedia.org/wiki/Bernoulli_number}.}.
(8) Consider permutation $w_k:= (2k+1,2k-1,\ldots,3,1,2k,2k-2,\ldots,4,2)
\in {\mathbb{S}}_{2k+1}$. Show that
\begin{gather*}
{\mathfrak{S}}_{w_{k}}^{(\beta-1)}(x_1=q,\,x_j=1,\, \forall \, j \ge 2) = q^{2k} (1+\beta)^{{n \choose 2}}.
\end{gather*}
(9) Consider permutations $\sigma_k^{+}=(1,3,5,\ldots,2k+1,2k+2,2k,\ldots,4,2)$ and $\sigma_{k}^{-}=(1,3,5,\ldots$, $2k+1,2k,2k-2,\ldots,4,2)$, and def\/ine polynomials
\begin{gather*}
S_k^{\pm}(q)= {\mathfrak{S}}_{\sigma_{k}^{\pm}} (x_1=q,\,x_j=1, \,\forall\, j \ge 2).
\end{gather*}
Show that
\begin{gather*}
S_k^{+}(0)={\rm VSASM}(k), \qquad S_k^{+}(1)={\rm VSASM}(k+1), \\
{\partial \over \partial q} S_k^{+}(q) \vert_{q=0}= 2k S_{k}^{+}(0),\qquad
{\rm Coef\/f}_{q^{k}}(S_k^{+}(q) )={\rm CSTCPP}(k+1),\\
S_k^{-}(0)={\rm CSTCPP}(k), \qquad S_k^{-}(1)={\rm CSTCPP}(k+1), \\
{\partial \over \partial q} S_k^{-}(q) \vert_{q=0}= (2k-1) S_{k}^{-}(0),\qquad
{\rm Coef\/f}_{q^{k}} (S_k^{-}(q) )={\rm VSASM}(k).
\end{gather*}
Let's observe that $\sigma_{k}^{\pm}= 1 \times \tau_{k-1}^{\pm}$, where
$\tau_k^{+}= (2,4,\ldots,2k,2k+1,2k-1,\ldots,3,1)$ and
$\tau_{k}^{-} = (2,4,\ldots,2k,2k-1,2k-3,\ldots,3,1)$. Therefore,
\begin{gather*}
{\mathfrak{S}}_{\tau_{k}^{\pm}}(x_1=q,\, x_j=1,\,
\forall j \ge 2) = q S_{k-1}^{\pm}(q).
\end{gather*}
Recall that ${\rm CSTCPP}(n)$ denotes the number of cyclically symmetric transpose
compliment plane partitions in a $2 n \times 2 n$ box, see, e.g., \cite[$A051255$]{SL},
and ${\rm VSASM}(n)$ denotes the number of alternating sign $(2n+1)
\times (2n+1)$ matrices symmetric the vertical axis, see, e.g., \cite[$A005156$]{SL}.
It might be well to point out that
\begin{gather*}
{\mathfrak{S}}_{\sigma_{n-1}^{+}}(x_1=x,\,x_i=1,\, \forall\, i \ge 2)= G_{2n-1,n-1}(x,y=1),\\
{\mathfrak{S}}_{\sigma_{n}^{-}}(x_1=x, \,x_i =1,\, \forall\, i \ge 2)= F_{2n,n-1}(x,y=1),
\end{gather*}
where (homogeneous) polynomials $G_{m,n}(x,y)$~and $F_{m,n}(x,y)$ are
def\/ined in~\cite{Py}, and related with integral solutions to Pascal's
hexagon relations
\begin{gather*}
f_{m-1,n} f_{m+1,n} +f_{m,n-1} f_{m,n+1}= f_{m-1,n-1} f_{m+1,n+1}, \qquad (m,n) \in \mathbb{Z}^2.
\end{gather*}
(10) Consider permutation
\begin{gather*}
u_n = \begin{pmatrix} 1& 2& \ldots& n& n+1& n+2& n+3 &\ldots & 2n \\
2& 4& \ldots& 2n& 1& 3& 5&\ldots& 2n-1 \end{pmatrix},
\end{gather*}
and set $u_n^{(k)}:= 1^{2k+1} \times u_{n}$.
Show that
\begin{gather*}
{\mathfrak G}_{u_{n}^{(k)}}^{(\beta-1)}(x_i=1,\,\forall\, i \ge 1) =
(1+\beta)^{{n+1 \choose 2}} {\mathfrak G}_{1^{k} \times w_{0}^{(n+1)}}^{((\beta)^2-1)}(x_i=1,\,\forall\, i \ge 1),
\end{gather*}
where $w_{0}^{(n+)}$ denotes the permutation $(n+1,n,n-1,\ldots,2,1)$.
(11) Let $n \ge 0$ be an integer.
Consider permutation $u_n= 1^n \times 321 \in {\mathbb{S}}_{3+ n}$. Show that
\begin{gather*} {\mathfrak{S}}_{u_{n}}(x_1=t,\,x_i=1,\, \forall\, i \ge 2) = {\frac{1}{4}}
{2 n+2 \choose 3}
+ {\frac{n}{2}} {2 n+2 \choose 1} t + {\frac{1}{2}}{2 n+2 \choose 1} t^2.
\end{gather*}
Consider permutation $v_n := 1^n \times 4321 \in {\mathbb{S}}_{n+4}$.
Show that
\begin{gather*}
{\mathfrak{S}}_{v_{n}}(x_1=t,\,x_i=1,\,\forall\, i \ge 2) \\
\qquad{}=
{\frac{1}{24}}{2n+4 \choose 5}
{2 n+2 \choose 1} +{\frac{1}{2}} {2n+4 \choose 5} t + {\frac{n}{4}} {2 n+4 \choose 3} t^2+ {\frac{1}{4}} {2n +4 \choose 3} t^3.
\end{gather*}
(12) Show that
\begin{gather*}
\sum_{(a,b,c) \in (\mathbb{Z}_{\ge 0})^3 } q^{a+b+c}
{a+b \brack b}_{q}
{a+c \brack c }_{q} {b+c \brack b }_{q} = {1 \over (q;q)_{\infty}^3} \left( \sum_{k \ge 2} (-1)^k {k \choose 2} q^{{k \choose 2}-1} \right).
\end{gather*}
It is not dif\/f\/icult to see that the left hand side sum of the above identity
counts the weighted number of plane partitions $\pi=(\pi_{ij})$ such that
\begin{gather*}
\pi_{i,j} \ge 0, \qquad \pi_{ij} \ge \max(\pi_{i+1,j},\pi_{i,j+1}), \qquad \pi_{ij} \le 1
\quad \text{if}\quad i \ge 2 \quad \text{and} \quad j \ge 2,
\end{gather*}
and the weight $wt(\pi):=\sum\limits_{i,j} \pi_{ij}$.
$(13)$ Let $\lambda=(\lambda_1 \ge \lambda_2 \ge \cdots \ge \lambda_p > 0)$
be a partition of size $n$. For an integer~$k$ such that $1 \le k \le n-p$
def\/ine a grassmannian permutation
\begin{gather*}
w_{\lambda}^{(k)} = [1, \ldots, k, \lambda_p+k+1, \lambda_{p-1}+k+2,\ldots,
\lambda_1+k+p, a_1,\ldots,a_{n-p-k}],
\end{gather*}
where we denote by $(a_1 < a_2 < \cdots < a_{n-k-p})$ the complement
$[1,n] \backslash (1, \ldots, k, \lambda_p+k+1, \lambda_{p-1}+k+2,\ldots,
\lambda_1+k+p)]$.
Show that the Grothendieck polynomial
\begin{gather*}
G_{\lambda}(\beta):={\mathfrak G}_{{w_{\lambda}}^{k}}^{\beta-1}(1^n)
\end{gather*}
is a polynomial of~$\beta$ with {\it nonnegative} coef\/f\/icients.
Clearly, $G_{\lambda}(1) = \dim V_{\lambda}^{\mathfrak{Gl}(k+\ell(\lambda))}$.
Find a combinatorial interpretations of polynomial $G_{\lambda}(\beta)$.
\end{Exercises}
Final remark, it follows from the seventh exercise listed above, that the
polynomials ${\mathfrak{S}}_{\sigma_{k}^{\pm}}^{(\beta)}(x_1=q,\,x_j=1,\, \forall\, j \ge 2)$ def\/ine a $(q,\beta)$-deformation of the number ${\rm VSASM}(k)$ (the case $\sigma_{k}^{+}$)
and the number ${\rm CSTCPP}(k)$ (the case $\sigma_{k}^{-}$), respectively.
\subsubsection{Specialization of Grothendieck polynomials}\label{section5.2.5}
Let $p$, $b$, $n$ and $i$, $2i < n $ be positive integers. Denote by
${\cal{T}}_{p,b,n}^{(i)}$ the {\it trapezoid}, i.e., a convex quadrangle having
vertices at the points
\begin{gather*}
(ip,i), \quad (ip,n-i), \quad (b+ip,i) \quad \text{and} \quad (b+(n-i)p,n-i).
\end{gather*}
\begin{Definition}\label{definition5.5} Denote by ${\rm FC}_{b,p,n}^{(i)}$ the set of lattice path from the
point $(ip,i)$ to that $(b+(n-i)p,n-i)$ with east steps
$E = (0,1)$ and north steps $N = (1,0)$, which are located inside of the
trapezoid ${\cal{T}}_{p,b,n}^{(i)}$.
If $\mathfrak{p} \in {\rm FC}_{b,p,n}^{(i)}$ is a path, we denote by
$p(\mathfrak{p})$ the number of {\it peaks}, i.e.,
\begin{gather*}
p(\mathfrak{p})= NE(\mathfrak{p})+E_{\rm in}(\mathfrak{p})+N_{\rm end}(\mathfrak{p}),
\end{gather*}
where $NE(\mathfrak{p})$ is equal to the number of steps $NE$ resting on
path $\mathfrak{p}$; $E_{\rm in}(\mathfrak{p})$~is equal to~$1$, if the
path $\mathfrak{p}$
{\it starts} with step~$E$ and~$0$ otherwise;
$N_{\rm end}(\mathfrak{p})$ is equal to $1$, if the path $\mathfrak{p}$
{\it ends} by the step~$N$ and~$0$ otherwise.
\end{Definition}
Note that the equality $N_{\rm end}(\mathfrak{p})=1$ may happened only in the case
$b=0$.
\begin{Definition} \label{definition5.6}
Denote by ${\rm FC}_{b,p,n}^{(k)}$ the set of $k$-tuples
${\mathfrak{P}}=(\mathfrak{p}_1,\ldots,\mathfrak{p}_{k})$ of
{\it non-crossing}
lattice paths, where for each $i=1,\ldots,k$, $\mathfrak{p}_i \in
{\rm FC}_{b,p,n}^{(i)}$.
\end{Definition}
Let
\begin{gather*}
{\rm FC}^{(k)}_{b,p,n}(\beta):= \sum_{\mathfrak{P} \in {\rm FC}_{b,p,n}^{(k)}}
\beta^{p(\mathfrak{P})}
\end{gather*}
denotes the generating function of the statistics $p(\mathfrak{P}):=
\sum\limits_{i=1}^{k} p(\mathfrak{p_{i}}) -k$.
\begin{Theorem}\label{theorem5.11} The following equality holds
\begin{gather*}
\mathfrak{G}_{\sigma_{n,k,p,b}}^{(\beta)}(x_1=1,x_2=1,\ldots)={\rm FC}_{p,b,n+k}^{(k)}(\beta +1),
\end{gather*}
where $\sigma_{n,k,p,b}$ is a unique grassmannian permutation with shape
$((n+1)^b,n^p,(n-1)^p,\ldots,1^{p})$ and flag $(\underbrace{k,\ldots,k)}_{np+b}$.
\end{Theorem}
\subsection[The ``longest element'' and Chan--Robbins--Yuen polytope]{The ``longest element'' and Chan--Robbins--Yuen polytope\footnote{Some results of this section, e.g., Theorems~\ref{theorem5.11} and~\ref{theorem5.12},
has been proved independently and in greater generality in~\cite{Me3}.}}\label{section5.3}
\subsubsection[The Chan--Robbins--Yuen polytope ${\cal{CRY}}_n$]{The Chan--Robbins--Yuen polytope $\boldsymbol{{\cal{CRY}}_n}$}\label{section5.3.1}
Assume additionally, cf.~\cite[Exercise~6.C8(d)]{ST3}, that the
condition~$(a)$ in Def\/inition~\ref{definition5.1} is replaced by that
\begin{enumerate}\itemsep=0pt
\item[$(a')$] $x_{ij}$ and $x_{kl}$ {\it commute} for all $i$, $j$, $k$
and~$l$.
\end{enumerate}
Consider the element $w_{0}^{(n)}:= \prod\limits_{1 \le i < j \le n} x_{ij}$. Let
us bring the element $w_{0}^{(n)}$ to the reduced form, that is, let us
consecutively apply the def\/ining relations~$(a')$~and~$(b)$ to the element
$w_{0}^{(n)}$ in {\it any} order until unable to do so. Denote the
resulting polynomial by $Q_n(x_{ij};\alpha, \beta)$.
Note that the polynomial itself {\it depends} on the order in
which the relations~$(a')$ and~$(b)$ are applied.
We denote by $Q_n(\beta)$ the specialization
\begin{gather*}
x_{ij}=1 \qquad \text{for all $i$ and $j$},
\end{gather*}
of the polynomial $Q_n(x_{ij}; \alpha=0,\beta)$.
\begin{Example}\label{example5.7}
\begin{gather*}
Q_3(\beta)=(2,1)=1+(\beta +1),\qquad
Q_4(\beta)=(10,13,4)=1+5(\beta +1)+ 4 (\beta +1)^2, \\
Q_5(\beta)=(140,336,280,92,9) =1+16 (\beta+1)+58 (\beta +1)^2+56 (\beta +1)^3 + 9(\beta +1)^4, \\
Q_6(\beta)=1+42 (\beta+1)+448 (\beta+1)^2+1674 (\beta+1)^3+2364 (\beta+1)^4\\
\hphantom{Q_6(\beta)=}{} +
1182 (\beta+1)^5+169 (\beta+1)^6, \\
Q_7(\beta)=(1,99,2569,25587,114005,242415,248817,118587,22924,1156)_{\beta+1}, \\
Q_8(\beta)=(1,219,12444,279616,2990335,16804401,52421688,93221276,94803125, \\
\hphantom{Q_8(\beta)=(}{}
53910939, 16163947, 2255749, 108900)_{\beta+1}.
\end{gather*}
\end{Example}
What one can say about the polynomial
$Q_n(\beta):= Q_n(x_{ij};\beta) \arrowvert_{x_{ij}=1, \,\forall \, i,j}$?
It is known, \cite[Exercise~6.C8(d)]{ST3}, that the constant term of the
polynomial $Q_n(\beta)$ is equal to the product of Catalan numbers
$\prod\limits_{j=1}^{n-1}C_{j}$. It is not dif\/f\/icult to see that if $n \ge 3$, then
${\rm Coef\/f}_{[\beta+1]}(Q_n(\beta))= 2^n-1-{n+1 \choose 2} $, see \cite[$A002662$]{SL}, for a number of combinatorial interpretations of the numbers $2^n-1-{n+1 \choose 2} $.
\begin{Theorem}\label{theorem5.12}
One has
\begin{gather*}
Q_n(\beta -1) = \left(\sum_{m \ge 0} \iota( {\cal{CRY}}_{n+1}, m)
\beta ^{m}\right)
(1-\beta )^{{n+1 \choose 2}+1},
\end{gather*}
where ${ \cal{CRY}}_m$ denotes the Chan--Robbins--Yuen polytope~{\rm \cite{CR,CRY}}, i.e., the convex polytope given by the following conditions:
\begin{gather*}
{\cal{CRY}}_m = \bigl\{ (a_{ij}) \in \operatorname{Mat}_{m \times m}(\mathbb{Z}_{\ge 0}) \bigr\}
\end{gather*}
such that
\begin{enumerate}\itemsep=0pt
\item[$(1)$] $\sum_{i} a_{ij} =1$, $\sum_{j} a_{ij} =1$,
\item[$(2)$] $a_{ij}=0$ if $j > i+1$.
\end{enumerate}
Here for any integral convex polytope ${\cal P} \subset \mathbb{Z}^{d}$,
$\iota( {\cal P}, n)$ denotes the number of integer points in the set
$n {\cal P} \cap \mathbb{Z}^{d}$.
\end{Theorem}
In particular, the polynomial $Q_n(\beta)$ does not depend on the
order in which the relations $(a')$ and $(b)$ have been applied.
Now let us denote by ${\widehat{Q}}_n(q,t; \alpha,\beta)$ the specialization
\begin{gather*}
x_{ij}=1, \quad i < j < n, \qquad \text{and}\qquad x_{i,n}=q \quad \text{if}\quad i=2,\ldots, n-1, \qquad x_{1,n}=t
\end{gather*}
of the (reduced) polynomial $Q_n(x_{ij};\alpha,\beta)$ obtained by applying
the relations $(a')$ and $(b)$ in a~{\it certain} order. The
polynomial $Q_n(x_{ij};\alpha,\beta)$ itself {\it depends} on
the order selected. To def\/ine polynomials which are frequently appear in
Section~\ref{section5}, we apply the rules $(a)$ and $(b)$ stated in Def\/inition~\ref{definition5.1} to a
given monomial $x_{i_{1},j_{1}} \cdots x_{i_{p},j_{p}} \in
{\widehat{{\rm ACYB}}}_n(\alpha, \beta)$ consequently according to the order in
which the monomial taken has been written. We set $Q_n(t,\alpha,\beta):=
{\widehat{Q}}_n(q=t,t;\alpha,\beta)$.
\begin{Conjecture}\label{conjecture5.6}
Let $n \ge 3$ and write
\begin{gather*}
Q_n(t=1; \alpha,\beta)=\sum_{k \ge 0}(1+\beta)^k c_{k,n}( \alpha),
\end{gather*}
then $c_{k,n}( \alpha) \in \mathbb{Z}_{\ge 0}[\alpha]$.
The polynomial $Q_n(t,\beta,\alpha=0)$ has degree $d_n:=[\frac{(n-1)^2}{4}]$ with respect to~$\beta$.
Write
\begin{gather*}
Q_n(t,\beta):= Q_n(t;\alpha=0,\beta)= t^{n-2} \sum_{k=0}^{d_{n}} c_{n}^{(k)}(t) \beta^{k}.
\end{gather*}
Then
$c_n^{(d_n)}(1) =a_{n}^2$ for some non-negative integer $a_n$.
Moreover, there exists a polynomial $a_n(t) \in \mathbb{N}[t] $ such that
\begin{gather*}
c_n^{(d_n)}(t) =a_n(1) a_n(t), \qquad a_{n}(0)= a_{n-1}.
\end{gather*}
The all roots of the polynomial $Q_n(\beta)$ belong to the set $\mathbb{R}_{< -1}$.
\end{Conjecture}
For example,
\begin{gather*}
Q_4(t=1;\alpha,\beta)=(1,5,4)_{\beta +1}+ \alpha (5,7)_{\beta +1} +3 \alpha^2, \\
Q_5(t=1; \alpha,\beta)= (1,16,58,56,9)_{\beta+1} + \alpha (16,109,146,29)_{\beta+1}\\
\hphantom{Q_5(t=1; \alpha,\beta)=}{} + \alpha^2 (51,125,34)_{\beta+1}+\alpha^3(35,17)_{\beta+1},\\
c_6^{(6)}=13(2,3,3,3,2), \qquad c_{7}^{(9)}(t)= 34 (3,5,6,6,6,5,3), \\
c_{8}^{(12)}(t)=330(13,27,37,43,45,45,43,37,27,13),\\
Q_4(t,\beta,\alpha=0) t^{-1} = t^2+(\beta+1)\big(3t+2t^2\big)+(\beta+1)^2(t+1)^2,\\
{\widehat{Q}}_4(q,t; \alpha=0,\beta)= \big(q t^2 + t^3 + 2qt^3 + q^2t^3 + q^3t^3 + t^4 + 2qt^4 + q^2t^4\big)\\
\hphantom{{\widehat{Q}}_4(q,t; \alpha=0,\beta)=}{} + \big(2qt^2 + 2t^3 + 3qt^3 + 2q^2t^3
+ 2t^4 + 2qt^4\big) \beta +\big(t^2+t^3\big)(q+t) \beta^2,\\
{\widehat{Q}}_5(q,t; \alpha=0,\beta)= \big(3 q^2 t + q^3 t + 5 q t^2 + 6 q^2 t^2 + 2 q^3 t^2 + 2 t^3 + 10 q t^3 +
10 q^2 t^3 + 6 q^3 t^3\\
\hphantom{{\widehat{Q}}_5(q,t; \alpha=0,\beta)=}{}
+ 3 q^4 t^3 + 3 q^5 t^3 + 2 q^6 t^3 + 3 t^4 +
11 q t^4 + 11 q^2 t^4 + 8 q^3 t^4 + 5 q^4 t^4+ 3 q^5 t^4 \\
\hphantom{{\widehat{Q}}_5(q,t; \alpha=0,\beta)=}{}
+ 3 t^5 +
9 q t^5 + 9 q^2 t^5 + 6 q^3 t^5 + 3 q^4 t^5 + 2 t^6 + 6 q t^6 + 6 q^2 t^6 +
2 q^3 t^6\big) \\
\hphantom{{\widehat{Q}}_5(q,t; \alpha=0,\beta)=}{}
+\big(9 q^2 t + 2 q^3 t + 17 q t^2 + 18 q^2 t^2 + 4 q^3 t^2 + 7 t^3 +
31 q t^3 + 29 q^2 t^3 \\
\hphantom{{\widehat{Q}}_5(q,t; \alpha=0,\beta)=}{}
+ 15 q^3 t^3 + 10 q^4 t^3 + 7 q^5 t^3 + 10 t^4 +
31 q t^4 + 29 q^2 t^4 + 18 q^3 t^4 \\
\hphantom{{\widehat{Q}}_5(q,t; \alpha=0,\beta)=}{}
+ 10 q^4 t^4 + 10 t^5 + 24 q t^5 +
21 q^2 t^5 + 10 q^3 t^5 + 6 t^6 + 12 q t^6 + 6 q^2 t^6\big) \beta \\
\hphantom{{\widehat{Q}}_5(q,t; \alpha=0,\beta)=}{}
+ \big(9 q^2 t + q^3 t + 21 q t^2 + 18 q^2 t^2 + 2 q^3 t^2 + 9 t^3 + 34 q t^3 +
28 q^2 t^3 \\
\hphantom{{\widehat{Q}}_5(q,t; \alpha=0,\beta)=}{}
+ 14 q^3 t^3 + 9 q^4 t^3 + 12 t^4 + 30 q t^4 + 24 q^2 t^4 +
12 q^3 t^4 + 12 t^5 + 21 q t^5 \\
\hphantom{{\widehat{Q}}_5(q,t; \alpha=0,\beta)=}{}
+ 12 q^2 t^5 + 6 t^6 + 6 q t^6\big) \beta^2 +
\big(3 q^2 t + 11 q t^2 + 6 q^2 t^2 + 5 t^3 + 15 q t^3 \\
\hphantom{{\widehat{Q}}_5(q,t; \alpha=0,\beta)=}{}
+ 10 q^2 t^3 + 5 q^3 t^3 +
6 t^4 + 11 q t^4 + 6 q^2 t^4 + 6 t^5 + 6 q t^5 + 2 t^6\big) \beta^3\\
\hphantom{{\widehat{Q}}_5(q,t; \alpha=0,\beta)=}{}
+
\big(2 q t^2 + t^3 + 2 q t^3 + q^2 t^3 + t^4 + q t^4 + t^5\big) \beta^4.
\end{gather*}
Note that polynomials ${\widehat{Q}}_n(q,t; \alpha=0,\beta=0)$ give rise to
a two parameters deformation of the product of Catalan's numbers $C_1C_2 \cdots C_{n-1}$. Are there combinatorial interpretations of these polynomials and
polynomials ${\widehat{Q}}_n(q,t; \alpha=0,\beta)$?
\begin{Comments} \label{comments5.8}
We {\it expect} that for each integer $n \ge 2$ the set
\begin{gather*}
\Psi_{n+1} := \left\{ w \in \mathbb{S}_{2n-1} \,|\, \mathfrak{S}_{w}(1)=
\prod_{j=1}^{n} {\rm Cat}_j \right\}
\end{gather*}
is {\it non empty}, whereas the set
\begin{gather*}
\left\{ w \in
\mathbb{S}_{2n-2} \,|\, \mathfrak{S}_{w}(1)= \prod_{j=1}^{n} {\rm Cat}_j \right\}
\end{gather*} is
{\it empty}. For example,
\begin{gather*}
\Psi_4 =\{ [1,5,3,4,2] \}, \qquad \Psi_5 =
\{ [1,5,7,3,2,6,4],~[1,5,4,7,2,6,3] \},\\
\Psi_6 = \big\{ w:=[1,3,2,8,6,9,4,5,7], w^{-1}, \dots \big\} , \qquad \Psi_7 = \{??? \},
\end{gather*}
but one can check that for $w=[2358,10,549,12,11] \in {\mathbb{S}}_{12}$
\begin{gather*}
{\mathfrak{S}}_{w}(1)= 776160 = \prod_{j=2}^6{\rm Cat}_j.
\end{gather*}
More generally, for any positive integer $N$ def\/ine
\begin{gather*}
\kappa(N) = \min \{ n \,|\, \exists \, w \in {\mathbb{S}}_n \ \text{such that} \
{\mathfrak{S}}_{w}(1)=N \}.
\end{gather*}
It is clear that $\kappa(N) \le N+1$.
\begin{Problem}\label{problem-page120} Compute the following numbers
\begin{gather*}
\kappa( n !), \quad \kappa\left(\prod_{j=1}^{n} {\rm Cat}_j\right), \quad \kappa({\rm ASM}(n)), \quad \kappa\big((n+1)^{n-1}\big).
\end{gather*}
\end{Problem}
For example, $ 10 \le \kappa({\rm ASM}(6)=7436) \le 12$. Indeed, take
$w=[716983254,10,12,11] \in \mathbb{S}_{12}$. One can show that
\begin{gather*}
{\mathfrak{S}}_{w}(x_1=t, \,x_i=1, \,\forall\, i \ge 2) = 13 t^6(t+10)(15 t +37),
\end{gather*}
so that ${\mathfrak{S}}_{w}(1) =ASM(6)$; $\kappa(6^4) =9$,~namely, one can take $w=[157364298]$.
\begin{Question}\label{question-page121}
Let $N$ be a positive integer. Does there exist a vexillary
$($grassmannian?$)$ permutation $w \in \mathbb{S}_{n}$ such that
$ n \le 2 \kappa(N)$ and $\mathfrak{S}_{w}(1) = N$?
\end{Question}
For example, $w=[1,4,5,6,8,3,5,7] \in \mathbb{S}_8$ is a grassmannian
permutation such that $\mathfrak{S}_{w}(1) =140$, and
$\mathfrak{R}_{w}(1,\beta)=(1,9,27,43,38,18,4)$.
\begin{Remark} \label{remark5.3}
We expect that for $n \ge 5$ there are {\it no} permutations
$w \in {\mathbb S}_{\infty}$ such that $Q_{n}(\beta) =\mathfrak{S}_{w}^{(\beta)}(1)$.
\end{Remark}
The numbers $\mathfrak{C}_n:= \prod\limits_{j=1}^{n} {\rm Cat}_j$ appear also
as the values of the {\it Kostant partition function} of the type $A_{n-1}$ on
some special vectors. Namely,
\begin{gather*}
\mathfrak{C}_n = K_{\Phi(1^{n})}(\gamma_n), \qquad \text{where}\quad \gamma_n=
\left(1,2,3,\ldots,n-1, -{n \choose 2}\right),
\end{gather*}
see, e.g., \cite[Exercise~6.C10]{ST3}, and \cite[pp.~173--178]{K1}. More generally
\cite[Exercise~g, p.~177, (7.25)]{K1}, one has
\begin{gather*}
K_{\Phi(1^{n})}(\gamma_{n,d}) = pp^{\delta_{n}}(d) \mathfrak{C}_{n-1} =
\prod_{j=d}^{n+d-2} {1 \over 2j+1}~{n+d+j \choose 2j},
\end{gather*}
where $\gamma_{n,d}=(d+1,d+2,\ldots,d+n-1, - n(2d+n-1)/2)$,
$pp^{\delta_n}(d)$ denotes the set of reversed (weak) plane partitions
bounded
by~$d$ and contained in the shape $\delta_n=(n-1,n-2,\ldots,1)$. Clearly,
$pp^{\delta_{n}}(1)= \prod\limits_{{1 \le i < j \le n}} {i+j+1 \over
i+j-1} =C_n$, where $C_n$ is the $n$-th Catalan number\footnote{For example,~if $n=3$, there exist~$5$ reverse (weak) plane
partitions of shape $\delta_3 =(2,1)$ bounded by~$1$, namely reverse plane
partitions $\left\{
\begin{pmatrix} 0& 0\\ 0 \end{pmatrix},\,
\begin{pmatrix} 0& 0 \\ 1 \end{pmatrix},\,
\begin{pmatrix} 0& 1 \\ 0& \end{pmatrix},\,
\begin{pmatrix} 0& 1\\ 1 \end{pmatrix},\,
\begin{pmatrix} 1& 1 \\ 1 \end{pmatrix} \right\}$.}.
\begin{Conjecture} \label{conjecture5.7}
For any permutation $w \in \mathbb{S}_n$ there exists a graph $\Gamma_{w}=
(V,E)$, possibly with multiple edges, such that the reduced volume
$\widetilde{{\rm vol}} ({\cal{F}}_{\Gamma_{w}})$ of the flow polytope
${\cal{F}}_{\Gamma_{w}}$, see, e.g.,~{\rm \cite{ST2}} for a definition of the former,
is equal to $\mathfrak{S}_{w}(1)$.
\end{Conjecture}
For a family of vexillary permutations $w_{n,p}$ of the shape $\lambda =
p \delta_{n+1}$ and f\/lag $\phi=(1,2,\ldots$, $n-1,n)$ the corresponding graphs
$\Gamma_{n,p}$ have been constructed in \cite[Section~6]{Me2}. In this case
the reduced volume of the f\/low polytope ${\cal{F}}_{\Gamma_{n,p}}$ is equal
to the Fuss--Catalan number
\begin{gather*}
{1 \over 1+(n+1)p} {(n+1)(p+1) \choose n+1}=
\mathfrak{S}_{w_{n,p}}(1),
\end{gather*}
cf.\ Corollary~\ref{corollary5.2}.
\end{Comments}
\begin{Exercises}\label{exercises5.8}\quad
\begin{enumerate}\itemsep=0pt
\item[$(a)$] Show that
the polynomial $R_n(t):= t^{1-n} Q_n(t;0,0)$ is symmetric
({\it unimodal?}), and $R_n(0)= \prod\limits_{k=1}^{n-2} {\rm Cat}_{k}$.
For example,
\begin{gather*}
R_4(t)=(1+t)\big(2+t+2t^2\big), \qquad R_5(t)= 2 (5,10,13,14,13,10,5)_{t},\\
R_6(t)= 10 (2,3,2)_{t} (7,7,10,13,10,13,10,7,7)_{t}, \qquad R_7(t)=30\big(196+\cdots+196 t^{15}\big).
\end{gather*}
Note that $R_n(1) =\prod\limits_{k=1}^{n-1} {\rm Cat}_k$.
\item[$(b)$] More generally, write as before,
\begin{gather*}
Q_n(t; 0,\beta)= t^{n-2}
\sum_{k \ge 0} c_{n}^{(k)}(t) \beta^{k}.
\end{gather*}
Show that the polynomials $c_n^{(k)}(t)$ are symmetric (unimodal?) for all~$k$ and~$n$.
\item[$(c)$] Consider a reduced polynomial ${\overline{R}}_n(\{x_{ij} \})$ of the
element
\begin{gather*}
\prod_{1 \le i < j \le n \atop (i,j) \not= (n-1,n)} x_{ij} \in
{\widehat{{\rm ACYB}}}(\alpha=\beta=0)^{ab},
\end{gather*}
see Def\/inition~\ref{definition5.1}. Here we assume additionally, that all elements
$\{x_{ij} \}$ are mutually commute. Def\/ine polynomial ${\widetilde{R}}_n(q,t)$ to be the following specialization
\begin{gather*}
x_{ij} \longrightarrow 1 \quad \text{if}\quad i < j <n-1, \qquad x_{i,n-1} \longrightarrow q, \qquad
x_{i,n} \longrightarrow t, \quad \forall\, i
\end{gather*}
of the polynomial ${\overline{R}}_n(\{x_{ij} \})$ in question.
Show that polynomials ${\widetilde{R}}_n(q,t)$ are well-def\/ined, and
\begin{gather*}
{\widetilde{R}}_n(q,t) ={\widetilde{R}}_n(t,q).
\end{gather*}
\end{enumerate}
\end{Exercises}
\begin{Examples}\label{examples5.2}
\begin{gather*}
R_4(t,\beta)= (2,3,3,2)_{t}+(4,5,4)_{t} \beta + (2,2)_{t} \beta^{2} ,\\
R_5(t,\beta)= (10,20,26,28,26,20,10)_{t} +(33,61,74,74,61,33)_{t} \beta +(39,65,72,65,39)_{t} \beta^2\\
\hphantom{R_5(t,\beta)=}{} +
(19,27,27,19)_{t} \beta^3 +(3,3,3)_{t} \beta^4 , \\
R_6(t,\beta)=
(140,350,550,700,790,820,790,700,550,350,140)_{t}\\
\hphantom{R_6(t,\beta)=}{} +
(686,1640,2478,3044,3322,3322,3044,2478,1640,686)_{t} \beta \\
\hphantom{R_6(t,\beta)=}{}
+
(1370, 3106,4480,5280,5537,5280,4480, 3106,1370)_{t} \beta^2\\
\hphantom{R_6(t,\beta)=}{}
+
(1420,3017,4113,4615,4615,4113,3017,1420)_{t} \beta^3\\
\hphantom{R_6(t,\beta)=}{}
+
(800,1565,1987,2105,1987,1565,800)_{t} \beta^4 +
(230,403,465,465, 403,230)_{t} \beta^5 \\
\hphantom{R_6(t,\beta)=}{}
+
(26,39,39,39,26)_{t} \beta^6,\\
R_{6}(1,\beta)= (5880, 22340, 34009, 26330, 10809, 2196, 169)_{\beta},\\
R_7(t,\beta)=(5880,17640,32340,47040,59790,69630,76230,79530,79530,76230 ,69630, \\
\hphantom{R_7(t,\beta)=(}{}
59790,47040,32340,17640,5880)_{t} +
(39980,116510,208196,295954,368410,\\
\hphantom{R_7(t,\beta)=(}{}
420850, 452226,462648,452226,420850,368410 , 295954,208196,116510,\\
\hphantom{R_7(t,\beta)=(}{}
39980)_{t} \beta
+ (118179,333345,578812,802004,975555,1090913,1147982, \\
\hphantom{R_7(t,\beta)=(}{}
1147982,1090913, 975555, 802004,578812,333345,118179)_{t} \beta^2 \\
\hphantom{R_7(t,\beta)=}{}
+ (198519,539551,906940,1221060,1447565,1580835,1624550,1580835,\\
\hphantom{R_7(t,\beta)=(}{}
1447565,1221060,
906940,539551,198519)_{t} \beta^3 \\
\hphantom{R_7(t,\beta)=}{} +
(207712,540840,875969,1141589,1314942,1398556,1398556,1314942,\\
\hphantom{R_7(t,\beta)=(}{}
1141589,875969,
540840,207712)_{t} \beta^4 \\
\hphantom{R_7(t,\beta)=}{}
+ (139320,344910,535107,671897,749338,773900,749338,671897,535107,\\
\hphantom{R_7(t,\beta)=(}{}
344910 ,
139320)_{t} \beta^5 +(59235,137985,203527,244815,263389,263389,244815,\\
\hphantom{R_7(t,\beta)=(}{}
203527,137985,59235)_{t} \beta^6 + (15119,32635,45333,51865,53691,51865, 45333,\\
\hphantom{R_7(t,\beta)=(}{}
32635,15119)_{t} \beta^7 +
(2034,3966,5132,5532,5532,5132,3966,2034) \beta^8\\
\hphantom{R_7(t,\beta)=}{}
+
(102,170,204,204,204,170,102)_{t} \beta^9,\\
R_{7}(1,\beta)= (776160,4266900,10093580,13413490,10959216,5655044, 1817902,\\
\hphantom{R_{7}(1,\beta)= (}{}
343595,33328,1156)_{\beta}.
\end{gather*}
\end{Examples}
\subsubsection[The Chan--Robbins--M\'{e}sz\'{a}ros polytope ${\cal{P}}_{n,m}$]{The Chan--Robbins--M\'{e}sz\'{a}ros polytope $\boldsymbol{{\cal{P}}_{n,m}}$}\label{section5.3.2}
Let $m \ge 0$ and $n \ge 2$ be integers, consider the reduced polynomial $Q_{n,m}(t, \beta)$ corresponding to the element
\begin{gather*}
M_{n.m}:=\left( \prod_{j=2}^{n} x_{1j} \right)^{m+1} \prod_{j=2}^{n-2}
\prod_{k=j+2}^{n} x_{jk}.
\end{gather*}
For example,
\begin{gather*}
Q_{2,4}(t,\beta)= (4,7,9,10,10,9,7,4)_{t} + (10,17,21,22,21,17,10)_{t} \beta \\
\hphantom{Q_{2,4}(t,\beta)=}{}
+(8,13,15,15,13,8 )_{t} \beta^2 + (2,3,3,3,2)_{t} \beta^3 , \\
Q_{2,4}(1,\beta)=
(60, 118, 72, 13)_{\beta} ,\\
Q_{2,5}(t,\beta)=(60,144,228,298,348,378,388,378,348,298,228,144,60)_{t} \\
\hphantom{Q_{2,5}(t,\beta)=}{}
+(262,614,948,1208,1378,1462,1462,1378,1208,948,614,262)_{t} \beta \\
\hphantom{Q_{2,5}(t,\beta)=}{}
+(458,1042,1560,1930,2142,2211,2142,1930,1560,1042,458)_{t} \beta^2 \\
\hphantom{Q_{2,5}(t,\beta)=}{}
+(405,887,1278,1526,1640,1640,1526,1278,887,405)_{t} \beta^4 \\
\hphantom{Q_{2,5}(t,\beta)=}{}
+(187,389,534,610,632,610,534,389,187)_{t} \beta^4 \\
\hphantom{Q_{2,5}(t,\beta)=}{}
+(41,79,102,110,110,102,79,41)_{t} \beta^5
+(3,5,6,6,6,5,3)_t \beta^6 , \\
Q_{2,5}(1,\beta) = (3300, 11744, 16475, 11472, 4072, 664, 34)_{\beta} , \\
Q_{2,6}(1,\beta)= (660660,3626584,8574762,11407812,9355194,4866708, 1589799 , \\
\hphantom{Q_{2,6}(1,\beta)= (}{}
310172, 32182, 1320)_{\beta} ,\\
Q_{2,7}(\beta)= (1,213,12145, 279189, 3102220,
18400252, 61726264, 120846096, 139463706, \\
\hphantom{Q_{2,7}(\beta)= (}{}
93866194, 5567810, 7053370, 626730, 16290 )_{\beta+1}.
\end{gather*}
\begin{Theorem}\label{theorem5.13}
One has
\begin{gather*}
Q_{m,n}(1,1) = \prod_{k=1}^{n-2} {\rm Cat}_{k} \prod_{1\le i < j \le n-1}
\frac{2(m+1)+i+j-1}{i+j-1},\\
\sum_{k \ge 0} \iota ({\cal{P}}_{n,m};k) \beta^k = \frac{Q_{m,n}(1,
\beta-1)}{(1-\beta)^{{n+1 \choose 2}+1}},
\end{gather*}
where $ {\cal{P}}_{n,m}$ denotes the generalized Chan--Robbins--Yuen polytope
defined in {\rm \cite{Me2}}, and for any integral convex polytope~$\cal{P}$,
$\iota({\cal{P}},k)$ denotes the Ehrhart polynomial of polytope~$\cal{P}$.
\end{Theorem}
\begin{Conjecture}\label{conjecture5.8} Let $n \ge 3$, $m \ge 0$ be integers, , write
\begin{gather*}
Q_{m,n}(t,\beta)=\sum_{k \ge 0} c_{m,n}^{(k)}(t) \beta^k, \qquad \text{and set}\qquad b(m,n):=
\max\big( k \,|\, c_{m,n}^{(k)}(t) \not= 0 \big).
\end{gather*}
Denote by ${\tilde{c}}_{m,n}(t)$ the polynomial obtained from that $c_{m,n}^{(b(m,n)}(t)$
by dividing the all coefficients of the latter on their GCD. Then
\begin{gather*}
{\tilde{c}}_{n,m}(t)=a_{n+m}(t),
\end{gather*}
where the polynomials $a_n(t):=c_{0,n}(t)$ have been defined in Conjecture~{\rm \ref{conjecture5.6}}.
\end{Conjecture}
For example,
\begin{gather*}
c_{2,5}(t)=4 a_7(t), \qquad c_{2,6}(t) = 10 a_8(t), \qquad c_{3,5}(t)= a_8(t), \\
c_{2,7}(t)= 10 (34,78,118,148,168,
178,181,178,168,148,118,78,34) \stackrel{?}{=} 10 a_9(t).
\end{gather*}
It is known \cite{K1, Me,Me-b} that
\begin{gather*}
\prod_{k=1}^{n-2} {\rm Cat}_{k} \prod_{1\le i < j \le n-1}
\frac{2(m+1)+i+j-1}{i+j-1}=\prod_{j=m+1}^{m+n-2} \frac{1}{2j+1} {n+m+j \choose 2 j}\\
\qquad{}
=
K_{A_{n+1}}\left(m+1,m+2,\ldots,n+m,-m n- {n \choose 2}\right) .
\end{gather*}
\begin{Conjecture}\label{conjecture5.9}
Let $ {\boldsymbol{a}}=(a_2,a_3,\ldots,a_n)$ be a sequence of non-negative
integers, consider the following element
\begin{gather*}
M_{({\boldsymbol{a}})} = \left( \prod_{j=2}^{n} x_{1j}^{a_{j}} \right)
\prod_{j=2}^{n-1} \left( \prod_{k=j+1}^{n} x_{jk} \right) .
\end{gather*}
Let $R_{{\boldsymbol{a}}}(t_1,\ldots,t_{n-1}, \alpha,\beta)$ be the following specialization
$x_{ij} \longrightarrow t_{j-1}$ for all $1 \le i < j \le n $
of the reduced polynomial $R_{{\boldsymbol{a}}}(x_{ij}) $ of monomial $M_{{\boldsymbol{a}}} \in
{ \widehat{{\rm ACYB}}}_n(\alpha,\beta)$.
Then the polynomial $R_{{\boldsymbol{a}}}(t_1,\ldots,t_{n-1}, \alpha,\beta)$ is well-defined, i.e., does not depend on an order in
which relations~$(a')$ and~$(b)$, Definition~{\rm \ref{definition5.1}}, have been applied.
\begin{gather*}
Q_{M_{\boldsymbol{a}}}(1,\beta=0)= K_{A_{n+1}}\left(a_{2}+1,a_{3}+2,\ldots,a_{n}+n-1,- {n \choose 2} - \sum_{j=2}^{n} a_j\right) .
\end{gather*}
Write
\begin{gather*}
Q_{M_{\boldsymbol{a}}}(t,\beta) =\sum_{k \ge 0} c_{ {\boldsymbol{a}}}^{(k)}(t) \beta^k.
\end{gather*}
The polynomials $c_{ {\boldsymbol{a}}}^{(k)}(t)$ are symmetric $($unimodal?$)$ for all~$k$.
\end{Conjecture}
\begin{Example} \label{example5.8}
Let's take $n=5$, ${\boldsymbol{a}}=(2,1,1,0)$. One can show that the value
of the Kostant partition function $K_{A_{5}}(3,3,4,4,-14)$ is equal to~$1967$. On the other hand, one has
\begin{gather*}
Q_{(2,1,1,0)}(t,\beta) t^{-3}= (50,118,183,233,263,273,263,233,183,118,
50)_{t} \\
\hphantom{Q_{(2,1,1,0)}(t,\beta) t^{-3}=}{}
+ (214,491,738,908,992,992,908,738,491,214)_{t} \beta\\
\hphantom{Q_{(2,1,1,0)}(t,\beta) t^{-3}=}{}
+ (365,808,1167,1379,1448, 1379 , 1167,808,365)_{t} \beta^2 \\
\hphantom{Q_{(2,1,1,0)}(t,\beta) t^{-3}=}{}
+ (313,661,906,1020,1020,906,661,313)_{t} \beta^3\\
\hphantom{Q_{(2,1,1,0)}(t,\beta) t^{-3}=}{}
+
(139,275,351,373,351,275,139)_{t} \beta^4\\
\hphantom{Q_{(2,1,1,0)}(t,\beta) t^{-3}=}{}
+(29,52,60,60,52,29)_{t} \beta^5 + (2,3,3,3,2)_{t} \beta^6 , \\
Q_{(2,1,1,0)}(1,\beta) = (1967,6686,8886,5800,1903,282,13) = (1,34,279,748,
688,204,13)_{\beta+1}.
\end{gather*}
It might be well to point out that since we know, see Theorem~\ref{theorem5.11}, that polynomials $Q_{M_{\boldsymbol{a}}}(1,\beta)$ in face are polynomials of~$\beta+1$
with non-negative integer coef\/f\/icients, we can treat the polynomial
${\widetilde{Q}}_{M_{\boldsymbol{a}}}(\beta):= Q_{M_{\boldsymbol{a}}}(1,\beta-1)$ as a~$\beta$-analogue of the Kostant partition function in the dominant chamber. It seems an interesting {\it problem} to f\/ind an interpretation of polynomials
${\widetilde{Q}}_{M_{\boldsymbol{a}}}(\beta)$ in the framework of the representation theory of Lie algebras. For example,
\begin{gather*}
{\widetilde{Q}}_{(2,1,1,0)}(\beta) = (1,34,279,748,688,204,13)_{\beta},\\
{\widetilde{Q}}_{(2,1,1,0)}(\beta =1)= 1967 = K_{A_{5}}(3,3,4,4,-14).
\end{gather*}
\end{Example}
\begin{Exercises} \label{exercises5.9}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] Show that
\begin{gather*}
R_{n}(t,-1) = t^{2(n-2)} R_{n-1}\big({-}t^{-1}, 1\big).
\end{gather*}
\item[(2)] Show that the ratio
\begin{gather*}
\frac{R_n(0,\beta)}{(1+\beta)^{n-2}}
\end{gather*}
is a polynomial in $(\beta+1)$~with non-negative coef\/f\/icients.
\item[(3)] Show that polynomial $R_n(t,1)$ has degree $e_n:= (n+1)(n-2)/2$, and
\begin{gather*}
{\rm Coef\/f}[t^{e_{n}}] R_n(t,1) =\prod_{k=1}^{n-1} {\rm Cat}_k .
\end{gather*}
\item[(4)] Show that
\begin{gather*}
{\widetilde{Q}}_{(n,2,3,0)}(\beta)= \left(1, 3 n+2, \binom{n+1}{2}+n, \binom{n+1}{3} +\binom{n}{2} \right)_{\beta},
\end{gather*}
and
\begin{gather*}
K_{A_{4}} (n,3,4,-n-7) = \frac{(n+2)(n+3)(n+9)}{6}.
\end{gather*}
\end{enumerate}
\end{Exercises}
\begin{Problems}\label{problems5.2}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] Assume additionally to the conditions $(a')$ and $(b)$ above that
\begin{gather*}
x_{ij}^2= \beta x_{ij} + 1 \qquad \text{if} \quad 1 \le i < j \le n.
\end{gather*}
What one can say about a reduced form of the element $w_{0}$ in this case?
\item[(2)] According to a result by S.~Matsumoto and J.~Novak~{\rm \cite{MN}}, if
$\pi \in {\mathbb S}_n$ is a permutation of the cyclic type $\lambda \vdash n$,
then the total number of primitive factorizations $($see definition in~{\rm \cite{MN})}
of~$\pi$ into product of $n-\ell (\lambda)$ transpositions, denoted by
$\operatorname{Prim}_{n-\ell (\lambda)}(\lambda)$, is equal to the product of Catalan numbers:
\begin{gather*}
\operatorname{Prim}_{n-\ell (\lambda)}(\lambda) = \prod_{i=1}^{\ell(\lambda)} {\rm Cat}_{\lambda_{i}-1}.
\end{gather*}
Recall that the Catalan number ${\rm Cat}_n:= C_n={1 \over n} {2n \choose n}$.
Now take $\lambda=(2,3,\ldots,n+1)$. Then
\begin{gather*}
Q_n(1)= \prod_{a=1}^{n} {\rm Cat}_a=\operatorname{Prim}_{{n \choose 2}}(\lambda).
\end{gather*}
Does there exist ``a~natural'' bijection between the primitive factorizations
and monomials which appear in the polynomial $Q_n(x_{ij};\beta)$?
\item[$(3)$] Compute in the algebra ${\widehat{{\rm ACYB}}}_n(\alpha,\beta)$ the
specialization
\begin{gather*}
x_{ij} \longrightarrow 1, \quad j < n,\qquad
x_{ij} \longrightarrow t,\quad 1 \le i < n,
\end{gather*}
denoted by $P_{w{_n}}(t,\alpha,\beta)$,
of the reduced polynomial $P_{s_{ij}}(\{x_{ij} \},\alpha, \beta)$
corresponding to the transposition
\begin{gather*}
s_{ij}:= \left( \prod_{k=i}^{j-2}
x_{k,k+1} \right) x_{j-1,j} \left( \prod_{k=j-2}^{i} x_{k,k+1} \right) \in
{\widehat{{\rm ACYB}}}_n(\alpha,\beta).
\end{gather*}
For example,
\begin{gather*}
P_{s_{14}}(t,\alpha,\beta)= t^{5}+3(1+\beta) t^{4}+
((3,5,2)_{\beta}+3 \alpha) t^{3}+ (2(1+\beta)^2+\alpha(5+4 \beta)) t^{2}\\
\hphantom{P_{s_{14}}(t,\alpha,\beta)=}{}+
((1+\beta((1+3 \alpha)+2 \alpha^{2}) t +\alpha + \alpha^{2}.
\end{gather*}
\end{enumerate}
\end{Problems}
\subsection{Reduced polynomials of certain monomials}\label{section5.4}
In this subsection we compute the reduced polynomials corresponding to {\it
dominant} monomials of the form
\begin{gather*}
x_{{\boldsymbol{m}}}:= x_{1,2}^{m_1} x_{23}^{m_2} \cdots x_{n-1,n}^{m_{n-1}}
\in \big({\widehat{{\rm ACYB}}}_n(\beta)\big)^{ab},
\end{gather*}
where ${\boldsymbol{m}} =(m_1 \ge m_2 \ge \cdots \ge m_{n-1} \ge 0)$
is a {\it partition}, and we apply the relations~$(a')$ and~$(b)$ in
the algebra $({\widehat{{\rm ACYB}}}_n(\beta))^{ab}$, see Def\/inition~\ref{definition5.1} and
Section~\ref{section5.3.1}, {\it successively}, starting from $x_{12}^{m_1}
x_{23}$.
\begin{Proposition} \label{proposition5.9} The function
\begin{gather*}
\mathbb{Z}_{\ge 0}^{n-1} \longrightarrow \mathbb{Z}_{\ge 0}^{n-1}, \qquad {\boldsymbol{m}}
\longrightarrow P_{{\boldsymbol{m}}}(t=1;\beta=1)
\end{gather*}
can be extended to a piece-wise polynomial function on the space
$\mathbb{R}_{\ge 0}^{n-1}$.
\end{Proposition}
We start with the study of powers of Coxeter elements. Namely, for powers
of {\it Coxeter} elements, one has\footnote{To simplify notation we set $P_{w}(\beta) := P_{w}(x_{ij}=1 ;\beta)$.}
\begin{gather*}
P_{(x_{12} x_{23})^2}(\beta)=(6,6,1), \qquad P_{(x_{12} x_{23} x_{34})^2}(\beta)=
(71,142,91,20,1) = (1,16,37,16,1)_{\beta+1} , \\
P_{(x_{12} x_{23} x_{34})^3}(\beta) = (1301,3903,4407,2309,555,51,1)= (1,45,315,579,315,45,1)_{\beta+1} , \\
P_{(x_{12} x_{23} x_{34} x_{45})^2}(\beta)=(1266,3798,4289,2248,541,50,1)=
(1,44,306,564,306,44,1)_{\beta+1} , \\
P_{(x_{12} x_{23} x_{34})^3}(\beta=1)=
12527 , \qquad P_{(x_{12} x_{23} x_{34})^4}(\beta=0)= 26599 , \\
P_{(x_{12} x_{23} x_{34})^4}(\beta=1)= 539601 ,\qquad
P_{(x_{12} x_{23} x_{34} x_{45})^2}(\beta=1)= 12193 , \\
P_{(x_{12} x_{23} x_{34} x_{45})^3}(\beta=0)=50000 ,\qquad
P_{(x_{12} x_{23} x_{34} x_{45})^3}(\beta=1)= 1090199 .
\end{gather*}
\begin{Lemma} \label{lemma5.3}
One has
\begin{gather*}
P_{x_{12}^n x_{23}^m}(\beta)= \sum_{k=0}^{\min(n,m)} {n+m-k
\choose m} {m \choose k} \beta^k = \sum_{k=0}^{\min(n,m)} {n \choose k}
{m \choose k} (1 +\beta)^k.
\end{gather*}
Moreover,
\begin{itemize}\itemsep=0pt
\item polynomial $P_{(x_{12} x_{23} \cdots x_{n-1,n})^m}(\beta-1)$
is a symmetric polynomial in $\beta$ with
non-negative coefficients.
\item polynomial $P_{x_{12}^{n} x_{23}^{m}}(\beta)$ counts the number
of $(n,m)$-Delannoy paths according to the number of $NE$ steps\footnote{Recall that a $(n,m)$-Delannoy path is a lattice paths from $(0,0)$ to
$(n,m)$ with steps $E=(1,0)$,
$N=(0,1)$ and $NE=(1,1)$ only.
For the def\/inition and examples of the
Delannoy paths and numbers, see \cite[$A001850$, $A008288$]{SL} and
\url{http://mathworld.wolfram.com/DelannoyNumber.html}. }.
\end{itemize}
\end{Lemma}
\begin{Proposition} \label{proposition5.10} Let $n$ and $k$, $0 \le k \le n$, be integers. The number
\begin{gather*}
P_{(x_{12} x_{23})^n (x_{34})^k}(\beta=0)
\end{gather*}
is equal to the number of~$ n$ up, $n$ down permutations in the
symmetric group ${\mathbb{S}}_{2n+k+1}$, see~{\rm \cite[$A229892$]{SL}} and
Exercises~{\rm \ref{exercises5.3}(2)}.
\end{Proposition}
\begin{Conjecture}\label{conjecture5.10} Let $n$, $m$, $k$ be nonnegative integers. Then the number
\begin{gather*}
P_{x_{12}^n x_{23}^m x_{34}^k}(\beta=0)
\end{gather*}
is equal to the number of~$n$ up, $m$ down and~$k$ up permutations in the
symmetric group~$\mathbb{S}_{n+m+k+1}$.
\end{Conjecture}
For example,
\begin{itemize}\itemsep=0pt
\item Take $n=2$, $k=0$, the six permutations in $\mathbb{S}_5$ with
$2$ up, $2$ down are {12543}, {13542}, {14532}, {23541}, {24531}, {34521}.
\item Take $n=3$, $k=1$, the twenty permutations in $\mathbb{S}_7$ with
$3$ up, $3$ down are {1237654}, {1247653}, {1257643}, {1267543}, {1347652}, {1357642},
{1367542}, {1457632}, {1467532}, {1567432}, {2347651}, \linebreak {2357641},
{2367541}, {2457631}, {2467531}, {2567431}, {3457621}, {3467521},
{3567421}, {4567321}, see \cite[$A229892$]{SL}.
\item Take $n=3$, $m=2$, $k=1$, the number of $3$ up, $2$ down and~$1$ up
permutations in~$\mathbb{S}_7$ is equal to $50=P_{321}(0)$: {1237645},
{1237546}, \dots, {4567312}.
\item Take $n=1$, $m=3$, $k=2$, the number of~$1$ up, $3$ down and~$2$ up
permutations in $\mathbb{S}_7$ is equal to $55=P_{132}(0)$, as it can be easily checked.
\end{itemize}
On the other hand, $P_{x_{12}^4 x_{23}^3 x_{34}^2 x_{45}}(\beta=0) =7203
< 7910$, where $7910$ is the number of $4$ up, $3$~down, $2$ up and $1$ down
permutations in the symmetric group $\mathbb{S}_{11}$.
\begin{Conjecture}\label{conjecture5.11}
Let $k_1,\ldots,k_{n-1}$ be a sequence of non-negative integer
numbers, consider monomial $M:=x_{12}^{k_{1}} x_{23}^{k_2}
\cdots x_{n-1,n}^{k_{n-1}}$. Then
reduced polynomial $P_{M}(\beta -1)$ is a unimodal
polynomial in $\beta$ with non-negative coefficients.
\end{Conjecture}
\begin{Example} \label{example5.9}
\begin{gather*}
P_{3,2,1}(\beta)= (1,14,27,8)_{\beta +1}=P_{1,2,3}(\beta) ,\qquad
P_{2,3,1}(\beta)= (1,15,30,9)_{\beta +1}=P_{1,3,2}(\beta) , \\
P_{3,1,2}(\beta)= (1,11,18,4)_{\beta +1}=P_{2,1,3}(\beta) , \\
P_{4,3,2,1}(\beta)= (1,74,837,2630,2708,885,68)_{\beta +1} , \qquad
P_{4,3,2,1}(0)= 7203 = 3 \cdot 7^4 , \\ P_{5,4,3,2,1}(\beta)=
(1,394,19177,270210,1485163,3638790,
4198361,2282942,\\
\hphantom{P_{5,4,3,2,1}(\beta)=(}{} 553828,51945,1300)_{\beta +1} ,\\
P_{5,4,3,2,1}(0)= 12502111 = 1019 \times 12269 .
\end{gather*}
\end{Example}
\begin{Exercises}\label{exercises5.10}\quad
(1) Show that if $ n \ge m$, then
\begin{gather*}
x_{ij}^{n} x_{jk}^{m} \Big{\arrowvert}_{x_{ij}=1=x_{jk}} = \sum_{a=0}^{n}
{m+a-1 \choose a} \left( \sum_{p=0}^{n-a} {m \choose p} \beta^{p} \right)
x_{ik}^{m+a} .
\end{gather*}
(2) Show that if $ n \ge m \ge k$, then
\begin{gather*}
P_{x_{12}^{n} x_{23}^{m} x_{34}^{k}}(\beta)=
P_{x_{12}^{n} x_{23}^{m}}(\beta)\\
\hphantom{P_{x_{12}^{n} x_{23}^{m} x_{34}^{k}}(\beta)=}{}+
\sum_{ a \ge 1 \atop b,p \ge 0} {m \choose p} {k \choose a} {a-1 \choose b} {n+1 \choose p+a-b} {m+a-1-b \choose a} (\beta+1)^{p+a}.
\end{gather*}
In particular, if $n \ge m \ge k$, then
\begin{gather*}
P_{x_{12}^{n} x_{23}^{m} x_{34}^{k}}(0)= {m+n \choose n}+
\sum_{a \ge 1} {k \choose a} \left(\sum_{b=1}^{a}
{m+n+1 \choose m+b} {a-1 \choose b-1} {m+b-1 \choose a} \right) .
\end{gather*}
Note that the set of relations from the item $(1)$ allows to give an
explicit formula for the polynomial $P_{M}(\beta)$ for any {\it dominant}
sequence $M=(m_1 \ge m_2 \ge \cdots \ge m_{k}) \in (\mathbb{Z}_{ > 0})^{k}$.
Namely,
\begin{gather*}
P_{M}(\beta+1) =
\sum_{\boldsymbol{a}} \prod_{j=2}^{k} {m_j+a_{j-1}-1 \choose a_{j-1} } \left(
\sum_{\boldsymbol{b}} \prod_{j=1}^{k-1} {m_{j+1} \choose b_{j} } \beta^{b_{j}} \right),
\end{gather*}
where the f\/irst sum runs over the following set ${\cal A}(M)$ of integer
sequences $ {\boldsymbol{a}} = (a_1,\ldots, a_{k-1}) $
\begin{gather*}
{\cal{A}}(M):= \{ 0 \le a_{j} \le m_{j}+a_{j-1}, \, j=1,\ldots, k-1 \},\qquad
a_{0}=0,
\end{gather*}
and the second sum runs over the set ${\cal B}(M)$ of all integer sequences
${\boldsymbol{b}} =(b_1,\ldots,b_{k-1})$
\begin{gather*}
{\cal B}(M) := \bigcup_{{\boldsymbol{a}} \in {\cal A}(M)} \{ 0 \le b_j \le
\min (m_{j+1},m_j-a_j+a_{j-1}) \}, \qquad j=1,\ldots,k-1.
\end{gather*}
(3) Show that
\begin{gather*}
\# \big|{\cal A}\big(n,1^{k-1}\big) \big| = {n+1 \over k}{2k+n \choose k-1} = f^{(n+k,k)},
\end{gather*}
where $f^{(n+k,k)}$ denotes the number of standard Young tableaux of shape
$(n+k,k)$. In particular, $ \#|{\cal A}(1^k)| = C_{k+1}$.
(4) Let $n \ge m \ge 1$ be integers and set $M=(n,m,1^{k})$.
Show that
\begin{gather*}
P_{M}(x_{ij}=1 ; \beta=0)= \sum_{p=0}^{n} {m+p+1 \over k}~{m+p-1 \choose p}
{m+2k+p \choose k-1} := P_k(n,m).
\end{gather*}
In particular, $P_{1}(n,m)= {n+m \choose n} +m {n+m+1 \choose n}$,
\begin{gather*}
P_{k}(n,1)={n+1 \over k+1} {2k+2+n \choose k} ,\qquad
P_{k}(2,2)= \big(79k^2+341k+360\big){(2k+2) ! \over k! (k+5)! }.
\end{gather*}
Let us remark that
\begin{gather*}
P_k(n,1)= {\frac{n+1}{n+k+2}} {\binom{2(k+1)+n}{k+1}}=F_{k+1}^{(2)}(n)=
D(k,1,n,2),
\end{gather*}
where the $D(k,1,n,2)$ and $F_{k+1}^{(2)}(n)$ are def\/ined in Section~\ref{section5.2.4}.
(5) Let $T \in {\rm STY}((n+k,k))$ be a standard Young tableau of shape
$(n+k,k)$.
Denote by~$r(T)$ the number of integers $j \in [1,n+k]$ such that the integer
$j$ belongs to the second row of tableau~$T$, whereas the number $j+1$ belongs
to the f\/irst row of~$T$.
Show that
\begin{gather*}
P_{x_{12}^{n} x_{23} \cdots x_{k+1,k+2}}(\beta -1) = \sum_{ T \in {\rm STY}((n+k,k))} \beta^{r(T)}.
\end{gather*}
(6) Let $M =(m_1,m_2,\ldots,m_{k-1}) \in \mathbb{Z}_{ > 0 }^{k-1}$ be a
composition. Denote by $\overleftarrow{M}$ the composition
$(m_{k-1},m_{k-2},\ldots,m_2,m_1)$, and set for short
\begin{gather*}
P_{M}(\beta):=
P_{\prod _{i=1}^{k-1} x_{i,i+1}^{m_i}}(x_{ij}=1; \beta).
\end{gather*}
Show that
$P_M(\beta)=P_{\overleftarrow{M}}(\beta)$.
Note that in general,
\begin{gather*}
P_{\prod\limits_{i=1}^{k-1}
x_{i,i+1}^{{m_i}}} (x_{ij};\beta) \not=
P_{\prod\limits_{i=1}^{k-1} x_{i,i+1}^{m_{k-i}}}(x_{ij}; \beta).
\end{gather*}
(7) Def\/ine polynomial $P_{M}(t,\beta)$ to be the following specialization
\begin{gather*}
x_{ij} \longrightarrow 1, \quad i < j <n, \qquad x_{in} \longrightarrow t, \quad i=1,\ldots,n-1
\end{gather*}
of a polynomial $P_{\prod\limits_{i=1}^{k-1} x_{i,i+1}^{m_i}}(x_{ij}; \beta)$.
Show that if $ n \ge m$, then
\begin{gather*}
P_{x_{12}^{n} x_{23}^m}(t,\beta)= \sum_{j=0}^{m} {m \choose j} \left(
\sum_{k=m-1}^{n+m-j-1} {k \choose m-1} t^{k-m+1} \right) \beta^j.
\end{gather*}
See Lemma~\ref{lemma5.2} for the case $t=1$.
(8) Def\/ine polynomials ${\widetilde{R}}_n(t)$ as follows
\begin{gather*}
{\widetilde{R}}_n(t) :=P_{(x_{12} x_{23} x_{34})^n} \big({-}t^{-1}, \beta= -1\big) (-t)^{3n}.
\end{gather*}
Show that polynomials ${\widetilde{R}}_n(t)$ have non-negative
coef\/f\/icients, and
\begin{gather*}
{\widetilde{R}}_n(0) =\frac{(3n)!}{6(n!)^3} .
\end{gather*}
$(9)$ Consider reduced polynomial $P_{n,2,2}(\beta)$ corresponding to
monomial $x_{12}^n (x_{23}x_{34})^2$ and
set ${\tilde{P}}_{n,2,2}(\beta):=P_{n,2,2}(\beta -1)$. Show
that
\begin{gather*}
{\tilde{P}}_{n,2,2}(\beta) \in \mathbb{N} [\beta] \qquad \text{and}\qquad
{\tilde{P}}_{n,2,2}(1) =T(n+5,3),
\end{gather*}
where the numbers $T(n,k)$ are def\/ined in \cite[$A110952$, $A001701$]{SL}.
\end{Exercises}
\begin{Conjecture} \label{conjecture5.12}
Let $\lambda$ be a partition. The element
$s_{\lambda} (\theta_1^{(n)},\ldots,\theta_{m}^{(n)})$ of the algebra
$3T_n^{(0)}$ can be written in this algebra as a sum of
\begin{gather*}
\left( \prod_{x \in \lambda} h(x) \right) \times
\dim {V_{\lambda'}}^{({\mathfrak {gl}}(n-m))} \times
\dim {V_{\lambda}}^{({\mathfrak {gl}}(m))}
\end{gather*}
monomials with all coefficients are equal to~$1$.
Here $s_{\lambda}(x_1,\ldots,x_m)$ denotes the Schur function corresponding
to the partition $\lambda$ and the set of variables $\{x_1,\ldots,x_m \}$;
for $x \in \lambda, $ $h(x)$ denotes the hook length corresponding to a~box~$x$; $V_{\lambda}^{({\mathfrak {gl}}(n))}$ denotes the
highest weight~$\lambda$ irreducible representation of the Lie algebra
${\mathfrak {gl}}(n)$.
\end{Conjecture}
\begin{Problems}\label{problems5.3}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] Define a bijection between monomials of the form
$\prod\limits_{a=1}^{s} x_{i_a,j_a}$ involved in the polynomial $P(x_{ij};\beta)$,
and dissections of a convex $(n+2)$-gon by $s$ diagonals, such that no two
diagonals intersect their interior.
\item[$(2)$] Describe permutations $w \in \mathbb{S}_n$ such that the
Grothendieck polynomial $\mathfrak{G}_{w}(t_1,\ldots,t_n)$ is equal to
the ``reduced polynomial'' for a some monomial in the associative
quasi-classical Yang--Baxter algebra $\widehat{{\rm ACYB}}_n(\beta)$.
\item[$(3)$] Study ``reduced polynomials'' corresponding to the monomials
\begin{itemize}\itemsep=0pt
\item transposition: $s_{1n}:= (x_{12}x_{23}
\cdots x_{n-2,n-1})^2 x_{n-1,n}$,
\item powers of the Coxeter element: $(x_{12}x_{23}\cdots
x_{n-1,n})^{k}$,
\end{itemize}
in the algebra $\widehat{{\rm ACYB}}_n(\alpha,\beta)^{ab}$.
\item[$(4)$] Construct a bijection between the set of $k$-dissections
of a convex $(n+k+1)$-gon and ``pipe dreams'' corresponding to the
Grothendieck polynomial
$\mathfrak{G}_{\pi_{k}^{(n)}}^{(\beta)}(x_1,\ldots,x_n)$.
As for a~de\-finition of ``pipe dreams'' for Grothendieck polynomials, see
{\rm \cite{KMY}} and {\rm \cite{FK1}}.
\end{enumerate}
\end{Problems}
\begin{Comments} \label{comments5.9}
We don't know any ``good'' combinatorial interpretation of polynomials which
appear in Problem~\ref{problems5.3}(3) for general~$n$ and~$k$. For example,
\begin{gather*}
P_{s_{13}}(x_{ij}=1;\beta)=(3,2)_{\beta},\qquad
P_{s_{14}}(x_{ij}=1;\beta)=(26,42,19,2)_{\beta}, \\
P_{s_{15}}(x_{ij}=1;\beta)=(381,988,917,362,55,2)_{\beta},\qquad
P_{s_{15}}(x_{ij}=1;1)= 2705.
\end{gather*}
On the other hand,
\begin{gather*}
P_{(x_{12}x_{23})^2 x_{34} (x_{45})^2}(x_{ij}=1;\beta)=(252,633,565,212,30,1),
\end{gather*}
that is in deciding on dif\/ferent reduced decompositions of the
transposition $s_{1n}$. one obtains in general dif\/ferent reduced polynomials.
One can compare these formulas for polynomials $P_{s_{ab}}(x_{ij}=1;\beta)$
with those for the $\beta$-Grothendieck polynomials corresponding to
transpositions $(a,b)$, see Comments~\ref{comments5.5}.
\end{Comments}
\subsubsection{Reduced polynomials, Motzkin and Riordan numbers}\label{section5.4.1}
In this subsection we investigate reduced polynomials associated with Coxeter
element $C_n=u_{12} u_{23} \cdots u_{n-1,n}$ in commutative algebra
${\widehat{{\rm ACYB}}}_n(\alpha, \beta)$ in more detail. Recall that this algebra
is generated over the ring $\mathbb{Z}[z,\alpha, \beta]$ by the set of elements
$\{u_{i,j}, \, 1 \le i < j \le n \}$ subject to the following relations
\begin{gather*}
u_{ij} u_{jk} = u_{ik} u_{ij}+u_{jk} u_{ik} + \beta u_{ik} + \alpha, \qquad i < j <k .
\end{gather*}
Show that
\begin{gather*}
P_n(1,1,\beta=-1) = M_n,
\end{gather*}
where $M_n$ denotes the $n$-th {\it Motzkin} number that is the number of
{\it Motzkin $n$-paths}: paths from $(0,0)$ to $(n,0)$ in an $n \times n$
grid using only steps $U=(1,1)$, $(1,0)$ and $(1,-1)$. It is also the
number of Dyck $(n+1)$-paths with no steps $UUU$, see \cite[$A001006$]{SL} for
a~wide variety of combinatorial interpretations, and vast literature
concerning the Motzkin numbers. For example,
\begin{gather*}
P_7(0,1,\beta = -1)= 36+37+24+18+5+6+0+1 = 127 =M_7.
\end{gather*}
Therefore we treat the polynomials
$P_n(t,\alpha,\beta=-1)$ as the $(t,\alpha)$-Motzkin numbers. For example,
\begin{gather*}
P_7(t,\alpha,\beta=-1)=t^7+6 \alpha t^5 + 5 \alpha t^4 +(0,4,14)_{\alpha} t^3 + (0,3,21)_{\alpha} t^2 + (0,2,21,14)_{\alpha} t\\
\qquad{} +(0,1,14,21)_{\alpha} = t^7+\alpha(1,2,3,4,5,6)_{t}+ \alpha^2(14,21,21,14)_{t} + \alpha^3(21,14)_{t}.
\end{gather*}
Therefore
\begin{gather*}
P_7(t,1,\beta=-1) = 1+21 \alpha+ 70 \alpha^2 + 35 \alpha^3, \qquad
P_7(1,1,\beta=-1)=127=M_7.
\end{gather*}
Show that
\begin{gather*}
P_n(0,1,\beta=-1)= A005043(n),
\end{gather*}
known as the {\it Riordan number}, or {\it Motzkin sum}~\cite{SL}. This
number, denoted by~$MS_n$, counts the number of Motzkin paths of length $n$
with no horizontal steps at level zero; it is also equal to the number of
Dyck paths of semilenght $n$ with no {\it peaks} at odd level, see \cite[$A005043$]{SL}
for a~bit more combinatorial interpretations, and literature
concerning the Motzkin sum or Riordan numbers. For example,
\begin{gather*}
P_7(t,1,-1) =
({\bf 36},37,24,18,5,6,0,1), \qquad 36 = MS_7.
\end{gather*}
Show that the Riordan number $MS_n$ is equal to the
number of underdiagonal paths from $(0,0)$ to the line $x=n-2$, using
only steps $(1,0)$, $(0,1)$ and $NE= (2,1)$ and beginning with the step
$NE=(2,1)$. Note that the number of such paths with no steps $NE$ is equal to
the Catalan number ${\rm Cat}_{n-1}$.
Let ${\cal{MS}} = \{ n \in \mathbb{N} \,|\, n= 2^{2k}(2 r+1)-1,\, k \ge 1, \, r \ge 0 \}$ be a subset of the set of all odd integers~\cite{DS}. Show that
$(a)$ $MS_n \equiv 1$ $({\rm mod}~2)$, if either $n \equiv 0$ $({\rm mod}~2)$ or $n \in
{\cal{MS}}_n$,
$(b)$ $MS_n \equiv 0$ $({\rm mod}~2)$, if $n$ is an odd integer and $ n \notin {\cal{MS}}$.
Show that
\begin{gather*}
\frac{P_n(0,\alpha,\beta)}{\alpha}\Big|_{\alpha=0} =N_{n-1}(\beta+1),
\end{gather*}
where as before, $N_{n}(t)$ denotes the Narayana polynomial.
Let us set
\begin{gather*}
P_{n}(0,\alpha,\beta)= \sum_{k \ge 0} c_k(\beta+1) \alpha^k.
\end{gather*}
Show that polynomials $c_k(\beta+1)$, $k \ge 0$ are symmetric (unimodal?) polynomials of the variable $\beta+1$.
Show that~\cite{SL}
\begin{gather*}
P_n(1,1,0) = A052709(n+1).
\end{gather*}
Show that~\cite{SL}
\begin{gather*}
P_n(0,1,0) = A052705(n)
\end{gather*}
that is the number of underdiagonal paths from $(0,0)$ to the line $x=n-2$,
using only steps $R=(1,0)$, $V=(0,1)$ and $NE=(2,1)$.
For example,
\begin{gather*}
P_7(0,10) = {\bf 36}+106+120+64+15+1= 342 =A052705(7).
\end{gather*}
Show that~\cite{SL}
\begin{gather*}
{\frac{\partial}{\partial \alpha}} P_n(t,\alpha,\beta) \Big|_{{\alpha=0,\atop \beta=0,} \atop t=1} = A05775(n-1),
\end{gather*}
that is the number of paths in the half-plane $x \ge 0$ from $(0,0)$ to $(n-1,2)$ or $(n-1,-3)$, and consisting of steps $U=(1,1)$, $D=(1,-1)$ and $H=(1,0)$.
For example,
\begin{gather*}
{\rm l.h.s.} =106+130+99+48+5+6 =427 =A05775(6).
\end{gather*}
Let us set
\begin{gather*}
P_n(t,\alpha, \beta=1):= \sum_{k,l \ge 0} c_{k,l}^{(n)} t^{k} \alpha^{l}.
\end{gather*}
Show that
\begin{gather*}
(a)\quad \sum_{k=1}^{n} c_{k,n-k}^{(n)} t^k \alpha^{n-k} = (t+ \alpha)^{n-1},\\
(b) \quad c_{k,n-k-1}^{(n)}= (k+1) \binom{n-1}{k+2}, \qquad 0 \le k \le n-3,\\
(c) \quad c_{1,0}^{(n)} = c_{0,0}^{(n)}+ (-1)^{n-1} , \qquad n \ge 3.
\end{gather*}
\subsubsection{Reduced polynomials, dissections and Lagrange inversion formula}\label{section5.4.2}
Let $ \{a_i,b_i,\beta_i,\alpha_i,\, 1 \le i \le n-1\} $ be a set of parameters, consider non commutative algebra generated over the ring $Z[\{a_i,b_i,\beta_i,\alpha_i \}_{1 \le i \le n-1}]$ by the set of generators $\{u_{ij},\, 1 \le i < j \le n \}$ subject to the set of relations
\begin{gather*}
u_{ij} u_{jk} =a_i u_{ik} u_{ij} + b_i u_{jk} u_{ik} + \beta u_{ik} +\alpha_i, \qquad 1 \le i < j < k \le n.
\end{gather*}
Consider reduced expression $R_n(\{u_{ij} \}_{1 \le i < j \le n})$ in the
above algebra which corresponds to the ``Coxeter element''
\begin{gather*}
C_n:= u_{12} u_{23} \cdots u_{n-1,n}.
\end{gather*}
Note that the reduced expression $R_n(\{u_{ij}\})$ is a linear combination of
noncommutative mono\-mials in the generators $\{u_{ij},\, 1 \le i < j \le n \}$
with coef\/f\/icients from the ring
\begin{gather*}
K_n:= \mathbb{Z}[\{a_i,b_i,\beta_i,\alpha_i\}_{1 \le i < n}].
\end{gather*}
Now to each monomial $U$ which appears in the reduced expression
$R_n(\{u_{ij}\})$ we associate a~{\it dissection} ${\cal D}:=
{\cal D}_{U}$ of a~convex $(n+1)$-gon as follows. First of
all let us label the vertices of a~convex $(n+1)$-gon selected, by the numbers
$n+1,n,\ldots,1$, written consequently and clockwise, starting from a f\/ixed
vertex, from here on named by $(n+1)$-vertex.
Next, let us take a monomial $ U= u_{i_{1},j_{1}} \cdots u_{i_{p}, j_{p}}$
which appears in the reduced expression $R_n(\{u_{ij} \})$ with coef\/f\/icient
$c(U) \in K_n$. We draw diagonals in a convex $(n+1)$-gon chosen which
connect vertices labeled correspondingly by numbers $i_{s}$ and $j_{s} +1$,
$s=1,\ldots, p$. It is clearly seen from the def\/ining relations in the algebra in question when being applied to the
Coxeter element above, that in fact, the diagonals we have drawn in a convex
$(n+1)$-gon selected, do not meet at interior points of our convex $(n+1)$-gon. Therefore, to each monomial $U$ which appears in the reduced polynomial
associated with the Coxeter element $C_n$ above, one can associate a~{\it dissecion} ${\cal D}:={\cal D}_{U}$ of a convex $(n+1)$-gon selected. Moreover, it
is not dif\/f\/icult to see (e.g., cf.~\cite{HN}) that there exists a natural
bijection $ U \Longleftrightarrow {\cal D}_{U}$ between monomials which
appear in the reduced expression $R_n(\{u_{ij}\})$ and the set of
dissections of a convex $(n+1)$-gon. As a corollary, to each dissection
${\cal D}:= {\cal D}_{U}$ of a~conves $(n+1)$-gon one can attache the element $c({\cal D}):= c(U) \in K_n$
which is equal to the coef\/f\/icient in front of monomial~$U$ in the reduced
expression corresponding to the Coxeter element~$C_n$.
To continue, let ${\boldsymbol{x}}=(x_1,\ldots,x_{n-1})$, $ {\boldsymbol{y}}=(y_1,\ldots,y_{n-1})$ and ${\boldsymbol{z}}=(z_1,\ldots,z_{n-1})$ be three sets of variables, and~${\cal D}$ be a~dissection of a convex $(n+1)$-gon. We associate with
dissection~${\cal D}$ a~monomial $m({\cal D}) \in K_n$ as follows
\begin{gather*}
m({\cal D }) := \prod_{k=1}^{n-1} x_{k}^{n(k)} y_{k}^{m(k)} z_{r(k)},
\end{gather*}
where $m(k):=m_{k}({\cal D})$ (resp.\ $r(k):=r_{k}({\cal D})$ and
$n(k):=n_{k}({\cal D})$) denotes the number of
(convex) $(m_{k}+2)$-gons constituent a dissection ${\cal D}$ taken (resp.\
the number of diagonals issue out of the vertex labeled by~$(n+1)$;
$n_{k}({\cal D}))$ stands for the number of (oriented) diagonals and edges which issue out of the vertex labeled by $k$, $k=1,\ldots,n$). Therefore we
associate with the reduced polynomial corresponding to the Coxeter element $u_{12}, \dots,u_{n-1,n}$ the following polynomial
\begin{gather*}
{\rm PL}_n({\boldsymbol{a}}, {\boldsymbol{b}}, {\boldsymbol{\beta}}, {\boldsymbol{\alpha}}, {\boldsymbol{x}}, {\boldsymbol{y}}, {\boldsymbol{z}})=
\sum_{\cal D} m({\cal D}) {c({\cal D})},
\end{gather*}
where the sum runs over all dissections ${\cal D}$ of a convex $(n+1)$-gon.
To begin with we set ${\boldsymbol{x}}={\bf 1}$ and consider the following specializations
\begin{gather*}
B_n({\boldsymbol{a}}, {\boldsymbol{y}}) ={\rm PL}_n({\boldsymbol{a}},{\boldsymbol{b}}={\bf 1},{\boldsymbol{\beta}}={\bf 1},{\boldsymbol{\alpha}}={\bf 0},{\boldsymbol{y}},{\boldsymbol{z}}={\bf 1}),\\
P_n({\boldsymbol{z}},{\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{\beta}}) = {\rm PL}_n({\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{\beta}}, {\boldsymbol{\alpha}} ={\bf 0}, {\boldsymbol{y}}={\bf 1},{\boldsymbol{z}}),
\end{gather*}
Show that
\begin{gather*}
B_{n-1}({\boldsymbol{a}},{\boldsymbol{y}})) = {\rm Coef\/f}_{t^{n}}
\big( z-f(t y_1,\ldots,t y_n) \big)^{ [ -1 ]},
\end{gather*}
where $f(y_1,\ldots,y_n) = \sum\limits_{k =1}^{n-1} y_k u^{k+1}$, and for any
formal power series $g(u)$, ${\frac{d}{du}} g(u) \vert_{u=0} =1$, we denote
by $g(u)^{[-1]}$ the Lagrange Inverse formal power series associated with that
$g(u)$ that is a~unique formal power series such that $g(g^{[-1]}(u)) = u =
g^{[-1]}(g(u))$.
Now let us recall the statement of Lagrange's inversion theorem. Namely, let
\begin{gather*}
f(x) =x - \sum_{k \ge 1} y_k x^{k+1}
\end{gather*}
be a formal power series. Then the inverse power series
$f^{[-1]}(u)$ is given by the following formula
\begin{gather*}
f^{[-1]}(y)= \sum_{n \ge 1} w_n u^n,
\end{gather*}
where
\begin{gather*}
w_{n} :=w_n(p_1,\ldots,p_n) = {\frac{1}{n+1}} \sum_{p_1,\ldots, p_n \ge 0 \atop \sum j p_j=n} \binom{n+\sum p_j}{n,p_1, \ldots,p_n} y_1^{p_{1}} y_2^{p_{2}} \cdots y_n^{p_{n}},
\end{gather*}
where if $N=m_1+ \cdots + m_n$, then
\begin{gather*}
\binom{N}{m_1,\ldots,m_n} = \frac{N !}{m_1 ! m_2 ! \cdots m_n !}
\end{gather*}
denotes the multinomial coef\/f\/icient.
Therefore, the coef\/f\/icient
\begin{gather*}
b_n(p_1,\ldots,p_n):={\frac{1}{n+1}} \binom{n+\sum p_j}{n,p_1, \ldots,p_n}, \qquad \sum_{j} j p_j =n
\end{gather*}
is equal to the number of dissections of a convex $(n+2)$-gon which contain
exactly $p_j$ convex $(j+2)$-gons, see, e.g.,~\cite{EE}. Equivalently, the
number $b_n(p_1,\ldots,p_n)$ is equal to the number of cells of the
associahedron ${\cal K}^{n-1}$ which are isomorphic to the cartesian product
$({\cal{K}}^{0})^{p_1} \times \cdots \times ({\cal{K}}^{n-1})^{p_n}$~\cite{JL1, JL2}.
Based on a natural and well-known bijection between the set of dissections
of a convex $(n+2)$-gon and the set of plane trees with $(n+1)$ ends and such
that the all other vertices have degree at least~$2$, see, e.g.,~\cite{ST1},
one can readily seen that the number $w_n(p_1,\ldots,p_n)$ def\/ined above under constraint $\sum_{j} j p_j= n$, is equal to the number of plane trees with $n+1$ ends and having~$p_j$ vertices of degree~$j+1$.
\begin{Example} \label{example5.10}
For short we set $B_n= {\rm PL}_n({\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{\beta}},{\boldsymbol{\alpha}}, {\boldsymbol{x}},{\boldsymbol{y}})$.
(1) Quadrangular:
\begin{gather*}
B_2= y_1^2(a_1 z_1+b_1 z_1 z_2) +y_2(\beta_1 z_1+ \alpha_1).
\end{gather*}
(2) Pentagon:
\begin{gather*}
B_3=y_1^3\big(a_1^2 z_1+a_1 b_1 z_1+a_2 b_1^2 z_1 z_2+
a_1 b_1 z_1 z_3+b_1^2 b_2 z_1 z_2 z_3\big)\\
\hphantom{B_3=}{}
+y_1 y_2\big(2 a_1 \beta_1 z_1+b_1 \beta_1 z_1+b_1^2 \beta_2 z_1 z_2+b_1 \beta_1 z_1 z_3+a_1 \alpha_1 b_1 \alpha_1+\alpha_1 z_3\big) \\
\hphantom{B_3=}{}
+y_3\big(\beta_1 \alpha_1+ \beta_1^2 z_1+b_1^2 \alpha_2 z_1\big).
\end{gather*}
(3) Hexagon:
\begin{gather*}
B_4= {y_1^4}\big(\big(a_1^3+2 a_1^2 b_1+a_1 a_2 b_1^2+a_1 b_1^2 b_2\big)
z_1\\
\hphantom{B_4=}{}
+a_1^2 b_1 b_2 z_1 z_2+a_2 b_1^3 b_2 z_1 z_2+a_1 a_3 b_1^2 z_1 z_3+a_1^2 b_1 z_1 z_4+a_1 b_1^2 z_1 z_4 +a_3 b_1^3 b_2^2 z_1 z_2 z_3\\
\hphantom{B_4=}{}
+a_2 b_1^2 b_2 z_1 z_2 z_4
+a_1 b_1^2 b_3 z_1 z_3 z_4+b_1^3 b_2^2 b_3 z_1 z_2 z_3 z_4\big)
+
{y_1^2 y_2}\big(a_1^2 \alpha_1+2 a_1 b_1 \alpha_1+a_2 b_1^2 \alpha_1\\
\hphantom{B_4=}{}
+b_1^2 b_2 \alpha_1+(3 a_1^2 b \beta_1+4 a_1 b_1 \beta_1+a_2 b_1^2 \beta_1+b_1^2+b_2 \beta_1+a_1 b_1^2 \beta_2) z_1+a_2 b_1^2 \beta_2 z_1 z_2\\
\hphantom{B_4=}{}
+b_1^3 b_2 \beta_2 z_1 z_2+a_2 b_1^3 \beta_2 z_1 z_2+a_1 b_1^2 b_3 z_1 z_3+a_1 b_1 \beta_1 z_1 z_3+a_3 b_1^2 z_1 z_3+b_1^2 \beta_1 z_1 z_4\\
\hphantom{B_4=}{}
+a_1 b_1 \beta_1 z_1 z_4+ b_1^2 b_2 \beta_3 z_1 z_2 z_3+b_1^3 b_2 \beta_2 z_1 z_2 z_4+b_1^2 b_3 \beta_1 z_1 z_3 z_4\big)+
{y_1 y_3}\big(a_1 \beta_1 \alpha_1\\
\hphantom{B_4=}{}
+2 b_1 \beta_1 \alpha_1+\big(2 a_1 \beta_1^2+2 b_1 \beta_1^2a_1 b_1^2 \alpha_3+a_2 b_1^2 \alpha_2+b_1^3 b_2 \alpha_2\big) z_1 +b_1^3 b_2 \alpha_3 z_1 z_2+b_1^3 \beta_2^2 z_1 z_2\\
\hphantom{B_4=}{}
+b_1^3 \alpha_3 z_1 z_4+b_3 \alpha_1 z_3 z_3 z_4+a_3 \alpha_1 z_3+a_1 \alpha_1 z_4+b_1 \alpha_1 z_4+\beta_1 \alpha_1 z_4\big)+
{y_2^2}\big(a_1 \beta_1 \alpha_1\\
\hphantom{B_4=}{}
+b_1^2 \beta_2 \alpha_2+\big(b_1^2\beta_1 \beta_2+a_1 \beta_1^2+a_1 \beta_1^2 \alpha_2\big) z_1 +\beta_3 \alpha_1 z_3+b_1 \beta_1 \beta_3 z_1 z_3\big)+
{y_4}(\alpha_1 \alpha_3+\beta_1^2 \alpha_1\\
\hphantom{B_4=}{}
+b_1^2 \alpha_1 \alpha_2 \big(b_1^2 \beta_1 \alpha_2+b_1^3 \beta_2 \alpha_2+\beta_1^3+ b_1^2 \beta_1 \alpha_3\big) z_1\big).
\end{gather*}
\emph{Special cases.}
Generalized {\it Schr\"{o}der} or {\it Lagrange} polynomials:
\begin{gather*}
P_n({\boldsymbol{a}}, {\boldsymbol{b}}, {\boldsymbol{\beta}}, {\boldsymbol{y}},{\boldsymbol{z}})= B_n \big |_{{\boldsymbol{\alpha}}={\bf 0}}.
\end{gather*}
For example,
\begin{gather*}
P_4({\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{y}})=
{y_1^4}\big(\big(a_1^3+2 a_1^2 b_1+a_1 a_2 b_1^2+a_1 b_1^2 b_2\big)
z_1+a_1^2 b_1 b_2 z_1 z_2+a_2 b_1^3 b_2 z_1 z_2+a_1 a_3 b_1^2 z_1 z_3\\
\hphantom{P_4({\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{y}})= }{}
+a_1^2 b_1 z_1 z_4+a_1 b_1^2 z_1 z_4 +a_3 b_1^3 b_2^2 z_1 z_2 z_3+a_2 b_1^2 b_2 z_1 z_2 z_4+a_1 b_1^2 b_3 z_1 z_3 z_4\\
\hphantom{P_4({\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{y}})= }{}
+b_1^3 b_2^2 b_3 z_1 z_2 z_3 z_4\big) +
{y_1^2 y_2}\big(\big(3 a_1^2 b \beta_1+4 a_1 b_1 \beta_1+a_2 b_1^2 \beta_1+b_1^2+b_2 \beta_1\\
\hphantom{P_4({\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{y}})= }{}
+a_1 b_1^2 \beta_2\big) z_1+a_2 b_1^2 \beta_2 z_1 z_2+b_1^3 b_2 \beta_2 z_1 z_2+a_2 b_1^3 \beta_2 z_1 z_2+a_1 b_1^2 b_3 z_1 z_3\\
\hphantom{P_4({\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{y}})= }{}
+a_1 b_1 \beta_1 z_1 z_3+a_3 b_1^2 z_1 z_3+b_1^2 \beta_1 z_1 z_4+a_1 b_1 \beta_1 z_1 z_4+ b_1^2 b_2 \beta_3 z_1 z_2 z_3\\
\hphantom{P_4({\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{y}})= }{}
+b_1^3 b_2 \beta_2 z_1 z_2 z_4+b_1^2 b_3 \beta_1 z_1 z_3 z_4\big) +
{y_1 y_3}\big(2 a_1 \beta_1^2+2 b_1 \beta_1^2 +b_1^3 \beta_2^2 z_1 z_2\\
\hphantom{P_4({\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{y}})= }{}
+b_1 \beta_1^2 z_1 z_4\big)+
{y_2^2}\big(\big(b_1^2\beta_1 \beta_2+a_1 \beta_1^2\big) z_1 + b_1 \beta_1 \beta_3 z_1 z_3\big)+
{y_4} \beta_1^3 z_1.
\end{gather*}
After the specialization $a_i=b_i=\beta_i=z_i=1$, $i=1,2,3,4$, one will obtain
\begin{gather*}
P_4({\boldsymbol{a}}={\bf 1},{\boldsymbol{b}}= {\bf 1}, {\boldsymbol{\beta}} = {\bf 1}, {\boldsymbol{y}}, {\boldsymbol{z}} ={\bf 1}) = 14 y_1^4+21 y_1^2 y_2+6 y_1 y_3+3 y_2^2+y_4.
\end{gather*}
Generalized {\it Narayana} polynomials:
\begin{gather*}
P_n({\boldsymbol{a}}, {\boldsymbol{b}}, {\boldsymbol{y}},{\boldsymbol{z}})= B_n \big |_{{\boldsymbol{\alpha}}={\bf 0} \atop {\boldsymbol{\beta}}={\bf 0}},\\
P_n({\boldsymbol{a}}, {\boldsymbol{b}}, {\boldsymbol{y}},{\boldsymbol{z}})= {y_1^4}\big(\big(a_1^3+2 a_1^2 b_1+a_1 a_2 b_1^2+a_1 b_1^2 b_2\big)
z_1+a_1^2 b_1 b_2 z_1 z_2+a_2 b_1^3 b_2 z_1 z_2\\
\hphantom{P_n({\boldsymbol{a}}, {\boldsymbol{b}}, {\boldsymbol{y}},{\boldsymbol{z}})=}{}
+a_1 a_3 b_1^2 z_1 z_3+a_1^2 b_1 z_1 z_4+a_1 b_1^2 z_1 z_4 +a_3 b_1^3 b_2^2 z_1 z_2 z_3+a_2 b_1^2 b_2 z_1 z_2 z_4\\
\hphantom{P_n({\boldsymbol{a}}, {\boldsymbol{b}}, {\boldsymbol{y}},{\boldsymbol{z}})=}{}
+a_1 b_1^2 b_3 z_1 z_3 z_4+b_1^3 b_2^2 b_3 z_1 z_2 z_3 z_4\big).
\end{gather*}
Generalized {\it Motzkin--Schr\"{o}der} polynomials:
\begin{gather*}
{\rm MS}_n({\boldsymbol{a}}, {\boldsymbol{b}}, {\boldsymbol{y}},{\boldsymbol{z}})= B_n \big |_{{\boldsymbol{a}}={\bf 0}}.
\end{gather*}
For example,
\begin{gather*}
{\rm MS}_4 ({\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{y}},{\boldsymbol{z}})= {y_1^1 y_2}
\big(a_1^2 \alpha_1 +2 a_1 b_1 \alpha_1+a_2 b_1^2 \alpha_1+b_1^2 b_2 \alpha_1\big) + {y_1 y_3} (a_1 \beta_1 \alpha_1+2 b_1 \beta_1 \alpha_1)\\
\hphantom{{\rm MS}_4 ({\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{y}},{\boldsymbol{z}})=}{}
+ {y_2^2}\big(a_1 \beta_1 \alpha_1+b_1^2 \beta_2 \alpha_2\big) + {y_4}\big( \alpha_1 \alpha_3+b_1^2 \alpha_1 \alpha_2 +\beta_1^2 \alpha_1\big).
\end{gather*}
Generalized {\it Motzkin} polynomials:
\begin{gather*}
M_n({\boldsymbol{b}}, {\boldsymbol{y}},{\boldsymbol{z}})= B_n \big |_{{\boldsymbol{a}}={\bf 0} \atop {\boldsymbol{\beta}} = {\bf 0}}.
\end{gather*}
For example,
\begin{gather*}
M_4({\boldsymbol{b}},{\boldsymbol{y}},{\boldsymbol{z}})= {y_1^4} b_1^3 b_2^2 b_3 z_1 z_2 z_3 z_4
+ {y_1^2 y_2} b_1^2 b_2 \alpha_1+ {y_1 y_3}\big(b_1^3 b_2 \alpha_2+b_1\alpha_1 z_4+b_1^3 b_2 z_1 z_3\\
\hphantom{M_4({\boldsymbol{b}},{\boldsymbol{y}},{\boldsymbol{z}})=}{}
+b_1^3 \alpha_2 z_1 z_4+b_3 \alpha_1 z_3 z_4\big)
+{y_4}\big(\alpha_2 \alpha_3+b_1^2 \alpha_1 \alpha_2\big).
\end{gather*}
Generalized {\it Motzkin--Riordan} polynomials:
\begin{gather*}
{\rm MR}_n({\boldsymbol{a}},{\boldsymbol{b}},{\boldsymbol{\beta}},{\boldsymbol{\alpha}},{\boldsymbol{y}})= B_n \big |_{{\boldsymbol{z}}={\bf 0}}.
\end{gather*}
Generalized {\it Riordan} polynomials:
\begin{gather*}
{\rm RI}_n({\boldsymbol{b}},{\boldsymbol{\alpha}},{\boldsymbol{y}})= B_n \big |_{{\boldsymbol{z}}={\bf 0} {\top {\boldsymbol{a}}={\bf 0}}\atop{\boldsymbol{\beta}}={\bf 0}}.
\end{gather*}
For example,
\begin{gather*}
{\rm RI}_4({\boldsymbol{b}},{\boldsymbol{\alpha}},{\boldsymbol{y}})= {y_1^2 y_2} b_1^2 b_2 \alpha_1+{y_4} \big(\alpha_2 \alpha_3+b_1^2 \alpha_1 \alpha_2\big).
\end{gather*}
\end{Example}
Let us set $B_n(y_1,\ldots,y_n) = B_n({\boldsymbol{a}}={\bf 1},{\boldsymbol{b}}={\bf 1},
{\boldsymbol{\beta}} ={\bf 1},{\boldsymbol{y}}) $. Let $\beta$ be a~new parameter.
Show that
\begin{gather*}
B(1,\beta,\ldots,\beta^{n-1})= \mathfrak{G}_{1 \times w_{0}^{(n-1)}}^{(\beta)}(\underbrace{1,\ldots,1}_{n}),
\end{gather*}
where $\mathfrak{G}_{w}^{(\beta)}(X)$ denotes the $\beta$-Grothendieck polynomial corresponding to a permutation $w \in \mathbb{S}_n$. In particular,
\begin{gather*}
B_n(\underbrace{1,\ldots,1}_n) = {\rm Sch}_n,
\end{gather*}
where ${\rm Sch}_n$ denotes the $n$-th Schr\"{o}der number, that is the numbers of
paths from $(0, 0)$ to $(2n, 0)$, using only steps northeast $U=(1,1)$ or
or $D=(1, - 1)$) or double $H= (2, 0)$, that never fall below the $x$-axis.
Assume that $n$ is devisible by an integer $d \ge 1$.
Show that if ${\boldsymbol{y}} = (y_j = \delta_{j+1,d})$, then
\begin{gather*}
B_n(0,\ldots,0,\underbrace{1}_{d-1},0,\ldots,0) = {\rm FC}_{n/d}^{(d+1)},
\end{gather*}
where ${\rm FC}_m^{p}$ denotes the Fuss--Catalan number, see, e.g., \cite{ST1}, and \cite[$A001764$]{SL} for a variety of combinatorial interpretations the Fuss--Catalan numbers ${\rm FC}_n^{(3)}$.
More generally, let $2 < d_1 < \cdots < d_k$ be a sequence of integers,
and set
\begin{gather*}
{\boldsymbol{y}}= ( \delta_{i+1,d_{j}},\, 1 \le j \le k).
\end{gather*}
Show that
the specialization $B_n({\boldsymbol{y}})$ counts the number of dissections of a convex $(n+2)$-gon on parts which are convex $(d+2)$-gons, where each~$d$
belongs to the set $\{d_1,\ldots,d_k \}$.
We would like to point out that the polynomials
\begin{gather*}
{\rm FS}_n^{(d)}:= {\rm Coef\/f}_{y_{d}^n} \big(P_{nd}({\boldsymbol{a}},{\boldsymbol{b}}, {\boldsymbol{\beta}}, {\boldsymbol{y}}=(\delta_{i+1,d}), {\boldsymbol{z}}) \big).
\end{gather*}
can be treated as a multi-parameter analogue of the Fuss--Catalan numbers ${\rm FC}_n^{(d+1)}$.
{\it Colored dissections} \cite{SW}.
A~colored dissection of a convex polygon is a dissection where
each $(d+1)$-gon appearing in the dissection can be colored by one of~$b_d$
possible colors\footnote{We assume that if $b_{d}=0$, then the dissection in question doesn't contain parts which are $(d+1)$-gons.}, $d \ge 2$~\cite{SW}.
Show~\cite{SW} that if $b_2,\ldots,b_n$ be a sequence of non-negative integers, $B_n(b_2,\ldots,\ldots,b_n)$ is equal to the number of colored dissections of a convex $(n+2)$-gon.
Consider the specialization $y_i=i-1$, $i=1,\ldots,n$. Show that
\begin{gather*}
B_n(y):= {\rm SL}(0,1,\ldots,n-1)={\rm Fine}(n+1),
\end{gather*}
where ${\rm Fine}(m)$ denotes the $m$-th {\it Fine number}, that is the number of
ordered rooted trees with $m$ edges having root of even degree~\cite[$A000957$]{SL}.
Therefore, the Fine number $Fine(n+1)$ counts the number of dissections of a convex $(n+2)$-gon such that each $(d+3)$-gon appearing in the dissection can be colored by $d$ possible colors, $d \ge 1$.
Consider the specialization $y_{3k+1}=1$, $y_{3k+2}=0$, $y_{3k+3}=-1$, $k \ge 0$.
Show that
\begin{gather*}
B_n(y_1, \ldots, y_n)= M_n,
\end{gather*}
where $M_n$ denotes the $n$-th {\it Motzkin} number \cite[$A001006$]{SL}.
Recall that it is the number of ways to draw any number of nonintersecting chord joining~$n$ labeled points on a circle. The number $M_n$ is also equals to the number of {\it Motzkin} paths, that is paths from $(0,0)$ to $(n,n)$ in the $n \times n$ grid using only steps $U=(1,1)$, $H=(1,0)$ and $D=(1,-1)$, see \cite[$A001006$]{SL}
for references and a wide variety of combinatorial interpretations
of Motzkin's numbers.
Consider the specialization $y_{3k+1}=0$, $y_{3k+2}= (-1)^{k}$,
$y_{3k+3} =(-1)^{k}$, $k \ge 0$. Show that
\begin{gather*}
B_n(y_1, \ldots, y_n)= {\rm MS}_n,
\end{gather*}
where ${\rm MS}_n$ denotes the {\it Motzkin sum} or {\it Riordan} number \cite[$A005043$]{SL}.
Recall that it is the number of Motzkin paths of length $n$ with no horizontal steps $H=(1,0)$ at level zero, see~\cite[$A005043$]{SL} for references and a wide variety of combinatorial interpretations of Riordan's numbers.
Consider the specialization $y_{2k+1}= (-1)^k$, $y_{2 k}= (-1)^{k+1}$,
$k \ge 0$. Show that~\cite{SL}
\begin{gather*}
B_n(y_1,\ldots,y_n) = A052709(n),
\end{gather*}
that is the number of underdiagonal lattice paths from $(0,0)$ to $(n-1,n-1)$
and such that each step is either $H=(1,0)$, $V=(0,1)$, or $D=(2,1)$.
Consider specialization $y_k= (-1)^{k} \frac{n !}{k !}$, $k \ge 1$.
Show that
\begin{gather*}
B_n(y_1, \ldots,y_n) = n^{n-2},
\end{gather*}
that is the number of {\it parking functions}, see, e.g., \cite{H, ST1}
and the literature quoted therein.
Consider the specialization $y_k = \frac{n !}{k !}$.
Show that~\cite{SL}
\begin{gather*}
B_n(y_1,\ldots,y_n)= A052894(n),
\end{gather*}
where $A052894(n)$ denotes the number of {\it Schr\"{o}der trees}\footnote{Schr\"{o}der trees have been introduced in a~paper by W.Y.C.~Chen~\cite{Chen1990}. Namely, these are trees for which the set of subtrees at
any vertex is endowed with the structure of ordered partition. Recall that
{\it an ordered partition} of a~set in which the blocks are linearly ordered~\cite{Chen1990}.}.
|
\section{Introduction}
\label{sect:intro}
Molecular gas has been detected through rotational CO transitions out to very high redshift, first in highly luminous sources experiencing ``bursty" star formation \citep[sub-millimetre galaxies \& QSOs; e.g.][]{greve05, riechers06, maiolino07} and subsequently in galaxies undergoing ``main-sequence" (MS) star formation \citep[SF; e.g.][]{daddi08, tacconi10, geach11}. Observations show that, when the universe was half its current age, MS galaxies had gas reservoirs approximately 10-fold larger than nearby spiral galaxies \citep[e.g.][]{daddi10a, tacconi13}.\\
It is not known, however, whether the gas content evolves in a synchronised fashion in both active and passive galaxies, i.e. following a ``universal" cosmic decline. Molecular gas is detected in roughly \nicefrac{1}{5} of all local ETGs \citep[slightly more in the field than in clusters; cf.][]{young11} and typically represents $\sim$1\% of the stellar mass of morphological types E/S0. Do ETG gas fractions increase with redshift? Intuition might suggest this as massive high-$z$ ETGs undergo mergers with gas-rich satellites and accrete a fraction of the latter's gas. \citet{gabor10} show that standard quenching mechanisms (gas consumption in mergers, AGN feedback, virial shocks) produce red sequence galaxies with bluer colors than observed at $z$\,$=$\,1-2 because their simulated spheroids retain substantial residual gas, resulting in excessive SF unless additional quenching mechanisms are invoked. Fragmentation-prone gas reservoirs may be kept from forming stars when embedded within a stellar spheroid \citep[``morphological quenching";][]{martig09}, implying that high-$z$ ETGs could host gas which cannot be tapped to form stars and has so far escaped attention. This is also consistent with the long gas consumption time scales \citet{saintonge12} find in nearby ETGs.\\
To obtain a first direct measurement of the currently unknown molecular gas reservoirs of distant passive galaxies we searched for CO($J$=2$\rightarrow$1) line emission from a massive elliptical galaxy at $z$\,$\sim$\,1.5 with the PdBI. Our target and observations are described in Section \ref{sect:obs}. Results and discussion follow in Sections \ref{sect:results} and \ref{sect:discussion}. We adopt a flat WMAP-7 cosmology \citep[$\Omega_m$\,=\,0.273 and $H_0$\,=\,70.4 km\,s$^{-1}$\,Mpc$^{-1}$;][]{larson11}.\newpage
\section{Target description and data}
\subsection{J100239.52+015659.1 - a pBzK-selected, massive, elliptical galaxy at $z$\,=\,1.4277}
\label{sect:target}
A stringent limit on the gas content of distant quiescent galaxies allowing a meaningful comparison with nearby ETGs is most readily achieved for high-mass galaxies. We selected the most massive galaxy -- J100239.52+015659.1 (pBzK-217431) -- of \citet[``O12" hereafter]{onodera12}. pBzK-217431 is a BzK-selected \citep{daddi04}, passively evolving, red elliptical galaxy in the COSMOS field with spectroscopic redshift\footnote{~The redshift error translates to a $\pm$0.6\,GHz frequency uncertainty ($\pm$5\,$\sigma$) of the redshifted CO($J$=2$\rightarrow$1) line, well within the 3.6\,GHz instantaneous bandwidth of the WideX correlator used here.}
$z$\,=\,1.4277$\pm$0.0015 (derived from a high-confidence detection of the 4000\,{\AA} break) and age $\sim$1.1\,Gyr \citep[corresponding to the average `build' redshift $z$\,$\sim$\,2.3 for the bulk of stars in ETGs in the O12 sample;][]{onodera15}.
With a stellar mass\footnote{~In accordance with literature on IMF variations \citep[e.g.][]{grillogobat10, conroydokkum12, cappellari12} we henceforth adopt a \citet{salpeter55} IMF for pBzK-217431 and a \citet{chabrier03} IMF for SFGs.} of 6.6$^{+0.5}_{-1.9}$\,$\times$\,$10^{11}$\,$M_{\odot}$ (from SED-fitting in O12) this object is among the most massive and luminous ($K_{\rm Vega}$\,=\,17.56) ETGs known at this redshift. It is 5-10 times more massive than star-forming galaxies (SFGs) at a similar redshift followed-up in CO by \citet{daddi10a} or the PHIBSS survey \citep{tacconi13}. pBzK-217431 lies within 1\,$\sigma$ of the local mass-size relation for elliptical galaxies \citep[e.g.][]{newman12} and has an optical surface brightness distribution consistent with the canonical de Vaucouleurs profile \citep[O12,][]{mancini10}. Stellar population modelling and SED-fitting indicate that its star formation rate (SFR) is $\lesssim$6\,$M_{\odot}$/yr \citep{onodera15}.
\subsection{IRAM PdBI 3\,mm-band CO($J$=2$\rightarrow$1) follow-up}
\label{sect:obs}
\begin{figure}
\epsscale{.94}
\centering
\plotone{f1.ps}
\caption{PdBI 3\,mm observations of pBzK-217431. {\it (a)} Spectrum extracted at the target position within $\pm$2000\,km/s of the expected frequency of redshifted CO($J$=2$\rightarrow$1). Dashed lines: $\pm$1\,$\sigma$ noise level for 40\,MHz channels. {\it (b)} Map of the emission integrated over the expected FWHM of the CO line (280\,MHz or 882\,km/s; see panel {\it (a)}, grey area). Solid (dotted) contours: positive (negative) 1 \& 2\,$\sigma$ noise level. Inset: synthesized beam. {\it (c)} HST $I$-band imaging of pBzK-217431 and surroundings (image size -- 30$''$$\times$30$''$; zoom-in size -- 2.5$''$$\times$2.5$''$). Blue/red circle: phase centre of the two separate PdBI pointings targeting the pBzK-217431 field (Section \ref{sect:obs}). Pointing centres are indicated with crosses in panel {\it (b)}.
\label{fig:basics}}
\end{figure}
pBzK-217431 was observed in several tracks between June 21-30, 2011 (IRAM project ID {\it V032}) and May 26-July 19, 2012 (project ID {\it V09E}), always with five PdBI antennae on-source, excluding one track in 2012 conducted with only four antennae. During both observation campaigns all tracks were carried out in the compact D-configuration (angular resolution: 6.5$''{\times}$5.7$''$), with the 3.6\,GHz WideX correlator tuned to the expected frequency of redshifted CO($J$=2$\rightarrow$1) at 94.96\,GHz ($\nu_{\rm rest}$\,=\,230.538\,GHz). In 2011 the phase centre of the observations lay on pBzK-217431 and in 2012 on a nearby dwarf-like galaxy with photometric redshift $z_{\rm phot}$\,=\,0.215 \citep{ilbert13}, offset $\sim$7$''$ from pBzK-217431 (Figure \ref{fig:basics}c). This shift was applied following the detection of a broad (760\,MHz; detected with $S/N$\,=\,6.5 in the $uv$-plane) emission feature associated with the neighboring galaxy in multiple tracks in 2011. Given the photometric redshift of the dwarf galaxy this could have represented a CO($J$=1$\rightarrow$0) line emission feature but the data gathered in 2012 suggested this was an artefact. We reduced the data from the two observing periods individually with the GILDAS software packages CLIC and MAPPING and relying on varying combinations of the calibrators 0851+202, 0906+015, 0923+392, 1005+066, 1040+244, 1055+018, 3C84, 3C273, 3C345, CRL618 and MWC349 for pointing and phase calibration, and on either 0923+392, 1005+066, 3C279, MWC349, 3C84 or 3C345 for flux calibration. Data sets {\it V032} and {\it V09E} were subsequently combined after correcting for primary beam attenuation and phase shifting between the two pointing centers. After flagging of poor visibilities the combined data set has a total 5-antenna (6-antenna) equivalent observing time of 11.59 (7.73)\,hr and reaches a noise level of 0.33\,mJy/beam per 40\,MHz channel (126\,km/s). This sensitivity was not sufficient to detect CO emission. Figure \ref{fig:basics}a shows the featureless spectrum binned in 40\,MHz channels, extracted at the location of pBzK-217431 and centred on the expected position of redshifted CO($J$=2$\rightarrow$1).
\section{Results}
\label{sect:results}
To set an upper limit on the CO($J$=2$\rightarrow$1) line flux we assume that -- as observed in most massive, local ETGs \citep{alatalo13} -- the molecular gas in pBzK-217431 has settled into a rotating disk with circular velocity $V_c\,=\,\kappa\sigma_{\star}$. Following the dynamical analysis on ATLAS$^{\rm 3D}$ ETGs by \citet{cappellari13}, we set $\kappa\,{\sim}\,1.64$ (i.e. intermediate between the maximal circular velocity and the circular velocity at the effective radius). We have no direct measurement of the velocity dispersion for pBzK-217431 but adopt the average\footnote{~A similar line width is inferred by assuming that velocity dispersion, mass and size are related as $\sigma^2_{\star}$\,$\propto$\,\ensuremath{M_{\star}}/$r_e$ and estimating the velocity dispersion of pBzK-217431 (\ensuremath{M_{\star}}\,=\,6.6$\times$10$^{11}$\,$M_{\odot}$, $r_e$\,=\,7.2\,kpc) based on the O12 sample median [$\sigma_{\star}$, \ensuremath{M_{\star}}, $r_e$]\,=\,[330\,km/s, 2.2$\times$10$^{11}$\,$M_{\odot}$, 2.4\,kpc]. This re-scaling procedure involving size measurements is only meaningful when all galaxies have similar structural properties; the median S\'ersic index and scatter of the O12 sample is 3.2$\pm$1.2, implying that pBzK-217431 lies within 0.5\,$\sigma$ of the sample median.} $\sigma_{\star}$\,=\,330\,km/s of the 1.4\,$<$\,$z$\,$<$\,2.1 ETG parent sample \citep{onodera15}. This value agrees well with velocity dispersions found by other high-$z$ ETG studies \citep[e.g.][]{newman10} and translates to an expected line width $2V_c\,{\sim}$\,910\,km/s (approx. seven spectral channels) if we assume a gas disk with inclination angle\footnote{~Line width and velocity-integrated flux $I_{\rm CO}$ vary with viewing angle as ${\rm sin}(i)$ and $\sqrt{{\rm sin}(i)}$, respectively. For an edge-on gas disk the upper limits on the CO-luminosity and H$_2$ mass would thus increase by $\sim$10\%.} $i\,{=}\,57.3^{\circ}$, i.e. the mean orientation for randomly oriented disks in 3D.\\
Figure \ref{fig:basics}b shows the flux distribution within $\sim$15$''$ of the target position after integration of the 3\,mm spectrum over seven channels (280\,MHz; 882\,km/s in velocity space) centred on the expected line frequency. While we see a $\sim$3\,$\sigma$ flux concentration 3$''$ north of pBzK-217431 in Figure \ref{fig:basics}b we do not consider this a reliable signature of weak line emission from our target. This is supported by the fact that the total number of $>$2\,$\sigma$ pixels in the displayed region is fully consistent with Gaussian statistics. At very low signal-to-noise the most robust estimate of the flux is generally obtained at the phase centre where we detect flux only at 1.2\,$\sigma$ significance. The 1\,$\sigma$ noise level for $I_{{\rm CO}(J{=}2{\rightarrow}1)}$ is 0.11\,Jy\,km/s, leading to a 3\,$\sigma$ upper line luminosity limit $L'_{{\rm CO}(J{=}2\rightarrow1)}$\,$<$\,8.8$\times$10$^{9}$\,K\,km/s\,pc$^2$ following \citet[equation 3]{solomonvandenbout05}.\\
For ATLAS$^{\rm 3D}$ ETGs CO($J$=2$\rightarrow$1) emission is consistent with thermal excitation \citep{young11} and can thus be used as a direct tracer of the H$_2$ content. Converting our line flux limit to a constraint on the gas mass of pBzK-217431 requires choosing a CO-to-H$_2$ conversion factor. ISM conditions in low-$z$ ETGs generally resemble those of local spiral galaxies (dense gas fractions -- \citealp{krips10}; H$_2$ surface densities -- \citealp{young11}; dust temperatures -- \citealp{crocker12}; low star formation efficiency, SFE -- \citealp{martig13}), leading us to adopt a Milky-Way-like \ensuremath{\alpha_{\rm CO}}\,$\sim$\,4.4\,$M_{\odot}$/(K\,km/s\,pc$^2$). A similar \ensuremath{\alpha_{\rm CO}}\ is inferred when considering a metallicity-dependent conversion factor \citep[e.g.][]{schruba12, genzel12} and using the slightly super-solar, average enrichment $Z/Z_{\odot}$\,$\sim$\,1.2 of ETGs of the O12 sample. Finally, we note that simulations of high-$z$ ellipticals with a range of gas fractions ($\sim$10-50\%) and post-processing with Large Velocity Gradient (LVG) models predict \ensuremath{\alpha_{\rm CO}}-values comparable to that of the Milky Way as well \citep{bournaud14}.\\
\begin{deluxetable}{lr}
\tabletypesize{\small}
\tablewidth{0pt}
\tablecaption{Source properties: pBzK-217431. \label{tab:prop}}
\tablehead{
\colhead{quantity/observable} &
\colhead{value}}
\startdata
R.A. [J2000] & 10$^h$02$^m$39.527$^s$\\
Dec. [J2000] & +01$^d$56$^m$59.12$^s$\\
$z_{\rm spec}$ & 1.4277$\pm$0.0015\\
\ensuremath{M_{\star}}\ [$M_{\odot}$] & 6.6$^{+0.5}_{-1.9}$\,$\times$\,$10^{11}$\\
$r_e$ [kpc] & 7.19$\pm$1.95\\
S\'ersic index $n$ & 3.8$\pm$0.6\\[0.5ex]
\hline\\[-2ex]
rms/40\,MHz [mJy] & 0.33\\
$I_{\rm CO(J{=}2\rightarrow1)}$ [Jy\,km/s] & $<$0.32\,$\sqrt{\left(\frac{\Delta v}{882\,{\rm km/s}}\right)}$\\
$L'_{\rm CO(J{=}2\rightarrow1)}$ [K\,km/s\,pc$^2$] & $<$8.8$\times$10$^{9}$\\
\ensuremath{M_{\rm mol.}}\ [$M_{\odot}$] & $<$3.9$\times$10$^{10}$\,$\left(\frac{\alpha_{\rm CO}}{4.4}\right)$\\
$f_{\rm gas}$ & $<$5.8\%
\enddata
\tablecomments{Properties listed above horizontal divider are from \citet{onodera15}. Upper limits are quoted at 3\,$\sigma$ significance.}
\end{deluxetable}
\begin{figure*}
\centering
\begin{tabular}{ccc}
\includegraphics[width=0.46\textwidth]{f2a.ps} & & \includegraphics[width=0.46\textwidth]{f2bc.ps}\\
\end{tabular}
\caption{CO emission and gas content of pBzK-217431 compared to ATLAS$^{\rm 3D}$ ETGs and 0\,$<$\,$z$\,$<$\,2.5 main-sequence (MS) galaxies. {\it (a)} 3\,$\sigma$ upper limit on the ratio of CO luminosity and stellar mass {\ensuremath{M_{\star}}} for pBzK-217431, plus typical $L'_{\rm CO}/\ensuremath{M_{\star}}$ ratios measured for ATLAS$^{\rm 3D}$ ETGs (light red -- median of CO-detected ETGs; dark red -- upper limit on sample median for {\it all} ATLAS$^{\rm 3D}$ ETGs) and literature MS galaxy samples. (See annotations for survey names. Author abbreviations: D10b -- \citealt{daddi10b}; G11 -- \citealt{geach11}). For MS galaxies we plot the median $L'_{\rm CO}/\ensuremath{M_{\star}}$ and redshift of the combined samples; error bars illustrate the 1\,$\sigma$ scatter in the data. Dark grey, solid line: expected redshift evolution of $L'_{\rm CO}/\ensuremath{M_{\star}}$ for an average MS galaxy with \ensuremath{M_{\star}}\,=\,5$\times$10$^{10}$\,\ensuremath{M_{\odot}}\ (Section \ref{sect:discussion_evo}). Dashed lines (edges of light grey shading): $L'_{\rm CO}/\ensuremath{M_{\star}}$ ratio for $\pm$1\,$\sigma$ ($\pm$2\,$\sigma$) outliers to the MS.\newline
{\it (b)} Molecular gas fraction $f_{\rm mol.}$ for ATLAS$^{\rm 3D}$ ETGs compared to the average relation between gas fraction and \ensuremath{M_{\star}}\ for massive $z$\,$\sim$\,0 MS galaxies. (Shading/lines as in panel {\it (a)}; in panels {\it (b)} and {\it (c)} the grey shading fades away at the cross-over mass, above which ETGs dominate SFGs by number at the respective redshift.)\newline
{\it (c)} As in panel {\it (b)} but highlighting the paucity of gas in pBzK-217431 with respect to typical gas fractions of massive MS galaxies at the same redshift. Dotted lines: $\pm$1\,$\sigma$ scatter of $f_{\rm mol.}$ in $z$\,$\sim$\,0 MS galaxies (cf. panel {\it (b)}).
\label{fig:fgas}}
\end{figure*}
Atomic (HI) to molecular gas ratios in nearby ETGs vary with environment and kinematic state. Based on the fifth nearest neighbour distribution of galaxies in the COSMOS field with a similar redshift $z$\,=\,1.43$\pm$0.2 as our target, pBzK-217431 lies in an average density environment (54$^{\rm th}$ percentile of the density distribution), where HI can dominate the gas phase for nearby ETGs \citep[e.g.][]{welch10}. Conversely, it is in a mass regime where locally HI-poor slow-rotators dominate the ETG population. Absent direct HI measurements, we assume that mass is the key factor and that CO emission traces the {\it entire} gas reservoir, similar to ETGs in the local sample of \citet{crocker11}. In this case, our observations constrain the gas mass of pBzK-217431 to be $M_{\rm gas}$\,$<$\,3.9$\times$10$^{10}$\,($\alpha_{\rm CO}$/4.4)\,$M_{\odot}$. This is comparable to gas masses in MS galaxies at 1\,$<$\,$z$\,$<$\,2, whereas pBzK-217431 is an order of magnitude more massive, resulting in a 3\,$\sigma$ upper limit\footnote{~Although $M_{\rm gas}$ appears in numerator and denominator of $f_{\rm gas}$, we can nevertheless define a meaningful upper limit on the gas fraction because the inverse $\nicefrac{1}{f_{\rm gas}}$\,=1+$\nicefrac{\ensuremath{M_{\star}}}{M_{\rm gas}}$ represents a well defined {\it lower} limit on $\nicefrac{1}{f_{\rm gas}}$ when, as here, only an upper limit on $M_{\rm gas}$ is available.} of $f_{\rm gas}$\,=\,$\nicefrac{M_{\rm gas}}{(M_{\star}+M_{\rm gas})}$\,$<$\,5.8\%, i.e. a factor 10 lower than most CO-detected MS galaxies at the peak epoch of galaxy assembly. The properties of pBzK-217431 are summarized in Table \ref{tab:prop}.
\section{Discussion}
\label{sect:discussion}
\subsection{The gas content of pBzK-217431 in the galaxy evolution context}
\label{sect:discussion_evo}
\begin{figure*}
\epsscale{.5}
\centering
\begin{tabular}{ccc}
\includegraphics[width=0.46\textwidth]{f3a.ps} & & \includegraphics[width=0.46\textwidth]{f3b.ps}\\
\end{tabular}
\caption{Location of pBzK-217431 in the integrated Schmidt-Kennicutt diagramme {\it (a)} and in SFE vs. $f_{\rm mol.}$ space {\it (b)}. Data points scattering around the `MS' SF law in {\it (a)} are literature main-sequence galaxies \citep[compiled in][Sa14]{sargent14}. Crosses (open circles): low- (high-)$z$ starbursts with measured \ensuremath{\alpha_{\rm CO}}\ and typically $\sim$15$\times$ higher SFE than MS galaxies. Indicative SFR estimates for pBzK-217431 are from average spectral/SED properties of $z$\,$\sim$\,1.5 ETG samples \citep[O12,][]{man14}. Colored symbols indicate the properties of simulated gas-rich/gas-poor ({\it purple}/{\it dark red}) ETGs and gas-rich/gas-poor ({\it blue}/{\it orange}) disks (see discussion in text). Error bars on simulated data points illustrate variations in the simulations of these quantities over a few 100\,Myr.
\label{fig:SFE}}
\end{figure*}
As a proxy for the redshift evolution of ETG gas fractions which involves no assumptions about \ensuremath{\alpha_{\rm CO}}\ we compare in Figure \ref{fig:fgas}a the \ensuremath{L'_{\rm CO}}/\ensuremath{M_{\star}}\ ratio of pBzK-217431 with sample averages for MS galaxies at $z$\,$<$\,2.5 and $z$\,$\sim$\,0 ETGs\footnote{~Masses for ATLAS$^{\rm 3D}$ ETGs are the product of the $r$-band luminosity and mass-to-light ratios from JAM modelling in \citet{cappellari12}.} from ATLAS$^{\rm 3D}$ \citep{young11}. Median \ensuremath{L'_{\rm CO}}/\ensuremath{M_{\star}}\ values of literature MS galaxy samples are plotted with spiral galaxy symbols: $z$\,$\sim$\,0 -- HERACLES \& COLD GASS \citep{leroy09, saintonge11}; 0.1\,$<$\,$z$\,$<$\,0.6 -- \citet{daddi10b, geach11} \& EGNoG \citep{bauermeister13}; 1.1\,$<$\,$z$\,$<$\,1.7 \& $z$\,$>$\,2 -- \citet{daddi10a} \& PHIBSS \citep{tacconi13}. The continuous evolution of the gas content in $z$\,$<$\,2.5 MS galaxies with \ensuremath{M_{\star}}\,$\sim$\,5$\times$10$^{10}$\,$M_{\odot}$, the typical mass of all these samples, is traced by the grey band, based on the scaling relations between \ensuremath{M_{\star}}, SFR, molecular gas mass $M_{\rm mol.}$ and \ensuremath{\alpha_{\rm CO}}\ calibrated in \citet{sargent14}. For ATLAS$^{\rm 3D}$ ETGs we plot the median of the subset of CO-detected ETGs (22\% of ATLAS$^{\rm 3D}$ ETGs) and an upper limit on the median \ensuremath{L'_{\rm CO}}/\ensuremath{M_{\star}}\ ratio of the overall sample, placed at the 84$^{\rm th}$ percentile of the \citet{kaplanmeier58} distribution which accounts for both the CO-detections and non-detections. The upper limit on the CO-deficit of pBzK-217431 relative to $z$\,$\sim$\,1.4 MS galaxies is $\sim$20 (see Figure \ref{fig:fgas}a), comparable to that of CO-detected ATLAS$^{\rm 3D}$ ETGs relative to local spirals. This limit is not stringent enough for a helpful comparison with the full ATLAS$^{\rm 3D}$ ETG sample median.\\
In the introduction we asked whether gas fractions in star-forming and quiescent galaxies evolved with time by equal proportions. A quantitative comparison must involve the fact that pBzK-217431 is about 10$\times$ as massive as ATLAS$^{\rm 3D}$ ETGs on average. Figures \ref{fig:fgas}b/c show that gas fractions of local ETGs vary inversely with \ensuremath{M_{\star}}, reaching $\lesssim$1\textperthousand\ at the highest masses. Our observations are not sufficiently deep to start testing whether, at fixed \ensuremath{M_{\star}}, $f_{\rm gas}$ for quiescent galaxies is constant over the last $\sim$10\,Gyr. Over this period gas fractions of normal galaxies with masses at or above the knee of the stellar mass function ($M^*$\,$\sim$\,10$^{11}$\,$M_{\odot}$) evolve by a factor 4-5 \citep[from a few to about 15-30\%; cf. Figure \ref{fig:fgas}b/c or][]{tacconi13, scoville14}. For pBzK-217431 and local ETGs the values are $<$6\% and $<$1\%, resp., such that we cannot exclude a similar $\sim$5-fold evolution of $f_{\rm gas}$ for ETGs based solely on our observations.\\
Combining our upper limit on the gas mass of pBzK-217431 with SFR-estimates from ETG stacked optical spectra and optical-to-IR SEDs \citep{man14, onodera15}, we can estimate its SFE. The average specific SFR of $<$\,10$^{-11}$\,yr$^{-1}$ for the ETG sample of O12 is consistent with the upper limit for the {\it dust-obscured} SFR in the largest mass bin of a larger sample of photometrically selected quiescent galaxies in \citet{man14}. Applied to pBzK-217431 this level implies a SFR of $<$5-10\,$M_{\odot}$/yr , while SED fitting \citep{man14} suggests SFR\,$\sim$\,0.4\,$M_{\odot}$/yr. The resulting, loose constraint on the gas depletion time $\tau_{\rm depl.}$\,=\,$\nicefrac{M_{\rm gas}}{\rm SFR}$ is $\lesssim$90\,($\alpha_{\rm CO}$/4.4)\,Gyr (see Figure \ref{fig:SFE}). Presuming that, similar to nearby ETGs \citep[e.g.][]{saintonge12, martig13}, high-$z$ ETGs have 2-5$\times$ lower SFEs than disk galaxies this means that a roughly 10-fold sensitivity increase compared to our PdBI observations would be required to detect an expected gas reservoir of order 10$^9$\,$M_{\odot}$. This highlights the need for alternative ISM tracers or deeper observations, e.g. with ALMA, in order to advance our knowledge of the cold gas in distant quiescent galaxies.
\subsection{Considerations on feedback and quenching}
\label{sect:discussion_feedback}
Galaxies as massive as pBzK-217431 at $z$\,=\,1.4277 are prime candidates for suffering so-called `mass quenching' \citep[e.g.][Figure 15]{peng10}. Various mechanisms potentially contribute to mass quenching, and more widespread observations of the ISM of quenched galaxies as presented here should ultimately reveal which (combination) of these is most important. Examples of internal processes are gravitational/morphological quenching \citep{martig09, genzel14}, AGN-feedback \citep[e.g.][]{granato04} or SF-activity \citep[e.g.][]{ceverinoklypin09}. Feedback suppresses SF by removing gas, while simulations of morphological quenching have shown that the SFR of ETGs can remain minimal for several Gyr, notwithstanding gas fractions of 5-10\% \citep[cf. Figure 3 in][]{martig09}.
Our non-detection for pBzK-217431 thus indicates that massive high-$z$ ellipticals at high redshift are truly gas-poor. We illustrate this with Figure \ref{fig:SFE}b which compares SFE(s) and molecular gas fraction(s) for pBzK-217431 and low- and high-$z$ MS galaxies with simulated $z$\,$\sim$\,2 ETGs (purple -- simulated ETG with $f_{\rm mol.}$\,$\sim$\,50\%; dark red -- simulated $f_{\rm mol.}$\,$\lesssim$10\%) having \ensuremath{M_{\star}}\,$\sim$4$\times$10$^{10}$ (open symbols) and $\sim$10$^{11}$\,$M_{\odot}$ (filled symbols). These simulations include those from \citet{bournaud14} and similar models with other parameters (galaxy type/gas fraction) with 3\,pc resolution. They are idealized closed-box models with a chosen initial stellar structure and gas fraction, evolved for a few dynamical times including self-gravity, hydrodynamics, cooling and stellar feedback to compute the phase-space distribution of the ISM and the associated SFE. In gas-rich simulated ETGs, morphological quenching is ineffective, as evidenced by the fact that these systems have the same SFEs as both observed and simulated gas-rich disk galaxies. Conversely, morphological quenching could be responsible for the small (comparable to the dispersion of the SF-law) SFE-offset between the gas-poor simulated ETGs and simulated $z$\,$\sim$\,0 disks (orange spirals). The low SFRs of massive, high-$z$ ellipticals are thus not the result of unusually low SFE caused by gravitational stabilization of gas reservoirs, but morphological quenching might shut down SF completely in already gas-poor ETGs. Note that the absence of gas need not be the outcome of active expulsion by AGN or stellar feedback. Sub-millimeter galaxies as the commonly assumed progenitors of early ellipticals usually have sufficiently short gas consumption times \citep[e.g.][]{bothwell13} for reservoir-depletion to well below the sensitivity of our observations of pBzK-217431.
\section{Summary}
\label{sect:summary}
We presented the outcome of a deep IRAM/PdBI 3\,mm search for CO($J$=2$\rightarrow$1) emission from pBzK-217431, a massive, passively evolving elliptical galaxy at $z$\,=\,1.4277. Though undetected in CO, we could infer a 3\,$\sigma$ upper limit of 5.8\% for its gas fraction. This is almost 10$\times$ below the typical $f_{\rm gas}$ of MS galaxies at the same epoch and comes close to the gas fractions of local ATLAS$^{\rm 3D}$ ETGs, although our limit is not stringent enough for a direct comparison with the most massive ellipticals in the nearby universe where $f_{\rm gas}$\,$\sim$\,1\textperthousand. Observational data on the gas content of quiescent galaxies not only tells us about the currently unknown contribution of such galaxies to the cold gas history of the Universe \citep[e.g.][Sargent et al., in prep.]{walter14}, it also provides insight into how/which quenching and feedback processes influence the ISM of the host galaxy \citep[e.g.][]{ciottiostriker97}. Our gas fraction measurement for pBzK-217431 leads us to speculate that moderate-sized, gravitationally stabilized gas reservoirs are not widespread among massive high-$z$ ellipticals and that their low SFR reflects a paucity of gas, caused by either feedback or exhaustion of their gas reservoir by prior SF activity.
\acknowledgments
We thank J. M. Winters, A. Cibinel, P.-A. Duc, M. Krips and the referee for helpful suggestions and/or data reduction support. MTS, ED, RG \& HD acknowledge funding from ERC-StG \#240039 and grant ANR-08-JCJC-0008, FB funding from ERC-StG \#257720.
|
\section{Introduction}
With the popularity of personal video cameras and multi-view video capturing devices, we are entering an era with rich amount of multimedia documents surrounding us. To interact with these videos, there have been increasing demands of understanding human activities in these videos. Although many research studies \cite{SIFTBag,randomforest,fusion,zhu09,MultimediaEvent12,CrossDataset,SIFT,Featurecorrelation12,MultiView,journals/tmm/DuanLC13} have been carried out to understand and retrieve large scale video contents, these works focus on high level semantics instead of describing human activities. There still exists a need to recognize fine-grained information for human activities, \emph{e.g.} body poses and temporal motions.
This paper targets on the challenge of understanding spatial and temporal variances of video phenomenons. More specifically, we are considering the following difficulties:
\vspace{-2mm}
\begin{itemize}
\setlength{\itemsep}{1pt}
\setlength{\parskip}{0pt}
\setlength{\parsep}{1pt}
\item A human body is composed of multiple parts, and different parts are associated with different motions.
\item Both the short term and long term motions may be fused by different background movement or camera motion, which brings the difficulties of accurately modeling the temporal characteristics for actions.
\item In real world videos, human actions often happen with uncertainties: some happen along with occlusions, some are caused by view/pose variance, or due to diverse actor appearances and motions.
\end{itemize}
Due to these difficulties, it is crucial to reduce the spatio-temporal ambiguity when modeling the human actions. Most of the previous studies were built on simplified action models while overlooking the detailed spatio-temporal structure information. Of these works, a large amount of studies were based on spatio-temporal interest points \cite{STIP,YuanLW11,SongTZCZL12}. Some researchers proposed to enrich the action model with the appearance information or context information \cite{ActionTemplate09,Shuicheng09}. Some other researchers learned temporal structures for action recognition \cite{SatkinH10,LatentTem_CVPR2012}. However, none of these works provides an effective model which can unify spatial and temporal information to infer the structure of human motion.
We aims to develop an effective configurable model, namely the Spatio-Temporal And-Or Graph (STAOG) for action recognition, which addresses the problems mentioned above. Our idea is partially motivated by the image grammar model~\cite{ZhuM06}, which hierarchically decomposes an image pattern with mixed and-nodes and or-nodes, as well as modeling rich structural variations of parts.
The challenges of generalizing And-Or graphs for action recognition are two-folds. First, the traditional And-Or graphs are limited in modeling the hierarchical configuration of spatio-temporal information. Because actions in videos are often more complicated than images, we need more powerful models for video problems. Second, videos require more efficient models that can be effectively learned from large amount of video information without elaborate supervision and initialization.
\begin{figure}[ptb]
\begin{center}
\includegraphics[width=3.5in]{framework}
\end{center}
\vspace{-4mm}
\caption{An example of the Spatio-Temporal And-Or Graph model, where the and-nodes represent compositions in either time or space, the or-nodes indicate structural alternatives, and the leaf-nodes (at the bottom) correspond to local part detectors. The links between leaf-nodes represent spatio-temporal contextual interactions.}
\label{fig:framework}
\end{figure}
To handle the first challenge, our STAOG model extends the traditional deformable graphical models by introducing switch variables in hierarchy, \emph{i.e.} or-nodes that explicitly specify structural reconfiguration. Both spatial and temporal interactions between action parts can be simultaneously learned. One action in the video can be treated as an ensemble of spatio-temporal compositions: a number of discrete temporal anchor frames, each of which is further decomposed into a layout of deformable parts. An example of the proposed STAOG model is illustrated in Fig.~\ref{fig:framework}. There are four layers in our model from bottom to top:
(1) The \emph{leaf-nodes} in the bottom layer represent a batch of local classifiers for detecting various action parts in every anchor frame, denoted by the solid circles in Fig.~\ref{fig:framework}. During detection, location displacement is allowed for each leaf-node to tackle the part deformation.
(2) The \emph{or-nodes} over the bottom are ``switch'' variables specifying the activation of their children leaf-nodes, denoted by the dashed circles. Each or-node is used to specify an appropriate selection from candidate action parts detected by the associating children leaf-nodes. In this way, by explicitly switching selections over leaf-nodes,
the or-nodes make our model reconfigurable during the inference of detection, which is the key to handle large action variabilities.
(3) The \emph{and-nodes} verify the holistic appearance of action within the anchor frame, (the rectangles in layer 2), and we thus consider it as the spatial and-node. It includes two aspects: (i) a global classifier with bag-of-features, and (ii) aggregated scores from its children or-nodes.
(4) The \emph{root-node} in the top can be viewed as an and-node in time, (the rectangle in top). Its definition is similar to the spatial and-node: (i) a classifier with global features in observed frames, (ii) aggregated scores over candidate temporal anchor frames, plus penalty for anchor frame displacements.
(5) The \emph{spatio-temporal contextual interactions}, \emph{e.g.} the curves (graph edges) among leaf-nodes in Fig.~\ref{fig:framework}, are defined based on informative contextual pairwise relations in either spatial or temporal domain. Note that the collaborative edges are imposed between leaf-nodes that are associated with different action parts. Their effectiveness will be particularly demonstrated in the experiments.
To overcome the second challenge, we present a novel weakly supervised learning algorithm for model training, inspired by the non-convex optimization techniques~\cite{cccp,NIPS12}. This algorithm trains the model in a dynamic manner: the model structure (\emph{e.g.} the configuration of leaf-nodes and or-nodes) is iteratively generated and reconfigured on the training data, with optimizing the multi-layer parameters. The other structure attributes (\emph{e.g.} the activation of leaf-nodes, and temporal deformation of anchor frames) are modeled with the latent variables and optimized simultaneously.
In the testing stage, we present an algorithm of cascaded search and verification for recognizing actions with the trained STAOG model. We first generate a set of hypotheses in both spatial and temporal compositions. (i) Spatial testing via the and-nodes. Within the input frame, all candidate action parts are found by leaf-nodes and several possible configurations (\emph{i.e.} spatial compositions of action parts) are produced with different specifications via the or-nodes. These configurations are also weighted via the and-node. (ii) Temporal testing via the root-node. The scores proposed by the and-nodes are aggregated via the root-node for a possible temporal composition. Several possible configurations are then produced as hypotheses represented by different latent variables. Finally, each hypothesis is globally verified with the spatial and temporal edges in the model.
This paper is organized as follows. Section 2 gives a brief review of related work. Then we present our STAOG model in Section 3, followed by a description of the inference procedure in Section 4. Section 5 presents a description for structural learning of our model. The experimental results and comparisons are exhibited in Section 6. Section 7 concludes this paper.
\section{Related work}
Traditional works for action recognition focused on developing informative features, such as spatio-temporal descriptors~\cite{STIP,DollarVSPETS05cuboids,zhu09},3D Gradient~\cite{3Dgradients}, 3D SIFT descriptors~\cite{SIFT} and motion features~\cite{TrajectoryModeling12,trajYan}, and the action classifier can be trained with labeled data. Most of these methods, however, are limited to periodic actions with clean background, such as running and jogging.
To address complex actions with cluttered background, several compositional or expressive models were proposed and achieve very impressive results~\cite{actioncontext,CrossDataset, actionbank,HMM,highlevel12,contextedges}. For example, Wang et al.~\cite{HMM} modeled the human action by a flexible constellation of parts conditioned on image observations and learned the parameters of an HCRF model in a max-margin framework, motivated by the recent progresses in object recognition and detection, \emph{e.g.} the deformable part model by Felzenszwalb et al.~\cite{iccp}. Yao et al.~\cite{ActionTemplate09} proposed to generate spatio-template action templates with the information projection method. Sadanand et al~\cite{actionbank} adopted high-level representations with a bank of individual action detectors. However, actions in video often involve much more information in both spatial and temporal domain, compared with image-based object recognition, and most of these studies do not explicitly localize parts of actions (actors) due to the computational burden. Moreover, structural configurations of these models are usually fixed, including a fixed number of part detectors as well as the predefined composition.
One unique characteristic of human action recognition problem lies in the temporal structure. A lot of works were proposed to build temporal structure models~\cite{HSMM05, ModelTemp,LatentTem_CVPR2012,SpatioTempGraph} based on discriminative and interesting motion segments of the video. Raptis et al.~\cite{DiscoverPart} extracted clusters of trajectories and proposed a graphical model to incorporate constraints for individual and group events. Albanese et al.~\cite{ProbabilisticPetriNet} represented temporal relations of activities using the probabilistic Petri Nets and integrated high-level reasoning approaches. Different from these approaches, we do not treat the whole temporal frames as units.Instead, we model temporal structure based on action parts with explicit relations and presents a solution to find both spatial and temporal configurations for dynamic activities.
Recently, the And-Or graph models~\cite{ZhuM06} have been discussed for several vision tasks such as object recognition~\cite{LinGrammar} and shape modeling~\cite{objectAndOrTree}. These works mainly focused on images instead of videos and do not take the temporal dynamic structure into account. The very recent work by Amer et al.~\cite{costsenstive} proposed to recognize activities with the spatio-temporal And-Or graph model, but they over-simplified the model training by manually fixing the model structure (\emph{i.e.} the layout of graph nodes).
It is worth mentioning that this paper learn the spatio-temporal graph without using any extra annotations or scripts. Research works which utilize rich annotation for event parsing and interpretation are beyond the scope of this work. In contrast, Marszalek et al.~\cite{actioncontext} explored the action contexts of natural dynamic scenes with movie scripts. Gupta et al.~\cite{Storyline09} proposed to learn a visually grounded storyline model from annotated videos, and Pei et al.~\cite{ZhuEventAnd} studied the event grammar model for daily activities based on a predefined set of unary and binary relations. Extra annotations are required for these studies.
\section{Spatio-Temporal And-Or Graph}
The STAOG model is defined as $\mathcal{G = (V,E)}$, where $\mathcal{V}$ represents the four types of nodes and $\mathcal{E}$ the graph edges as Fig.~\ref{fig:framework}. The root node in top verifies the temporal composition, which aggregates scores over anchor frames. Each and-node represents a temporal anchor frame for verifying spatial composition. The or-nodes are derived from each and-node, which are ``switch'' variables for specifying the activation of their children leaf-nodes. The number of leaf-nodes for each or-node is dynamically learned with an upper limit number $m$. For simplicity, we use $t= 1,\dots,T$ to index all and-nodes in the whole STAOG model, $i = 1,\dots,Z$ for or-nodes and $j = 1,\dots,n$ for leaf-nodes. We also index the child or-node of and-node $A_t$ as $i\in ch(t)$, and index the child leaf-node of or-node $U_i$ as $j \in ch(i)$. The spatio-temporal graph edges(\emph{i.e.} interactions) are defined between the leaf-nodes associated with different or-nodes. In this section, we describe two factors in detail: the spatio-temporal compositions, and the contextual interactions in both spatial and temporal domains.
\subsection{Spatio-Temporal Compositions}
\label{sec::Spatial}
We employ Laptev's 3-D corner detector~\cite{STIP} to detect interest points in video sequences, and each interest point is described by HoG (histogram of gradient) and HoF (histogram of optical flow)~\cite{realisticLap}. Furthermore, we generate a dictionary of spatio-temporal interest points' descriptors, clustered by the k-means method in training stage. Given a video sequence $X$, we first equally divide it into $T$ temporal segments. The center frame in each video segment is chosen as an initial anchor frame. Each anchor frame is further decomposed into a number of action parts. In our method, we define the action parts based on the video patch representation, \emph{i.e.} 3-D volumes spanning $\rho$ consecutive frames. Thus, for each anchor frame $I_t$, we observe a sequence of frames centered at $I_t$, and the sequence denoted as $\Lambda_t$ is treated as the input for anchor frame processing.
\begin{figure}[ptb]
\begin{center}
\epsfig{figure=spa,width=3.4in}
\end{center}
\caption{Illustration of spatial compositions. (a) The black boxes denote the initial positions of action parts. (b) The parts are exhibited which are associated with a set of leaf-nodes. Each $(d_x,d_y)$ indicates the location displacement determined by the model. (c) The activated leaf-nodes are highlighted by red and spatial contextual interactions are defined between the pairwise spatial adjacent leaf nodes.}
\label{fig:spa}
\end{figure}
\textbf{Leaf-node:} Each leaf-node $L_j$ represents a local classifier for detecting action parts within video patches and includes two terms: an appearance feature $\phi^l$ and the spatial displacement feature $\phi^s$. Within an anchor frame $I_t$, the features of action parts are described by the BoW histogram based on the generated dictionary. Assume the action part detected by $L_j$ is localized at position $p_j =(p_{j}^x,p_{j}^y)$ , then $\phi^l(\Lambda_t,p_j)$ is denoted as the appearance feature. During detection, the locations of action parts are allowed to be perturbed to tackle the spatial deformation. We incorporate the spatial displacement $\phi^s(q_{j},p_{j}) = (d_x,d_y)$ for each action part, which can be computed by maximizing the response of $L_j$ during inference; $q_j$, representing the initialized position of $L_j$, is set according to the center-point of the frame. Thus the spatial displacement is defined as $\phi^s(q_{j},p_{j}) = (d_x,d_y)$, where $(d_x,d_y)$ is the displacement.The response of $L_j$ is defined as,
\begin{equation}\begin{aligned}
R_{j}^l(\Lambda_t,p_j) = \omega_j^l\cdot\phi^l(\Lambda_t,p_j) - \omega_j^s\cdot\phi^s(q_{j},p_{j}),\label{eq:leafnode}
\end{aligned}\end{equation}
where $\omega_j^l$ is the parameter for the appearance feature and $\omega_j^s$ corresponds to the spatial deformation parameter.
\textbf{Or-node:} Each or-node $U_i$ is proposed to specify an appropriate candidate from its children leaf-nodes. For each leaf-node $L_j$ of $U_i$'s children, the indicator variable $v_j \in \{0,1\}$ represents whether it is activated or not and each or-node only selects one leaf-node. Briefly, we utilize the indicator vector $\mathbf{v}_i$ for the or-node $U_i$ and each element of $\mathbf{v}_i$ is an indicator variable $v_j$ of the leaf-node $L_j$. Intuitively, the significant intra-class variance caused by views, background clutters or actors can be captured by different spatial configurations that are determined with the or-nodes. The response of the or-node $U_i$ is defined as,
\begin{equation}\begin{aligned}
R_{i}^u(\Lambda_t, \mathbf{v}_i)=\sum_{j\in ch(i)}R_{j}^l(\Lambda_t,p_j)\cdot v_j,\label{eq:ornode}
\end{aligned}\end{equation}
\textbf{And-node:} Each and-node $A_t$ verifies the holistic appearance of action for the anchor frame $I_t$, and spatial composition of the or-nodes in its children. We define the configuration vector $V_t$ for all leaf-nodes within the anchor frame, which includes all indicator vectors $\mathbf{v}_i$ corresponding to its children or-nodes $U_{i}$. The response of the and-node $A_t$ is defined as,
\begin{equation}\begin{aligned}
\! R^a_t(\Lambda_t,V_t) =&\omega^a\cdot\phi^a(\Lambda_t) + \sum_{i\in ch(t)}R_{i}^u(\Lambda_t,\mathbf{v}_{i}),\label{eq:andnode}
\end{aligned}\end{equation}
where $\phi^a(\Lambda_t)$ is the BOW histogram globally extracted from the 3-D volume $\Lambda_t$ centered at $I_t$. The second term aggregates the response scores from all or-nodes of $A_t$'s children.
\textbf{Root-node:} The root-node is a global potential function that verifies the temporal compatibility of model, including three terms: the global BoW histogram of the video clip, aggregated scores of its children and-nodes, and temporal displacements of anchor frames. Fig.~\ref{fig:temp} illustrates the temporal composition by the root-node. We employ the root-node for searching for the best localizations of $T$ anchor frames. We introduce the latent variable $\Delta_t$ to indicate the temporal displacement of each anchor frame $I_t$, which will be calculated during inference. This implicitly carries the temporal ordering constraints which are crucial for discriminating human activities.
In particular, the temporal displacement penalty $\xi_t$ punishes the position of the and-node $A_t$ (\emph{i.e.} one anchor frame) shifting far away from the initial anchor point $\tau_t$ in time. Once $\Delta_t$ is optimized, the position of each anchor frame can be determined by $\tau_t + \Delta_t$ accordingly. We define $\xi_t$ by,
\begin{equation}\begin{aligned}
\xi_t = -\omega_t^{\tau}\cdot\Delta_t,
\end{aligned}\end{equation}
where $\omega_t^{\tau}$ is the corresponding parameter. The response of the root-node can be then defined as,
\vspace{-3mm}
\begin{equation}\begin{aligned}
\!\!R^r(X,\mathbf{V},\mathbf{\Delta})\!=\omega^r\cdot\phi^r(X) \!+ \sum^T_{t=1}R^a_t(\Lambda_t,V_t) + \xi_t,\label{eq:root}
\end{aligned}\end{equation}
where $\phi^r(X)$ is the BoW histogram feature extracted from the whole video sequence $X$. $\mathbf{V} = (V_1,\cdots,V_T)$ and $\mathbf{\Delta} = (\Delta_1,\cdots,\Delta_T)$ are latent variables in the model for specifying the spatial and temporal configurations.
\begin{figure}[ptb]
\begin{center}
\epsfig{figure=Interaction1,width = 3.2in}
\end{center}
\vspace{-4ex}
\caption{Illustration of the contextual relations for defining spatial edges in the STAOG model. We define the edges between spatial adjacent leaf-nodes with $8$ relations according to their spatial layout: above, below, left, right, near, far, clockwise and anti-clockwise.}
\label{fig:SpaInteraction}
\end{figure}
\subsection{Contextual Interactions}
\label{sec:context}
\textbf{Spatial interactions.} We impose spatial contextual interactions, \emph{i.e.} spatial edges, between pairwise spatially adjacent leaf-nodes in each anchor frame, as Fig.~\ref{fig:spa}(c) illustrates. Note that we only link the edges between a pair of leaf-nodes that are respectively associated with two different or-nodes.
For one edge connecting two leaf-nodes $(L_j,L_{j'})$, we define it with a set of informative relations, \emph{i.e.} a 8-bin binary feature $\varphi^s(L_j,L_{j'})$: \emph{above}, \emph{below}, \emph{left}, \emph{right}, \emph{near}, \emph{far}, \emph{clockwise} and \emph{anti-clockwise} between two adjacent leaf-nodes. The relations are visualized in Fig.~\ref{fig:SpaInteraction}. Suppose one edge connects two leaf-nodes $(L_j,L_{j'})$ which detect action parts at positions $p_j$ and $p_{j'}$ respectively. The centered red rectangle represents the location $p_j$, and the other red rectangles represent the adjacent parts. In the right chart of Fig.~\ref{fig:SpaInteraction}, the dashed line represents the initial layout of the two leaf-nodes, and the black solid line the adjusted actual layout during inference. Then we define the relations as,
\begin{itemize}
\setlength{\itemsep}{1pt}
\setlength{\parskip}{0pt}
\setlength{\parsep}{0pt}
\item \emph{near} or \emph{far}: If $p_{j'}$ is fallen into the outer dashed ellipse, it is near to $p_j$, \emph{i.e.} the bin of near is activated (\emph{i.e.} set as 1); otherwise it is far to $p_j$.\\
\item \emph{above}, \emph{below}, \emph{left} or \emph{right}: The corresponding bin is set as 1 only if the center of $p_{j'}$ is inside the corresponding dashed rectangles.\\
\item \emph{clockwise} or \emph{anti-clockwise}: one of the two relations is activated (\emph{i.e.} set as 1) according to the angle between the dashed line and the black solid line.\\
\vspace{-2ex}
\end{itemize}
The relations intuitively encode the spatial contexts of two action parts detected via the two leaf-nodes with respect to two different or-nodes. The response of the pairwise potentials can be parameterized as,
\vspace{-2mm}
\begin{equation}\begin{aligned}
\Gamma^s_{jj'} = \beta^s_{jj'}\cdot\varphi^s(L_j,L_{j'}),\label{eq:spaI}
\end{aligned}\end{equation}
where $\beta^s_{jj'}$ is the corresponding 8-bin parameter vector.
\textbf{Temporal interactions.} We also impose the edges in temporal domain in our model to represent the temporal interactions of action parts. The edges connect temporally adjacent leaf-nodes, illustrated in Fig.~\ref{fig:temp}. The edges are connected between any pair of leaf-nodes $(L_j,L_{j'})$ that belong to the same part within the two adjacent anchor frames respectively. A set of temporal relations are collected to concatenate a $4$-bin binary feature vector $\varphi^{\tau}(L_j,L_{j'})$
\begin{figure}[ptb]
\begin{center}
\epsfig{figure=temp,width=3.4in}
\end{center}
\vspace{-3mm}
\caption{Illustration of temporal compositions. The input video is decomposed into a number of discrete temporal anchor frames. The optimal position of each anchor frame is localized in $\Delta_t + \tau_t$, \emph{i.e.} the temporal displacement $\Delta_t$ plus the initial anchor point $\tau_t$. The temporal contextual interactions are defined in temporally adjacent action parts.}
\label{fig:temp}
\end{figure}
Specifically, we adopt four predicates: \emph{intersect}, \emph{after}, \emph{meets}, \emph{interrupt}, inspired by Allen's temporal predicates~\cite{AllenF94}~\cite{Spatiotemporalrelationship}. These predicates describe relations between two time intervals. The action part detected by one leaf-node $L_j$ for a specific anchor frame is described by the feature $\phi^l$ extracted from a 3-D volume with time span $\rho$. Note that we ignore some predicates of ordering such as \emph{before} and \emph{equals}, as the order of temporally adjacent anchor frames is supposed to be fixed. Assume that leaf-node $L_j$ is associated with the anchor frame localized at $\tau_t + \Delta_t$, (initial position plus displacement). The starting and ending time $(f_{L_j}^{start}, f_{L_j}^{end})$ of $L_j$ can be calculated as,
\vspace{-3ex}
\begin{equation}\begin{aligned}
f_{L_j}^{start} &= \tau_t + \Delta_t - \frac{\rho}{2},\\
f_{L_j}^{end} &= \tau_t + \Delta_t + \frac{\rho}{2}.
\end{aligned}\end{equation}
Then we can define the four temporal predicates for the two temporal adjacent leaf-nodes $(L_j,L_{j'})$ as,
\begin{equation}\begin{aligned}
\emph{intersect}(L_j, L_{j'}) &\Longleftrightarrow f_{L_{j'}}^{start} < f_{L_j}^{end},\\
\emph{after}(L_j, L_{j'}) &\Longleftrightarrow f_{L_j}^{end} < f_{L_{j'}}^{start} < f_{L_j}^{end} + \rho,\\
\emph{meets}(L_j, L_{j'}) &\Longleftrightarrow f_{L_{j'}}^{start} = f_{L_j}^{end},\\
\emph{interrupt}(L_j, L_{j'}) &\Longleftrightarrow f_{L_j}^{end}+ \rho< f_{L_{j'}}^{start}.
\end{aligned}\end{equation}
Thus, we define the response of one temporal edge linking two leaf-nodes accordingly,
\begin{equation}\begin{aligned}
\Gamma^{\tau}_{jj'} = \beta^{\tau}_{jj'}\cdot\varphi^{\tau}(L_j,L_{j'}),\label{eq:temI}
\end{aligned}\end{equation}
where $\beta^{\tau}_{jj'}$ is the corresponding 4-bin parameter. If the pairwise leaf-nodes $(L_j,L_{j'})$ satisfies the specific predicate, the corresponding bin is set to 1, otherwise 0.
Therefore, the overall response of the STAOG model is:
\begin{equation}\begin{aligned}
R^g(X,\mathbf{V},\mathbf{\Delta})&= R^r(X,\mathbf{V},\mathbf{\Delta})+ \sum^n_j(\sum_{j'\in \gamma^s(j)}\Gamma^s_{jj'}\cdot v_j\cdot v_{j'}\\&+ \sum_{j'\in \gamma^{\tau}(j)}\Gamma^{\tau}_{jj'}\cdot v_j\cdot v_{j'}),\label{eq:graph}
\end{aligned}\end{equation}
where $(\mathbf{V},\mathbf{\Delta})$ are the hidden variables in the STAOG model. The second term defines the spatio-temporal contextual interactions between the leaf-nodes. $\gamma^s(j)$ is denoted as the set of leaf-node $L_j$'s neighbors which are spatially adjacent to the leaf-node $L_j$, and $\gamma^{\tau}(j)$ is introduced for the leaf-nodes which are temporally adjacent to the leaf-node $L_j$. Intuitively, spatial interactions between action parts guarantee the spatial coherence, as well as temporal interactions embedding the temporal contextual relations. Briefly, we refer $\mathcal{L} = (\mathbf{V},\mathbf{\Delta})$ as the latent variables in the following. The Eq.\ref{eq:graph} can be briefly written as:
\begin{equation}\begin{aligned}
\!R^g(X,\mathcal{L}) = \psi\cdot\Phi(X,\mathcal{L}),\label{eq:global}
\end{aligned}\end{equation}
where $\psi$ includes the complete parameters of the STAOG model, and $\Phi(X,\mathcal{L})$ denotes the overall feature vector.
\section{Inference}
The inference task is to detect $T$ optimal temporal anchor frames for one video instance as well as the spatial composition of action parts within each anchor frame. In our approach, we perform a cascaded search that integrates three steps: spatial testing, temporal testing and global verification to maximize the global potential $R^g(X,\mathcal{L})$ defined in Eq.\ref{eq:graph}.
\textbf{Step 1. Spatial Testing via the and-nodes.}
The subgraph of the STAOG model, rooted at one and-node, can be viewed as the spatial composition classifier for localizing action parts in one frame. We first use all existing leaf-nodes to search for candidate actions parts. Assume the leaf-node $L_j$ associated with the frame $I_t$ detects the action part at the position $p_j^*$ by maximizing the response in Eq.\ref{eq:leafnode}. Each or-node is allowed to activate only one leaf-node, then a possible configuration consisting of action parts is decided by the indicator variables of the or-node, (\emph{i.e.} $\mathbf{v_i}$ for or-node $U_i$, indicating which leaf-node is activated). In this way, a set of possible configuration hypotheses $\{V_t\}$ are generated for further testing, which ensemble the hypotheses proposed by the or-nodes for the frame $I_t$. In practice, we limit the maximum number of hypotheses by setting a threshold on $R^a_t(\Lambda_t,V_t)$ in Eq.\ref{eq:andnode}.
\begin{small}
\begin{algorithm}[htb]
\caption{Inference Algorithm}
\label{alg:Inference}
\begin{algorithmic}\footnotesize
\INPUT ~~\\
A learned STAOG model $G$, the action parts detected by all leaf-nodes by maximizing the response in Eq.\ref{eq:leafnode}.
\INITIAL ~~\\
The set of possible hypotheses $l_t = \{\}$ for all $t\in{1,\dots T}$ anchor frames.
\INTERATION~~\\
\FORALL {$t = 1 \cdots T$ }
\STATE For each and-node $A_t$, a set of temporal displacement steps $\Sigma$ is predefined for sliding the possible anchor frames.
\FORALL {$\Delta_t \in \Sigma$ }
\STATE
\begin{itemize}
\setlength{\itemsep}{1pt}
\setlength{\parskip}{0pt}
\setlength{\parsep}{1pt}
\item[(a)] initialize the set of pair terms $Q = \{Q_1,\dots,Q_K\}$ for all or-nodes of $A_t$'s children.
\item[(b)] generate a set of pair terms $Q_i$ for each or-node $U_i$.
\FORALL {$U_i,i \in ch(t)$}
\FORALL{$L_j,j\in ch(i)$}
\STATE
$Q_i = Q_i \cup (i,j)$.
\ENDFOR
\ENDFOR~~\\
\item[(c)] obtain possible hypotheses $V_t = (\mathbf{v}_1,\dots,\mathbf{v}_K)$ by assembling the indicator variables of $K$ or-nodes according to the set $Q$.
\item[(d)] The set of possible hypotheses for each specific displacement $\Delta_t$ is constructed as $l_t = l_t \cup (\{V_t\},\Delta_t)$.
\end{itemize}
\ENDFOR
\ENDFOR~~\\
Assemble these hypothesis \{$l_t\}$ for all $T$ anchor frames orderly to generate the set of hypotheses sequence $l$. Each possible configuration $(\mathbf{V},\mathbf{\Delta})$ belongs to the set $l$. The global response of STAOG model can be calculated by Eq.\ref{eq:final}.
\OUTPUT ~~\\
\STATE
The latent variables $\mathbf{V},\mathbf{\Delta}$ and the final score $S_\psi(X)$.
\end{algorithmic}
\end{algorithm}
\end{small}
\textbf{Step 2. Temporal Testing via the root-node.}
We apply the spatial testing with the and-nodes to localize a number of candidate anchor frames over several frames. The scores over candidate anchor frames (proposed by the and-nodes) are aggregated via the root-node for a possible temporal composition. For efficiency, we utilize a fixed number of discrete steps $\Sigma$ for searching each anchor frame. Several possible hypotheses are then produced with different anchor frame determinations by sliding the discrete steps $\Sigma$. In addition, we re-weight the hypotheses at the root-node in Eq.\ref{eq:root} by considering the temporal displacements of anchor frames as well as the the global features over the video clip. Intuitively, the hypotheses are represented by the specified latent variables $(\mathbf{V},\mathbf{\Delta})$.
\textbf{Step 3. Global Verification.}
Given all the hypotheses from the root-node, we apply the global potential function defined in Eq.\ref{eq:graph} to validate the optimal detection. The objective of the global verification is to cope with the noisy local detections on leaf-nodes. It combines the score of the root-node with the responses of spatial and temporal contextual interactions (edges).
The optimal response $S_\psi(X)$ of the model as well as the latent variables $(\mathbf{V},\mathbf{\Delta})$ can be calculated as,
\begin{equation} \begin{aligned}
S_\psi(X) = \mathop{max}_{\mathbf{V},\mathbf{\Delta}}(\psi\cdot\Phi(X,\mathbf{V},\mathbf{\Delta})).\label{eq:final}
\end{aligned} \end{equation}
Algorithm \ref{alg:Inference} summarizes the overall algorithm of the inference.
\section{Structural Learning}
We formulate the structural learning of STAOG model as a joint optimization task for model structure and parameters. We solve this model by a novel latent learning method extended from the CCCP framework~\cite{cccp}. This algorithm iterates to train the model in a dynamic manner: the leaf-nodes can be automatically created or removed to reconfigure the model structure. The model structure is determined by latent variables $\mathcal{L} = (\mathbf{V},\mathbf{\Delta})$ that are inferred in each step.
\begin{figure*}[ptb]
\begin{center}
\epsfig{figure=learning,width = 1\textwidth}
\end{center}
\vspace{-4mm}
\caption{Illustration of discriminative structural learning. We reconfigure the model structure by re-arranging the feature vector, as the example illustrated. Parts of the STAOG model reconfigured in two iterations are shown in (a), where the left one represents the original model and the other one the new model. During this step, the new leaf-node associated with $U_7$ and $U_{11}$ are created and a leaf-node associated with $U_8$ is removed. Assume that we use 5 samples, $X_1,\dots,X_5$, for the structure learning. (b) shows the feature vectors detected by the same leaf-node are first grouped into one cluster, \emph{i.e.} one cluster for one leaf-node. (c) illustrates the feature rearrangement after clustering. For example, the feature vector of sample $X_1$ is grouped from cluster $\Omega_9$ into cluster $\Omega_{13}$, we move the feature bins $\pi_9^1$ into the bins corresponding to the leaf-node $L_9$. Cluster $\Omega_{14}$ is a newly generated cluster, we thus create a new leaf-node accordingly.}
\label{fig:learning}
\end{figure*}
Let $D = ((X_1,y_1),(X_2,y_2),\dots,(X_N,y_N))$ be a set of labeled training samples, with $y_{k} \in \{1,-1\}$. The feature vector for each sample $(X,y)$ is defined as,
\begin{equation} \begin{aligned}
\Phi(X,y,\mathcal{L}) = \begin{cases}
\Phi(X,\mathcal{L})& \text{if $y=+1$},\\
0& \text{if $y=-1$}.
\end{cases}
\end{aligned} \end{equation}
The temporal anchor frames and spatial configurations for each frame can be optimized by maximizing $R^g(X,\mathcal{L})$ in the inference procedure. We refer to $\mathcal{L} = (\mathbf{V},\mathbf{\Delta})$ as the latent variables, then we redefine Eq.\ref{eq:final} as,
\begin{equation} \begin{aligned}
S_\psi(X) = \mathop{max}_{y,\mathcal{L}}(\psi\cdot\Phi(X,y,\mathcal{L})),
\end{aligned} \end{equation}
The optimization of this function can be solved by the latent structural SVM framework,
\begin{equation} \begin{aligned}
\mathop{min}_\psi\frac{1}{2}\|\psi\|^2&+C\sum_{k=1}^N\lbrack \mathop{max}_{y,\mathcal{L}}(\psi\cdot\Phi(X_k,y,\mathcal{L}) + h(y_k,y))\\&- \mathop{max}_{\mathcal{L}}(\psi\cdot\Phi(X_k,y_k,\mathcal{L}))\rbrack,\label{eq:cost}
\end{aligned} \end{equation}
where $C$ is a penalty parameter set as 0.003 empirically and $h(y_k,y)$ is the cost function, where $h(y_k,y) = 0$ if $y_k = y$, otherwise 1. The optimization problem described above is not convex in general. Following the CCCP framework, we convert the function in Eq.\ref{eq:cost} into a convex and concave form as,
\begin{equation} \begin{aligned}\label{eq:cccp}
&\mathop{min}_\psi\lbrack\frac{1}{2}\|\psi\|^2 + C\sum_{k=1}^N \mathop{max}_{y,\mathcal{L}}(\psi\cdot\Phi(X_k,y,\mathcal{L}) + h(y_k,y))\rbrack\\
&- \lbrack C\sum_{k=1}^N \mathop{max}_{\mathcal{L}}(\psi\cdot\Phi(X_k,y_k,\mathcal{L}))\rbrack\\
&= \mathop{min}_\psi\lbrack f(\psi) - g(\psi)\rbrack,
\end{aligned} \end{equation}
where $f(\psi)$ represents the first two terms, and $g(\psi)$ the last term. This leads to an iterative learning algorithm that alternates estimating model parameters and the hidden variables $\mathcal{L}$. However, we still need to dynamically determine the graph configuration, \emph{i.e.} the production of leaf-nodes associated with or-nodes. An additional step for dynamically reconfiguring structure is added between the two original steps. The procedure is presented as follows.
\begin{small}
\begin{algorithm}[htb]
\caption{Learning algorithm for STAOG model}
\label{alg:Framwork}
\begin{algorithmic}\footnotesize
\INPUT ~~\\
Training samples,$D = ((X_1,y_1),(X_2,y_2),\dots,(X_N,y_N))$.
\OUTPUT ~~\\
The trained STAOG model .
\INITIAL ~~\\
\STATE
\begin{itemize}
\setlength{\itemsep}{1pt}
\setlength{\parskip}{0pt}
\setlength{\parsep}{1pt}
\item[1.] Initialize the positions of action parts and anchor frames for all samples.
\item[2.] Initialize the latent variables $\mathcal{L}$ and parameters $\psi$.
\end{itemize}
\MYWHILE
\STATE
\begin{itemize}
\setlength{\itemsep}{1pt}
\setlength{\parskip}{0pt}
\setlength{\parsep}{1pt}
\item[1.] Estimate the latent variable $\mathcal{L}^*_k = (\mathbf{V}^*_k,\mathbf{\Delta}^*_k)$ for each positive sample $(X_k,y_k)$ during inference.
\item[2.]
\begin{itemize}
\setlength{\itemsep}{1pt}
\setlength{\parskip}{0pt}
\setlength{\parsep}{1pt}
\item[(a)] Localize the anchor frames and action parts using the current latent variables $(\mathbf{V}^*,\mathbf{\Delta}^*)$.
\item[(b)] For each or-node $U_i$, we obtain a set of feature vectors of its children leaf-nodes for positive samples, and regroup the feature vectors by Spectral clustering.
\item[(c)] Reconfigure leaf-nodes according to clustering results to generate a new structure. Calculate the energy Eq.~\ref{eq:opt} as $E(\psi_t^d)$.
\end{itemize}
\item[3.] \begin{itemize}
\setlength{\itemsep}{1pt}
\setlength{\parskip}{0pt}
\setlength{\parsep}{1pt}
\IF{$E(\psi_t^d)<E(\psi_t)$}
\STATE
Accept the new model structure and estimate the parameters $\psi_{t+1}=\mathop{argmin}_\psi(f(\psi+\psi\cdot q_t^d))$.
\ELSE
\STATE
keep the previous model structure. Estimate the parameters $\psi_{t+1}=\mathop{argmin}_\psi(f(\psi+\psi\cdot q_t))$.
\ENDIF
\end{itemize}
\end{itemize}
\MYENDWHILE
The optimization function defined in Eq.~\ref{eq:cost} converges.
\end{algorithmic}
\end{algorithm}
\end{small}
(I) The model parameter $\psi_t$ obtained in the previous iteration is fixed. We find a hyperplane $q_t$ to upper bound the concave part $g(\psi)$ in Eq.~\ref{eq:cccp}. Specifically, $q_t$ is the derivative of $g(\psi)$. Thus, we have $-g(\psi)\leq-g(\psi_t) + (\psi-\psi_t)\cdot q_t,\forall\psi$. The optimal latent variable $\mathcal{L}_k^*$ is calculated by $\mathcal{L}_k^* = \mathop{argmax}_{\mathcal{L}}(\psi_t\cdot\Phi(X_k,y_k,\mathcal{L}))$ for each positive example. Then the hyperplane is constructed as $q_t= - C\sum_{k=1}^N\Phi(X_k,y_k,\mathcal{L}_k^*)$.
(II) In the second step, the STAOG model is adjusted by structural reconfiguration with applying the current model on training examples. The reconfiguration is performed for each part, \emph{i.e.} or-node, independently, with the fixed latent variable $\mathcal{L}^*_k = (V^*_k,\Delta^*_k)$. Note that each action part detected by one leaf-node is mapped to several feature bins at specific positions in the vector $\Phi(X_k,y_k,\mathcal{L}_k^*)$.
For each or-node $U_i$, we apply its children leaf-nodes for detecting parts in all positive samples. Assume that leaf-node $L_j$ detects an action part on the $k$-th sample using the feature vector $\pi_j^k$, which is a sub-vector of the complete feature vector of $k$-th sample, $\Phi(X_k,y_k,\mathcal{L}_k^*)$. And we obtain a set of feature vector $\{ \pi_j^k \}$ for all samples. The vectors detected by the same leaf-node are first grouped into one cluster, \emph{i.e.} one cluster for one leaf-node. We denote the cluster for the $j$-th leaf-node as $ \Omega_j$. Then we perform the spectral clustering algorithm with the Euclidean distance on vectors of all leaf-nodes of or-node $U_i$'s children for all positive samples, and the similar vectors are grouped together. We re-arrange the feature vectors for all samples based on the newly generated partition. For example, if the feature vector $\pi_j^k$ is grouped from $ \Omega_j $ into another cluster $ \Omega_j'$, we adjust the position of $\pi_j^k$ in $\Phi(X_k,y_k,\mathcal{L}_k^*)$, \emph{i.e.} by moving the feature bins into the position representing $j'$-th leaf-node. If $\Omega_j' $ is a newly generated cluster, we thus create a new leaf-node accordingly. By analogy, we remove one leaf-node if few samples are grouped into the corresponding cluster. In this way, the structure of $U_i$ is reconfigured with the feature vector re-arrangement. In practice, we constrain the extent of structural reconfiguration, \emph{i.e.} only few leaf-nodes can be created or removed in one iteration. We present a toy example in Fig.\ref{fig:learning} for illustration. In Fig.\ref{fig:learning}(a), a leaf-node associated with the or-node $U_7$ is created to better handle the intra-class variance; A leaf-node is removed if there is another similar one, (\emph{e.g.} the leaf-node associated with the or-node $U_8$). The sub-vector of $\pi_9^1$ of sample $X_1$ is grouped from cluster $\Omega_{9}$ to cluster $\Omega_{13}$, then the feature bins are moved from $\pi_9^1$ to $\pi_{13}^1$ as Fig.\ref{fig:learning}(b) shows.
After this structure reconfiguration, we obtain the new feature vector for each sample, $\Phi^d(X_k,y_k,\mathcal{L}_k^*)$, and the hyperplane is re-calculated as $q^d_t= - C\sum_{k=1}^N\Phi^d(X_k,y_k,\mathcal{L}_k^*)$, accordingly.
(III) The newly generated model structure can be represented by the feature vector $\Phi^d(X_k,y_k,\mathcal{L}_k^*)$. The model parameters can be learned by solving $\psi_{t}^d=\mathop{argmin}_\psi(f(\psi+\psi\cdot q_t^d))$. The optimization task in Eq.\ref{eq:cost} becomes,
\begin{equation} \begin{aligned}
& \mathop{min}_\psi\frac{1}{2}\|\psi\|^2 + C\sum_{k=1}^N\lbrack \mathop{max}_{y,\mathcal{L}}(\psi\cdot\Phi(X_k,y,\mathcal{L}) + h(y_k,y))\\
&- \psi\cdot\Phi^d(X_k,y_k,\mathcal{L}_k^*))\rbrack.\label{eq:opt}
\end{aligned} \end{equation}
This is a standard structural SVM problem, which can be solved in the cutting plane method and Sequential Minimal Optimization. The energy Eq.\ref{eq:opt} can be calculated by $E(\psi_t^d) = f(\psi_t^d)-g(\psi_t^d)$. We accept the new model structure until $E(\psi_t^d)<E(\psi_t)$ and $\psi_{t+1} = \psi_{t}^d$. Otherwise, we keep the model structure as in the previous iteration and the parameter vector is calculated by $\psi_{t+1}=\mathop{argmin}_\psi(f(\psi+\psi\cdot q_t))$. Thus, we ensure that the optimization function will decrease in each iteration. We repeat the 3-step iteration until convergence. Algorithm \ref{alg:Framwork} summarizes the overall algorithm of learning a STAOG model. In the case of multi-class classification, we use a one-against-rest approach and select the class with the highest score.
\section{Experiments}
We test our STAOG model on two different action recognition databases: UCF YouTube~\cite{youtube} and Olympics Sports~\cite{ModelTemp}. The video resolution is normalized to $320 \times 240$. The YouTube dataset contains $11$ action categories, which is challenging due to large variation in camera motion, object appearance, pose/view and large intra-class variability. We follow the standard setup using leave-one-out cross validation for a pre-defined set of 25 folds.Average accuracy over all classes is reported as performance measure. The Olympic Sports dataset consists of 16 different sports classes that contain complex motions going beyond simple punctual or repetitive actions. The challenges of Olympic sports arise from background clutters, viewpoints and complex sequence of primitive actions. Each action is performed only by a single actor and represents a temporal sequence of primitive actions (\emph{e.g.} triple-jumping, pole-vault and diving). We use the same train-test split setup and the average precision (AP) for each of the action classes as in ~\cite{LatentTem_CVPR2012}.
\subsection{Implementation}
\label{sec:imple}
\begin{small}
\begin{table*}
\centering
\begin{tabular}{| c | c | c | c | c | c | c | c | }
\hline\hline
& Liu et al~\cite{youtube}.
& Zhang et al~\cite{SpatioTemporalPhrases}.
& Ikizler-Cinbis et al~\cite{Ikizler-CinbisS10}.
& Dense trajectories~\cite{denseTraj}.
& Ours-1
& Ours-2
& Ours(full)\\
\hline
\hline
b\_shoot & 53.0\% & \textbf{98.0}\% & 48.5\% & 43.0\% & 58.4\% & 62.0\% & 77.9\%\\
bike & 73.0\% & 74.0\% & 75.2\% & \textbf{91.7\%} & 82.1\% & 87.3\% & 88.6\%\\
dive & 81.0\% & 80.0\% & 95.0\% & \textbf{99.0\%} & 98.4\% & 98.6\% & 98.8\%\\
golf & 86.0\% & 68.0\% & 95.0\% & {97.0\%} & 95.7\% & 95.3\% & \textbf{97.4\%}\\
h\_ride & 72.0\% & 65.0\% & 73.0\%& 85.0\% & 81.3\% & 86.0\% & \textbf{88.0\%}\\
s\_juggle & 54.0\% & 67.0\% & 53.0\% & 76.0\% & 66.0\% & 81.6\% & \textbf{82.2}\%\\
swing & 57.0\% & 71.0\% & 66.0\% & \textbf{88.0\%} & 85.2\% & 84.1\% & 85.4\%\\
t\_swing & 80.0\% & 68.0\% & 77.0\% & 71.0\% & 69.6\% & 80.0\% & \textbf{80.7\%}\\
t\_jump & 79.0\% & 80.0\% & 93.0\% & 94.0\% & 90.2\% & {94.7\%} & \textbf{95.8\%}\\
v\_spike & 73.3\% & 77.0\% & 85.0\% & 95.0\% & 89.7\% & 90.6\% & \textbf{96.4\%}\\
walk & 75.0\% & 54.0\% & 66.7\% & 87.0\% & 85.2\% & 86.2\% & \textbf{87.4}\%\\
\hline
\hline
Accuracy & 71.2\% & 72.9\% & 75.2 \% & 84.2\% & 82.0\%& \textbf{86.0\%} & \textbf{88.9\%}\\
\hline
\end{tabular}
\caption{Accuracy per action class and average accuracy for all classes on the YouTube dataset~\cite{youtube}.}\label{tab:youtube}
\end{table*}
\end{small}
\begin{figure*}[ptb]
\begin{center}
\epsfig{figure=part, width =6.5in}
\end{center}
\vspace{-4mm}
\caption{Example inference results on three different action models \emph{e.g.} basketball-layup (a), clean-and-jerk (b) and triple-jump(c) learned on the Olympic Sports dataset and each of action category includes two instances. The red boxes in each frame represent the discovered discriminative action parts. Our model successfully localizes the accurate anchor frames across the instances in the long action videos. In addition, it is noticed that the large intra-class variabilities can be captured by our model.}
\label{fig:part}
\end{figure*}
We fix the number of and-nodes (anchor frames) in the STAOG model as $T = 3$ for UCF YouTube dataset, and $T = 5$ for Olympic Sports dataset empirically. We need more anchor frames for Olympic dataset because the actions are more complex and last over more frames. The parameter $T$ can be roughly estimated by the action temporal complexity in general. The number of spatial layout for each anchor frame is fixed as $2 \times 2$, and thus $K = 4$ or-nodes for each anchor frame. There are at most $m = 4$ leaf-nodes associated to each or-node. We extract the interest points described by the HOG and HOF features by utilizing the code published in~\cite{realisticLap} beforehand. The size of one action part (within a 3-D volume) is empirically set to $60 \times 60$ pixels spanning $\rho=15$ frames. The dimension of the generated dictionary is set as $300$ for describing action parts in each anchor frame and the global features of the anchor frames. We set the discrete temporal steps searching for anchor frames: $\Sigma = [\pm{2}, \pm{4}, \pm{6}, \pm{8}, \pm{10}]$. The convergence of our learning algorithm usually takes $9\sim 10$ iterations.
\begin{small}
\begin{table}
\centering
\begin{tabular}{| c | c | c | c | }
\hline\hline
& Niebles et al~\cite{ModelTemp}.
& Tang et al~\cite{LatentTem_CVPR2012}.
& Ours(full)\\
\hline
\hline
h-jump & 27.0\% & 18.4\% & \textbf{35.6\%}\\
l-jump & 71.7\%& 81.8\% & \textbf{86.4\%} \\
t-jump & 10.1\%& 16.1\% & \textbf{36.2\%}\\
p-vault & \textbf{90.8}\% & 84.9\% & 84.3\% \\
g-vault & \textbf{86.1}\% & 85.7\% & 83.1\% \\
s-put & 37.3\% & 43.3\% & \textbf{56.8\%} \\
snatch & 54.2\% & {88.6\%} & \textbf{89.0\%}\\
c-jerk & 70.6\%& 78.2\% & \textbf{83.3\%}\\
j-throw & \textbf{85.0}\% & 79.5\% & 78.1\%\\
h-throw & 71.2\%& 70.5\% & \textbf{75.4\%}\\
d-throw & 47.3\% & 48.9\% & \textbf{53.3\%}\\
d-platform & \textbf{95.4}\% & 93.7\% & 92.8\%\\
d-board & \textbf{84.3}\% & {79.3\%} &76.5\%\\
basketball & 82.1\% & 85.5\% & \textbf{86.7\%}\\
bowling & 53.0\% & \textbf{64.3\%} & 62.0\%\\
t-serve & 33.4\% & 49.6\% &\textbf{62.3\%}\\
\hline
\hline
mAP & 62.5\% & 66.8\% & \textbf{71.4\%}\\
\hline
\end{tabular}
\vspace{-4mm}
\caption{Average Precision(AP) values on the Olympic Sports dataset~\cite{ModelTemp}.}\label{tab:olym}
\end{table}
\end{small}
The experiments are carried out on a PC with Core I5 3.0GHZ CPU and 4GB memory. The average CPU-time used to process a video from Olympic Sports dataset is $200$ seconds and $150$ for a video in the UCF YouTube dataset. In particular, it takes 7 seconds for processing one frame in Olympic Sports dataset, and 4 seconds for each frame in the UCF YouTube dataset on average. The efficiency of our method is slightly diverse with the densities of the feature points in the video sequence.
\subsection{Results and Comparisons}
Compared with the recently proposed methods on the YouTube dataset, our model outperforms the state-of-the-art: we achieve the accuracy of 88.9\% in YouTube dataset, the reported results of the competing algorithms are: 71.2\% in~\cite{youtube}, 72.9\% in ~\cite{SpatioTemporalPhrases},75.2\% in~\cite{Ikizler-CinbisS10} and 84.2\% in~\cite{denseTraj}. The accuracy scores for all categories are reported in Table~\ref{tab:youtube}. Our method outperforms on $7$ out of the $11$ categories which have relatively large intra-class variance or background disturbance. In the Olympic Sports dataset, we obtain better AP scores for $10$ out of the $16$ categories, and overall AP score $71.4\%$, better than the previous methods~\cite{LatentTem_CVPR2012,ModelTemp}. The competing method proposed by Tang et al.~\cite{LatentTem_CVPR2012} utilizes the variable-duration HMM for learning the temporal structure in the video. Our results show that our compositional model with the explicit spatial and temporal relations can achieve better performance. The detailed results are reported in Table~\ref{tab:olym}.
Figure.~\ref{fig:part} illustrates the inference results on three action categories from the Olympic Sports dataset, each of which includes two instances. Our model also localizes the action parts in the anchor frames, as they are discriminative in appearance and motion. These results demonstrate well the capability of our model, since the scenarios of actions contain the very realistic challenges in video action recognition. The spatial compositions defined over the and-nodes enable us to handle pose/view variations and background disturbances. The temporal compositions are effective to localize anchor frames in videos against various motion frequency, temporal locations and video length.
For further evaluation, we conduct three empirical analysis in different model settings as follows.
(I) We simplify the temporal compositions by discarding the displacement $\Delta_t$ for each anchor frame, \emph{i.e.} fixing the temporal structure. We report AP scores of this model setting in the fourth column of Table~\ref{tab:youtube}, named as "Ours-1". The average accuracy is $82\%$,$6.9\%$ less than our complete model.
(II) We also evaluate the benefit of spatio-temporal contextual interactions. Our model can be simplified into a tree structure by removing the interactions. The accuracies are shown in the sixth column of Table~\ref{tab:youtube}, named as "Ours-2". We can observe that the spatio-temporal contextual interactions make the accuracies increase $2.9\%$ on average. The increased performance of the interactions also speaks in favour for our model, as it shows that through better associations between the anchor frames and action parts, it is possible to achieve even better accuracies.
(III) One may be interested in how the performances are improved by introducing the or-nodes in spatial compositions, which is one of the key components in the STAOG model. In this experiment, we set different maximum numbers $m$ of leaf-nodes under the or-nodes, \emph{i.e.} how many leaf-nodes at most can be created for the model. We compare the results on the YouTube dataset in Fig.~\ref{fig:empirical_leafnodes}, and observe that $m = 4$ achieves the better results in general than $m=2$, $82.4\%$. In practice, the number $m$ is not sensitive, as the exact number of leaf-nodes is decided by the clustering on data during the structural learning.
\begin{figure}[ptb]
\begin{center}
\epsfig{figure=leafnodes, width = 3.4 in}
\end{center}
\vspace{-5ex}
\caption{Empirical analysis for different settings of spatial compositions, where we set different maximum numbers $m$ of leaf-nodes under the or-nodes. Each pillar represents the accuracy for one action category. The color indicates the results with different settings, $m = 2$ or $m = 4$.}
\label{fig:empirical_leafnodes}
\end{figure}
\section{Conclusion}
\vspace{-1mm}
This paper studies a novel hierarchical model for human action recognition, in the form of a configurable Spatio-Temporal And-Or Graph. This model is shown to handle well realistic challenges in action recognition. Moreover, we consider two aspects to improve our method. First, the model can be integrated with high level semantic information to represent multi-agent complex events. Second, we plan to speed up the algorithm for large-scale processing.
\vspace{-2mm}
\bibliographystyle{abbrv}
|
\section{Introduction}
Coherent superpositions of energy eigenstates are a powerful resource for quantum technologies, with applications to quantum metrology \cite{metrology,metrology2,metrology3}, computing \cite{computing,computing1, computing2,computing3},
and even biological systems \cite{biology}. Not surprisingly, great attention has been devoted to the characterization and quantification of coherence in a variety of contexts \cite{delocal,incoherence,reference,vaccaro2008tradeoff,marvian-spekkens1,marvian-spekkens2,spekkens-marvian-WAY,vaccaro2011particle,ahmadi,marvian-spekkensNat,rudolphPRX,aaberg-2014-prl,cwiklinski-2015-prl}.
In order to regard coherence as a resource, a canonical approach is to define
a set of \emph{free operations}, whose execution is regarded as inexpensive \cite{coecke}.
Depending on the physical scenario at hand, many choices are possible for the set of free operations, including~e.~g.~covariant \cite{reference,marvian-spekkens1,marvian-spekkens2,marvian-spekkensNat} and incoherent channels \cite{incoherence,delocal}. A uniquely motivated class of free operations consists of those operations that can be implemented in principle without consuming external energy sources---this is the case, for example, of the operations that can be achieved with passive optical elements like beam splitters, phase shifters, and mirrors. We refer to these operations as \emph{energy-preserving}.
Classically, energy-preserving operations can only transfer the system from a phase space point to another phase space point with equal energy.
In the quantum case, the situation is more interesting, since energy-preserving operations can act nontrivially on coherent superpositions. The goal of this paper is to explore the potential of energy-preserving operations for the manipulation of quantum coherence.
The typical question in a resource theory is whether or not a given input resource can be transformed into a desired output resource by means of free operations.
When a certain transformation of resources turns out to be impossible, a natural step further is to ask \emph{how well} it can be approximated with free operations. This step is similar in spirit to the step from the no-cloning theorem \cite{nocloning1,dieks} to the study of optimal cloning \cite{buzekhillery,gisinmassar}.
In this paper we will search for the optimal transformations of quantum coherence at zero energy cost. We will start from the basic task of transforming a pure superposition of energy eigenstates into another pure superposition. In this context, what is truly interesting is to consider not only deterministic operations, but also probabilistic operations arising from measurements that can be implemented at zero energy cost.
By allowing for a non-unit probability of success, we find out that the constraint of energy preservation can be stretched to a previously unsuspected extent. For example, we will see that a beam of $N$ atoms, each of them prepared in the superposition $ |S\> =(|E\> + |G\>)/\sqrt 2$ of the ground state and the first excited state, can be probabilistically transformed at no energy cost into a stronger beam of up to $N^2$ atoms in a state that is nearly identical to the state $|S\>^{\otimes 2N}$, up to an exponentially small error.
The ability to efficiently approximate forbidden transformations of coherence at zero energy cost is a new twist of the postselection approach widely applied in quantum information \cite{weakvalues,aaronson,Bergou-Hillery1,Bergou-Hillery2,BaganDis,FiurasekEst,BaganPRA,BaganPRL,GiulioNat,fiurasekclon,QCMC,chiribella-xie}.
Our first result is a complete characterization of the optimal energy-preserving transformations and of their ultimate performances. After having characterized the structure of the optimal processes with a fixed probability of success, we move to a different scenario, where the probability of success is not fixed a priori. Specifically, we consider adaptive protocols whereby the experimenter performs repeated rounds of probabilistic operations and is free to decide on the fly whether to be content with the result obtained so far or whether to continue further. In this scenario we design a recursive protocol, consisting of a sequence of binary measurements that produce at each step the best approximation of the target with the highest probability allowed by quantum mechanics. Subsequent iterations of our protocol lead to an increasing probability of success, but also to a degradation of the state of the system and, eventually, to the loss of the advantages of postselection.
This behaviour is a consequence of the inevitable tradeoff between the enhanced performance of probabilistic transformations and their reduced probability of success. The advantage of our recursive protocol is that it gives an explicit, ready-to-apply method to construct lower bounds to the optimal tradeoff curve between fidelity and probability of success, a problem that so far has been solved only in one case \cite{BaganPRL}. Remarkably, in this case we find out that our protocol reproduces the optimal tradeoff curve, provided that the input state is a superposition of sufficiently many energy levels. We conjecture that the protocol is asymptotically optimal also in those situations where the input and the target consist of many identical copies of pure states,e.~g.~as an example being the asymptotic cloning of quantum coherence. Independently of the validity of this conjecture, the importance of the recursive protocol can be best appreciated in all those cases where the optimal tradeoff curve is not explicitly known. To get even better lower bounds to the optimal curve, we finally introduce the operation of \emph{coherent coarse-graining}, which consists in joining different outcomes into a single quantum operation. Coherent coarse-graining allows one to keep the same probability of success of the outcomes that are joined, while increasing the fidelity with the target. Combined with our recursive protocol, this operation provides a canonical way to generate lower bounds to the optimal tradeoff between fidelity and success probability, whose exact calculation is generally hard to perform without resorting to numerical optimization.
To demonstrate the broad applicability of our methods, we illustrate the recursive protocol and its coherent coarse-graining in a number of concrete examples, including quantum phase estimation, energy-preserving amplification of coherent states, and the optimal design of correction operations for ancilla-driven computation. In addition to the applications presented explicitly in the paper, our results can be directly applied to most of the canonical problems of quantum information processing, such as optimal state discrimination, gate programming, entanglement conversion, universal NOT and universal transpose---whose implementation is significantly affected once one imposes the requirement that no energy should be drawn from the environment.
Finally, we go beyond the regime of pure state transitions and extend our results to transitions where mixed states are given as inputs. For such transitions, we prove upper bounds for the performances of both deterministic and probabilistic operations, providing conditions for the saturation of the bounds. This extension provides optimal strategies for the implementation of quantum tasks such as the purification of mixed states \cite{cirac-ekert-macchiavello,andersen} and the redistribution of quantum signals to multiple users \cite{dariano-macchiavello-perinotti-prl}.
The paper is organized as follows. Section \ref{sec:general} introduces the basic framework. In Section \ref{sec:pure} we characterize the optimal energy-preserving process.
Using this result, we construct the recursive protocol and study the operation of coherent coarse-graining in Section \ref{sec:network} and apply it to several tasks in quantum information processing (Section \ref{sec:examples}), including phase estimation (Subsections \ref{subsec:estimation} and \ref{subsec:maxcoh}), state cloning (Subsection \ref{subsec:cloning}), coherent light amplification (Subsection \ref{subsec:coherent}) and ancilla-driven computation (Subsection \ref{sec:unlearn}).The extension of our results to mixed states is discussed in Section \ref{sec:mixed}.
Finally, the conclusions are drawn in Section \ref{sec:conclusion}.
\section{The energy-preserving paradigm}\label{sec:general}
In this section we introduce the framework that will be adopted in the rest of the paper. We first present the class of energy-preserving channels, corresponding to the deterministic evolutions that can be implemented without drawing energy from the environment.
We then move to probabilistic operations, characterizing the stochastic evolutions that can be implemented at zero energy cost.
\subsection{Energy-preserving channels}\label{subsec:enpreschan}
Consider a quantum system interacting with the surrounding environment from time $t_1$ to time $t_2$ through an interaction Hamiltonian $H_{\rm int} (t)$, which we assume to be zero before $t_1$ and after $t_2$. As a result of the interaction, the system and the environment evolve jointly via the unitary operator
\begin{align}
U = \map T \exp \left\{-\frac{i}{\hbar} \, \int_{t_1}^{t_2} \, {\rm d} t \, \left[ H_{\rm sys} + H_{\rm env} + H_{\rm int} (t) \right]\right\} \, ,
\end{align}
where $\map T \exp$ denotes the time-ordered exponential and $H_{\rm sys}$ and $H_{\rm env}$ are the Hamiltonians of the system and of the environment, both assumed to be time-independent.
Regarding energy as a resource, we require the evolution to satisfy the condition
\begin{align}\label{totalenergy-pres}
U^\dag \, (H_{\rm sys} + H_{\rm env} ) \, U = H_{\rm sys} + H_{\rm env} \,
\end{align}
meaning that the total energy after $t_2$ is equal to the total energy before $t_1$. A sufficient condition for the validity of Eq. (\ref{totalenergy-pres}) is the commutation relation
\begin{align}
[ H_{\rm int}( t ), H_{\rm sys} + H_{\rm env}] = 0 \qquad \forall t\in [t_1,t_2] \, ,
\end{align}
which guarantees that the sum of the system and environment energies is a constant of motion during the entire evolution. Note that the interaction $H_{\rm int} (t)$ can be made arbitrarily weak by increasing the interaction time.
Among the interactions that conserve the total energy, we are interested in those that leave the energy of the environment untouched. Such interactions satisfy the additional condition
\begin{align}\label{environment-pres}
U^\dag H_{\rm env} U = H_{\rm env} \,
\end{align}
Clearly, the combination of Eqs. (\ref{totalenergy-pres}) and (\ref{environment-pres}) implies that the interaction has to preserve the energy of the system, namely
\begin{align}\label{energy-pres}
U^\dag H_{\rm sys} U = H_{\rm sys} \,
\end{align}
Assuming that the environment is initially in the state $\rho_{\rm env}$, the effective evolution of the system is described by the quantum channel (completely positive trace-preserving map) $\map M$ defined by
\begin{align}\label{stine}
\map M(\rho) = \operatorname{Tr}_{\rm env} \left[ U \, ( \rho \otimes \rho_{\rm env}) \, U^\dag\right]
\end{align}
where $\operatorname{Tr}_{\rm env}$ denotes the partial trace over the Hilbert space of the environment.
Clearly, the channel $\map M$ preserves the expectation value of the system's energy.
Even more strongly, condition (\ref{energy-pres})
implies that the channel $\map M$ preserves also the variance and all the momenta of the Hamiltonian, namely
\begin{align}\label{MdagH}
\map M^\dag \left(H^n_{\rm sys}\right) = H^n_{\rm sys} \qquad \forall n \in \mathbb{N} \, ,
\end{align}
where $\map M^\dag$ is the completely positive identity-preserving map describing the evolution in the Heisenberg picture, defined by
\begin{align}
\map M^\dag (A) = \operatorname{Tr}_{\rm env} [ \, (I_{\rm sys} \otimes \rho_{\rm env}) \, U^\dag \, (A\otimes I_{\rm env} ) \, U \, ] \, ,
\end{align}
for arbitrary operators $A$. When Eq. (\ref{MdagH}) is satisfied, we say that $\map M$ is an \emph{energy-preserving channel}.
The energy-preserving condition is equivalent to the requirement that the evolution $\map M$ do not affect the probability distribution of the energy,~i.~e. that one has the equality of probabilities
\begin{align}\label{energy}
\operatorname{Tr} [ P_E \, \map M(\rho) ] = \operatorname{Tr}[ P_E \, \rho] \, ,
\end{align}
where $\rho$ is an arbitrary state, $E$ is an arbitrary eigenvalue of $H_{\rm sys}$, and $P_E$ is the projector on the eigenspace corresponding to $E$.
It is easy to see that every energy-preserving channel $\map M$ is \emph{covariant} with respect to the free evolution of the system, that is, $\map M$ satisfies the condition
\begin{align}
\map M( U_t \cdot U_t^\dag ) = U_t \, \map M(\cdot) \, U_t^\dag \qquad \forall t\in \mathbb{R} \, ,
\end{align}
with $U_t= \exp[-itH_{\rm sys}/\hbar]$. However, the converse is not true in general: for example, a channel that discards the input state and re-prepares an eigenstate of the energy is covariant but not energy-preserving.
Physically, the difference between energy-preserving and covariant channels is that the latter preserve the total energy of the system and the environment, while the former preserve the individual energies of the system and of the environment. Further discussion on the relation between energy-preserving channels and other classes of channels, such as incoherent channels and Hadamard channels is presented in Appendix \ref{app:relations}.
Energy-preserving channels provide an economical way to process information in situations where energy becomes a scarce resource. For example, one can imagine a scenario where a microscopic machine has to perform a desired task on the system (like copying data, correcting for an error, or erasing information \cite{vaccaro2011information}) without changing the energy of the surrounding environment.
While scenarios of this kind are not a practical reality yet, prototype demonstrations of energy-preserving channels can be conceived for quantum systems with a high degree of control, such as e.g. ion traps \cite{ion1,ion2},
optical cavities \cite{cavity1,cavity2}, optical lattices \cite{lattice1,lattice2}, and optomechanical systems \cite{optomechanics1,optomechanics2}. Moreover, the condition that an observable is a constant of motion plays an important role even when the observable does not coincide with the energy. For example, implementations of logical gates in quantum dots often benefit from the existence of ``sweet spots", where charge noise is suppressed. However, the set of gates that can be performed at the sweet spot is limited: for example, in the three electron exchange-only spin qubit \cite{sweet1}, only the rotations around the $z$-axis can be performed at the sweet spot, while all other rotations incur into undesired noise \cite{sweet2}. Regarding the component of the spin along the $z$-axis as the ``energy'', we have that energy-preserving channels are exactly the operations that can be performed with suppressed noise.
Similar physics arises in superconducting flux qubits, where the sweet spot is with respect to magnetic noise \cite{sweet3}.
\subsection{Structure of the energy-preserving channels}\label{subsect:struct}
We have seen that every energy-preserving interaction induces an energy-preserving quantum channel, i.e. a channel satisfying Eq. (\ref{MdagH}).
The converse is also true:
given an energy-preserving channel $\map M$, one can always engineer an interaction between the system and the environment that conserves the total energy and does not draw energy from the environment. To establish this fact, note that the map $\map M^\dag$ satisfies the condition
\begin{align}\label{algpres}
\map M^\dag (A) = A \qquad \forall A\in\map A \, ,
\end{align}
where $\map A $ is the commutative algebra generated by the powers of the Hamiltonian. The algebra $\map A$ contains the identity and is closed under adjoint. Technically, algebras of this kind are known as unital $*$-algebras \cite{Bratteli_Robinson}.
For any such algebra, the maps that satisfy Eq. (\ref{algpres}) are characterized by a simple lemma:
\begin{lemma}[Lindblad \cite{Lindblad}]\label{lem:lindblad}
Let $\map M^\dag$ be an identity-preserving completely positive map, written in the Kraus form $\map M^\dag (A) = \sum_{k=1}^K M_k ^\dag AM_k $ and let $\map A$ be a unital $*$-algebra $\map A$.
The map $\map M^\dag$ preserves the elements of $\map A$ if and only if each Kraus operator $M_{k}$ belongs to the commutant of $\map A$,~i.~e.~to the set of operators
\begin{align*}
\map A' : = \{ B \in \mathcal B (\spc H) \, , \, [ A, B ] = 0 \quad \forall A \in \map A \, \} \, ,
\end{align*}
$\mathcal B (\spc H)$ denoting the set of bounded operators on $\spc H$.
\end{lemma}
In the case of the Abelian algebra generated by the powers of $H_{\rm sys}$, the commutation condition reduces to
\begin{align} [ M_k, H_{\rm sys}] = 0 \, ,
\end{align} meaning that each Kraus operator $M_k$ must be of the block diagonal form
\begin{align}\label{block}
M_k = \bigoplus_{E} \, P_E M_k P_E \, ,
\end{align}
with the sum running over spectrum of $H_{\rm sys}$. All throughout the paper we will assume the energy spectrum to be discrete.
As a consequence of the block diagonal form (\ref{block}),
the isometry $V : \spc H \to \spc H\otimes \spc H_{\rm env}$ defined by
\begin{align}\label{iso}
V = \sum_{k=1}^K M_k \otimes | \phi_k\> \, ,
\end{align}
where $ \{\, |\phi_k\>\,\}_{k=1}^K$ are orthonormal states,
satisfies the relation
\begin{align}\label{Vpres}
V H_{\rm sys} = (H_{\rm sys}\otimes I_{\rm env}) \, V \, .
\end{align}
In turn, Eq. (\ref{Vpres}) implies that the isometry $V$ can be realized via an interaction that preserves both the energy of the system and the energy of the environment:
\begin{theo}\label{theo:enpres'}
Let $\map M$ be a quantum channel transforming states on $\spc H$. Then, the following are equivalent:
\begin{enumerate}
\item $\map M$ is energy-preserving
\item $\map M$ can be realized via an interaction with the environment $\map M(\rho) = \operatorname{Tr}_{\rm env} [ U \, ( \rho \otimes |\phi_0\>\<\phi_0| ) \, U^\dag]$ such that $|\phi_0\>$ belongs to ground eigenspace of $H_{\rm env}$ and $U$ satisfies the conditions $U^\dag \, H_{\rm sys} \, U = H_{\rm sys} $ and $U^\dag \, H_{\rm env} \, U = H_{\rm env} $.
\end{enumerate}
\end{theo}
The proof is presented in Appendix \ref{app:proofV}.
Note the appearance of the Hamiltonian of the environment in the statement of the theorem. This is natural, because in general we are dealing with the evolution of an open system. In this situation, the Hamiltonian of the environment plays a crucial role, for it determines the minimum amount of energy one has to invest into the realization of the desired channel. Ideally, theorem \ref{theo:enpres'} guarantees that such amount can be reduced to zero in the case of energy-preserving channels. Specifically, the desired evolution can be engineered by initializing the environment in an eigenstate of its Hamiltonian and by turning on a coupling that preserves the individual energies of system and environment, keeping the latter inside the ground space for the whole time evolution. As a result, the evolution is implemented at zero energy cost, at the price of an entropy increase in the environment, which is generally left in a mixture of states with the same energy. In other words, the environment is only used passively as computational workspace in which information can be stored.
\subsection{Energy-preserving measurements}
While in the case of deterministic evolutions the notion of energy preservation is pretty straightforward, the situation is different for probabilistic transformations induced by quantum measurements.
In this section we introduce a notion of \emph{probabilistic energy-preserving transformations}, which characterizes those operations that can be implemented (in principle) without paying an energy cost.
Adopting Ozawa's model of the measurement process \cite{ozawa}, we view probabilistic evolutions as the result of a unitary interaction between the system and the environment, followed by the projective measurement of a ``meter observable" $O$ on the environment.
In this model, the preservation of the energy imposes constraints on the interaction as well as constraints on the measurement.
Like in the previous sections, we demand that the system-environment interaction preserve the total energy of the system and the environment [eq. (\ref{totalenergy-pres})] and do not change the energy of the environment [Eq. (\ref{environment-pres})].
As we argued in the previous paragraph, these requirements characterize the evolutions that can be implemented at zero energy cost.
Regarding the measurement, we demand that it should not disturb the energy the environment, or, equivalently, that the meter observable and the energy are compatible observables, namely
\begin{align}\label{compatible}
[O,H_{\rm env}] = 0 \, .
\end{align}
If this condition were not satisfied, the measurement of $O$ would collapse an eigenstate of the energy into a random eigenstate of $O$, thus altering the expectation value of the energy.
Observables that commute with the Hamiltonian of the environment are the prototype of measurements that can be performed without paying an energy cost. The conditions (\ref{totalenergy-pres}), (\ref{environment-pres}), and (\ref{compatible}) are the standard requirements put forward in the Wigner-Araki-Yanase theorem \cite{W,AY, ozawa2} and in all the works that followed up on it \cite{ozawaCNOT,computing3,spekkens-marvian-WAY,ahmadi}. In this context the commutation relation (\ref{compatible}) is known as \emph{Yanase's condition} \cite{ozawa2}. Very recently, Navascu\'es and Popescu adopted the same framework discussed here as the starting point to \emph{define} the energy requirements for the implementation of quantum measurements \cite{popescu}.
Let us analyze the probabilistic evolutions resulting from the requirement of zero energy cost. According to quantum measurement theory \cite{ozawa,lathi_book,holevo_stat_book,heinosaar_ziman}, the measurement of $O$ induces a stochastic evolution of the state of the system, described by a \emph{quantum instrument}, namely a collection of quantum operations
(completely positive trace non-increasing maps) $ \{ \map M_x\}_{x\in\set X}$ subject to the normalization condition
\begin{align}\label{normalization}
\sum_{x\in\set X} \operatorname{Tr} [ \map M_x (\rho)] = \operatorname{Tr} [\rho]
\end{align}
for every quantum state $\rho$. For a system prepared in the state $\rho$, the measurement generates the outcome $x\in \set X$ with probability
\begin{align}\label{pxrho}
p (x|\rho) = \operatorname{Tr} [ \map M_x (\rho)] \, ,
\end{align}
and, conditionally on outcome $x$, returns the system in the state
\begin{align}\label{rhoprime}
\rho_x' = \frac{\map M_x(\rho)}{\operatorname{Tr}[\map M_x (\rho)]} \, .
\end{align}
In the model considered here, the set of outcomes is the spectrum of the meter observable $O$ and the quantum operation $\map M_x$ is defined by
\begin{align}\label{qostine}
\map M_x (\rho) = \operatorname{Tr}_{\rm env} [ (I_{\rm sys} \otimes Q_x) \, U \,(\rho \otimes \rho_{\rm env}) \, U^\dag] \, ,
\end{align}
$Q_x$ being the projector on the eigenspace of $O$ with eigenvalue $x$. Note that, by summing over all possible outcomes, one obtains
\begin{align*}
\sum_{x\in\set X} \map M_x (\rho) & = \operatorname{Tr}_{\rm env} [ U \,(\rho \otimes \rho_{\rm env}) \, U^\dag] \\
& \equiv \map M (\rho)
\end{align*}
where $\map M$ is an energy-preserving channel. Energy preservation for $\map M$ simply follows from just the condition $ U^\dag \, H_{\rm sys} \, U = H_{\rm sys}$. Note that this conclusion is independent of the validity of Yanase's condition.
The physical model discussed so far motivates the following definition:
\begin{defi}
We say that a quantum instrument $\{\map M_x\}_{x\in\set X}$ is \emph{energy-preserving} if the average channel $\map M : = \sum_{x\in\set X} \, \map M_x$ is energy-preserving.
\end{defi}
Energy-preserving instruments are exactly the instruments that can be implemented (in principle) at zero energy cost. Precisely, every such instrument can be realized via an energy-preserving interaction, followed by the measurement of a meter observable that commutes with the energy:
\begin{theo}\label{theo:enpresinst}
Let $\{\map M_x\}_{x\in\set X}$ be a quantum instrument transforming states on $\spc H$. Then, $\{\map M_x\}_{x\in\set X}$ is energy-preserving if and only if one has
\[\map M_x(\rho) = \operatorname{Tr}_{\rm env} [ (I_{\rm sys} \otimes Q_x) U ( \rho \otimes |\phi_0\>\<\phi_0| ) U^\dag] \qquad \forall x\in\set X\]
where $|\phi_0\>$ is a ground state of the environment's Hamiltonian, $U$ satisfies the conditions $U^\dag \, H_{\rm sys} \, U = H_{\rm sys} $ and $U^\dag \, H_{\rm env} \, U = H_{\rm env} $, and $\{Q_x\}$ is a projective measurement satisfying Yanase's condition $[Q_x, H_{\rm env}]=0$ $\forall x\in\set X$.
\end{theo}
The proof can be found in Appendix \ref{app:proofV}, while a simple illustration of the result is shown in the following example:
\begin{eg}\label{eg:qubit}
Consider a system with two energy levels $E_0 = 0$ and $E_1 = \Delta E$, corresponding to the pure states $|0\>$ and $|1\>$, respectively. Clearly, the von Neumann instrument for the energy measurement---described by the quantum operations $\map M_x (\cdot) = |x\>\< x| \cdot |x\>\<x|$, $x= 0,1$---is energy-preserving. To implement this instrument, one can use as environment two identical copies of the system, choose the initial state $|\phi_0\> = |0\> |1\> $, and engineer a joint evolution $U$ satisfying
\begin{align*}
U |0\> |0\> |1\> & = |0\> |0\> |1\> \\
U |1\> |0\> |1\> & = |1\> |1\> |0\> \\
U |0\> |1\> |0\> & = |0\> |1\> |0\> \\
U |1\> |1\> |0\> & = |1\> |0\> |1\> \\
U |0\> |0\> |0\> & = |0\> |0\> |0\> \\
U |1\> |0\> |0\> & = |1\> |0\> |0\> \\
U |0\> |1\> |1\> & = |0\> |1\> |1\> \\
U |1\> |1\> |1\> & = |1\> |1\> |1\> \, .
\end{align*}
By measuring the meter observable $M = |1\>\< 1| \otimes |0\>\< 0|$ on the environment, one then obtains the instrument $\{\map M_0 , \map M_1 \}$ as effective evolution of the system.
\end{eg}
It is worth stressing that, despite the fact that the energy is preserved on average, its expectation value can fluctuate due to postselection. For instance, in Example \ref{eg:qubit} one can decide to postselect the output state $|1\>$.
With probability 1/2, the postselection will transform the state $|+\> = (|0\> + |1\>)/\sqrt 2$ into the state $|1\>$, whose energy is twice the expected energy of $|+\>$. Still, the transformation will take place at no energy cost, because both the interaction and the measurement of the meter observable preserve the energy. A further discussion on this point is provided in Appendix \ref{appendix:weak}.
Motivated by the above discussion, we put forward the following
\begin{defi}
A probabilistic transformation $\map M_0$ is \emph{energy-preserving} iff there exists an energy-preserving instrument $\{\map M_x\}_{x\in\set X}$ and an outcome $x_0\in\set X$ such that $\map M_0 = \map M_{x_0}$.
\end{defi}
Probabilistic energy-preserving transformations can be demonstrated in engineered quantum systems with high degree of control.
For example, a proposal for an experimental amplification of weak coherent states via probabilistic energy-preserving transformations was recently reported by Partanen \emph{et al} in Ref. \cite{preservingamp}, where high fidelity amplification was achieved using only passive optical elements.
\iffalse
In general, each CP map $\map M_x$ must be trace non-increasing, due to Eq. (\ref{normalization}). Trace non-increasing CP maps are known as quantum operations \cite{Kraus} and can be written in the Kraus representation
\begin{align}\label{kraus}
\map M_x (\rho) = \sum_{y\in\set Y_x} M_{xy} \, \rho \, M_{xy}^\dag
\end{align}
for suitable operators $M_{xy} : \spc H_{\rm in} \to \spc H_{\rm out}$, known as Kraus operators, and for a suitable index set $\set Y_x$.
Substituting in Eq. (\ref{pxrho}), the probability of the outcome $x$ can be expressed as
\begin{align}\label{Pi}
p (x|\rho) =\operatorname{Tr}\left[P_x \rho \right] \qquad P_x : = \sum_{y\in\set Y_x} M_{xy}^\dag M_{xy} \, .
\end{align}
Note that the normalization condition (\ref{normalization}) is equivalent to the requirement that the operators $\{ P_x\}_{x\in\set X}$ satisfy the condition
\begin{align}\label{POVM}
\sum_{x\in\set X} P_x = I_{\rm in} \, ,
\end{align}
where $I_{\rm in}$ is the identity on the input Hilbert space. In other words, the operators $\{ P_x\}_{x\in\set X}$ form a positive operator-valued measure (POVM).
\fi
\subsection{Energy-preserving filters}
Quantum measurements can be used to postselect a desired subset of evolutions. In this case, we refer to them as \emph{quantum filters}. Mathematically, a filter is described by an instrument $\{\map M_x\}_{x\in\set X}$ along with a partition of outcome set $\set X$ into two disjoint subsets, $X_{\rm succ}$ and $X_{\rm fail}$, which correspond to successful and unsuccessful instances, respectively. An \emph{energy-preserving filter} is simply a filter corresponding to an energy-preserving instrument.
We will now see that energy-preserving filters play a central role in the optimization of quantum operations. Their importance comes from a special relation with the class of instruments that are \emph{covariant} under time evolution \cite{holevo_stat_book},~i.~e.~instruments $\{\map M_x\}_{x\in\set X}$ such that every quantum operation $\map M_x$ satisfies the condition
\begin{align}\label{covariant}
\map M_x \left ( U_t \,\rho\, U_t^\dag \right) = V_t \, \map M_x ( \rho) \, V_t^\dag \qquad \forall t \in \mathbb{R}
\end{align}
where $\rho$ is an arbitrary input state and $\{ U_t~|~ t\in \mathbb{R}\}$ and $\{V_t ~|~ t\in\mathbb{R}\}$ are two unitary representations of the translation group, acting on the input and output of the instrument, respectively.
First of all, note that every energy-preserving instrument is covariant: indeed, the block form of the Kraus operators (\ref{block}) implies that every quantum operation $\map M_x$ satisfies the covariance condition with
\begin{align}\label{uvt}
U_t = V_t = \exp[-i t H_{\rm sys}/\hbar] \qquad \forall t\in \mathbb{R} \, .
\end{align}
An even more interesting relation between energy-preserving and covariant instruments is shown by the following proposition, proved in Appendix \ref{app:decompcov}:
\begin{prop}\label{prop:decompcov}
Every covariant instrument $\{\map M_x\}_{x\in\set X}$ transforming states on $\spc H_{\rm in}$ into states on $\spc H_{\rm out}$ can be decomposed as
$$ \map M_x = \map C_x \map P_x \qquad\forall x\in\set X $$ where
$\map C_x$ is a covariant channel transforming the input into the output system and $\{\map P_x\}_{x\in\set X}$ is a pure energy-preserving instrument, defined by
\begin{align}
\map P_x (\cdot) = \sqrt{ P_x} \cdot \sqrt {P_x} \, , \qquad P_x = \map M^\dag_x (I) \, .
\end{align}
\end{prop}
This result provides additional motivation to the study of energy-preserving filters. Indeed, there is a large class of tasks where the optimal quantum strategy is covariant---this is the case, e.g. of phase-covariant cloning \cite{qubit,fiurasekclon,GiulioNat}, phase-insensitive amplification \cite{insensitive}, and phase-estimation \cite{holevo_stat_book,helstrom}. Proposition \ref{prop:decompcov} establishes that energy-preserving filters are the canonical probabilistic element in all these tasks. The search for the optimal quantum operation is then split into two sub-problems: \emph{i)} the search for the optimal energy-preserving filter and \emph{ii)} the search for the optimal deterministic operation.
Now, the optimization of deterministic operations has been studied extensively in the literature, and the solution of problem \emph{ii)} is known in a number of relevant cases. In all these cases, the search for the optimal probabilistic operation is reduced to the search of the optimal energy-preserving filter. This is exactly the key question of this paper. A general method for the solution of the problem will be provided in the next section.
\section{Optimal energy-preserving operations} \label{sec:pure}
We are now ready to start the search for the optimal operations that transform a coherent superposition of energy eigenstates into another. In this section we first formalize the problem and then address the optimality question, providing the general form of the optimal energy-preserving operations.
\subsection{How well can we implement a desired state transition without investing energy?}
Regarding energy as a resource, a natural question is how well a desired task can be achieved without consuming the energy of the surrounding environment.
Consider the most basic task: transforming a pure input state $|\varphi\>$ into a target output state $ |\psi\>$.
For example, the input could be a weak coherent state with known amplitude but unknown phase, and the target could be another coherent state with the same phase but with larger amplitude. The problem of amplifying laser pulses using energy-preserving operations was recently studied in Ref. \cite{preservingamp}, where the authors showed that a nearly perfect amplification can be achieved probabilistically by exploiting the quantum fluctuations of the field, without drawing any energy from the outside. Another interesting example is quantum cloning: suppose that a spin 1/2 particle, immersed in a magnetic field, is prepared in a superposition of spin up and spin down. As a result, the particle will precede around the direction of the magnetic field and its state will evolve in time. How well can we copy the time information without tapping external energy sources?
Note that both in the amplification and in the cloning example, the input and the target states are drawn from a \emph{set} of states: more precisely, the problem is to transform the input state $ |\varphi_t\> = e^{-it H_{\rm sys}/\hbar} |\varphi\>$ into the targed state $ |\psi_t\> = e^{-it H_{\rm sys}/\hbar} |\psi\>$ for an arbitrary (and possibly unknown) value of the parameter $t$. However, since we require our operations to be energy-preserving, we do not need to optimize them for every value of $t$: indeed, every energy-preserving transformation that approximates the transition $ |\varphi\> \to |\psi\>$ will do equally well in approximating the transition $ |\varphi_t\> \to |\psi_t\>$, due to the covariance condition of Eqs. (\ref{covariant}) and (\ref{uvt}). This point is made clear if we measure the performance of our operations in terms of the fidelity between the target state and the actual output. For a probabilistic transformation $\map M_x$, the fidelity between target state and the actual output state is
\begin{align*}
F_{x,t} : = \< \psi_t | \rho_x' (t) |\psi_t\> \, , \qquad \rho_x'(t) = \frac{\map M_x ( |\varphi_t\>\<\varphi_t |}{\operatorname{Tr}[ \map M_x (|\varphi_t\>\<\varphi_t| ]} \, .
\end{align*}
Using the covariance condition it is immediate to see that $F_{x,t}$ is independent of $t$. Physically, this means that energy-preserving transformations perform equally well on all possible inputs that are connected by time evolution.
More generally, it is interesting to consider all the transformations corresponding to the successful instances of a quantum filter.
Averaging the fidelity over the successful instances, we obtain the value
\begin{align}
F = \sum_{x \in \set{X}_{\rm succ}} \, p( x | {\rm succ}) \, \< \psi_0 \,| \rho_x' (0)\, |\psi_0 \> \, ,
\end{align}
where $ p( x|{\rm succ}) $ is the conditional probability of obtaining $x$ given that a successful outcome has occurred.
Making the filter explicit, the average fidelity can be rewritten as
\begin{align}\label{F}
F = \frac{ \< \psi| \, \map M_{\rm succ} ( |\varphi\>\< \varphi|)\, |\psi\> }{p_{\rm succ}} \, ,
\end{align}
where $\map M_{\rm succ} $ is the quantum operation defined by
\begin{align}\label{msucc}
\map M_{\rm succ}: = \sum_{x\in\set X_{\rm succ}} \, \map M_x
\end{align} and $p_{\rm succ}$
is the probability of success, given by
\begin{align}
\nonumber p_{\rm succ} &= \operatorname{Tr} [ \map M_{\rm succ} ( |\varphi\>\< \varphi|)] \\
\nonumber &= \operatorname{Tr} [ \map M^\dag_{\rm succ} (I) \, |\varphi\>\< \varphi| ]\\
& = \<\varphi| \, P_{\rm succ} \, |\varphi\>\, ,
\label{psucc} \end{align}
having defined
$ P_{\rm succ} := \map M^\dag_{\rm succ} (I)$.
In a realistic situation one will be interested not only in maximizing the fidelity, but also in having a sufficiently high probability of success. Requiring the success probability to be larger than a given threshold, the problem becomes to find the energy-preserving quantum operation $\map M_{\rm succ}$ that maximizes the fidelity.
\subsection{Deterministic transitions: optimality of eigenstate alignment}
Let us consider first the case $p_{\rm succ} =1$. In this case, the optimization problem has a closed-form solution, corresponding to an operation that we call \emph{eigenstate alignment}. Given two superpositions of energy eigenstates, eigenstate alignment is a energy-preserving unitary operation that transforms the eigenstates in one superposition into the eigenstates in the other. Precisely, let us decompose the pure states $|\varphi\>$ and $|\psi\>$ as
\begin{align}\label{statedecomp}
|\varphi\>=\sum_{E}\sqrt{p_E}|\varphi_E\> \quad {\rm and} \quad |\psi\>=\sum_{E}\sqrt{q_E}|\psi_E\>
\end{align}
with
\begin{align*}
\begin{array}{rlrl} p_E & = \| P_E |\varphi \> \|^2 \, , & q_E & = \| P_E |\psi \>\|^2 \, ,\\
&&&\\
|\varphi_E\> &:= \frac{P_E|\varphi\>}{\|P_E|\varphi\>\|} \, , \qquad & |\psi_E\> &:= \frac{P_E|\psi\>}{\|P_E|\psi\>\|} \, ,
\end{array}
\end{align*}
$P_E$ being the projector on the eigenspace of $H_{\rm sys}$ with energy $E$. Then, we say that a unitary operator $U$ is an \emph{eigenstate alignment} of $|\varphi\>$ with $|\psi\>$ if $U$ is energy-preserving and
\begin{align}\label{eigenalign}
U |\varphi_E\> = |\psi_E\> \qquad \forall E : \, p_E \not = 0 \, , q_E \not = 0 \, .
\end{align}
It is immediate to see that a unitary eigenstate alignment exists for every pair of pure states. In general, eigenstate alignments are not unique, unless the spectrum of $H_{\rm sys}$ is non-degenerate and every energy $E$ in the spectrum satisfies the conditions $p_E \not = 0$ and $q_E \not = 0$. The importance of eigenstate alignment comes from the following
\begin{theo}\label{theo:EAoptimal}
For $p_{\rm succ} =1$, eigenstate alignment achieves the maximum fidelity for the transition $|\varphi\> \to |\psi\>$. The maximum fidelity is given by
\begin{align}F_{\rm det}=\left(\sum_E\sqrt{p_Eq_E}\right)^2 \, .
\end{align}
\end{theo}
The proof is provided in Appendix \ref{app:EAoptimal}. Theorem \ref{theo:EAoptimal} shows that the optimal energy-preserving channel can be chosen to be unitary without loss of generality. In this case, no interaction with the environment is needed.
We stress that the optimality of unitary transformations is a specific feature of the energy-preserving framework. Unitary transformations may not be optimal in the broader class of phase-covariant channels---for example, they are sometimes suboptimal for cloning qubit states on the equator of the Bloch sphere
\cite{qubit}.
\subsection{Probabilistic transitions: optimality of pure quantum operations}
Let us move to the general case $p_{\rm succ} \le1$.
\iffalse
In the search of the optimal filter, it is immediately clear that one can restrict the attention to filters with two outcomes.
This is because \emph{i)} both the fidelity and the success probability depend only of the map $\map M_{succ}$ defined in Eq. (\ref{msucc}) and \emph{ii)} if the original filter $\{\map M_x\}$ preserves the energy, so does the binary filter $\{\map M_{\rm succ}, \map M_{\rm fail}\}$ with
\begin{align} \map M_{\rm succ} = \sum_{x\in\set X_{\rm succ}} \, \map M_x \, , \quad \map M_{\rm fail} = \sum_{x\in \set X_{\rm fail}} \map M_x \, .
\end{align}
\fi
We now show that, without loss of generality, the optimal quantum operation $\map M_{\rm succ}$ can be chosen to be \emph{pure},~i.~e.~of the form
$\map M_{\rm succ} ( \cdot) = M_{\rm succ} \cdot M_{\rm succ}^\dag$
for some suitable operator $M_{\rm succ}$. To prove this result, we provide a construction that transforms any given quantum operation $\map M_{\rm succ}$ into a pure quantum operation $\map M_{\rm succ}'$ with the same probability of success and possibly a higher fidelity.
The construction is based on a new ingredient, which we name \emph{L\"uders reduction}.
\subsubsection{L\"uders reduction}
L\"uders reduction transforms a given quantum operation into a pure quantum operation with the same probability of success.
Specifically, the L\"uders reduction of a quantum operation $\map M$ is the pure quantum operation $\map P$ defined by
\begin{align}\label{luders}
\map P (\cdot) = \sqrt P \cdot \sqrt P \, , \quad P = \map M^\dag (I) \, .
\end{align}
When $P$ is a projector, the quantum operation $\map P$ coincides with the state transformation defined by L\"uders in his treatment of the measurement process \cite{luders}. When $P$ is not a projector, $\map P$ is often called the ``generalized L\"uders transformation" associated to $P$ \cite{busch}.
By construction, a quantum operation and its L\"uders reduction have the same probability:
Indeed, one has
\begin{align}
\nonumber \operatorname{Tr} [ \map P (\rho)] & = \operatorname{Tr} [ \sqrt {P} \rho \sqrt{P}] \\
\nonumber & = \operatorname{Tr} [ P \rho] \\
\nonumber & = \operatorname{Tr} \left[ \map M^\dag (I) \rho \right] \\
\label{ludersprob} & = \operatorname{Tr} [ \map M (\rho) ]
\end{align}
for every quantum state $\rho$.
Among the quantum operations that happens with the same probability as $\map M$, the L\"uders reduction can be characterized as the ``least noisy", meaning that every other quantum operations can be reproduced by applying a noisy channel to the output of $\map P$:
\begin{prop}\label{prop:ludersdecomp}
Every quantum operation $\map M$ can be decomposed as $ \map M = \map C \circ \map P $ where $\map P$ is the L\"uders reduction of $\map M$ and $\map C$ is a suitable quantum channel.
Moreover, if $\map M$ is energy-preserving, then also $\map P$ and $\map C$ are energy-preserving.
\end{prop}
The proof is provided in Appendix \ref{app:proofProp}. Using this result, the search of the optimal quantum operation is split into two different problems: the search of an optimal \emph{pure} operation $\map P$ and the search of the optimal \emph{deterministic} operation $\map C$. Note that this conclusion applies not apply only to the problem of transforming pure states, but also to the optimization of transitions involving mixed states.
\subsubsection{Increasing the fidelity without changing the success probability}
Combining L\"uders reduction and eigenstate alignment we can turn every quantum operation $\map M_{\rm succ}$ into a pure quantum operation $\map M_{\rm succ}'$ with the same success probability and a possibly higher fidelity.
The idea is simple: by proposition \ref{prop:ludersdecomp}, every energy-preserving quantum operation $\map M_{\rm succ}$
can be decomposed as
\begin{align*}
\map M_{\rm succ} = \map C \circ \map P_{\rm succ} \, ,
\end{align*}
where $\map C$ is an energy-preserving channel and $\map P_{\rm succ}$ is the L\"uders reduction given by
\begin{align*}
\map P_{\rm succ} (\cdot) = \sqrt{ P_{\rm succ}} \cdot \sqrt{ P_{\rm succ}} \, , \qquad P_{\rm succ} = \map M_{\rm succ}^\dag (I) \, .
\end{align*}
When the L\"uders reduction takes place, the input state $|\varphi\>$ is transformed into the pure state
\begin{align}\label{phiprime}
|\varphi'\> = \frac{ \sqrt{ P_{\rm succ} } |\varphi\>}{ \| \sqrt {P_{\rm succ}} |\varphi\> \|}
\, .
\end{align}
Now, the probability of success depends only on the operator $P_{\rm succ}$.
Fixing $P_{\rm succ}$, we know that the optimal energy-preserving channel for the transition $|\varphi'\> \to |\psi\>$ is given by eigenstate alignment (theorem \ref{theo:EAoptimal}). Hence, the fidelity for the quantum operation $\map M_{\rm succ}$ cannot be larger than the fidelity of the quantum operation\begin{align}\label{Msuccprime}
\map M_{\rm succ}' =\, \map U \circ \map P_{\rm succ} ,
\end{align}
where $\map U $ is the unitary channel corresponding to the eigenstate alignment of $|\varphi'\>$ with $|\psi\>$. Note that $\map M_{\rm succ}'$ is energy-preserving, because it is the composition of two energy-preserving operations. Summarizing, we have proven the following
\begin{prop}\label{prop:better}
For every energy-preserving operation $\map M_{\rm succ}$, the energy-preserving operation $\map M_{\rm succ}'$ defined in Eq. (\ref{Msuccprime}) has the same success probability and at least the same fidelity.
\end{prop}
Explicitly, the success probability and the fidelity of $\map M_{\rm succ}'$ are given by
\begin{align}
\label{psuccaa} p_{\rm succ}
&= \sum_E p_E \, \< \varphi_E | P_{\rm succ} | \varphi_E\>
\end{align}
and
\begin{align}
\label{Faa}
F = \left( \sum_E \sqrt{ p_E' q_E} \right)^2\, , \qquad p_E' = \frac{p_E \, \<\varphi_E | \, P_{\rm succ}\, |\varphi_E\>}{p_{\rm succ}} \, ,
\end{align}
the expression of the fidelity following from the application of Theorem \ref{theo:EAoptimal} to the transition $|\varphi'\>\to |\psi\>$.
Now, since turning a quantum operation into a pure quantum operation can only increase the fidelity, we proved the following
\begin{cor}[Optimality of pure quantum operations
For every fixed value of the success probability, the energy-preserving operation that maximizes the fidelity can be chosen to be pure without loss of generality.
\end{cor}
\iffalse
As an example, consider a system with non-degenerate energy spectrum and let $\map M$ be the quantum operation defined by
\begin{align*} \map M (\rho) = \sum_{E \ge E_0} \< E | \rho |E\> \, |E\>\<E|
\end{align*}
where $|E\>$ is the eigenstate with energy $E$ and $E_0$ is a suitable energy threshold. The L\"uders reduction of $\map M$ is the quantum operation $\map P$ defined by
\begin{align*}
\map P (\rho) = P_{0} \rho P_{0} \qquad P_{0} = \sum_{E \ge E_0} \, |E\>\< E| \,.
\end{align*}
Note that in this example $\map M$ destroys the coherence among energy eigenstates,
while its L\"uders reduction preserves the coherence among all eigenstates with energy above the threshold $E_0$.
\fi
\subsubsection{Optimal quantum operations from Lagrangian optimization} \label{subsec:optimal}
The optimization of the fidelity for given success probability can be completed by Lagrangian optimization.
Let us define the coefficients $$x_E:= \< \varphi_E | \, P_{\rm succ} \, |\varphi_E \> $$
which have values in the interval $[0,1]$ for every energy $E$.
With this definition, the probability of success (\ref{psuccaa}) and the fidelity (\ref{Faa}) can be expressed as
\begin{eqnarray}
p_{\rm succ}&=&\sum_E \, p_E x_E
\end{eqnarray}
and
\begin{eqnarray}\label{lagrangefid}
F=p_{\rm succ}^{-1}\left(\sum_E \sqrt{x_E p_Eq_E}\right)^2 \,
\end{eqnarray}
respectively.
By Lagrangian optimization, we obtain that the optimal filter has a simple structure: the energy spectrum is partitioned into two disjoint subsets, $\set S_0$ and $\set S_1$, and the coefficients of the optimal transformation are given by
\begin{eqnarray}\label{general}
x_E =
\left \{
\begin{array}{ll}
1 \qquad & E \in\set S_0 \\
\frac{p_{\rm succ} - p(\set S_0)}{ 1- q(\set S_0) } \, \frac{q_E}{p_E} \qquad &E \in\set S_1
\end{array}
\right.
\end{eqnarray}
where $p(\set S_0) : = \sum_{E\in\set S_0} p_E$ and $q (\set S_0) : = \sum_{E\in\set S_0} q_E$.
In other words, the optimal filter is completely determined by the choice of the set $\set S_0$. Inserting Eq. (\ref{general}) into Eq. (\ref{lagrangefid}), the maximization of the fidelity is reduced to the maximization of the quantity
\begin{align}
\Omega [ S_0] = \sum_{E\in\set S_0} \sqrt{p_E q_E} + \sqrt{ [ p_{\rm succ} - p (\set S_0)] \, [1-q(\set S_0)]} \, .
\end{align}
Examples of quantum operations of the form (\ref{general}) can be found in Ref. \cite{BaganPRL}, which focused on the specific problem of phase-estimation.
More examples will be provided in section \ref{sec:examples}.
\iffalse
\begin{proof}
Let us expand the state $|\varphi\>$ as
$$|\varphi\>=\sum_{E}\sqrt{p_E}|\varphi_E\> \qquad |\varphi_E\>:= \frac{P_E|\varphi\>}{\|P_E|\varphi\>\|} \, ,$$
where $P_E$ the projector on the energy eigenspace with eigenvalue $E$ and $p_E=\|P_E|\varphi\>\|^2$. Similarly, we expand $|\psi\>$ as $|\psi\>=\sum_E\sqrt{q_E}|\psi_E\>$.
Consider a generic energy-preserving binary filter $\{\map M_{\rm succ}, \map M_{\rm fail}\}$, and expand the transformation $\map M_{\rm succ}$ in the Kraus form $\map{M}_{\rm succ}(\cdot)=\sum_{k}M_k\cdot M_k^\dagger$. By the energy-preserving condition [Eq. (\ref{block})] we have $$M_k=\bigoplus_E M_E^{(k)}$$ with $M_E^{(k)}:=P_EM_kP_E$. Fixing the probability of success to be $p_{\rm succ}$, we have the upper bound
\begin{eqnarray}
\nonumber F\cdot p_{\rm succ}&=&\sum_k \left|\<\psi|M^{(k)}|\varphi\>\right|^2\\
\nonumber &=&\sum_k\left|\sum_E \sqrt{p_Eq_E}\<\psi_E|M^{(k)}_E|\varphi_E\>\right|^2\\
\nonumber &\le&\sum_k\left(\sum_E \sqrt{p_Eq_E}\left|\<\psi_E|M^{(k)}_E|\varphi_E\>\right|\right)^2\\
\label{uuupper} &\le&\sum_k\left(\sum_E \sqrt{p_Eq_E}\left\|M^{(k)}_E|\varphi_E\>\right\|\right)^2 \, .\end{eqnarray}
Now, let us define the operator
\begin{eqnarray}\label{coarse}
M_E:= \sqrt{\sum_k M_E^{(k)\dagger}M_E^{(k)}}
\end{eqnarray}
and let us set $a_E : = \sqrt { \| M_E |\varphi_E \> \| } $ and $b_{E,k} : = \| M_E^{(k)} |\varphi_E\> \| / a_E$. With this notation, the upper bound (\ref{uuupper}) becomes
\begin{align}
\nonumber F\cdot p_{\rm succ} & \le \sum_k\left(\sum_E \sqrt{p_Eq_E} \, a_E \, b_{E,k} \right)^2 \\
\nonumber &\le \sum_k \left (\sum_E \sqrt{p_Eq_E} \, a^2_E \right ) \, \left ( \sum_{E'} \sqrt{p_{E'}q_{E'}} \, b^2_{E',k} \right) \\
\nonumber & = \left (\sum_E \sqrt{p_Eq_E} \, a^2_E \right )^2 \\
\label{bound}& = \left (\sum_E \sqrt{p_Eq_E} \, \| M_E |\varphi_E\> \| \right )^2 \,
\end{align}
having used the Cauchy-Schwarz inequality in the second step. The upper bound can be achieved by choosing a filter with successful Kraus operation $\map M'_{\rm succ} (\cdot) = M' \cdot M^{\prime \dag} $ defined by
\begin{eqnarray}\label{coarseDEF}
M' = \bigoplus_E M'_E \, , \quad M'_E:= U_E M_E
\end{eqnarray}
where $U_E$ is any unitary operator that maps the subspace $\spc{H}_E : = P_E \spc H$ to itself and transforms $ M_E |\varphi_E\>/\| M_E |\varphi_E\>\|$ into $|\psi_E\>$. Therefore we can construct the energy-preserving binary pure filter as $\{\map M'_{\rm succ}, \map M_{\rm fail}\}$. It is straightforward to prove that $\|M'_E|\varphi_E\>\|^2=\sum_k \|M_E^{(k)}|\varphi_E\>\|^2$, which means that the probability of success for the filter $\map M_{\rm succ}'$ is equal to the probability of success of $\map M_{\rm succ}$.
\end{proof}
\fi
\medskip
\subsection{The ultimate limits of probabilistic energy-preserving processes}
So far we considered the optimization of the fidelity for fixed success probability. We now remove the constraint and focus only on the maximization of the fidelity. The problem is interesting because it highlights the quantum limits to what is ultimately possible \cite{calsamiglia-2014-natphys}.
The ultimate limits for energy-preserving operations is characterized by the following
\begin{prop}\label{cor:maxfid}
For finite dimension Hilbert space $\spc H \simeq \mathbb{C}^d$, the maximum of the fidelity over all energy-preserving quantum operations is
\begin{align}\label{fmax} F_{\max} = \sum_{E \in \mathsf{Sp} (\varphi) \cap \mathsf{Sp}(\psi) } \, q_E \, ,
\end{align}
where $\mathsf{Sp}(\varphi)$ denotes the energy spectrum of the state $|\varphi\>$, defined as
$$ \mathsf{Sp} (\varphi) : = \{ E ~|~ \< \varphi | P_E |\varphi\> \not = 0\} \, .$$
For a quantum operation achieving fidelity $F_{\max}$ the maximum probability of success is given by
\begin{align}\label{psuccmax} p^{\max}_{\rm succ} = \left( \min_{E \in \mathsf{Sp}(\varphi) \cap \mathsf{Sp}(\psi) } ~ \frac {p_E}{q_E} \right) \, F_{\max} \, .
\end{align}
The quantum operation achieving maximum fidelity with maximum probability is pure and its Kraus operator satisfies the condition
\begin{align}\label{Moptimized} M |\varphi_E\> = \left[ \min_{E' \in \mathsf{Sp}(\varphi) \cap \mathsf{Sp}(\psi) } \sqrt{ \, \frac { p_{E'} }{q_{E'}} }\right] \, \sqrt{\frac{q_E}{ p_E } } ~ |\psi_E\>
\end{align}
for every energy $E$ in $\mathsf{Sp}(\varphi)$.
\end{prop}
The proof is provided in Appendix \ref{app:ultimate}.
\section{Multiround recursive protocols}\label{sec:network}
In the previous section we provided a recipe to find the protocol that achieves maximum fidelity for a fixed value of the success probability. By definition, the resulting protocol is taylor-made to that specific value of the probability. However, in many situations it is useful to have a more flexible protocol, where the experimenter can make successive attempts at realizing the desired transformation and is free to decide on the fly when to stop.
In this section we analyze protocols of this form, which we refer to as \emph{recursive protocols}.
Under the energy-preserving constraint, we identify the protocol that produces the best possible approximation of the target state at each step.
\subsection{The optimal recursive protocol}\label{subsec:RUS}
Given a sequence of $K$ binary filters
with outcomes $\{\rm succ, \rm fail\}$, consider the protocol defined by the following instructions:
\begin{enumerate}
\item Set $k=1$.
\item If $k\le K$, then apply the $k$-th filter; else terminate.
\item If the outcome is $x= \rm succ$, then terminate.
\item If the outcome is $x=\rm fail$, then replace $k$ with $k+1$ and go back to instruction 1.
\end{enumerate}
Recursive protocols of this form are an example of ``quantum loop programs", studied in Ref. \cite{loop}. All these protocols can be can be visualized as decision trees of the following form
\begin{equation*}
\begin{aligned}\Qcircuit @C=1em @R=0.2em @!R
{
& \multigate{1}{ {\rm Filter~ 1}} & \qw\poloFantasmaCn{\rm succ}&\qw& &\\
& \pureghost{ {\rm Filter~ 1} } &\qw\poloFantasmaCn{\rm fail}&\qw &\multigate{1}{ {\rm Filter~ 2} } &\qw\poloFantasmaCn{\rm succ} &\qw& \\
& & & &\pureghost{ {\rm Filter ~2}} & \qw\poloFantasmaCn{\rm fail}&\qw&\\
& & & & & &\vdots&\\
& & & & & & \qw\poloFantasmaCn{\rm fail} &\multigate{1}{ {\rm Filter~ K} }&\qw\poloFantasmaCn{\rm succ}& \qw \\
& & & & & & &\pureghost{ {\rm Filter ~K}} & \qw\poloFantasmaCn{\rm fail} &\qw
}
\end{aligned}
\end{equation*}
Protocols of the above form have been employed for different purposes, including entanglement concentration \cite{divincenzo}, implementation of quantum gates \cite{RUS1,RUS2,RUS3} and ancilla-driven computation \cite{ADQC}. One such protocol that is particularly similar to ours is \emph{quantum rejection sampling} \cite{rejectionsampling}. There, the goal is to generate a target superposition $|\psi\>$ using a black box $U_\varphi$ that prepares another superposition $|\varphi\>$ from a fixed state $|0\>$. The difference between rejection sampling and our problem is that in our case we do not have the black box $U_\varphi$ at disposal, but only the coherent superposition $|\varphi\>$, which is a strictly weaker resource due to Nielsen and Chuang's no-programming theorem \cite{nielsen1997programmable}.
In our case, the goal of the protocol is to transform a coherent superposition of energy eigenstates into another.
Of course, at each step there will be a tradeoff between the fidelity with the target and the probability of success. In the simplest scenario, the protocol can be designed to attain the absolute maximum of the fidelity at each round, and to do so with maximum probability of success.
An experimenter following such a protocol will have the guarantee that the best possible performance is achieved in each individual round.
We consider the transition $|\varphi\> \to |\psi\>$ in the case of states $|\varphi\>$ and $|\psi\>$ in a finite dimensional Hilbert space $\spc H \simeq \mathbb{C}^d$, $d< \infty$, or, more generally, states whose energy spectra intersect on a finite set of points, with $| \mathsf{Sp} (\varphi) \cap \mathsf{Sp}(\psi)| \le d$ For the optimal protocol we make a list of desiderata in decreasing order of priority:
for every $k\in\{1,\dots, K-1\}$
\begin{enumerate}
\item at the $k$-th round, the successful quantum operation should transform the input state $\rho^{(k)}$ into the target $ |\psi\>$ with maximum fidelity
\item the optimal transition $\rho^{(k)} \to |\psi\>\<\psi|$ must be achieved with maximum probability of success
\item the unsuccessful quantum operation at the $k$-th round should produce the state $\rho^{(k+1)}$ that leads to maximum fidelity for the transformation $\rho^{(k+1)} \to |\psi\>\<\psi|$ at the $(k+1)$-th round
\item at the final round ($k=K$) the successful quantum operation should achieve maximum fidelity with maximum probability and, conditional on the fulfillment of this requirement, the unsuccessful quantum operation should achieve maximum fidelity.
\end{enumerate}
The derivation of the optimal protocol is rather technical and is provided in Appendix \ref{app:derivation}. In the following we present the final result of the optimization and discuss the implications of our findings.
At the $k$-th round, the optimal binary filter consists of two pure quantum operations, $\map B^{(k)}_{\rm succ} (\cdot) = B^{(k)}_{\rm succ} \cdot B^{(k) \dag}_{\rm succ} $ and $\map B^{(k)}_{\rm fail} (\cdot) = B^{(k)}_{\rm fail } \cdot B^{(k)\dag}_{\rm fail}$. Since all quantum operations are pure, the state of the system is pure at every round. The input state at
the $k$-th round, denoted by $|\varphi^{(k)}\>$, can be expanded as
$$\left|\varphi^{(k)}\right\> = \sum_{E} \sqrt{ p^{(k)}_{E}} \, \left |\varphi_{E} \right \> \, , $$
where the energy eigenstates are the same as in the decomposition of $|\varphi\>$.
With this notation, the successful quantum operation determined in an essentially unique way by the condition
\begin{align}
\nonumber B^{(k)}_{\rm succ} \left| \varphi_E \right \> = & \left[ \min_{E' \in \mathsf{Sp}(\varphi^{(k)}) \cap \mathsf{Sp}(\psi) } \sqrt{ \, \frac { p^{(k)}_{E'} }{q_{E'}} }\right] \sqrt{\frac{q_E}{ p^{(k)}_E } } ~ |\psi_E\> \\
\label{bksucc} & \qquad \qquad \qquad \qquad\quad \forall E \in \mathsf{Sp}(\varphi^{(k)})\, .
\end{align}
Here the only freedom is in the definition of the operator in the subspace spanned by energy eigenstates outside the spectrum of $|\varphi^{(k)}\>$. The form of Eq. (\ref{bksucc}) follows directly from the requirements 1 and 2 in our list of desiderata (cf. proposition \ref{cor:maxfid}). Similarly, the unsuccessful quantum operation is determined in an essentially unique way by the condition
\begin{align}\label{bkfail}
B^{(k)}_{\rm fail} = \sqrt{ I- B^{(k)\dag}_{\rm succ} B^{(k)}_{\rm succ}} \, .
\end{align}
The form of Eq. (\ref{bkfail}) follows from the requirement 3 in our list. Remarkably, the quantum operation $\map B^{(k)}_{\rm fail}$ does not maximize only the fidelity achievable from the input state $\rho^{(k+1)}$, but also the \emph{probability} that maximum fidelity is achieved.
\subsection{Fidelity and success probability}
The optimal protocol is specified recursively by equations \eqref{bksucc} and \eqref{bkfail}. Making the dependence on the input and target states explicit, it is possible to derive closed formulas for the fidelity and the success probability.
To this purpose, we introduce some notation. For every energy $E \in \mathsf{Sp} (\varphi) \cap \mathsf{Sp}(\psi)$, we define the ratio
$ p_E/q_E $ and arrange its possible values in increasing order as
\[ 0<r_1 < r_2 < \dots < r_L \, , \] where $r_L$ is the maximum ratio. Clearly, by the assumption of finite dimensionality, $L $ satisfies the relation
$$L \le | \mathsf{Sp} (\varphi) \cap \mathsf{Sp}(\psi) | \le d \, .$$
For every possible value $r_i$, we consider the set of energy eigenvalues $\set R_i$ defined as
\begin{align}\label{R}
\set R_i := \left \{ E \in \mathsf{Sp}(\varphi) \cap \mathsf{Sp} (\psi) ~ \left |~ \frac {p_E}{q_E} = r_i \right. \right\}
\end{align}
and we denote the union of the first $k$ sets by
\begin{align}\label{U}
\set U_k : = \bigcup_{i=1}^k \, \set R_i \, .
\end{align}
With this definition, the fidelity and the success probability at the $k$-th step can be expressed as
\begin{align}\label{fidelity1}
F_{\max}^{(k)}
\sum_{E \in \mathsf{Sp}(\varphi) \setminus \set U_{k-1} } \, q_{E}
\, .
\end{align}
and
\begin{eqnarray}\label{prob1}
p_{\rm succ}^{(k)}= (r_k-r_{k-1}) \, \cdot \, F^{(k)}_{\max} \, ,
\end{eqnarray}
respectively. The proof is presented in Appendix \ref{app:Mk}. Note that the fidelity is strictly decreasing with $k$, reaching zero for $k = L$. In other words, it is useless to consider protocols with more than $L$ rounds.
\iffalse The decrease of the fidelity with time is the main difference between our protocol and the repeat-until-success protocols used for the implementation of quantum gates \cite{RUS1,RUS2,RUS3}. In our case the desired state transformation may be impossible under the energy-preserving constraint, which implies that the subsequent iterations of the protocol will always be affected by an error. ``Success'' refers here to the success in generating the best possible approximation of the target compatibly with the restriction that the energy must be preserved on average.
\fi
The explicit expressions given by Eqs. (\ref{fidelity1}) and (\ref{prob1}) turn out to be very useful
for studying the tradeoff between fidelity and success probability.
%
Indeed, they allow us to evaluate the probability that the protocol succeeds in one of the first $T$ rounds, given by
\begin{eqnarray}
\nonumber
p_{{\rm succ}} (T) &:=&\sum_{k=1}^{T}p_{\rm succ}^{(k)} \\
& =& \sum\limits_{E\in \set U_{T-1} } p_E + r_T \, F_{\rm max}^{(T)}
\label{Ptot}
\, ,
\end{eqnarray}
and to observe its scaling with the average fidelity achieved in the first $T$ steps, given by
\begin{eqnarray}
F(T)&:= \frac { \sum_{k=1}^{T} \, p_{\rm succ}^{(k)} ~F_{\max}^{(k)} } { p_{\rm succ}(T) } \, .
\end{eqnarray}
The tradeoff curve between $F(T)$ and $p_{\rm succ} (T)$ will be illustrated in section \ref{sec:examples} for a number of concrete examples.
\subsection{Output states and termination time of the protocol}
In addition to the fidelity and success probability, it is useful to know what states are produced at every step of the protocol. Assuming that the total number of rounds is upper bounded as $K\le L$, the explicit expression of the output state produced at the $k$-th round can be obtained as follows.
We regard the recursive protocol as a quantum instrument, with outcomes in the set $\{1,\dots , K+1\} $. The outcome $k$ corresponds to the pure quantum operation with Kraus operator
\begin{eqnarray}\label{Mk}
M_k:= \left \{
\begin{array}{ll}
B^{(k)}_{\rm succ}B_{\rm fail}^{(k-1)}\cdots B_{\rm fail}^{(1)} \quad & k=1,\dots, K, \\
& \\
B^{(K)}_{\rm fail}B_{\rm fail}^{(K-1)}\cdots B_{\rm fail}^{(1)}& k = K+1
\end{array} \, .
\right.
\end{eqnarray}
For $k \le K$, the Kraus operators are characterized in Appendix \ref{app:Mk}, where we prove the following relation
\begin{align}\label{solution1}
M_k |\varphi_E\> =
\left\{
\begin{array}{ll}
0 \qquad & E\in \set U_{k-1} \\ \\
\sqrt{ (r_k -r_{k-1}) \, \frac{q_E}{p_E} }\, |\psi_E\> & E \in \mathsf{Sp}(\varphi) \setminus \set U_{k-1} \, ,
\end{array}
\right.
\end{align}
having set $r_0: = 0$.
Hence, the output state in case of success at the $k$-th round is given by
\begin{align}
\nonumber | \psi^{(k)} \> & : = \frac{ M_k |\varphi\>}{ \| M_k |\varphi\> \|} \\
\label{outputk} & = \frac 1 {\sqrt{ F_{\max}^{(k)}}} \sum_{E \in \mathsf{Sp}(\varphi) \setminus \set U_{k-1} } \sqrt{ q_E} \, |\psi_E\> \, .
\end{align}
Note that $|\psi^{(k)}\>$ is a truncated version of the target state, in which the energy spectrum has been deprived of all the values in $\set U_{k-1}$ and of all the values that are not in in the spectrum of $|\varphi\>$. The energy spectrum of the output state is eroded from one step to the next: each iteration of the protocol produces a state with a strictly lesser amount of coherence in the energy eigenbasis. Due to the assumption of finite dimensionality, the process of erosion terminates in a finite number of steps, equal to $T_{\max}$.
Protocols with more than $T_{\max}$ rounds terminate after $T_{\max}$ steps, meaning that the probability of success satisfies
$$ p_{\rm succ} (T) = 1 \qquad \forall\ T > T_{\max} \, .$$
The fact that the protocol is guaranteed to terminate in a finite time is an appealing feature.
It is worth stressing that the termination time $T_{\max}$ is upper bounded by the number of distinct energy levels of the system, which can be much smaller than the dimension of the Hilbert space, as in the following
\begin{eg}
Consider the case of $N$ identical non-interacting systems of dimension $d$. In this case the total Hamiltonian is the sum of the single-system Hamiltonians, and its number of energy levels is upper bounded by the number of partitions of $N$ into $d$ non-negative numbers (see e.g. \cite{GiulioNat}). We then have that the number of rounds needed to terminate is upper bounded as
$$ T_{\max} \le \begin{pmatrix} d+N-1 \\ N \end{pmatrix} < (N+1)^{d-1} \, , $$
i.~e.~ by a polynomial in $N$. Even if the probability of success in the first round is exponentially small in $N$, as in the case of quantum super-replication \cite{GiulioNat}, the recursive protocol is guaranteed to reach unit probability in a polynomial number of iterations.
\end{eg}
\subsection{Increasing the fidelity of the recursive protocol}
At every iteration of the recursive protocol, the total probability of success increases, while the average fidelity decreases. In general, the relation between fidelity and probability of success is not optimal, because the histories leading to successful outcomes are mixed incoherently: at the $T$-th step, the successful quantum operation has the form
\begin{align}
\map M^{(T)} (\rho) = \sum_{k=1}^T \, M_k \rho M_k^\dag \, ,
\end{align}
where $M_k$ are the Kraus operators defined in Eq. (\ref{Mk}). Now, we have a systematic method to increase the fidelity while keeping the same success probability: the goal is accomplished by taking the L\"uders reduction of $\map M^{(T)}$ and performing eigenstate alignment on the output. The following paragraphs highlight the main features of this method.
\subsubsection{Coherent coarse-graining}
L\"uders reduction transforms the quantum operation $\map M^{(T)}$ into the pure quantum operation
$$\map P^{(T)} (\cdot) = \sqrt{ P^{(T)}} \,(\cdot )\, \sqrt{ P^{(T)}} $$ with
$$ P^{(T)} = \map M^{(T) \dag} (I) = \sum_{k=1}^T \, M_k^\dag M_k\,. $$
The technique of joining different quantum operations into a single one is an important tool and as such deserves a name:
\begin{defi}
We call $\map P^{(T)}$ the \emph{coherent coarse-graining} of the quantum operations $\{\map M_k~|~ k =1 , \dots, T\}$.
\end{defi}
An intuitive way to visualize coherent coarse-graining is through a generalization of the double slit experiment involving multiple slits \cite{sorkin1994quantum}. When detectors are placed on the slits, the passage of a particle through the $k$-th slit will trigger the occurrence of the quantum operation $\map M_k$. When the detectors at the first $T$ slits are removed, the passage through these slits will result into the coherently coarse-grained operation $\map P^{(T)}$.
\subsubsection{Eigenstate alignment}
By construction, coherent coarse-graining does not change the probability of success. The fidelity is then increased by eigenstate alignment, achieved by any energy-preserving unitary $U$ such that
$$ U |\varphi_E\> = |\psi_E\> \qquad \forall E \in \mathsf{Sp}(\varphi) \cap \mathsf{Sp}(\psi) \, .$$
Note that the operation of eigenstate alignment does not depend on how many rounds of the protocols are coarse-grained. The operation could be performed even before the filter is applied, provided that one suitably adapts the definition of the filter.
When combined, coherent coarse-graining and eigenstate alignment yield the pure quantum operation
\begin{align}\label{coarsesolution}
\map M^{(T)\prime} (\cdot) = M^{(T)\prime} \cdot M^{(T)\prime\dag} \, , \qquad M^{(T)\prime} : = U \sqrt{ \sum_{k=1}^T M_k^\dag M_k} \, ,
\end{align}
whose action on the energy eigenstates is given by
\begin{align}\label{MT}
M^{(T)\prime}|\varphi_E\> =
\left\{
\begin{array}{ll}
|\psi_E\> \qquad & E \in \set U_T \\ \\
\sqrt{ r_T \, \frac {q_E}{p_E}} |\psi_E\> & E \not \in \mathsf{Sp} (\varphi) \setminus \set U_T \, ,
\end{array}
\right.
\end{align}
having used Eq. (\ref{solution1}) for the Kraus operators of the original protocol. Note that the above equation implies that the operator $M^{(T) \prime}$ satisfies the Lagrange extremality condition (\ref{general}).
For $T$ larger than the termination time $T_{\max}$, our construction eventually yields the \emph{optimal} energy-preserving channel for the transition $|\varphi\> \to |\psi\>$ (cf. Theorem \ref{theo:EAoptimal}).
\subsubsection{The performance of the coherently coarse-grained protocol}
Taking eigenstate alignment as an obliged step in the optimal operation, we refer to the quantum operation $\map M^{(T)\prime}$ as the \emph{coherent coarse-graining} of (the first $T$ steps of) the protocol. By construction, the probability of success of the quantum operation $\map M^{(T)\prime}$ is equal to the original success probability $p_{\rm succ}(T)$ given in Eq. (\ref{prob1}).
On the other hand, the fidelity can be evaluated explicitly by using Eq. (\ref{MT}), which yields
\begin{align}\label{F'}
F' (T)=\frac{\left[ \sum\limits_{E\in\set U_T }\sqrt{p_E q_E} + \sqrt{ r_T } \, \sum\limits_{ E \in \mathsf{Sp}(\varphi) \setminus \set U_T } q_E \right]^2}{ \sum\limits_{E\in\set{U}_T} p_E + r_T \sum\limits_{E\in \mathsf{Sp}(\varphi) \setminus \set U_T}q_E}.
\end{align}
Performing the operation of coherent coarse-graining for different values of $T$ one can obtain a sequence of filters that approximate the optimal curve of the fidelity-probability tradeoff. The improvement due to coherent coarse-graining will be illustrated in the next section with a number of concrete examples.
\section{Applications}\label{sec:examples}
In this section we apply the recursive protocol and the method of coherent coarse-graining to the tasks of parameter estimation, cloning of quantum clocks, phase-insensitive amplification of coherent states, and approximate correction in ancilla-driven quantum computation.
\subsection{Probabilistic phase estimation with multiple copies}\label{subsec:estimation}
Here we apply the recursive protocol to the task of phase estimation \cite{holevo_stat_book,helstrom}.
When the phase is encoded in a quantum state in a suboptimal way, one can try to improve the precision of phase estimation by first transforming the state into the optimal state. Of course, such transition cannot take place deterministically---for otherwise the original state would have been already optimal.
However, a probabilistic protocol can produce good approximations of the optimal state and, conditionally on the success of the probabilistic transformations, it can enable an improved phase estimation. In the following we will use our recursive protocol for this purpose.
To illustrate the idea, we consider the simple case where the phase is encoded into the state of a two-level quantum system (qubit), as
\[|\varphi_\theta\> = e^{-i\theta Z} |\varphi\> \qquad \theta \in [0,2\pi) \, , \]
with $ Z = |0\>\<0| - |1\>\<1|$ and $|\varphi\> =(|0\>+e^{i\theta}|1\>)/\sqrt 2$.
We assume that $N$ identical copies of the state are available and search for the optimal strategy to estimate $\theta$.
To quantify the precision, we use the gain function $G(\theta, \hat \theta)$ defined by \cite{holevo_stat_book}
\begin{align*}
G(\theta,\hat{\theta}):=\frac{1+\cos(\theta-\hat{\theta})}{2} \, .
\end{align*}
Note that the gain funcgion assigns a larger gain when the estimate $\hat \theta$ is closer to the true value $\theta$, attaining the maximum value $1$ if and only if $\hat \theta = \theta$.
Then, the goal is to find the estimation strategy that maximizes the average gain
\begin{equation}\label{GDEF}
\begin{split}
\<G\>&:=\int\frac{d\theta}{2\pi}\int\frac{d\hat{\theta}}{2\pi} \, G(\theta,\hat{\theta}) ~ \< \varphi_\theta|^{\otimes N} \, E_{\hat \theta} \, |\varphi_\theta\>^{\otimes N}
\end{split} \, .
\end{equation}
where $\left\{E_{\hat{\theta}} \, \right\}$ is the Positive Operator-Valued Measure (POVM) describing the estimation strategy.
For phase estimation with pure states, the optimal POVM has been derived by Holevo \cite{holevo_stat_book}. Specifically, for pure states of the form
\[ |\Phi_\theta \> = \sum_{n=0}^{N} c_n \, e^{-i\theta n} \, |n\> \, ,\]
Holevo's POVM yields the gain
\begin{equation}\label{G}
\begin{split}
\<G_{\rm det}\>&=\frac{1}{2}+\frac{1}{2}\<\Phi_0| \Delta|\Phi_0\>\\
\Delta_{mn}&=\frac{1}{2}\left[\delta_{m(n-1)}+\delta_{m(n+1)}\right] \, .
\end{split}
\end{equation}
In our case, the above expression yields the value
\begin{align}
\label{detG1}
\<G_{\rm det} \> &=\frac12+\frac{1}{2^{N+1}}\sum_{n=0}^{N-1}\sqrt{{N\choose n}{N\choose n+1}} \\
\nonumber
& = 1- O\left(\frac 1 N\right) \, .
\end{align}
Now, when the unknown phase shift $e^{-i\theta Z}$ is probed $N$ times, one can obtain a much better estimate by preparing the optimal input state, which in this case is the ``sine-state" \cite{sin,berry-wiseman-2000-prl}
\begin{align}\label{EstTarget}
|\varphi_{{\rm opt},\theta}\>=\sqrt{\frac{2}{N+1}} \, \sum_{n=0}^{N}\sin\left(\frac{n\pi}{N+1}\right) e^{i\theta n} \, |n\> \, .
\end{align}
This state achieves the Heisenberg scaling $\<G \> = 1- O(1/N^2)$.
In the following, we will use our recursive protocol to transform the state $|\varphi_\theta\>^{\otimes N}$ into the optimal state $|\varphi_{\rm opt,\theta}\> $ and we will use the output states of our protocol for phase estimation.
Note that the output states of our protocol are pure at every step. Thanks to this fact, we can apply Holevo's recipe (\ref{G}) to compute the optimal gain. Precisely, the gain at the $k$-th step is given by
\begin{align}\label{Gn}
\left\<G^{(k)} \right\>=\frac{1}{2}+\frac{1}{2} \, \left\<\psi^{(k)} \right| \, \Delta \,\left|\psi^{(k)} \right\> \, ,
\end{align}
where $\left |\psi^{(k)} \right\>$ is the output state at the $k$-th step, given by Eq. (\ref{outputk}). Averaging over the first $T$ steps we obtain the gain
\begin{align}
\< G_T \>&:=\frac{\sum_{k=1}^{T}p_{\rm succ}^{(k)}\<G^{(k)}\>}{p_{\rm succ}(T)} \label{Gav1}
\end{align}
where $p_{\rm succ}^{(k)}$ is the probability of achieving success at the $k$-th step and $p_{\rm succ} (T) = \sum_{k=1}^T \, p_{\rm succ}^{(k)}$. The value of the gain can be explicitly calculated using Eqs. (\ref{Ptot}), (\ref{outputk}), and (\ref{solution1}). In Figure \ref{fig:EstTradeoff} we show the estimation gain for $N=30$ copies of the input state and for $K= 27 $ iterations of the recursive protocol.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\linewidth]{EstTradeoff1.pdf}\\
\caption{{\bf Probabilistic phase estimation via the recursive protocol and its coherent coarse-graining.} The figure shows the tradeoff between success probability and average gain for phase estimation with the qubit state $|\varphi_\theta\>^{\otimes N}$ for $N=30$. The green solid line (with numerics represented by red dots) shows the tradeoff between estimation gain and success probability for a recursive protocol with $K= 27$ rounds, with the $T$-th point corresponding to the first $T$ steps. The blue solid line (with numerics represented by the black dots) shows the tradeoff for filters generated by coherent coarse-graining, with the $T$-th point corresponding to the coherent coarse-graining of the first $T$ steps. Note that the gain for the coherent coarse-graining remains higher than the optimal deterministic estimation's gain (the black dashed line) even when the protocol becomes ``almost deterministic" (i.e. the probability of success tends to one), while the gain for the recursive protocol drops down quickly with the growth of the success probability. }
\label{fig:EstTradeoff}
\end{figure}
The performance of recursive protocol can be compared with the performances of its coherent coarse-graining. By coherently coarse-graining over the first $T$ rounds, we obtain the average gain given by
\begin{align}\label{Gavcoarse1}
\<G'_{T}\>=\frac{1}{2}+\frac{1}{2} \, \left\< \psi' (T ) \right| \Delta \left|\psi'(T) \right\> \, ,
\end{align}
with
\[\qquad \left|\psi'(T) \right\>
= \frac{ M^{\prime (T) } |\varphi\>^{\otimes N} }{ \| M^{\prime (T)} |\varphi\>^{\otimes N} \| }\]
and $M^{(T)}$ as in Eq. (\ref{coarsesolution}). The estimation gain of the coherent coarse-graining are plotted in Figure \ref{fig:EstTradeoff}.
In addition, Figure \ref{fig:EstGandP} shows the scaling of the gain and the success probability with the number of copies $N$.
\begin{figure}[t!]
\centering
\subfigure[]{\label{fig:EstGain1}
\includegraphics[width=0.9\linewidth]{EstGain1.pdf}
}
\subfigure[]{\label{fig:EstProb1}
\includegraphics[width=0.9\linewidth]{EstProb1.pdf}
}
\caption{{\bf Scaling of the gain and success probability for coherently coarse-grained protocols. } Figure \ref{fig:EstGain1} shows the estimation gain as a function of the number of copies $N$, for three coherently coarse-grained protocols corresponding to different values of $T$, including $T=1$ (black line with black dots), $T=3$ (green line with red dots) and $T=5$ (blue line with purple dots). The dashed line with black dots represents the optimal deterministic gain $\<G_{\rm det}\>$. Figure \ref{fig:EstProb1} shows the decrease of the total success probability as a function of $N$ for different values of $T$, including $T=1$ (black line with black dots), $T=3$ (green line with red dots) and $T=5$ (blue line with purple dots). }
\label{fig:EstGandP}
\end{figure}
\subsection{Converting coherence into metrological precision}\label{subsec:maxcoh}
In the previous paragraph we analyzed the problem of phase estimation with $N$ identical qubits.
Here we will consider a one-shot scenario, where the phase has to be estimated from a single copy of the state
\begin{align}\label{EstResource}
|\varphi_\theta\>:=\frac{1}{\sqrt{N}}\sum_{n=0}^{N-1} \, e^{-in\theta} \, |n\> \, ,
\end{align}
consisting of a uniform superposition of the eigenstates of the Hamiltonian $H = \sum_{n=1}^{N-1} \, n\, |n\>\<n|$. The above state is the \emph{maximally coherent state} \cite{incoherence} in the energy eigenbasis, that is, it is the most valuable state in the resource theory of coherence. Interestingly, it is \emph{not} the optimal state for phase estimation.
Indeed, the estimation gain for the maximally coherent state can be evaluated with Eq. (\ref{G}), which in this case yields
\begin{align}\label{detG}
\<G_{\rm det} \>=1-\frac1 { 2N} \, .
\end{align}
When the number of copies is asymptotically large, the gain approaches its maximum value with the standard quantum limit scaling $1/N$, rather than the Heisenberg scaling $1/N^2$.
We now explore how the maximally coherent state can be transformed into the optimal state for phase estimation. The performances of the recursive protocols and of its coherent coarse-graining can be evaluated using Eqs. (\ref{Gav1}) and (\ref{Gavcoarse1}).
When the number of iterations $T$ is small compared to the number of energy levels $N$, the average gain has the simple analytical expression
\begin{align}\label{asyG}
\<G_T\>=1-\frac{\pi^2}{2N^2}\left[T(T-1)+\frac12\right]+O\left[\left(\frac{T}{N}\right)^{3}\right] \, .
\end{align}
Note that the gain exhibits Heisenberg scaling with the number of energy levels $N$, with a constant that grows quadratically with the number of rounds $T$. The success probability can also be evaluated analytically in the regime $ N \gg T$ and its value is given by
\begin{align}\label{asyP}
p_{\rm succ}(T)=\frac{1}{2}+\frac{\pi^2}{N^2}\left[T(T-1)+\frac18\right]+O\left[\left(\frac{T}{N}\right)^{3}\right] \, .
\end{align}
From the above expressions, one can clearly see the tradeoff between gain and success probability, which can be made explicit in the tradeoff curve
\begin{align}\label{curve}
\<G_T\>=1-\frac{3\pi^2}{16N^2}-\frac{p_{\rm succ}(T)-1/2}{2} \, , \qquad N \gg T \, .
\end{align}
In Figure \ref{fig:EstTradeoff} we illustrate the tradeoff between the probability of success and the average gain for $N =61$. The recursive protocol manages to increase the probability of success by approximately $30\%$ from the first round to the 14-th, while keeping the average gain above the deterministic gain. In Figure \ref{fig:EstGandP} we show the scaling of the gain and the success probability with the dimension $N$ for different values of $T$.
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\linewidth]{EstTradeoff.pdf}\\
\caption{{\bf Probabilistic phase estimation via the recursive protocol and its coherent coarse-graining.} The figure shows the tradeoff between success probability and average gain for phase estimation with maximally coherent states with $N=61$. The green solid line (with numerics represented by red dots) shows the probability-gain tradeoff for $K= 17$ rounds of the recursive protocol. At the first round the protocol reaches the maximum possible gain, equal to $G_{\max} = 99.9\%$, in agreement with the analytical expression of Eq. (\ref{asyG}). The blue solid line (with numerics represented by the black dots) shows the tradeoff for filters generated by coherent coarse-graining, with the $T$-th point corresponding to the coherent coarse-graining of the first $T$ steps of the recursive protocol. For the first $K=17$ rounds the estimation gain of coherent coarse-graining remains approximately equal to $G_{\max} = 99.9\%$, although eventually it is bound to decrease to the optimal deterministic value $\< G_{\rm det}\> = 99.2\%$ (black dashed line). }
\label{fig:EstTradeoff}
\end{figure}
\begin{figure}[h!]
\centering
\subfigure[]{\label{fig:EstGain}
\includegraphics[width=0.9\linewidth]{EstGain.pdf}
}
\subfigure[]{\label{fig:EstProb}
\includegraphics[width=0.9\linewidth]{EstProb.pdf}
}
\caption{{\bf Scaling of the estimation gain and success probability for the recursive protocol.} Figures \ref{fig:EstGain} and \ref{fig:EstProb} show the average gain $G$ and the total success probability (\ref{fig:EstProb}) as a function of the energy scale $N$ for different values of $T$, including $T=1$ (black line with black dots), $T=2$ (green line with red dots) and $T=3$ (blue line with purple dots). }
\label{fig:EstGandP}
\end{figure}
Let us evaluate now the performance of coherent coarse-graining. In the $N\gg T$ regime, the gain has the analytical expression
\begin{align*}
\< \, G_T' \, \> =1-\frac{\pi^2}{4N^2}\, \left \{1+4 \, \left[\, p_{\rm succ}(T)-\frac 12\, \right]^2 \right\} \, .
\end{align*}
The tradeoff between estimation gain and probability of success is illustrated in Figure \ref{fig:EstTradeoff} for $N=61$.
Also in this case, one can observe that coherent coarse-graining offers a better tradeoff curve than the recursive protocol.
Finally, it is interesting to compare the performance of the coherent coarse-graining with the optimal tradeoff curve between gain and probability of success, which is known explicitly in the case under consideration \cite{BaganPRL}. Remarkably, the comparison shows that for large $N$ the coherent coarse-graining yields exactly the optimal estimation strategy of Ref. \cite{BaganPRL}. In other words, in this case the coherent coarse-graining of our recursive protocol is asymptotically optimal.
At this point, a natural question is whether coherent coarse-graining \emph{always} gives the optimal fidelity/probability tradeoff. The answer turns out to be negative: by evaluating Eq. (\ref{Gavcoarse1}) for small values of $N$ (e.g. $N=10$) we find out that the average gain of the coherent coarse-graining sometimes falls below the threshold of the optimal deterministic gain in Eq. (\ref{detG}), clearly indicating sub-optimality in the non-asymptotic regime.
\subsection{Energy-preserving cloning of quantum coherence}\label{subsec:cloning}
Here we consider the problem of quantum cloning \cite{rmp,cerfreview}, where the task is to transform $N$ identical copies of an unknown quantum state into a larger number $M \ge N$ of approximate copies.
In most cases, the problem has been addressed without imposing any constraint on the cloning process, except for its compatibility with the laws of quantum mechanics. Instead, here we consider copy machines that have to work without any supply of energy for the outside. Consider for example a scenario where one wants to clone the state of a quantum clock \cite{GiulioNat}, given by
$$ |\psi_t\> = e^{ - it H/\hbar} |\psi\> \, , $$
where $H = H^\dag$ is a suitable Hamiltonian. Here the time parameter $t$ is assumed to be unknown and the copy machine is required to work equally well for every value of $t$. In order to produce copies without requiring energy from the outside, the machine has to process the $N$ input clocks jointly with a state of $M-N$ ``blank clocks", which provide no information about time, but possess sufficient energy to enable the desired transition. Indeed, in order to approximate $M$ perfect copies of the state $|\psi_t\>$ the machine should at least be able to produce output states that have energy close to $M \< \psi| H|\psi\>$, meaning that the blank clocks should have energy close to $(M-N) \< \psi| H|\psi\>$. The problem of energy-preserving cloning of clock states is equivalent to the problem of cloning coherence: denoting by $ |\beta\> $ the blank state, the cloning machine attempts at converting the state $ |\psi\>^{\otimes N} \otimes |\beta\>$ into the state $ |\psi\>^{\otimes M}$. Choosing the blank state to be an eigenstate of the energy, we have that maximizing the fidelity for the transition $ |\psi\>^{\otimes N} \otimes |\beta\> \xrightarrow {} |\psi\>^{\otimes M}$ under the energy-preserving restriction is equivalent to maximizing the fidelity of cloning for every instant of time.
In the following we analyze in detail the simplest example of energy-preserving cloning of quantum coherence: we consider $N$ two-level systems, each of them with Hamiltonian $H=\frac{\hbar\omega}{2} \, Z $ and initially prepared in the coherent superposition $|+\> = (|0\>+|1\>)/\sqrt{2}$.
Th question is how well one can produce $M > N$ approximate copies without paying an energy cost. For simplicity, we assume that the difference $M-N$ is even: under this assumption, we can choose the blank state to be an energy eigenstate with energy \emph{exactly} equal to zero. Specifically, we choose the state $|\beta\>= |M-N, 0\>$, belonging to the symmetric eigenbasis
\begin{align*}
|L, m\>:=\frac{ \sum_{\pi\in\set{S}_{L}} U_\pi|0\>^{\otimes (L+m)/2} \, \otimes \, |1\>^{\otimes (L-m)/2}}{ \sqrt{ L! \, [( L+m)/2]! \, [ (L-m)/2]!}} \, ,
\end{align*}
where $\set{S}_{L}$ denotes the group of permutations of a $L$-element set and $U_{\pi}$ is the unitary that permutes $L$ Hilbert spaces according to the permutation $\pi$.
We now apply our recursive protocol, producing at each step an approximation of the desired $M$-copy state. Let us expand the states $ |\psi\>^{\otimes N} $ and $|\psi\>^{\otimes M}$ as
$$|\psi\>^{\otimes L}=2^{-L/2}\sum_{m=-L}^{L}\sqrt{{L\choose \frac{L-m}2}} \, |L,m\> \qquad L = M,N \, , $$
then use the formulas for the fidelity and success probability derived in Section~\ref{sec:network}.
At the first step of the protocol, the successful quantum operation produces an output state with the maximum possible fidelity, given by
$$F^{(1)}_{\max}=\frac{1}{2^M}\sum_{n=-N}^{N}{M\choose \frac{M-n}{2}}.$$
The above fidelity turns out to be equal to the absolute maximum of the fidelity achievable over all covariant quantum operations, derived by Fiura\`sek in Ref. \cite{fiurasekclon}. For large $N$, the fidelity is close to 1 whenever $M$ is small compared to $N^2$, thus allowing one to achieve quantum super-replication \cite{GiulioNat}. It is well known that the price of super-replication is a probability of success vanishing exponentially fast with $N$ \cite{GiulioNat}. The main interest of our recursive protocol lies in the fact that it allows us to increase the probability of success. In a protocol with $K>1$ steps, the average fidelity decreases at each step, while the probability of success increases. The tradeoff between the fidelity and the probability of success is illustrated in Figure~\ref{fig:qubit} for the case of $N= 80$, $M = 400$ and $K= 32$, using Eqs. (\ref{fidelity1}) and (\ref{prob1}). In addition, we compare the fidelity of the recursive protocol with that of its coherent coarse-graining, given by Eq. (\ref{F'}). As already observed, the coherent coarse-graining achieves a higher fidelity, while keeping the same success probability. In the figure we also plot the optimal fidelity in the deterministic case (black dashed line in Figure~\ref{fig:qubit}). The deterministic fidelity (derived in theorem \ref{theo:EAoptimal}) coincides with the fidelity for phase-covariant cloning \cite{qubit}, meaning that the optimal cloner can be realized in an energy-preserving fashion.
Figure \ref{fig:qubit} well illustrates the advantages of the recursive protocol. At the first round the fidelity is very high, but the success probability has the extremely tiny value $p_{\rm succ}^{(1)} = 6\times10^{-20}$. The subsequent rounds of the protocol increase the success probability dramatically, reaching a probability of approximately $23\%$ at the 31-th step. The fidelity for the recursive protocol remains higher than the optimal deterministic fidelity up to almost the very last step. An even better performance is attained through coherent coarse-graining.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{Qubits.pdf}\\
\caption{{\bf Energy-preserving cloning of quantum clocks via the recursive protocol and its coherent coarse-graining.} The figure shows the tradeoff between success probability and fidelity for the $N$-to-$M$ cloning of the clock state $|\psi_t\> = (e^{i\omega t/2}|0\>+e^{-i\omega t/2}|1\>)/\sqrt{2}$ with $M=400$ and $N=80$. The green solid line (with numerics represented by red dots) shows the probability-fidelity tradeoff for a recursive protocol with $K= 32$ rounds. The blue solid line (with numerics represented by the black dots) shows the tradeoff for filters generated by coherent coarse-graining, with the $T$-th point corresponding to the coherent coarse-graining of the first $T$ steps of the recursive protocol. Finally, the black dashed line represents the fidelity for the optimal deterministic cloning protocol. Notice that the recursive protocol maintains fidelity larger than the optimal deterministic fidelity for all steps up to the last. }
\label{fig:qubit}
\end{figure}
\subsection{Passive probabilistic amplification of coherent light}\label{subsec:coherent}
In quantum optics the energy-preserving instruments are those that preserve the average photon number. In the single-mode scenario, the number observable is non-degenerate and the energy-preserving quantum operations have diagonal Kraus operators in the Fock basis $\{|n\>\}$. In the following we consider the application of the recursive protocol to the amplification of the coherent state of light
\begin{align*}
&|r_1 \>\xrightarrow{ \quad }|r_2 \> \qquad 0\le r_1\le r_2 \, .
\end{align*}
Note that, since we require the amplification map to be part of a number-preserving quantum instrument, our protocol defines a phase-insensitive amplifier \cite{insensitive}, which works equally well for the transition
\begin{align*}
&|r_1 e^{i\theta} \>\xrightarrow{ \quad }|r_2 e^{i\theta} \> \qquad 0\le r_1 \le r_2 \, ,
\end{align*}
where $\theta$ is an arbitrary angle.
Amplifying a coherent state without increasing its photon number seems to be a daunting task. However, the fact that the number is preserved only on average grants us the opportunity to reach high fidelity in a probabilistic fashion. In the case of amplifiers, the tradeoff between success probability and fidelity is essentially a tradeoff between success probability and photon number modulation.
Since the Hilbert space is infinite dimensional, our recursive protocol cannot be applied directly. To overcome the obstacle, we define a threshold $N$ and assume that the successful operations project the input state inside the subspace spanned by Fock states with number smaller than $N$. Practically, for $N \gg r_2^2$, the projection can be done without affecting the fidelity. The fidelity at the $k$-th round, given by Eq. (\ref{fidelity1}), can be lower bounded as
\begin{align}\label{coherentfid}
F^{(k)}_{\max}\ge1-e^{-r_2^2}\left(\frac{r_2^2e}{N-k+1}\right)^{N-k+1}
\end{align}
when $N-k+1>r_2^2$.
On the other hand, the probability of success in Eq. (\ref{prob1}) can be expressed as
\begin{align}\label{coherentprob}
p_{\rm succ}^{(k)}=\left\{\begin{matrix} e^{r_2^2-r_1^2}\left(\frac{r_1}{r_2}\right)^{2N} \, F^{(1)}_{\max}& \quad k=1\\
\\
e^{r_2^2-r_1^2}\left(\frac{r_1}{r_2}\right)^{2N-2k+2}\left[1-\left(\frac{r_1}{r_2}\right)^2\right] \, F^{(k)}_{\max}&\quad k>1 \end{matrix}\right.
\end{align}
\iffalse
Finally the total probability of success, derived by plugging Eq. (\ref{coherentprob}) into Eq. (\ref{Ptot}), reads
\begin{align}
p_{\rm succ}=p_{\rm succ}^{(1)}\cdot\left(\frac{r_2}{r_1}\right)^{2(K-1)} \, .
\end{align}
It shows that the recursive protocol increases the probability of success exponentially while the average fidelity $F$ remains almost unity, when the number of rounds $K$ is not large.
\fi
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{Coherent.pdf}\\
\caption{{\bf Energy-preserving amplification of coherent light via the recursive protocol and its coherent coarse graining.} The figure shows the tradeoff between success probability and the average fidelity for the amplification $|r_1 \>\to|r_2 \> $ with $r_1=1$, $r_2=1.5$ and $N=80$. The green solid line (with numerics represented by red dots) shows the probability-fidelity tradeoff for a recursive protocol with $K=81$ rounds. The blue solid line (with numerics represented by the black dots) shows the tradeoff for filters generated by coherent coarse-graining. Note that the difference between the two curves becomes large as the success probability tends to one. For unit probability the recursive protocol and its coherent coarse-graining give fidelities $F_{\rm det} = 49.9\%$ and $F_{ {\rm det}}' = |\<r_1|r_2\>|^2 =77.9\%$, respectively. }
\label{fig:amplify}
\end{figure}
Interestingly, the successful quantum operation at the first round of our protocol ($k=1$) coincides with the optimal probabilistic amplifier for coherent states with known amplitude \cite{fiurasekclon,caves}, which indeed can be implemented with energy-preserving operations.
Specifically, evaluating Eqs. (\ref{coherentfid}) and (\ref{coherentprob}) for $k=1$ one retrieves the expressions for the optimal fidelity and success probability appearing in Eqs. (6.24) and (6.36) of Ref. \cite{caves}
For the subsequent rounds of the recursive protocol ($k>1$), the input state is not coherent anymore and the successful quantum operation differs from the optimal coherent-state amplifier.
In Figure \ref{fig:amplify} we show the performance the recursive protocol and its coherent coarse-graining for the amplification of coherent states from $r_1=1$ to $r_2=1.5$. The threshold in the Fock space is chosen to be $N=80$ and the protocol is applied recursively for $K=81$ rounds. From the plot it can be seen that the filters generated by coherent coarse-graining reach a relatively high fidelity, compatibly with the strong constraint set by the number-preserving condition. For instance, the coherent coarse-graining of the recursive protocol with $K=80$ succeeds with probability $p_{\rm succ} =79.6\%$ and reaches fidelity $F= 83.9\%$. When the probability reaches 1, the fidelities of the recursive protocol and its coherent coarse graining become $F_{\rm det} = 49.9\%$ and $F_{ {\rm det}}' = |\<r_1|r_2\>|^2 =77.9\%$, respectively. The latter is well above the fidelity of the optimal amplifier for \emph{arbitrary} coherent states, which is given by $F_{\rm universal}= 4/9 $ \cite{namikirapid,chiribella-xie}.
\subsection{Energy-preserving correction in ancilla-driven quantum computation}\label{sec:unlearn}
In ancilla-driven quantum computation \cite{ADQC} the evolution of the system is determined by the outcomes of measurements on the ancilla. Ideally, the goal is to implement measurements that induce unitary gates on the system. To achieve this goal, the measurements should not not extract any information about the state of the system: the probability of each outcome should be the probability that a particular unitary gate is applied to the system \cite{LostFound}. However, in many non-ideal situations the measurement extracts some information, thus inducing a non-unitary evolution on the system. When this is the case, one can attempt to correct the unwanted non-unitarity by performing additional measurements.
This type of correction has been studied in Refs. \cite{tradeoff,unlearn},
where a number of different strategies have been proposed.
Here we consider the problem in the energy-preserving setting: suppose that a quantum system with $d$ non-degenerate energy levels interacts with an ancilla via an energy-preserving unitary evolution. Then, the ancilla undergoes the measurement of an observable that is compatible with the energy. As a result, the system evolves randomly according to an energy-preserving instrument $\{\map M_x\}_{x\in\set X}$. We assume that the measurement on the ancilla is a rank-one projective measurement and, therefore, the quantum operations $\{\map M_x\}$ are pure. For every given $x \in\set X$, the problem is to correct the quantum operation $\map M_x$, making it as close as possible to a desired energy-preserving unitary gate $U_x$. As a correction we allow ourselves to use an energy-preserving filter, with quantum operations $\left\{\map N^{(x)}_{\rm succ}, \map N^{(x)}_{\rm fail}\right\}$. Due to the presence of the filter, an initial pure state $|\eta\>$ is transformed probabilistically into the pure state
$$|\eta_x\> = \frac{ N^{(x)}_{\rm succ} \, M_x \, |\eta\> }{\| N^{(x)}_{\rm succ} \, M_x \, |\eta\> \|} \, .$$
To evaluate the quality of the correction, we consider the fidelity between $|\eta_x\>$ and the target state $ U_x |\eta\>$, averaging over all possible pure input states.
Assuming that initially the state $|\eta\>$ is drawn at random according to the Haar measure, the conditional probability distribution over the pure states is given by
$$ p(\eta| x, {\rm succ} ) \, {\rm d} \eta = \lambda_x \, \left\| N^{(x)}_{\rm succ} \, M_x \, |\eta\> \right\|^2 \, {\rm d} \eta \, ,
$$
where $M_x$ and $ N^{(x)}_{\rm succ}$ are the Kraus operators of $\map M_x$ and $\map N^{(x)}_{\rm succ}$, respectively, and $\lambda_x $ is the normalization constant $\lambda_x: = \left( \int \| N^{(x)}_{\rm succ} \, M_x \, |\eta'\> \|^2 \, {\rm d} \eta' \right)^{-1}$.
Hence, the average fidelity over all pure states is given by
\begin{align}\label{Fx}
\nonumber F_x & : = \int {\rm d} \eta \, p(\eta| x, {\rm succ}) \, \left | \< \eta_x| U_x |\eta\> \right|^2 \\
\nonumber & = \frac{ \int {\rm d} \eta \, | \< \eta| U_x^\dag N^{(x)}_{\rm succ} M_x |\eta\> |^2 }{ \int{\rm d} \eta' \| N^{(x)}_{\rm succ} \, M_x \, |\eta'\> \|^2 } \\
& = \frac{ F_{0}^{(x)} \cdot d+1 }{ d +1 }
\end{align}
where $F_{0}^{(x)}$ is the fidelity given by
$$ F_{0}^{(x)} = \frac{ \left | \< e_0 | \, U_x^\dag N^{(x)}_{\rm succ} M_x \,|e_0 \>\right|^2}{\| N^{(x)}_{\rm succ} M_x |e_0\> \|^2} \qquad |e_0\> = \frac{ \sum_{n=1}^d |n\>}{\sqrt d}\,. $$
Maximizing the average fidelity is then equivalent to finding the optimal quantum operation for the transformation
$$ |\varphi_x\> := \frac{ M_x |e_0 \>}{ \| M_x |e_0 \> \|} ~ \xrightarrow{\quad} ~ |\psi_x\> : = U_x |e_0\> \, . $$
The maximization under the energy-preserving constraint is exactly the problem solved in this paper. In particular, for every outcome $x$ we can use our recursive protocol to obtain a high-fidelity approximation of the desired transformation. In this context, it is immediate to realize that our protocol provides an approximate correction strategy, with the property that the overall quantum operation acts exactly like the target gate $U_x$ in a suitable subspace, whose dimension shrinks at every step.
For concreteness, let us see explicitly how the protocol works in a concrete example. We choose the quantum operation $\map M_x$ with Kraus operator $ M_x = \sum_{n=1}^d \, \mu^{n/2} \, |n\>\<n|$.
The fidelity at the $k$-th step is given by
$$ F^{(k)}_x = \frac{d+2 - k}{d+1} \, ,$$
while the probability of success, averaged over all pure states, is given by
$$ p_{\rm succ}^{(k)} =\left\{\begin{array}{ll} \, \mu^{d-k}(1-\mu)^2(d+1-k)/(1-\mu^d)&\quad k>1 \, \\ \\
\mu^{d-1}(1-\mu)d/(1-\mu^d)&\quad k=1\,.\end{array}\right.
$$
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{Unlearning.pdf}\\
\caption{{\bf Energy-preserving correction of a quantum operation via the recursive protocol and its coherent coarse graining.} The figure shows the tradeoff between success probability and average fidelity for unlearning the quantum operation with Kraus operator $ M_x = \sum_{n=1}^d \, \mu^{n/2} \, |n\>\<n|$ with $d=100$ and $\mu=0.9$. The green solid line (with numerics represented by red dots) shows the probability-fidelity tradeoff for a recursive protocol with $K= 70$ rounds. The blue solid line (with numerics represented by the black dots) shows the tradeoff for filters generated by coherent coarse-graining.}
\label{fig:unlearn}
\end{figure}
The features of the recursive protocol and of its coherent coarse-graning are illustrated in figure \ref{fig:unlearn} for $d=100$, $\mu=0.9$ and $K=70$. The probability of success increases from a very small value ($p_{\rm succ}^{(1)} = 3\times10^{-4}$) to approximately $14\%$ at the 68-th step, at the cost of a reduced fidelity.
\section{Extension to mixed states} \label{sec:mixed}
So far we considered transitions between pure states.
However, for many practical applications it is important to consider transformations where the input states are mixed,~e.~g.~due to the presence of decoherence. Normally the target state is still pure, since ideally one would like to remove the noise from the output. This is the case for tasks like mixed state purification \cite{cirac-ekert-macchiavello,andersen}, super-broadcasting \cite{dariano-macchiavello-perinotti-prl}, and for the evaluation of the corresponding quantum benchmarks \cite{polzik-NJP}. In all these cases, our techniques provide general upper bounds on the fidelity of the optimal energy-preserving operations. The bounds are achievable for a quite large class of states, which includes all the thermal states associated to stoquastic Hamiltonians \cite{bravyi2008thecomplexity,bravyi2009complexity}, such as the Hamiltonians of flux qubits in Josephson junctions \cite{burkard2004multilevel} and the Hamiltonians of (bosonic) Bose-Einstein condensates \cite{ceperley1995path}.
\subsection{Deterministic transitions and eigenstate alignment}
Given a generic mixed state $\rho$, we search for the best energy-preserving approximation of the transition $\rho \to |\psi\>\<\psi|$, where $|\psi\>$ is a target pure state. To this purpose, it is convenient to decompose the initial state into blocks corresponding to the different energy eigenvalues, namely
\begin{align}\label{blockrho}
\rho = \sum_{E,E'} \, \rho_{E,E'} \qquad \rho_{E,E'} = P_E \rho P_{E'} \, .
\end{align}
With this notation, we have the following
\begin{theo}\label{theo:EAoptimal_mixed}
For $p_{\rm succ} =1$, every energy-preserving approximation of the transition $\rho \to |\psi\>\<\psi|$ satisfies the bound
\begin{align}\label{mixedbound}
F_{\rm det}\le \sum_{E,E' \in \mathsf{Sp}(\rho)}\sqrt{q_Eq_{E'}} \, \| \rho_{E,E'} \|_1 \, ,
\end{align}
where $ \| A \|_1 = \operatorname{Tr} [ \sqrt{ A^\dag A} ]$ is the trace norm, $q_E = \| P_E |\psi \>\|^2 $ and $\mathsf{Sp}(\rho)$ is the \emph{energy spectrum of $\rho$}, defined as
\begin{align}
\mathsf{Sp} (\rho) = \{ E~|~ P_E \rho P_E \not = 0 \, \}.
\end{align}
The bound is achievable if the input state $\rho$ is \emph{block positive},~that is, if there exist orthonormal bases for the energy eigenspaces such that, when the matrix elements are taken in those bases, each matrix $[\rho_{E,E'}]$ is positive semidefinite.
In this case, every quantum channel $\map A$ satisfying the condition
\begin{align}\label{EA_mixed}
\map A (\rho) = \sum_{E,E'} \, \|\rho_{E,E'}\|_1 \, |\psi_E\>\<\psi_{E'}| \qquad \forall E,E' \in \mathsf{Sp}(\rho)
\end{align}
is optimal.
\end{theo}
Note that, when the input state is pure, the bound (\ref{mixedbound}) coincides with the optimal fidelity.
We refer to every channel $\map A$ satisfying Eq. (\ref{EA_mixed}) as an \emph{eigenstate alignment of the mixed state $\rho$ to the pure state $|\psi\>$}. Note that eigenstate alignment may not be a unitary operation anymore, because it may have to send different eigenstates with energy $E$ to the fixed eigenstate $|\psi_E\>$.
The proof of theorem \ref{theo:EAoptimal_mixed} is provided in Appendix \ref{app:optprob_mixed}. Three applications of the theorem are as follows:
\begin{eg}[Non-degenerate Hamiltonians] When the Hamiltonian $H$ is non-degenerate, the bound (\ref{mixedbound}) becomes
\begin{align}
F\le \sum_{E,E'} \, \sqrt{ q_E q_{E'}} \, | \< \varphi_E | \rho | \varphi_{E'} \> | \, ,
\end{align}
where $\{|\varphi_E\>\}$ is the energy eigenbasis. Note that, since the Hamiltonian is non-degenerate, the choice of eigenbasis is unique up to phase transformations $|\varphi_E\> \mapsto |\varphi_E'\> = e^{i\theta_E} \, |\varphi_E\>$.
The bound is achievable if, for a suitable choice of phases, one has
\begin{align}\label{pureinphase}
\< \varphi_E |\rho_E |\varphi_{E'} \> \ge 0 \qquad \forall E, E'
\end{align}
Mixed states of this form were called \emph{pure in phase} by D'Ariano \emph{et al} \cite{d2000isotropic}, who considered them in the context of phase estimation. Such states play an important role in the area of quantum Hamiltonian complexity, where they arise as thermal states of \emph{stoquastic Hamiltonians} \cite{bravyi2008thecomplexity,bravyi2009complexity},~i.~e. Hamiltonians with non-positive matrix elements in a given basis.
Physically, we can consider a scenario where the system starts in the thermal state of a stoquastic Hamilonian and subsequently undergoes a rapid change of Hamiltonian to a diagonal one, making the initial thermal state a non-trivial resource.
\end{eg}
\begin{eg}[Pure states subject to random time evolution]
Suppose that the system, initially prepared in a pure state $|\varphi\>$, has evolved under its free Hamiltonian for a time $t$ which is not perfectly known,~e.~g.~due to the finite time resolution of the clocks available in the laboratory.
Then, the system is effectively described by the mixed state
\[ \rho = \int {\rm d} t \, \pi(t) \, U_t |\varphi\>\<\varphi| U_{t}^\dag \, , \qquad U_\tau = e^{-i t H_{\rm sys}/\hbar }\, , \]
$\pi(t)$ being the probability distribution for $t$.
In this case, the bound (\ref{mixedbound}) becomes
\[ F \le \sum_{E,E'} \, \sqrt{q_E q_{E'} p_E p_{E'}} \, \left| \widetilde{\pi} (E-E')\right| \, , \]
where $p_E$ and $q_E$ are defined as in Eq. (\ref{statedecomp}) and $\widetilde {\pi}$ is the Fourier transform of $\pi$.
If the Fourier transform is positive (i.~e.~if the noise has positive spectrum), then the bound is attainable by every unitary operation $U$ satisfying the eigenstate alignment condition [Eq. (\ref{eigenalign})].
\end{eg}
\begin{eg}[Multiple copies of qubit mixed states]\label{eg:mixed}
Consider a system of $N$ non-interacting qubits, each having the same Hamiltonian $H = E_0 \, Z/2$. Then, the total Hamiltonian of the system is degenerate and has the block diagonal form
\[ H_{\rm sys} = E_0 \, \bigoplus_{l} \, \left( J_z^{(l)} \otimes I_{d^{(N)}_l}\right) \, , \]
where $l$ is the quantum number of the total angular momentum, $J^{(l)}_z$ is the $z$ component of the angular momentum operator in the subspace with quantum number $l$, and $I_{d^{(N)}_l}$ is the identity on a multiplicity space $\spc M_l^{(N)}$, of dimension
$$d_l^{(N)}=\frac{4l+2}{N+2l+2}{N\choose N/2+l}$$ (see e.g. Ref. \cite{ChiribellaDAriano05}).
From the above decomposition it is clear that the eigenvalues of the energy are given by $E_m = E_0 m$, where $m$ are the eigenvalues of the $z$-component of the total angular momentum operator. A basis for the corresponding eigenspace is given by the vectors
\[|\varphi_{m,l,n}\> = |l,m\> \otimes | \mu_{l,n} \> \, ,\]
where $l$ goes from $|m|$ to $N/2$, $|l,m\>$ is the eigenstate of $J^{(l)}_z$ with eigenvalue $m$, and $\{|\mu_{l,n}\>~|~ n= 1, \dots , d_l^{(N)}\}$ is a basis for the multiplicity space.
Now, suppose that each qubit is initially prepared in the state
\[\omega = \frac{e^{\beta \, X}}{\operatorname{Tr}[e^{\beta\, X}]} \, , \qquad \beta \ge 0 \, , \quad X = \begin{pmatrix} 0&1\\
1&0 \end{pmatrix} \, . \]
With this choice, the state $\omega^{\otimes N}$ satisfies the condition for the achievability of the bound (\ref{mixedbound}): indeed, one has
\begin{align}
\nonumber
P_{E_m} \omega^{\otimes N} P_{E_{m'}} & = \bigoplus_l \, \frac{\< l,m | e^{2\beta J_x^{(l)}} |l,m'\> }{{\operatorname{Tr}[e^{\beta X}]^{N}}} \\
& \qquad \qquad \, \times \left( |l,m\>\<l,m'| \otimes I_{d_l^{(N)}}\right) \, ,
\label{block}
\end{align}
and $ \< l,m | e^{2\beta J_x^{(l)}} |l,m'\>\ge 0$ since the matrix $J_x^{(l)}$ has positive matrix elements.
Hence, the matrix elements of the operator $P_{E_m} \rho P_{E_{m'}} $ in the basis $\{ |l,m\> \otimes |\mu_{l,n}\> ~|~ l \ge \max\{ |m|, |m'|\} \, , n = 1, \dots d_l^{(N)} \}$ form a non-negative matrix.
For the transition $\omega^{\otimes N} \to |\psi\>\<\psi|$, eigenstate alignment is not a unitary operation, because all basis vectors with energy $E_m$ are transformed into $|\psi_{E_m\>}$.
\end{eg}
\subsection{The ultimate probabilistic fidelity}
We now provide the exact value of the maximum fidelity for the transition $\rho \to |\psi\>\<\psi|$ when no restriction is imposed on the probability of success.
\begin{theo}\label{prop:optprob_mixed}
For a finite dimensional Hilbert space, the maximum fidelity over all energy-preserving operations is given by
\begin{align}
\nonumber
&F_{\max} = \left \| A \right\|_{\infty} \, , \\
&A = \sum_{E,E'} \, \sqrt{q_E q_{E'}} \, |\psi_E\>\<\psi_{E'}| \otimes \left( \rho_{E,E}^{T}\right)^{-\frac 12} \rho_{E,E'}^T \left( \rho_{E',E'}^{T}\right)^{-\frac 12} \, ,
\label{mixedFmax}
\end{align}
where $ \| A \|_\infty = \max_{ \| |\psi\> \|=1} \, \| A |\psi\> \|$ denotes the operator norm and $\rho^T$ denotes the transpose of $\rho$.
For a quantum operation achieving fidelity $F_{\max}$, the maximum probability of success is equal to
\begin{align}\label{maxprobmixed}
p^{\max}_{\rm succ} = \max_{\sigma} \min_E \frac 1 { \left \| \left( \rho_{E,E}^T\right)^{-1/2} \sigma_E \left(\rho_{E,E}^T\right)^{-1/2} \right\|_\infty \ } \, ,
\end{align}
where the maximum $\max_\sigma$ runs over all density matrices $\sigma$ with support contained in the eigenspace of $A$ with maximum eigenvalue and $\sigma_E : = \operatorname{Tr}_{\rm 1} [ (P_E \otimes I) \sigma ]$, $\operatorname{Tr}_1$ denoting the partial trace over the first Hilbert space.
\iffalse
For block positive states, a simple lower bound is
\begin{align}\label{gammalow}
F_{\max} \ge \gamma_{\max}
\end{align}
where $\gamma_{\max}$ is the maximum eigenvalue of the matrix $\Gamma$ defined by
\[ \Gamma_{E,E'} = \frac{ \| \rho_{E,E'} \|_1}{ \sqrt{ \operatorname{Tr} [ \rho_{E,E} ] \, \operatorname{Tr} [ \rho_{E',E'}]} } \, \sqrt{q_E q_{E'}} \, , \qquad E,E' \in \mathsf{Sp}(\rho) \, . \]
The maximum probability of a quantum operation achieving the bound is
\begin{align}
p_{\max} = \min_{E \in \mathsf{Sp}(\rho)} \, \frac{ \| \rho_{EE} \|_1}{ v_E^2} \, ,
\end{align}
where $\bs v = (v_E)_{E\in\mathsf{Sp}(\rho)}$ is the normalized eigenvector of $\Gamma$ corresponding to the maximum eigenvalue.
\fi
\end{theo}
The proof, provided in appendix \ref{app:optprob_mixed}, includes the explicit construction of the optimal strategy and an expression for the maximum probability of success.
\subsection{Purification of coherence at zero energy cost}
We now illustrate the application of our techniques to the concrete problem of purifying a mixed state \cite{cirac-ekert-macchiavello,andersen}. Suppose that we are given $N$ identical quantum systems, each prepared in the same mixed state, which happens to possess a non-zero amount of coherence in across different energy levels. Can we collect the coherence present in the $N$ systems and concentrate it in a single system? And can we do it without drawing energy from the outside? Mathematically, the task is to implement the transition
\[ \omega^{\otimes N} \to |\phi\>\<\phi| \otimes |\beta\>\<\beta| \, ,\]
where $\omega$ is the initial mixed state, $|\phi\> = \sum_{n=1}^d \, |n\> /\sqrt d$ is the maximally coherent state, and $|\beta\> $ is an eigenstate of the energy, in which $N-1$ systems are meant to be left. In the following we discuss the qubit case ($d=2$) and we choose the mixed state to be
\[\omega=\frac{e^{\beta X}}{\operatorname{Tr}\left[e^{\beta X}\right]} \, . \]
This state can be thought as the thermal state of the initial Hamiltonian $ H_{\rm in}= - X$ and represents a non-trivial resource if the Hamiltonian is suddenly changed into $H = Z$.
For simplicity, we choose $N$ to be odd, so that $|\beta\>$ can be chosen to be an eigenstate with zero energy.
Let us consider first the deterministic transitions. The performance of the optimal energy-preserving channel is determined by theorem \ref{theo:EAoptimal_mixed}, which leads to the
expression
\begin{align*}
F_{\rm det} = \frac 1 2 \sum_{m,m' =\pm \frac 12 }\, \sum_{l =\frac 12}^{\frac N2} \, \frac{ d_l^{(N)} \, \< l,m | e^{2\beta J_x^{(l)}} |l,m'\> }{{\operatorname{Tr}[e^{\beta X}]^{N}}} \, .
\end{align*}
Here the r.h.s. follows from Eqs. (\ref{mixedbound}) and (\ref{block}), using the fact that $\omega^{\otimes N}$ is block positive, as observed in Example \ref{eg:mixed}. The optimal deterministic fidelity is plotted in Figure \ref{fig:Fmixed} for various values of $N$ and $\beta$.
Note that, quite counterintuitively, the deterministic fidelity \emph{decreases} with the growth of $N$. The origin of this behaviour can be found in the constraint of energy preservation. Essentially, a deterministic energy preserving operation cannot do anything better than realigning the blocks corresponding the values $m = \pm 1/2$ to the corresponding eigenstates. However, when $N$ grows, the blocks are spread over an increasing number of values of $m$, so that the weight of the $m= \pm 1/2$ component becomes less and less significant. As a result, the deterministic fidelity vanishes in the limit $N\to \infty$. While the state $\omega^{\otimes N}$ contains an increasing amount of coherence, collecting this coherence from the high-energy blocks requires an exchange of energy with the surrounding environment.
For probabilistic strategies, the situation is different: the limitation due to energy-preservation can be partially lifted and the fidelity approaches unit as $N$ increases. To evaluate the maximum fidelity, we have to compute the norm of the operator $A$ defined in Eq. (\ref{mixedFmax}). In the case at hand, we have
\begin{align*} A & = \frac 12 \, \sum_{m,m' = \pm \frac 12} \sum_{l=\frac 12}^{\frac N2}\frac{\<l,m| e^{2\beta J_x^{(l)}} |l,m'\>}{\sqrt {\<l,m| e^{2\beta J_x^{(l)}} |l,m\> \<l,m'| e^{2\beta J_x^{(l)}} |l,m'\> } } \\
& \qquad \qquad \qquad \, \left|\frac 12, m\right\>\left\<\frac 12, m' \right| \otimes |l,m\>\<l,m'| \otimes I_{d_l^{(N)}} \, .
\end{align*}
Taking the operator norm, we then obtain \[ F_{\rm prob} = \max_{l \in \left\{ \frac 12\, , \dots \, , \frac N 2 \right\} } \frac{ 1+ a_l}2 \, , \]
with
\[ \qquad a_l = \frac{\<l,\frac 12| e^{2\beta J_x^{(l)}} |l,-\frac 12\>}{\sqrt {\<l,\frac 12| e^{2\beta J_x^{(l)}} |l,\frac 12\> \<l,-\frac 12| e^{2\beta J_x^{(l)}} |l,-\frac 12\> } } \]
The optimal probabilistic fidelity is plotted in Figure \ref{fig:Fmixed} for different values of $N$ and $\beta$.
\begin{figure}[h!]
\centering
\subfigure[]
{\label{fig:Fmixed}
\includegraphics[width=0.9\linewidth]{Fmixed.pdf}
}
\subfigure[]{\label{fig:Pmixed}
\includegraphics[width=0.9\linewidth]{Pmixed.pdf}
}
\caption{{\bf Optimal purification of coherence via energy-preserving operations.} Figure \ref{fig:Fmixed} shows the optimal fidelities as function of the number of input copies $N$. Here different colors represent different values of $\beta$---specifically, $\beta=0.4$ (blue line with purple dots), $\beta=0.8$ (black line with black dots) and $\beta=1.2$ (green line with red dots). The dashed lines represent the deterministic fidelities, while the solid lines represent the probabilistic fidelities. While the probabilistic fidelities increase with $N$, the deterministic fidelities decrease, due to the restriction imposed by energy preservation.
Figure \ref{fig:Pmixed} shows the maximum success probability for the quantum operations with maximum fidelity. The color code is the same as in Figure \ref{fig:Fmixed}.
}
\label{fig:mixed}
\end{figure}
Finally, theorem \ref{prop:optprob_mixed} allows us to evaluate the maximum probability of success for the quantum operations that achieve maximum fidelity. According to the theorem, it is enough to characterize the density matrices that have support inside the eigenspace of $A$ with maximum eigenvalue. Such density matrices have the form $\sigma = |\Phi_{N/2}\>\< \Phi_{N/2}| \otimes \tau$, where $|\Phi_{N/2}\>$ is the maximally entangled state
\[|\Phi_{N/2}\> = \left( \left|\frac12, \frac12\right\> \left | \frac N 2,\frac12\right\> + \left| \frac 12,-\frac 12 \right\> \left | \frac N2, -\frac 12\right\>\right)/\sqrt 2 \, .\]
Hence, we have the relation
\begin{align*} \sigma_m & = \operatorname{Tr}_1\left [ \left( \left| \frac 12, m \right \> \left\< \frac 12 , m \right| \otimes I \right) \sigma\right] \\
& = \frac 12 \, \left|\frac N 2,m\right\>\left\<\frac N2,m\right| \otimes \tau
\end{align*}
which can be inserted into Eq. (\ref{maxprobmixed}), yielding
\begin{align*}
p_{\rm succ}^{\max}
& = \max_\tau \min_{m= \pm \frac 12} \frac {2 \left\< l, m \right | e^{2 \beta J_x^{(l)} } \, \left| l,m \right\> }{ \operatorname{Tr} [ e^{\beta X}]^N \, \| \tau \|_\infty } \\
& = \frac {2 d_l^{(N)} \left\< l, -\frac 12 \right | e^{2 \beta J_x^{(l)} } \, \left| l,-\frac 12 \right\> }{ \operatorname{Tr} [ e^{\beta X}]^N } \, .
\end{align*}
A plot of the probability of success as a function of $N$ and $\beta$ is shown in Figure \ref{fig:Pmixed}.
\section{Conclusions}\label{sec:conclusion}
We introduced the class of quantum operations that can in principle be implemented at zero energy cost. Within this class, we addressed the search for the optimal operations implementing a desired state transition. We considered operations that can generally be probabilistic, showing that the limitations arising from the preservation of the energy can be lifted to a surprising extent at the price of a reduced probability of success.
Our investigation started from the problem of transforming given input state into a desired output state. To solve this problem, we developed two general techniques, dubbed eigenstate alignment and L\"uders reduction.
Eigenstate alignment provides the best deterministic way to transform pure states at zero energy cost.
L\"uders reduction applies more generally to the optimization of energy-preserving quantum operations. Essentially, it allows one to break down every quantum operation into the product of a pure probabilistic part followed by a deterministic part.
Employing these techniques, we reduced the search for best energy-preserving transformations of pure states to a simple Lagrangian optimization.
The characterization of the optimal energy-preserving transformations of pure states allowed us to construct a multi-round recursive protocol that achieves maximum fidelity with maximum success probability in each round.
The probability of success of the protocol increases from one round to the next and, for a system with finite energy spectrum, the protocol terminates deterministically in a finite number of steps.
Our protocol can be easily applied to every desired transformation of pure states, allowing one to find lower bounds to the optimal tradeoff curve between fidelity and probability, whose exact expression is known only in few cases.
As an illustration of the versatility of the protocol, we applied it to a number of concrete tasks, including quantum phase estimation, cloning of coherence, energy-preserving amplification, and ancilla-driven computation at zero energy cost.
To further improve the bounds on the fidelity-probability tradeoff, we applied the techniques of L\"uders reduction and eigenstate alignment to the different histories resulting from subsequent rounds of the recursive protocol. Specifically, we introduced an operation of coherent coarse-graining, whereby a set of quantum operations are joined into a pure quantum operation, with the same probability of occurrence of the original set and, typically, with a higher fidelity with the target.
Remarkably, when applied to the problem of phase estimation with maximally coherent states, coherent coarse-graining yields points that lie exactly on the optimal tradeoff curve, provided that the number of energy levels is sufficiently large. Characterizing all the situations in which our method provides the optimal tradeoff curve is an open problem. In general, we expect optimality to be achieved asymptotically in scenarios where the energy distribution of the state is sufficiently regular, including~e.~g.~the cases of
phase estimation and quantum cloning in the asymptotic regime \cite{ourNJP}.
\iffalse
A range of applications of the energy-preserving recursive protocol have been shown in Section \ref{sec:examples}. As one of the nontrivial applications, energy-preserving quantum cloning was defined and illustrated with a concrete example of cloning two-level systems. The filters adapted not only modify the amplitude distribution of the input state, but also perform the energy-preserving eigenbasis transformation $|N, m\>|M-N, 0\>\to |M, m\>$ for arbitrary $m$. We notice that this transformation coincides with the optimal phase covariant cloning channel of equatorial qubits \cite{qubit} asymptotically, thus we conclude that the optimal phase-covariant cloning can be achieved by energy-preserving instruments in this case. It is sufficient to consider energy-preserving protocols, which need not include extra energy into the system, when cloning some families of quantum states.
\fi
In this paper we provided a comprehensive study of the optimal quantum information processing under the energy-preserving constraint.
A related avenue of future research is the study of optimal quantum information processing under general conservation laws.
The techniques developed in this paper are already adapted to search of optimal quantum evolutions that preserve an \emph{algebra} of quantum observables, such as the algebra generated by the angular momentum operators.
Interactions that preserve the angular momentum have recently attracted attention in the implementation of quantum gates and quantum measurements \cite{ozawa,ozawaCNOT,ozawa2,spekkens-marvian-WAY, ahmadi}, although the characterization of the optimal operations is still an open problem. In this context, our result suggests a new strategy to approach the optimization, by considering probabilistic modulation of the amplitudes of the wave-function in sectors with different angular momentum. Also in this case, our results allow one to construct first a recursive protocol and to increase its fidelity through the operation of coherent-coarse graining. While such generalizations are beyond the scope of the present paper, it is our hope that our work will pave the way to a systematic optimization of quantum operations under arbitrary conservation laws.
\medskip
\acknowledgments{We thank the anonymous referees of this paper for a number of comments that stimulated presentation improvements and a the extension of the results to mixed states. This work is supported by the Foundational Questions Institute (FQXi-RFP3-1325), the National Natural Science Foundation of China (11450110096, 11350110207), and by the 1000 Youth Fellowship Program of China.
}
\bibliographystyle{apsrev}
|
\section{Introduction}
The theory of discrete-time dynamical systems studies qualitative and quantitative properties of orbits of points in a space under iterations of a given map. Various conditions can be imposed upon the map to simplify the orbit structures, which in turn lead to results about the dynamical system under consideration. One such well-known condition is expansiveness. Roughly speaking, a map is expansive if no two distinct orbits stay close forever. Expansiveness plays an important role in the exploitation of hyperbolicity in smooth dynamical systems, and in complex dynamics in particular (see for example, \cite{Ma87} and \cite{PU10}).
In the context of continuous maps on compact metric spaces, there are two weaker notions of expansion, called \emph{$h$-expansiveness} and \emph{asymptotic $h$-expansiveness}, introduced by R.~Bowen \cite{Bow72} and M.~Misiurewicz \cite{Mi73}, respectively. Forward-expansiveness implies $h$-expan\-siveness, which in turn implies asymptotic $h$-expansiveness \cite{Mi76}. Both of these weak notions of expansion play important roles in the study of smooth dynamical systems (see \cite{Bu11, DFPV12, DM09, DN05, LVY13}). Moreover, any smooth map on a compact Riemannian manifold is as\-ymp\-totically $h$-expansive \cite{Bu97}. Recently, N.-P.~Chung and G.~Zhang extended these concepts to the context of a continuous action of a countable discrete sofic group on a compact metric space \cite{CZ14}.
The dynamical systems that we study in this paper are induced by \emph{expanding Thurston maps}, which are a priori not differentiable. Thurston maps are branched covering maps on the sphere $S^2$ that generalize rational maps with finitely many postcritical points on the Riemann sphere. More precisely, a (non-homeomorphic) branched covering map $f\: S^2\rightarrow S^2$ is a Thurston map if it has finitely many critical points each of which is preperiodic. These maps arose in W.~P.~Thurston's characterization of postcritically-finite rational maps (see \cite{DH93}). See Section~\ref{sctThurstonMap} for a more detailed introduction to Thurston maps.
Inspired by the analogy to Cannon's conjecture in geometric group theory (see for example, \cite[Section~5 and Section~6]{Bon06}), M.~Bonk and D.~Meyer investigated extensively properties of expanding Thurston maps \cite{BM10}. (For a precise formulation of these analogies via the so-called Sullivan's dictionary, see \cite[Section~1]{HP09}.) Such maps share many common features of rational maps. For example, for each such map $f$, there are exactly $1+\deg f$ fixed points, counted with a natural weight induced by the local degree at each point, where $\deg f$ denotes the topological degree of the map $f$ \cite{Li13}; there exists a unique measure of maximal entropy (see for example, \cite{BM10}), with respect to which iterated preimages and periodic points are equidistributed in some appropriate sense (see for example, \cite{Li13}). More generally, for each \emph{potential} $\phi \: S^2 \rightarrow \R$ that is H\"older continuous with respect to some natural metric induced by $f$, there exists a unique equilibrium state, with respect to which iterated preimages are equidistributed \cite{Li14}.
P.~Ha\"issinsky and K.~Pilgrim investigated branched covering maps in a more general context \cite{HP09}. We will focus on expanding Thurston maps in this paper.
Let $(X,d)$ be a compact metric space, and $g\: X\rightarrow X$ a continuous map on $X$. Denote, for $\epsilon>0$ and $x\in X$,
$$
\Phi_\epsilon (x) = \{ y\in X \,|\, d(g^n(x),g^n(y)) \leq \epsilon \mbox{ for all } n \geq 0 \}.
$$
The map $g$ is called \defn{forward expansive} if there exists $\epsilon >0$ such that $\Phi_\epsilon(x)=\{x\}$ for all $x\in X$. By R.~Bowen's definition in \cite{Bow72}, the map $g$ is \defn{$h$-expansive} if there exists $\epsilon>0$ such that the topological entropy $h_{\operatorname{top}}(g|_{\Phi_\epsilon(x)})=h_{\operatorname{top}}(g,\Phi_\epsilon(x))$ of $g$ restricted to $\Phi_\epsilon(x)$ is $0$ for all $x\in X$. One can also formulate asymptotic $h$-expansiveness in a similar spirit, see for example, \cite[Section~2]{Mi76}. However, in this paper, we will adopt equivalent formulations from \cite{Do11}. See Section~\ref{subsctAsymHExpConcepts} for details.
Another way to formulate forward expansiveness is via distance expansion. We say that $g\: X\rightarrow X$ is \defn{distance-expanding} (with respect to the metric $d$) if there exist constants $\lambda>1$, $\eta>0$, and $n\in\N$ such that for all $x,y\in X$ with $d(x,y)\leq \eta$, we have $d(g^n(x),g^n(y))\geq \lambda d(x,y)$. If $g$ is forward expansive, then there exists a metric $\rho$ on $X$ such that the metrics $d$ and $\rho$ induce the same topology on $X$ and $g$ is distance-expanding with respect to $\rho$ (see for example, \cite[Theorem~4.6.1]{PU10}). Conversely, if $g$ is distance-expanding, then it is forward expansive (see for example, \cite[Theorem~4.1.1]{PU10}). So roughly speaking, if $g$ is forward expansive, then the distance between two points that are close enough grows exponentially under forward iterations of $g$.
Since a Thurston map, by definition, has to be a branched covering map, we can always find two distinct points that are arbitrarily close to a critical point (thus arbitrarily close to each other) and that are mapped to the same point. Thus a Thurston map cannot be forward expansive. In order to impose some expansion condition, it is then natural to consider backward orbits. We say that a Thurston map is expanding if for any two points $x,y\in S^2$, their preimages under iterations of the map gets closer and closer. See Definition~\ref{defExpanding} for a precise formulation.
The expansion property of expanding Thurston maps seems to be rather strong. However, as a part of our first main theorem below, we will show that no expanding Thurston map is $h$-expansive.
\begin{theorem} \label{thmWeakExpansion}
Let $f\: S^2 \rightarrow S^2$ be an expanding Thurston map. Then $f$ is asymptotically $h$-expansive if and only if $f$ has no periodic critical points. Moreover, $f$ is not $h$-expansive.
\end{theorem}
When R.~Bowen introduced $h$-expansiveness in \cite{Bow72}, he mentioned that no diffeomorphism of a compact manifold was known to be not $h$-expansive. M.~Misiurewicz then produced an example of a diffeomorphism that is not asymptotically $h$-expansive \cite{Mi73}. M.~Lyubich showed that any rational map is asymptotically $h$-expansive \cite{Ly83}. J.~Buzzi established asymptotic $h$-expan\-sive\-ness of any $C^\infty$-map on a compact Riemannian manifold \cite{Bu97}. Examples of $C^\infty$-maps that are not $h$-expansive were given by M.~J.~Pacifico and J.~L.~Vieitez \cite{PV08}. Our Theorem~\ref{thmWeakExpansion} implies that any rational expanding Thurston map (i.e., any postcritically-finite rational map whose Julia set is the whole sphere (see \cite[Proposition~19.1]{BM10})) is not $h$-expansive.
Expanding Thurston maps may be the first example of a class of a priori non-differentiable maps that are not $h$-expansive but may be asymptotically $h$-expansive depending on the property of orbits of critical points.
As an immediate consequence of Theorem~\ref{thmWeakExpansion} and the result of J.~Buzzi \cite{Bu97} mentioned above, we get the following corollary, which partially answers a question of K.~Pilgrim (see Problem~2 in \cite[Section~21]{BM10}).
\begin{cor} \label{corNotConjDiff}
An expanding Thurston map with at least one periodic critical point cannot be conjugate to a $C^\infty$-map on the Euclidean $2$-sphere.
\end{cor}
\smallskip
Our real motivation to investigate Theorem~\ref{thmWeakExpansion} comes from another basic theme in the study of dynamical systems, namely, the investigation of the measure-theoretic entropy and measure-theoretic pressure, and their maximizing measures known as the measures of maximal entropy and equilibrium states, respectively.
For a continuous map on a compact metric space, we can consider the topological pressure as a weighted version of the topological entropy, with the weight induced by a real-valued continuous function, called \emph{potential}. The Variational Principle identifies the topological pressure with the supremum of its measure-theoretic counterpart, the measure-theoretic pressure, over all invariant Borel probability measures \cite{Bow75, Wa76}. Under additional regularity assumptions on the map and the potential, one gets existence and uniqueness of an invariant Borel probability measure maximizing measure-theoretic pressure, called the equilibrium state for the given map and the potential. When the potential is $0$, the corresponding equilibrium state is known as the measure of maximal entropy. Often periodic points and iterated preimages are equidistributed in some appropriate sense with respect to such measures. See Section~\ref{subsctThermDynForm} for concepts mentioned here.
The existence, uniqueness, and various properties of equilibrium states have been studied in many different contexts (see for example, \cite{Bow75, Ru89, Pr90, KH95, Zi96, MauU03, BS03, Ol03, Yu03, PU10, MayU10}).
M.~Misiurewicz showed that asymptotic $h$-expansiveness guarantees that the measure-theoretic entropy $\mu\mapsto h_\mu(f)$ is upper semi-continuous \cite{Mi76}. We then get the following corollary from Theorem~\ref{thmWeakExpansion}.
\begin{cor} \label{corUSC}
Let $f\: S^2 \rightarrow S^2$ be an expanding Thurston map without periodic critical points. Then the measure-theoretic entropy $h_\mu(f)$ considered as a function of $\mu$ on the space $\mathcal{M}(S^2,f)$ of $f$-invariant Borel probability measures is upper semi-continuous. Here $\mathcal{M}(S^2,f)$ is equipped with the weak$^*$ topology.
\end{cor}
Recall that if $X$ is a metric space, a function $h\: X\rightarrow [-\infty,+\infty]$ is \defn{upper semi-continuous} if $\limsup_{y\to x} h(y) \leq h(x)$ for all $x\in \X$.
In \cite{Li14}, we established the existence and uniqueness of the equilibrium state for an expanding Thurston map and a given real-valued H\"older continuous potential. Here the sphere $S^2$ is equipped with a natural metric induced by $f$, called a visual metric. (See Theorem~\ref{thmExistenceUniquenessES}.) The tools we used in \cite{Li14} are from the \emph{thermodynamical formalism}. Neither Theorem~\ref{thmWeakExpansion} nor Corollary~\ref{corUSC} was used there. Note that Corollary~\ref{corUSC} implies a partially stronger existence result than the one obtained in \cite{Li14}.
\begin{theorem} \label{thmExistenceES}
Let $f\:S^2 \rightarrow S^2$ be an expanding Thurston map without periodic critical points and $\psi \in C(S^2)$ be a real-valued continuous function on $S^2$ (with respect to the standard topology). Then there exists at least one equilibrium state for the map $f$ and the potential $\psi$.
\end{theorem}
See Section~\ref{subsctThermDynForm} for a quick proof after necessary definitions are given precisely.
Once we know the existence and uniqueness of the equilibrium states, one natural question to ask is how periodic points and iterated preimages are distributed with respect to such measures. We know that for an expanding Thurston map, iterated preimages and preperiodic points (and in particular, periodic points) are equidistributated with respect to the unique measure of maximal entropy (see \cite{Li13} and \cite{HP09}).
Some versions of equidistribution of iterated preimages with respect to the unique equilibrium state for an expanding Thurston map and a H\"older continuous potential were obtained in \cite{Li14}. We record them in Proposition~\ref{propWeakConvPreImgWithWeight}. However, similar results for periodic points were inaccessible by the methods used in \cite{Li14} due to technical difficulties arising from the existence of critical points.
In this paper, thanks to Theorem~\ref{thmWeakExpansion} and Corollary~\ref{corUSC}, rather than trying to establish the equidistribution of periodic points directly, we derive some stronger results using a general framework devised by Y.~Kifer \cite{Ki90}. More precisely, we obtain \emph{level-2 large deviation principles} for periodic points with respect to equilibrium states in the context of expanding Thurston maps without periodic critical points and H\"older continuous potentials. We use a variant of Y.~Kifer's result formulated by H.~Comman and J.~Rivera-Letelier \cite{CRL11}, which is recorded in Theorem~\ref{thmAbsLargeDeviationPrincipleCRL} for the convenience of the reader. For related results on large deviation principles in the context of rational maps on the Riemann sphere under additional assumptions, see \cite{PSh96, PSr07, XF07, PRL11, Com09, CRL11}.
Denote the space of Borel probability measures on a compact metric space $X$ equipped with the weak$^*$ topology by $\mathcal{P}(X)$. A sequence $\{\Omega_n \}_{n\in\N}$ of Borel probability measures on $\mathcal{P}(X)$ is said to satisfy a \defn{level-2 large deviation principle with rate function $I$} if for each closed subset $\mathfrak{F}$ of $\mathcal{P}(X)$ and each open subset $\mathfrak{G}$ of $\mathcal{P}(X)$ we have
$$
\limsup\limits_{n\to+\infty} \frac1n \log \Omega_n(\mathfrak{F}) \leq - \inf\{I(x) \,|\, x\in\mathfrak{F} \},
$$
and
$$
\liminf\limits_{n\to+\infty} \frac1n \log \Omega_n(\mathfrak{G}) \geq - \inf\{I(x) \,|\, x\in\mathfrak{G} \}.
$$
We refer the reader to \cite[Section~2.5]{CRL11} and the references therein for a more systematic introduction to the theory of large deviation principles.
In order to apply Theorem~\ref{thmAbsLargeDeviationPrincipleCRL}, we just need to verify three conditions:
\begin{enumerate}
\smallskip
\item[(1)] The existence and uniqueness of the equilibrium state.
\smallskip
\item[(2)] Some characterization of the topological pressure (see Proposition~\ref{propPressureLimitPeriodicPts} and Proposition~\ref{propPressureLimitPreImgs}).
\smallskip
\item[(3)] The upper semi-continuity of the measure-theoretic entropy.
\end{enumerate}
The first condition is established in \cite{Li14}. The second condition is weaker than the equidistribution results, and is within reach. The last condition is known for expanding Thurston maps without periodic critical points by Corollary~\ref{corUSC}. Thus we get the following level-2 large deviation principles.
\begin{theorem} \label{thmLDP}
Let $f\: S^2\rightarrow S^2$ be an expanding Thurston map with no periodic critical points, and $d$ a visual metric on $S^2$ for $f$. Let $\mathcal{P}(S^2)$ denote the space of Borel probability measures on $S^2$ equipped with the weak$^*$ topology. Let $\phi$ be a real-valued H\"older continuous function on $(S^2,d)$, and $\mu_\phi$ be the unique equilibrium state for the map $f$ and the potential $\phi$.
For each $n\in\N$, let $W_n\: S^2\rightarrow\mathcal{P}(S^2)$ be the continuous function defined by
$$
W_n(x) = \frac1n \sum\limits_{i=0}^{n-1} \delta_{f^i(x)},
$$
and denote $S_n\phi(x) = \sum\limits_{i=0}^{n-1} \phi\(f^i(x)\)$ for $x\in S^2$. Fix an arbitrary sequence of functions $\{w_n\: S^2\rightarrow \R\}_{n\in\N}$ satisfying $w_n(x)\in [1,\deg_{f^n}(x)]$ for each $n\in\N$ and each $x\in S^2$. We consider the following sequences of Borel probability measures on $\mathcal{P}(S^2)$:
{\bf Iterated preimages:} Given a sequence $\{x_n\}_{n\in\N}$ of points in $S^2$, for each $n\in\N$, put
$$
\Omega_n (x_n) = \sum\limits_{y\in f^{-n}(x_n)} \frac{ w_n(y) \exp(S_n\phi(y))}{\sum_{z\in f^{-n}(x_n)} w_n(z) \exp(S_n\phi(z)) }\delta_{W_n(y)}.
$$
{\bf Periodic points:} For each $n\in\N$, put
$$
\Omega_n = \sum\limits_{x=f^n(x)} \frac{ w_n(x) \exp(S_n\phi(x))}{\sum_{y=f^n(y)} w_n(y) \exp(S_n\phi(y)) }\delta_{W_n(x)}.
$$
Then each of the sequences $\{\Omega_n(x_n)\}_{n\in\N}$ and $\{\Omega_n\}_{n\in\N}$ converges to $\delta_{\mu_\phi}$ in the weak$^*$ topology, and satisfies a large deviation principle with rate function $I^\phi \: \mathcal{P}(S^2) \rightarrow [0,+\infty]$ given by
\begin{equation} \label{eqRateFnI}
I^\phi (\mu) =
\left\{
\begin{array}{ll}
P(f,\phi) -\int\! \phi \,\mathrm{d}\mu - h_\mu(f) & \mbox{if } \mu\in\mathcal{M}(S^2,f); \\
+\infty & \mbox{if } \mu\in\mathcal{P}(S^2) \setminus \mathcal{M}(S^2,f).
\end{array}
\right.
\end{equation}
Furthermore, for each convex open subset $\mathfrak{G}$ of $\mathcal{P}(S^2)$ containing some invariant measure, we have
\begin{equation} \label{eqInfI}
-\inf\limits_{\mathfrak{G}} I^\phi = \lim\limits_{n\to+\infty} \frac{1}{n} \log \Omega_n (x_n)(\mathfrak{G}) =\lim\limits_{n\to+\infty} \frac{1}{n} \log \Omega_n(\mathfrak{G})
\end{equation}
and (\ref{eqInfI}) remains true with $\mathfrak{G}$ replaced by its closure $\overline{\mathfrak{G}}$.
\end{theorem}
As an immediate consequence, we get the following corollary. See Section~\ref{subsctEquidistr} for the proof.
\begin{cor} \label{corMeasTheoPressure}
Let $f\: S^2\rightarrow S^2$ be an expanding Thurston map with no periodic critical points, and $d$ a visual metric on $S^2$ for $f$. Let $\phi$ be a real-valued H\"older continuous function on $(S^2,d)$, and $\mu_\phi$ be the unique equilibrium state for the map $f$ and the potential $\phi$. Given a sequence $\{x_n\}_{n\in\N}$ of points in $S^2$. Fix an arbitrary sequence of functions $\{w_n\: S^2\rightarrow \R\}_{n\in\N}$ satisfying $w_n(x)\in [1,\deg_{f^n}(x)]$ for each $n\in\N$ and each $x\in S^2$.
Then for each $\mu\in\mathcal{M}(S^2,f)$, and each convex local basis $G_\mu$ of $\mathcal{P}(S^2)$ at $\mu$, we have
\begin{align}
& h_\mu(f) + \int\! \phi \,\mathrm{d}\mu \notag \\
= & \inf \Bigg\{ \lim\limits_{n\to+\infty} \frac{1}{n} \log \sum\limits_{y\in f^{-n}(x_n), \, W_n(y)\in\mathfrak{G}} w_n(y) e^{S_n\phi(y)} \,\Bigg|\, \mathfrak{G}\in G_\mu \Bigg\} \label{eqMeasTheoPressure}\\
= & \inf \Bigg\{ \lim\limits_{n\to+\infty} \frac{1}{n} \log \sum\limits_{x=f^n(x), \, W_n(x)\in\mathfrak{G}} w_n(x) e^{S_n\phi(x)} \,\Bigg|\, \mathfrak{G}\in G_\mu \Bigg\}. \notag
\end{align}
Here $W_n$ and $S_n\phi$ are as defined in Theorem~\ref{thmLDP}.
\end{cor}
As mentioned above, equidistribution results follow from corresponding level-2 large deviation principles.
\begin{cor} \label{corEquidistr}
Let $f\: S^2\rightarrow S^2$ be an expanding Thurston map with no periodic critical points, and $d$ a visual metric on $S^2$ for $f$. Let $\phi$ be a real-valued H\"older continuous function on $(S^2,d)$, and $\mu_\phi$ be the unique equilibrium state for the map $f$ and the potential $\phi$. Fix an arbitrary sequence of functions $\{w_n\: S^2\rightarrow \R\}_{n\in\N}$ satisfying $w_n(x)\in [1,\deg_{f^n}(x)]$ for each $n\in\N$ and each $x\in S^2$.
We consider the following sequences of Borel probability measures on $S^2$:
{\bf Iterated preimages:} Given a sequence $\{x_n\}_{n\in\N}$ of points in $S^2$, for each $n\in\N$, put
$$
\nu_n = \sum\limits_{y\in f^{-n}(x_n)} \frac{ w_n(y) \exp(S_n\phi(y))}{\sum_{z\in f^{-n}(x_n)} w_n(z) \exp(S_n\phi(z)) }\frac1n \sum\limits_{i=0}^{n-1} \delta_{f^i(y)},
$$
{\bf Periodic points:} For each $n\in\N$, put
$$
\eta_n = \sum\limits_{x=f^n(x)} \frac{ w_n(x) \exp(S_n\phi(x))}{\sum_{y=f^n(y)} w_n(y) \exp(S_n\phi(y)) }\frac1n \sum\limits_{i=0}^{n-1} \delta_{f^i(x)}.
$$
Then as $n\longrightarrow +\infty$,
\begin{equation*}
\nu_n \stackrel{w^*}{\longrightarrow} \mu_\phi,\quad \text{and} \quad
\eta_n \stackrel{w^*}{\longrightarrow} \mu_\phi.
\end{equation*}
Here $S_n$ is defined as in Theorem~\ref{thmLDP}.
\end{cor}
\begin{rem}
Since $S_n\phi(f^i(x))=S_n\phi(x)$ for $i\in\N$ if $f^n(x)=x$, we get
$$
\eta_n = \sum\limits_{x=f^n(x)} \frac{ \frac{S_n w_n(x)}{n} \exp(S_n\phi(x))}{\sum_{y=f^n(y)} w_n(y) \exp(S_n\phi(y)) }\delta_x,
$$
for $n\in\N$. In particular, when $w_n(\cdot)\equiv 1$,
$$
\eta_n = \sum\limits_{x=f^n(x)} \frac{ \exp(S_n\phi(x))}{\sum_{y=f^n(y)} \exp(S_n\phi(y)) }\delta_x;
$$
when $w_n(x)=\deg_{f^n}(x)$, since $\deg_{f^n}(f^i(x))=\deg_{f^n}(x)$ for $i\in\N$ if $f^n(x)=x$, we have
$$
\eta_n = \sum\limits_{x=f^n(x)} \frac{ \deg_{f^n}(x)\exp(S_n\phi(x))}{\sum_{y=f^n(y)} \deg_{f^n}(y) \exp(S_n\phi(y)) }\delta_x.
$$
\end{rem}
See Section~\ref{subsctEquidistr} for the proof of Corollary~\ref{corEquidistr}. Note that the part of Corollary~\ref{corEquidistr} on iterated preimages generalizes (\ref{eqWeakConvPreImgToMuSumWithWeight}) and (\ref{eqWeakConvPreImgToMuTildeWithWeight}) in Proposition~\ref{propWeakConvPreImgWithWeight} in the context of expanding Thurston maps without periodic critical points. We also remark that our results Corollary~\ref{corUSC} through Corollary~\ref{corEquidistr} are only known in this context. In particular, the following questions for expanding Thurston maps $f\: S^2\rightarrow S^2$ with at least one periodic critical point are still open.
\begin{question}
Is the measure-theoretic entropy $\mu\mapsto h_\mu(f)$ upper semi-continuous?
\end{question}
\begin{question}
Are iterated preimages and periodic points equidistributed with respect to the unique equilibrium state for a H\"older continuous potential?
\end{question}
Note that regarding Question~2, we know that iterated preimages, counted with local degree, are equidistributed with respect to the equilibrium state by (\ref{eqWeakConvPreImgToMuSumWithWeight}) in Proposition~\ref{propWeakConvPreImgWithWeight}. If Question~1 can be answered positively, then the mechanism of Theorem~\ref{thmAbsLargeDeviationPrincipleCRL} works and we get that the equidistribution of periodic points from the corresponding large deviation principle. However, for iterated preimages without counting local degree, (i.e., when $w_n(\cdot)\neq \deg_{f^n}(\cdot)$ in Corollary~\ref{corEquidistr}, and in particular, when $w_n(\cdot)\equiv 1$,) the verification of Condition~(2) mentioned earlier for Theorem~\ref{thmAbsLargeDeviationPrincipleCRL} to apply still remains unknown. Compare (\ref{eqPressureLimitPreImgsWDeg}) and (\ref{eqPressureLimitPreImgsWODeg}) in Proposition~\ref{propPressureLimitPreImgs}.
\smallskip
We will now give a brief description of the structure of this paper.
After fixing some notation in Section~\ref{sctNotation}, we give a quick review of Thurston maps in Section~\ref{sctThurstonMap}. We direct the reader to \cite[Section~3]{Li14} for a more detailed introduction to such maps and the terminology that we use in this paper. However, we do record explicitly most of the results from \cite{BM10,Li13,Li14} that will be used in this paper.
In Section~\ref{sctAssumptions}, we state the assumptions on some of the objects in this paper, which we are going to repeatedly refer to later as \emph{the Assumptions}. Note that these assumptions are the same as those in \cite[Section~4]{Li14}.
Section~\ref{sctAsymHExp} is devoted to the investigation of the weak expansion properties of expanding Thurston maps and the proof of Theorem~\ref{thmWeakExpansion}.
We first introduce basic concepts in Section~\ref{subsctAsymHExpConcepts}. We review the notion of topological conditional entropy $h(g|\lambda)$ of a continuous map $g\: X\rightarrow X$ (on a compact metric space $X$) given an open cover $\lambda$ of $X$, and the notion of topological tail entropy $h^*(g)$ of $g$. The latter was first introduced by M.~Misiurewicz under the name ``topological conditional entropy'' \cite{Mi73,Mi76}. We adopt the terminology and formulations by T.~Downarowicz in \cite{Do11}. We then define $h$-expansiveness and asymptotic $h$-expansiveness using these notions.
In Section~\ref{subsctWeakExpLemmas}, we prove four lemmas that will be used in the proof of the asymptotic $h$-expansiveness of expanding Thurston maps without periodic critical points. Lemma~\ref{lmLocInjectAwayFromCrit} states that any expanding Thurston map is uniformly locally injective away from the critical points, in the sense that if one fixes such a map $f$ and a visual metric $d$ on $S^2$ for $f$, then for each $\delta>0$ sufficiently small and each $x\in S^2$, the map $f$ is injective on the $\delta$-ball centered at $x$ as long as $x$ is not in a $\tau(\delta)$-ball of any critical point of $f$, where $\tau(\delta)$ can be made arbitrarily small if one lets $\delta$ go to $0$. In Lemma~\ref{lmTileInFlower} we prove a few properties of flowers in the cell decompositions of $S^2$ induced by an expanding Thurston map and some special $f$-invariant Jordan curve. Lemma~\ref{lmCoverByFlowers} gives a covering lemma to cover sets of the form $\bigcap\limits_{i=0}^n f^{-i}(W_i)$ by $(m+n)$-flowers, where $m\in\N_0$, $n\in\N$, and each $W_i$ is an $m$-flower. Finally, we review some basic concepts in graph theory, and provide a simple upper bound of number of leaves of certain trees in Lemma~\ref{lmTree}. Note that we will not use any nontrivial facts from graph theory in this paper.
Section~\ref{subsctProofWeakExp} consists of the proof of Theorem~\ref{thmWeakExpansion} in the form of three separate theorems. Namely, we show in Theorem~\ref{thmAsympHExpWOPC} the asymptotic $h$-expansiveness of expanding Thurston maps without periodic critical points. The proof relies on a quantitative upper bound of the frequency for an orbit under such a map to get close to the set of critical points. Lemma~\ref{lmTree} and terminology from graph theory is used here to make the statements in the proof precise. We then prove in Theorem~\ref{thmNotAsympHExp} and Theorem~\ref{thmNotHExp} the lack of asymptotic $h$-expansiveness of expanding Thurston maps with periodic critical points and the lack of $h$-expansiveness of expanding Thurston maps without periodic critical points, respectively, by explicit constructions of periodic sequences $\{v_i\}_{i\in\N}$ of $m$-vertices for which one can give lower bounds for the numbers of open sets in the open cover $\bigvee\limits_{j=0}^{n-1} f^{-j}\(\W^m\)$ needed to cover the set $\bigcap\limits_{j=0}^{n-1} f^{-j}(W^m(v_{n-j}))$, for $l,m,n\in\N$ sufficiently large. Here $W^m(v_{n-j})$ denotes the $m$-flower of $v_{n-j}$ (see (\ref{defFlower})), and $\W^m$ is the set of all $m$-flowers (see (\ref{defSetNFlower})). These lower bounds lead to the conclusion that the topological tail entropy and topological conditional entropy, respectively, are strictly positive, proving the corresponding theorems (compare with Defintion~\ref{defHExp} and Definition~\ref{defAsympHExp}). The periodic sequence $\{v_i\}_{i\in\N}$ of $m$-vertices in the proof of Theorem~\ref{thmNotAsympHExp} shadows a certain infinite backward pseudo-orbit in such a way that each period of $\{v_i\}_{i\in\N}$ begins with a backward orbit starting at a critical point $p$ which is a fixed point of $f$, and approaching $p$ as the index $i$ increases, and then ends with a constant sequence staying at $p$. The fact that the constant part of each period of $\{v_i\}_{i\in\N}$ can be made arbitrarily long is essential here and is not true if $f$ has no periodic critical points. The periodic sequence $\{v_i\}_{i\in\N_0}$ of $m$-vertices in the proof of Theorem~\ref{thmNotHExp} shadows a certain infinite backward pseudo-orbit in such a way that each period of $\{v_i\}_{i\in\N_0}$ begins with a backward orbit starting at $f(p)$ and $p$, and approaching $f(p)$ as the index $i$ increases, and then ends with $f(p)$. In this case $p$ is a critical point whose image $f(p)$ is a fixed point. In both constructions, we may need to consider an iterate of $f$ for the existence of $p$ with the required properties. Combining Theorems~\ref{thmAsympHExpWOPC}, \ref{thmNotAsympHExp}, and \ref{thmNotHExp}, we get Theorem~\ref{thmWeakExpansion}.
Section~\ref{sctLDP} is devoted to the study of large deviation principles and equidistribution results for periodic points and iterated preimages of expanding Thurston maps without periodic critical points. The idea is to apply a general framework devised by Y.~Kifer \cite{Ki90} to obtain level-2 large deviation principles, and to derive the equidistribution results as consequences.
In Section~\ref{subsctThermDynForm}, we review briefly the theory of thermodynamical formalism and recall relevant concepts and results in this theory from \cite{Li14} in the context of expanding Thurston maps and H\"older continuous potentials. After the necessary concepts are introduced, we provide a quick proof of Theorem~\ref{thmExistenceES}, which asserts the existence of equilibrium states for expanding Thurston maps without periodic critical points and given continuous potentials.
In Section~\ref{subsctLDP}, we give a brief review of level-2 large deviation principles in our context. We record the theorem of Y.~Kifer \cite{Ki90}, reformulated by H.~Comman and J.~Rivera-Letelier \cite{CRL11}, on level-2 large deviation principles. This result, stated in Theorem~\ref{thmAbsLargeDeviationPrincipleCRL}, will be applied later to our context.
After proving and recording several technical lemmas in Section~\ref{subsctLDPLemmas}, we generalize some characterization of topological pressure in Section~\ref{subsctTopPres} in our context. More precisely, we use equidistribution results for iterated preimages from \cite{Li14} recorded in Proposition~\ref{propWeakConvPreImgWithWeight} to show in Proposition~\ref{propPressureLimitPreImgs} and Proposition~\ref{propPressureLimitPeriodicPts} that
\begin{equation} \label{eqTopPresCharact}
P(f,\phi)=\lim\limits_{n\to +\infty} \frac{1}{n} \log \sum w_n(y) \exp(S_n\phi(y)),
\end{equation}
where the sum is taken over preimages under $f^n$ in Proposition~\ref{propPressureLimitPreImgs}, and over periodic points in Proposition~\ref{propPressureLimitPeriodicPts}, the potential $\phi\:S^2 \rightarrow \R$ is H\"older continuous with respect to a visual metric $d$, and the weight $w_n(y)\in [1,\deg_{f^n}(y)]$ for $n\in\N$ and $y\in S^2$. We note that for periodic points, the equation (\ref{eqTopPresCharact}) is established in Proposition~\ref{propPressureLimitPeriodicPts} for all expanding Thurston maps, but for iterated preimages, we only obtain (\ref{eqTopPresCharact}) for expanding Thurston maps without periodic critical points in Proposition~\ref{propPressureLimitPreImgs}.
In Section~\ref{subsctProofLDP}, by applying Theorem~\ref{thmAbsLargeDeviationPrincipleCRL} to give a proof of Theorem~\ref{thmLDP}, we finally establish level-2 large deviation principles in the context of expanding Thurston maps without periodic critical points and given H\"older continuous potentials.
Section~\ref{subsctEquidistr} consists of the proofs of Corollary~\ref{corMeasTheoPressure} and Corollary~\ref{corEquidistr}. We first obtain characterizations of the measure-theoretic pressure in terms of the infimum of certain limits involving periodic points and iterated preimages (Corollary~\ref{corMeasTheoPressure}). Such characterizations are then used in the proof of the equidistribution results (Corollary~\ref{corEquidistr}).
\bigskip
\noindent
\textbf{Acknowledgments.} The author wants to express his gratitude to the Institut Henri Poincar\'e for the kind hospitality during his stay in Paris from January to March 2014, when a major part of this work was carried out. The author also would like to thank N.-P.~Chung for explaining his work on weak expansiveness for actions of sofic groups. Last but not least, the author wants to express his deepest gratitude to M.~Bonk for his patient teaching and guidance as the advisor of the author.
\section{Notation} \label{sctNotation}
Let $\C$ be the complex plane and $\widehat{\C}$ be the Riemann sphere. We use the convention that $\N=\{1,2,3,\dots\}$ and $\N_0 = \{0\} \cup \N$. As usual, the symbol $\log$ denotes the logarithm to the base $e$.
The cardinality of a set $A$ is denoted by $\card{A}$. For $x\in\R$, we define $\lfloor x\rfloor$ as the greatest integer $\leq x$, and $\lceil x \rceil$ the smallest integer $\geq x$.
Let $g\: X\rightarrow Y$ be a function between two sets $X$ and $Y$. We denote the restriction of $g$ to a subset $Z$ of $X$ by $g|_Z$.
Let $(X,d)$ be a metric space. For subsets $A,B\subseteq X$, we set $d(A,B)=\inf \{d(x,y)\,|\, x\in A,\,y\in B\}$, and $d(A,x)=d(x,A)=d(A,\{x\})$ for $x\in X$. For each subset $Y\subseteq X$, we denote the diameter of $Y$ by $\diam_d(Y)=\sup\{d(x,y)\,|\,x,y\in Y\}$, the interior of $Y$ by $\inter Y$, the closure of $Y$ by $\overline Y$, and the characteristic function of $Y$ by $\mathbbm{1}_Y$, which maps each $x\in Y$ to $1\in\R$. For each $r>0$, we define $N^r_d(A)$ to be the open $r$-neighborhood $\{y\in X \,|\, d(y,A)<r\}$ of $A$, and $\overline{N^r_d}(A)$ the closed $r$-neighborhood $\{y\in X \,|\, d(y,A)\leq r\}$ of $A$. For $x\in X$, we denote the open ball of radius $r$ centered at $x$ by $B_d(x, r)$.
We set $C(X)$ to be the space of continuous functions from $X$ to $\R$, by $\mathcal{M}(X)$ the set of finite signed Borel measures, and $\mathcal{P}(X)$ the set of Borel probability measures on $X$. For $\mu\in\mathcal{M}(X)$, we use $\Norm{\mu}$ to denote the total variation norm of $\mu$, $\operatorname{supp} \mu$ the support of $\mu$, and
$$
\langle \mu,u \rangle = \int \! u \,\mathrm{d}\mu
$$
for each $u\inC(S^2)$. For a point $x\in X$, we define $\delta_x$ as the Dirac measure supported on $\{x\}$. For $g\inC(X)$ we set $\mathcal{M}(X,g)$ to be the set of $g$-invariant Borel probability measures on $X$. If we do not specify otherwise, we equip $C(X)$ with the uniform norm $\Norm{\cdot}_\infty$, and equip both $\mathcal{M}(X)$ and $\mathcal{M}(X,g)$ with the weak$^*$ topology.
The space of real-valued H\"{o}lder continuous functions with an exponent $\alpha\in (0,1]$ on a compact metric space $(X,d)$ is denoted as $\Holder{\alpha}(X,d)$. For given $f\: X \rightarrow X$ and $\varphi \in C(X)$, we define
\begin{equation} \label{eqDefSnPt}
S_n \varphi (x) = \sum\limits_{j=0}^{n-1} \varphi(f^j(x))
\end{equation}
and
\begin{equation} \label{eqDefWn}
W_n(x) = \frac{1}{n} \sum\limits_{j=0}^{n-1} \delta_{f^j(x)}
\end{equation}
for $x\in X$ and $n\in\N_0$. Note that when $n=0$, by definition we always have $S_0 \varphi = 0$, and by convention $W_0=0$.
\section{Thurston maps} \label{sctThurstonMap}
This section serves as a minimal review for expanding Thurston maps. Most of the definitions and results here were discussed in \cite[Section~3]{Li14}. The reader is encouraged to read Section~3 in \cite{Li14} for a quick introduction to expanding Thurston maps and the terminology that we use in this paper. For a more thorough treatment of the subject, we refer to \cite{BM10}.
Let $S^2$ denote an oriented topological $2$-sphere. A continuous map $f\:S^2\rightarrow S^2$ is called a \defn{branched covering map} on $S^2$ if for each point $x\in S^2$, there exists a positive integer $d\in \N$, open neighborhoods $U$ of $x$ and $V$ of $y=f(x)$, open neighborhoods $U'$ and $V'$ of $0$ in $\widehat{\C}$, and orientation-preserving homeomorphisms $\varphi\:U\rightarrow U'$ and $\eta\:V\rightarrow V'$ such that $\varphi(x)=0$, $\eta(y)=0$, and
$$
(\eta\circ f\circ\varphi^{-1})(z)=z^d
$$
for each $z\in U'$. The positive integer $d$ above is called the \defn{local degree} of $f$ at $x$ and is denoted by $\deg_f (x)$. The \defn{degree} of $f$ is
\begin{equation} \label{eqDeg=SumLocalDegree}
\deg f=\sum\limits_{x\in f^{-1}(y)} \deg_f (x)
\end{equation}
for $y\in S^2$ and is independent of $y$. If $f\:S^2\rightarrow S^2$ and $g\:S^2\rightarrow S^2$ are two branched covering maps on $S^2$, then so is $f\circ g$, and
\begin{equation} \label{eqLocalDegreeProduct}
\deg_{f\circ g}(x) = \deg_g(x)\deg_f(g(x)), \qquad \text{for each } x\in S^2.
\end{equation}
A point $x\in S^2$ is a \defn{critical point} of $f$ if $\deg_f(x) \geq 2$. The set of critical points of $f$ is denoted by $\operatorname{crit} f$. A point $y\in S^2$ is a \defn{postcritical point} of $f$ if $y = f^n(x)$ for some $x\in\operatorname{crit} f$ and $n\in\N$. The set of postcritical points of $f$ is denoted by $\operatorname{post} f$. Note that $\operatorname{post} f=\operatorname{post} f^n$ for all $n\in\N$.
\begin{definition} [Thurston maps] \label{defThurstonMap}
A Thurston map is a branched covering map $f\:S^2\rightarrow S^2$ on $S^2$ with $\deg f\geq 2$ and $\card(\operatorname{post} f)<+\infty$.
\end{definition}
Let $f\:S^2 \rightarrow S^2$ be a Thurston map, and $\mathcal{C}\subseteq S^2$ be a Jordan curve containing $\operatorname{post} f$. Then the pair $f$ and $\mathcal{C}$ induces natural \emph{cell decompositions} (see \cite[Definition~3.2]{Li14}) $\mathbf{D}^n(f,\mathcal{C})$ of $S^2$, for $n\in\N_0$, such that
$$
\mathbf{D}^n(f,\mathcal{C})=\X^n(f,\mathcal{C}) \cup \E^n(f,\mathcal{C}) \cup \overline\V^n(f,\mathcal{C})
$$
consisting of \emph{$n$-cells}, where the set $\X^n(f,\mathcal{C})$ consists of \emph{$n$-tiles}, the set $\E^n(f,\mathcal{C})$ consists of \emph{$n$-edges}, and $\overline\V^n(f,\mathcal{C}) = \{ \{x\} \,|\, x\in \V^n(f,\mathcal{C})\}$ where the set $\V^n(f,\mathcal{C})$ consists of \emph{$n$-vertices}. The \defn{interior} of an $n$-cell is denoted by $\inte(c)$ (see the discussion preceding Definition~3.2 in \cite{Li14}). The \defn{$k$-skeleton}, for $k\in\{0,1,2\}$, of $\mathbf{D}^n(f,\mathcal{C})$ is the union of all $n$-cells of dimension $k$ in this cell decomposition.
We record Proposition~6.1 of \cite{BM10} here in order to summarize properties of the cell decompositions $\mathbf{D}^n(f,\mathcal{C})$ defined above.
\begin{prop}[M.~Bonk \& D.~Meyer, 2010] \label{propCellDecomp}
Let $k,n\in \N_0$, let $f\: S^2\rightarrow S^2$ be a Thurston map, $\mathcal{C}\subseteq S^2$ be a Jordan curve with $\operatorname{post} f \subseteq \mathcal{C}$, and $m=\card(\operatorname{post} f)$.
\smallskip
\begin{itemize}
\smallskip
\item[(i)] The map $f^k$ is cellular for $(\mathbf{D}^{n+k}(f,\mathcal{C}), \mathbf{D}^n(f,\mathcal{C}))$. In particular, if $c$ is any $(n+k)$-cell, then $f^k(c)$ is an $n$-cell, and $f^k|_c$ is a homeomorphism of $c$ onto $f^k(c)$.
\smallskip
\item[(ii)] Let $c$ be an $n$-cell. Then $f^{-k}(c)$ is equal to the union of all
$(n+k)$-cells $c'$ with $f^k(c')=c$.
\smallskip
\item[(iii)] The $1$-skeleton of $\mathbf{D}^n(f,\mathcal{C})$ is equal to $f^{-n}(\mathcal{C})$. The $0$-skeleton of $\mathbf{D}^n(f,\mathcal{C})$ is the set $\V^n(f,\mathcal{C})=f^{-n}(\operatorname{post} f )$, and we have $\V^n(f,\mathcal{C}) \subseteq \V^{n+k}(f,\mathcal{C})$.
\smallskip
\item[(iv)] $\card(\X^n(f,\mathcal{C}))=2(\deg f)^n$, $\card(\E^n(f,\mathcal{C}))=m(\deg f)^n$, and $\card (\V^n(f,\mathcal{C})) \leq m (\deg f)^n$.
\smallskip
\item[(v)] The $n$-edges are precisely the closures of the connected components of $f^{-n}(\mathcal{C})\setminus f^{-n}(\operatorname{post} f )$. The $n$-tiles are precisely the closures of the connected components of $S^2\setminus f^{-n}(\mathcal{C})$.
\smallskip
\item[(vi)] Every $n$-tile is an $m$-gon, i.e., the number of $n$-edges and the number of $n$-vertices contained in its boundary are equal to $m$.
\end{itemize}
\end{prop}
From now on, if the map $f$ and the Jordan curve $\mathcal{C}$ are clear from the context, we will sometimes omit $(f,\mathcal{C})$ in the notation above.
If we fix the cell decomposition $\mathbf{D}^n(f,\mathcal{C})$, $n\in\N_0$, we can define for each $v\in \V^n$ the \defn{$n$-flower of $v$} as
\begin{equation} \label{defFlower}
W^n(v) = \bigcup \{\inte (c) \,|\, c\in \mathbf{D}^n,\, v\in c \}.
\end{equation}
Note that flowers are open (in the standard topology on $S^2$). Let $\overline{W}^n(v)$ be the closure of $W^n(v)$. We define the \defn{set of all $n$-flowers} by
\begin{equation} \label{defSetNFlower}
\W^n = \{W^n(v) \,|\, v\in\V^n\}.
\end{equation}
\begin{rem} \label{rmFlower}
For $n\in\N_0$ and $v\in\V^n$, we have
$$
\overline{W}^n(v)=X_1\cup X_2\cup \cdots \cup X_m,
$$
where $m=2\deg_{f^n}(v)$, and $X_1, X_2, \dots X_m$ are all the $n$-tiles that contains $v$ as a vertex (see \cite[Lemma~7.2]{BM10}). Moreover, each flower is mapped under $f$ to another flower in such a way that is similar to the map $z\mapsto z^k$ on the complex plane. More precisely, for $n\in\N_0$ and $v\in \V^{n+1}$, there exists orientation preserving homeomorphisms $\varphi\: W^{n+1}(v) \rightarrow D$ and $\eta\: W^{n}(f(v)) \rightarrow D$ such that $D$ is the unit disk on $\C$, $\varphi(v)=0$, $\eta(f(v))=0$, and
$$
(\eta\circ f \circ \varphi^{-1}) (z) = z^k
$$
for all $z\in D$, where $k=\deg_f(v)$. Let $\overline{W}^{n+1}(v)= X_1\cup X_2\cup \cdots \cup X_m$ and $\overline{W}^n(f(v))= X'_1\cup X'_2\cup \cdots \cup X'_{m'}$, where $X_1, X_2, \dots X_m$ are all the $(n+1)$-tiles that contains $v$ as a vertex, listed counterclockwise, and $X'_1, X'_2, \dots X'_{m'}$ are all the $n$-tiles that contains $f(v)$ as a vertex, listed counterclockwise, and $f(X_1)=X'_1$. Then $m= m'k$, and $f(X_i)=X'_j$ if $i\equiv j \pmod{k}$, where $k=\deg_f(v)$. (See also Case~3 of the proof of Lemma~5.2 in \cite{BM10} for more details.)
\end{rem}
\begin{definition} [Expansion] \label{defExpanding}
A Thurston map $f\:S^2\rightarrow S^2$ is called \defn{expanding} if there exist a metric $d$ on $S^2$ that induces the standard topology on $S^2$ and a Jordan curve $\mathcal{C}\subseteq S^2$ containing $\operatorname{post} f$ such that
\begin{equation*}
\lim\limits_{n\to+\infty}\max \{\diam_d(X) \,|\, X\in \X^n(f,\mathcal{C})\}=0.
\end{equation*}
\end{definition}
It is clear that if $f\: S^2\rightarrow S^2$ is an expanding Thurston map, then so is $f^n\: S^2\rightarrow S^2$, for $n\in \N$.
For an expanding Thurston map $f$, we can fix a particular metric $d$ on $S^2$ called a \emph{visual metric for $f$} . For the existence and properties of such metrics, see \cite[Chapter~8]{BM10}. For a fixed expanding Thurston map, each visual metric corresponds to a unique \emph{expansion factor} $\Lambda >1$. One major advantage of a visual metric $d$ is that in $(S^2,d)$ we have good quantitative control over the sizes of the cells in the cell decompositions discussed above (see \cite[Lemma~8.10]{BM10}).
\begin{lemma}[M.~Bonk \& D.~Meyer, 2010] \label{lmCellBoundsBM}
Let $f\:S^2 \rightarrow S^2$ be an expanding Thurston map, and $\mathcal{C} \subseteq S^2$ be a Jordan curve containing $\operatorname{post} f$. Let $d$ be a visual metric on $S^2$ for $f$ with expansion factor $\Lambda>1$. Then there exists a constant $C\geq 1$ such that for all $n$-edges and all $n$-tiles $\tau$ with $n\in\N_0$, we have $C^{-1} \Lambda^{-n} \leq \diam_d(\tau) \leq C\Lambda^{-n}$.
\end{lemma}
In addition, we will need the fact that a visual metric $d$ induces the standard topology on $S^2$ (\cite[Proposition~8.9]{BM10}) and the fact that the metric space $(S^2,d)$ is \emph{linearly locally connected} (\cite[Proposition~16.3]{BM10}).
A Jordan curve $\mathcal{C}\subseteq S^2$ is \defn{$f$-invariant} if $f(\mathcal{C})\subseteq \mathcal{C}$. For each $f$-invariant Jordan curve $\mathcal{C}\subseteq S^2$ containing $\operatorname{post} f$, the partition $(\mathbf{D}^1,\mathbf{D}^0)$ is a \emph{cellular Markov partition} for $f$ (see \cite[Definition~3.4]{Li14}). M.~Bonk and D.~Meyer \cite[Theorem~1.2]{BM10} proved that there exists an $f^n$-invariant Jordan curve $\mathcal{C}\subseteq S^2$ containing $\operatorname{post}{f}$ for each sufficiently large $n$ depending on $f$. We proved a slightly stronger version of this result in \cite[Lemma~3.12]{Li13} which we record in the following lemma.
\begin{lemma} \label{lmCexistsL}
Let $f\:S^2\rightarrow S^2$ be an expanding Thurston map, and $\widetilde{\mathcal{C}}\subseteq S^2$ be a Jordan curve with $\operatorname{post} f\subseteq \widetilde{\mathcal{C}}$. Then there exists an integer $N(f,\widetilde{\mathcal{C}}) \in \N$ such that for each $n\geq N(f,\widetilde{\mathcal{C}})$ there exists an $f^n$-invariant Jordan curve $\mathcal{C}$ isotopic to $\widetilde{\mathcal{C}}$ rel.\ $\operatorname{post} f$ such that no $n$-tile in $\mathbf{D}^n(f,\mathcal{C})$ joins opposite sides of $\mathcal{C}$.
\end{lemma}
\begin{definition}[Joining opposite sides] \label{defConnectop}
Fix a Thurston map $f$ with $\card(\operatorname{post} f) \geq 3$ and an $f$-invariant Jordan curve $\mathcal{C}$ containing $\operatorname{post} f$. A set $K\subseteq S^2$ \defn{joins opposite sides} of $\mathcal{C}$ if $K$ meets two disjoint $0$-edges when $\card( \operatorname{post} f)\geq 4$, or $K$ meets all three $0$-edges when $\card(\operatorname{post} f)=3$.
\end{definition}
Note that $\card (\operatorname{post} f) \geq 3$ for each expanding Thurston map $f$ \cite[Corollary~6.4]{BM10}.
We proved in \cite[Lemma~3.14]{Li13} the following easy lemma.
\begin{lemma} \label{lmPreImageDense}
Let $f\:S^2\rightarrow S^2$ be an expanding Thurston map. Then for each $p\in S^2$, the set $\bigcup\limits_{n=1}^{+\infty} f^{-n}(p)$ is dense in $S^2$, and
\begin{equation}
\lim\limits_{n\to +\infty} \card(f^{-n}(p)) = +\infty.
\end{equation}
\end{lemma}
Expanding Thurston maps are Lipschitz with respect to a visual metric \cite[Lemma~3.12]{Li14}.
\begin{lemma} \label{lmLipschitz}
Let $f\:S^2 \rightarrow S^2$ be an expanding Thurston map, and $d$ be a visual metric on $S^2$ for $f$. Then $f$ is Lipschitz with respect to $d$.
\end{lemma}
We established the following generalization of \cite[Lemma~16.1]{BM10} in \cite[Lemma~3.13]{Li14}.
\begin{lemma} \label{lmMetricDistortion}
Let $f\:S^2 \rightarrow S^2$ be an expanding Thurston map, and $\mathcal{C} \subseteq S^2$ be a Jordan curve that satisfies $\operatorname{post} f \subseteq \mathcal{C}$ and $f^{n_\mathcal{C}}(\mathcal{C})\subseteq\mathcal{C}$ for some $n_\mathcal{C}\in\N$. Let $d$ be a visual metric on $S^2$ for $f$ with expansion factor $\Lambda>1$. Then there exists a constant $C_0 > 1$, depending only on $f$, $d$, $\mathcal{C}$, and $n_\mathcal{C}$, with the following property:
If $k,n\in\N_0$, $X^{n+k}\in\X^{n+k}(f,\mathcal{C})$, and $x,y\in X^{n+k}$, then
\begin{equation} \label{eqMetricDistortion}
\frac{1}{C_0} d(x,y) \leq \frac{d(f^n(x),f^n(y))}{\Lambda^n} \leq C_0 d(x,y).
\end{equation}
\end{lemma}
\section{The Assumptions} \label{sctAssumptions}
We state below the hypothesis under which we will develop our theory in most parts of this paper. We will repeatedly refer to such assumptions in the later sections.
\begin{assumptions}
\quad
\begin{enumerate}
\smallskip
\item $f\:S^2 \rightarrow S^2$ is an expanding Thurston map.
\smallskip
\item $\mathcal{C}\subseteq S^2$ is a Jordan curve containing $\operatorname{post} f$ with the property that there exists $n_\mathcal{C}\in\N$ such that $f^{n_\mathcal{C}} (\mathcal{C})\subseteq \mathcal{C}$ and $f^m(\mathcal{C})\nsubseteq \mathcal{C}$ for each $m\in\{1,2,\dots,n_\mathcal{C}-1\}$.
\smallskip
\item $d$ is a visual metric on $S^2$ for $f$ with expansion factor $\Lambda>1$ and a linear local connectivity constant $L\geq 1$.
\smallskip
\item $\phi\in \Holder{\alpha}(S^2,d)$ is a real-valued H\"{o}lder continuous function with an exponent $\alpha\in(0,1]$.
\end{enumerate}
\end{assumptions}
Observe that by Lemma~\ref{lmCexistsL}, for each $f$ in (1), there exists at least one Jordan curve $\mathcal{C}$ that satisfies (2). Since for a fixed $f$, the number $n_\mathcal{C}$ is uniquely determined by $\mathcal{C}$ in (2), in the remaining part of the paper we will say that a quantity depends on $\mathcal{C}$ even if it also depends on $n_\mathcal{C}$.
Recall that the expansion factor $\Lambda$ of a visual metric $d$ on $S^2$ for $f$ is uniquely determined by $d$ and $f$. We will say that a quantity depends on $f$ and $d$ if it depends on $\Lambda$.
Note that even though the value of $L$ is not uniquely determined by the metric $d$, in the remainder of this paper, for each visual metric $d$ on $S^2$ for $f$, we will fix a choice of linear local connectivity constant $L$. We will say that a quantity depends on the visual metric $d$ without mentioning the dependence on $L$, even though if we had not fixed a choice of $L$, it would have depended on $L$ as well.
In the discussion below, depending on the conditions we will need, we will sometimes say ``Let $f$, $\mathcal{C}$, $d$, $\phi$, $\alpha$ satisfy the Assumptions.'', and sometimes say ``Let $f$ and $d$ satisfy the Assumptions.'', etc.
\section{Asymptotic $h$-Expansiveness} \label{sctAsymHExp}
\subsection{Basic concepts} \label{subsctAsymHExpConcepts}
We first review some concepts from dynamical systems. We refer the reader to \cite[Chapter~3]{PU10}, \cite[Chapter~9]{Wa82} or \cite[Chapter~20]{KH95} for more detailed studies of these concepts.
Let $(X,d)$ be a compact metric space and $g\:X\rightarrow X$ a continuous map.
A \defn{cover} of $X$ is a collection $\xi=\{A_j \,|\, j\in J\}$ of subsets of $X$ with the property that $\bigcup\xi = X$, where $J$ is an index set. The cover $\xi$ is an \defn{open cover} if $A_j$ is an open set for each $j\in J$. The cover $\xi$ is \defn{finite} if the index set $J$ is a finite set.
A \defn{measurable partition} $\xi$ of $X$ is a cover $\xi=\{A_j\,|\,j\in J\}$ of $X$ consisting of countably many mutually disjoint Borel sets $A_j$, $j\in J$, where $J$ is a countable index set.
Let $\xi=\{A_j\,|\,j\in J\}$ and $\eta=\{B_k\,|\,k\in K\}$ be two covers of $X$, where $J$ and $K$ are the corresponding index sets. We say $\xi$ is a \defn{refinement} of $\eta$ if for each $A_j\in\xi$, there exists $B_k\in\eta$ such that $A_j\subseteq B_k$. The \defn{common refinement} $\xi \vee \eta$ of $\xi$ and $\eta$ defined as
$$
\xi \vee \eta = \{A_j\cap B_k \,|\, j\in J,\, k\in K\}
$$
is also a cover. Note that if $\xi$ and $\eta$ are both open covers (resp., measurable partitions), then $\xi \vee \eta$ is also an open cover (resp., a measurable partition). Define $g^{-1}(\xi)=\{g^{-1}(A_j) \,|\,j\in J\}$, and denote for $n\in\N$,
$$
\xi^n_g= \bigvee\limits_{j=0}^{n-1} g^{-j}(\xi) = \xi\vee g^{-1}(\xi)\vee\cdots\vee g^{-(n-1)}(\xi).
$$
We adopt the following definition from \cite[Remark~6.1.7]{Do11}.
\begin{definition} [Refining sequences of open covers] \label{defRefSeqOpenCover}
A sequence of open covers $\{\xi_i\}_{i\in\N_0}$ of a compact metric space $X$ is a \defn{refining sequence of open covers} of $X$ if the following conditions are satisfied
\begin{enumerate}
\smallskip
\item[(i)] $\xi_{i+1}$ is a refinement of $\xi_i$ for each $i\in\N_0$.
\smallskip
\item[(ii)] For each open cover $\eta$ of $X$, there exists $j\in\N$ such that $\xi_i$ is a refinement of $\eta$ for each $i\geq j$.
\end{enumerate}
\end{definition}
By the Lebesgue Number Lemma (\cite[Lemma~27.5]{Mu00}), it is clear that for a compact metric space, refining sequences of open covers always exist.
The topological tail entropy was first introduced by M.~Misiurewicz under the name ``topological conditional entropy'' \cite{Mi73, Mi76}. We adopt the terminology in \cite{Do11} (see \cite[Remark~6.3.18]{Do11}).
\begin{definition}[Topological conditional entropy and topological tail entropy] \label{defTopTailEntropy}
Let $(X,d)$ be a compact metric space and $g\:X\rightarrow X$ a continuous map. The \defn{topological conditional entropy} $h(g|\lambda)$ of $g$ given $\lambda$, for some open cover $\lambda$, is
\begin{equation} \label{eqDefTopCondEntropy}
h(g|\lambda)= \lim\limits_{l\to+\infty} \lim\limits_{n\to+\infty} \frac{1}{n} H \( \bigvee\limits_{i=0}^{n-1} g^{-i}\(\xi_l\) \Bigg| \bigvee\limits_{j=0}^{n-1} g^{-j}\(\lambda\) \),
\end{equation}
where $\{\xi_l\}_{l\in\N_0}$ is an arbitrary refining sequence of open covers, and for each pair of open covers $\xi$ and $\eta$,
\begin{equation} \label{eqH(xi|eta)}
H(\xi | \eta) = \log \Big(\max_{A\in\eta} \Big\{\min\Big\{\card \xi_A \,\Big|\, \xi_A \subseteq \xi,\, A \subseteq \bigcup \xi_A \Big\}\Big\} \Big)
\end{equation}
is the logarithm of the minimal number of sets from $\xi$ sufficient to cover any set in $\eta$.
The \defn{topological tail entropy} $h^*(g)$ of $g$ is defined by
\begin{equation} \label{eqDefTopTailEntropy}
h^*(g) = \lim\limits_{m\to+\infty} \lim\limits_{l\to+\infty} \lim\limits_{n\to+\infty} \frac{1}{n} H \( \bigvee\limits_{i=0}^{n-1} g^{-i}\(\xi_l\) \Bigg| \bigvee\limits_{j=0}^{n-1} g^{-j}\(\eta_m\) \),
\end{equation}
where $\{\xi_l\}_{l\in\N_0}$ and $\{\eta_m\}_{m\in\N_0}$ are two arbitrary refining sequences of open covers, and $H$ is as defined in (\ref{eqH(xi|eta)}).
\end{definition}
\begin{remark}
The topological entropy of $g$ (see Section~\ref{subsctThermDynForm}) is $h_{\operatorname{top}}(g) = h(g|\{X\})$, where $\{X\}$ is the open cover of $X$ consisting of only one open set $X$. See for example, \cite[Section~6.1]{Do11}.
\end{remark}
The limits in (\ref{eqDefTopCondEntropy}) and (\ref{eqDefTopTailEntropy}) always exist, and both $h(g|\lambda)$ and $h^*(g)$ are independent of the choices of refining sequences of open covers $\{\xi_l\}_{l\in\N_0}$ and $\{\eta_m\}_{m\in\N_0}$, see \cite[Section~6.3]{Do11}, especially the comments after \cite[Definition~6.3.14]{Do11}.
The topological tail entropy $h^*$ is also well-behaved under iterations, as it satisfies
\begin{equation} \label{eqH*gn=nH*g}
h^*(g^n) = n h^*(g)
\end{equation}
for each $n\in\N$ and each continuous map $g\:X\rightarrow X$ on a compact metric space $X$ (\cite[Proposition~3.1]{Mi76}).
The concept of $h$-expansiveness was introduced by R.~Bowen in \cite{Bow72}. We adopt the formulation in \cite{Mi76} (see also \cite{Do11}).
\begin{definition}[$h$-expansiveness] \label{defHExp}
A continuous map $g\: X\rightarrow X$ on a compact metric space $X$ is called \defn{$h$-expansive} if there exists a finite open cover $\lambda$ of $X$ such that $h(g|\lambda)=0$.
\end{definition}
A weaker property was then introduced by M.~Misiurewicz in \cite{Mi73} (see also \cite{Mi76, Do11}).
\begin{definition}[Asymptotic $h$-expansiveness] \label{defAsympHExp}
We say that a continuous map $g\: X\rightarrow X$ on a compact metric space $X$ is \defn{asymptotically $h$-expansive} if $h^*(g)=0$.
\end{definition}
\subsection{Technical lemmas} \label{subsctWeakExpLemmas}
Now we go back to the dynamical system $(S^2,f)$ where $f$ is an expanding Thurston map.
We need the following four lemmas for the proof of the asymptotic $h$-expansiveness of expanding Thurston maps with no periodic critical points.
\begin{lemma}[Uniform local injectivity away from the critical points] \label{lmLocInjectAwayFromCrit}
Let $f$, $d$ satisfies the Assumptions. Then there exists a number $\delta_0 \in (0,1]$ and a function $\tau\: (0,\delta_0]\rightarrow (0,+\infty)$ with the following properties:
\begin{enumerate}
\smallskip
\item[(i)] $\lim\limits_{\delta \to 0} \tau(\delta) = 0$.
\smallskip
\item[(ii)] For each $\delta \leq \delta_0$, the map $f$ restricted to any open ball of radius $\delta$ centered outside the $\tau(\delta)$-neighborhood of $\operatorname{crit} f$ is injective, i.e., $f|_{B_d(x,\delta)}$ is injective for each $x \in S^2 \setminus N_d^{\tau(\delta)}(\operatorname{crit} f)$.
\end{enumerate}
\end{lemma}
This lemma is straightforward to verify, but for the sake of completeness, we include the proof here.
\begin{proof}
We first define a function $r\: S^2\setminus \operatorname{crit} f \rightarrow (0,+\infty)$ in the following way
\begin{equation*}
r(x) = \sup \{ R>0 \,|\, f|_{B_d(x,R)} \text{ is injective} \},
\end{equation*}
for $x\in S^2\setminus \operatorname{crit} f$. Note that $r(x) \leq d(x,\operatorname{crit} f) <+\infty$ for each $x\in S^2\setminus \operatorname{crit} f$. We also observe that the supremum is attained, since otherwise, suppose $f(y)=f(z)$ for some $y,z\in B(x,r(x))$, then $f$ is not injective on the ball $B(x,R_0)$ containing $y$ and $z$ with $R_0 = \frac12(r(x) + \max\{d(x,y),\,d(x,z)\}) <r(x)$, a contradiction.
We claim that $r$ is continuous.
Indeed, let $\{x_i\}_{i\in\N}$ be a sequence of points in $S^2$ and $x\in S^2$ with the property that $\lim\limits_{i\to+\infty} x_i=x$. For each $i\in\N$, if $r(x_i) - d(x_i,x)>0$, then $B(x,r(x_i)-d(x_i,x)) \subseteq B(x_i,r(x_i))$. So $r(x) \geq r(x_i) - d(x_i,x)$. Thus
$$
r(x) \geq \limsup\limits_{i\to+\infty} (r(x_i)-d(x_i,x)) =\limsup\limits_{i\to+\infty} r(x_i).
$$
On the other hand, for each $i\in \N$, if $r(x_i) - d(x_i,x)>0$, then $B(x_i,r(x)-d(x_i,x)) \subseteq B(x,r(x))$. So $r(x_i) \geq r(x) - d(x_i,x)$. Thus
$$
\liminf\limits_{i\to+\infty} r(x_i) \geq \liminf\limits_{i\to+\infty}(r(x)-d(x_i,x)) =r(x).
$$
Hence $r(x)=\lim\limits_{i\to+\infty} r(x_i)$. So $r$ is continuous and the claim is proved.
Next, we fix a sufficiently small number $t_0>0$ with $S^2\setminus N_d^{t_0} (\operatorname{crit} f) \neq \emptyset$. We define a function $\sigma \: (0,t_0] \rightarrow (0,+\infty)$ by setting
\begin{equation*}
\sigma(t) = \inf \{ r(x) \,|\, x\in S^2\setminus N_d^t (\operatorname{crit} f) \}
\end{equation*}
for $t\in (0,t_0]$. We observe that $\sigma$ is continuous and non-decreasing. Since $r(x)\leq d(x,\operatorname{crit} f)$ for each $x\in S^2\setminus \operatorname{crit} f$, we can conclude that $\lim\limits_{t\to 0} \sigma(t) = 0$. By the definition of $\sigma$, we get that $f|_{B_d(x,\sigma(t))}$ is injective, for $t\in (0,t_0]$ and $x\in S^2\setminus N_d^t(\operatorname{crit} f)$.
Finally, we construct $\tau\: (0,\delta_0]\rightarrow (0,+\infty)$, where $\delta_0 = \min \{1,\sigma(t_0)\}$ by setting
\begin{equation} \label{eqTau}
\tau (\delta) = \inf \{ t\in (0,t_0] \,|\, \sigma(t) \geq \delta \}
\end{equation}
for each $\delta\in(0,\delta_0]$. We note that $\lim\limits_{\delta \to 0} \tau(\delta) = 0$.
For $\delta\in(0,\delta_0]$ and $t\in (\tau(\delta),t_0]$, we have $\sigma(t) \geq \delta$ by (\ref{eqTau}) and the fact that $\sigma$ is non-decreasing. Since $\sigma$ is continuous on $(0,t_0]$, we get $ \sigma(\tau(\delta)) \geq \delta$. For each $x\in S^2 \setminus N_d^{\tau(\delta)} (\operatorname{crit} f)$, we know from the definition of $\sigma$ that $f|_{B_d(x,\sigma(\tau(\delta)))}$ is injective. Therefore $f|_{B_d(x,\delta)}$ is injective.
\end{proof}
\begin{lemma} \label{lmTileInFlower}
Let $f$ and $\mathcal{C}$ satisfy the Assumptions. Fix $m,n\in\N_0$ with $m<n$. If $f(\mathcal{C})\subseteq \mathcal{C}$ and no $1$-tile in $\X^1(f,\mathcal{C})$ joins opposite sides of $\mathcal{C}$, then the following statements hold:
\begin{enumerate}
\smallskip
\item[(i)] For each $n$-vertex $v\in \V^n(f,\mathcal{C})$ and each $m$-vertex $w\in\V^m(f,\mathcal{C})$, if $v\notin \overline{W}^m(w)$, then $W^m(w)\cap W^n(v)=\emptyset$.
\smallskip
\item[(ii)] For each $n$-tile $X^n\in \X^n(f,\mathcal{C})$, there exists an $m$-vertex $v^m\in \V^m(f,\mathcal{C})$ such that $X^n\subseteq W^m(v^m)$.
\smallskip
\item[(iii)] For each pair of distinct $m$-vertices $p,q\in\V^m(f,\mathcal{C})$, $\overline{W}^{n+1}(p) \cap \overline{W}^{n+1}(q) = \emptyset$.
\end{enumerate}
\end{lemma}
Recall that $W^n$ is defined in (\ref{defFlower}) and $\overline{W}^n(p)$ is the closure of $W^n(p)$. Note that a flower is an open set (see \cite[Lemma~7.2]{BM10}) and by definition a tile is a closed set.
\begin{proof}
We first observe that in order to prove any of the statements in the lemma, it suffices to assume $n=m+1$. So we will assume, without loss of generality, that $n=m+1$.
\smallskip
(i) Since $v\notin\overline{W}^m(w)$, by (\ref{defFlower}) we get that $v\notin c$ for each $m$-cell $c\in\mathbf{D}^m$ with $w\in c$. Since $f(\mathcal{C})=\mathcal{C}$, for each $n$-cell $c'\in\mathbf{D}^n$ and each $m$-cell $c\in\mathbf{D}^m$, if $c\cap \inte(c')\neq \emptyset$, then $c'\subseteq c$ (see Lemma~4.3 and the proof of Lemma~4.7 in \cite{BM10}). Thus $c\cap \inte(c')= \emptyset$ for $c\in\mathbf{D}^m$ and $c'\in\mathbf{D}^n$ with $w\in c$ and $v\in c'$. So $W^m(w)\cap W^n(v)=\emptyset$ by (\ref{defFlower}).
\smallskip
(ii) Let $X^m \in \X^m$ be the unique $m$-tile with $X^n\subseteq X^m$. Depending on the location of $X^n$ in $X^m$, it suffices to prove statement~(ii) in the following cases:
\begin{enumerate}
\smallskip
\item[(1)] Assume that $X^n \subseteq \inte (X^m)$. Then $X^n \subseteq W^m(v^m)$ for any $v^m \in X^m \cap \V^m$.
\smallskip
\item[(2)] Assume that $\emptyset \neq X^n\cap e \subseteq \inte (e)$ for some $m$-edge $e\in \E^m$ with $e\subseteq X^m$. Then since no $1$-tile joins opposite sides of $\mathcal{C}$, by Proposition~\ref{propCellDecomp}(i), either $X^n \cap \partial X^m \subseteq \inte (e)$ or there exists $e'\in \E^m$ such that $X^n \cap \partial X^m \subseteq \inte (e) \cup \inte (e')$ and $e\cap e' = \{v\}$ for some $v\in\V^m$. In the former case, choose any $v^m \in e\cap \V^m$; and in the latter case, let $v^m=v$. Then $X^n \subseteq W^m(v^m)$.
\smallskip
\item[(3)] Assume $X^n\cap \V^m \neq \emptyset$. Since no $1$-tile joins opposite sides of $\mathcal{C}$, by Proposition~\ref{propCellDecomp}(i), there exists some $m$-vertex $v^m\in \V^m$ such that $X^n \cap \V^m =\{v^m\}$. Let $e,e'\in \E^m$ be the two $m$-edges that satisfy $e\cup e' \subseteq X^m$ and $e\cap e' = \{v^m\}$. Then by Proposition~\ref{propCellDecomp}(i) and the assumption that no $1$-tile joins opposite sides of $\mathcal{C}$, we get that $X^n\cap\partial X^m \subseteq \{v^m\} \cup \inte (e) \cup \inte (e')$. Thus $X^n \subseteq W^m(v^m)$.
\end{enumerate}
\smallskip
(iii) We observe that since no $1$-tile in $\X^1$ joins opposite sides of $\mathcal{C}$ and $f(\mathcal{C})\subseteq \mathcal{C}$, by Proposition~\ref{propCellDecomp}(i), each $(k+1)$-tile $X^{k+1}$ contains at most one $k$-vertex, for $k\in\N_0$.
Let $p,q\in\V^m$ be distinct. Then by Remark~\ref{rmFlower} and the observation above, we know $q\notin\overline{W}^n(p)$. So by part~(i), we get $W^n(p)\cap W^{n+1}(q) = \emptyset$. Since flowers are open sets, we have $W^n(p)\cap \overline{W}^{n+1}(q) = \emptyset$. It suffices to prove that $\overline{W}^{n+1}(p) \subseteq W^n(p)$. Indeed this inclusion is true; for otherwise, there exists an $(n+1)$-tile $X^{n+1}\subseteq \overline{W}^{n+1}(p)$ and a point $x\in \overline{W}^n(p) \setminus W^n(p)$ such that $\{x,p\}\subseteq X^{n+1}$. By (\ref{defFlower}) and applying Proposition~\ref{propCellDecomp}(i), we get a contradiction to the assumption that no $1$-tile in $\X^1$ joins opposite sides of $\mathcal{C}$.
\end{proof}
Let $f\: S^2\rightarrow S^2$ be an expanding Thurston map, and $\mathcal{C}\subseteq S^2$ a Jordan curve containing $\operatorname{post} f$ such that $f(\mathcal{C}) \subseteq \mathcal{C}$. We denote, for $m\in\N_0$, $n\in\N$, $q\in S^2$, and $q_i\in \V^m(f,\mathcal{C})$ for $i\in\{0,1,\dots,n-1\}$,
\begin{align} \label{eqDefEm}
& E_m(q_0,q_1,\dots,q_{n-1};q) \notag \\
= & \big\{ x\in f^{-n}(q) \,\big|\, f^i(x)\in \overline{W}^m(q_i), i\in \{0,1,\dots,n-1\} \big\}\\
= & f^{-n}(q) \cap \bigg( \bigcap\limits_{i=0}^{n-1} f^{-i}\(\overline{W}^m(q_i)\) \bigg), \notag
\end{align}
where $\overline{W}^m(q_i)$ is the closure of the $m$-flower $W^m(q_i)$ as defined in Section~\ref{sctThurstonMap}.
\begin{lemma} \label{lmCoverByFlowers}
Let $f\: S^2\rightarrow S^2$ be an expanding Thurston map, and $\mathcal{C}\subseteq S^2$ a Jordan curve containing $\operatorname{post} f$ such that $f(\mathcal{C}) \subseteq \mathcal{C}$. Then
\begin{equation}
\bigcap\limits_{i=0}^n f^{-i} (W^m(p_i)) \subseteq \bigcup\limits_{x\in E_m(p_0,p_1,\dots,p_{n-1};p_n)} W^{m+n} (x),
\end{equation}
for $m\in\N_0$, $n\in\N$, and $p_i\in\V^m(f,\mathcal{C})$ for $i\in \{0,1,\dots,n\}$. Here $E_m$ is defined in (\ref{eqDefEm}).
\end{lemma}
\begin{proof}
We prove the lemma by induction on $n\in\N$.
For $n=1$, we know that for all $p_0,p_1\in \V^m(f,\mathcal{C})$,
\begin{align*}
W^m(p_0) \cap f^{-1}(W^m(p_1))
\subseteq & \bigcup \big\{ W^{m+1}(x) \,\big|\, x\in f^{-1}(p_1),x\in\overline{W}^m(p_0) \big\} \\
= & \bigcup\limits_{x\in E_m(p_0;p_1)} W^{m+1}(x)
\end{align*}
by (\ref{eqDefEm}) and the fact that $W^{m+1}(x)\cap W^m(p_0) = \emptyset$ if both $x\in \V^{m+1}(f,\mathcal{C})$ and $x\notin \overline{W}^m(p_0)$ are satisfied (see Lemma~\ref{lmTileInFlower}(i)).
We now assume that the lemma holds for $n=l$ for some $l\in\N$.
We fix a point $p_i\in \V^m(f,\mathcal{C})$ for each $i\in \{0,1,\dots,l,l+1\}$. Then
\begin{equation*}
\bigcap\limits_{i=0}^{l+1} f^{-i} (W^m(p_i)) = W^m(p_0) \cap f^{-1} \bigg(\bigcap\limits_{i=1}^{l+1} f^{-(i-1)}(W^m(p_i))\bigg).
\end{equation*}
By induction hypothesis, the right-hand side of the above equation is a subset of
\begin{align*}
& W^m(p_0) \cap f^{-1} \bigg( \bigcup\limits_{x\in E_m(p_1,p_2,\dots,p_l;p_{l+1})} W^{m+l}(x) \bigg) \\
= & \bigcup\limits_{x\in E_m(p_1,p_2,\dots,p_l;p_{l+1})} \( W^{m}(p_0) \cap f^{-1}\(W^{m+l}(x)\) \) \\
\subseteq & \bigcup\limits_{x\in E_m(p_1,p_2,\dots,p_l;p_{l+1})} \( \bigcup\big\{ W^{m+l+1}(y) \,|\, y\in f^{-1}(x), y\in \overline{W}^m(p_0)\big\} \) \\
= & \bigcup\limits_{x\in E_m(p_1,p_2,\dots,p_l;p_{l+1})} \bigcup\limits_{y\in E_m(p_0;x)} W^{m+l+1}(y),
\end{align*}
where the last two lines is due to (\ref{eqDefEm}) and the fact that $W^{m+l+1}(y)\cap W^m(p_0) = \emptyset$ if both $y\in \V^{m+l+1}(f,\mathcal{C})$ and $y\notin \overline{W}^m(p_0)$ are satisfied (see Lemma~\ref{lmTileInFlower}(i)).
We claim that
\begin{equation*}
\bigcup\limits_{x\in E_m(p_1,p_2,\dots,p_l;p_{l+1})} E_m(p_0;x) = E_m(p_0,p_1,\dots,p_l; p_{l+1}).
\end{equation*}
Assuming the claim, we then get
\begin{equation*}
\bigcap\limits_{i=0}^{l+1} f^{-i} (W^m(p_i)) \subseteq \bigcup\limits_{x\in E_m(p_0,p_1,\dots,p_l;p_{l+1})} W^{m+l+1}(y).
\end{equation*}
Thus it suffices to prove the claim now. Indeed, by (\ref{eqDefEm}),
\begin{align*}
& \bigcup\limits_{x\in E_m(p_1,p_2,\dots,p_l;p_{l+1})} E_m(p_0;x) \\
= & \bigg\{ y\in f^{-1}(x) \,\bigg|\, y\in \overline{W}^m(p_0), x\in f^{-l}(p_{l+1}) \cap \bigg( \bigcap\limits_{i=1}^l f^{-i+1} \( \overline{W}^m(p_i) \) \bigg) \bigg\} \\
= & \bigg\{ y\in f^{-l-1}(p_{l+1}) \,\bigg|\, y\in \overline{W}^m(p_0), f(y) \in \bigcap\limits_{i=1}^l f^{-i+1} \( \overline{W}^m(p_i) \) \bigg\} \\
= & E_m(p_0,p_1,\dots,p_l;p_{l+1}).
\end{align*}
The induction step is now complete.
\end{proof}
\smallskip
We now review the notions of a simple directed graph and of a finite rooted tree that will be used in the proof of Theorem~\ref{thmAsympHExpWOPC}. Since the only purpose of such notions is to make the statements and proofs precise, and we will not use any nontrivial facts from graph theory, we adopt here a simplified approach to define relevant concepts as quickly as possible (compare \cite{BJG09}).
A \defn{simple directed graph} $\mathcal{G}=(\mathcal{V}(\mathcal{G}),\mathcal{E}(\mathcal{G}))$ is made up from a \defn{set of vertices} $\mathcal{V}(\mathcal{G})$ and a \defn{set of directed edges}
$$
\mathcal{E}(\mathcal{G}) \subseteq \mathcal{V}(\mathcal{G})\times \mathcal{V}(\mathcal{G}) \setminus \{(v,v)\,|\,v\in\mathcal{V}(\mathcal{G})\}.
$$
A simple directed graph $\mathcal{G}$ is \defn{finite} if $\card\mathcal{V}(\mathcal{G}) < +\infty$. Two \defn{vertices} $v,w\in \mathcal{V}(\mathcal{G})$ are \defn{connected by a directed edge $(v,w)$} if $(v,w)\in\mathcal{E}(\mathcal{G})$. If $e=(v,w)\in\mathcal{E}(\mathcal{G})$, then we call $v$ the \defn{initial vertex} of $e$, denoted by $i(e)$, and $w$ the \defn{terminal vertex} of $e$, denoted by $t(e)$. The \defn{indegree} of a vertex $v\in\mathcal{V}(\mathcal{G})$ is $d^-(v)=\card\{w\in\mathcal{V}(\mathcal{G}) \,|\, (w,v)\in \mathcal{E}(\mathcal{G}) \}$, and the \defn{outdegree} of $v$ is $d^+(v)=\card\{w\in\mathcal{V}(\mathcal{G}) \,|\, (v,w)\in \mathcal{E}(\mathcal{G}) \}$. A \defn{path} from a vertex $v\in\mathcal{V}(\mathcal{G})$ to a vertex $w\in\mathcal{V}(\mathcal{G})$ is a finite sequence of vertices $v=v_0,v_1,v_2,\dots,v_{n-1},v_n=w$ such that $(v_i,v_{i+1})\in\mathcal{E}(\mathcal{G})$ for each $i\in\{0,1,\dots,n-1\}$. The \defn{length} of such a path is $n$. The \defn{distance from $v$ to $w$} is the minimal length of all paths from $v$ to $w$. By convention, the distance from $v$ to $v$ is $0$, and if there is no path from $v$ to $w$ for $v\neq w$, then the distance from $v$ to $w$ is $\infty$. If the distance of $v$ to $w$ is $n\in\N_0$, then we say that $w$ is at a distance $n$ from $v$.
A finite simple directed graph $\T$ is a \defn{a finite rooted tree} if there exists a vertex $r\in\mathcal{V}(\T)$ such that for each vertex $v\in\mathcal{V}(\T)\setminus \{r\}$ there exists a unique path from $r$ to $v$. We call such a simple directed graph a \defn{finite rooted tree with root $r$}, and $r$ the \defn{root} of $\T$. Note that a finite rooted tree has a unique root. A vertex $v$ of a finite rooted tree $\T$ is called a \defn{leaf} (of $\T$) if $d^+(v)=0$. If $(v,w)\in \mathcal{E}(\T)$, then $w$ is said to be a \defn{child} of $v$.
\begin{lemma}[A bound for the number of leaves] \label{lmTree}
Let $\T$ be a finite rooted tree with root $r$ whose leaves are all at the same distance from $r$. Assume that there exist constants $c,k\in\N$ with the following properties:
\begin{enumerate}
\smallskip
\item[(i)] $d^+(x) \leq c$ for each vertex $x\in\mathcal{V}(\T)$,
\smallskip
\item[(ii)] for each leaf $v$, the number of vertices $w$ with $d^+(w) \geq 2$ in the path from $r$ to $v$ is at most $k$.
\end{enumerate}
Then then number of leaves of $\T$ is at most $c^k$.
\end{lemma}
\begin{proof}
Let $N\in\N_0$ be the distance from $r$ to any leaf of $\T$. For each $n\in\N_0$, we define $\mathcal{V}_n$ as the set of vertices of $\T$ at distance $n$ from $r$. It is clear that a vertex $v\in\mathcal{V}(\T)$ is a leaf of $\T$ if and only if $v\in \mathcal{V}_N$.
We can recursively construct a function $h\: \mathcal{V}(\T) \rightarrow \mathcal{L}$ by setting $h(r)=1$, and for each $v\in \mathcal{V}(\T)$, defining $h(v)= \frac{h(w)}{d^+(w)}$, where $w\in\mathcal{V}(\T)$ is the unique vertex with $(w,v)\in\mathcal{E}(\T)$. See Figure~\ref{figTreeSplit}.
\begin{figure}
\centering
\begin{overpic}
[width=6cm,
tics=20]{tree_eg.eps}
\put(64,132){$1$}
\put(64,110){$1$}
\put(-7,69){$\frac{1}{3}$}
\put(-7,35){$\frac{1}{3}$}
\put(0,-10){$\frac{1}{3}$}
\put(48,69){$\frac{1}{3}$}
\put(48,35){$\frac{1}{3}$}
\put(37,-10){$\frac{1}{6}$}
\put(66,-10){$\frac{1}{6}$}
\put(117,69){$\frac{1}{3}$}
\put(80,-10){$\frac{1}{6}$}
\put(73,35){$\frac{1}{6}$}
\put(141,36){$\frac{1}{6}$}
\put(104,-10){$\frac{1}{24}$}
\put(124,-10){$\frac{1}{24}$}
\put(139,-10){$\frac{1}{24}$}
\put(162,-10){$\frac{1}{24}$}
\end{overpic}
\caption{The function $h$ for a finite rooted tree.}
\label{figTreeSplit}
\end{figure}
By the two properties in the hypothesis, we have $h(v) \geq c^{-k}$ for each leaf $v \in \mathcal{V}(\T)$ of $\T$. On the other hand, it is easy to see from induction that $\sum\limits_{w\in \mathcal{V}_n} h(w) = 1$ for each $n\in\{0,1,\dots,N\}$. In particular, we have $\sum\limits_{w\in \mathcal{V}_N} h(w) = 1$. Thus $\card \mathcal{V}_N \leq c^k$. Therefore, the number of leaves of $\T$ is at most $c^k$.
\end{proof}
\subsection{Proof of Theorem~\ref{thmWeakExpansion}} \label{subsctProofWeakExp}
We split Theorem~\ref{thmWeakExpansion} into three parts and prove each one separately here.
\begin{theorem} \label{thmAsympHExpWOPC}
An expanding Thurston map $f\: S^2 \rightarrow S^2$ with no periodic critical points is asymptotically $h$-expansive.
\end{theorem}
\begin{proof}
We need to show $h^*(f)=0$. By (\ref{eqH*gn=nH*g}), it suffices to prove that $f^i$ is asymptotically $h$-expansive for some $i\in\N$. Note that by (\ref{eqLocalDegreeProduct}), $f^i$ has no periodic critical points for each $i\in\N$ if $f$ does not. Thus by Lemma~\ref{lmCexistsL}, we can assume, without loss of generality, that there exists a Jordan curve $\mathcal{C}\subseteq S^2$ containing $\operatorname{post} f$ such that $f(\mathcal{C})\subseteq\mathcal{C}$, and no $1$-tile joins opposite sides of $\mathcal{C}$. We consider the cell decompositions of $S^2$ induced by $f$ and $\mathcal{C}$ in this proof.
Recall that $\W^i$ defined in (\ref{defSetNFlower}) denotes the set of all $i$-flowers $W^i(p)$, $p\in\V^i$, for each $i\in\N_0$.
Since $f$ is expanding, it is easy to see from Lemma~\ref{lmCellBoundsBM}, Proposition~\ref{propCellDecomp}, and the Lebesgue Number Lemma (\cite[Lemma~27.5]{Mu00}) that $\{ \W^i \}_{i\in\N_0}$ forms a refining sequence of open covers of $S^2$ (see Definition~\ref{defRefSeqOpenCover}). Thus it suffices to prove that
\begin{align} \label{eqTailEntropy=0}
& h^*(f) = \\
& \lim\limits_{m\to+\infty} \lim\limits_{l\to+\infty} \lim\limits_{n\to+\infty} \frac{1}{n} H \( \bigvee\limits_{i=0}^{n-1} f^{-i}\(\W^l\) \Bigg| \bigvee\limits_{j=0}^{n-1} f^{-j}\(\W^m\) \) = 0. \notag
\end{align}
See (\ref{eqH(xi|eta)}) for the definition of $H$.
\smallskip
We now fix arbitrary $n,m,l\in N$ that satisfy $m+n>l>m$.
The plan for the proof is the following. We will first obtain an upper bound for the number of $(m+n-1)$-flowers needed to cover each element $A$ in the cover $\bigvee\limits_{j=0}^{n-1} f^{-j}\(\W^m\)$ of $S^2$. By Lemma~\ref{lmCoverByFlowers}, it suffices to find an upper bound for $\card E_m(p_0,p_1,\dots,p_{n-2};p_{n-1})$ for $p_0,p_1,\dots,p_{n-1}\in\V^m$. We identify $E_m(p_0,p_1,\dots,p_{n-2};p_{n-1})$ with the set of leaves of a certain rooted tree. By Lemma~\ref{lmTree}, we will only need to bound the number of vertices with more than one child in each path connecting the root with some leave. This can be achieved after one observes that for an expanding Thurston map with no periodic critical points, the frequency for an orbit getting near the set of critical points is bounded from above. After this main step, we will then find an upper bound for the number of $(l+n)$-tiles needed to cover $A$. By observing that each $(l+n)$-tile is a subset of some element in $\bigvee\limits_{j=0}^{n-1} f^{-j}\(\W^l\)$, we will finally obtain a suitable upper bound for $H \( \bigvee\limits_{i=0}^{n-1} f^{-i}\(\W^l\) \Bigg| \bigvee\limits_{j=0}^{n-1} f^{-j}\(\W^m\) \)$ which leads to (\ref{eqTailEntropy=0}).
\smallskip
Let $A\in \bigvee\limits_{j=0}^{n-1} f^{-j}(\W^m)$, say
\begin{equation} \label{eqA_WOPC}
A= \bigcap\limits_{i=0}^{n-1} f^{-i}(W^m(p_i))
\end{equation}
where $p_0,p_1,\dots,p_{n-1} \in \V^m$. By Lemma~\ref{lmCoverByFlowers},
\begin{equation}
A \subseteq \bigcup\limits_{x\in E_m(p_0,p_1,\dots,p_{n-2};p_{n-1})} W^{m+n-1}(x),
\end{equation}
where $E_m$ is defined in (\ref{eqDefEm}).
We can construct a rooted tree $\T$ from $E_m(p_0,p_1,\dots,p_{n-2}; p_{n-1})$ as a simple directed graph. The set $\mathcal{V}(\T)$ of vertices of $\T$ is
\begin{equation*}
\mathcal{V}(\T) = \bigcup\limits_{i=0}^{n-1} \big\{ (f^i(x), n-1-i) \in S^2\times \N_0 \,\big|\, x\in E_m(p_0,p_1,\dots,p_{n-2}; p_{n-1}) \big\}.
\end{equation*}
Two vertices $(x,i),(y,j)\in \mathcal{V}(\T)$ are connected by a directed edge $((x,i),(y,j)) \in \mathcal{E}(\V)$ if and only if $f(y)=x$ and $j=i+1$. Clearly the simple directed graph $\T$ constructed this way is a finite rooted tree with root $(p_{n-1},0)\in\mathcal{V}(\T)$.
Observe that if a vertex $(x,i)\in \mathcal{V}(\T)$ is a leaf of $\T$, then $x\in f^{-n+1}(p_{n-1})$ and $i=n-1$.
\begin{figure}
\centering
\begin{overpic}
[width=6cm,
tics=20]{tree_eg.eps}
\put(64,132){$(p_{n-1},0)$}
\put(141,36){$(v,3)$}
\end{overpic}
\caption{An example of $\T$ with $n=5$ and $c(v,3)=4$.}
\label{figTree_WOPC}
\end{figure}
For each $(x,i)\in\mathcal{V}(\T)$, we write $c(x,i) = d^+((x,i))$, i.e.,
\begin{equation} \label{eqDefC_NoChildren_WOPC}
c(x,i) = \card \{ (y,i+1) \in\mathcal{V}(\T) \,|\, f(y)=x \}.
\end{equation}
We make the convention that for each $x\in S^2$ and each $i\in\Z$, if $(x,i)\notin \mathcal{V}(\T)$, then $c(x,i)=-1$. See Figure~\ref{figTree_WOPC} for an example of $\T$.
Recall that by (\ref{eqDefEm}),
\begin{align*}
& E_m(p_0,p_1,\dots,p_{n-2};p_{n-1}) \\
= & \big\{ y\in f^{-n+1}(p_{n-1}) \,\big|\, f^i(y) \in \overline{W}^m(p_i),i\in\{0,1,\dots,n-2\} \big\}.
\end{align*}
So if $(x,i)\in\mathcal{V}(\T)$, then $c(x,i)$ is at most the number of distinct preimages of $x$ under $f$ contained in $\overline{W}^m(p_{i+1})$. Thus
\begin{equation} \label{eqBoundForNoChildren_WOPC}
0\leq c(x,i) \leq \deg f \text{ for } (x,i)\in\mathcal{V}(\T).
\end{equation}
Fix a visual metric $d$ on $S^2$ for $f$ with expansion factor $\Lambda >1$. The map $f$ is Lipschitz with respect to $d$ (see Lemma~\ref{lmLipschitz}). Then there exists a constant $K\geq 1$ depending only on $f$ and $d$ such that $d(f(x),f(y)) \leq K d(x,y)$ for $x,y\in S^2$. We may assume that $K\geq 2$.
Define
$$
N_c=\max \{\min \{i\in\N \,|\, f^j(x) \notin \operatorname{crit} f \text{ if } j\geq i\} \,|\, x\in \operatorname{crit} f \}.
$$
The maximum is taken over a finite set of integers since $f$ has no periodic critical points. So $N_c\in\N$. Note that by definition, if $x\in\operatorname{crit} f$, then $f^i(x) \in \operatorname{post} f\setminus \operatorname{crit} f$ for each $i\geq N_c$. Denote the shortest distance between a critical point and the set $\operatorname{post} f\setminus \operatorname{crit} f$ by
$$
D_c=\min \{d(x,y) \,|\, x\in \operatorname{post} f\setminus \operatorname{crit} f, y\in \operatorname{crit} f \}.
$$
Then $D_c \in (0,+\infty)$ since both $\operatorname{post} f \setminus \operatorname{crit} f$ and $\operatorname{crit} f$ are nonempty finite sets.
\smallskip
We now proceed to find an upper bound for
\begin{equation*}
\card \big\{ i\in \{0,1,\dots,n-1\} \,\big|\, c(f^i(z),n-1-i) \geq 2 \big\}
\end{equation*}
for each $(z,n-1)\in\mathcal{V}(\T)$, uniform in $(z,n-1)$. Recall that $z\in f^{-n+1} (p_{n-1})$ for each $(z,n-1)\in\mathcal{V}(\T)$. We fix such a point $z$.
\begin{figure}
\centering
\begin{overpic}
[width=12cm,
tics=20]{tree.eps}
\put(210,332){$p_{n-1}$}
\put(305,330){$\overline{W}^m(p_{n-1})$}
\put(305,305){$\overline{W}^m(p_{n-2})$}
\put(305,278){$\overline{W}^m(p_{n-3})$}
\put(280,175){$\overline{W}^m(p_{k})$}
\put(300,138){$\overline{W}^m(p_{k-1})$}
\put(320,43){$\overline{W}^m(p_{1})$}
\put(320,-5){$\overline{W}^m(p_{0})$}
\put(235,278){$f^{n-3}(z)$}
\put(220,175){$f^k(z)$}
\put(207,145){$f^{k-1}(z)$}
\put(225,3){$z$}
\put(40,270){$*_3$}
\put(267,140){$*_2$}
\put(-7,7){$*_1$}
\put(80,8){$_{<r}$}
\put(243,138){$_{<r}$}
\put(135,285){$_{<r}$}
\end{overpic}
\caption{$*_1,*_2,*_3\in\operatorname{crit} f$, $r=\tau(3C\Lambda^{-m})$, and $c(f^k(z),n-1-i)=2$.}
\label{figLocInjectAwayFromCrit_WOPC}
\end{figure}
In order to find an upper bound, we first define, for each $i\in\N$ sufficiently large,
\begin{equation} \label{eqDefMm_WOPC}
M_i = \bigg\lfloor \log_K \(\frac{D_c - \tau(3C\Lambda^{-i})}{\tau(3C\Lambda^{-i})} \) \bigg\rfloor -2,
\end{equation}
where the function $\tau$ is from Lemma~\ref{lmLocInjectAwayFromCrit}, and $C\geq 1$ is a constant depending only on $f$, $\mathcal{C}$, and $d$ from Lemma~\ref{lmCellBoundsBM}. Note that $\tau(3C\Lambda^{-i})\longrightarrow 0$ as $i\longrightarrow+\infty$ (Lemma~\ref{lmLocInjectAwayFromCrit}), thus $M_i$ is well-defined for $i$ sufficiently large, and
\begin{equation} \label{eqM_mToInfty_WOPC}
\lim\limits_{i\to+\infty} M_i = +\infty.
\end{equation}
We assume that $m$ is sufficiently large such that the following conditions are both satisfied:
\begin{enumerate}
\smallskip
\item[(i)] $m > \log_\Lambda \big(\frac{3C}{\delta_0}\big)$,
\smallskip
\item[(ii)] $M_m > N_c$,
\end{enumerate}
where $\delta_0\in(0,1]$ is a constant that depends only on $f$ and $d$ from Lemma~\ref{lmLocInjectAwayFromCrit}. Note that by Lemma~\ref{lmCellBoundsBM}, each $m$-flower is of diameter at most $2C\Lambda^{-m}$. Thus condition~(i) implies that for each $v\in\V^m$, each pair of points $x,y \in \overline{W}^m(v)$ satisfy $d(x,y)< 3C\Lambda^{-m} < \delta_0$.
Fix $k\in\{0,1,\dots,n-1\}$ with $c\(f^k(z),n-1-k\) \geq 2$. Then $k\neq 0$ and the number of distinct points in $\overline{W}^m(p_{k-1})$ that are mapped to $f^k(z)$ under $f$ is at least $c\(f^k(z),n-1-k\) \geq 2$. Thus $f$ is not injective on $\overline{W}^m(p_{k-1})$. See Figure~\ref{figLocInjectAwayFromCrit_WOPC}. By Lemma~\ref{lmCellBoundsBM}, $\diam_d\(\overline{W}^m(p_{k-1})\)\leq 2 C\Lambda^{-m}$. Since $f^{k-1}(z)\in \overline{W}^m(p_{k-1})$, the map $f$ is not injective on $B_d\(f^{k-1}(z),3C\Lambda^{-m}\)$. Then since $3C\Lambda^{-m}<\delta_0$, by Lemma~\ref{lmLocInjectAwayFromCrit},
\begin{equation*}
d\(f^{k-1}(z),\operatorname{crit} f\) < \tau\(3C\Lambda^{-m}\).
\end{equation*}
Choose $w\in\operatorname{crit} f$ that satisfies
$
d\(f^{k-1}(z),w\) < \tau\(3C\Lambda^{-m}\).
$
Then for each $j\in \N_0$,
\begin{equation} \label{eqPfThmAsympHExp1_WOPC}
d\(f^{k+j-1}(z),f^j(w) \) < K^j \tau\(3C\Lambda^{-m}\).
\end{equation}
We will show that in the sequence $f^k(z),f^{k+1}(z),\dots,f^{k+M_m}(z)$, the number of terms $f^{k+j}(z)$, $0\leq j\leq M_m$, for which the vertex
$$
\(f^{k+j}(z), n-1-k-j\) \in \mathcal{V}(\T)
$$
has at least two children is bounded above by $N_c$, i.e.,
\begin{equation} \label{eqNcEveryMm_WOPC}
\card \big\{ j\in\{0,1,\dots,M_m\} \,\big|\, c\(f^{k+j}(z), n-1-k-j\)\geq 2 \big\} \leq N_c.
\end{equation}
Note that $M_m$ is defined in (\ref{eqDefMm_WOPC}). Here we use the convention that for each $x\in S^2$ and each $i\in \Z$, if $(x,i)\notin \mathcal{V}(\T)$, then $c(x,i)=-1$.
Indeed, for each $j\in\{N_c,N_c +1,\dots,\min\{M_m,n-1-k\}\}$, we have $f^j(w)\in \operatorname{post} f\setminus \operatorname{crit} f$. Note that here $M_m>N_c$ by condition~(ii) on $m$. Thus by (\ref{eqPfThmAsympHExp1_WOPC}) and (\ref{eqDefMm_WOPC}),
\begin{align*}
d\(f^{k+j-1}(z),\operatorname{crit} f \) & \geq d\(\operatorname{crit} f, f^j(w)\) - d\( f^j(w), f^{k+j-1}(z) \)\\
& \geq D_c - K^j \tau( 3C\Lambda^{-m}) \\
& \geq D_c - K^{M_m} \tau( 3C\Lambda^{-m}) \\
& \geq D_c - \( \frac{D_c - \tau(3C\Lambda^{-m})}{\tau(3C\Lambda^{-m})} \) \tau(3C\Lambda^{-m} )\\
& = \tau(3C\Lambda^{-m} ).
\end{align*}
Hence by Lemma~\ref{lmLocInjectAwayFromCrit}, the restriction of $f$ to $B_d\(f^{k+j-1}(z), 3C\Lambda^{-m}\)$ is injective. Note that $f^{k+j-1}(z) \in \overline{W}^m(p_{k+j-1})$, and by Lemma~\ref{lmCellBoundsBM}, $\diam_d \(\overline{W}^m(p_{k+j-1}) \) \leq 2C\Lambda^{-m}$. So $f$ is injective on $\overline{W}^m(p_{k+j-1})$. Thus
$$
c\(f^{k+j}(z), n-1-k-j\) = 1
$$
for each $j\in\{N_c,N_c +1,\dots,\min\{M_m,n-1-k\}\}$. Hence
$$
c\(f^{k+j}(z), n-1-i-j\) \in \{1,-1\}
$$
for each $j\in\{N_c,N_c +1,\dots,M_m\}$. Then (\ref{eqNcEveryMm_WOPC}) holds.
Thus we get that
\begin{equation} \label{eqSplitNoBound_WOPC}
\card \big\{ i\in \{0,1,\dots,n-1\} \,\big|\, c(f^i(z),n-1-i) \geq 2 \big\} \leq N_c \bigg\lceil \frac{n}{M_m} \bigg\rceil
\end{equation}
for each $(z,n-1)\in\mathcal{V}(\T)$.
Hence by (\ref{eqSplitNoBound_WOPC}), (\ref{eqBoundForNoChildren_WOPC}), and Lemma~\ref{lmTree}, we can conclude that the number of leaves of $\T$ is at most $(\deg f)^{N_c \(\frac{n}{M_m} +1\)}$, or equivalently,
\begin{equation}
\card E_m(p_0,p_1,\dots,p_{n-2};p_{n-1}) \leq (\deg f)^{N_c \(\frac{n}{M_m} +1\)}.
\end{equation}
We have obtained an upper bound for the number of $(m+n-1)$-flowers needed to cover $A$. Next, we will find an upper bound for the number of $(m+n-1)$-tiles, and consequently, an upper bound for the number of $(l+n)$-tiles, needed to cover $A$.
Denote the maximum number of $i$-tiles contained in the closure of any $i$-flower, over all $i\in\N_0$, by $W_f$, i.e.,
\begin{equation*}
W_f = \sup \big\{ \card \big\{ X^i\in\X^i \,\big|\, X^i\subseteq \overline{W}^j(v) \big\} \,\big|\, j\in\N_0,v\in\V^j \big\}.
\end{equation*}
Observe that $W_f = \sup \{ 2 \deg_{f^i}(v) \,|\, i\in\N_0, \, v\in\V^i \}$. Since $f$ has no periodic critical points, it follows from \cite[Lemma~17.1]{BM10} that $W_f$ is a finite number that only depends on $f$.
Thus we can cover $A$ in (\ref{eqA_WOPC}) by a collection of $(m+n-1)$-tiles of cardinality at most $W_f (\deg f)^{N_c \(\frac{n}{M_m} +1\)}$.
\smallskip
On the other hand, we claim that each $(l+n)$-tile $X^{l+n}\in \X^{l+n}$ is a subset of at least one element in the open cover $\bigvee\limits_{i=0}^{n-1} f^{-i} (\W^l)$ of $S^2$. To prove the claim, we first fix an $(l+n)$-tile $X^{l+n} \in \X^{l+n}$. By Proposition~\ref{propCellDecomp}(ii) and Lemma~\ref{lmTileInFlower}(ii), for each $i\in\{0,1,\dots,n-1\}$, there exists an $l$-vertex $v_i\in \V^l$ such that $f^i\(X^{l+n}\) \subseteq W^l(v_i)$. Thus
$$
X^{l+n} \subseteq \bigcap\limits_{i=0}^{n-1} f^{-i} \( W^l(v_i)\).
$$
The proof for the claim is complete.
\smallskip
Note that for each $(m+n-1)$-tile $X^{m+n-1}\in \X^{m+n-1}$, the collection
$$
\big\{ X^{l+n} \in \X^{l+n} \,\big|\, X^{l+n}\subseteq X^{m+n-1} \big\}
$$
forms a cover of $X^{m+n-1}$, and has cardinality at most $(2\deg f)^{l-m+1}$, which follows immediately from Proposition~\ref{propCellDecomp}.
Hence, we get that for each element $A$ of $\bigvee\limits_{j=0}^{n-1} f^{-j} (\W^m)$, we can find a cover of $A$ consisting of elements of $\bigvee\limits_{i=0}^{n-1} f^{-i} (\W^l)$ in such a way that the cardinality of the cover is at most $(2\deg f)^{l-m+1} W_f (\deg f)^{N_c \(\frac{n}{M_m} +1\)}$.
We conclude that
\begin{align*}
h^*(f) & = \lim\limits_{m\to+\infty} \lim\limits_{l\to+\infty} \lim\limits_{n\to+\infty} \frac{1}{n} \log \( (2\deg f)^{l-m+1} W_f (\deg f)^{N_c \(\frac{n}{M_m} +1\)} \) \\
& = \lim\limits_{m\to+\infty} \lim\limits_{l\to+\infty} \lim\limits_{n\to+\infty} \frac{1}{n} N_c \(\frac{n}{M_m} +1\) \log(\deg f) \\
& = \lim\limits_{m\to+\infty} \frac{N_c \log(\deg f)}{M_m} \\
& = 0.
\end{align*}
The last equality follows from (\ref{eqM_mToInfty_WOPC}).
\end{proof}
Recall that a point $x\in S^2$ is a periodic point of $f\: S^2\rightarrow S^2$ with period $n$ if $f^n(x)=x$ and $f^i(x)\neq x$ for each $i\in\{1,2,\dots,n-1\}$.
\begin{theorem} \label{thmNotAsympHExp}
An expanding Thurston map $f\: S^2\rightarrow S^2$ with at least one periodic critical point is not asymptotically $h$-expansive.
\end{theorem}
\begin{proof}
We need to show $h^*(f)>0$. By (\ref{eqH*gn=nH*g}), it suffices to prove that $f^i$ is not asymptotically $h$-expansive for some $i\in\N$. Note that by (\ref{eqLocalDegreeProduct}), if a point $x\in S^2$ is a periodic critical point of $f^i$ for some $i\in\N$, then it is a periodic point of $f$ and there exists $j\in\N_0$ such that $f^j(x)$ is a periodic critical point of $f$. Thus each periodic critical point of $f^\tau$ is a fixed point of $f^\tau$ if $\tau\in\N$ is a common multiple of the periods of all the periodic critical points of $f$. Hence by Lemma~\ref{lmCexistsL}, we can assume, without loss of generality, that there exists a Jordan curve $\mathcal{C}\subseteq S^2$ containing $\operatorname{post} f$ such that $f(\mathcal{C})\subseteq\mathcal{C}$, and no $1$-tile joins opposite sides of $\mathcal{C}$, and each periodic critical point of $f$ is a fixed point of $f$.
Let $p$ be a critical point of $f$ that is fixed by $f$.
In addition, we can assume, without loss of generality, that $f^{-1}(p)\setminus \mathcal{C} \neq \emptyset$. Indeed, by Lemma~\ref{lmPreImageDense}, there exists $j\in\N$ such that $f^{-j}(p)\setminus \mathcal{C} \neq \emptyset$. We replace $f$ by $f^j$, and observe that by (\ref{eqLocalDegreeProduct}) and the fact that each periodic critical point of $f$ is a fixed point of $f$, the set of periodic critical points of $f$ and that of $f^j$ coincide. Note that for the new map and its invariant curve $\mathcal{C}$, no $1$-tile joins opposite sides of $\mathcal{C}$, and each periodic critical point is a fixed point.
From now on, we consider the cell decompositions of $S^2$ induced by $f$ and $\mathcal{C}$ in this proof.
\smallskip
Recall that for $i\in\N_0$, we denote by $\W^i$ as in (\ref{defSetNFlower}) the set of all $i$-flowers $W^i(p)$ where $p\in\V^i$.
Since $f$ is expanding, it is easy to see from Lemma~\ref{lmCellBoundsBM}, Proposition~\ref{propCellDecomp}, and the Lebesgue Number Lemma (\cite[Lemma~27.5]{Mu00}) that $\{ \W^i \}_{i\in\N_0}$ forms a refining sequence of open covers of $S^2$ (see Definition~\ref{defRefSeqOpenCover}). Thus it suffices to prove that
\begin{equation*}
h^*(f) = \lim\limits_{m\to+\infty} \lim\limits_{l\to+\infty} \lim\limits_{n\to+\infty} \frac{1}{n} H \( \bigvee\limits_{i=0}^{n-1} f^{-i}\(\W^l\) \Bigg| \bigvee\limits_{j=0}^{n-1} f^{-j}\(\W^m\) \) > 0.
\end{equation*}
See (\ref{eqH(xi|eta)}) for the definition of $H$.
Our plan is to construct a sequence $\{v_i\}_{i\in\N}$ of $m$-vertices such that for each $n\in\N$, the number of elements in $\bigvee\limits_{i=0}^{n-1} f^{-i}\(\W^l\)$ needed to cover $B_n=\bigcap\limits_{j=0}^{n-1} f^{-j} (W^m(v_{n-j}))$ can be bounded from below in such a way that $h^*(f)>0$ follows immediately. More precisely, we observe that the more connected components $B_n$ has, the harder to cover $B_n$. So we will choose $\{v_i\}_{i\in\N}$ as a periodic sequence of $m$-vertices shadowing an infinite backward pseudo-orbit under iterations of $f$ in such a way that each period of $\{v_i\}_{i\in\N}$ begins with a backward orbit starting at $p$ and approaching $p$ as the index $i$ increases, and then ends with a constant sequence staying at $p$. By a recursive construction, we keep track of each $B_n$ by a finite subset $V_n\subseteq B_n$ with the property that $\card(A\cap V_n)\leq 1$ for each $A \in \bigvee\limits_{i=0}^{n-1} f^{-i}\(\W^l\)$. A quantitative control of the size of $V_n$ leads to the conclusion that $h^*(f)>0$. The fact that the constant part of each period of $\{v_i\}_{i\in\N}$ can be made arbitrarily long is essential here and is not true if $f$ has no periodic critical points.
\smallskip
For this we fix $m,l\in\N$ with $l>m+100$.
Let $k=\deg_f(p)$. Then $k>1$.
Define $q_0 = p$ and choose $q_1\in f^{-1}(p)\setminus \mathcal{C}$. Then $q_1$ is necessarily a $1$-vertex, but not a $0$-vertex, i.e., $q_1\in\V^1\setminus \V^0$. Since $q_1\notin \mathcal{C}$, we have $q_1\in W^0(p)$. By (\ref{defFlower}), the only $2$-vertex contained in $W^2(p)$ is $p$. So $q_1\in W^0(p)\setminus W^2(p)$. Since $f\(W^i(p)\) = W^{i-1}(p)$ for each $i\in\N$ (see Remark~\ref{rmFlower}), we can recursively choose $q_j\in\V^j$ for $j\in\{2,3,\dots,m\}$ such that
\begin{enumerate}
\smallskip
\item[(i)] $f(q_j) = q_{j-1}$,
\smallskip
\item[(ii)] $q_j\in W^{j-1}(p) \setminus W^{j+1}(p)$.
\end{enumerate}
We define a singleton set $Q_j = \{ q_j \}$.
We set $q_m^1=q_m$.
Next, we choose recursively, for each $j\in\{m+1,m+2,\dots, l-2\}$, a set $Q_j$ with $\card Q_j = k^{j-m}$ consisting of distinct points $q_j^i\in\V^j$, $i\in\{1,2,\dots,k^{j-m}\}$, such that
\begin{enumerate}
\smallskip
\item[(i)] $f(Q_j)= Q_{j-1}$,
\smallskip
\item[(ii)] $Q_j \subseteq W^{j-1}(p)\setminus W^{j+1}(p)$.
\end{enumerate}
Note by Remark~\ref{rmFlower}, it is clear that these two properties uniquely determines $Q_j$ from $Q_{j-1}$.
Finally, we construct recursively, for $j\in\{l-1,l,l+1\}$, a set $Q_j$ with $\card Q_j = k^{l-2-m}$ consisting of distinct points $ q_j^i\in\V^j$, $i\in\big\{1,2,\dots, k^{l-2-m} \big\}$, such that
\begin{enumerate}
\smallskip
\item[(i)] $f(q_j^i)= q_{j-1}^i$,
\smallskip
\item[(ii)] $Q_j \subseteq W^{j-1}(p)\setminus W^{j+1}(p)$.
\end{enumerate}
\smallskip
We will now construct recursively, for each $n=(l+1)s+r$, with $s\in\N_0$ and $r\in\{0,1,\dots,l\}$, an $m$-vertex $v_n\in\V^m$ and a set of $n$-vertices $V_n\subseteq \V^n$ such that the following properties are satisfied:
\begin{enumerate}
\smallskip
\item[(1)] $V_n\subseteq W^m(v_n)$ for $n\in \N_0$;
\smallskip
\item[(2)] $f\(V_n\) = V_{n-1}$ for $n\in\N$;
\smallskip
\item[(3)] For $s\in\N_0$, and
\begin{enumerate}
\smallskip
\item[(i)] for $r=0$, $V_{(l+1)s+r} \subseteq W^l(p)$,
\smallskip
\item[(ii)] for $r\in\{1,2,\dots, m\}$, $V_{(l+1)s+r} \subseteq W^{l+1}\(v_{(l+1)s+r}\)$,
\smallskip
\item[(iii)] for $r\in\{m+1,m+2,\dots, l-2\}$, there exists, for each $i\in\{1,2,\dots,k^{r-m}\}$, a subset $V^i_{(l+1)s+r}$ of $V_{(l+1)s+r}$ such that
\begin{enumerate}
\smallskip
\item[(a)] $V^i_{(l+1)s+r} \cap V^j_{(l+1)s+r} = \emptyset$ for $1\leq i< j\leq k^{r-m}$,
\smallskip
\item[(b)] $\bigcup\limits_{i=1}^{k^{r-m}} V^i_{(l+1)s+r} = V_{(l+1)s+r}$,
\smallskip
\item[(c)] $V^i_{(l+1)s+r} \subseteq W^{l+1} \(q_r^i\)$,
\end{enumerate}
\smallskip
\item[(iv)] for $r\in\{l-1,l\}$, there exists, for each $i\in\{1,2,\dots,k^{l-2-m}\}$, a subset $V^i_{(l+1)s+r}$ of $V_{(l+1)s+r}$ such that
\begin{enumerate}
\smallskip
\item[(a)] $V^i_{(l+1)s+r} \cap V^j_{(l+1)s+r} = \emptyset$ for $1\leq i< j\leq k^{l-2-m}$,
\smallskip
\item[(b)] $\bigcup\limits_{i=1}^{k^{l-2-m}} V^i_{(l+1)s+r} = V_{(l+1)s+r}$,
\smallskip
\item[(c)] $V^i_{(l+1)s+r} \subseteq W^{l+1} \(q_r^i\)$;
\end{enumerate}
\end{enumerate}
\smallskip
\item[(4)] for $n\in\N_0$, $A\in\bigvee\limits_{i=0}^{n-1} f^{-i}\(\W^l\)$, and $x,y\in V_n$ with $x\neq y$, we have $\{x,y\}\nsubseteq A$.
\end{enumerate}
\smallskip
We start our construction by first defining $v_n\in \V^m$ for each $n\in\N$. For $s\in\N_0$ and $r\in\{0,1,\dots, m\}$, set $v_{(l+1)s+r} = q_r$. For $s\in\N_0$ and $r\in\{m+1,m+2,\dots, l\}$, set $v_{(l+1)s+r} = p$.
We now define $V_n$ recursively.
Let $V_0= \{q_0\}$. Clearly $V_0$ satisfies properties~(1) through (4).
Assume that $V_n$ is defined and satisfies properties~(1) through (4) for each $n\in\{0,1,\dots, (l+1)s+r \}$, where $s\in\N_0$ and $r\in\{0,1,\dots,l\}$. We continue our construction in the following cases depending on $r$.
\smallskip
\emph{Case 1.} Assume $r\in\{0,1,\dots,m-1\}$. Then $v_{(l+1)s+r}=q_r$ and $v_{(l+1)s+r+1}=q_{r+1}$.
Since $f\(W^{l+1}\(q_{r+1}\)\) = W^l\(q_r\)$ (see Remark~\ref{rmFlower}), and $V_{(l+1)s+r}\subseteq W^l(q_r)$ by the induction hypothesis, we can choose, for each $x\in V_{(l+1)s+r}$, a point $x'\in W^{l+1}(q_{r+1})$ such that $f(x')=x$. Then define $V_{(l+1)s+r+1}$ to be the collection of all such chosen $x'$ that corresponds to $x\in V_{(l+1)s+r}$. Note that
$$
\card V_{(l+1)s+r+1} = \card V_{(l+1)s+r}.
$$
All properties required for $V_{(l+1)s+r+1}$ in the induction step are trivial to verify. We only consider the last property here. Indeed, suppose that $x,y\in V_{(l+1)s+r+1}$ satisfy that $x\neq y$ and $\{x,y\}\subseteq A$ for some $A\in\bigvee\limits_{i=0}^{(l+1)s+r} f^{-i}\(\W^l\)$. Then by construction $f(x),f(y)$, and $f(A)$ satisfy
\begin{enumerate}
\smallskip
\item[(a)] $f(A)\subseteq B$ for some $B\in \bigvee\limits_{i=0}^{(l+1)s+r-1} f^{-i}\(\W^l\)$,
\smallskip
\item[(b)] $f(x), f(y)\in V_{(l+1)s+r}$, and $f(x)\neq f(y)$,
\smallskip
\item[(c)] $\{f(x),f(y)\} \subseteq f(A)\subseteq B$.
\end{enumerate}
This contradicts property~(4) for $V_{(l+1)s+r}$ in the induction hypothesis.
\smallskip
\emph{Case 2.} Assume $r\in\{m,m+1,\dots,l-3\}$. Then $v_{(l+1)s+r+1}=p$, $v_{(l+1)s+m} = q_m$, and when $r\neq m$, we have $v_{(l+1)s+r}=p$.
If $r=m$, we define $V^1_{(l+1)s+r}=V_{(l+1)s+r}$. Recall that $q_m^1=q_m$.
Note that for each $i\in\{1,2,\dots,k^{r+1-m}\}$, $f\(W^{l+2}\(q_{r+1}^i\)\) = W^{l+1}\(q_r^j\)$ for some $j\in\{1,2,\dots,k^{r-m}\}$ (see Remark~\ref{rmFlower}), and $V^j_{(l+1)s+r} \subseteq W^{l+1}\(q^j_r\)$ by the induction hypothesis. For each $j\in\{1,2,\dots,k^{r-m}\}$, each $x\in V^j_{(l+1)s+r}$, and each $i\in\{1,2,\dots,k^{r+1-m}\}$ with $f\(W^{l+2}\(q_{r+1}^i\)\) = W^{l+1}\(q_r^j\)$, we can choose a point $x'\in W^{l+2}\(q_{r+1}^i\)$ such that $f(x')=x$. Then define $V^i_{(l+1)s+r+1}$ to be the collection of all such chosen $x'$ that corresponds to $x\in V^j_{(l+1)s+r}$. Set $V_{(l+1)s+r+1}= \bigcup\limits_{i=1}^{k^{r+1-m}} V^i_{(l+1)s+r+1}$.
Since $Q_{r+1} \subseteq \V^{r+1} \cap W^m(p)$, $r\in\{m,m+1,\dots,l-3\}$, $l>m+100$, and no $1$-tile joins opposite sides of $\mathcal{C}$, we get that
\begin{enumerate}
\smallskip
\item[(a)] for $i,j\in\{ 1,2,\dots,k^{r+1-m}\}$ with $i\neq j$, by Lemma~\ref{lmTileInFlower}(iii),
$$
W^{l+2}\(q^i_{r+1}\) \cap W^{l+2}\(q^j_{r+1}\) =\emptyset,
$$
and so $V^i_{(l+1)s+r+1} \cap V^j_{(l+1)s+r+1} = \emptyset$,
\smallskip
\item[(b)] $V_{(l+1)s+r+1} \subseteq W^m(p)$.
\end{enumerate}
Thus
$$
\card V_{(l+1)s+r+1} = k \card V_{(l+1)s+r}.
$$
We only need to verify property~(4) required for $V_{(l+1)s+r+1}$ in the induction step now. Indeed, suppose that $x,y\in V_{(l+1)s+r+1}$ with $x\neq y$ and $\{x,y\}\subseteq A$ for some $A\in\bigvee\limits_{a=0}^{(l+1)s+r} f^{-a}\(\W^l\)$. Then $A\subseteq W^l(v^l)$ for some $v^l\in \V^l$. By construction, there exist $i,j \in \{1,2,\dots,k^{r+1-m}\}$ such that $x\in W^{l+2}\(q^i_{r+1}\)$ and $y\in W^{l+2}\(q^j_{r+1}\)$. Note that $q^i_{r+1}, q^j_{r+1} \in \V^{r+1}$, $r\in\{m,m+1,\dots,l-3\}$, and $l>m+100$. So $q^i_{r+1}, q^j_{r+1} \in \V^{l-2}$. Since $x\in W^l(v^l)\cap W^{l+2}\(q^i_{r+1}\)$, we get $q^i_{r+1}\in \overline{W}^l(v_l)$ by Lemma~\ref{lmTileInFlower}(i), and thus $v^l\in \overline{W}^l(q^i_{r+1})$. Similarly $v^l\in \overline{W}^l(q^j_{r+1})$. Since $q^i_{r+1}, q^j_{r+1} \in \V^{l-2}$ and no $1$-tile joins opposite sides of $\mathcal{C}$, we get from Lemma~\ref{lmTileInFlower}(iii) that $q^i_{r+1}=q^j_{r+1}$, i.e., $i=j$. Thus $f(x)\neq f(y)$ by construction. But then $f(x),f(y)$, and $f(A)$ satisfy
\begin{enumerate}
\smallskip
\item[(a)] $f(A)\subseteq B$ for some $B\in \bigvee\limits_{a=0}^{(l+1)s+r-1} f^{-a}\(\W^l\)$,
\smallskip
\item[(b)] $f(x), f(y)\in V_{(l+1)s+r}$, and $f(x)\neq f(y)$,
\smallskip
\item[(c)] $\{f(x),f(y)\} \subseteq f(A)\subseteq B$.
\end{enumerate}
This contradicts property~(4) for $V_{(l+1)s+r}$ in the induction hypothesis.
\smallskip
\emph{Case 3.} Assume $r\in\{l-2,l-1,l\}$, then $v_{(l+1)s+r+1}=v_{(l+1)s+r}=p$.
Note that for each $i\in\{1,2,\dots,k^{l-2-m}\}$, $f\(W^{l+2}\(q_{r+1}^i\)\) = W^{l+1}\(q_r^i\)$ (see Remark~\ref{rmFlower}), and $V^i_{(l+1)s+r} \subseteq W^{l+1}\(q^i_r\)$ by the induction hypothesis. For each $j\in\{1,2,\dots,k^{l-2-m}\}$ and each $x\in V^i_{(l+1)s+r}$, we can choose a point $x'\in W^{l+2}\(q_{r+1}^i\)$ such that $f(x')=x$. Then define $V^i_{(l+1)s+r+1}$ to be the collection of all such chosen $x'$ that corresponds to $x\in V^i_{(l+1)s+r}$. Set $V_{(l+1)s+r+1}= \bigcup\limits_{i=1}^{k^{l-2-m}} V^i_{(l+1)s+r+1}$.
Since $Q_{r+1} \subseteq \V^{r+1} \cap W^r(p)$, $r\in\{l-2,l-1,l\}$, and $l>m+100$, we get that
\begin{enumerate}
\smallskip
\item[(a)] for $i,j\in\{ 1,2,\dots,k^{l-2-m}\}$ with $i\neq j$,
$$
f\big(V^i_{(l+1)s+r+1}\big) \cap f\big(V^j_{(l+1)s+r+1}\big) = V^i_{(l+1)s+r} \cap V^j_{(l+1)s+r}=\emptyset
$$
(by the induction hypothesis), and so
$$
V^i_{(l+1)s+r+1} \cap V^j_{(l+1)s+r+1} = \emptyset,
$$
\smallskip
\item[(b)] $V_{(l+1)s+r+1} \subseteq W^m(p)$,
\smallskip
\item[(c)] if $r=l$, then $V_{(l+1)s+r+1} \subseteq W^l(p)$.
\end{enumerate}
Thus
$$
\card V_{(l+1)s+r+1} = \card V_{(l+1)s+r}.
$$
We only need to verify the last property required for $V_{(l+1)s+r+1}$ in the induction step now. Indeed, suppose that $x,y\in V_{(l+1)s+r+1}$ with $x\neq y$ and $\{x,y\}\subseteq A$ for some $A\in\bigvee\limits_{i=0}^{(l+1)s+r} f^{-i}\(\W^l\)$. Then by construction $f(x),f(y)$, and $f(A)$ satisfy
\begin{enumerate}
\smallskip
\item[(a)] $f(A)\subseteq B$ for some $B\in \bigvee\limits_{i=0}^{(l+1)s+r-1} f^{-i}\(\W^l\)$,
\smallskip
\item[(b)] $f(x), f(y)\in V_{(l+1)s+r}$, and $f(x)\neq f(y)$,
\smallskip
\item[(c)] $\{f(x),f(y)\} \subseteq f(A)\subseteq B$.
\end{enumerate}
This contradicts property~(4) for $V_{(l+1)s+r}$ in the induction hypothesis.
\smallskip
The recursive construction and the inductive proof of the properties of the construction are now complete.
Note that by our construction, we have
\begin{equation} \label{eqCardUn}
\card V_{(l+1)s} = k^{(l-m-2)s}, \qquad s\in \N.
\end{equation}
\smallskip
For each $s\in\N$, we consider
$$
B_{(l+1)s}= \bigcap\limits_{j=0}^{(l+1)s-1} f^{-j} \(W^m\(v_{(l+1)s - j}\) \) \in \bigvee\limits_{j=0}^{(l+1)s-1} f^{-j} \(\W^m \).
$$
Then $V_{(l+1)s} \subseteq B_{(l+1)s}$ by properties~(1) and (2) of the construction. On the other hand, by property~(4), if $\mathcal{A}\subseteq \bigvee\limits_{j=0}^{(l+1)s-1} f^{-j} \(\W^l \)$ satisfies
$$
\bigcup \mathcal{A} \supseteq B_{(l+1)s} \supseteq V_{(l+1)s}.
$$
So $\card \mathcal{A} \geq \card V_{(l+1)s}$.
Thus by (\ref{eqDefTopTailEntropy}), (\ref{eqH(xi|eta)}), and (\ref{eqCardUn}),
\begin{align*}
h^*(f) & = \lim\limits_{m\to+\infty} \lim\limits_{l\to+\infty} \lim\limits_{n\to+\infty} \frac{1}{n} H \( \bigvee\limits_{i=0}^{n-1} f^{-i}\(\W^l\) \Bigg| \bigvee\limits_{j=0}^{n-1} f^{-j}\(\W^m\) \) \\
& \geq \liminf\limits_{m\to+\infty} \liminf\limits_{l\to+\infty} \liminf\limits_{s\to+\infty} \frac{1}{(l+1)s} \log \(k^{(l-m-2)s}\) \\
& = \liminf\limits_{m\to+\infty} \liminf\limits_{l\to+\infty} \frac{l-m-2}{l+1} \log k \\
& = \log k \\
& > 0.
\end{align*}
Therefore, the map $f$ is not asymptotically $h$-expansive.
\end{proof}
\begin{lemma} \label{lmHExpIterates}
Let $g\: X\rightarrow X$ be a continuous map on a compact metric space $(X,d)$. If $g$ is $h$-expansive then so is $g^n$ for each $n\in\N$.
\end{lemma}
The converse can also be easily established, i.e., if $g^n$ is $h$-expansive for some $n\in\N$, then so is $g$. But we will not need it in this paper.
\begin{proof}
We first observe from Definition~\ref{defRefSeqOpenCover} that if $\{\xi_l\}_{l\in\N_0}$ is a refining sequence of open covers, then so is $\{\xi^n_l\}_{l\in\N_0}$ for each $n\in\N$, where $\xi^n_l = \bigvee\limits_{i=0}^{n-1} g^{-i}\(\xi_l\)$. We also note that given an open cover $\lambda$ of $X$, we have
$$
\bigvee\limits_{i=0}^{mn-1} g^{-i}\(\lambda\) =\bigvee\limits_{j=0}^{m-1} \(g^n\)^{-j}\(\lambda^n\)
$$
for $n,m\in \N$, where $\lambda^n = \bigvee\limits_{k=0}^{n-1} g^{-k}\(\lambda\)$.
Assume that $g$ is $h$-expansive, then $h(g|\lambda)=0$ for some finite open cover $\lambda$ of $X$. Thus for each $n\in\N$,
\begin{align*}
h(g|\lambda) = & \lim\limits_{l\to+\infty} \lim\limits_{m\to+\infty} \frac{1}{mn} H \( \bigvee\limits_{i=0}^{mn-1} g^{-i}\(\xi_l\) \Bigg| \bigvee\limits_{j=0}^{mn-1} g^{-j}\(\lambda\) \) \\
= & \frac1n \lim\limits_{l\to+\infty} \lim\limits_{m\to+\infty} \frac{1}{m} H \( \bigvee\limits_{i=0}^{m-1} \(g^n\)^{-i}\(\xi^n_l\) \Bigg| \bigvee\limits_{j=0}^{m-1} \(g^n\)^{-j}\(\lambda^n\) \) \\
= & \frac1n h\(g^n|\lambda^n\),
\end{align*}
where $\xi^n_l$, $\lambda^n$ are defined as above. Note that $\lambda^n$ is also a finite open cover of $X$. Therefore $h\(g^n|\lambda^n\) = 0$, i.e., $g^n$ is $h$-expansive.
\end{proof}
The proof of the following theorem is similar to that of Theorem~\ref{thmNotAsympHExp}, and slightly simpler. However, due to subtle differences in both notation and constructions, we include the proof for the convenience of the reader.
\begin{theorem} \label{thmNotHExp}
No expanding Thurston map is $h$-expansive.
\end{theorem}
\begin{proof}
Let $f$ be an expanding Thurston map.
By Theorem~\ref{thmNotAsympHExp} and the fact that if $f$ is $h$-expanding then it is asymptotically $h$-expansive (see \cite[Corollary~2.1]{Mi76}), we can assume that $f$ has no periodic critical points.
Note that by (\ref{eqLocalDegreeProduct}), if a point $x\in S^2$ is a periodic critical point of $f^i$ for some $i\in\N$, then there exists $j\in\N_0$ such that $f^j(x)$ is a periodic critical point of $f$. So $f^i$ has no periodic critical points for $i\in\N$.
By Lemma~\ref{lmHExpIterates}, it suffices to prove that there exists $i\in\N$ such that $f^i$ is not $h$-expansive. Thus by Lemma~\ref{lmCexistsL}, we can assume, without loss of generality, that there exists a Jordan curve $\mathcal{C}\subseteq S^2$ containing $\operatorname{post} f$ such that $f(\mathcal{C})\subseteq \mathcal{C}$ and no $1$-tile joins opposite sides of $\mathcal{C}$.
In addition, we can assume, without loss of generality, that there exists a critical point $p\in\operatorname{crit} f \setminus \mathcal{C}$ with $f^2(p)=f(p) \neq p$. Indeed, we can choose any critical point $p_0\in \operatorname{crit} f$, then $f^{2i}(p_0)=f^i(p_0)\neq p_0$ for some $i\in\N$ since $f$ has no periodic critical points. By Lemma~\ref{lmPreImageDense}, there exist $j\in\N$ and $p\in f^{-ij}(p_0)\setminus \mathcal{C}$. We replace $f$ by $f^{i(j+1)}$. Note that for this new map $f$, we have $p\in\operatorname{crit} f \setminus \mathcal{C}$, $f^2(p)=f(p)\neq p$, $f(\mathcal{C})\subseteq \mathcal{C}$ and no $1$-tile joins opposite sides of $\mathcal{C}$.
Let $k=\deg_f(p)$. Then $k>1$.
From now on, we consider the cell decompositions of $S^2$ induced by $f$ and $\mathcal{C}$ in this proof.
\smallskip
Recall that $\W^i$ defined in (\ref{defSetNFlower}) denotes the set of all $i$-flowers $W^i(v)$, $v\in\V^i$, for each $i\in\N_0$.
Since $f$ is expanding, it is easy to see from Lemma~\ref{lmCellBoundsBM}, Proposition~\ref{propCellDecomp}, and the Lebesgue Number Lemma (\cite[Lemma~27.5]{Mu00}) that $\{ \W^i \}_{i\in\N_0}$ forms a refining sequence of open covers of $S^2$ (see Definition~\ref{defRefSeqOpenCover}). Thus it suffices to prove that
\begin{equation*}
h(f|\W^m)=\lim\limits_{l\to+\infty} \lim\limits_{n\to+\infty} \frac{1}{n} H \( \bigvee\limits_{i=0}^{n-1} f^{-i}\(\W^l\) \Bigg| \bigvee\limits_{j=0}^{n-1} f^{-j}\(\W^m\) \) > 0
\end{equation*}
for each $m\in\N$ sufficient large. See (\ref{eqH(xi|eta)}) for the definition of $H$.
Our plan is to construct a sequence $\{v_i\}_{i\in\N_0}$ of $m$-vertices such that for each $n\in\N_0$, the number of elements in $\bigvee\limits_{i=0}^{n-1} f^{-i}\(\W^l\)$ needed to cover $B_n=\bigcap\limits_{j=0}^{n-1} f^{-j} (W^m(v_{n-j}))$ can be bounded from below in such a way that $h(f|\W^m)>0$ follows immediately. More precisely, we observe that the more connected components $B_n$ has, the harder to cover $B_n$. So we will choose $\{v_i\}_{i\in\N_0}$ as a periodic sequence of $m$-vertices shadowing an infinite backward pseudo-orbit under iterations of $f$ in such a way that each period of $\{v_i\}_{i\in\N_0}$ begins with a backward orbit starting at $f(p)$ and $p$, and approaching $f(p)$ as the index $i$ increases, and then ends with $f(p)$. By a recursive construction, we keep track of each $B_n$ by a finite subset $V_n\subseteq B_n$ with the property that $\card(A\cap V_n)\leq 1$ for each $A \in \bigvee\limits_{i=0}^{n-1} f^{-i}\(\W^l\)$. A quantitative control of the size of $V_n$ leads to the conclusion that $h(f|\W^m)>0$ for some $m$ sufficiently large.
\smallskip
For this we fix $m,l\in\N$ with $l>2m+100 >200$.
Define $q_1 = p$. Then $q_1$ is necessarily a $1$-vertex, but not a $0$-vertex, i.e., $q_1\in \V^1 \setminus \V^0$. Since $q_1=p\notin \mathcal{C}$, we have $q_1\in W^0(f(p))$. By (\ref{defFlower}), the only $2$-vertex contained in $W^2(f(p))$ is $f(p)$. So $q_1 \in W^0(f(p)) \setminus W^2(f(p))$. Since $f(W^i(f(p))) = W^{i-1}(f(p))$ for each $i\in \N$ (see Remark~\ref{rmFlower}), we can recursively choose $q_j\in\V^j$ for each $j\in\{2,3,\dots,m+2\}$ such that
\begin{enumerate}
\smallskip
\item[(i)] $f(q_j) = q_{j-1}$,
\smallskip
\item[(ii)] $q_j \in W^{j-1}(f(p)) \setminus W^{j+1}(f(p))$.
\end{enumerate}
Set $q_0 = q_{m+2}$.
Since $f(W^i(p)) = W^{i-1}(f(p))$ for each $i\in \N$, and $k=\deg_f(p)>1$, we can choose distinct points $p_i\in\V^{m+3}$, $i\in\{1,2,\dots,k\}$, such that
\begin{enumerate}
\smallskip
\item[(i)] $f(p_i)= q_{m+2}$,
\smallskip
\item[(ii)] $p_i \in W^{m+2}(p) \setminus W^{m+4}(p)$.
\end{enumerate}
We will now construct recursively, for each $n=(m+2)s+r$ with $s\in\N_0$ and $r\in\{0,1,\dots,m+1\}$, an $m$-vertex $v_n\in\V^m$ and a set of $n$-vertices $V_n\subseteq \V^n$ such that for each $n\in\N_0$, the following properties are satisfied:
\begin{enumerate}
\smallskip
\item[(1)] $V_n \subseteq W^m(v_n)$;
\smallskip
\item[(2)] $f(V_n) = V_{n-1}$ if $n\neq 0$;
\smallskip
\item[(3)]
\begin{enumerate}
\item[(i)] $V_n \subseteq W^{m+1+r}(q_r)$ if $n=(m+2)s+r$ for some $s\in\N_0$ and some $r\in\{1,2,\dots,m+1\}$,
\smallskip
\item[(ii)] $V_n\subseteq W^{m+1+m+2}(q_0)$ if $n=(m+2)s$ for some $s\in\N_0$;
\end{enumerate}
\smallskip
\item[(4)] $\card V_n = k^{ \lceil \frac{n}{m+2} \rceil}$;
\smallskip
\item[(5)] for $A\in \bigvee\limits_{i=0}^{n-1} f^{-i} \(\W^l\)$ and $x,y\in V_n$ with $x\neq y$, we have $\{x,y\}\nsubseteq A$.
\end{enumerate}
We start our construction by first defining $v_n\in\V^m$ for each $n\in\N_0$. For $s\in\N_0$ and $r\in\{1,2,\dots,m\}$, set $v_{(m+2)s+r} = q_r$. For $s\in\N_0$ and $r\in\{0, m+1\}$, set $v_{(m+2)s+r} = f(p)$.
We now define $V_n$ recursively.
Let $V_0 = \{q_{m+2}\}$. Clearly $V_0$ satisfies properties~(1) through (5) in the induction step.
Assume that $V_n$ is defined and satisfies properties~(1) through (5) for each $n\in\{0,1,\dots,(m+2)s+r\}$, where $s\in\N_0$ and $r\in\{0,1,\dots,m+1\}$, we continue our construction in the following cases depending on $r$.
\smallskip
\emph{Case 1.} Assume $r=0$. Then $v_{(m+2)s+r} = f(p)$ and $v_{(m+2)s+r+1} = q_1=p$.
Note that $V_{(m+2)s+r} \subseteq W^{2m+3}(q_r)$ by the induction hypothesis, $q_r = q_{m+2} \in W^{m+1}(f(p))$, $f(p_i) = q_r$, and $f\(W^{2m+4}(p_i)\) = W^{2m+3}(q_r)$ for each $i\in\{1,2,\dots,k\}$ (see Remark~\ref{rmFlower}). Fix an arbitrary $i\in\{1,2,\dots,k\}$. We can choose, for each $x\in V_{(m+2)s+r}$, a point $x'\in W^{2m+4}(p_i)$ such that $f(x')=x$. Then define $V^i_{(m+2)s+r+1}$ to be the collection of all such chosen $x'$ that corresponds to $x\in V_{(m+2)s+r}$. Set
$$
V_{(m+2)s+r+1} = \bigcup\limits_{i=1}^k V^i_{(m+2)s+r+1}.
$$
Since $p_i \in W^{m+2}(p)$ and $V^i_{(m+2)s+r+1} \subseteq W^{2m+4}(p_i)$, we get that $V^i_{(m+2)s+r+1} \subseteq W^{m+2}(p)$. So $V_{(m+2)s+r+1} \subseteq W^{m+2}(p) \subseteq W^m(p)$. Since $v_{(m+2)s+r+1} = q_1=p$, properties~(1) and (3) are verified. Property~(2) is clear from the construction.
To establish property~(4), it suffices to show that $V^i_{(m+2)s+r+1} \cap V^j_{(m+2)s+r+1} = \emptyset$ for $1\leq i<j \leq k$. Indeed, since $V^i_{(m+2)s+r+1} \subseteq W^{2m+4}(p_i)$ and $V^j_{(m+2)s+r+1} \subseteq W^{2m+4}(p_j)$, it suffices to prove that $\overline{W}^{2m+4}(p_i)\cap \overline{W}^{2m+4}(p_j) = \emptyset$. Suppose that $\overline{W}^{2m+4}(p_i)\cap \overline{W}^{2m+4}(p_j) \neq \emptyset$, then since no $1$-tile joins opposite sides of $\mathcal{C}$, and $p_i,p_j\in\V^{m+3}$, we get from Lemma~\ref{lmTileInFlower}(iii) that $p_i=p_j$, i.e., $i=j$. But $i<j$, a contradiction.
We only need to verify property~(5) now. Indeed, suppose that distinct points $x,y\in V_{(m+2)s+r+1}$ satisfy $\{x,y\}\subseteq A$ for some $A\in\bigvee\limits_{a=0}^{(m+2)s+r} f^{-a}\(\W^l\)$. Then $A\subseteq W^l(v^l)$ for some $v^l\in\V^l$. By construction, there exist $i,j\in\{1,2,\dots,k\}$ such that $x\in W^{2m+4}(p_i)$ and $y\in W^{2m+4}(p_j)$. Since $l> 2m+100$ and $x\in W^l(v^l)\cap W^{2m+4}(p_i)$, we get $v^l \in \overline{W}^{2m+4}(p_i)$ by Lemma~\ref{lmTileInFlower}(i). Similarly $v^l\in \overline{W}^{2m+4}(p_j)$. Then by the argument above, we get that $p_i=p_j$, i.e., $i=j$. Thus $f(x)\neq f(y)$ by construction. But then $f(x),f(y)$, and $f(A)$ satisfy
\begin{enumerate}
\smallskip
\item[(a)] $f(A)\subseteq B$ for some $B\in \bigvee\limits_{a=0}^{(m+2)s+r-1} f^{-a}\(\W^l\)$,
\smallskip
\item[(b)] $f(x), f(y)\in V_{(m+2)s+r}$, and $f(x)\neq f(y)$,
\smallskip
\item[(c)] $\{f(x),f(y)\} \subseteq f(A)\subseteq B$.
\end{enumerate}
This contradicts property~(5) for $V_{(m+2)s+r}$ in the induction hypothesis.
\smallskip
\emph{Case 2.} Assume $r\neq 0$, i.e., $r\in\{1,2,\dots,m+1\}$.
Note that $V_{(m+2)s+r} \subseteq W^{m+1+r} (q_r)$, $f(q_{r+1})=q_r$, and by Remark~\ref{rmFlower}, $
f\(W^{m+1+r+1}(q_{r+1})\) = W^{m+1+r}(q_r)$. We can choose, for each $x\in V_{(m+2)s+r}$, a point $x'\in W^{m+1+r+1}(q_{r+1})$ such that $f(x')=x$. Then define $V_{(m+2)s+r+1}$ to be the collection of all such chosen $x'$ that corresponds to $x\in V_{(m+2)s+r}$. Properties~(2), (3), and (4) are clear from the construction. To establish property~(1) in the case when $r\in\{1,2,\dots,m-1\}$, we recall that $v_{(m+2)s+r+1} = q_{r+1}$. For the case when $r\in\{m,m+1\}$, we note that $V_{(m+2)s+r+1} \subseteq W^{m+1+r+1}(q_{r+1})$ and $q_{r+1}\in W^r(f(p))$, so
$$
V_{(m+2)s+r+1} \subseteq W^{2m}(q_{r+1}) \subseteq W^m(f(p)) = W^m\(v_{(m+2)s+r+1}\).
$$
We only need to verify property~(5) now. Indeed, suppose that distinct points $x,y\in V_{(m+2)s+r+1}$ satisfy $\{x,y\}\subseteq A$ for some $A\in\bigvee\limits_{i=0}^{(m+2)s+r} f^{-i}\(\W^l\)$. Then by construction $f(x),f(y)$, and $f(A)$ satisfy
\begin{enumerate}
\smallskip
\item[(a)] $f(A)\subseteq B$ for some $B\in \bigvee\limits_{i=0}^{(m+2)s+r-1} f^{-i}\(\W^l\)$,
\smallskip
\item[(b)] $f(x), f(y)\in V_{(m+2)s+r}$, and $f(x)\neq f(y)$,
\smallskip
\item[(c)] $\{f(x),f(y)\} \subseteq f(A)\subseteq B$.
\end{enumerate}
This contradicts property~(5) for $V_{(m+2)s+r}$ in the induction hypothesis.
\smallskip
The recursive construction and the inductive proof of the properties of the construction are now complete.
\smallskip
For each $s\in\N$, we consider
$$
B_{(m+2)s}= \bigcap\limits_{j=0}^{(m+2)s-1} f^{-j} \(W^m\(v_{(m+2)s - j}\) \) \in \bigvee\limits_{j=0}^{(m+2)s-1} f^{-j} \(\W^m \).
$$
Then $V_{(m+2)s} \subseteq B_{(m+2)s}$ by properties~(1) and (2) of the construction. On the other hand, by property~(5), if $\mathcal{A}\subseteq \bigvee\limits_{j=0}^{(m+2)s-1} f^{-j} \(\W^l \)$ satisfies
$$
\bigcup \mathcal{A} \supseteq B_{(m+2)s} \supseteq V_{(m+2)s}.
$$
So $
\card \mathcal{A} \geq \card V_{(m+2)s} = k^s$, where the equality follows from property~(4).
Thus by (\ref{eqH(xi|eta)}),
\begin{align*}
h(f|\W^m)= & \lim\limits_{l\to+\infty} \lim\limits_{n\to+\infty} \frac{1}{n} H \( \bigvee\limits_{i=0}^{n-1} f^{-i}\(\W^l\) \Bigg| \bigvee\limits_{j=0}^{n-1} f^{-j}\(\W^m\) \) \\
\geq & \liminf\limits_{l\to+\infty} \liminf\limits_{s\to+\infty} \frac{1}{(m+2)s} \log \(k^s\)
= \frac{\log k}{m+2}
> 0.
\end{align*}
Therefore, the map $f$ is not $h$-expansive.
\end{proof}
\section{Large deviation principles} \label{sctLDP}
\subsection{Thermodynamical formalism} \label{subsctThermDynForm}
We review some key concepts from thermodynamical formalism. For a more careful introduction in our context, see \cite[Section~5]{Li14}. We refer the reader to \cite[Chapter~3]{PU10}, \cite[Chapter~9]{Wa82} or \cite[Chapter~20]{KH95} for a detailed study of these concepts for general dynamical systems.
Let $(X,d)$ be a compact metric space and $g\:X\rightarrow X$ a continuous map. For $n\in\N$ and $x,y\in X$,
$$
d^n_g(x,y)=\operatorname{max}\big\{ \hspace{-0.2mm} d\!\(g^k(x),g^k(y)\) \hspace{-0.3mm} \big| k\in\{0,1,\dots,n-1\} \!\big\}
$$
defines a new metric on $X$. A set $F\subseteq X$ is \defn{$(n,\epsilon)$-separated}, for some $n\in\N$ and $\epsilon>0$, if for each pair of distinct points $x,y\in F$, we have $d^n_g(x,y)\geq \epsilon$. For $\epsilon > 0$ and $n\in\N$, let $F_n(\epsilon)$ be a maximal (in the sense of inclusion) $(n,\epsilon)$-separated set in $X$.
For each $\psi \inC(X)$, the following limits exist and are equal, and we denote the limits by $P(g,\psi)$ (see for example, \cite[Theorem~3.3.2]{PU10}):
\begin{align} \label{defTopPressure}
P(g,\psi) & = \lim \limits_{\epsilon\to 0} \limsup\limits_{n\to+\infty} \frac{1}{n} \log \sum\limits_{x\in F_n(\epsilon)} \exp(S_n \psi(x)) \notag \\
& = \lim \limits_{\epsilon\to 0} \liminf\limits_{n\to+\infty} \frac{1}{n} \log \sum\limits_{x\in F_n(\epsilon)} \exp(S_n \psi(x)),
\end{align}
where $S_n \psi (x) = \sum\limits_{j=0}^{n-1} \psi(g^j(x))$ is defined in (\ref{eqDefSnPt}). We call $P(g,\psi)$ the \defn{topological pressure} of $g$ with respect to the \emph{potential} $\psi$. When $\psi\equiv 0$, we call $P(g,0)=h_{\operatorname{top}}(g)$ the \defn{topological entropy} of $g$.
Let $\mu\in \mathcal{M}(X,g)$. We denote by $h_\mu(g)$ the \emph{measure-theoretic entropy} of $g$ for $\mu$. Then for each $\psi\inC(X)$, the \defn{measure-theoretic pressure} $P_\mu(g,\psi)$ of $g$ for the measure $\mu$ and the potential $\psi$ is
\begin{equation} \label{defMeasTheoPressure}
P_\mu(g,\psi)= h_\mu (g) + \int \! \psi \,\mathrm{d}\mu.
\end{equation}
By the Variational Principle (see for example, \cite[Theorem~3.4.1]{PU10}), we have that for each $\psi\inC(X)$,
\begin{equation} \label{eqVPPressure}
P(g,\psi)=\sup\{P_\mu(g,\psi)\,|\,\mu\in \mathcal{M}(X,g)\},
\end{equation}
and in particular,
\begin{equation} \label{eqVPEntropy}
h_{\operatorname{top}}(g)=\sup\{h_\mu(g)\,|\,\mu\in \mathcal{M}(X,g)\}.
\end{equation}
A measure $\mu$ that attains the supremum in (\ref{eqVPPressure}) is called an \defn{equilibrium state} for the transformation $g$ and the potential $\psi$. A measure $\mu$ that attains the supremum in (\ref{eqVPEntropy}) is called a \defn{measure of maximal entropy} of $g$.
By the work of P.~Ha\"{\i}ssinsky and K.~Pilgrim \cite{HP09}, and M.~Bonk and D.~Meyer \cite{BM10}, we know that there exists a unique measure of maximal entropy $\mu_f$ for an expanding Thurston map $f$, and that
\begin{equation} \label{eqTopEntropy}
h_{\operatorname{top}} (f)=\log(\deg f).
\end{equation}
The existence and uniqueness of the equilibrium state for an expanding Thurston map $f$ and a H\"older continuous potential $\phi$ is established in \cite[Theorem~1.1]{Li14}.
\begin{theorem} \label{thmExistenceUniquenessES}
Let $f\:S^2 \rightarrow S^2$ be an expanding Thurston map and $d$ be a visual metric on $S^2$ for $f$. Let $\phi$ be a real-valued H\"{o}lder continuous function on $S^2$ with respect to the metric $d$. Then there exists a unique equilibrium state $\mu_\phi$ for the map $f$ and the potential $\phi$.
\end{theorem}
Let $f$, $d$, $\phi$, $\alpha$ satisfy the Assumptions. Recall that the main tool we used in \cite{Li14} to prove this theorem is the Ruelle operator $\mathcal{L}_\phi$. We summarize relevant definitions and facts about the Ruelle operator below and refer the reader to \cite[Section~5]{Li14} for a detailed discussion.
Let $\psi\in C(S^2)$ be a continuous function. Recall that $f\: S^2 \rightarrow S^2$ is an expanding Thurston map. The \defn{Ruelle operator} $\mathcal{L}_\psi$ on $C(S^2)$ is defined as the following
\begin{equation} \label{eqDefRuelleOp}
\mathcal{L}_\psi(u)(x)= \sum\limits_{y\in f^{-1}(x)} \deg_f(y) u(y) \exp(\psi(y)),
\end{equation}
for each $u\in C(S^2)$. Note that $\mathcal{L}_\psi$ is a well-defined, positive, continuous operator on $C(S^2)$. The adjoint operator $\mathcal{L}_\psi^*\: C^*(S^2)\rightarrow C^*(S^2)$ of $\mathcal{L}_\psi$ acts on the dual space $C^*(S^2)$ of the Banach space $C(S^2)$. We identify $C^*(S^2)$ with the space $\mathcal{M}(S^2)$ of finite signed Borel measures on $S^2$ by the Riesz representation theorem.
\begin{proof}[Proof of Theorem~\ref{thmExistenceES}]
By Alaoglu's theorem, the space $\mathcal{M}(S^2,f)$ of $f$-invariant Borel probability measures equipped with the weak$^*$ topology is compact. Since the measure-theoretic entropy $\mu\mapsto h_\mu(f)$ is upper semi-continuous by Corollary~\ref{corUSC}, so is $\mu\mapsto P_\mu(f,\psi)$ by (\ref{defMeasTheoPressure}). Thus $\mu\mapsto P_\mu(f,\psi)$ attains its supremum over $\mathcal{M}(S^2,f)$ at a measure $\mu_\psi$, which by the Variational Principle (\ref{eqVPPressure}) is an equilibrium state for the map $f$ and the potential $\psi$.
\end{proof}
\subsection{Level-2 large deviation principles} \label{subsctLDP}
Let $X$ be a compact metrizable topological space. Recall that $\mathcal{P}(X)$ is the set of Borel probability measures on $X$. We equip $\mathcal{P}(X)$ with the weak$^*$ topology. Note this topology is metrizable (see for example, \cite[Theorem~5.1]{Con85}). Let $I\: \mathcal{P}(X) \rightarrow [0,+\infty]$ be a \defn{lower semi-continuous function}, i.e., $I$ satisfy the condition that $\liminf_{y\to x} I(y)\geq I(x)$ for all $x\in \mathcal{P}(X)$.
A sequence $\{\Omega_n \}_{n\in\N}$ of Borel probability measures on $\mathcal{P}(X)$ is said to satisfy a \defn{large deviation principle with rate function $I$} if for each closed subset $\mathfrak{F}$ of $\mathcal{P}(X)$ and each open subset $\mathfrak{G}$ of $\mathcal{P}(X)$ we have
$$
\limsup\limits_{n\to+\infty} \frac1n \log \Omega_n(\mathfrak{F}) \leq - \inf\{I(x) \,|\, x\in\mathfrak{F} \},
$$
and
$$
\liminf\limits_{n\to+\infty} \frac1n \log \Omega_n(\mathfrak{G}) \geq - \inf\{I(x) \,|\, x\in\mathfrak{G} \}.
$$
We will apply the following theorem due to Y.~Kifer \cite[Theorem~4.3]{Ki90}, reformulated by H.~Comman and J.~Rivera-Letelier \cite[Theorem~C]{CRL11}.
\begin{theorem}[Y.~Kifer, 1990; H.~Comman \& J.~Rivera-Letelier, 2011] \label{thmAbsLargeDeviationPrincipleCRL}
Let $X$ be a compact metrizable topological space, and let $g\: X\rightarrow X$ be a continuous map. Fix $\phi\inC(X)$, and let $H$ be a dense vector subspace of $C(X)$ with respect to the uniform norm. Let $I^\phi \: \mathcal{P}(X) \rightarrow [0,+\infty]$ be the function defined by
\begin{equation*}
I^\phi (\mu) =
\left\{
\begin{array}{ll}
P(g,\phi) -\int\! \phi \,\mathrm{d}\mu - h_\mu(g) & \mbox{if } \mu\in\mathcal{M}(X,g); \\
+\infty & \mbox{if } \mu\in\mathcal{P}(X) \setminus \mathcal{M}(X,g).
\end{array}
\right.
\end{equation*}
We assume the following conditions are satisfied:
\begin{enumerate}
\smallskip
\item[(i)] The measure-theoretic entropy $h_\mu(g)$ of $g$, as a function of $\mu$ defined on $\mathcal{M}(X,g)$ (equipped with the weak$^*$ topology), is finite and upper semi-continuous.
\smallskip
\item[(ii)] For each $\psi \in H$, there exists a unique equilibrium state for the map $g$ and the potential $\phi+\psi$.
\end{enumerate}
Then every sequence $\{ \Omega_n \}_{n\in\N}$ of Borel probability measures on $\mathcal{P}(X)$ such that for each $\psi \in H$,
\begin{equation} \label{eqThmCCondition}
\lim\limits_{n\to+\infty} \frac{1}{n} \log \int_{\mathcal{P}(X)} \! \exp \(n\int\!\psi \,\mathrm{d}\mu \) \,\mathrm{d}\Omega_n(\mu) = P(g,\phi+\psi) - P(g,\phi),
\end{equation}
satisfies a large deviation principle with rate function $I^\phi$, and it converges in the weak$^*$ topology to the Dirac measure supported on the unique equilibrium state for the map $g$ and the potential $\phi$. Furthermore, for each convex open subset $\mathfrak{G}$ of $\mathcal{P}(X)$ containing some invariant measure, we have
\begin{equation*}
\lim\limits_{n\to+\infty} \frac{1}{n} \log \Omega_n(\mathfrak{G})
= \lim\limits_{n\to+\infty} \frac{1}{n} \log \Omega_n\big(\overline{\mathfrak{G}}\big)
= -\inf\limits_{\mathfrak{G}} I^\phi
= -\inf\limits_{\overline{\mathfrak{G}}} I^\phi.
\end{equation*}
\end{theorem}
Recall that $P(g,\phi)$ is the topological pressure of the map $g$ with respect to the potential $\phi$.
In our context, $X=S^2$, the map $g=f$ where $f\: S^2\rightarrow S^2$ is an expanding Thurston map with no periodic critical points. Fix a visual metric $d$ on $S^2$ for $f$. The function $\phi$ is a real-valued H\"older continuous function with an exponent $\alpha\in (0,1]$. Then $H=\Holder{\alpha}(S^2,d)$ is the space of real-valued H\"older continuous functions with the exponent $\alpha$ on $(S^2,d)$. Note that $\Holder{\alpha}(S^2,d)$ is dense in $C(S^2)$ (equipped with the uniform norm) (see for example, \cite[Lemma~6.12]{Li14}). Condition~(i) is satisfied by Corollary~\ref{corUSC}. Condition~(ii) is guaranteed by Theorem~\ref{thmExistenceUniquenessES}. Thus we just need to verify (\ref{eqThmCCondition}) for the sequences we will consider in this section.
\subsection{Technical lemmas} \label{subsctLDPLemmas}
\begin{lemma} \label{lmAtMostOneCritPerTile}
Let $f$, $\mathcal{C}$, $d$, $\Lambda$ satisfy the Assumptions. Then there exists $N_0\in\N$ such that for each $n\geq N_0$ and each $n$-tile $X^n\in\X^n(f,\mathcal{C})$, the number of fixed points of $f^n$ contained in $X^n$ is at most $1$.
\end{lemma}
\begin{proof}
By Lemma~\ref{lmMetricDistortion}, for each $i\in\N$, each $i$-tile $X^i\in\X^i(f,\mathcal{C})$, and each pair of points $x,y\in X^i$, we have
$$
d\(f^i(x),f^i(y)\) \geq \frac{\Lambda^i}{C_0} d(x,y),
$$
where $C_0>1$ is a constant depending only on $f$ and $d$ from Lemma~\ref{lmMetricDistortion}.
We choose $N_0\in\N$ such that $\Lambda^{N_0} > C_0$.
Let $n\geq N_0$ and $X^n\in\X^n(f,\mathcal{C})$. Suppose two distinct points $p,q\in X$ satisfy $f^n(p)=p$ and $f^n(q)=q$. Then
$$
1=\frac{d\(f^n(p),f^n(q)\)}{d(p,q)} \geq \frac{\Lambda^n}{C_0} >1,
$$
a contradiction. This completes the proof.
\end{proof}
This lemma in some sense generalizes Lemma~4.3 in \cite{Li13} to Jordan curves that are not necessarily $f$-invariant, but $f^{n_c}$-invariant for some $n_c\in\N$. The conclusions of both lemmas hold when $n$ is sufficiently large, which is a combinatorial condition in Lemma~4.3 in \cite{Li13} and a metric condition for Lemma~\ref{lmAtMostOneCritPerTile} here. The proof of Lemma~\ref{lmAtMostOneCritPerTile} is simpler, but the proof of Lemma~4.3 in \cite{Li13} is more self-contained.
By Lemma~5.1 and Lemma~5.2 in \cite{Li14}, we have the following distortion bounds.
\begin{lemma} \label{lmSnPhiBound}
Let $f$, $\mathcal{C}$, $d$, $L$, $\Lambda$, $\phi$, $\alpha$ satisfy the Assumptions. Then there exist constants $C_1>0$ and $C_2\geq 1$ depending only on $f$, $\mathcal{C}$, $d$, $\phi$, and $\alpha$ such that
\begin{equation} \label{eqSnPhiBound}
\Abs{S_m\phi(x)-S_m\phi(y)} \leq C_1 d(f^m(x),f^m(y))^\alpha,
\end{equation}
and
\begin{equation} \label{eqSigmaExpSnPhiBound}
\frac{\sum\limits_{x'\in f^{-l}(x_0)} \deg_{f^l}(x') \exp (S_l\phi(x'))}{\sum\limits_{y'\in f^{-l}(y_0)} \deg_{f^l}(y') \exp (S_l\phi(y'))}\leq C_2,
\end{equation}
for $n,m,l\in\N_0$ with $m\leq n $, $X^n\in\X^n(f,\mathcal{C})$, $x,y\in X^n$, and $x_0,y_0\in S^2$.
\end{lemma}
Note that due to the convention described in Section~\ref{sctAssumptions}, we do not say that $C_1$ and $C_2$ depend on $\Lambda$ or $n_\mathcal{C}$.
We record the following well-known lemma, sometimes known as the Portmanteau Theorem, and refer the reader to \cite[Theorem~2.1]{Bi99} for a proof.
\begin{lemma} \label{lmPortmanteau}
Let $(X,d)$ be a compact metric space, and $\mu$ and $\mu_i$, for $i\in\N$, be Borel probability measures on $X$. Then the following are equivalent:
\begin{enumerate}
\smallskip
\item[(i)] $\mu_i \stackrel{w^*}{\longrightarrow} \mu$ as $i\longrightarrow +\infty$;
\smallskip
\item[(ii)] $\limsup\limits_{i\to+\infty} \mu_i(F) \leq \mu(F)$ for each closed set $F\subseteq X$;
\smallskip
\item[(iii)] $\liminf\limits_{i\to+\infty} \mu_i(G) \geq \mu(G)$ for each open set $G\subseteq X$;
\smallskip
\item[(iv)] $\lim\limits_{i\to+\infty} \mu_i(B)= \mu(B)$ for each Borel set $B\subseteq X$ with $\mu(\partial B) = 0$.
\end{enumerate}
\end{lemma}
\subsection{Characterizations of the pressure $P(f,\phi)$} \label{subsctTopPres}
Let $f$, $d$, $\phi$, $\alpha$ satisfy the Assumptions. We denote by $m_\phi$ the unique eigenmeasure of $\mathcal{L}^*_\phi$, i.e., the unique Borel probability measure on $S^2$ that satisfies $\mathcal{L}^*_\phi(m_\phi)=c m_\phi$ for some constant $c\in \R$ (compare \cite[Theorem~5.11 and Corollary~6.10]{Li14}). As in \cite{Li14}, we denote $\overline\phi=\phi-P(f,\phi)$, and
\begin{equation} \label{eqDefPhiTilde}
\widetilde\phi = \phi - P(f,\phi) +\log u_\phi - \log (u_\phi\circ f),
\end{equation}
where $u_\phi$ is the unique eigenfunction, upto scalar multiplication, of $\mathcal{L}_{\overline\phi}$ corresponding to the eigenvalue $1$, which is also the uniform limit of the sequence of continuous functions $\Big\{ \frac{1}{n} \sum\limits_{j=0}^{n-1} \mathcal{L}^j_{\overline\phi} (\mathbbm{1}_{S^2}) \Big\}_{n\in\N}$ (compare Theorem~5.16 and the remark that follows it in \cite{Li14}).
We record the following result on equidistribution of preimages with respect to the equilibrium state $\mu_\phi$ and the measure $m_\phi$ (see \cite[Proposition~9.1]{Li14}). We will only use (\ref{eqWeakConvPreImgToMWithWeight}) in this paper, namely, in the proof of Proposition~\ref{propPressureLimitPeriodicPts}.
\begin{prop} \label{propWeakConvPreImgWithWeight}
Let $f$, $d$, $\phi$, $\alpha$ satisfy the Assumptions. Let $\mu_\phi$ be the unique equilibrium state for $f$ and $\phi$, let $m_\phi$ be the unique eigenmeasure of $\mathcal{L}^*_\phi$, and $\widetilde\phi$ as defined in (\ref{eqDefPhiTilde}). For each sequence $\{ x_n \}_{n\in\N}$ of points in $S^2$, we define the Borel probability measures
\begin{align}
\xi_n & = \frac{1}{Z_n(\phi)} \sum\limits_{y\in f^{-n}(x_n)} \deg_{f^n} (y) \exp\(S_n\phi(y)\) \delta_y, \\
\widehat\xi_n & = \frac{1}{Z_n(\phi)} \sum\limits_{y\in f^{-n}(x_n)} \deg_{f^n} (y) \exp\(S_n\phi(y)\) \frac{1}{n} \sum\limits_{i=0}^{n-1} \delta_{f^i(y)},\\
\widetilde\xi_n & = \frac{1}{Z_n\big(\widetilde\phi \hspace{0.5mm} \big)} \sum\limits_{y\in f^{-n}(x_n)} \deg_{f^n} (y) \exp(S_n\widetilde\phi(y)) \delta_y,
\end{align}
for each $n\in\N_0$, where $Z_n(\psi) = \sum\limits_{y\in f^{-n}(x_n)} \deg_{f^n} (y) \exp\(S_n\psi(y)\)$, for $\psi\inC(S^2)$. Then
\begin{equation} \label{eqWeakConvPreImgToMWithWeight}
\xi_n \stackrel{w^*}{\longrightarrow} m_\phi \text{ as } n\longrightarrow +\infty,
\end{equation}
\begin{equation} \label{eqWeakConvPreImgToMuSumWithWeight}
\widehat\xi_n \stackrel{w^*}{\longrightarrow} \mu_\phi \text{ as } n\longrightarrow +\infty,
\end{equation}
\begin{equation} \label{eqWeakConvPreImgToMuTildeWithWeight}
\widetilde\xi_n \stackrel{w^*}{\longrightarrow} \mu_\phi \text{ as } n\longrightarrow +\infty.
\end{equation}
\end{prop}
We now prove a slight generalization of Proposition~5.17 in \cite{Li14}.
\begin{prop} \label{propPressureLimitPreImgs}
Let $f$, $d$, $\phi$, $\alpha$ satisfy the Assumptions. Then for each sequence $\{x_n\}_{n\in\N}$ in $S^2$, we have
\begin{equation} \label{eqPressureLimitPreImgsWDeg}
P(f,\phi) = \lim\limits_{n\to +\infty} \frac{1}{n} \log \sum\limits_{y\in f^{-n}(x_n)} \deg_{f^n}(y) \exp (S_n\phi(y)).
\end{equation}
If we also assume that $f$ has no periodic critical points, then for an arbitrary sequence of functions $\{w_n\: S^2\rightarrow \R\}_{n\in\N}$ satisfying $w_n(x)\in [1,\deg_{f^n}(x)]$ for each $n\in\N$ and each $x\in S^2$, we have
\begin{equation} \label{eqPressureLimitPreImgsWODeg}
P(f,\phi) = \lim\limits_{n\to +\infty} \frac{1}{n} \log \sum\limits_{y\in f^{-n}(x_n)} w_n(y) \exp (S_n\phi(y)).
\end{equation}
\end{prop}
\begin{proof}
We fix a Jordan curve $\mathcal{C}\subseteq S^2$ that satisfies the Assumptions (see \cite[Theorem~1.2]{BM10} or Lemma~\ref{lmCexistsL} for the existence of such $\mathcal{C}$). By Proposition~5.17 in \cite{Li14}, for each $x\in S^2$ we have
$$
P(f,\phi) = \lim\limits_{n\to +\infty} \frac{1}{n} \log \sum\limits_{y\in f^{-n}(x)} \deg_{f^n}(y) \exp (S_n\phi(y)).
$$
Combining this equation with (\ref{eqSigmaExpSnPhiBound}) in Lemma~\ref{lmSnPhiBound}, we get (\ref{eqPressureLimitPreImgsWDeg}).
Assume now that $f$ has no periodic critical points. Then there exists a finite number $M\in\N$ that depends only on $f$ such that $\deg_{f^n}(x) \leq M$ for $n\in\N_0$ and $x\in S^2$ \cite[Lemma~17.1]{BM10}. Thus for each $n\in\N$,
$$
1 \leq \frac{\sum\limits_{y\in f^{-n}(x_n)} \deg_{f^n}(y) \exp (S_n\phi(y))}{\sum\limits_{y\in f^{-n}(x_n)} w_n(y) \exp (S_n\phi(y))} \leq M.
$$
Hence (\ref{eqPressureLimitPreImgsWODeg}) follows from (\ref{eqPressureLimitPreImgsWDeg}).
\end{proof}
While Proposition~\ref{propPressureLimitPreImgs} is a statement for itereated preimages, the next proposition is for periodic points. Let $P_{1,f^n}=\{x\in S^2 \,|\, f^n(x)=x\}$ for $n\in\N$.
\begin{prop} \label{propPressureLimitPeriodicPts}
Let $f$, $d$, $\phi$, $\alpha$ satisfy the Assumptions. Fix an arbitrary sequence of functions $\{w_n\: S^2\rightarrow \R\}_{n\in\N}$ satisfying $w_n(x)\in [1,\deg_{f^n}(x)]$ for each $n\in\N$ and each $x\in S^2$. Then
\begin{equation} \label{eqPressureLimitPeriodicPts}
P(f,\phi) = \lim\limits_{n\to +\infty} \frac{1}{n} \log \sum\limits_{x\in P_{1,f^n}} w_n(x) \exp (S_n\phi(x)).
\end{equation}
\end{prop}
\begin{proof}
We fix a Jordan curve $\mathcal{C}\subseteq S^2$ that satisfies the Assumptions (see \cite[Theorem~1.2]{BM10} or Lemma~\ref{lmCexistsL} for the existence of such $\mathcal{C}$).
We first prove (\ref{eqPressureLimitPeriodicPts}). By Proposition~\ref{propPressureLimitPreImgs}, it suffices to prove that there exist $C>1$ and $z\in S^2$ such that for each $n\in\N$ sufficiently large,
\begin{equation} \label{eqSumsComparable}
\frac{1}{C} \leq \frac{\sum\limits_{x\in P_{1,f^n}} w_n(x) \exp (S_n\phi(x))}{\sum\limits_{x\in f^{-n}(z)}\deg_{f^n}(x) \exp (S_n\phi(x))} \leq C.
\end{equation}
We fix a $0$-edge $e_0 \subseteq \mathcal{C}$ and a point $z\in \inte (e_0)$.
By Proposition~7.1 in \cite{Li14}, $m_\phi(\mathcal{C}) = 0$. By the continuity of $m_\phi$, we can find $\delta >0$ such that
\begin{equation}
m_\phi \big( \overline{N^\delta_d} (\mathcal{C})\big) < \frac{1}{100}.
\end{equation}
Note that $\deg_{f^n}(y)=1$ if $f^n(y)=z$ for $n\in\N$. We define, for each $n\in\N_0$, the probability measure
\begin{align} \label{eqDistrPreImgs}
\nu_n
& = \sum\limits_{x\in f^{-n}(z)} \frac{\deg_{f^n} (x)\exp\(S_n\phi(x)\)}{\sum_{y\in f^{-n}(z)} \deg_{f^n} (y) \exp\(S_n\phi(y)\)} \delta_x \notag \\
& = \sum\limits_{x\in f^{-n}(z)} \frac{\exp\(S_n\phi(x)\)}{\sum_{y\in f^{-n}(z)} \exp\(S_n\phi(y)\)} \delta_x ,
\end{align}
Let $N_0\in \N$ be the constant from Lemma~\ref{lmAtMostOneCritPerTile}. By (\ref{eqWeakConvPreImgToMWithWeight}) in Proposition~\ref{propWeakConvPreImgWithWeight}, $\nu_n\stackrel{w^*}{\longrightarrow} m_\phi$ as $n\longrightarrow +\infty$. So by Lemma~\ref{lmPortmanteau}, we can choose $N_1>N_0$ such that for each $n\in\N$ with $n>N_1$, we have
\begin{equation} \label{eqNbhdCFewPreImgs}
\nu_n \big( \overline{N^\delta_d} (\mathcal{C})\big) < \frac{1}{10}.
\end{equation}
By Lemma~\ref{lmCellBoundsBM}, it is clear that we can choose $N_2>N_1$ such that for each $n\in\N$ with $n>N_2$, and each $n$-tile $X^n\in\X^n$,
\begin{equation} \label{eqDiamNTileSmall}
\diam_d(X^n) < \frac{\delta}{10}.
\end{equation}
We observe that for each $i\in\N$, we can pair a white $i$-tile $X_w^i\in\X_w^i$ and a black $i$-tile $X_b^i\in\X_b^i$ whose intersection $X_w^i\cap X_b^i$ is an $i$-edge contained in $f^{-i}(e_0)$. There are a total of $d^i$ such pairs and each $i$-tile is in exactly one such pair. We denote by $\mathbf{P}_i$ the collection of the unions $X_w^i\cup X_b^i$ of such pairs, i.e.,
$$
\mathbf{P}_i=\{X_w^i\cup X_b^i \,|\, X_w^i\in\X_w^i,X_b^i\in\X_b^i, X_w^i\cap X_b^i\cap f^{-i}(e_0) \in \E^i \}.
$$
We denote $\mathbf{P}^\delta_i=\{A\in \mathbf{P}_i \,|\, A \setminus N^\delta_d(\mathcal{C}) \neq \emptyset\}$.
We now fix an integer $n> N_2$.
Then $\mathbf{P}^\delta_n$ forms a cover of $S^2\setminus \overline{N^\delta_d}(\mathcal{C})$. For each $A\in\mathbf{P}^\delta_n$, by (\ref{eqDiamNTileSmall}) we have $A\cap \mathcal{C} =\emptyset$. So $A\subseteq \inte X^0_w$ or $A\subseteq \inte X^0_b$, where $X^0_w$ and $X^0_b$ are the white $0$-tile and the black $0$-tile in $\X^0$, respectively. So by Brouwer's Fixed Point Theorem (see for example, \cite[Theorem~1.9]{Ha02}) and Lemma~\ref{lmAtMostOneCritPerTile}, we can define a function $p\:\mathbf{P}^\delta_n\rightarrow P_{1,f^n}$ in such a way that $p(A)$ is the unique fixed point of $f^n$ contained in $A$. (For example, if $A\in\mathbf{P}^\delta_n$ is the union of a black $n$-tile $X^n_b$ and a white $n$-tile $X^n_w$ and is a subset of the interior of the black $0$-tile, then there is no fixed point of $f^n$ in $X^n_w$, and by applying Brouwer's Fixed Point Theorem to the inverse of $f^n$ restricted to $X^n_b$, we get a fixed point $x\in X^n_b$ of $f^n$, which is the unique fixed point of $f^n$ in $X^n_b$ by Lemma~\ref{lmAtMostOneCritPerTile}.) Moreover, for each $A\in\mathbf{P}^\delta_n$, $p(A)\in \inter A$, so $\deg_{f^n}(p(A)) =1 = w_n(p(A))$. In general, by Lemma~\ref{lmAtMostOneCritPerTile}, each $A\in\mathbf{P}_n$ contains at most $2$ fixed points of $f^n$.
We also define a function $q\: \mathbf{P}_n \rightarrow f^{-n}(z)$ in such a way that $q(A)$ is the unique preimage of $z$ under $f^n$ that is contained in $A$, for each $A\in\mathbf{P}_n$ (see Proposition~\ref{propCellDecomp}). We note that if $X^n_w\in\X^n_w$ and $X^n_b\in\X^n_b$ are the $n$-tiles that satisfy $X^n_w\cup X^n_b =A\in\mathbf{P}_n$ and $e_n=X^n_w\cap X^n_b$, then $q(A)\in e_n$. Thus in particular, $\deg_{f^n}(q(A))=1$ for each $A\in\mathbf{P}_n$.
Hence by construction, we have
\begin{equation} \label{eqSplitSumPreImgs}
\sum\limits_{x\in f^{-n}(z)} e^{S_n\phi(x)} = \sum\limits_{A\in \mathbf{P}^\delta_n} e^{S_n\phi(q(A))} + \sum\limits_{A\in \mathbf{P}_n\setminus\mathbf{P}^\delta_n} e^{S_n\phi(q(A))},
\end{equation}
and
\begin{align} \label{eqSplitSumPerPts}
& \sum\limits_{A\in \mathbf{P}^\delta_n} e^{S_n\phi(p(A))} \leq \sum\limits_{x\in P_{1,f^n}} w_n(x) e^{S_n\phi(x)}\\
\leq & \sum\limits_{A\in \mathbf{P}^\delta_n} e^{S_n\phi(p(A))} + \sum\limits_{A\in \mathbf{P}_n\setminus\mathbf{P}^\delta_n} \sum\limits_{x\in A\cap P_{1,f^n}} e^{S_n\phi(x)}. \notag
\end{align}
The last inequality in (\ref{eqSplitSumPerPts}) is due to the fact that if $x\in P_{1,f^n}$ satisfies $\deg_{f^n}(x) \geq 2$, then $x\in\V^n$ with $x\notin \bigcup \mathbf{P}^\delta_n$, and the number of $A\in \mathbf{P}_n$ that contains $x$ is at least $\deg_{f^n}(x)$ (and at most $2\deg_{f^n}(x)$).
By (\ref{eqSnPhiBound}) in Lemma~\ref{lmSnPhiBound}, we get
\begin{equation} \label{eqQuotSumAwayC}
\frac{1}{C_3} \leq \frac{\sum\limits_{A\in \mathbf{P}^\delta_n} e^{S_n\phi(p(A))}}{\sum\limits_{A\in \mathbf{P}^\delta_n} e^{S_n\phi(q(A))}} \leq C_3,
\end{equation}
and since in addition, $\card(A\cap P_{1,f^n}) \leq 2$ for $A\in\mathbf{P}_n$ by Lemma~\ref{lmAtMostOneCritPerTile}, we have
\begin{equation} \label{eqQuotSumNearC}
\frac{\sum\limits_{A\in \mathbf{P}_n \setminus \mathbf{P}^\delta_n} \sum\limits_{x\in A\cap P_{1,f^n}} e^{S_n\phi(x)}}{\sum\limits_{A\in \mathbf{P}_n \setminus \mathbf{P}^\delta_n} e^{S_n\phi(q(A))}} \leq 2 C_3,
\end{equation}
where
$$
C_3 = \exp\(C_1 \(\diam_d(S^2)\)^\alpha\),
$$
and $C_1>0$ is a constant from Lemma~\ref{lmSnPhiBound}. Both $C_1$ and $C_3$ depend only on $f$, $\mathcal{C}$, $d$, $\phi$, and $\alpha$.
By (\ref{eqSplitSumPreImgs}), (\ref{eqDistrPreImgs}), and (\ref{eqNbhdCFewPreImgs}), we get
\begin{equation} \label{eqSumPreImgsComparable}
\sum\limits_{x\in f^{-n}(z)} e^{S_n\phi(x)}
\geq \sum\limits_{A\in \mathbf{P}^\delta_n} e^{S_n\phi(q(A))}
\geq \frac{9}{10} \sum\limits_{x\in f^{-n}(z)} e^{S_n\phi(x)}.
\end{equation}
Hence, by (\ref{eqSplitSumPerPts}), (\ref{eqQuotSumAwayC}), and (\ref{eqSumPreImgsComparable}), we have
\begin{equation*}
\frac{\sum\limits_{x\in P_{1,f^n}} w_n(x) e^{S_n\phi(x)}}{\sum\limits_{x\in f^{-n}(z)}\deg_{f^n}(x) e^{S_n\phi(x)}}
\geq \frac{\sum\limits_{A\in \mathbf{P}^\delta_n} e^{S_n\phi(p(A))}}{\frac{10}{9} \sum\limits_{A\in \mathbf{P}^\delta_n} e^{S_n\phi(q(A))} }
\geq \frac{9}{10C_3}.
\end{equation*}
On the other hand, by (\ref{eqSplitSumPreImgs}), (\ref{eqSplitSumPerPts}), (\ref{eqQuotSumAwayC}), (\ref{eqQuotSumNearC}), and (\ref{eqSumPreImgsComparable}), we get
\begin{align*}
& \frac{\sum\limits_{x\in P_{1,f^n}} w_n(x) e^{S_n\phi(x)}}{\sum\limits_{x\in f^{-n}(z)}\deg_{f^n}(x) e^{S_n\phi(x)}} \leq \frac{ \sum\limits_{A\in \mathbf{P}^\delta_n} e^{S_n\phi(p(A))} + \sum\limits_{A\in \mathbf{P}_n\setminus\mathbf{P}^\delta_n} \sum\limits_{x\in A\cap P_{1,f^n}} e^{S_n\phi(x)} }{ \sum\limits_{x\in f^{-n}(z)} e^{S_n\phi(x)} } \\
\leq & \frac{\sum\limits_{A\in \mathbf{P}^\delta_n} e^{S_n\phi(p(A))}}{\sum\limits_{A\in \mathbf{P}^\delta_n} e^{S_n\phi(q(A))}} + \frac{ \sum\limits_{A\in \mathbf{P}_n\setminus\mathbf{P}^\delta_n} \sum\limits_{x\in A\cap P_{1,f^n}} e^{S_n\phi(x)} }{ 10 \sum\limits_{A\in \mathbf{P}_n\setminus\mathbf{P}^\delta_n} e^{S_n\phi(q(A))} } \leq C_3 + \frac{2}{10}C_3.
\end{align*}
Thus (\ref{eqSumsComparable}) holds if we choose $C=2 C_3$ and $n>N_2$. The proof is now complete.
\end{proof}
\subsection{Proof of large deviation principles} \label{subsctProofLDP}
\begin{proof}[Proof of Theorem~\ref{thmLDP}]
Let $\phi \in \Holder{\alpha}(S^2,d)$ for some $\alpha\in(0,1]$.
We apply Theorem~\ref{thmAbsLargeDeviationPrincipleCRL} with $X=S^2$, $g=f$, and $H=\Holder{\alpha}(S^2,d)$. Note that $\Holder{\alpha}(S^2,d)$ is dense in $C(S^2)$ with respect to the uniform norm (see for example, \cite[Lemma~6.12]{Li14}). Theorem~\ref{thmExistenceUniquenessES} implies Condition~(ii) in the hypothesis of Theorem~\ref{thmAbsLargeDeviationPrincipleCRL}. Condition~(i) follows from Corollary~\ref{corUSC}, (\ref{eqVPEntropy}), and the fact that $h_{\operatorname{top}}(f) = \log(\deg f)$ \cite[Corollary~20.8]{BM10}.
It now suffices to verify (\ref{eqThmCCondition}) for each of the sequences $\{\Omega_n(x_n)\}_{n\in\N}$ and $\{\Omega_n\}_{n\in\N}$ of Borel probability measures on $\mathcal{P}(S^2)$.
Fix an arbitrary $\psi \in \Holder{\alpha}(S^2,d)$.
By (\ref{eqPressureLimitPreImgsWODeg}) in Proposition~\ref{propPressureLimitPreImgs},
\begin{align*}
& \lim\limits_{n\to+\infty} \frac{1}{n} \log \int_{\mathcal{P}(S^2)} \! \exp \(n\int\!\psi \,\mathrm{d}\mu \) \,\mathrm{d}\Omega_n(x_n)(\mu) \\
= & \lim\limits_{n\to+\infty} \frac{1}{n} \log \sum\limits_{y\in f^{-n}(x_n)} \frac{ w_n(y) \exp(S_n\phi(y)) }{\sum_{z\in f^{-n}(x_n)} w_n(z) \exp(S_n\phi(z))} e^{\sum_{i=0}^{n-1} \psi\(f^i(y)\) } \\
= & \lim\limits_{n\to+\infty} \frac{1}{n} \Bigg( \log \sum\limits_{y\in f^{-n}(x_n)} w_n(y) e^{S_n(\phi+\psi)(y)} - \log \sum\limits_{z\in f^{-n}(x_n)} w_n(z) e^{S_n(\phi)(z)} \Bigg) \\
= & P(f,\phi+\psi) - P(f,\phi).
\end{align*}
Similarly, by (\ref{eqPressureLimitPeriodicPts}) in Proposition~\ref{propPressureLimitPeriodicPts}, we get
$$
P(f,\phi+\psi) - P(f,\phi)
= \lim\limits_{n\to+\infty} \frac{1}{n} \log \int_{\mathcal{P}(S^2)} \! \exp \(n\int\!\psi \,\mathrm{d}\mu \) \,\mathrm{d}\Omega_n(\mu)
$$
The theorem now follows from Theorem~\ref{thmAbsLargeDeviationPrincipleCRL}.
\end{proof}
\subsection{Equidistribution with respect to the equilibrium state} \label{subsctEquidistr}
\begin{proof}[Proof of Corollary~\ref{corMeasTheoPressure}]
We prove the first equality in (\ref{eqMeasTheoPressure}) now.
Fix $\mu \in \mathcal{M}(S^2,f)$ and a convex local basis $G_\mu$ at $\mu$. By (\ref{eqRateFnI}) and the upper semi-continuity of $h_\mu(f)$ (Corollary~\ref{corUSC}), we get
$$
-I^\phi(\mu) = \inf\limits_{\mathfrak{G}\inG_\mu}\bigg(\sup\limits_\mathfrak{G} (-I^\phi)\bigg) = \inf\limits_{\mathfrak{G}\inG_\mu}\(-\inf\limits_\mathfrak{G} I^\phi\).
$$
Then by (\ref{eqRateFnI}) and (\ref{eqInfI}),
\begin{align*}
& -P(f,\phi) + \int\!\phi\,\mathrm{d}\mu + h_\mu(f) = -I^\phi(\mu) = \inf\limits_{\mathfrak{G}\inG_\mu}\(-\inf\limits_\mathfrak{G} I^\phi\) \\
=& \inf\limits_{\mathfrak{G}\inG_\mu} \Bigg\{ \lim\limits_{n\to+\infty} \frac{1}{n}\log \sum\limits_{y\in f^{-n}(x_n),\, W_n(y)\in\mathfrak{G}} \frac{w_n(y)\exp(S_n\phi(y))}{Z_n(\phi)} \Bigg\},
\end{align*}
where we write $Z_n(\phi)= \sum\limits_{z\in f^{-n}(x_n)} w_n(z) \exp (S_n\phi(z))$. By (\ref{eqPressureLimitPreImgsWODeg}) in Proposition~\ref{propPressureLimitPreImgs}, we have $P(f,\phi)=\lim\limits_{n\to+\infty} \frac{1}{n} \log Z_n(\phi)$. Thus the first equality in (\ref{eqMeasTheoPressure}) follows.
\smallskip
By similar arguments, with (\ref{eqPressureLimitPreImgsWDeg}) in Proposition~\ref{propPressureLimitPreImgs} replaced by (\ref{eqPressureLimitPeriodicPts}) in Proposition~\ref{propPressureLimitPeriodicPts}, we get the second equality in (\ref{eqMeasTheoPressure}).
\end{proof}
\begin{proof}[Proof of Corollary~\ref{corEquidistr}]
Recall that $W_n(x)=\frac{1}{n}\sum\limits_{i=0}^{n-1} \delta_{f^i(x)} \in \mathcal{P}(S^2)$ for $x\in S^2$ and $n\in\N$ as defined in (\ref{eqDefWn}). We write
$$
Z^+_n(\mathfrak{G})=\sum\limits_{y\in f^{-n}(x_n), \, W_n(y)\in\mathfrak{G}} \deg_{f^n}(y)\exp(S_n\phi(y))
$$
and
$$
Z^-_n(\mathfrak{G})=\sum\limits_{y\in f^{-n}(x_n), \, W_n(y)\notin\mathfrak{G}} \deg_{f^n}(y)\exp(S_n\phi(y))
$$
for each $n\in\N$ and each open set $\mathfrak{G}\subseteq \mathcal{P}(S^2)$.
Let $G_{\mu_\phi}$ be a convex local basis of $\mathcal{P}(S^2)$ at $\mu_\phi$. Fix an arbitrary convex open set $\mathfrak{G}\inG_{\mu_\phi}$.
By the uniqueness of the equilibrium state in our context and Corollary~\ref{corMeasTheoPressure}, we get that for each $\mu\in\mathcal{P}(S^2)$, there exist numbers $a_\mu < P(f,\phi)$ and $N_\mu \in\N$ and an open neighborhood $\mathfrak{U}_\mu \in \mathcal{P}(S^2)\setminus \{\mu_\phi\}$ containing $\mu$ such that for each $n>N_\mu$,
\begin{equation} \label{eqZn<}
Z_n^+(\mathfrak{U}_\mu) \leq \exp(n a_\mu).
\end{equation}
Since $\mathcal{P}(S^2)$ is compact in the weak$^*$ topology by Alaoglu's theorem, so is $\mathcal{P}(S^2)\setminus \mathfrak{G}$. Thus there exists a finite set $\{\mu_i\,|\, i\in I\} \subseteq \mathcal{P}(S^2)\setminus \mathfrak{G}$ such that
\begin{equation} \label{eqUcover}
\mathcal{P}(S^2)\setminus \mathfrak{G} \subseteq \bigcup\limits_{i\in I} \mathfrak{U}_{\mu_i}.
\end{equation}
Here $I$ is a finite index set. Let $a=\max\{ a_{\mu_i} \,|\, i\in I\}$. Note that $a<P(f,\phi)$. By Corollary~\ref{corMeasTheoPressure} with $\mu=\mu_\phi$, we get that
\begin{equation} \label{eqPfcorEquidistr1}
P(f,\phi_0) \leq \lim\limits_{n\to+\infty} \frac{1}{n} \log Z^+_n(\mathfrak{G}).
\end{equation}
Combining (\ref{eqPfcorEquidistr1}) with (\ref{eqPressureLimitPreImgsWDeg}) in Proposition~\ref{propPressureLimitPreImgs}, we get that the equality holds in (\ref{eqPfcorEquidistr1}). So there exist numbers $b\in (a,P(f,\phi))$ and $N\geq \max\{ N_i \,|\, i\in I\}$ such that for each $n> N$,
\begin{equation} \label{eqZn>}
Z_n^+(\mathfrak{G}) \geq \exp(nb).
\end{equation}
We claim that every subsequential limit of $\{\nu_n\}_{n\in\N}$ in the weak$^*$ topology lies in the closure $\overline{\mathfrak{G}}$ of $\mathfrak{G}$. Assuming that the claim holds, then since $\mathfrak{G}\inG_{\mu_\phi}$ is arbitrary, we get that any subsequential limit of $\{\nu_n\}_{n\in\N}$ in the weak$^*$ topology is $\mu_\phi$, i.e., $\nu_n \stackrel{w^*}{\longrightarrow} \mu_\phi$ as $n\longrightarrow +\infty$.
\smallskip
We now prove the claim. We first observe that for each $n\in\N$,
\begin{align*}
\nu_n & = \sum\limits_{y\in f^{-n}(x_n)} \frac{w_n(y) \exp(S_n\phi(y))}{ Z^+_n(\mathfrak{G}) + Z^-_n(\mathfrak{G})} W_n(y) \\
& = \frac{Z^+_n(\mathfrak{G})}{Z^+_n(\mathfrak{G}) + Z^-_n(\mathfrak{G})} \nu'_n + \sum\limits_{y\in f^{-n}(x_n),\,W_n(y)\notin\mathfrak{G}} \frac{w_n(y) e^{S_n\phi(y)}}{ Z^+_n(\mathfrak{G}) + Z^-_n(\mathfrak{G})} W_n(y),
\end{align*}
where $\nu'_n = \sum\limits_{y\in f^{-n}(x_n),\,W_n(y)\in\mathfrak{G}} \frac{w_n(y) \exp(S_n\phi(y))}{ Z^+_n(\mathfrak{G})} W_n(y)$.
Note that since $a<b$, by (\ref{eqUcover}), (\ref{eqZn<}), and (\ref{eqZn>}),
$$
0\leq \lim\limits_{n\to+\infty} \frac{Z_n^-(\mathfrak{G})}{Z_n^+(\mathfrak{G})} \leq \lim\limits_{n\to+\infty} \frac{\sum_{i\in I} Z_n^+(\mathfrak{U}_i)}{Z_n^+(\mathfrak{G})} \leq \lim\limits_{n\to+\infty} \frac{\card(I)\exp(na)}{\exp(nb)} = 0.
$$
So $\lim\limits_{n\to+\infty} \frac{Z^+_n(\mathfrak{G})}{Z^+_n(\mathfrak{G}) + Z^-_n(\mathfrak{G})} = 1$, and that the total variation
\begin{align*}
& \Bigg\|{\sum\limits_{y\in f^{-n}(x_n),\,W_n(y)\notin\mathfrak{G}} \frac{w_n(y) \exp(S_n\phi(y))}{ Z^+_n(\mathfrak{G}) + Z^-_n(\mathfrak{G})} W_n(y)} \Bigg\| \\
\leq & \frac{\sum\limits_{y\in f^{-n}(x_n),\,W_n(y)\notin\mathfrak{G}} w_n(y) \exp(S_n\phi(y)) \Norm{W_n(y)}}{Z^+_n(\mathfrak{G}) + Z^-_n(\mathfrak{G})} \\
\leq & \frac{Z^-_n(\mathfrak{G})}{Z^+_n(\mathfrak{G}) + Z^-_n(\mathfrak{G})} \longrightarrow 0
\end{align*}
as $n\longrightarrow +\infty$. Thus a measure is a subsequential limit of $\{\nu_n\}_{n\in\N}$ if and only if it is a subsequential limit of $\{\nu'_n\}_{n\in\N}$. Note that $v'_n$ is a convex combination of measures in $\mathfrak{G}$, and $\mathfrak{G}$ is convex, so $\nu'_n\in\mathfrak{G}$, for $n\in\N$. Hence each subsequential limit of $\{\nu_n\}_{n\in\N}$ lies in the closure $\overline{\mathfrak{G}}$ of $\mathfrak{G}$. The proof of the claim is complete now.
\smallskip
By similar arguments as in the proof of the convergence of $\{\nu_n\}_{n\in\N}$ above, with (\ref{eqPressureLimitPreImgsWODeg}) in Proposition~\ref{propPressureLimitPreImgs} replaced by (\ref{eqPressureLimitPeriodicPts}) in Proposition~\ref{propPressureLimitPeriodicPts}, we get that $\eta_n \stackrel{w^*}{\longrightarrow} \mu_\phi$ as $n\longrightarrow +\infty$.
\end{proof}
|
\section{Some examples and best-practices}
\label{sec:intro}
For internal references use label-refs: see section~\ref{sec:intro}.
Bibliographic citations can be done with cite: refs.~\cite{a,b,c}.
When possible, align equations on the equal sign. The package
\texttt{amsmath} is already loaded. See \eqref{eq:x}.
\begin{equation}
\label{eq:x}
\begin{split}
x &= 1 \,,
\qquad
y = 2 \,,
\\
z &= 3 \,.
\end{split}
\end{equation}
Also, watch out for the punctuation at the end of the equations.
If you want some equations without the tag (number), please use the available
starred-environments. For example:
\begin{equation*}
x = 1
\end{equation*}
The amsmath package has many features. For example, you can use use
\texttt{subequations} environment:
\begin{subequations}\label{eq:y}
\begin{align}
\label{eq:y:1}
a & = 1
\\
\label{eq:y:2}
b & = 2
\end{align}
and it will continue to operate across the text also.
\begin{equation}
\label{eq:y:3}
c = 3
\end{equation}
\end{subequations}
The references will work as you'd expect: \eqref{eq:y:1},
\eqref{eq:y:2} and \eqref{eq:y:3} are all part of \eqref{eq:y}.
A similar solution is available for figures via the \texttt{subfigure}
package (not loaded by default and not shown here).
All figures and tables should be referenced in the text and should be
placed at the top of the page where they are first cited or in
subsequent pages. Positioning them in the source file
after the paragraph where you first reference them usually yield good
results. See figure~\ref{fig:i} and table~\ref{tab:i}.
\begin{figure}[tbp]
\centering
\includegraphics[width=.45\textwidth,trim=0 380 0 200,clip]{img1.pdf}
\hfill
\includegraphics[width=.45\textwidth,origin=c,angle=180]{img2.pdf}
\caption{\label{fig:i} Always give a caption.}
\end{figure}
\begin{table}[tbp]
\centering
\begin{tabular}{|lr|c|}
\hline
x&y&x and y\\
\hline
a & b & a and b\\
1 & 2 & 1 and 2\\
$\alpha$ & $\beta$ & $\alpha$ and $\beta$\\
\hline
\end{tabular}
\caption{\label{tab:i} We prefer to have borders around the tables.}
\end{table}
We discourage the use of inline figures (wrapfigure), as they may be
difficult to position if the page layout changes.
We suggest not to abbreviate: ``section'', ``appendix'', ``figure''
and ``table'', but ``eq.'' and ``ref.'' are welcome. Also, please do
not use \texttt{\textbackslash emph} or \texttt{\textbackslash it} for
latin abbreviaitons: i.e., et al., e.g., vs., etc.
\section{Sections}
\subsection{And subsequent}
\subsubsection{Sub-sections}
\paragraph{Up to paragraphs.} We find that having more levels usually
reduces the clarity of the article. Also, we strongly discourage the
use of non-numbered sections (e.g.~\texttt{\textbackslash
subsubsection*}). Please also see the use of
``\texttt{\textbackslash texorpdfstring\{\}\{\}}'' to avoid warnings
from the hyperref package when you have math in the section titles
\section{Introduction}
Cosmological observations point to a primordial universe that can be effectively described by an approximate de Sitter phase. The details of this inflationary phase are to a large extent still unknown, but the most recent Planck data does rule out a large fraction of parameter space, giving us some hints about the underlying physics \cite{Planck:2013}. From a theoretical point of view it is probably fair to say that inflation is poorly understood. Indeed, few convincing theoretical constraints on the inflationary parameter space exist that would identify natural, UV consistent, models. Trying to embed inflation into a fundamental description like string theory is notoriously difficult, but the last decade has seen considerable phenomenological progress in that direction (for a thorough discussion see \cite{BauMcA:2014} and references therein). The absence of a guiding principle which is able to rule out a significant fraction of inflationary models, combined with the attractive features of eternal inflation, has fueled the idea that perhaps our inflationary universe is just one realization in a huge landscape of bubble universes that are continuously being produced as a consequence of a stochastically varying scalar field during a phase of eternal inflation\cite{Vilenkin:1983, Susskind:2003}.
This exotic possibility makes the general prediction that the spatial sections in our universe should be hyperbolic on the largest scales that start probing the boundary of the bubble. Clearly this general prediction is hard, if not impossible, to verify because the primordial inflationary expansion typically redshifts the negative curvature scale far beyond the observable universe \cite{Guth:2012}, although in the inflationary landscape relatively short phases of slow-roll inflation might be preferred \cite{Freivogel:2006}, which could lead to observable consequences in the CMB temperature correlations at low multipoles \cite{YamLinNarSasTan:2011}. More particular predictions include a bubble universe that might, under fortuitous conditions, provide an explanation for the low power anomaly at low $l$ due to the steepening of the slope right after penetration of the barrier \cite{BouHarSen:2013}. Another potential consequence of observational interest is that bubble universes can in principle collide \cite{Kleban:2009, Kleban:2011, Kleban:2012}, which could leave definite non-isotropic signatures in the Cosmic Microwave Background sky. Unfortunately the chance for a collision to have taken place in our past is small and model-dependent, and therefore not seeing this effect will not be able to rule out a multiverse origin \cite{Peiris:2011}.
Recently, another rather generic consequence of a multiverse origin has been explored \cite{Agullo:2010, Komatsu:2013, Kanno:2014}. In the context of an inflationary landscape one would expect the initial vacuum state for quantum fluctuations in a single inflationary bubble to be entangled with the rest of the eternally inflating universe, leading to a mixed state inside the bubble. Since the Cosmic Microwave Background temperature anisotropies (as well as the large scale structure distribution) is probing the statistics of these inflationary quantum fluctuations, one could imagine uncovering evidence in favor of a mixed initial state that would support the idea that our universe originated from false vacuum decay. This idea warrants a careful study of the actual (observational) potential to constrain departures away from a standard pure initial state for inflationary quantum fluctuations and how these departures relate to the global vacuum state of eternal inflation.
In this work we present a first step, triggered by some recent work in this direction \cite{Kanno:2014}, in clarifying the connection between the vacuum state of the false, eternally inflating, vacuum and potential departures from the standard Bunch--Davies vacuum state in a hyperbolic bubble. As a starting point we identify a limit that connects hyperbolic coordinates and its naturally associated vacuum to the de Sitter invariant Bunch--Davies state on planar sections. Following up on older work in the context of open inflation, after selecting the global Bunch--Davies vacuum we then use the mixed reduced density matrix defined in a single hyperbolic coordinate patch of de Sitter space to explicitly show that the statistics of inflationary quantum fluctuations are indistinguishable from the standard planar Bunch--Davies predictions at late times and we explain why this should not come as a surprise. The mixed nature of the initial state in hyperbolic de Sitter space in fact ensures that all predictions match with those of the pure Bunch--Davies state in the planar de Sitter space that one started out with. The mixed nature of the initial hyperbolic vacuum state therefore, in this case, does not imply observable departures from the standard planar Bunch-Davies predictions for the power- and bi-spectrum. Selecting instead the pure hyperbolic vacuum does lead to differences that are however strongly suppressed in the curvature scale, as we again show explicitly by computing the power-- and bi--spectrum for scalar field vacuum fluctuations. We end with a discussion on the implications in the context of (open) inflationary models and some remaining open questions that we hope to return to in future work.
\section{A family of de Sitter hyperbolic sections and their vacua}
We will start by constructing a limit in the family of de Sitter hyperbolic sections that reduces to the planar description of de Sitter space. This allows us to explicitly see the distinction between the natural vacuum state on a generic hyperbolic section and the Bunch-Davies vacuum state as defined using planar coordinates, which can be understood as a singular (but well-defined) limit of the natural hyperbolic vacuum.
\subsection{A generalized hyperbolic embedding}
Let us remind the reader that $4$-dimensional de Sitter space can be defined as the embedding surface in $5$-dimensional flat space defined by
\begin{equation}
-X_0^2 + X_1^2 + X_2^2 + X_3^2 + X_4^2 = 1 \,
\label{def-ds}
\end{equation}
where the de Sitter curvature length scale has been normalized to one. This embedding equation is clearly invariant under $SO(1,4)$ transformations, corresponding to the isometry group of $dS_4$.
For our purposes we will be interested in two different coordinate sets on this embedding surface that both belong to the class of isotropic and homogeneous FRLW spaces. The standard coordinate set used for describing inflation is the planar one, identifying flat spatial sections. In what follows, we will suppress the two (spatial) coordinates $X_3$ and $X_4$ for purposes of efficiency, effectively suppressing an $S^2$ in the de Sitter space. The planar coordinates are defined as
\begin{eqnarray}
X_0 + X_1 &=& e^{t_p} \, \nonumber \\
X_0 - X_1 &=& \left( r_p^2 \, e^{2t_p} -1 \right) e^{-t_p} \, \\
X_2 &=& r_p \, e^{t_p} \, \nonumber
\label{planar}
\end{eqnarray}
leading to the well-known planar expression for the induced de Sitter metric
\begin{equation}
ds^2 = -dt_p^2 + e^{2t_p} \left[ dr_p^2 + r_p^2 \, d\Omega_2^2 \right] \,
\label{planarmetric}
\end{equation}
with $-\infty < t_p < +\infty$ and $r>0$, and we reinserted the $S^2$ part. Since $X_0+X_1 \geq 0$ these coordinates only cover the upper half diagonal part in the $X_0$ versus $X_1$ plane. Besides the obvious $SO(3)$ isometries, the boost symmetries of the embedding space are realized on the planar metric as an isometry involving a particular combination of time translation and spatial scaling\footnote{More precisely, it corresponds with the isometry $t\rightarrow t+\gamma$ and $r\rightarrow e^{-\gamma}r$ of the planar de Sitter metric.}. Because inflation redshifts away any existing spatial curvature present initially, this coordinate set should be an excellent approximation to derive the late-time effects of a sustained phase of cosmological inflation. Nevertheless, one could imagine a situation where our universe has originated from a tunneling event out of an eternally inflating false vacuum\footnote{Moreover, in the nineties models of open inflation were of particular interest, independent of whether their origin was due to tunneling \cite{BucGolTur:1994, LytSte:1990}.}. The nucleated bubble would have negatively curved spatial sections \cite{ColLuc:1980, Gott:1982}, leading to the hyperbolic coordinate set
\begin{eqnarray}
X_0 + X_1 &=& \cosh{t_h} +\sinh{t_h} \, \cosh{r_h} \, \nonumber \\
X_0 - X_1 &=& - \cosh{t_h} +\sinh{t_h} \, \cosh{r_h} \, \\
X_2 &=& \sinh{t_h} \, \sinh{r_h} \, . \nonumber
\label{hyperbolic}
\end{eqnarray}
With these identifications the induced hyperbolic de Sitter metric reads
\begin{equation}
ds^2 = -dt_h^2 + \sinh{t_h}^2 \left[ dr_h^2 + \sinh^2{r_h} \, d\Omega_2^2 \right] \,
\label{hyperbolicmetric}
\end{equation}
where $0 \leq t_h<+\infty$ and $r_h>0$, and as before we included the full $S^2$ that was left out in the embedding coordinate identification. The coordinate singularity at $t_h=0$ can be interpreted in the context of false vacuum decay as the creation of the open inflationary bubble. Note that $t_h=0$ corresponds to $X_0=0$ (and $X_1=1$, $X_2=0$): the bubble nucleation time from the point of view of the embedding space. The spatial sections correspond to constant negative curvature slices that exhibit an $SO(1,3)$ isometry. We will in fact be interested in a one-parameter generalization of this hyperbolic coordinate embedding, obtained by boosting in the $X_0$--$X_1$ plane of the $5$--dimensional embedding space. Combined with rotations these transformation allow one to move the `nucleation' time of the hyperbolic bubble to any specific point on the embedding surface. Just performing a Lorentz boost in the $X_0$--$X_1$ plane will change the nucleation time (and position in $X_1$), which yields the following generalized hyperbolic coordinate set
\begin{eqnarray}
X_0 + X_1 &=& e^{-\gamma} \, \left[ \cosh{t_h} +\sinh{t_h} \, \cosh{r_h} \right] \, \nonumber \\
X_0 - X_1 &=& e^{\gamma} \, \left[ -\cosh{t_h} +\sinh{t_h} \, \cosh{r_h} \right]\, \\
X_2 &=& \sinh{t_h} \, \sinh{r_h} \, \nonumber
\label{gammahyperbolic}
\end{eqnarray}
where $\gamma$ is the boost parameter. This generalized hyperbolic solution of the embedding equation will of course lead to the same induced metric, but the nucleation time and position of the associated bubble in the embedding space have now shifted to $X_0 = -\sinh{\gamma}$ and $X_1= \cosh{\gamma}$ respectively. Moreover, since $t_h \geq 0$ one finds that $X_0 + X_1 \geq e^{-\gamma}$, restricting the hyperbolic section to the upper right diagonal part in the $X_0$ versus $X_1$ plane, which overlaps with, but for any finite $\gamma$ is smaller than, the part of de Sitter covered by planar coordinates. This is depicted in figure \ref{fig:hyperbolicboost}. One can verify that in the limit of infinite $\gamma$ the planar and hyperbolic coordinates cover the same region of de Sitter space, which is consistent with the observation that in this limit the hyperbolic nucleation time in the embedding space is shifted to $X_0 \rightarrow -\infty$.
\begin{figure}[ht]
\centering
\begin{minipage}[t]{\textwidth}
\centering
\includegraphics[width=0.5\textwidth]{fig4.pdf}\caption{Conformal diagram of $dS_4$ with the left- and right-hyperbolic patch as the upper-left resp. upper-right triangles. The dashed line is the unboosted situation $\gamma=0$. For finite $\gamma$ (solid line), we see that that nucleation time of the left bubble gets pushed to earlier times, and vice versa for the right bubble. In the limit of $\gamma\rightarrow \infty$, we can see that the left bubble will cover the entire upper-left triangle of the conformal diagram, coinciding with the planar patch.}\label{fig:hyperbolicboost}
\end{minipage}
\end{figure}
The generalization of the hyperbolic coordinate set introduced above allows us to explicitly relate the planar and hyperbolic sections of de Sitter space. Since the two coordinate sets cover the same region in the $\gamma \rightarrow \infty$ limit, there should exist a one-to-one mapping between the coordinates in that limit. More precisely, we would like to introduce a new set of hyperbolic coordinates that are to be kept fixed in the limit $\gamma \rightarrow \infty$, and that in the limit exactly reproduce the planar coordinate embedding solution. Note that any (constant) shift or rescaling of the hyperbolic embedding coordinates is still a solution of the embedding equation, but will change the expression for the induced metric. Since we expect the range of the hyperbolic time coordinate to be extended to $-\infty$ and the negative curvature to be scaled away, we redefine
\begin{equation}
\tilde t_h \equiv t_h -\gamma ~ ; ~ \tilde r_h \equiv \frac{1}{2} r_h \, e^\gamma \, .
\label{fixedcoord}
\end{equation}
This leaves us with the following generalized hyperbolic solution to the embedding equation
\begin{eqnarray}
X_0 + X_1 &=& e^{-\gamma} \, \left[ \cosh{(\tilde t_h +\gamma)} +\sinh{(\tilde t_h + \gamma)} \, \cosh{(2\tilde r_h \, e^{-\gamma})} \right] \, \nonumber \\
X_0 - X_1 &=& - e^{\gamma} \, \left[ \cosh{(\tilde t_h + \gamma)} - \sinh{(\tilde t_h +\gamma)} \, \cosh{(2\tilde r_h \, e^{-\gamma})} \right]\, \\
X_2 &=& \sinh{(\tilde t_h +\gamma)} \, \sinh{(2\tilde r_h \, e^{-\gamma})} \, \nonumber
\label{gammahyperbolictilde}
\end{eqnarray}
where $-\gamma \leq \tilde t_h < +\infty$. For finite $\gamma$ the shift in hyperbolic time and the rescaling of the hyperbolic radius (or equivalently the inverse rescaling of hyperbolic momentum) does obviously not affect any hyperbolic patch observables, but it does allow one to analyze the infinite boost limit in a simple and useful way. The induced hyperbolic metric now reads
\begin{eqnarray}
ds^2 &=& -d\tilde t_h^2 + \sinh{(\tilde t_h + \gamma)}^2 \left[ 4 e^{-2\gamma} \, d\tilde r_h^2 + \sinh{(2 \tilde r_h e^{-\gamma})}^2 \, d\Omega_2^2 \right] \label{hyperbolictildemetric} \\
&\overset{\gamma \rightarrow \infty }= & -d\tilde t_h^2 + e^{2 \tilde t_h} \left[d\tilde r_h^2 + \tilde r_h^2 \, d\Omega_2^2 \right] \, . \nonumber
\end{eqnarray}
In the second line we performed the limit $\gamma \rightarrow \infty$, keeping $\tilde t_h$ and $\tilde r_h$ fixed, showing that (\ref{gammahyperbolictilde}) exactly reduces to the planar embedding solution (\ref{planar}). Note that all the $\gamma$ dependence in the induced metric is removed and one is left with precisely the planar line-element (\ref{planarmetric}) in terms of the coordinates $\tilde t_h$ and $\tilde r_h$.
For any finite boost parameter $\gamma$ global de Sitter space is covered by two (adjacent) hyperbolic sections, see figure \ref{fig:hyperbolicboost}. The other hyperbolic embedding can be obtained by changing the sign of $X_1$, resulting in the interchange of the expressions for $X_0+X_1$ and $X_0-X_1$ in (\ref{hyperbolic}).
Acting with the same boost on this second hyperbolic embedding results in the opposite effect, moving the nucleation time to $X_0 \rightarrow +\infty$. The opposite minus infinity boost should instead reduce to another planar section (with $X_0+X_1$ and $X_0-X_1$ in (\ref{planar}) interchanged), suggesting that the redefined coordinates in this case should read
\begin{equation}
\tilde t_h \equiv t_h +\gamma ~ ; ~ \tilde r_h \equiv \frac{1}{2} r_h \, e^{-\gamma} \, .
\label{fixedcoord2}
\end{equation}
Putting this together we obtain for the adjacent hyperbolic section the following generalized embedding
\begin{eqnarray}
X_0 + X_1 &=& - e^{-\gamma} \, \left[ \cosh{(\tilde t_h - \gamma)} - \sinh{(\tilde t_h -\gamma)} \, \cosh{(2\tilde r_h \, e^{\gamma})} \right]
\, \nonumber \\
X_0 - X_1 &=& e^{\gamma} \, \left[ \cosh{(\tilde t_h - \gamma)} + \sinh{(\tilde t_h -\gamma)} \, \cosh{(2\tilde r_h \, e^{\gamma})} \right]
\, \\
X_2 &=& \sinh{(\tilde t_h -\gamma)} \, \sinh{(2\tilde r_h \, e^{\gamma})} \, \nonumber
\label{gammahyperbolictilde2}
\end{eqnarray}
where now $\gamma \leq \tilde t_h < +\infty$. The induced hyperbolic metric in this case is obtained by just replacing $\gamma$ with $-\gamma$ in (\ref{hyperbolictildemetric}). By construction the limit $\gamma \rightarrow \infty$ should instead collapse and in a sense remove the adjacent hyperbolic section. Clearly, in the opposite $\gamma \rightarrow -\infty$ the roles of the two hyperbolic sections are reversed.
Having established this explicit relation between hyperbolic and planar coordinates, we can now use it to better understand and connect their respective vacua, which should be different for any finite value of $\gamma$. In particular, the planar Bunch--Davies state is known to be equivalent to the unique and de Sitter invariant Euclidean vacuum\footnote{The invariance of the Bunch--Davies vacuum under de Sitter isometries strictly speaking fails for massless fields, but since this subtlety does not affect our results we will ignore it from now on.}. On the other hand, any pure hyperbolic vacuum state is defined on a negatively curved spatial slice that is not a de Sitter Cauchy surface. This means that the Bunch--Davies state in a single hyperbolic patch can only be described by an appropriately defined mixed state; see figures \ref{fig:hyperbolicpatches} and \ref{fig:lefthyperbolicpatch}. The mixed state defined on one of the two (conjugate) hyperbolic sections reproducing the Bunch--Davies state was first constructed in \cite{Sasaki:1995} and was subsequently used in \cite{MalPim:2012} to compute the reduced density matrix and the corresponding entanglement entropy for a single hyperbolic section.
\begin{figure}[h!]
\centering
\begin{minipage}[t]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig1.pdf}
\caption{Conformal diagram of $dS_4$ with the left- and right-hyperbolic patch shown. As neither patch contains a Cauchy slice of the full $dS_4$, restricting the Bunch-Davies vacuum to one of them will yield a mixed state.}\label{fig:hyperbolicpatches}
\end{minipage}
\quad
\begin{minipage}[t]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig2.pdf}
\caption{An observer confined to live in the left-hyperbolic patch (a bubble universe) can define his own pure hyperbolic vacuum. This state will differ significantly from the mixed state resulting from a restriction of the Bunch-Davies state to this bubble.}\label{fig:lefthyperbolicpatch}
\end{minipage}
\end{figure}
One application of the one-parameter family of hyperbolic de Sitter foliations is that one can confirm that the natural choice for a hyperbolic vacuum reduces to the planar Bunch--Davies state in the limit $\gamma \rightarrow \infty$. Secondly, one could attempt to generalize the entangled expression for wavefunctions of the Bunch--Davies state, with support on both the left and right hyperbolic section, and work out its dependence on the embedding boost parameter. In the $\gamma \rightarrow \infty$ limit this should reduce to the pure planar Bunch--Davies state, implying that the reduced density matrix carries some non-trivial $\gamma$ dependence to make sure the associated entanglement entropy vanishes in the strict $\gamma \rightarrow \infty$ limit.
To summarize, we established that the infinite boost limit of a hyperbolic de Sitter patch (and as a consequence also its corresponding vacuum state) reduces to the planar de Sitter patch (and the Bunch--Davies vacuum). This appears to be similar to an observation made in \cite{GreParScha:2006} where the static vacuum, understood as the empty state for a corresponding free-falling observer, was also argued to reduce to the Bunch--Davies state in the infinite boost limit. Note that to each hyperbolic patch one can associate a free-falling observer in one of the two center regions in between the hyperbolic patches that never intersects either one of them. These time-like curves are indeed connected to each other by the same embedding space boosts \cite{ParVer:2004}. To complete the argument one needs to confirm that the static vacuum state associated to this free-falling observer is connected to the hyperbolic vacuum state. Note that (for $\gamma=0$) the center region in between the hyperbolic patches is usually covered by coordinates that are obtained from the hyperbolic coordinates as follows $t_h = i (t_C - \frac{\pi}{2})$ and $r_h = r_C + i \frac{\pi}{2}$, resulting in the following center region metric
\begin{equation}
ds^2= dt_C^2 + \cos{t_C}^2 \left[-dr_C^2 + \cosh{r_C}^2 d\Omega^2\right] \, ,
\end{equation}
where $r_C$ is now a time-like coordinate. Each of the two center regions clearly identifies a causal diamond belonging to the free-falling observer of interest. To make this explicit, one notices that the coordinate transformation $r_s \equiv \sin{t_C}$ and $t_S \equiv r_C$ indeed reproduces the static patch metric, upon ignoring the two-dimensional sphere\footnote{That the $S^2$ part does not reproduce the standard static patch expression can be understood by realizing that this static patch region is rotated by an angle $\pi/2$ with respect to the standard embedding. This affects the $S^2$ angles, which have to be transformed as well in order to obtain the complete static patch metric.}. This establishes a map from the hyperbolic to the static patch of a specific free-falling observer, relating their respective vacua and their behavior in the infinite (embedding) boost limit.
\subsection{The hyperbolic vacuum}\label{sectionhyperbolic}
After having established a limit to obtain the planar embedding and coordinates, let us now remind the reader of the standard positive frequency modes on a single hyperbolic patch \cite{Sasaki:1995}, as if it were the entire universe (see figure \ref{fig:lefthyperbolicpatch}). The scalar wave equation for the hyperbolic patch of de Sitter \eqref{hyperbolicmetric} reads
\begin{equation}
\left[\frac{1}{\sinh^3 t}\partial_t \sinh^3 t \partial_t - \frac{1}{\sinh^2 t} \nabla_{\mathcal{H}^3}^2 + m^2 \right]\phi=0
\label{eq:hyperboliceom}
\end{equation}
where we defined the Laplacian on the three-hyperboloid
\begin{equation}
\nabla_{\mathcal{H}^3}^2=\frac{1}{\sinh^2 r}\partial_r (\sinh^2 r \partial_r)+\frac{1}{\sinh^2 r}\nabla_{\mathcal{S}^2}^2 .
\label{eq:hyperboliclaplacian}
\end{equation}
A natural set of solutions to the hyperbolic equations of motion \eqref{eq:hyperboliceom} is given by
\begin{equation}\label{Hmodes}
\frac{1}{\sinh(t)}P^{ip}_{\nu-\frac{1}{2}}\left(\cosh (t)\right)Y_{plm}(r,\Omega)
\end{equation}
where it is customary to define $\nu = \sqrt{\frac{9}{4}-\frac{m^2}{H^2}}$. The quantum numbers $l,m$ label the usual $SO(3)$ irreps, and together with the continuous quantum number $p$ it completely specifies the hyperbolic momentum. Furthermore, $P^{ip}_{\nu-\frac{1}{2}}$ are the associated Legendre functions of the second kind and the $Y_{plm}$ are the orthonormal eigenfunctions of the hyperbolic Laplacian \eqref{eq:hyperboliclaplacian}
\begin{equation}
\nabla_{\mathcal{H}^3}^2 Y_{plm}(r,\Omega)=-(1+p^2)Y_{plm}(r,\Omega).
\end{equation}
For $\nu > \frac{1}{2}$, there is in fact a supplementary set of solutions with $p = i(\nu-\frac{1}{2})$ \cite{Sasaki:1995}. These so-called ``supercurvature modes" will not be of interest for the purposes that are considered here, where our main focus will be on potential signatures in the large (``subcurvature'') momentum limit. We refer to \cite{Cohn:1998} for an interesting account on the role and interpretation of these supercurvature modes.
Switching to conformal time $\eta$ we can write the metric of the hyperbolic slice as
\begin{equation}
\begin{aligned}
ds^2 &= \sinh^2(t(\eta))\left(-d\eta^2+dr^2+\sinh^2(r)d\Omega_2^2 \right),
\end{aligned}
\end{equation}
where $\eta = \ln (\tanh(\frac{t}{2}))$, or equivalently $\cosh t = -\frac{1}{\tanh \eta}$ and $-\infty<\eta<0$. In terms of the conformal time $\eta$ we find that in the far past $\eta \rightarrow -\infty$ and in the limit of large momenta $p \gg 1$ one obtains
\begin{equation}
\begin{aligned}
P^{ip}_{1}\left(-\frac{1}{\tanh \eta}\right) \propto e^{-ip\eta}\left(1-\frac{i}{p \tanh \eta}\right) \rightarrow e^{-i p\eta},
\end{aligned}
\end{equation}
so we can identify these mode functions with the ``natural hyperbolic vacuum": they define a state that is empty in the far past for large momenta, approaching the standard vacuum description in flat space. As expected, in the limit $\gamma \rightarrow \infty$ that we introduced in the previous section (\ref{fixedcoord2}) the mode functions reduce to the standard Bunch--Davies mode functions in flat slicing, explicitly connecting the hyperbolic and planar patch vacua in this limit
\begin{equation}
\begin{aligned}
\lim_{\gamma \rightarrow \infty} P^{ip}_{1}(\cosh (\tilde{t}+\gamma)) &\propto e^{-ip\tilde{\eta}}\left(1-\frac{i}{p\tilde{\eta}}\right).
\end{aligned}
\end{equation}
For all the details we refer the reader to the appendix \ref{app:modefunctionscalar}, but it should be clear that the tildes on the coordinates in the above equation relate to the redefined hyperbolic coordinates that are kept fixed in the infinite boost limit.
The mode functions (\ref{Hmodes}) must of course be properly normalized, enforcing $[\hat{b}_{plm},\hat{b}_{plm}^{\dagger}] = \delta_{ll'} \delta_{mm'} \delta(p-p')$, implying the following Klein-Gordon inner product
\begin{equation}
\begin{aligned}
\langle \phi_{plm},\phi_{plm}\rangle_{\text{KG}} &= \delta_{ll'} \delta_{mm'} \delta(p-p')
\end{aligned}
\end{equation}
giving (see appendix \ref{KGnorm})
\begin{equation}\label{Pnorm}
\begin{aligned}
N_{P^p}^2 &\equiv \langle P^{ip},P^{ip} \rangle_{\text{KG}} \\
&= \frac{2\sinh(\pi p)}{\pi}.
\end{aligned}
\end{equation}
With the help of (\ref{Pnorm}) we can now express the field operator in a single hyperbolic patch as (keeping in mind that we are ignoring supercurvature modes)
\begin{equation}\label{Hexpansion}
\begin{aligned}
\phi(t,r,\Omega) &= \int_0^\infty dp \sum_{l=0}^\infty \sum_{m=-l}^l\frac{1}{N_{P^p}}\frac{1}{\sinh(t)}\left(\hat{b}_{plm}P^{ip}_{\nu-\frac{1}{2}}\left(\cosh (t)\right)Y_{plm}(r,\Omega)+\text{h.c.}\right)
\end{aligned}
\end{equation}
defining the natural hyperbolic vacuum state $|\Omega_H\rangle$ as
\begin{equation}\label{Hvacuum}
\begin{aligned}
\hat{b}_{plm}|\Omega_H\rangle &= 0 &&\forall p,l,m.
\end{aligned}
\end{equation}
This hyperbolic vacuum state can be understood as a natural choice in an isolated (stand-alone) open inflationary universe, as it is empty in the far past and reduces to the planar Bunch--Davies vacuum state in the infinite boost limit. As we have elaborated upon in the previous section, because a similar statement can be made for the vacuum in a static patch \cite{GreParScha:2006}, and because the region in between the two hyperbolic patches contains a causal diamond region, this state can be analytically continued to the static (empty) vacuum for a free-falling observer that never intersects the two adjacent hyperbolic de Sitter patches.
Clearly this state is very different from the unique de Sitter invariant Bunch--Davies vacuum for generic $\gamma$, so one would expect anomalous behavior similar to what happens in the de Sitter static vacuum or the flat Rindler vacuum. To that end let us analyze the behavior of the energy momentum tensor in a (generic) hyperbolic vacuum. Note that for a flat Rindler wedge in lightcone coordinates $(u,v)$, there is a horizon at $u=0$ and the Fulling-Rindler vacuum $|0_{FR}\rangle$ corresponds to the empty state in a single wedge. In that case it is well-known that the $T_{uu}$ component of the energy momentum tensor (with the usual UV-divergence removed by subtracting the UV-divergent expectation value of $T_{uu}$ in the Minkowski vacuum $|0_M\rangle$) diverges as one approaches the horizon: $\langle T_{uu}\rangle_{FR} - \langle T_{uu}\rangle_{M} = -\frac{1}{48\pi}\frac{1}{u^2}$ in $1+1$ dimensions for $u>0$ (for a nice derivation of this result see \cite{Parentani:1993}).
A similar analysis can be done for the energy momentum tensor in the hyperbolic de Sitter patch, where the global de Sitter invariant vacuum state is now the Bunch-Davies vacuum $|\Omega_{BD}\rangle$. The obvious difference with the Rindler wedge is the absence of a timelike Killing vector. In addition, the $t=0$ surface is a (light-) cone, so a better analogy is with Milne space, to which the de Sitter hyperbolic section reduces for small $t$. In any case, we will use the same regularization procedure, restricting to the minimally coupled massless case $\nu = \frac{3}{2}$. The most convenient method to calculate components of the energy momentum tensor makes use of the Wightman function $G^{+}(x,x',t,t')$ and specifically we will look at the following contribution
\begin{equation}
\begin{aligned}
\langle (\partial_{\alpha}\phi)^2(x,t) \rangle &= \lim_{x',t'\rightarrow x,t}\partial_{\alpha}\partial_{\alpha}G(x,x',t,t').
\end{aligned}
\end{equation}
The Wightman function for the Bunch-Davies state is well known, but here we use the expression in terms of an integral over the hyperbolic momentum $p$ as given in \cite{Sasaki:1995}. This allows us to consistently regulate the UV-divergence of $\langle T_{\mu \nu}\rangle$ in the two states of interest. In appendix \ref{EMT} we show that the difference $\langle T_{tt}\rangle_{H}-\langle T_{tt}\rangle_{BD}$ is UV-finite and diverges as $t\rightarrow 0$
\begin{equation}
\begin{aligned}
\langle T_{tt}\rangle_{H}-\langle T_{tt}\rangle_{BD} &= -\frac{11}{240\pi^2}\frac{1}{t^4}+O\left(\frac{1}{t^2}\right).
\end{aligned}
\end{equation}
So we conclude that the hyperbolic vacuum $|\Omega_H\rangle$ has singular properties that are completely analogous to the Minkowski Fulling-Rindler, Milne and de Sitter static vacuum, see also \cite{Eme:2014}. The energy momentum tensor diverges in the limit $t\rightarrow 0$, so infinite energy seems to be required to prepare the state at the (singular) origin. A complete description all the way until $t=0$ is therefore obviously inconsistent, but strictly speaking that does not need to be fatal in a cosmological setting, in the sense that in a stand-alone open universe this might be interpreted as the Big Bang singularity.
Of course, arguably the most natural and well-behaved choice for an initial state on hyperbolic de Sitter sections is the de Sitter invariant Bunch--Davies state, to which we turn next.
\subsection{The Bunch--Davies state in the hyperbolic patch}
Here we will just briefly summarize the results of \cite{Sasaki:1995} and \cite{MalPim:2012}. More details can be found in those papers and in appendix \ref{Modes}. The most important observation is that mode functions of one of the hyperbolic patches (\ref{Hmodes}) do not correspond to regular mode functions on the full (Euclidean) de Sitter space. In \cite{Sasaki:1995} the hyperbolic mode functions are analytically continued to the other hyperbolic patch, allowing them to construct a set of regular mode functions that can cover all of de Sitter space as follows
\begin{equation}\label{continued}
\begin{aligned}
\chi_p^{(R)}&= \left\{ \begin{matrix}
P^{ip}_{\nu-\frac{1}{2}}(z) &&\text{for} \ z \in R \\
\frac{i \sin(\pi (\nu-\frac{1}{2}))}{\sinh(p\pi)} P^{ip}_{\nu-\frac{1}{2}}(z)
+\frac{i \sin(\pi(ip+\nu-\frac{1}{2}))e^{}}{\sinh(\pi p)}\frac{\Gamma[\nu+\frac{1}{2}+ip]}{\Gamma[\nu+\frac{1}{2}-ip]}P^{-ip}_{\nu-\frac{1}{2}}(z) &&\text{for} \ z \in L
\end{matrix}\right.
\end{aligned}
\end{equation}
These mode functions do not yet describe the Euclidean or Bunch--Davies vacuum, which can for instance be concluded by the fact that they are not (anti-)symmetric under the transformation $R \leftrightarrow L$. It turns out that the linear combinations $\chi_R\pm\chi_L$ correspond to the proper mode functions associated with the Euclidean or Bunch--Davies vacuum, as was proven by computing the Wightman function \cite{Sasaki:1995}. The (still to be normalized) mode functions are linear combinations of the associated Legendre functions
\begin{equation}\label{dSmodes}
\begin{aligned}
\chi_{p,\sigma} &=
\left\{
\begin{array}{lr}
\alpha^{\sigma}_{p,R} P^{ip}_{\nu-\frac{1}{2}}(z)+\beta^{\sigma}_{p,R}
P^{-ip}_{\nu-\frac{1}{2}}(z) &\text{for} \ x \in R \\
\alpha^{\sigma}_{p,L} P^{ip}_{\nu-\frac{1}{2}}(z)+\beta^{\sigma}_{p,L}
P^{-ip}_{\nu-\frac{1}{2}}(z) &\text{for} \ x \in L
\end{array}
\right.
\end{aligned}\end{equation}
where $\sigma = \pm1$\footnote{$\sigma = \pm 1$ is related to the combination $\chi_P^{(L)}\pm \chi_P^{(R)}$.} and $z = \cosh(t)$; the expressions for the $\alpha$'s and $\beta$'s are given in \eqref{alfabeta}. We stress that the associated Legendre functions $P$ in (\ref{dSmodes}) do not have to be analytically continued any further\footnote{They are constituents of $\chi_{L}$ and $\chi_{R}$, which are already regular everywhere. This is different from \cite{Kanno:2014}.}. The full field expansion, with creation and annihilation operators $\hat{a}_{\sigma plm}$ satisfying $[\hat{a}_{\sigma plm},\hat{a}_{\sigma' p'l'm'}^{\dagger}] = \delta_{\sigma \sigma'}\delta_{ll'} \delta_{mm'} \delta(p-p')$, is given by:
\begin{equation}\label{dSexpansion}
\begin{aligned}
\phi(t,r,\Omega) &= \int dp \sum_{\sigma=\pm1}\sum_{l,m}\frac{1}{N_{\chi^{p\sigma}}}\left(\hat{a}_{\sigma plm}\chi_{p,\sigma}(z)Y_{plm}(r,\Omega)+\text{h.c.}\right)
\end{aligned}
\end{equation}
where $N_{\chi^{p\sigma}}$ is the Klein--Gordon norm consistent with the commutation relations\footnote{Strictly speaking the expression (\ref{dSexpansion}) is incomplete, since we should also include the ``zero mode". For our purposes this will however not affect the results.}. In appendix \ref{KGnorm} we show that the normalization $N_{\chi^{p\sigma}}$ is given by
\begin{equation}\label{dSnorm}
\begin{aligned}
N_{\chi^{p\sigma}}^2 &\equiv \langle\chi_{p,\sigma}Y_{plm},\chi_{p, \sigma}Y_{ plm} \rangle_{\text{KG}} \\
&= \sum_{q = L,R}\left(\alpha^{\sigma}_{p,q}\bar{\alpha}^{\sigma}_{p,q}-\beta^{\sigma}_{p,q}\bar{\beta}^{\sigma}_{p,q}\right)N^2_{P^p}
\end{aligned}
\end{equation}
where $\bar{\alpha},\bar{\beta}$ denote the complex conjugates of $\alpha,\beta$. We conclude that the Bunch--Davies vacuum state, in a single hyperbolic patch, is defined as\footnote{As before we ignore the supercurvature modes.}
\begin{equation}
\begin{aligned}
\hat{a}_{\sigma plm}|\Omega_{BD}\rangle &= 0 &&\forall \, \sigma, p,l,m.
\end{aligned}
\end{equation}
Now let us describe the relation between the creation and annihilation operators of the modes (\ref{dSmodes}) and the creation and annihilation operators of hyperbolic modes (\ref{Hmodes}). Both field expansions (\ref{dSexpansion}) and (\ref{Hexpansion}) are linear combinations of associated Legendre functions. We can find the relation between the $\hat{a}_{\sigma plm}$ and the $\hat{b}_{qplm}$ ($q = L,R$) by comparing the coefficients
\begin{equation}\label{relationab}
\begin{aligned}
\hat{b}_{q plm} &= \sum_{\sigma=\pm 1}\frac{N_{P^p}}{N_{\chi_{p,\sigma}}}\left(\alpha^{\sigma}_{p,q}\hat{a}_{\sigma plm}+\bar{\beta}^{\sigma}_{p,q}\hat{a}^{\dagger}_{\sigma pl-m}\right).
\end{aligned}
\end{equation}
Given (\ref{dSnorm},\ref{relationab}), they enforce
\begin{equation}\label{commutation}
\begin{aligned}
\left.\begin{array}{l}
\left[ \hat{a}_{\sigma plm},\hat{a}^{\dagger}_{\sigma'p'l'm'} \right] = \delta(p-p')\delta_{\sigma \sigma'}\delta_{mm'}\delta_{ll'} \\ \\
\left[\hat{a}_{\sigma plm},\hat{a}_{\sigma'p'l'm'}\right] = 0
\end{array}\right\}
\Leftrightarrow \left\{
\begin{array}{l}
\left[\hat{b}_{qplm},\hat{b}^{\dagger}_{q'p'l'm'}\right] = \delta(p-p')\delta_{mm'}\delta_{ll'}\delta_{qq'} \\ \\
\left[\hat{b}_{qplm},\hat{b}_{q'p'l'm'}\right] = 0
\end{array}\right..
\end{aligned}
\end{equation}
For more details we refer to \cite{Sasaki:1995} and the appendices. The above relationship confirms that the normalizations (\ref{Pnorm}) and (\ref{dSnorm}) are consistent and in particular that the right normalization for the hyperbolic mode functions is given by (\ref{Pnorm}). This will be of importance when comparing the predictions for the power-- and bi--spectrum of the two different states under consideration: the pure hyperbolic vacuum and the Bunch--Davies state (as we will do in section \ref{sec:correlators}). The latter is a mixed state from the point of view of a single hyperbolic patch, due to the entanglement between the two hyperbolic patches in the Bunch--Davies vacuum.
Let us here remind the reader that we would like to compare the predictions for the expectation values of (scalar field) quantum fluctuations in the two different states that were introduced above. A priori different initial states give different predictions for the cosmic microwave background temperature anisotropies and the large scale structure distribution. We should stress that we are technically not considering an actual bubble nucleation event, where more intricate and model-dependent bubble wall physics could lead to additional effects \cite{Sasaki:1996, Gar:1998}, see also \cite{HawHerTur:2000}. Instead, we will work under the assumption that the two states that were introduced capture an essential difference that is generic: the entangled nature of the Bunch--Davies vacuum implies a mixed initial state, whereas the hyperbolic vacuum corresponds to a pure state on a single hyperbolic section. Different (presumably more realistic) states in open universes have been considered in the past \cite{Sasaki:1996, Komatsu:2013}, but these states appear to be pure (and excited) hyperbolic states, so they do not capture any effects related to the mixed nature of the initial state. In the process we hope to clear up some confusion that might have arisen and that could also have consequences for more realistic bubble states that were considered in the past.
We should add that one might anticipate the differences between the two states to only become visible at small hyperbolic momentum $p \lesssim 1$, i.e. scales comparable to the hyperbolic curvature. However, even small curvature suppressed changes in the initial state might be enhanced in the (nonlinear) bi--spectrum, as has been pointed out and analyzed in \cite{Holman:2007, Meerburg:2009} and for a certain generic type of mixed state in \cite{Agullo:2010}. This motivates our particular interest in computing the bi--spectrum, comparing the different states to the planar Bunch--Davies result. But first let us review some general facts regarding correlators in de Sitter space and summarize the results for the two--point functions.
\section{Correlators in hyperbolic de Sitter space}\label{sec:correlators}
Making use of the previously established relations between the de Sitter invariant Bunch--Davies vacuum and the hyperbolic vacuum state, we will compute both the Bunch--Davies and the hyperbolic vacuum power--spectra of scalar field quantum fluctuations\footnote{See \cite{Eme:2014} for related work on the response of Unruh detectors.}. The Bunch--Davies result can also be calculated using a reduced density matrix formalism in the hyperbolic patch. Let us from the outset emphasize that within our basic de Sitter set-up, even though the Bunch--Davies state is mixed from the hyperbolic patch perspective, all hyperbolic Bunch--Davies correlators should match the (hyperbolic coordinate transformed) planar Bunch--Davies correlators. As a consequence one can rule out large deviations of Bunch--Davies hyperbolic correlators at late time and large momenta (when the hyperbolic coordinates reduce to planar coordinates) as compared to the planar Bunch--Davies correlators. That leaves the (pure) hyperbolic vacuum as the potentially more interesting state to consider, as far as enhanced initial state effects with respect to the planar Bunch--Davies state are concerned.
Let us start by pointing out that the field operator $\phi_p$ evaluated on points in the left hyperbolic patch is trivial in the right hyperbolic patch
\begin{equation}
\begin{aligned}
\phi_p(x) &= \phi_{L,p}(x)\otimes \mathbf{I}_R &&\text{for} \ x \in \text{L}.
\end{aligned}
\end{equation}
As a consequence, de Sitter $n$-point functions of fields $\phi_p$ in the Bunch--Davies state, evaluated on points in the left hyperbolic patch, can be calculated either using the full global description or by using a reduced density matrix
$ \hat{\rho}_L = \text{Tr}_{\mathcal{H}_R}\left\{ |\Omega_{\text{BD}}\rangle\langle\Omega_{\Omega_{BD}}|\right\}$, and their results should agree. This is shown explicitly in appendix \ref{app:reduceddensitymatrixspectrum}.
By defining the $\hat{b}_{qplm}$ as in (\ref{relationab}), we can write the field operator for arbitrary values of $p,l,m$ as
\begin{equation}
\begin{aligned}
\phi_{plm}(x) &= \hat{b}_{Lplm}\frac{1}{N_{P^p}}P^{ip}_{\nu-\frac{1}{2},L}Y_{plm}+\text{h.c.} \\
&+\hat{b}_{Rplm}\frac{1}{N_{P^p}}P^{ip}_{\nu-\frac{1}{2},R}Y_{plm}+\text{h.c.},
\end{aligned}
\end{equation}
where
\begin{equation}
\begin{aligned}
P^{ip}_{\nu-\frac{1}{2},L} &= \left\{\begin{array}{ll}
P^{ip}_{\nu-\frac{1}{2}}(t) &\text{for} \ t \in L \\
0 &\text{for} \ t \in R \\
\end{array}\right.
\end{aligned}
\end{equation}
and vice versa for $P^{ip}_{\nu-\frac{1}{2},R}$. Although these functions are not mode functions on a full Cauchy slice covering the de Sitter space, we are allowed to express the field in terms of them. Note that the $\hat{b}_{L}$ and $\hat{b}_{R}$ operators mutually commute. To make explicit that the field operator decomposes in the left and right hyperbolic patches, we write
\begin{equation}\label{explicit}
\begin{aligned}
\phi_{plm}(x) &= \left(\hat{b}_{Lplm}\frac{1}{N_{P^p}}P^{ip}_{\nu-\frac{1}{2},L}Y_{plm}+\text{h.c.}\right)\otimes \mathbf{I}_R \\
&+\mathbf{I}_L\otimes \left(\hat{b}_{Rplm}\frac{1}{N_{P^p}}P^{ip}_{\nu-\frac{1}{2},R}Y_{plm}+\text{h.c.}\right).
\end{aligned}
\end{equation}
Note that the expansion of a scalar field in Minkowski spacetime in terms of left and right Rindler wedge modes is similar (see for instance \cite{Parentani:1993}).
The restriction of the operator (\ref{explicit}) to points in the left hyperbolic patch is by definition equal to the full operator evaluated on points in the left hyperbolic patch.
Note that if the field operator evaluated on points in the left hyperbolic patch would also have support on the right hyperbolic patch, it would not make sense to do a density matrix calculation as done above.
Clearly therefore Bunch--Davies scalar field correlators should be the same, independent of whether one uses hyperbolic or planar coordinates. Of course, since the de Sitter invariant length is expressed differently in terms of planar or hyperbolic coordinates, the functional dependence of the equal (hyperbolic) time correlators will look different. Since the difference between planar and hyperbolic coordinates vanishes in the late time and large momentum limit, the the planar and the hyperbolic Bunch--Davies correlators match in that limit and small modifications are suppressed in the hyperbolic curvature scale. We conclude that hyperbolic Bunch--Davies correlators can be computed either using a global de Sitter description (for which the Bunch--Davies state is a pure initial state) or by considering a single de Sitter hyperbolic patch (for which the Bunch--Davies state equals a mixed initial state).
After these important preliminaries let us now proceed by computing the power spectrum of a massless scalar field in a hyperbolic coordinate patch in the hyperbolic vacuum and Bunch--Davies initial state respectively, as a function of the hyperbolic momentum $p$.
\subsection{Power--spectrum results}
For most of the details we refer to the appendix \ref{app:powerspectramasslessfield}. Here we will just quote the main results. For the two point function in the hyperbolic vacuum state we find
\begin{equation}
\begin{aligned}
\langle \Omega_H|\phi_p\phi_p'|\Omega_H\rangle
&= \delta(p-p')\frac{H^2}{\sinh^2(t)}\frac{p}{4\pi^2}\frac{\cosh^2(t)+p^2}{p^2+1}
\end{aligned}
\end{equation}
where we have used the completeness relation of the eigenfunctions of the hyperbolic Laplacian\footnote{$\sum_{lm}\left|Y_{plm}\right|^2 = \frac{p^2}{2\pi^2}.$} and the commutation relations. At late times $t\rightarrow \infty$ this approaches
\begin{equation}
\begin{aligned}
\langle \Omega_H|\phi_p\phi_p'|\Omega_H\rangle &\rightarrow \frac{H^2p}{4\pi^2(p^2+1)}
\end{aligned}
\end{equation}
and the appropriately normalized power spectrum (at late times) equals
\begin{equation}\label{psH}
\begin{aligned}
\Delta^2_{\phi,H}(p) = \frac{H^2}{4\pi^2}\frac{p^2}{p^2+1}.
\end{aligned}
\end{equation}
The same (hyperbolic coordinate patch) two point function in the Bunch--Davies vacuum is instead found to be equal to
\begin{equation}\label{2ptdirect2}
\begin{aligned}
\langle \Omega_{BD}|\phi_p\phi_{p'}|\Omega_{BD}\rangle &= \delta(p-p')\frac{H^2}{\sinh^2(t)}\frac{p}{4\pi^2}\frac{\cosh^2(t)+p^2}{p^2+1}\coth(\pi p).
\end{aligned}
\end{equation}
This result can either be obtained from a direct calculation using the global Bunch--Davies vacuum construction and restricting to one of the hyperbolic coordinate patches \cite{Sasaki:1995}, or from a (mixed) density matrix calculation in a single hyperbolic coordinate patch, using the explicit expression for the density matrix as reported in \cite{MalPim:2012}, as we confirm in appendix \ref{app:reduceddensitymatrixspectrum}.
As alluded to earlier, the reason for this expression to not exactly reproduce the scale-invariant planar coordinate result for the scalar field power spectrum in the Bunch--Davies vacuum is that different coordinates are used. As the hyperbolic and planar coordinates are the same at late times and for small distances, the late-time power spectra at large momentum should be the same as well. At late times $t\rightarrow \infty$ we find
\begin{equation}
\begin{aligned}
\langle \Omega_{BD}|\phi_p\phi_{p'}|\Omega_{BD}\rangle &\rightarrow \frac{H^2p}{4\pi^2(p^2+1)}\coth(\pi p).
\end{aligned}
\end{equation}
Correspondingly, the power spectrum (at late times) is given by
\begin{equation}\label{psBD}
\begin{aligned}
\Delta^2_{\phi,BD}(p) =\frac{H^2}{4\pi^2}\frac{p^2}{p^2+1}\coth(\pi p).
\end{aligned}
\end{equation}
\begin{figure}[ht]
\centering
\begin{minipage}[t]{\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{powerspectra.pdf}\caption{The power spectra (logarithmic scale) for the hyperbolic vacuum (blue) and the Bunch--Davies vacuum (red), as function of the hyperbolic momentum $p$ with $H=1$. The dashed line indicates the scale--invariant planar Bunch--Davies result $\frac{H^2}{4 \pi^2}$. }\label{fig1}
\end{minipage}
\end{figure}
Looking at these power spectra we indeed find that for $p \gg 1$, when $\coth\pi p\approx 1$ and $p^2+1 \approx p^2$, the hyperbolic vacuum as well as the hyperbolic Bunch--Davies result matches the standard scale invariant planar Bunch--Davies result $\frac{H^2}{4\pi^2}$. They only start to differ from each other and the standard planar Bunch--Davies expression for sufficiently small momenta $p \lesssim 1$ (see figure (\ref{fig1})). Note that although the corresponding wavelengths are expected to lie far outside our observable window, given the fact that they correspond to length scales longer or comparable to the hyperbolic curvature scale, for both hyperbolic states the power is suppressed as compared to the standard planar Bunch--Davies result. As these departures from the standard planar result become evident, one should keep in mind that the difference found in the hyperbolic Bunch--Davies result can be attributed to a coordinate change, whereas (part of) the change in the hyperbolic vacuum power spectrum is related to an initial state modification. Although this difference might seem unimportant at this point, when considering the bi--spectrum in the next section it is a relevant distinction, since the bi--spectrum has been found to be particularly sensitive to changes in the initial state.
Finally let us also remark that the bubble state put forward in \cite{Sasaki:1996} and more recently used in \cite{Komatsu:2013} does resemble the Bunch--Davies state in the sense that the expectation value of the number operator on hyperbolic sections agrees (giving thermal occupation numbers in terms of co-moving momentum), but in one aspect is crucially different due to the fact that it has been constructed as a Bogoliubov transformation of the hyperbolic vacuum. As a consequence the state necessarily corresponds to a pure state (and inherits the singular properties of the hyperbolic vacuum) and it explains why their result for the power spectrum does not agree with our hyperbolic Bunch--Davies result that necessarily corresponds to a mixed state.
We conclude that, independent of the particular initial state under consideration, any power--spectrum signatures of an open inflationary universe are confined to the curvature scale, which has to be several orders of magnitude larger than the largest observable length scale in the universe. Although the initial hyperbolic state is mixed when assuming a (globally defined) planar Bunch--Davies state, the power--spectrum results in this admittedly basic set-up in which all bubble wall physics is ignored, do not show large deviations, as should be expected. In fact, this statement is true for general hyperbolic $n$-point correlators in the Bunch--Davies vacuum. Potentially enhanced bi--spectrum results due to initial state excitations, as compared to the standard planar Bunch--Davies result, can therefore only be expected assuming the pure hyperbolic vacuum as the initial state in the hyperbolic patch.
\subsection{The bi--spectrum in the hyperbolic vacuum}
Let us next consider the bi--spectrum of scalar density perturbations in the hyperbolic vacuum, which can be viewed (for any finite boost parameter $\gamma$) as some sort of excited initial state with respect to the standard planar Bunch--Davies vacuum. We will mainly be interested in the so--called squeezed limit, for which previous work uncovered enhanced results for excited (planar) Bunch--Davies intial states \cite{Holman:2007, Meerburg:2009, Agullo:2010, Holman:2012}. Moreover, as before it should be reasonable to work in the sub--curvature approximation $p \gg 1 $, for which the hyperbolic momenta $p_i$ are approximately equal to the standard (flat) wavenumbers $k_i$ up to curvature suppressed corrections.
To compute the hyperbolic bi--spectrum we need the action to third order in the scalar density perturbation $\varphi$ and the hyperbolic curvature introduces some new ingredients as compared to the planar calculation \cite{Maldacena:2002}, which have been carefully dealt with in \cite{Komatsu:2013}. The state of interest in \cite{Komatsu:2013} is however the bubble state constructed first in \cite{Sasaki:1996}, which is related to the hyperbolic vacuum by means of a Bogoliubov transformation and crucially differs from the Bunch--Davies vacuum. The bubble state can be viewed as a (well-motivated and first principles derived) initial state modification with respect to the hyperbolic vacuum, explaining their interest in trying to identify enhanced features in the bi--spectrum. Note that their computation is not applicable to initial state modifications of the (planar) Bunch--Davies vacuum. Such an interpretation would only be valid in the infinite boost limit, when the hyperbolic vacuum reduces to the planar Bunch--Davies vacuum. As we explained, in our pure de Sitter set-up it is the hyperbolic vacuum initial state that (for finite boost parameter $\gamma$) corresponds to an excited state as compared to the planar Bunch--Davies vacuum and could therefore potentially display interesting bi--spectrum enhancements. Instead the hyperbolic Bunch--Davies bi--spectrum equals the planar Bunch--Davies bi--spectrum and all apparent changes can be related to the coordinate change from planar to hyperbolic. So here we will be interested in computing the results for bi--spectrum in the hyperbolic vacuum, which can fortunately be extracted straightforwardly from the results in \cite{Komatsu:2013}.
So let us first briefly review the basic results reported in \cite{Komatsu:2013} and then apply them to our case of interest. We will be interested in a massless minimally coupled scalar field, for which the positive and negative frequency modes defined in a single hyperbolic patch, can be nicely expressed as
\begin{eqnarray}\label{eq:modefunctions}
u_p (\eta) & = & H \frac{\cosh \eta + ip \sinh \eta}{\sqrt{2p(1+p^2)}}e^{-ip\eta}, \nonumber \\
v_p (\eta) & = & H \frac{\cosh \eta - ip \sinh \eta}{\sqrt{2p(1+p^2)}}e^{ip\eta}.
\end{eqnarray}
The perturbed metric in one of the hyperbolic patches, is written in the ADM formalism as
\begin{eqnarray}
ds^2 = - N^2 dt^2 + h_{ij}\left( dx^i + N^i dt \right)\left( dx^j + N^j dt \right),
\end{eqnarray}
where, as usual, $N$ is the lapse function, $N^i$ is the shift and $h_{ij} = a^2 (t) e^{2 \zeta} \gamma _{ij} $ is the spatial metric, with curvature perturbation $\zeta$. We will be interested in gravitationally induced nonlinearities on the scalar density perturbation, requiring that we need to introduce (generic) slow-roll evolution of the background scalar $\phi(t)$ in order to couple the scalar inflaton field to the scalar density perturbation. Now, plugging the above metric into the action for the scalar degree of freedom (assuming slow-roll evolution) and solving the constraint equations order by order, one can obtain the quadratic and cubic (and higher) action. The quadratic action for scalar perturbations, in the flat gauge $\zeta = 0$, to leading order in the slow roll parameters is
\begin{eqnarray}
S^{(2)} = \int dt d^3x a(t)^3 \sqrt{\gamma} \left( \frac{1}{2} \dot{\varphi}^2 - \frac{1}{2 a(t)^2} \partial _i \varphi \partial ^i \varphi \right),
\end{eqnarray}
where $\partial _i$ is the covariant derivative with respect to $\gamma _{ij}$. Up to this point this should all be familiar, so let us now turn to the cubic action on a hyperbolic patch. As explained in \cite{Komatsu:2013} the dominant term in the third order Lagrangian is found to be
\begin{eqnarray}\label{eq:dominantterm}
\mathcal{L} ^{(3)} & = & - \sqrt{\gamma} a^{5} \dot{\phi} \left( \left( \partial ^{2} - 3 \right)^{-1}\dot{\varphi}_c \right) \dot{\varphi}_c^{2},
\end{eqnarray}
The new ingredient due to the hyperbolic curvature in this action is the $-3$ term. The $\varphi _c$ is (as usual) the redefined field $\left( \varphi \rightarrow \varphi _c \right)$, defined as
\begin{eqnarray}
\varphi = \varphi _c + \frac{\dot{\phi}}{4 (\dot{a}/a)} \left[ \left( \partial ^2 - 3 \right)^{-1} \partial _i\varphi _c \partial ^i \varphi _c \right],
\end{eqnarray}
which does not affect the quadratic action and has removed terms in the cubic action that are proportional to the equations of motion. The bi--spectra of $\varphi$ and $\varphi _c$ are then related as follows
\begin{eqnarray}
\left \langle \varphi (x_1) \varphi (x_2) \varphi (x_3) \right \rangle & = & \left \langle \varphi _c (x_1) \varphi _c (x_2) \varphi _c (x_3) \right \rangle \nonumber \\
& + & \frac{\dot{\phi}}{4(\dot{a}/a)} \left[ \left( \partial ^{2} - 3 \right)^{-1} \left \langle \partial _i \varphi _c(x_1) \varphi _c (x_2) \right \rangle \left \langle \partial ^i \varphi _c(x_1) \varphi _c (x_3) \right \rangle + \text{permutations} \right] \nonumber \\
\end{eqnarray}
In harmonic space, introducing the geometrical factor
\begin{equation}
\int d^3x \sqrt{\gamma} Y_{p_1 l_1 m_1}(x) Y_{p_2 l_2 m_2}(x) Y_{p_3 l_3 m_3}(x) \equiv {\cal F}^{l_1 l_2 l_3}_{p_1 p_2 p_3} {\cal G}^{m_1 m_2 m_3}_{l_1 l_2 l_3} \, ,
\end{equation}
the bi--spectrum $B(p_1, p_2, p_3)$ is defined as
\begin{equation}
\left \langle \varphi_{p_1 l_1 m_1} \varphi_{p_2 l_2 m_2} \varphi_{p_3 l_3 m_3} \right \rangle = B(p_1, p_2, p_3) \, {\cal F}^{l_1 l_2 l_3}_{p_1 p_2 p_3} {\cal G}^{m_1 m_2 m_3}_{l_1 l_2 l_3} \, .
\end{equation}
Of the two bi--spectrum contributions due to the field redefinition, the first term is then computed using eq.~(\ref{eq:dominantterm}) in the in--in formalism, where we already assumed the sub-curvature limit allowing one to replace the momentum $p$ with the planar momentum $k$
\begin{eqnarray}\label{eq:bispectrum}
B(k_1,k_2,k_3) & = & 2 \text{Re} \left[ i v_{k_1}(0) v_{k_2}(0) v_{k_3}(0) \left( \int_{-\infty}^{0} d \eta \frac{2 a^6 \dot{\phi}}{k_1 ^2 + 4} \dot{u}_{k_1}(\eta) \dot{u}_{k_2}(\eta) \dot{u}_{k_3}(\eta) \right) + 1 \leftrightarrow 2 + 1 \leftrightarrow 3 \right]. \nonumber \\
\end{eqnarray}
In a hyperbolic (quasi) de Sitter space the integral over conformal time is naturally divided into two eras: from $-\infty$ to $-1$, where the curvature term is dominant and from $-1$ to $0$ corresponding to standard inflationary expansion. Assuming slow-roll evolution in both eras, the expressions for $\dot{\phi}$ in terms of the slow-roll parameter $\epsilon$ change when transitioning from the curvature-dominated era into the inflationary era.
The mode functions defining the hyperbolic vacuum $| \Omega_H \rangle$ in the sub--curvature approximation $p \gg 1 $, in the different eras, read as follows
\begin{eqnarray}
u_k(\eta) &\simeq& -\frac{i}{a(\eta)\sqrt{2k}} e^{-ik\eta} \quad (-\infty < \eta \lesssim -1) \, \nonumber \\
u_k(\eta) &\simeq& -\frac{H}{\sqrt{2k^3}} (1+ik\eta) e^{-ik\eta} \quad (-1 \lesssim \eta < 0) \, \nonumber \\
\dot{u}_k (\eta) &=& \frac{\partial u_k}{\partial \eta} \frac{1}{a(\eta)} \simeq \frac{1}{a^{2}(\eta)} \sqrt{\frac{k}{2}} e^{-ik \eta}, \nonumber \\
v_{k}(0) & \simeq & \frac{H}{\sqrt{2k^3}}.
\end{eqnarray}
Using these expressions to do the integral in (\ref{eq:bispectrum})and ignoring effects due to the discontinuity in the transition between eras, one obtains \cite{Komatsu:2013}
\begin{eqnarray}
B(k_1, k_2, k_3) & = & \frac{2 \sqrt{2\epsilon}H^4}{4 k_1 k_2 k_3\left( k_1 + k_2 + k_3 \right)} \left( \frac{1}{ k_1 ^2} + \frac{1}{k_2 ^2} + \frac{1}{ k_3 ^2} \right).
\label{hyperbolic-bispectrum}
\end{eqnarray}
This result agrees with the standard planar Bunch--Davies result in single field inflation. For completion the additional contribution from the field redefinition equals
\begin{eqnarray}
B^{\text{redef}} (k_1 , k_2 , k_3) & = & \frac{\dot{\phi}}{4H} \frac{\mathbf{k_1} \cdot \mathbf{k_2}}{k_3 ^2 + 4} \frac{H ^2}{2 k_1 ^3} \frac{H ^2}{2 k_2 ^3} + \text{permutations} \, ,
\end{eqnarray}
but as shown in \cite{Komatsu:2013} the contribution from this term is subdominant in the squeezed limit $k_3 \ll k_1 \simeq k_2 = k$ and therefore can be ignored when looking for other enhanced contributions in the squeezed limit, as expected for initial state modifications. The enhancement of initial state modifications in the squeezed limit, as compared to the hyperbolic vacuum, was indeed observed after adding the specific negative frequency term to the mode functions that describes the bubble initial state \cite{Komatsu:2013}. However the Bogoliubov coefficient describing this excited bubble state is exponentially suppressed in the hyperbolic momentum over the curvature scale (which is normalized to one) $e^{-\pi p}$, meaning that these effects for the observable modes in the sky are exponentially suppressed and undetectable.
After having reviewed the hyperbolic bi--spectrum calculation in \cite{Komatsu:2013}, we will now connect these results to our case of interest. To start out with, the quoted result for the bi--spectrum (\ref{hyperbolic-bispectrum}) is of course the hyperbolic vacuum bi--spectrum in the late time, sub-curvature limit, which agrees with the standard planar Bunch--Davies result. Clearly then, no (modified initial state) enhancement with respect to the planar Bunch--Davies result is found in the late time and large momentum limit. The squeezed enhancement that was revealed in \cite{Komatsu:2013} for the excited bubble state, albeit exponentially suppressed in momentum, is with respect to the hyperbolic vacuum and can only be interpreted as enhancement with respect to the planar Bunch--Davies vacuum in the late time and large momentum limit. As we noted, the late time and large momentum limit effectively corresponds to the action of the infinite boost in the embedding space for which the hyperbolic vacuum indeed matches the planar Bunch--Davies vacuum state, explaining the result. The excited bubble vacuum, obtained through a Bogoliubov rotation from the hyperbolic vacuum and therefore corresponding to a pure state, will also not agree with the hyperbolic Bunch--Davies result. Instead the hyperbolic bi--spectrum in the Bunch--Davies state has to match the planar Bunch--Davies result (written in hyperbolic coordinates), obviously excluding a (squeezed) enhancement.
Our first conclusion is therefore that the bi--spectrum in the hyperbolic vacuum is clearly (and unsurprisingly) not enhanced with respect to the planar Bunch--Davies result in the late time and sub-curvature limit. In fact it is straightforward to (perturbatively) extend this conclusion beyond the strict sub-curvature or large momentum limit by analyzing the leading correction. The first correction in the large momentum approximation is obtained from an asymptotic expansion of the hyperbolic mode function at late times
\begin{equation}
u_p (\eta \ll 1, p \gg 1) \approx \frac{H}{\sqrt{2p^3}} \, (1 + ip \eta) \, e^{-ip\eta} \, \left( 1-\frac{1}{2p^2} + {\cal O}(1/p^4) \right) \, . \\
\label{pexpansion}
\end{equation}
Noting the last term (in brackets) one observes that the leading large momentum correction will give additional contributions suppressed in the large momentum limit with at least one factor of $1/p^2$\footnote{Re-installing the hyperbolic curvature scale, which was set to one, this term would be explicitly dimensionless and read $k_c/p^2$.}, but leaving the relative momentum dependence in tact. As a consequence enhancements in the squeezed momentum limit as compared to the planar Bunch--Davies result are excluded, as that would typically require additional terms featuring negative frequencies. It should be clear that corrections suppressed as $1/p^2$, without additional enhancements in particular momentum configurations, does not constitute an interesting non-Gaussian signature of an open inflationary universe.
We conclude that, unfortunately, the bi--spectrum in the hyperbolic vacuum does not produce interesting enhancements that could be searched for. This conclusion is in fact corroborated by effectively constructing the hyperbolic vacuum as an excited state on top of the (global) planar Bunch--Davies vacuum (see appendix \ref{app:Hvacuum}). Effectively one then discovers that the relevant Bogoliubov coefficients are suppressed exponentially in momentum, ensuring that these effects will not be observable. As emphasized, the same is true for the bi--spectrum in the hyperbolic Bunch--Davies initial state. So even though the hyperbolic vacuum can be thought of as an excited state with respect to standard planar Bunch--Davies vacuum, it is not of a type that gives rise to large (enhanced) corrections in the bi--spectrum as compared to the standard planar Bunch--Davies result.
\section{Conclusions}
Before summarizing our results, let us remind the reader once more that our motivation was to carefully study the relation between the hyperbolic and planar coordinate patches and their corresponding states in (mostly) pure de Sitter space. We believe these results to be of interest, and partially applicable, in the context of de Sitter false vacuum decay, but it is also clear that in that case a more complete analysis should include (model-dependent) wall physics that will affect the details. Instead we concentrated on a general and qualitative difference between two examples of initial states on a hyperbolic section of de Sitter space: the pure hyperbolic vacuum and the (entangled) de Sitter invariant Bunch--Davies state, corresponding to a mixed state. We first of all noted that the pure hyperbolic vacuum is formally inconsistent, due to the energy momentum tensor becoming singular at the null boundary of the hyperbolic section. This issue should plague all pure hyperbolic states, in particular also the bubble state introduced in \cite{Sasaki:1996}, which is obtained from the hyperbolic vacuum by a (unitary) Bogoliubov transformation.
The Bunch--Davies vacuum is qualitatively different, giving rise to a mixed state on a single hyperbolic section. The fact that the Bunch--Davies state is mixed on a hyperbolic section could a priori imply interesting non-Gaussian signatures \cite{Kanno:2014} as compared to the bi--spectrum in the hyperbolic vacuum. But as we pointed out all planar $n$-point correlators in the Bunch--Davies state, with all points in the hyperbolic section, should equal the hyperbolic $n$-point Bunch--Davies correlators. Any differences are generated by the coordinate change from planar to hyperbolic, instead of arising due to changes in the initial state. So even though the Bunch--Davies state is a mixed state on a hyperbolic section, certainly implying different results as compared to the hyperbolic vacuum, it will reproduce all planar Bunch--Davies correlators, up to coordinate changes. Since the hyperbolic and planar coordinates coincide in the late time and large momentum (sub-curvature) limit all effects due to the coordinate change should disappear and the leading corrections will not be enhanced.
We stressed that instead the hyperbolic vacuum should generically be considered as an excited state as compared to the Bunch--Davies vacuum. As a consequence that should be the state of interest for computing the bi--spectrum and look for enhanced features. Using the hyperbolic coordinate embedding we explicitly constructed a family of hyperbolic solutions that reduces to the planar coordinates in the infinite boost limit, as such providing a limiting relation between the hyperbolic vacuum and the planar Bunch--Davies vacuum. As a corollary we also argued that the hyperbolic vacuum can be mapped to a specific static vacuum, implying that the static vacuum should also reduce to (a sector of) the Bunch--Davies state in the infinite boost limit, as was first noted in \cite{GreParScha:2006}. Again, this limiting behavior implies that in the late time and large momentum limit, the bi--spectrum results for the hyperbolic vacuum should agree with the standard planar Bunch--Davies result and using the results of \cite{Komatsu:2013} this was indeed confirmed. Looking at the leading correction in the large momentum expansion, we verified that no enhancement in particular momentum configurations is generated and that the corrections are at least suppressed as $1/p^2$. These type of curvature suppressed Non-Gaussian corrections will clearly be impossible to detect. So unfortunately, on the basis of our analysis here, we should conclude that no detectable signals of an open inflationary universe in the fluctuation statistics is expected on small sub-curvature scales.
To summarize the two hyperbolic states introduced, one of them mixed (Bunch--Davies) and the other one pure (hyperbolic vacuum), make almost identical predictions in the late time sub-curvature limit. In fact, in the infinite boost limit the states become formally identical to the planar Bunch--Davies vacuum. For the mixed Bunch--Davies state this seems to imply that the density matrix $\rho_{BD}$ should depend on the boost parameter $\gamma$. Correspondingly, the associated von Neumann entropy $\text{Tr} (-\rho_{\text{BD}} \, \ln \rho_{\text{BD}})$ of the mixed Bunch--Davies state on the hyperbolic section should depend on the boost parameter $\gamma$ to ensure that the entropy vanishes in the infinite boost limit. This density matrix was computed in \cite{MalPim:2012} and it would be of interest to consider the generalization for non-zero boost parameter $\gamma$. Although one might think the density matrix and corresponding entropy to be boost invariant, this is not entirely obvious and the above observation does indeed suggest it might not be, perhaps in some subtle (singular) way. The dependence on the boost parameter should be such that it is invariant under $\gamma \rightarrow -\gamma$, effectively interchanging the two hyperbolic sections. Since the boost dependence can be implemented through a simple rescaling on the left hyperbolic momenta (and a time shift) \ref{fixedcoord} and the inverse rescaling on the right hyperbolic momenta (and time shift) it should be possible to trace the boost dependence of the Bunch--Davies state in terms of the left and right hyperbolic modes. It should then be straightforward to construct the corresponding density matrix and confirm that the density matrix and corresponding entropy become trivial in the infinite boost limit.
In more realistic scenarios, trying to incorporate the bubble nucleation dynamics, an initial state has been proposed that can be constructed by performing a unitary Bogoliubov transformation on the hyperbolic vacuum state \cite{Sasaki:1996}. As a consequence this state appears to be pure, which at first sight contradicts the general expectation that the initial state inside a bubble, after false vacuum decay, should be mixed. Although the pure bubble state is in certain aspects similar to the Bunch--Davies vacuum, the correlation functions will be different as compared to the Bunch--Davies hyperbolic correlation functions. In addition, as in the hyperbolic vacuum, the expectation value of the stress tensor in this pure hyperbolic bubble state should become singular as one approaches the bubble wall. It might be of interest to revisit the original construction and see how it can be adjusted to construct a mixed hyperbolic initial state instead. We hope to address this and some of the other remaining questions in future work.
\acknowledgments
We thank I-Sheng Yang and Ben Freivogel for useful discussions. This work is part of the Delta ITP consortium, a program of the Netherlands Organisation for Scientific Research (NWO) that is funded by the Dutch Ministry of Education, Culture and Science (OCW). This work is also supported in part by the Foundation for Fundamental Research (FOM), which is part of NWO.
\begin{appendix}
\section{Mode functions}\label{Modes}
\subsection{Solutions to the hyperbolic equation of motion}
The metric for both the left and right hyperbolic patch is given by:
\begin{equation}
\begin{aligned}
ds^2 &= \frac{1}{H^2}\left(-dt^2+\sinh^2t\left(dr^2+\sinh^2r \ d\Omega_2^2\right)\right)
\end{aligned}
\end{equation}
where the coordinates $t,r,\phi,\theta$ are dimensionless and $c=1$. The action for a massive non-interacting minimally coupled scalar field $\phi$ is given by:
\begin{equation}
\begin{aligned}
S &= -\frac{1}{2}\int \sqrt{-g} \ d^4x \left(g^{\mu\nu}\partial_{\mu}\phi\partial_{\nu}\phi+m^2\phi^2\right).
\end{aligned}
\end{equation}
The action of a conformally coupled scalar field can be written in the Einstein frame with effective mass $m^2 = 2H^2$. The equation of motion is given by
\begin{equation}
\begin{aligned}
0 &= \left({\frac{1}{\sinh^3(t)}}\frac{\partial}{\partial t}\sinh^3(t)
\frac{\partial}{\partial t}
-\frac{1}{\sinh^2(t)} \nabla_{\mathcal{H}^3}^2+\frac{m^2}{H^2}\right)\phi \\
&= \left({\frac{1}{\sinh^3(t)}}\frac{\partial}{\partial t}\sinh^3(t)
\frac{\partial}{\partial t}
-\frac{1}{\sinh^2(t)} \nabla_{\mathcal{H}^3}^2+\frac{9}{4}-\nu^2\right)\phi
\end{aligned}
\end{equation}
where $\nu$ is defined as $\nu=\sqrt{\frac{9}{4}-\frac{m^2}{H^2}}$ and $\nabla_{\mathcal{H}^3}^2$ is the Laplacian on the hyperboloid $\mathcal{H}^3$. We will use $\nu' = \nu -\frac{1}{2}$, consistent with \cite{Sasaki:1995}, such that $\nu' = 1$ corresponds to the massless minimally coupled case and $\nu' = 0$ corresponds to the massless conformally coupled case for which the effective mass is $m^2 = 2H^2$.
The eigenfunctions $Y_{plm}$ of the Laplacian $\nabla_{\mathcal{H}^3}^2$ on the hyperboloid $\mathcal{H}^3$, that are regular in $r=0$, are given by \cite{Sasaki:1995}:
\begin{eqnarray}
-\nabla_{\mathcal{H}^3}^2Y_{plm} &=& (p^2+1)Y_{plm} \nonumber\\
Y_{plm}(r,\Omega)&=&f_{pl}(r)Y_{lm}(\Omega)
\nonumber\\
f_{pl}(r)&=&
\frac{\Gamma(ip+l+1)}{\Gamma(ip+1)}\frac{p}{\sqrt{\sinh r}}
P^{-l-1/2}_{ip-1/2}(\cosh r)
\nonumber\\
&=&
(-1)^l\sqrt{\frac{2}{\pi}}\,\frac{\Gamma(-ip+1)}{\Gamma(-ip+l+1)}\,
\sinh^lr \frac{d^l}{d(\cosh r)^l}
\left(\frac{\sin pr}{\sinh r}\right)
\label{fpl}
\end{eqnarray}
where $Y_{lm}(\Omega)$ is the normalized
spherical harmonic function on the unit two-sphere,
$\Gamma(z)$ is the Gamma function and $P^{\nu}_{\mu}(z)$ is
the associated Legendre function of the first kind \cite{Magnus}.
The mode functions that correspond to the natural hyperbolic vacuum are given by
\begin{equation}
\begin{aligned}
\left\{\begin{array}{ll}
\frac{H}{\sinh t}P^{ip}_{\nu'}\left(\cosh t\right) &\text{positive energy modes} \\
\frac{H}{\sinh t}P^{-ip}_{\nu'}\left(\cosh t\right)
&\text{negative energy modes}
\end{array}
: p\geq 0\right\}.
\end{aligned}
\end{equation}
Mode functions on a Cauchy slice of de Sitter, must be regular and consist of linear combinations of the hyperbolic mode functions in the left and right hyperbolic patches \cite{Sasaki:1995}. The mode functions that correspond to the Bunch--Davies state are given in \cite{Sasaki:1995}:
\begin{equation}\label{tanakaexpans}
\begin{aligned}
\chi_{p,\sigma} &=
\left\{
\begin{array}{ll}
\displaystyle
\left(
{e^{\pi p}-\sigma e^{-i\pi\nu'}\over \Gamma(\nu'+ip +1)}
P^{ip}_{\nu'}(z)-
{e^{-\pi p}-\sigma e^{-i\pi\nu'}\over \Gamma(\nu'-ip +1)}
P^{-ip}_{\nu'}(z)\right) &\text{for} \ x \in R
\\
\displaystyle
\left(
{\sigma e^{\pi p}- e^{-i\pi\nu'}\over \Gamma(\nu'+ip +1)}
P^{ip}_{\nu'}(z)-
{\sigma e^{-\pi p}- e^{-i\pi\nu'}\over \Gamma(\nu'-ip +1)}
P^{-ip}_{\nu'}(z)\right) &\text{for} \ x \in L
\end{array}
\right. \\
&= \left\{
\begin{array}{lr}
\alpha^{\sigma}_{p,R} P^{ip}_{\nu'}(z)+\beta^{\sigma}_{p,R}
P^{-ip}_{\nu'}(z) &\text{for} \ x \in R \\
\alpha^{\sigma}_{p,L} P^{ip}_{\nu'}(z)+\beta^{\sigma}_{p,L}
P^{-ip}_{\nu'}(z) &\text{for} \ x \in L
\end{array}
\right.
\end{aligned}
\end{equation}
where $z = \cosh t$ and the constants $\alpha^{\sigma}_{p,q}$ and $\beta^{\sigma}_{p,q}$ are defined as:
\begin{equation}\label{alfabeta}
\begin{aligned}
\begin{array}{ll}
\alpha^{\sigma}_{p,L} = \sigma {e^{\pi p}-\sigma e^{-i\pi\nu'}\over \Gamma(\nu'+ip +1)} & \alpha^{\sigma}_{p,R} = {e^{\pi p}-\sigma e^{-i\pi\nu'}\over \Gamma(\nu'+ip +1)} \\
\beta^{\sigma}_{p,L} = -\sigma
{e^{-\pi p}-\sigma e^{-i\pi\nu'}\over \Gamma(\nu'-ip +1)} & \beta^{\sigma}_{p,R} = -
{e^{-\pi p}-\sigma e^{-i\pi\nu'}\over \Gamma(\nu'-ip +1)}
\end{array}.
\end{aligned}
\end{equation}
These mode functions must be normalized through the Klein--Gordon normalization (see section \ref{KGnorm}).
\subsection{Klein--Gordon normalization}\label{KGnorm}
\subsubsection*{Hyperbolic modes}
We normalize the hyperbolic modes on the hyperbolic patch using the variable $z = \cosh t$ and using the orthonormality of the $Y_{plm}$:
\begin{equation}\label{NP}
\begin{aligned}
N_{P^p}^2 &\equiv \langle\phi_{plm},\phi_{plm} \rangle_{\text{KG}} \\
&= i \int_{\Sigma}d\Sigma^{\mu}\left(\phi_{plm}\partial_{\mu} \phi_{plm}^*-\phi_{plm}^*\partial_{\mu} \phi_{plm}\right) \\
&= i \sinh^3 t \left(\frac{P^{ip}_{\nu'}(\cosh t)}{\sinh t}\partial_t\left( \frac{P^{-ip}_{\nu'}(\cosh t)}{\sinh t}\right)-\frac{P^{-ip}_{\nu'}(\cosh t)}{\sinh t}\partial_t \left(\frac{P^{ip}_{\nu'}(\cosh t)}{\sinh t }\right)\right) \\
&= i (z^2-1)\left(P^{ip}_{\nu'}(z)\partial_z P^{-ip}_{\nu'}(z)-P^{-ip}_{\nu'}(z)\partial_z P^{ip}_{\nu'}(z)\right).
\end{aligned}
\end{equation}
For the minimally coupled massless case $\nu' = 1$ we have:
\begin{equation}\label{pnorm}
\begin{aligned}
N^2_{P^p} &= i (z^2-1) P^{ip}_1(z)P^{-ip}_1(z)\times \left(\frac{1}{z-ip}+\frac{ip}{2}\frac{1}{1+z}+\frac{ip}{2}\frac{1}{1-z}- \frac{1}{z+ip}+\frac{ip}{2}\frac{1}{1+z}+\frac{ip}{2}\frac{1}{1-z} \right) \\
&= \frac{2p}{\left|\Gamma[1+ip]\right|^2} \\
&= \frac{2\sinh(\pi p)}{\pi}.
\end{aligned}
\end{equation}
In fact, for $\nu' \neq 1$ this normalization is also valid. In \cite{Sasaki:1995} it is shown that one can expand the mode functions in the $t\rightarrow 0$ regime:
\begin{equation}
\begin{aligned}
\frac{1}{\sinh t}P^{ip}_{\nu'}(\cosh t) &\approx \frac{2^{ip}}{\Gamma[1-ip]}t^{ip-1}.
\end{aligned}
\end{equation}
Using this expansion in (\ref{NP}) also results into the normalization (\ref{pnorm}). This normalization is valid for any $t$ by the properties of the Klein--Gordon normalization.
\subsubsection*{Bunch--Davies modes}
The Bunch--Davies modes are given in terms of linear combinations of the hyperbolic modes in (\ref{tanakaexpans}). Schematically we have (using the orthogonality of the hyperbolic mode functions):
\begin{equation}\label{Nchi}
\begin{aligned}
N^2_{\chi^{\sigma,p}} &= \langle \chi_{\sigma, p} ,\chi_{\sigma, p} \rangle \\
&= \sum_{q=L,R} \sum_{q'=L,R} \langle\left(\alpha^{\sigma}_{p,q}P^{p,q}+\beta^{\sigma}_{p,q}\bar{P}^{p,q}\right),\left(\alpha^{\sigma}_{p,q'}P^{p,q'}+\beta^{\sigma}_{p,q'}\bar{P}^{p,q'}\right)\rangle_{KG} \\
&= \sum_{q,q'=L,R}\left(\begin{array}{l} \alpha^{\sigma}_{p,q}\bar{\alpha}^{\sigma}_{p,q'}\langle P^{p,q},P^{p,q'}\rangle_{KG} + \alpha^{\sigma}_{p,q}\bar{\beta}^{\sigma}_{p,q'}\langle P^{p,q},\bar{P}^{p,q'}\rangle_{KG} \\ +\bar{\alpha}^{\sigma}_{p,q'}\beta^{\sigma}_{p,q}\langle \bar{P}^{p,q},P^{p,q'}\rangle_{KG}+\beta^{\sigma}_{p,q}\bar{\beta}^{\sigma}_{p,q'}\langle \bar{P}^{p,q},\bar{P}^{p,q'}\rangle_{KG}
\end{array}\right) \\
&= N^2_{P^p}\sum_{q=L,R}\left(\alpha^{\sigma}_{p,q}\bar{\alpha}^{\sigma}_{p,q}-\beta^{\sigma}_{p,q}\bar{\beta}^{\sigma}_{p,q}\right),
\end{aligned}
\end{equation}
where we used:
\begin{equation}
\begin{aligned}
\langle P^{p,q},P^{p,q'}\rangle_{KG} &= -\langle \bar{P}^{p,q},\bar{P}^{p,q'}\rangle_{KG} &&= \delta_{qq'}N^2_{P^p}, \\
\langle P^{p,q},\bar{P}^{p,q'}\rangle_{KG} &= \langle \bar{P}^{p,q},P^{p,q'}\rangle_{KG} &&= 0.
\end{aligned}
\end{equation}
Using (\ref{alfabeta}) we find:
\begin{equation}\label{aabb}
\begin{aligned}
\sum_{q=L,R}\left(\alpha^{\sigma}_{p,q}\bar{\alpha}^{\sigma}_{p,q}-\beta^{\sigma}_{p,q}\bar{\beta}^{\sigma}_{p,q}\right) &= \frac{8\sinh \pi p \left(\cosh \pi p -\sigma \cos \pi\nu'\right)}{\left|\Gamma[\nu'+ip+1]\right|^2} .
\end{aligned}
\end{equation}
So finally we can substitute (\ref{aabb}) into (\ref{Nchi}):
\begin{equation}
\begin{aligned}
N^2_{\chi^{\sigma,p}} &= N^2_{P^p} \sum_{q=L,R}\left(\alpha^{\sigma}_{p,q}\bar{\alpha}^{\sigma}_{p,q}-\beta^{\sigma}_{p,q}\bar{\beta}^{\sigma}_{p,q}\right) \\
&= \frac{2\sinh \pi p}{\pi}\times \frac{8\sinh \pi p \left(\cosh \pi p -\sigma \cos \pi\nu'\right)}{\left|\Gamma[\nu'+ip+1]\right|^2} \\
&= \frac{16 \sinh^2 \pi p \left(\cosh \pi p -\sigma \cos \pi\nu'\right)}{\pi \left|\Gamma[\nu'+ip+1]\right|^2}.
\end{aligned}
\end{equation}
This is consistent with \cite{Sasaki:1995}, but note that we included an extra factor of $2\sinh \pi p$ into the normalization, in order to simplify the expressions (\ref{alfabeta}). The normalized mode functions are the same as in \cite{Sasaki:1995}, of course.
\subsection{Mode functions for the massless scalar field}\label{app:modefunctionscalar}
Since we are mostly concerned with the massless minimally coupled scalar field ($\nu'=1$), we state the normalized mode functions for that case explicitly in hyperbolic time coordinate $t$ and in conformal time $\eta = \ln \tanh \frac{t}{2}$:
\begin{equation}\label{masslessmodes}
\begin{aligned}
\frac{1}{N_{P^p}}\frac{H}{\sinh t }P^{ip}_{1}(\cosh t) &= \frac{H}{\sqrt{2p(p^2+1)}}\left(\coth\frac{t}{2}\right)^{\frac{ip}{2}}\left(p \ \text{csch} \ t+ i\coth t\right)\\
&= \frac{H}{\sqrt{2p(p^2+1)}}e^{-ip\eta}\left(p\sinh \eta-i\cosh \eta \right),
\end{aligned}
\end{equation}
where we have chosen a convenient phase factor in the normalization, that does not affect the physics. The conformal time $\eta$ is defined as:
\begin{equation}
\begin{aligned}
ds^2 &= \frac{1}{H^2}\left(-dt^2+\sinh^2t\left(dr^2+\sinh^2r d\Omega^2_2\right)\right) \\
&= \frac{\sinh^2(t(\eta))}{H^2}\left(-d\eta^2+dr^2+\sinh^2r d\Omega^2_2\right) \\
\Rightarrow \eta&= \int \frac{dt}{\sinh t} \\
&= \ln \tanh \frac{t}{2}.
\end{aligned}
\end{equation}
Other useful relations between ``hyperbolic" time $t$ and ``conformal" time $\eta$ are:
\begin{equation}
\begin{aligned}
\sinh t &= -\frac{1}{\sinh \eta},
&&\cosh t = -\coth \eta.
\end{aligned}
\end{equation}
At early times $\eta \rightarrow -\infty$ the mode function for the massless minimally coupled scalar field (\ref{masslessmodes}) behaves like a positive energy mode function:
\begin{equation}
\begin{aligned}
\frac{H}{\sqrt{2p(p^2+1)}}e^{-ip\eta}\left(p\sinh \eta-i\cosh \eta \right) &= \sinh\eta \left(\frac{H(p+i)}{\sqrt{2p(p^2+1)}}e^{-ip\eta}+O(e^{2\eta})\right) \\
&\approx \sinh\eta \frac{H(p+i)}{\sqrt{2p(p^2+1)}}e^{-ip\eta}.
\end{aligned}
\end{equation}
The infinite boost limit $\gamma\rightarrow \infty$ (\ref{fixedcoord}) corresponds to the large $t$ (or small $\eta$) and large momentum limit. In particular in terms of conformal time, the rescaling for small $\eta$ reads $\eta \rightarrow \eta \, e^{-\gamma}$, implying that the combination $p \, \eta$ is invariant in the limit. This gives
\begin{equation}
\begin{aligned}
e^{-ip\eta}\left(p\sinh \eta-i\cosh \eta \right) &= p \sinh\eta \, e^{-ip\eta}\left(1-\frac{i}{p \eta}+O(\eta)\right) \\
&\approx e^{-ip\eta}\left(p\eta -i\right).
\end{aligned}
\end{equation}
This is exactly the mode function for a massless scalar field in the flat de Sitter slicing (up to the appropriate normalization).
\section{Power spectra for the massless field}\label{app:powerspectramasslessfield}
\subsection{Direct calculation}
\textbf{Power spectrum in hyperbolic vacuum}\\
The power spectrum in the hyperbolic vacuum can be computed in a straightforward way
\begin{equation}
\begin{aligned}
\langle \Omega_{H}|\phi_p\phi_p'|\Omega_H\rangle &= \frac{H^2}{\sinh^2(t)}\sum_{lm}\sum_{l'm'}\frac{Y_{plm}Y^*_{p'l'm'}}{N_{P^p}N_{P^{p'}}} \\
&\times \langle \Omega_{H}|\left(\hat{b}_{plm}P^{ip}_{\nu'}+\hat{b}^{\dagger}_{plm}P^{-ip}_{\nu'}\right)\left(\hat{b}_{p'l'm'}P^{ip'}_{\nu'}+\hat{b}^{\dagger}_{p'l'm'}P^{-ip'}_{\nu'}\right)|\Omega_H\rangle \\
&= \frac{H^2}{\sinh^2(t)}\sum_{lm}\sum_{l'm'}Y_{plm}Y^*_{p'l'm'}\frac{P^{ip}_{\nu'}P^{-ip'}_{\nu'}}{N_{P^p}N_{P^{p'}}} \langle \Omega_{H}|\hat{b}_{plm}\hat{b}^{\dagger}_{p'l'm'}|\Omega_H\rangle \ \text{using} \ \ \hat{b}_{plm}|\Omega_H\rangle = 0 \\
&= \delta(p-p')\frac{H^2}{\sinh^2(t)}\sum_{lm} \left|Y_{plm}\right|^2\frac{\left|P^{ip}_{\nu'}\right|^2}{N^2_{P^p}} \ \text{using} \left[\hat{b}_{plm},\hat{b}^{\dagger}_{p'l'm'}\right] = \delta(p-p')\delta_{ll'}\delta_{mm'} \\
&= \delta(p-p')\frac{H^2}{\sinh^2(t)}\frac{p^2}{2\pi^2}\frac{\left|P^{ip}_{\nu'}\right|^2}{N^2_{P^p}},
\end{aligned}
\end{equation}
where we used the completeness relation for the $Y_{plm}$ in the last step. For the massless minimally coupled case $\nu' = 1$ we have:
\begin{equation}
\begin{aligned}
\langle \Omega_{H}|\phi_p\phi_p'|\Omega_H\rangle &= \delta(p-p')\frac{H^2}{\sinh^2(t)}\frac{p}{4\pi^2}\frac{\cosh^2(t)+p^2}{(p^2+1)},
\end{aligned}
\end{equation}
and for large $t\rightarrow \infty$
\begin{equation}
\begin{aligned}
\langle \Omega_{H}|\phi_p\phi_p'|\Omega_H\rangle &\rightarrow \delta(p-p')\frac{H^2}{4\pi^2}\frac{p}{(p^2+1)}.
\end{aligned}
\end{equation}
The power spectrum for the massless minimally coupled scalar field is given by:
\begin{equation}
\begin{aligned}
\langle \Omega_{H}|\phi^2|\Omega_H\rangle &= \int dp \int dp' \langle \Omega_{H}|\phi_p\phi_p'|\Omega_H\rangle \\
&= \int d \ln p \frac{H^2}{4\pi^2}\frac{p^2}{p^2+1} \\
\Rightarrow \Delta^2_{\phi,H}(p) &= \frac{H^2}{4\pi^2}\frac{p^2}{p^2+1}.
\end{aligned}
\end{equation}
\textbf{Power spectrum in Bunch--Davies vacuum}\\
The computation is similar to the previous case:
\begin{equation}\label{2ptdirect3}
\begin{aligned}
\langle \Omega_{BD}|\phi_p\phi_{p'}|\Omega_{BD}\rangle &= \frac{H^2}{\sinh^2 t}\sum_{lm l'm'} \sum_{\sigma\sigma'} \frac{Y_{plm}Y^*_{p'l'm'}}{N_{\chi_{p,\sigma}}N_{\chi_{p',\sigma'}}} \\
&\times \langle \Omega_{BD}|\left(\hat{a}_{\sigma plm}\chi_{p,\sigma}+ \hat{a}^{\dagger}_{\sigma plm}\bar{\chi}_{p,\sigma}\right)\left(\hat{a}_{\sigma'p'l'm'}\chi_{p',\sigma'}+\hat{a}^{\dagger}_{\sigma'p'l'm'}\bar{\chi
}_{p',\sigma'}\right)|\Omega_{BD}\rangle \\
&= \frac{H^2}{\sinh^2 t}\sum_{lm l'm'} Y_{plm}Y^*_{p'l'm'} \sum_{\sigma\sigma'}\frac{\chi_{p,\sigma}\bar{\chi
}_{p',\sigma'}}{N_{\chi_{p,\sigma}}N_{\chi_{p',\sigma'}}} \langle \Omega_{BD}|\hat{a}_{\sigma plm}\hat{a}^{\dagger}_{\sigma'p'l'm'}|\Omega_{BD}\rangle \\
&= \delta(p-p')\frac{H^2}{\sinh^2 t}\sum_{lm}|Y_{plm}|^2\sum_{\sigma}\left|\frac{\chi_{p,\sigma}}{N_{\chi_{p,\sigma}}}\right|^2 \ \text{using} \ [\hat{a}_{\sigma plm},\hat{a}_{\sigma' p'l'm'}^{\dagger}] \\
&= \delta_{\sigma \sigma'}\delta_{ll'} \delta_{mm'} \delta(p-p') \\
&= \delta(p-p')\frac{H^2}{\sinh^2 t}\frac{p^2}{2\pi^2}\sum_{\sigma}\left(\begin{array}{l}
\left(\frac{\alpha_{p,L}^{\sigma}\bar{\alpha}^{\sigma}_{p,L}}{N^2_{\chi^{p\sigma}}}+\frac{\beta_{p,L}^{\sigma}\bar{\beta}^{\sigma}_{p,L}}{N^2_{\chi^{p,\sigma}}}\right)\left|P^{ip}_{\nu'}\right|^2 \\
+\frac{\alpha_{p,L}^{\sigma}\bar{\beta}_{p,L}^{\sigma}}{N^2_{\chi^{p\sigma}}}P^{ip}_{\nu'}P^{ip}_{\nu'}+\frac{\bar{\alpha}^{\sigma}_{p,L}\beta_{p,L}^{\sigma}}{N^2_{\chi^{p\sigma}}}P^{-ip}_{\nu'}P^{-ip}_{\nu'} \end{array}\right).
\end{aligned}
\end{equation}
In the last step we used the completeness relation for $Y_{plm}$ and the expansion of $\chi$ in terms of the associated Legendre polynomials (\ref{tanakaexpans},\ref{alfabeta}). Here we will compute the spectrum for the massless scalar field $(\nu'=1)$. For the massless minimally coupled scalar ($\nu'=1$) the cross terms involving $P^{ip}P^{ip}$ and $P^{-ip}P^{-ip}$ vanish:
\begin{equation}
\begin{aligned}
\sum_{\sigma}\frac{\alpha_{p,L}^{\sigma}\bar{\beta}_{p,L}^{\sigma}}{
N^2_{\chi^{p\sigma}}} &\propto \sum_{\sigma}\frac{(e^{\pi p}+\sigma)(e^{-\pi p}+\sigma)}{\cosh \pi p +\sigma} \propto \sum_{\sigma}\sigma = 0,
\end{aligned}
\end{equation}
and similarly for the term involving $P^{-ip}P^{-ip}$. So for the massless minimally coupled case $(\nu'=1)$ we have:
\begin{equation}\label{2ptmasslessdirect}
\begin{aligned}
\langle \Omega_{BD}|\phi_p\phi_{p'}|\Omega_{BD}\rangle
&= \delta(p-p')\frac{H^2}{\sinh^2 t}\frac{p^2}{2\pi^2}\sum_{\sigma}
\left(\frac{\alpha_{p,L}^{\sigma}\bar{\alpha}^{\sigma}_{p,L}+\beta_{p,L}^{\sigma}\bar{\beta}^{\sigma}_{p,L}}{N^2_{\chi^{p,\sigma}}}\right)\left|P^{ip}_{1}\right|^2 \\
&= \delta(p-p')\frac{H^2}{\sinh^2 t}\frac{p^2}{2\pi^2}\sum_{\sigma}\frac{\pi}{16\sinh^2\pi p}\frac{(e^{\pi p}+\sigma)^2+(e^{-\pi p}+\sigma)^2}{\cosh\pi p+\sigma}\left|P^{ip}_{1}\right|^2 \\
&= \delta(p-p')\frac{H^2}{\sinh^2 t}\frac{p^2}{2\pi^2}\frac{\pi \cosh(\pi p)}{2\sinh^2(\pi p)}\left|P^{ip}_1\right|^2 \\
&= \delta(p-p')\frac{H^2}{\sinh^2(t)}\frac{p(\cosh^2(t)+p^2)}{4\pi^2(p^2+1)}\coth(\pi p).
\end{aligned}
\end{equation}
For large $t\rightarrow \infty$ we have:
\begin{equation}
\begin{aligned}
\langle \Omega_{BD}|\phi_p\phi_{p'}|\Omega_{BD}\rangle
&= \delta(p-p')\frac{H^2}{4\pi^2} \frac{p\coth(\pi p)}{p^2+1}
\end{aligned}
\end{equation}
and
\begin{equation}
\begin{aligned}
\langle \Omega_{BD}|\phi^2|\Omega_{BD}\rangle &= \int dp \int dp' \langle \Omega_{BD}|\phi_p\phi_{p'}|\Omega_{BD}\rangle \\
&= \frac{H^2}{4\pi^2}\int dp \frac{p \coth \pi p}{p^2+1} +\text{supercurvature modes} \\
&= \frac{H^2}{4\pi^2}\int d\ln p \ \ \frac{p^2 \coth \pi p}{p^2+1} +\text{supercurvature modes}.
\end{aligned}
\end{equation}
The power spectrum is given by
\begin{equation}\label{BDdirect}
\begin{aligned}
\Delta^2_{\phi,BD}(p) &= \frac{H^2}{4\pi^2}\frac{p^2 \coth\pi p}{p^2+1}
\end{aligned}
\end{equation}
which reduces for $p\gg 1$ to an approximately scale invariant spectrum:
\begin{equation}
\begin{aligned}
\Delta^2_\phi(p) &\approx \frac{H^2}{4\pi^2}.
\end{aligned}
\end{equation}
\subsection{Reduced density matrix calculation}\label{app:reduceddensitymatrixspectrum}
In this section we derive the power spectrum in the Bunch-Davies state using an alternative method. We consider the reduced density matrix that remains after having traced out the degrees of freedom in the right hyperbolic patch. We find the same answer as in the direct calculation (\ref{BDdirect},\ref{2ptmasslessdirect}).
The reduced density matrix for the left hyperbolic patch has been calculated by Maldacena and Pimentel \cite{MalPim:2012} and is given by:
\begin{equation}\label{dm}
\begin{aligned}
\hat{\rho}_{L,p,l,m} &= Tr_{H_R} \left\{ |\Omega_{BD} \rangle \langle \Omega_{BD} | \right\} \\
&= (1-|\gamma_p)|^2)\sum_{n=0}^\infty |\gamma_p|^{2n} |n ; p,l,m\rangle
\langle n; p,l,m |
\end{aligned}
\end{equation}
where for the \emph{massless} scalar field $\gamma_p$ and $|n ; p,l,m\rangle$ are given by\footnote{For the massive scalar field Maldacena and Pimentel apply a Bogoliubov transformation on the set of $\hat{b}_{plm}$ operators to bring $\hat{\rho}_L$ in the form of (\ref{dm}).}:
\begin{equation}\label{gamma}
\begin{aligned}
\gamma_p(m=0) &= ie^{-\pi p} \\
|n ; p,l,m\rangle &= \frac{(\hat{b}^{\dagger}_{plm})^n}{\sqrt{n!}} |\Omega_{H}\rangle.
\end{aligned}
\end{equation}
For a point in the left hyperbolic wedge $x\in L$ the two point function is given by (\ref{Hexpansion}):
\begin{equation}
\begin{aligned}
\langle \Omega_{BD}|\phi_p\phi_{p'}|\Omega_{BD}\rangle &=
\text{Tr}_{\mathcal{H}_L}\left\{\phi_p\phi_{p'}\hat{\rho}_{L}\right\} \\
&= \delta(p-p')\frac{H^2}{\sinh^2(t)}\sum_{lm}|Y_{plm}|^{2} \frac{|P^{ip}_1|^2}{N^2_{P^p}} \\
&\times (1-|\gamma_p|^2)\sum_n |\gamma_p|^{2n}\left(2n+1\right) \\
&= \delta(p-p')\frac{H^2}{\sinh^2(t)}\frac{p^2}{2\pi^2} \frac{\cosh^2(t)+p^2}{2p(p^2+1)} \frac{1+|\gamma_p|^2}{1-|\gamma_p|^2} \\
&= \delta(p-p')\frac{H^2}{\sinh^2(t)}\frac{p}{4\pi^2} \frac{\cosh^2(t)+p^2}{(p^2+1)} \coth(\pi p)
\end{aligned}
\end{equation}
which is equal to the result of the direct calculation (\ref{2ptmasslessdirect}).
\section{Divergence of the energy momentum tensor}\label{EMT}
As is the case in the Fulling-Rindler vacuum, the energy momentum tensor diverges at the null boundary of the hyperbolic patch. One could construct lightcone coordinates $u = \eta-r$ and $v = \eta+r$ in order to calculate $T_{uu}$. Equivalently, we consider the leading divergence of $T_{tt}$ in the $t\rightarrow 0$ limit, which is more convenient.
\begin{equation}
\begin{aligned}
T_{tt} &= (\partial_t \phi)^2-\frac{1}{2}g_{tt}g^{\sigma \rho}(\partial_{\sigma}\phi)(\partial_{\rho}\phi)
\end{aligned}
\end{equation}
For the massless case we have:
\begin{equation}
\begin{aligned}
\langle T_{tt} \rangle &= \frac{1}{2}\langle (\partial_t\phi)^2 + g^{rr}(\partial_r\phi)^2+g^{\theta\theta}(\partial_{\theta}\phi)^2+g^{\phi\phi}(\partial_{\phi\phi}\phi)^2\rangle.
\end{aligned}
\end{equation}
One can calculate this directly using the hyperbolic mode functions (\ref{Hmodes}) and the density matrix (\ref{dm}) for the Bunch-Davies expectation value $\langle T_{\mu\nu}\rangle_{BD}$. Equivalently, for the leading order term we can use the Wightman functions $G^{+}(x,x')$ as given in \cite{Sasaki:1995}:
\begin{equation}
\begin{aligned}
\langle (\partial_{t}\phi)^2 \rangle &= \lim_{t' \rightarrow t}\partial_{t}\partial_{t'}G^{+}(t,t'),
\end{aligned}
\end{equation}
and similarly for the other coordinates. Note that the contribution of the supercurvature modes to the Wightman function only leads to subleading divergences\footnote{For $\nu' > 0$ the supercurvature mode contribution to the Wightman function is \cite{Sasaki:1995}: \begin{equation} G^{+}_{*}(t,t',\zeta) = \frac{H^2}{4\pi^{\frac{5}{2}}}\Gamma[-\nu'+1]\Gamma[\nu'+\frac{3}{2}]\frac{\sinh (\nu')\zeta}{\sinh \zeta}\left(\sinh t \sinh t'\right)^{\nu'-1}.\end{equation} For the minimally coupled massless case $\nu = \frac{3}{2}$ the supercurvature mode becomes time-independent. The contribution to the energy momentum tensor is of subleading order.}. The contribution of the subcurvature modes to the Wightman function for the massless $\nu' =1$ case is given by \cite{Sasaki:1995}:
\begin{equation}
\begin{aligned}
G^{+}(t,t',\zeta) &= \frac{H^2}{\sinh t \sinh t'}\frac{1}{8\pi^2}\int_{-\infty}^{\infty}dp \frac{\sin p \zeta}{\sinh \zeta}\frac{e^{\pi p}}{\sinh \pi p}\frac{(\cosh t+ip)(\cosh t'- ip)}{1+p^2}\left(\frac{\tanh \frac{t'}{2}}{\tanh\frac{t}{2}}\right)^{ip}, \\
\zeta &= \cosh r \cosh r'-\sinh r\sinh r'\left(\cos \theta \cos\theta'+\sin \theta \sin\theta'\cos(\phi -\phi' )\right).
\end{aligned}
\end{equation}
One can check the following:
\begin{equation}\label{factors}
\begin{aligned}
\langle (\partial_t\phi)^2 \rangle &= \lim_{t'\rightarrow t}\partial_{t}\partial_{t'}G^{+}(t,t',\zeta) &&= \frac{H^2}{4\pi^2}\int_0^{p_F}dp \ p(p^2+1)\coth(\pi p)\frac{1}{t^4}+O\left(\frac{1}{t^2}\right) \\
\langle (\partial_r\phi)^2 \rangle &= \lim_{r'\rightarrow r}\partial_{r}\partial_{r'}G^{+}(t,t,\zeta) &&= \frac{H^2}{4\pi^2}\int_0^{p_F}dp \ p(p^2+1)\coth(\pi p)\frac{1}{t^2}+O\left(t^0\right) \\
\langle (\partial_{\theta}\phi)^2 \rangle &= \lim_{\theta'\rightarrow \theta}\partial_{\theta}\partial_{\theta'}G^{+}(t,t,\zeta) &&= \sinh^2r \frac{H^2}{4\pi^2}\int_0^{p_F}dp \ p(p^2+1)\coth(\pi p)\frac{1}{t^2}+O\left(t^0\right) \\
\langle (\partial_{\phi}\phi)^2 \rangle &= \lim_{\phi'\rightarrow \phi}\partial_{\phi}\partial_{\phi'}G^{+}(t,t,\zeta) &&= \sin^2\theta \sinh^2r \frac{H^2}{4\pi^2}\int_0^{p_F}dp \ p(p^2+1)\coth(\pi p)\frac{1}{t^2}+O\left(t^0\right).
\end{aligned}
\end{equation}
Note that all these are divergent as $t\rightarrow 0$, but they also show the usual UV-divergence. The UV-divergence is regulated by a cutoff $ p_F$. The difference $\langle T_{tt}\rangle_H-\langle T_{tt}\rangle_{BD}$ will be UV-finite. We combine the components (\ref{factors}) to obtain $\langle T_{tt}\rangle_{BD}$. The expectation value $\langle T_{tt}\rangle_H$ is obtained by replacing $\coth \pi p \rightarrow 1$, where we use the expectation value $\langle \hat{b}^{\dagger}_{plm}\hat{b}_{plm}+\mathbf{I}\rangle$ in the two different states \footnote{We can calculate $\langle \hat{b}^{\dagger}_{plm}\hat{b}_{plm}+\mathbf{I}\rangle$ by using the density matrix (\ref{dm})}:
\begin{equation}
\begin{aligned}
\langle \hat{b}^{\dagger}_{plm}\hat{b}_{plm}+\mathbf{I}\rangle_{BD} &=\coth \pi p \\
\langle \hat{b}^{\dagger}_{plm}\hat{b}_{plm}+\mathbf{I}\rangle_{H} &= 1.
\end{aligned}
\end{equation}
Finally, we calculate the difference $\langle T_{tt}\rangle_H-\langle T_{tt}\rangle_{BD}$:
\begin{equation}
\begin{aligned}
\langle T_{tt}\rangle_H-\langle T_{tt}\rangle_{BD} &= \frac{H^4}{2\pi^2}\int_0^{\infty} dp \ p(p^2+1)\left(1-\coth \pi p\right) \frac{1}{t^4} + O\left(\frac{1}{t^2}\right) \\
&= -\frac{11}{240\pi}\frac{1}{t^4}+O\left(\frac{1}{t^2}\right).
\end{aligned}
\end{equation}
Note that we took the cutoff $p_F$ to infinity and obtain a UV-finite integral.
\section{The hyperbolic vacuum embedded in the Bunch--Davies state}\label{app:Hvacuum}
From \cite{MalPim:2012} we have for the massless case $\nu' = 1$:
\begin{equation}
\begin{aligned}
|\Omega_{\text{BD}}\rangle &= \left(\otimes_{plm} e^{\gamma_p \hat{b}_{Lplm}^{\dagger}\otimes \hat{b}^{\dagger}_{Rplm}}\right)|\Omega_{H,L}\rangle\otimes |\Omega_{H,R}\rangle
\end{aligned}
\end{equation}
or suppressing the indices $p,l,m$:
\begin{equation}
\begin{aligned}
|\Omega_{\text{BD}}\rangle &= e^{\gamma \hat{b}_{L}^{\dagger}\otimes \hat{b}^{\dagger}_{R}}|\Omega_{H,L}\rangle\otimes |\Omega_{H,R}\rangle
\end{aligned}
\end{equation}
with $\gamma = ie^{-\pi p}$.
The left hyperbolic vacuum $|\Omega_{H,L}\rangle$ is not a state of the full system; we need information about the state in the right hyperbolic patch as well. The simplest way to embed the left hyperbolic vacuum in the full Hilbert space, we can consider the simple and symmetric state $|\Omega_{H,L}\rangle\otimes |\Omega_{H,R}\rangle$. This state is \emph{not} the natural vacuum state (the Bunch--Davies state) for the full de Sitter space. \\
\textbf{Proposition:}
\begin{equation}\label{prop}
\begin{aligned}
|\Omega_{H,L}\rangle\otimes |\Omega_{H,R}\rangle &\propto e^{-|\gamma_p| \ \hat{a}^{\dagger}_{+}\otimes \ \hat{a}^{\dagger}_{-}}|\Omega_{BD}\rangle.
\end{aligned}
\end{equation}
\textbf{Proof:} \\
We will show that the right hand side of (\ref{prop}) vanishes when we act with any of the $\hat{b}_{Llpm}$ annihilation operators. We use the expression for the hyperbolic annihilation operator $\hat{b}_{Llpm}$ in terms of the creation and annihilation for Bunch-Davies modes (\ref{relationab}), suppressing from now on the labels $p,l,m$:
\begin{equation}\label{basimple}
\begin{aligned}
\hat{b}_{L} &= \sum_{\sigma}\frac{N_{P}}{N_{\chi_{\sigma}}}\left(\alpha^{\sigma}_{L}\hat{a}_{\sigma}+\bar{\beta}^{\sigma}_{L}\hat{a}^{\dagger}_{\sigma}\right).
\end{aligned}
\end{equation}
We want to show that $\hat{b}_{L}$ acting on the RHS of (\ref{prop}) vanishes:
\begin{equation}\label{duurtlang}
\begin{aligned}
\sum_{\sigma}\frac{N_{P}}{N_{\chi_{\sigma}}}\left(\alpha^{\sigma}_{L}\hat{a}_{\sigma}+\bar{\beta}^{\sigma}_{L}\hat{a}^{\dagger}_{\sigma}\right) e^{-|\gamma_p| \hat{a}^{\dagger}_{+}\otimes \hat{a}^{\dagger}_{-}}|\Omega_{BD}\rangle \stackrel{?}{=} 0.
\end{aligned}
\end{equation}
Consider the annihilation operator $\hat{a}_{\sigma}$ acting on (\ref{prop}):
\begin{equation}
\begin{aligned}
\hat{a}_{\pm}e^{-|\gamma_p|\hat{a}^{\dagger}_{+}\hat{a}^{\dagger}_{-}}|\Omega_{BD}\rangle &= \left[\hat{a}_{\pm} \ , \ e^{-|\gamma_p| \hat{a}^{\dagger}_{+}\hat{a}^{\dagger}_{-}}\right]|\Omega_{BD}\rangle \\
&= -\left[\hat{a}_{\pm} \ , \ |\gamma_p|\hat{a}^{\dagger}_{+}\hat{a}^{\dagger}_{-} \right] e^{-|\gamma_p| \hat{a}^{\dagger}_{+}\hat{a}^{\dagger}_{-}}|\Omega_{BD}\rangle \\
&= -|\gamma_p| \hat{a}^{\dagger}_{\mp} e^{-|\gamma_p| \hat{a}^{\dagger}_{+}\hat{a}^{\dagger}_{-}}|\Omega_{BD}\rangle.
\end{aligned}
\end{equation}
Substituting this result in (\ref{duurtlang}) gives:
\begin{equation}\label{duurtlanger}
\begin{aligned}
\sum_{\sigma}\frac{N_{P}}{N_{\chi_{\sigma}}}\left(\alpha^{\sigma}_{L}\hat{a}_{\sigma}+\bar{\beta}^{\sigma}_{L}\hat{a}^{\dagger}_{\sigma}\right) e^{|\gamma_p| \hat{a}^{\dagger}_{+}\otimes \hat{a}^{\dagger}_{-}}|\Omega_{BD}\rangle = N_{P}\sum_{\sigma}\left(-\frac{\alpha^{-\sigma}_{L}}{N_{\chi_{-\sigma}}}|\gamma|+\frac{\bar{\beta}^{\sigma}_{L}}{N_{\chi_{\sigma}}}\right)\hat{a}^{\dagger}_{\sigma} e^{|\gamma_p| \hat{a}^{\dagger}_{+}\otimes \hat{a}^{\dagger}_{-}}|\Omega_{BD}\rangle.
\end{aligned}
\end{equation}
It is easy to check that the quantity between brackets on the RHS of (\ref{duurtlanger}) vanishes for both $\sigma = \pm 1$. This finalizes the proof:
\begin{equation}
\hat{b}_{L}e^{-|\gamma|\hat{a}^{\dagger}_{+}\otimes \hat{a}^{\dagger}_{-}}|\Omega_{BD}\rangle = 0 \ \ \ \ \ \ \ \forall p,l,m.
\end{equation}
The state (\ref{prop}) is pure. Note that the symmetric and antisymmetric modes corresponding to $\sigma=\pm1$ are entangled with each other and their reduced density matrices are thermal.
\end{appendix}
|
\section{Introduction}
Alarming statistics about distracted driving can be found on the official US government website about distracted driving \cite{distraction.gov}. In 2010, 18\% of injury crashes were distraction-related. 3331 people were killed in 2011 in a crash involving a distracted driver, and distraction is responsible for 11\% of fatal crashes of drivers under the age of twenty. These statistics are even more worrying as the number of possible distractions within a car keeps increasing. The large number of displays and new infotainment devices in cars has made the problem more critical. \\
The National Highway Transportation Safety Administration (NHTSA) has defined distracted driving as ``an activity that could divert a person's attention away from the primary task of driving''. It is commonly classified into 3 categories, namely
\begin{itemize}
\item Manual distraction: The driver takes his hands off the wheel. This includes for example text messaging, eating, using a navigation system, or adjusting radio.
\item Visual distraction: The driver takes his eyes off the road, for example reading or watching a video.
\item Cognitive distraction: The driver's mind is not fully focused on driving. This can happen when the driver is talking to other passengers, texting, or simply thinking.
\end{itemize}
The influence of distraction on drivers performance has been widely studied \cite{stutts2003driver}, \cite{strayer2004profiles}, \cite{blueprint}, \cite{Trafficsafetyfacts}, \cite{olson2009driver}, and interesting facts have come to light: cell phones use represent 18\% of distracted fatal driver accident in North America. Indeed, cell phone conversations induce a high level of cognitive distraction, thus reducing the brain activity related to driving by 37\% (which might be worse than ingesting alcohol). Handsfree cell phones have not been found particularly safer that hand-held use. More importantly, text messaging requires visual, manual, and cognitive attention at the same time, making it the most dangerous distraction. It was found that text messaging takes the driver's eyes off the road for 4.6 seconds, which is sufficient to drive the length of a football field completely blind. The crash risk when text messaging is twenty-three times worse than driving with no distraction. \\
All these facts suggest that drivers should be aware of the risk, but it is also the car manufacturer's responsibility to offer intelligent assistance tools to avoid driver distraction, and to limit crash risks. This issue is still an open problem, as the variety of actions, the differences between drivers and outdoor conditions make this task extremely challenging.\\
Our approach aims at determining first if a driver is distracted or not, and in the case he is, the system should be able to recognize the type of distraction. Based on computer vision techniques, we propose four different modules for features extraction, focusing on arm position, face orientation, facial expression and eye behavior. We propose two strategies to combine the output information from each module: an AdaBoost classifier with temporal smoothing, and a Hidden Markov Model-based classifier.\\
This paper is organized as follows: Section \ref{sec:lit_rev} is an overview of existing methods and commercial products. Section \ref{sec:modules} presents in detail each module for assessing driver distraction. Section \ref{sec:fusion} describes our fusion strategies for distraction recognition, section \ref{sec:experimental_results} shows our main experimental results and demonstrates the efficiency of our approach. Last section \ref{sec:conclusion} is a concluding section, discussing our results and highlights our future work.
\section{Literature review}
\label{sec:lit_rev}
In this section, we review the existing studies that have been carried out in the field of driver inattention, and we discuss their strengths and shortcomings. We only present here approaches that were useful for our study, or having a significant impact on the community. For a more extensive survey, some literature reviews can be found in: \cite{williamson2005review}, \cite{young2007driver}, \cite{dong2011driver}.\\
First of all, an important distinction should be made between driver inattention, driver fatigue and driver distraction. They all alter driver performance, but they are not caused by the same factors and can have various effects on the driver behavior. Fatigue is related to drowsiness and is affecting the driver because of physical, physiological or psychological reasons. Distraction was defined earlier and is related to an object, a person, an idea or an event that diverts the driver. Fatigue and distraction are both considered as driver inattention. A precise definition and relationship analysis between those terms has been attempted by Regan et al. \cite{regan2011driver}. Our work deals exclusively with driver distraction, and most of existing methods apply either for driver fatigue or driver inattention in general. As many techniques, features and fusion strategies are similar for fatigue, inattention and distraction, our review considers both fatigue and distraction detection methods.\\
Three main categories of system have been used for determining driver inattention, and a few studies have used a hybrid approach to combine them:
\subsubsection{Physiological sensors}
This approach detects physiological features such as brain activity, heart rate or hand moist \cite{shiwu2011active}, \cite{lal2002driver}. In particular, electroencephalograph (EEG) has been found to be a valid, objective and accurate measure of driver inattention using $\delta$, $\theta$, $\alpha$, $\beta$ and $\sigma$ brain wave activity. However, physiological sensors are intrusive and cannot be used inside a car for commercial applications. They can be used as ground truth for studies, but they do not represent a realistic solution for driver inattention monitoring.
\subsubsection{Driver performance}
This approach uses external information and indicators of driver performance to infer the level of inattention. This includes, for example lateral position, steering wheel movements or pedal activity. For example, Raynney et al. \cite{ranney2008driver} suggested that distraction involved an sensitive lack of vehicle control, such as drifting from a side of the road, or unexpected speed changes. In 2007, Volvo introduced the Driver Alert Control system \cite{Volvo}, monitoring the road continuously with a camera and alerting the driver in case of dangerous behavior. These methods are correlated with driver inattention \cite{pilutti1997identification}, but they are also affected by external factors such as driver experience, road type, weather and outdoor light. Moreover, the measures rely on long term statistics and the system is unable to predict immediate dangers such as micro-sleep events.
\subsubsection{Computer vision}
The third approach, probably the most popular, relies on visual driver features. When inattentive, the driver's face and body show characteristic behaviors. Placing a camera in front of the driver and analysing his face expressions and movements makes a lot of sense and is considered as an efficient way for assessing driver inattention. In particular, the so-called eyes-off-road glance duration and head-off-road glance time are recognized as valid measures for visual distraction and can be assessed using an embedded camera \cite{angell2006driver}. For driver fatigue, PERCLOS (percentage of eye closure) \cite{dinges1998perclos} is considered as the best correlated physiological feature. Other behaviors such as yawning \cite{ji2006probabilistic}, \cite{Ji2004},
\cite{bergasa2008analysing} or nodding \cite{Bergasa2006}, \cite{Senaratne2007} are also popular features, widely used in the field. Existing systems usually rely on a hardware setup that can be a simple color camera \cite{smith2003determining}, \cite{DOrazio2007}, \cite{Senaratne2007}, an infrared camera able to alternate between bright and dark pupil effect (useful for eye detection and robust to illumination variations) \cite{Bergasa2006}, \cite{ji2006probabilistic}, \cite{Ji2004}, \cite{craye2013multi} or a set of cameras to improve face orientation estimation \cite{bergasa2008analysing}. Recently, the Microsoft Kinect sensor has received particular attention \cite{limulti2012} as it provides both color camera, depth map and comes with a powerful face tracker. Image processing and computer vision tools are then used for extracting indicators of inattention. The key components are face, eyes and mouth detection. This is the starting point for feature extraction such as yawning, nodding, face motion, gaze estimation, or blink detection. Statistics such as PERCLOS, gaze distribution, yawning or nodding frequencies can be computed based on the features to infer driver inattention. Last, a fusion module is designed in order to merge the information and infer the level of inattention. Most popular fusion techniques are fuzzy logic \cite{Bergasa2006}, \cite{damousis2008fuzzy}, \cite{Senaratne2007}, Bayesian networks \cite{Ji2004} and dynamic Bayesian networks \cite{ji2006probabilistic}, neural networks \cite{limulti2012} or simple decision rules \cite{smith2003determining}, \cite{bergasa2008analysing}.
Recently, commercial products, such as Eye Alert \cite{eyeAlert} have been conceived to detect driver inattention using computer vision and emit a warning in case of dangerous situation.
\subsubsection{Hybrid systems}
Last, a few approaches use a combination of the three aforementioned techniques. For example, Daza et al. \cite{daza2011drowsiness} used both driver (PERCLOS, nodding, etc.) and driving (lane variation, steering wheel movements, etc.) data to assess fatigue. Fletcher et al. \cite{fletcher2005correlating} have successfully merged driver gaze and road information to detect if a driver was missing any road sign. Last Li et al. \cite{limulti2012} have used both computer vision, steering wheel signal and pulse oxymeter to infer driver fatigue.
\subsubsection{Proposed approach and contributions}
Among state of the art techniques, most of them focus on driver fatigue detection, sometimes extended to driver inattention. To our knowledge, no serious study has been done on distraction only, trying to detect the type of action the driver is accomplishing. Determining the type of driver distraction provides higher level information than just the level of distraction. It could be used for number of applications related to intelligent transportation systems. For inter vehicles communication, providing the action a driver is doing can be more explicit and useful than statistics on how distracted that driver is. Even for smart cars, detecting the type of distraction enables statistics computation on the driver's behaviour that could further help the vehicle in keeping the driver safe.\\
Our system is based on a Kinect sensor. Originally conceived for entertainment and video game applications, it has quickly become a good tool for the computer vision community. Indeed, not only it was the first low-cost depth sensor for general public, but it also came with a very polished SDK, giving developers a large range of possibilities. For example, the SDK provides a quite efficient skeletal tracking algorithm and tools for gesture recognition. In our case, the RGBD (RGB-depth) data is very helpful for driver segmentation as well as face detection and tracking. To the best of our knowledge, only Li et al. \cite{limulti2012} have published work making use of the Kinect for car safety applications.\\
As most of existing systems rely solely on driver's face behavior, we also use driver's gesture to help us is our inference task. Unlike traditional approaches, our sensor is placed in such a way that driver's upper body is visible. Thus, we can extract driver arms position and motion. This feature will be of major help for determining driver distraction.
\section{Feature extraction}
\label{sec:modules}
This section explains how to extract features from each body and face components. We divide our task into four independent modules, namely (1) arm position estimation, (2) face orientation, (3) facial features - called animation units (AUs) - such as mouth shape and eyebrow raising, and (4) gaze estimation and eye closure. Each module uses either depth data, color data, or both. They are fused later on via different fusion schemes to determine the type of distraction. The next sections describe in detail the realization of each module.
\subsection{Arm position}
\begin{figure*}
\centering
\includegraphics[scale=0.6]{arm_position.pdf}
\caption{Main blocks of the arm position module. The input is the raw depth data. After processing, we extract features and use an AdaBoost classifier to output four scores representing the arm position.}
\label{im:arm_position}
\end{figure*}
Much effort have been deployed by Microsoft on providing an efficient skeleton tracking to help developers in gesture recognition, both for the entire and the half body (in seated mode) \cite{held20123d}. However, the algorithm has been tested on our recorded sequences, and has shown poor results, even in seated mode. The main reasons for this is the shape of the seat of our simulator, which is hard to differenciate from the body, and the relatively small motion of the driver during a driving session. Other methods and source code are available for arm tracking or pose estimation, such as \cite{PuppetParade}, \cite{zhu2007constrained}, or \cite{shotton2013real}, but none of these approach was satisfying for us, either because they required a large training set, or because we obtained poor performance on our sequences.
For this reason, we decided to build our own arm detection system within the framework of machine learning for arm position recognition. Our features are based on the depth map data and represent the orientation of the rims of the right arm of the driver. Figure \ref{im:arm_position} shows the main steps of our method.\\
\subsubsection{Background removal}
First of all, a background/foreground segmentation is applied to isolate the driver from the driving seat and background. Kinect depth data already contains 3 bits of player data, representing the location of each player, thus providing rough segmentation. As a more accurate segmentation is required in our case, we have used a background subtraction-based approach. First, short video sequences of the seat without the driver is recorded, and an average depth map of the seat only is obtained. Then, this depth map is subtracted from the depth map of driving sequences, and image thresholding is applied to remove the seat and the background. We then combine Kinect segmentation with our technique, and we clean the irrelevent blobs by keeping the biggest one only.\\
\subsubsection{Feature extraction}
Based on the segmented depth map, we now need to find features for discriminating arm position. Most of the time, the Kinect records the driver with a frontal view, and the right arm is therefore on the right side of the body. Based on this assumption, we extract features based on foreground contours. We first apply to the binary foreground image the marching squares algorithm - the 2D version of the marching cubes \cite{lorensen1987marching} - which provides us an ordered list of contour pixels (in this list, each pixel is preceded and followed by neighboring pixels in the image). Then, we only consider the pixels of the right side of the body by removing from the list the pixels on the left hand side. Those pixels represent the rim of the right head, shoulder, arm and body of the driver. \\
We then cut this ``half'' contour list into twenty successive segments, composed of the same number of pixels. Each pixel of the list is associated with a 3D point (given image coordinates and depth value), such that each segment of the list corresponds to a 3D point cloud. We then extract for each segment the main axis - a 3D vector - of the associated point cloud by principal component analysis. Putting together all the main axis, we obtain a $20 \times 3$ feature vector.\\
Using only the right contour from the frontal view is not enough, as arm and especially forearm might not always be detected. To overcome this situation, we apply the aforementioned technique to what we call the profile view: each pixel of the depth map reprensents a point in 3D world coordinates. Suppose that the $\vec{X}$ and $\vec{Y}$ coordinates are the image pixel coordinates, and $\vec{Z}$ coordinate is the (depth) value of the pixel. Thus, the depth map corresponding to the frontal view is the projection of the 3D world points onto the $(\vec{X},\vec{Y})$ plan. The profile view would then be the projection on the $(\vec{Y},\vec{Z})$ plan. Figure \ref{im:feature_examples} is an example of a segmented depth map and associated profile view. From this profile view, we apply contour detection and feature extraction just as described above. We now have 120 features. On figure \ref{im:feature_examples}, examples of features extracted for different poses are shown. Using the profile view makes possible the detection of the forearm even when the arm is in front of the body.\\
\begin{figure*}
\centering
\begin{tabular}{cccc}
&\includegraphics[scale=0.25]{frontal.pdf} &
\includegraphics[scale=0.25]{profile.pdf}& \\
& Frontal view & Profile view & \\
\includegraphics[scale=0.25]{arm_forward.pdf} &
\includegraphics[scale=0.25]{arm_right.pdf} &
\includegraphics[scale=0.25]{arm_up.pdf} &
\includegraphics[scale=0.25]{arm_down.pdf} \\
(a) Forward & (b) Right & (c) Up & (d) Down\\
\end{tabular}
\caption{First row: an example of frontal and associated profile views. Second row: examples of arm positions and associated features. Red dots are projections of point clouds' local orientations from the frontal view and blue dots from the profile view. As can be seen, profile view features are particularly useful in the case of up and down positions.}
\label{im:feature_examples}
\end{figure*}
\subsubsection{Arm position output}
After feature extraction, the arm position module outputs an estimation of the arm position among four possible states: arm up, arm down, arm right, and arm forward. In that regard, we use a machine learning technique.
Among the 120 extracted features, some are discriminative, some are useless, and some might even be contradictory. Selecting or weighting the features is therefore necessary for good classification. We decided to use an AdaBoost \cite{freund1997decision}, which is a very appropriate tool in this regard. Nevertheless, AdaBoost only solves two-class problems, so we use a 1-vs-all approach: we train four sub-classifiers, each of them specialized for one class (i.e., using as positive examples some samples for a specific class, and as negative examples, some samples from any other class). The estimated position would be the one with highest value among the four sub-classifiers. The actual output of the module is not exactly the estimated position, but the score of each sub-classifier. Indeed, this four-values information is better-suited for the fusion stage.
\subsection{Eyes behavior}
Eye behavior is a very important feature for driver distraction, including both eye gaze and eye blinking. Therefore, this module aims to localize the iris of the driver and deduce his gaze position, while detecting whether the eye is open or closed. For this module, we only rely on the Kinect color stream. Figure \ref{im:eye_flowchart} shows a summary of the module.\\
\begin{figure*}
\centering
\includegraphics[scale=0.55]{eye_flowchart.pdf}
\caption{Main steps of the eye behavior module}
\label{im:eye_flowchart}
\end{figure*}
\subsubsection{Iris localization}
Iris detection first relies on finding the eye corners positions. This is done using the face tracking algorithm provided by the Kinect SDK, further described in Section \ref{sec:facetracking}. We then use a robust iris detection method based on cost function maximization and spatio-temporal considerations. Given the eye corners locations, we isolate the eye by creating a squared patch with the same width as the eye corner distance. To make sure that our parameters are scale-invariant, we resize this square to a $60\times 60$ patch. Our cost function is based on the response of three different filters described below.\\
The first one is the circular Hough transform that has been widely used in iris detection problems, for example in \cite{1384533} or \cite{kawaguchi2003iris}. We choose a low edge detection threshold to make sure that the iris contour is detected in any situation, and we use a varying radius from six to nine pixels. Thus, we make sure that the iris will always be responsive to the filter.\\
The second filter was used in Zhang \textit{et al.} \cite{zhang2008new}, and relies on the circular Gabor filter. It provides significant impulse response to round objects, which is appropriate for iris detection. For that we convolve our eye template with the following Gabor kernel:
\begin{equation}
G(x,y) = g(x,y)exp(2i\pi F \sqrt{x^2+y^2})
\end{equation}
where $F$ is the radial frequency, set to 0.0884 and $g$ is the Gaussian envelope, defined as:
\begin{equation}
g(x,y) = \frac{1}{\sqrt{2\pi \sigma^2}}exp(-\frac{x^2+y^2}{2\sigma^2})
\end{equation}
with $\sigma$ the variance, set to 4.5 in our system.
The third filter is inspired by Kawaguchi \textit{et al.} \cite{kawaguchi2003iris}. It relies on the high intensity difference between the iris and its immediate neighborhood. For this, we convolve the eye template with the masks $R_1$, $R_2$, $R_3$ represented in Figure \ref{im:masks}. The radius of mask $R_1$ is 6, and for masks $R_2$ and $R_3$ is 15. We obtain three transformed eye templates, called $C_1$, $C_2$, and $C_3$. We combine them to obtain our separability response using the following formula:
\begin{equation}
S(x,y) = \frac{C_2(x,y)-C_1(x,y)}{C_1(x,y)} + \frac{C_3(x,y)-C_1(x,y)}{C_1(x,y)}
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=0.4]{masks_iris.pdf}
\caption{Masks $R_1$, $R_2$ and $R_3$ for separability measure. Dark areas represent pixel value 1 and white areas represent pixel value 0}
\label{im:masks}
\end{figure}
Last, we normalize our three filter responses, sum them up, and take the maximum value as our iris center estimate.
The detection is now quite accurate, but a few mistakes can be avoided and corrected using spatio-temporal information. We therefore use a simple spatio temporal consistency checking to improve the detection performance. For each template, we look at the predicted position of each four filter (including the sum of the three filters) and the previous estimated locations. We then apply simple rules based on distances and eye position to remove irrelevent detections.
\subsubsection{Gaze estimation}
Now that the iris is detected, we generate a feature set based on gaze estimation. Gaze can be estimated from a 3D model of the face determining the 3D-world orientation, and the iris location given the eyeballs' sizes and positions \cite{heyma-11}. We could use such an approach, but we need only a rough estimation of where the driver is looking, and statistical measures of the eye gaze distributions. For that reason, the features we extract are simply the relative position of the iris to the eyes' corners: from the face tracing, we extract the eye corner positions, and we generate the 4 dimensional feature vector as follows:
\begin{equation}
\begin{pmatrix} x_{l}\\y_{l}\\x_r\\y_r\end{pmatrix} =
\begin{pmatrix} \frac{X_l-C_l^{(l)}}{C_l^{(l)}-C_{l}^{(r)}} \\ \\ \frac{X_r-C_r^{(l)}}{C_r^{(l)}-C_{r}^{(r)}} \end{pmatrix}
\end{equation}
with $X_l$,$X_r$ the left (right) iris position and $C_i^{j}$ ($i,j \in {left, right}$) the $j$ corner of the $i$ eye.
\subsubsection{Eye closure detection}
We use an additional feature which determines whether the driver's irises are visible (eyes open) or not (eyes closed). This will be helpful when taking into account the amount of time the driver is not looking at the road and the potential danger this represents. A simple yet efficient approach for this is to construct a database of open and closed eyes, and to apply template matching or classification techniques \cite{Choi2011},\cite{arai2011comparative}, \cite{grauman2003communication}. We use a simple classification approach. When pupil position is estimated, we create a small iris template, centered at the iris location. We normalize the grayscale to make it more illumination-insensitive. We create a subset of around 2000 eyes templates, and we manually label each of them as whether it corresponds to an open eye with visible pupil, or a closed eye (or non-visible pupil). Using this dataset, we train an SVM classifier using an RBF kernel with $\sigma$ = 13. Last, we use the SVM for each session, and we add to the output of the module the SVM score (not the output label) for each eye and each frame.\\
Figure \ref{im:iris_detection_result} shows examples of pupil detection to qualitatively illustrate the performance of our detector.
\begin{figure*}
\centering
\begin{tabular}{cccc}
\includegraphics[scale=0.35]{results2150.pdf} &
\includegraphics[scale=0.35]{results2200.pdf} &
\includegraphics[scale=0.35]{results1120.pdf}&
\includegraphics[scale=0.35]{results1310.pdf} \\
\includegraphics[scale=0.35]{results2520.pdf} &
\includegraphics[scale=0.35]{results2230.pdf} &
\includegraphics[scale=0.35]{results2610.pdf} &
\includegraphics[scale=0.35]{results2560.pdf} \\
\end{tabular}
\caption{Examples of iris detection involving various poses, expressions, and image qualities.}
\label{im:iris_detection_result}
\end{figure*}
\subsection{Head pose and facial expression}
\label{sec:facetracking}
The Kinect SDK provides very useful features and cutting-edge algorithms to help developers. The face tracking algorithm \cite{faceTracking} is one of them, and reveals itself to be particularity useful for our work. It uses cooperative fusion of the depth map and the color image to first estimate the root of the head, and then provides a robust and accurate face model. It relies on the active appearance model \cite{edwards1998interpreting} extended to 3D coordinates. The raw output of the face tracking is a set of 100 vertices and 121 triangles forming a mesh of the face. The face was tracked in most situations, but failed in case of rapid and significant face rotations, or when occluded by an object (typically when drinking). A few seconds were also required most of the time to correctly initialize the face model and fit it to the face. In the case where face tracking was not providing any output, the eye behavior module was deactivated as no eye location could be found.\\
Upper level information is also available from the face tracking, making the output much more meaningful and useful for our tasks. We use that high level information as the output of two of our modules.
\subsubsection{Face orientation} Based on 3D vertices coordinates, face tracking can provide head orientation angles and head center 3D position. For our work, we extract only the head orientation, namely the pitch, roll, and yaw angles, which are values between -180 and 180 degrees. The position depended too much on the driver's height and did not help in the classification task.
\subsubsection{Facial expression} Face tracking also provides six animation units (AUs) based on the definition of the Candide3 model \cite{ahlberg2001candide}. AUs are expressed as coefficients and represent how strongly distorted features of the face are. We extract only mouth-related AUs: upper lip raiser (AU10), jaw lowerer (AU26/27), lip stretcher (AU20), and lip corner depressor (AU13/15). Other AUs are eyebrows-related and did not help in the recognition task. More detailed information about AUs provided by the face tracker can be found on the Microsoft website \cite{faceTracking}. The output of the module is a set of 6 AUs extracted from the face model (if any).
Figure \ref{im:faceTracking} shows sample images and associated face tracking results. As can be seen, tracking is efficient in a number of situations.
\begin{figure*}
\centering
\begin{tabular}{ccc}
\includegraphics[scale=0.22]{face10.jpg} &
\includegraphics[scale=0.22]{face3.jpg} &
\includegraphics[scale=0.22]{face4.jpg} \\
(a) & (b) & (c) \\
\includegraphics[scale=0.22]{face5.jpg} &
\includegraphics[scale=0.22]{face6.jpg} &
\includegraphics[scale=0.22]{face7.jpg} \\
(d) & (e) & (f) \\
\end{tabular}
\caption{Examples of face tracking under different face poses (a),(b),(c) and facial expressions (d), (e), (f)}
\label{im:faceTracking}
\end{figure*}
\section{Fusion}
\label{sec:fusion}
In this section, we explain how to merge the aforementioned module outputs, and deduce the type of distractions taken by the driver.
Detecting distraction can be assimilated to a gesture recognition a gesture recognition problem. As gesture recognition is getting more and more popular in the fields of entertainment and security, in-vehicle applications also exists, for example for automatic commands \cite{westeyn2003georgia}. However, our goal is slightly different and somehow more difficult in our case, as the driver is not cooperative when taking the action. Plus, there is no single way to accomplish an action which might be unexpectedly interrupted because of road constraints. Even for a human being, isolating and recognizing actions inside a car is not always obvious.
\subsection{Dataset}
A set of video sequences was collected on our experimental platform. Five distractive tasks were recorded and manually labelled for training and evaluation. Features were extracted based on the modules of Section \ref{sec:modules}, on a frame by frame basis. They were then concatenated such that a seventeen-features vector was representing each frame. More about the conditions of the experiments has been described in Section \ref{sec:experimental_results}.
\subsection{Feature extraction and temporal smoothing}
\label{sec:smooth}
Classification can be achieved using each frame independently, but a much better performance was obtained using temporal considerations. For this, we applied a running median filter, using a hundred-sample sliding window. This corresponds to a window length of approximately six seconds, which means that the beginning of an action was predicted with a delay of three seconds. We also computed the running standard deviation within the hundred-samples window, providing information about the temporal variation of each feature. The standard deviation information has shown better performance and stability compared to regular speed and acceleration features ($\delta$ and $\delta^2$ features). As a result, we obtained thirty-four features per sample. We can now train a classifier with a portion of this dataset to obtain a distraction recognition system.
\subsection{AdaBoost classifier}
Depending on the type of distraction we want to identify, the extracted features are either strongly discriminative or unsignificant. In that regard, AdaBoost can provide an appropriate solution. Briefly, AdaBoost (for Adaptive Boosting) is able to turn a set of T weak classifiers into a stronger one, by linearly combining each of them in an optimal way.
\begin{equation}
H(x) = sign(\sum_{t = 1}^T{\alpha_th_t(x)})
\end{equation}
At each iteration ($T$ in total), a weak classifier $h_t$ is selected from among a family of classifiers, namely, the one minimizing the weighted error rate. The weights and $\alpha_t$ are updated based on the minimum error rate. The weak classifiers we use are decision trees, which, by definition, are trained to select the most representative features maximizing the information gain.
As distraction identification is a multi-class classification problem, we used a 1-vs-all approach. Each class was trained using a simple real AdaBoost, initialized with a decision tree of depth four and 300 iterations.
The temporal aspect of the dataset can also help refining the frame-by-frame classification; therefore, we used an additional temporal filtering: we first classified each sample independently, and we replaced each estimated $S$ output by the mode (most frequent output) over a hundred-samples sliding window centered at $S$. This way, isolated misclassifications were removed.
\subsection{Hidden Markov Model }
Because of the temporal aspect of our dataset, stochastic models can be applied. Similar to speech recognition tasks, gesture recognition can be successfully achieved using Hidden Markov models (HMMs) \cite{lee1999hmm}, conditional random fields \cite{wang2006hidden}, or recursive types of classifiers, making the classification time-dependent \cite{murakami1991gesture}.
We chose to use HMMs, which provide a high-level of flexibility for modeling the structure of an observation sequence. HMMs allow recognition of the hidden state sequence by the dependency of the observation sequence on the hidden states and the topology of the state transitions. It is now acknowledged that the use of HMM is fundamental in temporal pattern recognition.
We have developed our system using the HMM Toolkit (HTK) \cite{young2002htk}. The labels associated with each samples were the five possible states of distraction described in Section
We train a different Markov model for each class, and used the Viterbi algorithm to decide which state each sample belongs to. We tried several configurations, varying the number of hidden states (from five to twenty) and the type of data (raw or smoothed, with or without $\delta$ and $\delta^2$ features).
\section{Experimental results}
\label{sec:experimental_results}
In this section, we demonstrate the efficiency of our system. First, a description of the experimental setup and dataset acquisition is described, then we evaluate the distraction detection and recognition using AdaBoost and Hidden Markov models.
\subsection{Experimental setup and data collection}
The data collection was obtained using a driving simulator. The simulator is equipped with monitors, steering wheel and driving software City Car driving \cite{CCD}. We used a Kinect sensor to record color video, depth map and audio. Steering wheel and pedal position as well as hear rate were also recorded. Last the entire driving session was captured based on a 1 fps screenshot of the monitors. In this paper, the only sensor used is the Kinect, but the other sensors will be used in a future work.\\
A total of 8 drivers were asked to participate in the study. Drivers
were either men of women, from different countries, either wearing
glasses or not with age varying from 24 to 40 years old. They were
all experimented drivers, but using their car at varying frequencies.
Each driver was recorded during four 15 minutes sessions. Two sessions
were in the early morning, when driver was awake and alert, and two
sessions in the late evening, at the end of a full business day, when
the driver was tired. Each session was either on a highway with low
traffic, or in the city with higher traffic. During each session,
the driver had to follow a well-defined procedure involving several
tasks putting him into distracted situations and miming visual signs.
During a driving session, about half of the time was normal driving,
and the other half was distracted driving.
More precisely, the drivers were asked to accomplish five distracting
tasks during each driving session, namely
\begin{itemize}
\item making a phone call
\item drinking
\item sending an SMS
\item looking at an object inside the vehicle (either a map or adjusting
the radio)
\item driving normally
\end{itemize}
As a result, we have obtained around eight hours of driving sequences, that were manually labelled for analysis. The dataset was evaluated using cross validation: The performance
on each driver was evaluated separately by taking the data related
to the driver as the testing set, and the other drivers as the training
set. A ground truth was manually labeled for each session and was
available each second. The performance of the whole system was measured
by the total average accuracy, calculated according to the following
formula:
\begin{equation}
Total\, Average\, Accuracy=\frac{Number\, of\, correct\, decisions}{Total\, number\, of\, frames}\label{eq:Total_Average_Accuracy}
\end{equation}
\subsection{Results}
A first approach for classification is to use a time-independent classifier, and add temporal refinement to increase frame by-frame-accuracy. For practical reasons, we use an AdaBoost classifier again using a 1-vs-all approach. Each class is trained using a simple real AdaBoost initialized with a decision tree of depth four and 300 iterations.
We evaluate the action recognition using AdaBoost and HMM classifiers. For each driver, we evaluate the distraction recognition capacity by training a classifier using all driver sessions except the driver to be evaluated, and testing using all sessions involving this driver. For AdaBoost, the best accuracy was obtained using a Real AdaBoost algorithm with initialization based on a decision tree of depth four and 300 iterations. For HMM, the optimal parameters were a ten-states automaton with a single Gaussian mixture for modeling each node, and using the smoothed data described in Section \ref{sec:smooth} rather than the original raw data. No significant improvement was found when using $\delta$ and $\delta^2$ features.\\
Table \ref{table_overall} presents the overall accuracy of the AdaBoost and HMM classifiers for each driver. We provide the accuracy for the five classes and the accuracy for distraction detection only: in this case, we have merged all the classes involving distraction into a single class, and compared it with the normal driving class. Average accuracies are quite close between AdaBoost and HMM (85.05\% and 84.78\%), but the results for each driver may vary. More specifically, HMM is outperforms AdaBoost for most of the drivers, but for a few drivers, HMM performs significantly worse than AdaBoost. This may be due to instabilities related to the high dimensionality of the features and the number of states compared to the size of our training set. Increasing the number of drivers might solve this issue. Also, one might be surprised that HMM does not perform as superbly as it does for the gesture recognition task. Again, this is because the driver is not cooperative in this case, and there is no single way to accomplish a distractive task. For example, when drinking, a driver can put the container down between each swallow, or just keep it in the hand. When phoning, the driver can place the phone between his head and his shoulder, or keep it in hand. Moreover, actions can be interrupted suddenly, because the driver needs both hands on the wheel to turn, or greater focus on the road to avoid accidents. All these constraints make the actions less similar in time, and significantly limits the HMM performance. Using a bigger dataset might eventually improve the classification results. As regards distraction-only accuracy (89.84\% and 89.64\%), results suggest that our system can successfully detect whether a driver is actually distracted or not.\\
\begin{table*}
\caption{Accuracies of AdaBoost and HMM classifiers for distraction recognition (5 classes) and distraction detection (2 classes)}
\label{table_overall}
\centering
\begin{tabular}{|>{\centering\arraybackslash}p{2.2cm}|>{\centering\arraybackslash}p{2.2cm}|>{\centering\arraybackslash}p{2.2cm}|>{\centering\arraybackslash}p{2.2cm}|>{\centering\arraybackslash}p{2.2cm}|}
\hline
& \multicolumn{2}{c|}{\bf AdaBoost} & \multicolumn{2}{c|}{\bf HMM} \\
\hline
Subject & Distraction recognition & Distraction detection & Distraction recognition & Distraction detection \\
\hline
1 & 89.36 & 91.44 & 84.00 & 92.31 \\
\hline
2 & 86.02 & 90.05 & 87.41 & 91.16 \\
\hline
3 & 87.38 & 90.95 & 90.19 & 96.41 \\
\hline
4 & 85.48 & 94.72 & 81.85 & 84.65 \\
\hline
5 & 81.14 & 85.05 & 83.68 & 83.82 \\
\hline
6 & 80.94 & 87.7 & 81.52 & 89.53 \\
\hline
\bf{Average} & \bf{85.05} & \bf{89.84} & \bf{84.78} & \bf{89.64} \\
\hline
\end{tabular}
\end{table*}
Another test we did was to train the AdaBoost classifier with each module separately and evaluate their inference capacity (See table \ref{tab:accuPerFeatures}). We found that features related with arm position were by far the most discriminative, probably because normal driving position was easy to detect, and represented almost 40\% of a driving session. Other features were also useful, but provided efficient recognition only for specific actions. For example, we found that face orientation was very discriminative for drinking and text messaging, face expression helped a lot differentiating phone call and drinking and eyes behaviour was efficient for text messaging and normal driving. AdaBoost was an appropriate choice of classifier, as decision trees as weak classifiers were doing feature selection for each type of action.
\begin{table*}
\centering
\caption{Accuracies per features}
\label{tab:accuPerFeatures}
\begin{tabular}{|>{\centering\arraybackslash}p{2cm}|>{\centering\arraybackslash}p{2cm}|>{\centering\arraybackslash}p{2cm}|>{\centering\arraybackslash}p{2cm}|>{\centering\arraybackslash}p{2cm}|>{\centering\arraybackslash}p{2cm}|}
\hline
Subject & Arm & Orientation & Expression & Eyes & All \\
\hline
1 & 76.49 & 55.9 & 54.31 & 65.7 & 89.36 \\
\hline
2 & 70.04 & 56.35 & 55.45 & 61.72 & 86.02\\
\hline
3 & 85.99 & 66.8 & 56.48 & 38.6 & 87.38 \\
\hline
4 & 69.80 & 63.63 & 55.09 & 59.95 & 85.48\\
\hline
5 & 74.55 & 66.6 & 56.30 & 59.06 & 81.14\\
\hline
6 & 81.01 & 77.87 & 64.25 & 63.85 & 80.94 \\
\hline
\end{tabular}
\end{table*}
The overall accuracy is a good indicator of system performance, but it does not say how each class is correctly detected. Table \ref{tab:each_class} provides a few classification metrics for each class, based on the average of each driver performance. For each class, extremelly high accuracy is due to an unbalanced testing set. Indeed, a single action (except normal driving) represents between 10\% and 20\% of the entire sequence. Therefore, recall and precision measures are better indicators of the classification capacity. From the table, it is clear that phone call and normal driving were quite successfully detected. Drinking was a little behind, mainly because the action was sometimes very fast (the driver just swallowed for a few seconds and put the drink down) and attenuated by temporal smoothing. The worst performance was for object distraction, probably because this action required neither huge visual nor cognitive attention. In that regard, the action was pretty similar to normal driving and therefore hard to detect, even for a human being.
In order to get more insight about those results, figure \ref{im:result_classification} displays the frame by frame classification for a given sequence. Ground truth is the blue lines, and estimated class is the red one. In this example, phone call and text message are accurately detected, drinking comes with a few false positives and object distraction is often considered as normal driving. We have added a few frames to provide a better visualization of why detection was successful or not. Correct detections are represented with green frames and arrow, whereas false detections are in red. The drinking false positives are often due to strong arm movements (when changing gear during phone call for example). We believe that a better temporal analysis could remove that type of false positive detections. For object distraction misclassification, the sample frames show that the driver does not look extremely distracted, and it can be hard to say if he is actually adjusting the radio or just having his hand on the gear lever. Fortunately object distraction is the less demanding distraction and therefore the less dangerous, making the misclassification in that case less critical.
\begin{table*}
\centering
\caption{Classification measures for each class}
\label{tab:each_class}
\begin{tabular}{|>{\centering\arraybackslash}p{2cm}|>{\centering\arraybackslash}p{2cm}|>{\centering\arraybackslash}p{2cm}|>{\centering\arraybackslash}p{2cm}|>{\centering\arraybackslash}p{2cm}|>{\centering\arraybackslash}p{2cm}|}
\hline
Subject & Phone call & Text message & Drinking & Object distraction & Normal driving \\
\hline
Accuracy & 95.54 & 96.24 & 95.55 & 92.79 & 89.98 \\
\hline
Sensitivity /recall & 80.56 & 73.63 & 86.14 & 24.78 & 96.00\\
\hline
Specificity & 97.90 & 98.65 & 96.19 & 98.91 & 87.38 \\
\hline
Precision & 87.51 & 89.22 & 51.45 & 68.28 & 85.48\\
\hline
f measure & 81.71 & 72.04 & 60.79 & 26.30 & 81.14\\
\hline
g-means & 87.98 & 78.32 & 90.91 & 40.16 & 80.94 \\
\hline
\end{tabular}
\end{table*}
\begin{figure*}
\includegraphics[scale = 0.35]{result_classification.pdf}
\caption{Results of action recognition for a given sequence. Ground truth (red) and estimated actions (blue) are displayed for each frame (x axis). Sample frames are displayed to illustrate the correct and wrong classification results for some actions. Green frames means correctly detected, red frame means wrong detection}
\label{im:result_classification}
\end{figure*}
\section{Conclusion and future work}
\label{sec:conclusion}
We have successfully built an inattention detection and recognition tool based on Kinect sensor, computer vision and machine learning. Based on four modules extracting data from arm position, face orientation, face expression and eye behaviour, we have constructed two types of classifiers to perform inattention action recognition. Based on data collected with a driving simulator, we were able to evaluate our work, and results show that our method is accurate and might be used in a real car.\\
Compared to existing approaches aiming to detect inattention or provide a level of inattention, our system outputs higher level information, more suitable for context aware human-machine interaction. Not only can it be used for immediate driver safety, but also for long term statistics about the driver habits, or for inter vehicles communications systems. Moreover, the different modules we have constructed are extremely flexible and could be used for other type of statistics computation and inference.\\
Future work directions are as follows: first, we believe that action recognition could be improved using more temporal information. For example, drinking and phone call are sometimes mixed up and alternating in successive frames, whereas it is very unlikely in practice that a driver is doing both at the same time. Such mistakes could be avoided. Next, the modules we have designed could allow fatigue detection using PERCLOS, nodding and yawning frequencies for example. Therefore, efforts will be put on fatigue detection in future work. We also plan to use background and environment information to assess the level of risk on the road. We might use a Bayesian network and later on dynamic Bayesian network in order to fuse multiple sources related to the driver (age, driving experience, fatigue) and the environment (time of the day, road type, outside traffic, vehicle speed) in addition to distraction type to assess the level of danger the driver is exposed to. Moreover, during the recorded sessions, we also have used additional sensors such as microphone, heart rate monitor or steering wheel signals. We will work on integrating those signals to the system. Last, we also plan to record more driving sessions, involving additional actions, per say interactions and chatting with other passengers.
|
\section{Introduction and statement of main results}\label{section:intro}
This work can be regarded as a continuation of some of the investigations initiated by the first author and Saff in \cite{LopSaff}, where greedy energy sequences with respect to general kernels were introduced and many properties of these sequences were obtained. Greedy energy sequences are defined in \cite{LopSaff} as follows. Let $X$ be a locally compact space containing infinitely many points, and let $k: X\times X\longrightarrow \mathbb{R}\cup\{+\infty\}$ be a symmetric and lower semicontinuous function. The function $k$ is referred to as the \emph{kernel}. Given a compact set $A\subset X$, a sequence $(a_{n})_{n=0}^{\infty}\subset A$ is called a \emph{greedy $k$-energy sequence} on $A$ if it is generated in the following way:
\begin{itemize}
\item[$\bullet$] The first point $a_{0}$ is selected arbitrarily on $A$.
\item[$\bullet$] Assuming that $a_{0},\ldots,a_{n}$ have been selected, $a_{n+1}$ is chosen to satisfy
\begin{equation}\label{eq:alggreedy}
\sum_{i=0}^{n} k(a_{n+1},a_{i})=\inf_{x\in A} \sum_{i=0}^{n} k(x,a_{i})
\end{equation}
for every $n\geq 0$.
\end{itemize}
The existence of a point $a_{n+1}$ satisfying \eqref{eq:alggreedy} follows from the lower semicontinuity of $k$, but of course the choice of $a_{n+1}$ may not be unique in general. As in \cite{LopSaff}, we will use here the notation $\alpha_{N}:=(a_{0},\ldots,a_{N-1})$ to denote the first $N$ points of the sequence $(a_{n})_{n=0}^{\infty}$.
If $X=\mathbb{C}$ and we use the logarithmic kernel $k(x,y)=-\log |x-y|$ (here and below $|x-y|$ indicates the Euclidean distance between $x$ and $y$), then the above algorithm generates the classical Leja sequences on compact subsets of the complex plane. If $X=\mathbb{R}^{p}$, $p\geq 2$, and $k(x,y)=|x-y|^{-s}$ is the Riesz $s$-kernel with parameter $s>0$, then the sequences obtained are the greedy $s$-energy sequences on compact subsets of $\mathbb{R}^{p}$.
The main interest for the study of greedy energy sequences in \cite{LopSaff} was to compare the asymptotic behavior of the configurations $\alpha_{N}$ with the asymptotic behavior of optimal energy configurations. The study of the asymptotic properties of optimal energy configurations has been a leitmotif in the work of Saff, and his efforts have produced a tremendous advancement in the theory of discrete minimal energy problems. In this work we will be referring frequently to optimal energy configurations and their asymptotic properties, so we define them next.
Given a set $\omega=\{x_{1},\ldots,x_{N}\}$ of $N$ ($N\geq 2$) points in $X$, not necessarily distinct, we write $\card(\omega)=N$ and we define the \emph{discrete energy} of $\omega$ with respect to $k$ by
\begin{equation}\label{def:energy}
E(\omega):=\sum_{1\leq i\neq j\leq N}k(x_{i},x_{j})=2\sum_{1\leq i<j\leq N} k(x_{i},x_{j}).
\end{equation}
Now assume that $A\subset X$ is a compact set and $N\geq 2$ is an integer. A set $\omega_{N}\subset A$ is an \emph{optimal $N$-point configuration} on $A$ if
\[
E(\omega_{N})=\inf\{E(\omega): \omega\subset A, \card(\omega)=N\},
\]
that is, $\omega_{N}$ has the lowest possible energy among all $N$-point configurations on $A$. Note that the existence of optimal energy configurations is guaranteed by the lower semicontinuity of $k$.
Part of the results in \cite{LopSaff} were obtained in the context of the unit circle $S^{1}$ and the Riesz $s$-kernel. It was shown in \cite{LopSaff} that in terms of first-order asymptotics, there is a difference in the behavior of greedy $s$-energy sequences and optimal $N$-point configurations when $s>1$ (see \cite[Proposition 2.6]{LopSaff}) which is not present in the case $s\leq 1$. Moreover, this difference takes place in the more general context of rectifiable Jordan arcs or curves, see \cite[Theorem 2.9]{LopSaff}. It was also shown in \cite{LopSaff} that on $S^{1}$ and in the case $0<s\leq 1$, the second-order asymptotic behavior of greedy $s$-energy sequences is no longer the same as that of optimal $N$-point configurations (see \cite[Propositions 2.4 and 2.7]{LopSaff}). These differences will be explained in detail below.
In this paper we investigate more deeply the asymptotic behavior of Leja sequences and greedy $s$-energy sequences on $S^{1}$ from the energy point of view. Consequently we are able to refine some of the results in \cite{LopSaff} mentioned above. We first describe the results we obtain for Leja sequences and later for greedy $s$-energy sequences. The results we obtain have also motivated some conjectures for general sequences on $S^{1}$ that we state in Section~\ref{section:conjectures}.
\subsection{Results for Leja sequences on the unit circle}
Recall that a Leja sequence $(a_{n})_{n=0}^{\infty}$ on an infinite compact set $K\subset\mathbb{C}$ is a sequence that is constructed by choosing an arbitrary $a_{0}\in K$, and selecting each subsequent $a_{n+1}\in K$ such that
\begin{equation}\label{Lejaproperty}
\sum_{i=0}^{n}\log\frac{1}{|a_{n+1}-a_{i}|}=\inf_{z\in K}\sum_{i=0}^{n}\log\frac{1}{|z-a_{i}|},\qquad n\geq 0.
\end{equation}
Equivalently, for every $n\geq 0$, $a_{n+1}$ maximizes the product $\prod_{i=0}^{n}|z-a_{i}|$ on $K$. Leja sequences are named after F. Leja in recognition of his work \cite{Leja}, although they were first introduced by A. Edrei in \cite{Edrei}. These sequences have attracted some interest in recent years, especially concerning the study of their interpolation properties, see e.g. \cite{BiaCal, CalManh, Chkifa, Reichel, TaylorTotik}. Not many works have been devoted to the study of the energy and distribution of Leja sequences; some of these are \cite{Gotz, Pritsker, CD, Lop, LopSaff}.
Given a configuration $\omega=\{x_{1},\ldots,x_{N}\}$ of $N\geq 2$ distinct points in the complex plane $\mathbb{C}$, we will denote by $E_{0}(\omega)$ its logarithmic energy, that is,
\[
E_{0}(\omega):=\sum_{1\leq i\neq j\leq N}\log\frac{1}{|x_{i}-x_{j}|}=2\sum_{1\leq i<j\leq N}\log\frac{1}{|x_{i}-x_{j}|},
\]
see \eqref{def:energy}.
Let $(a_{n})_{n=0}^{\infty}$ be a Leja sequence on a compact set $K\subset\mathbb{C}$, and recall that $\alpha_{N}=(a_{0},\ldots,a_{N-1})$ denotes the $N$-tuple of the first $N$ points of this sequence. A well-known result in logarithmic potential theory that can be consulted in \cite[Theorem V.1.1]{SaffTotik} asserts that if $K$ is non-polar (i.e., $K$ supports a positive measure with finite logarithmic energy), then
\begin{equation}\label{eq:firstorderasymp}
\lim_{N\rightarrow\infty}\frac{E_{0}(\alpha_{N})}{N^{2}}=\inf_{\mu\in\mathcal{P}(K)}\iint\log\frac{1}{|z-w|}\,d\mu(w)\,d\mu(z),
\end{equation}
where $\mathcal{P}(K)$ is the set of probability measures supported on $K$. Moreover, the sequence of point configurations $\alpha_{N}$ has as limiting distribution the equilibrium measure on $K$, which is the unique probability measure $\nu$ on $K$ satisfying the extremal property
\begin{equation}\label{contlogenergy}
\iint\log\frac{1}{|z-w|}\,d\nu(w)\,d\nu(z)=\inf_{\mu\in \mathcal{P}(K)}\iint\log\frac{1}{|z-w|}\,d\mu(w)\,d\mu(z).
\end{equation}
The asymptotic result \eqref{eq:firstorderasymp} was first proved for Fekete sets on $K$ by Fekete and Szeg\H{o}, see \cite[Theorem 5.5.2]{Ran}. Fekete sets on $K$ consisting of $N\geq 2$ points are exactly optimal $N$-point configurations on $K$ relative to the logarithmic kernel; that is, they are configurations $\omega_{N}\subset K$ satisfying
\[
E_{0}(\omega_{N})=\inf\{E_{0}(\omega): \omega\subset K, \card(\omega)=N\}.
\]
Before we state our results for Leja sequences on $S^{1}$ we describe some basic properties of these sequences. Firstly, it is clear that a rotation (by multiplication with $\rho\in\mathbb{C}$, $|\rho|=1$) will neither destroy the Leja sequence property \eqref{Lejaproperty} nor change the logarithmic energy of the configurations $\alpha_{N}$. So, it suffices to consider Leja sequences starting with initial point $1$. Following the terminology used in \cite{BiaCal} and \cite{CalManh}, we will refer to the configurations $\alpha_{N}$ as \emph{$N$-Leja sections}.
Leja sequences on $S^{1}$ can be described in detail by the following properties obtained by L. Bialas-Ciez and J.-P. Calvi in \cite[Theorem 5]{BiaCal}, see also \cite[Lemma 4.2]{LopSaff}. Let us define first the notation $(A,B)=(a_{0},\ldots,a_{N-1},b_{0},\ldots,b_{M-1})$ for an $N$-tuple $A=(a_{0},\ldots,a_{N-1})$ and an $M$-tuple $B=(b_{0},\ldots,b_{M-1})$.
Then:
\begin{itemize}
\item[$1)$] Any $2^{n}$-Leja section is formed by the $2^{n}$th roots of unity.
\item[$2)$] Given any $2^{n+1}$-Leja section $\alpha_{2^{n+1}}$ containing the $2^{n}$-Leja section $\alpha_{2^{n}}$ as its first $2^{n}$ points, there exists a $2^{n}$th root $\rho$ of $-1$ and a $2^{n}$-Leja section $\beta_{2^{n}}$ such that $\alpha_{2^{n+1}}=(\alpha_{2^{n}},\rho\,\beta_{2^{n}})$.
\item[$3)$] Iterating $2)$, it is easily seen that for any $k$-Leja section $\alpha_{k}$ with $k=2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{t}}$, $n_{1}>n_{2}>\cdots >n_{t}\geq0$, there exists for each $i=1,\ldots,t$ a $2^{n_{i}}$-Leja section $\alpha^{i}_{2^{n_{i}}}$ (with initial point $1$) such that
\[
\alpha_{k}= (\alpha^{1}_{2^{n_{1}}},\rho_{1}\,\alpha^{2}_{2^{n_{2}}},\rho_{1}\rho_{2}\,\alpha^{3}_{2^{n_{3}}},\ldots, \left(\prod_{i=1}^{t-1}\rho_{i}\right)\alpha^{t}_{2^{n_{t}}}),
\]
for some numbers $\rho_{i}$ that are $2^{n_{i}}$th roots of $-1$. In other words, any Leja section is composed of rotated Leja sections of smaller size.
\end{itemize}
Concerning the asymptotic behavior of $E_{0}(\alpha_{N})$ for Leja sequences on the unit circle, the asymptotic formula \eqref{eq:firstorderasymp} applied in this context gives
\[
\lim_{N\rightarrow\infty}\frac{E_{0}(\alpha_{N})}{N^2}=0,
\]
since the equilibrium measure on $S^{1}$ is the normalized arclength measure and its logarithmic energy \eqref{contlogenergy} is zero. In this paper we prove the following.
\begin{theorem}\label{theo:firstorderasymp}
If $(a_{n})_{n=0}^{\infty}$ is a Leja sequence on $S^{1}$, then for the sequence $\alpha_{N}=(a_{0},\ldots,a_{N-1})$ we have
\begin{equation}\label{firstorderlimit}
\lim_{N\rightarrow\infty}\frac{E_{0}(\alpha_{N})}{N\log N}=-1.
\end{equation}
\end{theorem}
We note that each Fekete set on $S^{1}$ with $N\geq 2$ points is a rotated copy of the $N$th roots of unity having logarithmic energy $-N\log N$. Therefore $E_{0}(\omega)\geq -N\log N$ for any $N$-point configuration $\omega$ on $S^{1}$. A refinement of \eqref{firstorderlimit} is the following second order estimate.
\begin{theorem}\label{theo:secondorderest}
Under the same assumptions as in Theorem $\ref{theo:firstorderasymp}$, for every $N$ we have
\begin{equation}\label{secondorderest}
0\leq \frac{E_{0}(\alpha_{N})+N\log(N)}{N}<\log(4/3).
\end{equation}
The upper bound in \eqref{secondorderest} is best possible since
\begin{equation}\label{ulsecondorder}
\limsup_{N\rightarrow\infty}\frac{E_{0}(\alpha_{N})+N\log N}{N}=\log(4/3).
\end{equation}
\end{theorem}
Observe that if $N$ is a power of $2$, then the lower bound in \eqref{secondorderest} is attained. The estimates in \eqref{secondorderest} imply \eqref{firstorderlimit}, but we shall provide a direct proof of \eqref{firstorderlimit} not using \eqref{secondorderest}.
\subsection{Results for greedy $s$-energy sequences on the unit circle}
Let $s>0$ and let $\omega=\{x_{1},\ldots,x_{N}\}\subset\mathbb{C}$ be a configuration of $N\geq 2$ distinct points. We denote by $E_{s}(\omega)$ the Riesz $s$-energy of $\omega$, that is,
\[
E_{s}(\omega):= \sum_{1\leq i\neq j\leq N}\frac{1}{|x_{i}-x_{j}|^{s}}= 2\sum_{1\leq i< j\leq N}\frac{1}{|x_{i}-x_{j}|^{s}}.
\]
In this paper we shall also analyze the asymptotic behavior of the Riesz $s$-energy of the first $N$ points of a greedy $s$-energy sequence on $S^{1}$. Recall that by definition, such sequences $(a_{n})_{n=0}^{\infty}\subset S^{1}$ are obtained by choosing an arbitrary $a_{0}\in S^{1}$ and selecting each subsequent $a_{n+1}\in S^{1}$, $n\geq 0$, such that
\[
\sum_{i=0}^{n}\frac{1}{|a_{n+1}-a_{i}|^{s}}=\inf_{z\in S^{1}}\sum_{i=0}^{n}\frac{1}{|z-a_{i}|^{s}}.
\]
The first important observation we make is that for any $s>0$, greedy $s$-energy sequences coincide with Leja sequences on $S^{1}$ due to the symmetry of the circle. This can be easily deduced from an induction argument that uses Lemmas 4.1 and 4.2 from \cite{LopSaff} which we will omit here.
Also, we emphasize that for any fixed $s>0$, if $\omega_{N}$ is an optimal $N$-point configuration on $S^{1}$ that minimizes the Riesz $s$-energy, i.e., $\omega_{N}$ satisfies
\[
E_{s}(\omega_{N})=\inf\{E_{s}(\omega): \omega\subset S^{1}, \card(\omega)=N\},
\]
then $\omega_{N}$ is again formed by $N$ equally spaced points, see \cite{Gotz2}.
Following the notation used in \cite{BHS}, we will denote by $\mathcal{L}_{s}(N)$ the Riesz $s$-energy ($s>0$) of $N$ equally spaced points on the unit circle, i.e.,
\[
\mathcal{L}_{s}(N):=E_{s}(\{z_{k,N}\}_{k=1}^{N}),\qquad z_{k,N}:=\exp(2\pi i (k-1)/N),\quad k=1,\ldots,N.
\]
It is easy to see that
\[
\mathcal{L}_{s}(N)=2^{-s} N \sum_{k=1}^{N-1}\left(\sin\frac{k\pi}{N}\right)^{-s},\qquad N\geq 2,
\]
using $|e^{i\xi}-e^{i\theta}|=2 |\sin (\frac{\xi-\theta}{2})|$. By convention we set $\mathcal{L}_{s}(1)=0$.
If the Riesz parameter $s$ satisfies $0<s<1$, one can still use potential theory, as in the logarithmic case, to study the asymptotic behavior of $\mathcal{L}_{s}(N)$ and the Riesz $s$-energy of greedy $s$-energy configurations. The following first-order asymptotic results are known and can be proved using the same techniques. We have
\begin{equation}\label{firstorderRiesz}
\lim_{N\rightarrow\infty}\frac{\mathcal{L}_{s}(N)}{N^{2}}=\lim_{N\rightarrow\infty}\frac{E_{s}(\alpha_{N})}{N^{2}}=I_{s}(\sigma),
\end{equation}
where $\sigma$ is the normalized arc length measure on $S^{1}$, which minimizes the energy
\begin{equation}\label{def:continuousRieszenergy}
I_{s}(\mu):=\iint\frac{1}{|x-y|^{s}}\,d\mu(x)\,d\mu(y)
\end{equation}
among all probability measures on $S^{1}$. For a proof of \eqref{firstorderRiesz}, see \cite{Landkof,LopSaff}. The limiting value in \eqref{firstorderRiesz} is given by
\[
I_{s}(\sigma)=\frac{1}{2\pi}\int_{0}^{2\pi}\frac{1}{|1-e^{i\phi}|^s}\,d\phi=2^{-s}\,\frac{\Gamma((1-s)/2)}{\sqrt{\pi}\,\Gamma(1-s/2)},
\]
cf. \cite[2.5.3.1]{Prud}.
In terms of second-order asymptotics for the sequence $\mathcal{L}_{s}(N)$, the following limit holds (see \cite{BHS}):
\begin{equation}\label{eq:secorderasympRieszFekete}
\lim_{N\rightarrow\infty}\frac{\mathcal{L}_{s}(N)-I_{s}(\sigma) N^{2}}{N^{1+s}}=\frac{2\zeta(s)}{(2\pi)^{s}},
\end{equation}
where $\zeta(s)$ is the analytic extension of the classical Riemann zeta function. It should be noted that in the range $s\in(0,1)$ we have $\zeta(s)<0$. In contrast to \eqref{eq:secorderasympRieszFekete}, it was shown in \cite[Corollary 2.5]{LopSaff} that in the case of greedy $s$-energy sequences on $S^{1}$ and the corresponding configurations $\alpha_{N}$, the sequence
\[
\left(\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}\right)_{N}
\]
is not convergent. In this paper we look more closely at this sequence.
In order to state our results in the Riesz setting, we need to introduce certain notations and definitions.
\begin{definition}\label{def:Thetap}
Let $p\geq 1$ be a fixed integer. We let $\Theta_{p}\subset [0,1]^{p}$ denote the set of all vectors $\vec{\theta}=(\theta_{1},\theta_{2},\ldots,\theta_{p})$ for which there exists an infinite sequence $\mathcal{N}$ of integers $N=2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{p}}$, $n_{1}>n_{2}>\cdots>n_{p}\geq 0$, satisfying
\begin{equation}\label{eq:conddefthetavec}
\lim_{N\in\mathcal{N}}\frac{2^{n_{i}}}{N}=\theta_{i},\qquad \mbox{for all}\,\,i=1,\ldots,p.
\end{equation}
\end{definition}
\smallskip
Note that if $(\theta_{1},\ldots,\theta_{p})\in\Theta_{p}$, then
\begin{equation}\label{eq:propvecinTheta}
\sum_{i=1}^{p}\theta_{i}=1.
\end{equation}
On $\Theta_{p}\times [0,\infty)$ we define the following function
\begin{equation}\label{def:H}
H((\theta_{1},\ldots,\theta_{p});s):=\sum_{k=1}^{p}\theta_{k}^{s}\,\Big(2 (2^s-1) \left(\sum_{j=k+1}^{p}\theta_{j}\right)+\theta_{k}\Big).
\end{equation}
It follows from \eqref{eq:propvecinTheta} that for any $\vec{\theta}=(\theta_{1},\ldots,\theta_{p})\in\Theta_{p}$ we have
\[
H(\vec{\theta};0)=H(\vec{\theta};1)=1.
\]
In Section~\ref{section:RieszenergyLeja} we give some further remarks about the sets $\Theta_{p}$ and the functions $H$ in \eqref{def:H}.
The graphs of some functions $H$ associated with three vectors $\vec{\theta}$ are shown in Figure~\ref{graphsHfunctions}.
\begin{figure}[h]
\centering
\includegraphics[totalheight=2.5in,keepaspectratio]{grafico3.eps}
\caption{In increasing order, we show the graphs of the functions \eqref{def:H} associated with the vectors $\vec{\theta}=(16/21, 4/21, 1/21)$, $\vec{\theta}=(4/5,1/5)$ and $\vec{\theta}=(2/3,1/3)$, respectively.}\label{graphsHfunctions}
\end{figure}
\begin{definition}\label{def:functionsh}
Let $0<s<1$ be fixed. Using the function \eqref{def:H} we introduce the notations
\begin{align}
\underline{h}_{p}(s) & :=\inf_{\vec{\theta}\in\Theta_{p}} H(\vec{\theta};s),\qquad p\in\mathbb{N},\label{def:hplower:1}\\
\underline{h}(s) & :=\inf_{p\in\mathbb{N}} \underline{h}_{p}(s).\label{def:hplower:2}
\end{align}
Similarly, for $s>1$ fixed we define
\begin{align}
\overline{h}_{p}(s) & :=\sup_{\vec{\theta}\in\Theta_{p}} H(\vec{\theta};s),\qquad p\in\mathbb{N},\label{def:hpupper:1}\\
\overline{h}(s) & :=\sup_{p\in\mathbb{N}} \overline{h}_{p}(s).\label{def:hpupper:2}
\end{align}
\end{definition}
Unfortunately we have not found an explicit expression of the functions $\underline{h}(s)$ and $\overline{h}(s)$. Our next result is the following.
\begin{theorem}\label{theo:secorderasympRieszenergy:1}
Let $s\in(0,1)$ be fixed, and let $(a_{n})_{n=0}^{\infty}$ be a greedy $s$-energy sequence on $S^{1}$. Then, for the sequence of configurations $\alpha_{N}=(a_{n})_{n=0}^{N-1}$ we have
\begin{equation}\label{eq:secorderasympRieszenergy:1}
\limsup_{N\rightarrow\infty}\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}=\underline{h}(s)\,\frac{2\zeta(s)}{(2\pi)^{s}},
\end{equation}
where $\zeta(s)$ is the analytic extension of the classical Riemann zeta function, and $\underline{h}(s)$ is defined in \eqref{def:hplower:1}--\eqref{def:hplower:2}. We also have
\begin{equation}\label{secondorderasympRieszenergy:2}
\liminf_{N\rightarrow\infty}\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}=\frac{2\zeta(s)}{(2\pi)^{s}}.
\end{equation}
In particular, the sequence $\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}$ is not convergent since $\underline{h}(s)<1$ for every $s\in (0,1)$.
\end{theorem}
In contrast to the case $s\in(0,1)$, if $s\geq 1$ potential-theoretic tools are no longer available to study the asymptotic behavior of $\mathcal{L}_{s}(N)$ or $E_{s}(\alpha_{N})$. This is due to the fact that in this case the continuous Riesz $s$-energy \eqref{def:continuousRieszenergy} of any probability measure $\mu$ on $S^{1}$ is infinite.
As a particular case of a general result for rectifiable Jordan curves in $\mathbb{R}^{d}$ proved in \cite[Theorem 3.2]{MMRS}, we know that if $s>1$ then
\begin{equation}\label{eq:asympfirstorderRieszsupercrit}
\lim_{N\rightarrow\infty}\frac{\mathcal{L}_{s}(N)}{N^{1+s}}=\frac{2\zeta(s)}{(2\pi)^{s}},
\end{equation}
where $\zeta(s)=\sum_{n=1}^{\infty}n^{-s}$ denotes now the classical Riemann zeta function.
Concerning greedy $s$-energy sequences, we have the following result, analogous to Theorem~\ref{theo:secorderasympRieszenergy:1}.
\begin{theorem}\label{theo:firstorderasympRieszenergy}
Let $s>1$ be fixed, and let $(a_{n})_{n=0}^{\infty}$ be a greedy $s$-sequence on $S^{1}$. Then, for the sequence of configurations $\alpha_{N}=(a_{n})_{n=0}^{N-1}$ we have
\begin{equation}\label{eq:firstorderasympRieszenergy:1}
\limsup_{N\rightarrow\infty}\frac{E_{s}(\alpha_{N})}{N^{1+s}}=\overline{h}(s)\,\frac{2\zeta(s)}{(2\pi)^{s}},
\end{equation}
where $\zeta(s)$ is the classical Riemann zeta function, and $\overline{h}(s)$ is defined in \eqref{def:hpupper:1}--\eqref{def:hpupper:2}. We also have
\begin{equation}\label{eq:firstorderasympRieszenergy:2}
\liminf_{N\rightarrow\infty}\frac{E_{s}(\alpha_{N})}{N^{1+s}}=\frac{2\zeta(s)}{(2\pi)^{s}}.
\end{equation}
In particular, the sequence $\frac{E_{s}(\alpha_{N})}{N^{1+s}}$ is not convergent since $\overline{h}(s)>1$ for every $s>1$.
\end{theorem}
We remark that in \cite[Proposition 2.6]{LopSaff} it was already shown that the sequence $\frac{E_{s}(\alpha_{N})}{N^{1+s}}$ is not convergent. We also want to emphasize that the following result, related with \eqref{eq:firstorderasympRieszenergy:1}, can be deduced from \cite[Theorem 2.9]{LopSaff}. If $(x_{n})_{n=0}^{\infty}\subset S^{1}$ is any sequence of pairwise distinct points on the unit circle and $s>1$, then for the sequence of configurations $\omega_{N}=\{x_{0},\ldots,x_{N-1}\}$ we have
\[
\limsup_{N\rightarrow\infty}\frac{E_{s}(\omega_{N})}{N^{1+s}}>\frac{2\zeta(s)}{(2\pi)^{s}}.
\]
We finally consider the critical case $s=1$. As a corollary of \cite[Theorem 3.2]{MMRS} and \cite[Theorem 2.10]{LopSaff} we know that
\[
\lim_{N\rightarrow\infty}\frac{\mathcal{L}_{1}(N)}{N^2 \log N}=\lim_{N\rightarrow\infty}\frac{E_{1}(\alpha_{N})}{N^2 \log N}=\frac{1}{\pi},
\]
for any greedy $s$-energy sequence ($s=1$) on $S^{1}$ and the corresponding configurations $\alpha_{N}$. Moreover, we have the following second-order asymptotics (see \cite{BHS}):
\begin{equation}\label{eq:secorderFeketecrit}
\lim_{N\rightarrow\infty}\frac{\mathcal{L}_{1}(N)-\frac{1}{\pi}N^2 \log N}{N^2}=\frac{1}{\pi}(\gamma-\log(\pi/2)),
\end{equation}
where $\gamma=\lim_{N\rightarrow\infty} (1+\frac{1}{2}+\cdots+\frac{1}{N}-\log N)$ denotes the Euler-Mascheroni constant.
In our next result we consider the corresponding second-order expression
\[
\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}\,N^{2}\,\log N}{N^2}.
\]
In order to state this result we need some definitions.
For $\vec{\theta}=(\theta_{1},\ldots,\theta)\in\Theta_{p}$ we let
\begin{equation}\label{def:K}
K(\theta_{1},\ldots,\theta_{p}):= 2\log 2+\sum_{k=1}^{p}\theta_{k}^2\, \log \left(\theta_{k}/4\right)+2\sum_{k=1}^{p-1}\left(\sum_{j=k+1}^{p}\theta_{j}\right) \theta_{k}\log \theta_{k},
\end{equation}
where if $\theta_{k}=0$, we understand in \eqref{def:K} that $\theta_{k} \log \theta_{k}=0$. Let
\begin{equation}\label{def:kappa}
\kappa:=\sup_{p\in\mathbb{N}}\, \sup_{\vec{\theta}\in\Theta_{p}} K(\vec{\theta}).
\end{equation}
\begin{theorem}\label{theo:criticalcase}
Let $(a_{n})_{n=0}^{\infty}$ be a greedy $s$-energy sequence on $S^{1}$ for $s=1$. Then, for the sequence of configurations $\alpha_{N}=(a_{n})_{n=0}^{N-1}$ we have
\begin{equation}\label{secondorderasympcritcase:1}
\limsup_{N\rightarrow\infty}\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}N^2 \log N}{N^2}=\frac{1}{\pi}(\gamma-\log(\pi/2)+\kappa),
\end{equation}
where $\kappa$ is the constant in \eqref{def:kappa}. We also have
\begin{equation}\label{secondorderasympcritcase:2}
\liminf_{N\rightarrow\infty}\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}N^2 \log N}{N^2}=\frac{1}{\pi}(\gamma-\log(\pi/2)).
\end{equation}
In particular, the sequence $\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}N^2 \log N}{N^2}$ is not convergent since $\kappa>0$.
\end{theorem}
We remark that in \cite[Corollary 2.8]{LopSaff} it was already shown that the sequence $\frac{E_{1}(\alpha_{N})-\frac{1}{\pi} N^{2}\log N}{N^2}$ is not convergent.
This paper is organized as follows. In Section~\ref{section:conjectures} we formulate some conjectures for general sequences on the unit circle. In Section~\ref{section:firstorderasymp} we prove Theorem~\ref{theo:firstorderasymp}, and in Section~\ref{section:secondorderest} we prove Theorem~\ref{theo:secondorderest}. In Section~\ref{section:RieszenergyLeja} we give the proofs of the results in the Riesz setting.
\section{Some conjectures}\label{section:conjectures}
From the energy point of view, it is clear that Leja sequences and greedy $s$-energy sequences are special within the class of general sequences on the unit circle, as each point in the sequence is selected in an optimal way. In fact, we can also define the point $a_{n}$ in a greedy $s$-energy sequence as a point satisfying
\[
E_{s}(\{a_{0},\ldots,a_{n-1},a_{n}\})=\inf_{x\in S^{1}} E_{s}(\{a_{0},\ldots,a_{n-1},x\}),\qquad n\geq 1.
\]
Because of this property, it is reasonable to expect that greedy sequences provide the lowest upper limit for the normalized energy expressions that have been described above. We state this as a conjecture.
\begin{conjecture}
Let $(x_{n})_{n=0}^{\infty}\subset S^{1}$ be an arbitrary sequence on $S^{1}$ such that $x_{i}\neq x_{j}$ for every $i\neq j$, and let $\omega_{N}=\{x_{0},\ldots,x_{N-1}\}$, $N\geq 2$. Then
\[
\limsup_{N\rightarrow\infty}\frac{E_{0}(\omega_{N})+N\log N}{N}\geq \log(4/3);
\]
for $s\in(0,1)$,
\[
\limsup_{N\rightarrow\infty}\frac{E_{s}(\omega_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}\geq \underline{h}(s)\,\frac{2\zeta(s)}{(2\pi)^{s}};
\]
for $s>1$,
\[
\limsup_{N\rightarrow\infty}\frac{E_{s}(\omega_{N})}{N^{1+s}}\geq \overline{h}(s)\,\frac{2\zeta(s)}{(2\pi)^{s}};
\]
and for $s=1$,
\[
\limsup_{N\rightarrow\infty}\frac{E_{s}(\omega_{N})-\frac{1}{\pi}N^{2}\log N}{N^2}\geq \frac{1}{\pi}(\gamma-\log(\pi/2)+\kappa),
\]
where the expressions on the right-hand sides of the last three inequalities are the same as those appearing
in \eqref{eq:secorderasympRieszenergy:1}, \eqref{eq:firstorderasympRieszenergy:1} and \eqref{secondorderasympcritcase:1}.
\end{conjecture}
\section{First-order asymptotics in the logarithmic case}\label{section:firstorderasymp}
In this section we prove Theorem~\ref{theo:firstorderasymp}, but first we give a preliminary discussion and prove an auxiliary result.
Let $(a_{n})_{n=0}^{\infty}$ be a Leja sequence on $S^{1}$. It was shown in \cite[Lemma 4]{CalManh} that if $k=2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{t}}$ with $n_{1}>n_{2}>\cdots> n_{t}\geq 0$, then
\begin{equation}\label{eq:minprod}
\prod_{i=0}^{k-1}|a_{k}-a_{i}|=2^{t}.
\end{equation}
Let $U_{k}$ denote the discrete potential
\[
U_{k}(z):=\sum_{i=0}^{k-1}\log\frac{1}{|z-a_{i}|}.
\]
Then
\[
E_{0}(\alpha_{N})=2\sum_{0\leq i<k\leq N-1}\log\frac{1}{|a_{i}-a_{k}|}=2\sum_{k=1}^{N-1}\sum_{i=0}^{k-1}\log\frac{1}{|a_{k}-a_{i}|}=2\sum_{k=1}^{N-1}U_{k}(a_{k}).
\]
If $\tau(k)$ is the integer with the property
\begin{equation}\label{def:tau}
k=2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{\tau(k)}},\qquad n_{1}>n_{2}>\cdots> n_{\tau(k)}\geq 0,
\end{equation}
then according to \eqref{eq:minprod},
\[
U_{k}(a_{k})=-\log\, (2^{\tau(k)}),
\]
and therefore
\begin{equation}\label{energy}
E_{0}(\alpha_{N})=-2\,\log(2)\,\sum_{k=1}^{N-1}\tau(k).
\end{equation}
Note that $\tau(k)$ is the number of ones in the binary representation of $k$, so it satisfies the following properties:
\[
\tau(2^{n})=1,\qquad n\geq 0,
\]
and if $n_{1}>n_{2}>\cdots>n_{k}$, then
\begin{equation}\label{property}
\tau(2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{k}}+m)=k+\tau(m),\qquad 1\leq m\leq 2^{n_{k}}-1.
\end{equation}
Recall that the logarithmic energy of the configuration formed by $N$ equally spaced points in $S^{1}$ equals $-N\log N$. Since the configuration $\alpha_{2^{n}}$ consists of $2^{n}$ equally spaced points, we have
\begin{equation}\label{energyequalspace}
E_{0}(\alpha_{2^{n}})=-2^{n}\log(2^{n}).
\end{equation}
In particular, \eqref{energy} and \eqref{energyequalspace} give
\begin{equation}\label{sum}
n 2^{n-1}=\sum_{k=1}^{2^{n}-1}\tau(k).
\end{equation}
More generally, we have the following.
\begin{lemma}\label{lemmasum}
Assume that
\begin{equation}\label{eq:bin:N}
N=2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{t}},\qquad n_{1}>n_{2}>\cdots>n_{t}\geq 0.
\end{equation}
Then
\begin{equation}\label{sum:sk}
\sum_{k=1}^{N-1}\tau(k)=\sum_{i=1}^{t}\left(n_{i}+2(i-1)\right)\,2^{n_{i}-1}.
\end{equation}
\end{lemma}
\begin{proof}
The proof is by induction on $t$. If $t=1$ then \eqref{sum:sk} is exactly \eqref{sum}. Applying (\ref{property}) and (\ref{sum}) we obtain
\begin{align*}
\sum_{k=2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{t-1}}}^{2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{t}}-1} \tau(k) & =\tau(2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{t-1}})
+\sum_{m=1}^{2^{n_{t}}-1}\tau(2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{t-1}}+m)\\
& =t-1+\sum_{m=1}^{2^{n_{t}}-1}(\tau(m)+t-1)=(t-1)\, 2^{n_{t}}+n_{t}\, 2^{n_{t}-1}.
\end{align*}
So \eqref{sum:sk} now follows easily by induction applying the previous computations and
\[
\sum_{k=1}^{2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{t}}-1}\tau(k)
=\sum_{k=1}^{2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{t-1}}-1}\tau(k)
+\sum_{k=2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{t-1}}}^{2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{t}}-1}\tau(k).
\]
\end{proof}
\noindent \emph{Proof of Theorem~$\ref{theo:firstorderasymp}$.} From \eqref{sum:sk} and (\ref{energy}) it follows that if $N=2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{\tau(N)}}$, with $n_{1}>n_{2}>\cdots>n_{\tau(N)}\geq 0,$ then
\begin{equation}\label{energydecomp}
E_{0}(\alpha_{N})=\Lambda_{N,1}+\Lambda_{N,2},
\end{equation}
where
\begin{align*}
\Lambda_{N,1} & :=-\log(2)(n_{1} 2^{n_{1}}+n_{2} 2^{n_{2}}+\cdots+n_{\tau(N)} 2^{n_{\tau(N)}}),\\
\Lambda_{N,2} & :=-\log(2)(2^{n_{2}+1}+2\cdot 2^{n_{3}+1}+\cdots+(\tau(N)-1) 2^{n_{\tau(N)}+1}).
\end{align*}
We first justify that
\begin{equation}\label{limit1}
\lim_{N\rightarrow\infty}\frac{\Lambda_{N,2}}{N\log N}=0.
\end{equation}
Indeed, we have
\begin{gather*}
-\frac{1}{2\log(2)}\frac{\Lambda_{N,2}}{N\log N}=\frac{2^{n_{2}}+2\cdot 2^{n_{3}}+\cdots+(\tau(N)-1)2^{n_{\tau(N)}}}{(2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{\tau(N)}})\log N}\\
=\frac{2^{n_{2}-n_{1}}+2\cdot 2^{n_{3}-n_{1}}+\cdots+(\tau(N)-1)2^{n_{\tau(N)}-n_{1}}}{(1+2^{n_{2}-n_{1}}+\cdots+2^{n_{\tau(N)}-n_{1}})\log N}\\
\leq \frac{\frac{1}{2}+2\big(\frac{1}{2}\big)^{2}+\cdots+(\tau(N)-1)\big(\frac{1}{2}\big)^{\tau(N)-1}}{\log N}.
\end{gather*}
The numerator in the last expression is bounded by $\sum_{n=1}^{\infty}n 2^{-n}=2$ and (\ref{limit1}) follows.
We now show that
\begin{equation}\label{limit2}
\lim_{N\rightarrow\infty}\frac{\Lambda_{N,1}}{N\log N}=-1,
\end{equation}
hence \eqref{firstorderlimit} will follow from \eqref{energydecomp}, \eqref{limit1} and \eqref{limit2}.
We write
\begin{gather}
-\frac{\Lambda_{N,1}}{N\log N}=\frac{\log(2)(n_{1} 2^{n_{1}}+n_{2} 2^{n_{2}}+\cdots+n_{\tau(N)} 2^{n_{\tau(N)}})}{(2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{\tau(N)}})\log (2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{\tau(N)}})}\notag\\
=\frac{\log(2)(n_{1}+n_{2} 2^{n_{2}-n_{1}}+\cdots+n_{\tau(N)} 2^{n_{\tau(N)}-n_{1}})}{(1+2^{n_{2}-n_{1}}+\cdots+2^{n_{\tau(N)}-n_{1}})\log (2^{n_{1}}(1+2^{n_{2}-n_{1}}+\cdots+2^{n_{\tau(N)}-n_{1}}))}\label{expression:Lambdan1}\\
=\frac{\log(2)(n_{1}+n_{2} 2^{n_{2}-n_{1}}+\cdots+n_{\tau(N)} 2^{n_{\tau(N)}-n_{1}})}{(1+2^{n_{2}-n_{1}}+\cdots+2^{n_{\tau(N)}-n_{1}})\{n_{1}\log 2+\log (1+2^{n_{2}-n_{1}}+\cdots+2^{n_{\tau(N)}-n_{1}})\}}.\notag
\end{gather}
Since $n_{1}>n_{2}>\cdots>n_{\tau(N)}$, we have $1+2^{n_{2}-n_{1}}+\cdots+2^{n_{\tau(N)}-n_{1}}< \sum_{m=0}^{\infty}2^{-m}=2$. Therefore
\begin{equation}\label{normlimit}
\lim_{N\rightarrow\infty}\frac{(1+2^{n_{2}-n_{1}}+\cdots+2^{n_{\tau(N)}-n_{1}})\log(1+2^{n_{2}-n_{1}}+\cdots+2^{n_{\tau(N)}-n_{1}})}{ n_{1}+n_{2} 2^{n_{2}-n_{1}}+\cdots+n_{\tau(N)} 2^{n_{\tau(N)}-n_{1}}}=0,
\end{equation}
due to the fact that $n_{1}\rightarrow\infty$ as $N\rightarrow\infty$.
Now we write
\begin{gather}
\frac{n_{1}\log (2)\, (1+2^{n_{2}-n_{1}}+\cdots+2^{n_{\tau(N)}-n_{1}})}{\log (2)\, (n_{1}+n_{2} 2^{n_{2}-n_{1}}+\cdots+n_{\tau(N)} 2^{n_{\tau(N)}-n_{1}})}\notag\\
=\frac{1+2^{n_{2}-n_{1}}+\cdots+2^{n_{\tau(N)}-n_{1}}}{1+\big(\frac{n_{2}}{n_{1}}\big)\,2^{n_{2}-n_{1}}+\cdots+\big(\frac{n_{\tau(N)}}{n_{1}}\big)\,2^{n_{\tau(N)}-n_{1}}}=:\frac{c_{N}}{d_{N}}.\label{def:anbn}
\end{gather}
In order to prove that $c_{N}/d_{N}\rightarrow 1$ it suffices to show that $c_{N}-d_{N}\rightarrow 0$. We have
\[
c_{N}-d_{N}=\left(1-\frac{n_{2}}{n_{1}}\right)\,2^{n_{2}-n_{1}}+\left(1-\frac{n_{3}}{n_{1}}\right)\,2^{n_{3}-n_{1}}
+\cdots+\left(1-\frac{n_{\tau(N)}}{n_{1}}\right)\,2^{n_{\tau(N)}-n_{1}}.
\]
One can prove that this expression approaches zero applying Lebesgue's dominated convergence theorem. On $\mathbb{N}=\{1,2,\ldots\}$ we define a sequence of functions $(f_{N})_{N}$ as follows:
\[
f_{N}(m)=\begin{cases}
(1-\frac{n_{i}}{n_{1}})\, 2^{n_{i}-n_{1}} & \mbox{if}\,\,m=n_{1}-n_{i}\,\,\mbox{for some $i$, $2\leq i\leq \tau(N)$},\\
0 & \mbox{otherwise}.
\end{cases}
\]
The function $f_{N}$ is well-defined, and clearly
\[
c_{N}-d_{N}=\sum_{m=1}^{\infty} f_{N}(m).
\]
For each fixed $m$, $f_{N}(m)=0$ or $f_{N}(m)=(m/n_{1})\, 2^{-m}$. In any case, since $n_{1}\rightarrow\infty$ we have $f_{N}(m)\rightarrow 0$ as $N\rightarrow\infty$. Hence Lebesgue's theorem gives
\[
\lim_{N\rightarrow\infty} c_{N}-d_{N}=\sum_{m=1}^{\infty} 0=0.
\]
Since $c_{N}/d_{N}\rightarrow 1$, \eqref{limit2} follows from \eqref{expression:Lambdan1}, \eqref{normlimit} and \eqref{def:anbn}. \hfill $\Box$
\begin{remark}
We would like to emphasize that if $(x_{n})_{n=0}^{\infty}$ is any sequence of pairwise distinct points in $S^{1}$, then it is clear that for the sequence of configurations $\omega_{N}=\{x_{0},x_{1},\ldots,x_{N-1}\}$ we have
\begin{equation}\label{ulgeneralseq}
\limsup_{N\rightarrow\infty}\frac{E_{0}(\omega_{N})}{N\log N}\geq -1.
\end{equation}
If we have equality in \eqref{ulgeneralseq}, then
\begin{equation}\label{lgeneralseq}
\lim_{N\rightarrow\infty}\frac{E_{0}(\omega_{N})}{N\log N}=-1
\end{equation}
and the sequence $(x_{n})_{n=0}^{\infty}$ will be asymptotically uniformly distributed; that is, we have the weak-star convergence
\begin{equation}\label{weakstarconvunif}
\frac{1}{N}\sum_{x\in\omega_{N}}\delta_{x}\xrightarrow[N\rightarrow\infty]{*}\sigma,
\end{equation}
where $\sigma$ denotes the normalized arc length measure on $S^{1}$. Indeed, if \eqref{lgeneralseq} holds, then
\[
\lim_{N\rightarrow\infty}\frac{E_{0}(\omega_{N})}{N^{2}}=0=\iint\log\frac{1}{|x-y|}\,d\sigma(x)\,d\sigma(y),
\]
and this implies \eqref{weakstarconvunif} by a standard argument in potential theory, see \cite{Ran}.
\end{remark}
\section{Second-order estimates in the logarithmic case}\label{section:secondorderest}
\noindent\emph{Proof of Theorem~$\ref{theo:secondorderest}$.} The inequality on the left-hand side of \eqref{secondorderest} is obvious. If $N$ has the binary representation \eqref{eq:bin:N}, then in virtue of \eqref{energy} and \eqref{sum:sk} we have
\begin{equation}\label{exp:modenergy}
\frac{E_{0}(\alpha_{N})+N\log N}{N}=-\log (2)\,\frac{\sum_{i=1}^{t}(n_{i}+2i-2)\,2^{n_{i}}}{\sum_{i=1}^{t}2^{n_{i}}}+\log\left(\sum_{i=1}^{t}2^{n_{i}}\right),
\end{equation}
hence the inequality on the right-hand side of \eqref{secondorderest} is the same as
\[
-\log (2)\,\frac{\sum_{i=1}^{t}(n_{i}+2i-2)\,2^{n_{i}}}{\sum_{i=1}^{t}2^{n_{i}}}+\log\left(\sum_{i=1}^{t}2^{n_{i}}\right)<2\log 2-\log 3.
\]
Simplifying we obtain that this is equivalent to
\[
\log\left(3\sum_{i=1}^{t}2^{n_{i}}\right)<\log(2)\,\frac{\sum_{i=1}^{t}(n_{i}+2i) 2^{n_{i}}}{\sum_{i=1}^{t}2^{n_{i}}}.
\]
Hence we want to show that
\begin{equation}\label{ineqsecondorder}
3\sum_{i=1}^{t} 2^{n_{i}}<2^{c_{N}},\qquad c_{N}=\frac{\sum_{i=1}^{t}(n_{i}+2i)\, 2^{n_{i}}}{\sum_{i=1}^{t}2^{n_{i}}}.
\end{equation}
In order to prove \eqref{ineqsecondorder} we apply the following inequality, which can be found in \cite[page 78]{HLP}: For any collection of positive numbers $(b_{i})_{i=1}^{t}$ and $(p_{i})_{i=1}^{t}$ we have
\begin{equation}\label{ineqHLP}
\frac{\sum_{i=1}^{t}p_{i}\,b_{i}}{\sum_{i=1}^{t}p_{i}}\leq \exp\left(\frac{\sum_{i=1}^{t}p_{i}\,b_{i}\,\log(b_{i})}{\sum_{i=1}^{t}p_{i}\,b_{i}}\right),
\end{equation}
with equality only if all the $b$'s are equal. The inequality \eqref{ineqHLP} is obtained applying Jensen's inequality to the convex function $x\log x$. Taking in \eqref{ineqHLP} the values
\[
p_{i}=2^{-2i},\qquad b_{i}=2^{n_{i}+2i},\qquad i=1,\ldots,t,
\]
we obtain after simplification the expression
\[
\frac{\sum_{i=1}^{t}2^{n_{i}}}{\sum_{i=1}^{t} 4^{-i}}<2^{c_{N}},
\]
which gives \eqref{ineqsecondorder}.
In order to prove \eqref{ulsecondorder}, it suffices now to show that for the subsequence
\begin{equation}\label{choiceN}
N=N(k)=\sum_{j=0}^{k} 4^{j}=\frac{4^{k+1}-1}{3}
\end{equation}
one gets
\[
\lim_{k\rightarrow\infty}\frac{E_{0}(\alpha_{N})+N\log N}{N}=\log\left(\frac{4}{3}\right).
\]
For this it is convenient to rewrite \eqref{exp:modenergy} as
\begin{equation}\label{modenergyrewrite}
\frac{E_{0}(\alpha_{N})+N\log N}{N}=\log(2)\,\frac{\sum_{i=2}^{t}(n_{1}-n_{i}+2-2i)\,2^{n_{i}-n_{1}}}{\sum_{i=1}^{t}2^{n_{i}-n_{1}}}+\log\left(\sum_{i=1}^{t}2^{n_{i}-n_{1}}\right).
\end{equation}
For the choice \eqref{choiceN} of $N$ we have $n_{1}-n_{i}=2i-2$, hence the first term vanishes, while the second term approaches $\log(4/3)$.\hfill $\Box$
\bigskip
An interesting property of the sequence analyzed in Theorem \ref{theo:secondorderest} is the fact that
\[
\frac{E_{0}(\alpha_{N})+N\log N}{N}=\frac{E_{0}(\alpha_{2N})+2N\log (2N)}{2N},\qquad \mbox{for all}\,\,N\geq 1,
\]
which can be easily checked using \eqref{modenergyrewrite}. This property explains the ``periodic" behavior of the sequence $\left(\frac{E_{0}(\alpha_{N})+N\log N}{N}\right)_{N}$ that can be observed in Fig.~\ref{fig:energylog} below, with increasing ``periods" of length $2^{n}$.
\begin{figure}[h]
\centering
\includegraphics[totalheight=3in,keepaspectratio]{sequencelogfirst4200.eps}
\caption{This figure shows the first $4200$ points of the sequence $\left(E_{0}(\alpha_{N})+N\log N)/N\right)_{N}$.}
\label{fig:energylog}
\end{figure}
\section{Proofs of results in the Riesz setting}\label{section:RieszenergyLeja}
We begin with a formula that expresses the Riesz $s$-energy of the first $N$ points $\alpha_{N}$ of a greedy $s$-energy sequence on the unit circle in terms of the binary representation of $N$.
\begin{proposition}\label{prop:Reiszenergybinary}
Let $(a_{n})_{n=0}^{\infty}$ be a greedy $s$-energy sequence on $S^{1}$, and let $\alpha_{N}=(a_{n})_{n=0}^{N-1}$. Assume that $N$ has the binary representation \eqref{eq:bin:N}. Then
\begin{equation}\label{Rieszenergybinary}
E_{s}(\alpha_{N})=\sum_{k=1}^{t-1}\left(\sum_{j=k+1}^{t} 2^{n_{j}-n_{k}}\right)\mathcal{L}_{s}(2^{n_{k}+1})+\sum_{k=1}^{t}\left(1-\sum_{j=k+1}^{t} 2^{n_{j}-n_{k}+1}\right)\mathcal{L}_{s}(2^{n_{k}}),
\end{equation}
understanding $\sum_{t+1}^{t}$ as empty sum.
\end{proposition}
\begin{proof}
The proof of \eqref{Rieszenergybinary} is obtained from a repeated application of the following simple property. If $A$ and $B$ are two finite sets of points on the unit circle with $A\cap B=\emptyset$, then
\begin{equation}\label{Rieszdecompformula}
E_{s}(A\cup B)=E_{s}(A)+E_{s}(B)+2\sum_{y\in B}\sum_{x\in A}|x-y|^{-s}.
\end{equation}
Let $(a_{n})_{n=0}^\infty$ be a greedy $s$-energy sequence on $S^{1}$. Recall that this sequence also has the structure of a Leja sequence, and hence the properties $1)$--$3)$ described in Sec.~\ref{section:intro} are also applicable for this sequence. Let $A_{1}$ denote the set formed by the first $2^{n_{1}}$ points of the sequence and $B_{1}$ denote the next $N-2^{n_{1}}=2^{n_{2}}+\cdots+2^{n_{t}}$ points of the sequence, i.e.,
\[
A_{1}:=(a_{n})_{n=0}^{2^{n_{1}-1}},\qquad B_{1}:=(a_{n})_{n=2^{n_{1}}}^{N-1}.
\]
Since the points in $A_{1}$ are equally spaced, we have $E_{s}(A_{1})=\mathcal{L}_{s}(2^{n_{1}})$. Any $y\in B_{1}$ lies in the midpoint of one of the $2^{n_{1}}$ arcs determined by the points of $A_{1}$. So clearly $\sum_{x\in A_{1}}|x-y|^{-s}$ is independent of $y$, and we can write this expression as the difference
\begin{align*}
\sum_{x\in A_{1}}|x-y|^{-s}= & 2^{-s}\sum_{j=1}^{2^{n_{1}+1}-1}\left(\sin \frac{\pi j}{2^{n_{1}+1}}\right)^{-s}-2^{-s}\sum_{j=1}^{2^{n_{1}}-1}\left(\sin \frac{\pi j}{2^{n_{1}}}\right)^{-s}\\
= & \frac{1}{2^{n_{1}+1}}\,\mathcal{L}_{s}(2^{n_{1}+1})-\frac{1}{2^{n_{1}}}\,\mathcal{L}_{s}(2^{n_{1}}).
\end{align*}
We conclude from \eqref{Rieszdecompformula} and the computation above that
\[
E_{s}(\alpha_{N})=\left(\sum_{j=2}^{t} 2^{n_{j}-n_{1}}\right)\mathcal{L}_{s}(2^{n_{1}+1})+\left(1-\sum_{j=2}^{t}2^{n_{j}-n_{1}+1}\right)\mathcal{L}_{s}(2^{n_{1}})+E_{s}(B_{1}).
\]
Now we can apply this argument to the set $B_{1}$, since this set itself has the structure of the first $N-2^{n_{1}}$ points of a greedy sequence, see \cite[Theorem 5]{BiaCal} and \cite[Lemma 4.2]{LopSaff}. In particular, if we make the partition $B_{1}=A_{2}\cup B_{2}$, where $A_{2}$ is the set formed by the first $2^{n_{2}}$ points in $B_{1}$ and $B_{2}$ is the set formed by the remaining $N-2^{n_{1}}-2^{n_{2}}=2^{n_{3}}+\cdots+2^{n_{t}}$ points in $B_{1}$, then again we have that $A_{2}$ is formed by equally spaced points and any point of $B_{2}$ lies in the midpoint of one of the $2^{n_{2}}$ arcs determined by the points of $A_{2}$. Hence as before we get
\[
E_{s}(B_{1})=\left(\sum_{j=3}^{t} 2^{n_{j}-n_{2}}\right)\mathcal{L}_{s}(2^{n_{2}+1})+\left(1-\sum_{j=3}^{t}2^{n_{j}-n_{2}+1}\right)\mathcal{L}_{s}(2^{n_{2}})+E_{s}(B_{2}),
\]
and so
\[
E_{s}(\alpha_{N})=\sum_{k=1}^{2}\left(\sum_{j=k+1}^{t} 2^{n_{j}-n_{k}}\right)\mathcal{L}_{s}(2^{n_{k}+1})+\sum_{k=1}^{2}\left(1-\sum_{j=k+1}^{t} 2^{n_{j}-n_{k}+1}\right)\mathcal{L}_{s}(2^{n_{k}})+E_{s}(B_{2}).
\]
Applying this argument repeatedly it is clear that we arrive at \eqref{Rieszenergybinary}.
\end{proof}
Before giving the proofs of the results in the Riesz setting, we make some remarks concerning the sets $\Theta_{p}$ and the functions $H$ defined in \eqref{def:H}.
The reader can easily check that an alternative way to define the set $\Theta_{p}$ is the following. This set consists of all vectors $\vec{\theta}=(\theta_{1},\ldots,\theta_{p})$ that can be written in the form
\begin{equation}\label{def:vectorsinThetap}
\vec{\theta}=\left(\frac{2^{t_{1}}}{M}, \frac{2^{t_{2}}}{M},\ldots, \frac{2^{t_{r-1}}}{M}, \frac{1}{M}, 0,\ldots, 0\right),
\end{equation}
where $M=2^{t_{1}}+2^{t_{2}}+\cdots+2^{t_{r-1}}+1$ is an odd integer with $t_{1}>t_{2}>\cdots>t_{r-1}>0$ and $1\leq r\leq p$. The number of zeros in \eqref{def:vectorsinThetap} is then $p-r$, if they appear. In particular we see that the set $\Theta_{p}$ can be regarded as a subset of $\Theta_{p+1}$, for all $p$. We preferred to give the Definition~\ref{def:Thetap} for $\Theta_{p}$ instead of the one described here since we are only going to make use of the limiting property \eqref{eq:conddefthetavec}.
It follows from \eqref{def:H} that if $\vec{\theta}=(\theta_{1},\ldots,\theta_{p})$ satisfies the condition $\theta_{k}\geq 2 \sum_{j=k+1}^{p}\theta_{j}$ for all $k=1,\ldots,p-1$, then $H(\vec{\theta};s)$ is convex as a function of $s$ since in this case we can write it as a positive linear combination of convex functions.
\subsection{Second-order asymptotics in the Riesz case for $0<s<1$}
Below we will make use of a fortunate relation between the coefficients appearing in \eqref{Rieszenergybinary}, the arguments of $\mathcal{L}_{s}$ in this formula, and $N^{2}$. The reader can easily check that for $N$ as in \eqref{eq:bin:N} we have
\begin{equation}\label{expansionN2}
N^{2}=\sum_{k=1}^{t-1}\left(\sum_{j=k+1}^{t} 2^{n_{j}-n_{k}}\right)\,2^{2(n_{k}+1)}+\sum_{k=1}^{t}\left(1-\sum_{j=k+1}^{t} 2^{n_{j}-n_{k}+1}\right)\,2^{2 n_{k}}.
\end{equation}
So if we introduce the notation
\begin{equation}\label{def:Rs}
\mathcal{R}_{s}(N):=\frac{\mathcal{L}_{s}(N)-I_{s}(\sigma) N^{2}}{N^{1+s}},\qquad 0<s<1,
\end{equation}
it follows from \eqref{Rieszenergybinary}and \eqref{expansionN2} that
\begin{align}
\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}= & \sum_{k=1}^{t-1}\left(\frac{2^{n_{k}+1}}{N}\right)^{1+s}\left(\sum_{j=k+1}^{t}2^{n_{j}-n_{k}}\right)\mathcal{R}_{s}(2^{n_{k}+1})\notag\\
+ & \sum_{k=1}^{t}\left(\frac{2^{n_{k}}}{N}\right)^{1+s}\left(1-\sum_{j=k+1}^{t}2^{n_{j}-n_{k}+1}\right)\mathcal{R}_{s}(2^{n_{k}}).\label{eq:secorderRieszenergy}
\end{align}
See an illustration of the sequence $\left(\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}\right)_{N}$ in Fig.~\ref{fig:energyonehalf} below.
\begin{figure}[h]
\centering
\includegraphics[totalheight=3in,keepaspectratio]{sequencesonehalffirst4200.eps}
\caption{This is a plot of the first $4200$ points of the sequence $\left((E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2})/N^{1+s}\right)_{N}$ in the case $s=1/2$.}
\label{fig:energyonehalf}
\end{figure}
\bigskip
\noindent\emph{Proof of Theorem~$\ref{theo:secorderasympRieszenergy:1}$.}
We first prove the inequality ``$\geq$" in \eqref{eq:secorderasympRieszenergy:1}, which is straightforward. Let $p\in\mathbb{N}$ be arbitrary and fix a vector $\vec{\theta}=(\theta_{1},\ldots,\theta_{p})\in\Theta_{p}$. By Definition~\ref{def:Thetap}, there exists an infinite sequence $\mathcal{N}$ of integers of the form $N=2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{p}}$, $n_{1}>n_{2}>\cdots>n_{p}\geq 0$ such that \eqref{eq:conddefthetavec} holds. Applying \eqref{eq:secorderRieszenergy} we have
\begin{gather*}
\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}\\
=\sum_{k=1}^{p-1}\left(\frac{2^{n_{k}+1}}{N}\right)^{1+s}\left(\sum_{j=k+1}^{p}2^{n_{j}-n_{k}}\right)\mathcal{R}_{s}(2^{n_{k}+1})+\sum_{k=1}^{p}\left(\frac{2^{n_{k}}}{N}\right)^{1+s}\left(1-\sum_{j=k+1}^{p}2^{n_{j}-n_{k}+1}\right)\mathcal{R}_{s}(2^{n_{k}})\\
=\sum_{k=1}^{p-1}\left(\frac{2^{n_{k}+1}}{N}\right)^{s}\left(\sum_{j=k+1}^{p}\frac{2^{n_{j}+1}}{N}\right)\mathcal{R}_{s}(2^{n_{k}+1})+\sum_{k=1}^{p}\left(\frac{2^{n_{k}}}{N}\right)^{s}\left(\frac{2^{n_{k}}}{N}-\sum_{j=k+1}^{p}\frac{2^{n_{j}+1}}{N}\right)\mathcal{R}_{s}(2^{n_{k}}).
\end{gather*}
Using now \eqref{eq:conddefthetavec}, \eqref{def:Rs} and \eqref{eq:secorderasympRieszFekete}, we get
\begin{equation}\label{eq:limitsubseq}
\lim_{N\in\mathcal{N}}\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}
=\left(\sum_{k=1}^{p-1}(2\theta_{k})^{s} \sum_{j=k+1}^{p}2\theta_{j}+\sum_{k=1}^{p}\theta_{k}^{s}\,\Large(\theta_{k}-\sum_{j=k+1}^{p}2\theta_{j}\Large)\right)\frac{2\zeta(s)}{(2\pi)^{s}}.
\end{equation}
Here we have taken into account that if for some particular $k=1,\ldots,p$, the sequence $2^{n_{k}}$ does not approach infinity, then $2^{n_{k}}/N$ approaches $\theta_{k}=0$ and therefore we still have $(2^{n_{k}}/N)^{s}\,\mathcal{R}_{s}(2^{n_{k}})\rightarrow \theta_{k}^{s}\, 2\zeta(s)/(2\pi)^{s}=0$. The first factor on the right-hand side of \eqref{eq:limitsubseq} is exactly $H(\vec{\theta};s)$, and therefore
\[
\limsup_{N\rightarrow\infty}\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}\geq H(\vec{\theta};s)\,\frac{2\zeta(s)}{(2\pi)^{s}}.
\]
Since $p$ and $\vec{\theta}$ were arbitrary, we obtain the desired inequality.
Now we prove the converse inequality in \eqref{eq:secorderasympRieszenergy:1}. Let $\mathcal{N}\subset\mathbb{N}$ be an infinite sequence for which the sequence $\left(\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}\right)_{N\in\mathcal{N}}$ converges, and we shall show that
\begin{equation}\label{eq:ineqtobeproved}
\lim_{N\in\mathcal{N}}\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}\leq \underline{h}(s)\,\frac{2\zeta(s)}{(2\pi)^{s}}.
\end{equation}
Assume first that there exists $p\geq 1$ such that an infinite number of integers $N\in\mathcal{N}$ satisfy the property $\tau(N)=p$, cf. \eqref{def:tau}. Then, taking a subsequence $\widetilde{\mathcal{N}}$ of $\mathcal{N}$ if necessary, such that the integers $N=2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{p}}\in\widetilde{\mathcal{N}}$ satisfy
\[
\lim_{N\in\widetilde{N}}\frac{2^{n_{i}}}{N}=\theta_{i},\qquad \mbox{for all}\,\,i=1,\ldots,p,
\]
we get
\[
\lim_{N\in\mathcal{N}}\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}=\lim_{N\in\widetilde{\mathcal{N}}}\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}=H((\theta_{1},\ldots,\theta_{p});s)\,\frac{2\zeta(s)}{(2\pi)^{s}},
\]
and therefore \eqref{eq:ineqtobeproved} holds.
So let us assume now that such an integer $p$ does not exist. This means that we assume now that $\tau(N)\rightarrow\infty$ as $N\rightarrow\infty$ in the sequence $\mathcal{N}$. Let us rewrite, for $N=2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{\tau(N)}}\in\mathcal{N}$, $n_{1}>n_{2}>\cdots>n_{\tau(N)}\geq 0$, the expression
\begin{gather}
\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}\notag\\
=\sum_{k=1}^{\tau(N)}\left(\frac{2^{n_{k}}}{N}\right)^{1+s}\left[2^{s+1}\,\mathcal{R}_{s}(2^{n_{k}+1})\,\sum_{j=k+1}^{\tau(N)}2^{n_{j}-n_{k}}+\mathcal{R}_{s}(2^{n_{k}})\left(1-\sum_{j=k+1}^{\tau(N)} 2^{n_{j}-n_{k}+1}\right)\right].\label{eq:rewritingsecorderenergy}
\end{gather}
and let us introduce the notation
\begin{equation}\label{def:lambdaNk}
\lambda_{N,k}:=2^{s+1}\,\mathcal{R}_{s}(2^{n_{k}+1})\,\sum_{j=k+1}^{\tau(N)}2^{n_{j}-n_{k}}+\mathcal{R}_{s}(2^{n_{k}})\left(1-\sum_{j=k+1}^{\tau(N)} 2^{n_{j}-n_{k}+1}\right).
\end{equation}
Since the sequence $(\mathcal{R}_{s}(N))_{N}$ is bounded, it is evident that there exists an absolute constant $C_{1}>0$ independent of $N$, such that
\begin{equation}\label{bound1}
\left|\lambda_{N,k}\right|\leq C_{1},\qquad \mbox{for all}\,\,N\in\mathcal{N}\,\,\mbox{and}\,\,k=1,\ldots,\tau(N).
\end{equation}
On the other hand, we have the following simple estimate for each $N=2^{n_{1}}+\cdots+2^{n_{\tau(N)}}\in\mathcal{N}$,
\begin{equation}\label{bound2}
\frac{2^{n_{k}}}{N}\leq \frac{2^{n_{k}}}{2^{n_{1}}}=2^{n_{k}-n_{1}}\leq 2^{-(k-1)},\qquad k=1,\ldots,\tau(N).
\end{equation}
Now let $0<\epsilon<1$ be fixed. It follows from \eqref{bound2} that there exists $M=M(\epsilon)\in\mathbb{N}$ independent of $N$ such that
\begin{equation}\label{bound4}
\sum_{k=M+1}^{\tau(N)}\frac{2^{n_{k}}}{N}<\epsilon,\qquad \mbox{for all}\,\,N\in\mathcal{N}.
\end{equation}
hence \eqref{bound1} and \eqref{bound4} give
\begin{equation}\label{bound3}
\sum_{k=M+1}^{\tau(N)}\left(\frac{2^{n_{k}}}{N}\right)^{1+s}|\lambda_{N,k}|<C_{1}\epsilon,\qquad \mbox{for all}\,\,N\in\mathcal{N}.
\end{equation}
Applying \eqref{eq:rewritingsecorderenergy} and \eqref{def:lambdaNk} we can write
\begin{equation}\label{energydecomp:1}
\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}=S_{N,M,1}+S_{N,M,2},
\end{equation}
where
\begin{equation}\label{def:SNM1SNM2}
S_{N,M,1}:=\sum_{k=1}^{M}\left(\frac{2^{n_{k}}}{N}\right)^{1+s}\lambda_{N,k},\qquad
S_{N,M,2}:=\sum_{k=M+1}^{\tau(N)}\left(\frac{2^{n_{k}}}{N}\right)^{1+s}\lambda_{N,k},
\end{equation}
hence by \eqref{bound3} we have
\begin{equation}\label{bound5}
|S_{N,M,2}|\leq C_{1}\epsilon,\qquad \mbox{for all}\,\,N\in\mathcal{N}.
\end{equation}
Now we focus on the sum $S_{N,M,1}$. First we rewrite $\lambda_{N,k}$ in the form
\begin{gather}
\lambda_{N,k}=\mathcal{R}_{s}(2^{n_{k}})+2^{-n_{k}+1}\,(2^{s}\,\mathcal{R}_{s}(2^{n_{k}+1})-\mathcal{R}_{s}(2^{n_{k}}))\,\sum_{j=k+1}^{\tau(N)}2^{n_{j}}\label{def:lambdaNk:2}\\
=\mathcal{R}_{s}(2^{n_{k}})+2^{-n_{k}+1}\,(2^{s}\,\mathcal{R}_{s}(2^{n_{k}+1})-\mathcal{R}_{s}(2^{n_{k}}))\,\left(\sum_{j=k+1}^{M}2^{n_{j}}+\sum_{j=M+1}^{\tau(N)}2^{n_{j}}\right).\notag
\end{gather}
This shows that we can write
\begin{equation}\label{energydecomp:2}
S_{N,M,1}=D_{N,M,1}+D_{N,M,2},
\end{equation}
where
\begin{align}
D_{N,M,1} & :=\sum_{k=1}^{M}\left(\frac{2^{n_{k}}}{N}\right)^{1+s}\left[\mathcal{R}_{s}(2^{n_{k}})+2^{-n_{k}+1}\,(2^{s}\,\mathcal{R}_{s}(2^{n_{k}+1})-\mathcal{R}_{s}(2^{n_{k}}))\,\sum_{j=k+1}^{M}2^{n_{j}}\right],\label{def:DNM1}\\
D_{N,M,2} & :=\left(\sum_{j=M+1}^{\tau(N)} 2^{n_{j}}\right) \sum_{k=1}^{M}\left(\frac{2^{n_{k}}}{N}\right)^{1+s}
2^{-n_{k}+1}\,(2^{s}\,\mathcal{R}_{s}(2^{n_{k}+1})-\mathcal{R}_{s}(2^{n_{k}})). \label{def:DNM2}
\end{align}
Let's first estimate the sum \eqref{def:DNM2}. We have
\begin{gather*}
D_{N,M,2}=\left(\sum_{j=M+1}^{\tau(N)}\frac{2^{n_{j}}}{N}\right)\sum_{k=1}^{M}\left(\frac{2^{n_{k}}}{N}\right)^{s} (2^{s+1}\,\mathcal{R}_{s}(2^{n_{k}+1})-2\,\mathcal{R}_{s}(2^{n_{k}})).
\end{gather*}
Using \eqref{bound2} and the boundedness of the sequence $(\mathcal{R}_{s}(N))_{N}$, we find that there exists an absolute constant $C_{2}>0$ such that
\[
\left|\sum_{k=1}^{M}\left(\frac{2^{n_{k}}}{N}\right)^{s} (2^{s+1}\,\mathcal{R}_{s}(2^{n_{k}+1})-2\,\mathcal{R}_{s}(2^{n_{k}}))\right|<C_{2},\qquad \mbox{for all}\,\,N\in\mathcal{N}.
\]
This estimate and \eqref{bound4} shows that
\begin{equation}\label{bound6}
|D_{N,M,2}|<C_{2}\epsilon,\qquad \mbox{for all}\,\,N\in\mathcal{N}.
\end{equation}
Finally we analyze the sum \eqref{def:DNM1}. Introducing the notation
\[
\widetilde{\lambda}_{N,k}:=\mathcal{R}_{s}(2^{n_{k}})+2^{-n_{k}+1}\,(2^{s}\,\mathcal{R}_{s}(2^{n_{k}+1})-\mathcal{R}_{s}(2^{n_{k}}))\,\sum_{j=k+1}^{M}2^{n_{j}},
\]
we can write
\begin{equation}\label{energydecomp:3}
D_{N,M,1}=\sum_{k=1}^{M}\left(\frac{2^{n_{k}}}{N}\right)^{1+s} \widetilde{\lambda}_{N,k}=E_{N,M,1}+E_{N,M,2},
\end{equation}
where
\begin{align*}
E_{N,M,1} & :=\sum_{k=1}^{M}\left(\frac{2^{n_{k}}}{2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{M}}}\right)^{1+s}\widetilde{\lambda}_{N,k},\\
E_{N,M,2} & :=\left(\left(\frac{2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{M}}}{N}\right)^{1+s}-1\right) \sum_{k=1}^{M}\left(\frac{2^{n_{k}}}{2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{M}}}\right)^{1+s} \widetilde{\lambda}_{N,k}.
\end{align*}
Again the numbers $\widetilde{\lambda}_{N,k}$ are uniformly bounded and we have
\begin{gather*}
\sum_{k=1}^{M}\left(\frac{2^{n_{k}}}{2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{M}}}\right)^{1+s}\leq \sum_{k=1}^{M}\frac{2^{n_{k}}}{2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{M}}}=1,\\
\left|\left(\frac{2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{M}}}{N}\right)^{1+s}-1\right|\leq 1-(1-\epsilon)^{1+s},
\end{gather*}
where in the latter inequality we used \eqref{bound4}. We conclude that
\begin{equation}\label{bound7}
|E_{N,M,2}|\leq C_{3}\, (1-(1-\epsilon)^{1+s}),
\end{equation}
for some constant $C_{3}>0$.
Note that the expression $E_{N,M,1}$ is exactly as in \eqref{eq:rewritingsecorderenergy} but with $N$ replaced by $2^{n_{1}}+\cdots+2^{n_{M}}$ and $\tau(N)$ replaced by $M$. Therefore, as before we can find a subsequence $\widetilde{\mathcal{N}}$ of $\mathcal{N}$ such that
\begin{equation}\label{cond:subseqNtilde}
\lim_{N\in\widetilde{\mathcal{N}}}\frac{2^{n_{i}}}{2^{n_{1}}+\cdots+2^{n_{M}}}=\theta_{i},\qquad \mbox{for all}\,\,i=1,\ldots,M,
\end{equation}
and consequently
\begin{equation}\label{finallimit}
\lim_{N\in\widetilde{\mathcal{N}}} E_{N,M,1}=H((\theta_{1},\ldots,\theta_{M});s)\,\frac{2\zeta(s)}{(2\pi)^{s}}.
\end{equation}
Applying now the relations \eqref{energydecomp:1}, \eqref{energydecomp:2}, \eqref{energydecomp:3} and the bounds \eqref{bound5}, \eqref{bound6}, \eqref{bound7} and \eqref{finallimit}, we conclude that
\begin{gather*}
\lim_{N\in\mathcal{N}}\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}
=\lim_{N\in\widetilde{\mathcal{N}}}\frac{E_{s}(\alpha_{N})-I_{s}(\sigma) N^{2}}{N^{1+s}}=\lim_{N\in\widetilde{\mathcal{N}}}\left(E_{N,M,1}+E_{N,M,2}+D_{N,M,2}+S_{N,M,2}\right)\\
\leq H((\theta_{1},\ldots,\theta_{M});s)\,\frac{2\zeta(s)}{(2\pi)^{s}}
+C_{3}(1-(1-\epsilon)^{1+s})+C_{2}\epsilon+C_{1}\epsilon\\
\leq \underline{h}(s)\,\frac{2\zeta(s)}{(2\pi)^{s}}+C_{3}(1-(1-\epsilon)^{1+s})+C_{2}\epsilon+C_{1}\epsilon.
\end{gather*}
This inequality holds for an arbitrary $\epsilon>0$, so we obtain \eqref{eq:ineqtobeproved}. This finishes the proof of \eqref{eq:secorderasympRieszenergy:1}.
The asymptotic formula \eqref{secondorderasympRieszenergy:2} follows from the inequality $E_{s}(\alpha_{N})\geq \mathcal{L}_{s}(N)$, which is valid for every $N$ and is an equality for all $N$ of the form $N=2^{n}$, and the asymptotic formula \eqref{eq:secorderasympRieszFekete}.\hfill$\Box$
\subsection{First-order asymptotics in the Riesz case for $s>1$}
\noindent\emph{Proof of Theorem~$\ref{theo:firstorderasympRieszenergy}$.} The proofs of \eqref{eq:firstorderasympRieszenergy:1} and \eqref{eq:firstorderasympRieszenergy:2} are identical to the proofs of the corresponding formulas in Theorem~\ref{theo:secorderasympRieszenergy:1}. The reader only needs to use, instead of \eqref{eq:secorderRieszenergy}, the formula
\begin{align*}
\frac{E_{s}(\alpha_{N})}{N^{1+s}}= &
\sum_{k=1}^{t-1}\left(\frac{2^{n_{k}+1}}{N}\right)^{1+s}\left(\sum_{j=k+1}^{t}2^{n_{j}-n_{k}}\right)\frac{\mathcal{L}_{s}(2^{n_{k}+1})}{(2^{n_{k}+1})^{1+s}}\\
+ & \sum_{k=1}^{t}\left(\frac{2^{n_{k}}}{N}\right)^{1+s}\left(1-\sum_{j=k+1}^{t}2^{n_{j}-n_{k}+1}\right)\frac{\mathcal{L}_{s}(2^{n_{k}})}{(2^{n_{k}})^{1+s}},
\end{align*}
which follows from \eqref{Rieszenergybinary}, and use \eqref{eq:asympfirstorderRieszsupercrit} instead of \eqref{eq:secorderasympRieszFekete}.\hfill$\Box$
\bigskip
Figure~\ref{fig:energysonehalf} below displays the first $4200$ points of the sequence $\left(E_{s}(\alpha_{N})/N^{1+s}\right)_{N}$ in the case $s=2$.
\begin{figure}[h]
\centering
\includegraphics[totalheight=3in,keepaspectratio]{sequences2first4200.eps}
\caption{The first $4200$ points of the sequence $\left(E_{s}(\alpha_{N})/N^{1+s}\right)_{N}$ in the case $s=2$.}
\label{fig:energysonehalf}
\end{figure}
\subsection{Second-order asymptotics in the Riesz case for $s=1$}
\noindent\emph{Proof of Theorem~$\ref{theo:criticalcase}$.} Below we will use the notation
\begin{equation}\label{def:R1}
\mathcal{R}_{1}(N):=\frac{\mathcal{L}_{1}(N)-\frac{1}{\pi}N^{2}\log N}{N^2}.
\end{equation}
If $N=2^{n_{1}}+2^{n_{2}}+\cdots+2^{n_{p}}$ in decreasing order of powers, applying \eqref{expansionN2} we can write conveniently
\begin{align*}
N^{2} \log N & =\sum_{k=1}^{p-1}\left(\sum_{j=k+1}^{p}2^{n_{j}-n_{k}}\right) \left(2^{2(n_{k}+1)} \log\left(2^{n_{k}+1}\right)+2^{2(n_{k}+1)} \log\left(\frac{N}{2^{n_{k}+1}}\right)\right)\\
& +\sum_{k=1}^{p}\left(1-\sum_{j=k+1}^{p} 2^{n_{j}-n_{k}+1}\right)\left(2^{2n_{k}}\log \left(2^{2n_{k}}\right)+2^{2n_{k}} \log\left(\frac{N}{2^{2n_{k}}}\right)\right).
\end{align*}
Hence, applying \eqref{Rieszenergybinary} for $s=1$ and \eqref{def:R1} we obtain
\begin{align}\label{eq:EnergyNcritcase}
\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}N^2\log N}{N^{2}} & =\frac{1}{N^2}\sum_{k=1}^{p-1}\left(\sum_{j=k+1}^{p}2^{n_{j}+n_{k}+2}\right)
\left(\mathcal{R}_{1}\left(2^{n_{k}+1}\right)+\frac{1}{\pi}\log\left(\frac{2^{n_{k}+1}}{N}\right)\right)\notag\\
& +\frac{1}{N^2}\sum_{k=1}^{p}\left(2^{2n_{k}}-\sum_{j=k+1}^{p}2^{n_{j}+n_{k}+1}\right)\left(\mathcal{R}_{1}(2^{n_{k}})+\frac{1}{\pi}\log\left(\frac{2^{n_{k}}}{N}\right)\right).
\end{align}
The proof of \eqref{secondorderasympcritcase:1} follows the same guidelines of the proof of \eqref{eq:secorderasympRieszenergy:1}. To prove the inequality ``$\geq$" in \eqref{secondorderasympcritcase:1}, we take an arbitrary $\vec{\theta}=(\theta_{1},\ldots,\theta_{p})\in\Theta_{p}$, and we let $\mathcal{N}$ be a sequence of integers $N=2^{n_{1}}+\cdots+2^{n_{p}}$ as in Definition~\ref{def:Thetap} satisfying \eqref{eq:conddefthetavec}. If we call $L=\frac{1}{\pi}(\gamma-\log (\pi/2))$ and apply \eqref{eq:conddefthetavec} and \eqref{eq:secorderFeketecrit}, it follows from \eqref{eq:EnergyNcritcase} that
\begin{gather}
\lim_{N\in\mathcal{N}}\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}N^{2} \log N}{N^2}\label{limitfinitetau}\\
=\sum_{k=1}^{p-1}\left(\sum_{j=k+1}^{p}4 \theta_{k} \theta_{j}\right)(L+\frac{1}{\pi}\log(2\theta_{k}))+\sum_{k=1}^{p}(\theta_{k}^{2}-\sum_{j=k+1}^{p}2 \theta_{k} \theta_{j})(L+\frac{1}{\pi}\log \theta_{k})\notag\\
=L\left(\sum_{k=1}^{p}\theta_{k}^{2}+2\sum_{k=1}^{p-1}\sum_{j=k+1}^{p}\theta_{k} \theta_{j}\right)
+\frac{4}{\pi}\sum_{k=1}^{p-1}\left(\sum_{j=k+1}^{p}\theta_{j}\right)\theta_{k} \log(2\theta_{k})
+\frac{1}{\pi}\sum_{k=1}^{p}(\theta_{k}^2-\sum_{j=k+1}^{p}2\theta_{k}\theta_{j})\log \theta_{k}\notag\\
=L+\frac{1}{\pi}\,K(\theta_{1},\ldots,\theta_{p})\leq L+\frac{\kappa}{\pi},\notag
\end{gather}
where we used the fact that
\[
\sum_{k=1}^{p}\theta_{k}^{2}+2\sum_{k=1}^{p-1}\sum_{j=k+1}^{p}\theta_{k} \theta_{j}=(\theta_{1}+\cdots+\theta_{p})^2=1.
\]
This proves the desired inequality.
The proof of the converse inequality in \eqref{secondorderasympcritcase:1} is similar to the one given for \eqref{eq:ineqtobeproved}, so we will make reference to that proof below. Let $\mathcal{N}\subset\mathbb{N}$ be an infinite sequence for which $\left(\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}N^2 \log N}{N^2}\right)_{N\in\mathcal{N}}$ converges and we shall show that
\begin{equation}\label{eq:ineqtobeproved2}
\lim_{N\in\mathcal{N}}\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}N^2 \log N}{N^2}\leq L+\frac{\kappa}{\pi}.
\end{equation}
As in the proof of \eqref{eq:ineqtobeproved}, if there exists $p\geq 1$ such that an infinite number of $N\in\mathcal{N}$ satisfy $\tau(N)=p$, then it is clear that \eqref{eq:ineqtobeproved2} holds.
So we assume now that $\tau(N)\rightarrow\infty$ as $N\rightarrow\infty$ in the sequence $\mathcal{N}$. We have, for $N=\sum_{k=1}^{\tau(N)} 2^{n_{k}}$,
\begin{align}
\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}N^2 \log N}{N^2}= & \sum_{k=1}^{\tau(N)}\left(\frac{2^{n_{k}}}{N}\right)^{2}\left(\mathcal{R}_{1}(2^{n_{k}})+2^{-n_{k}+1}\,(2\,\mathcal{R}_{1}(2^{n_{k}+1})-\mathcal{R}_{1}(2^{n_{k}})) \sum_{j=k+1}^{\tau(N)}2^{n_{j}}\right)\notag\\
+ & \sum_{k=1}^{\tau(N)}\left(\frac{2^{n_{k}}}{N}\right)^{2}\left(r(2^{n_{k}})+2^{-n_{k}+1}\,(2\,r(2^{n_{k}+1})-r(2^{n_{k}})) \sum_{j=k+1}^{\tau(N)}2^{n_{j}}\right),\label{eq:decompE1norm}
\end{align}
where we use the notation
\[
r(2^{n_{k}})=\frac{1}{\pi} \log(2^{n_{k}}/N),\qquad r(2^{n_{k}+1})=\frac{1}{\pi} \log(2^{n_{k}+1}/N).
\]
Let $\epsilon>0$ be arbitrary, and choose $M\in\mathbb{N}$ sufficiently large so that \eqref{bound4} holds. Let $\lambda_{N,k}$ denote the expression in \eqref{def:lambdaNk:2} with $s=1$, and let
\[
\rho_{N,k}:=r(2^{n_{k}})+2^{-n_{k}+1}\,(2\,r(2^{n_{k}+1})-r(2^{n_{k}})) \sum_{j=k+1}^{\tau(N)}2^{n_{j}}.
\]
We see from \eqref{eq:decompE1norm} that we can write
\begin{equation}\label{eq:decompE1norm:2}
\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}N^2 \log N}{N^2}=S_{N,M,1}+S_{N,M,2}+S_{N,M,3}+S_{N,M,4},
\end{equation}
where $S_{N,M,1}$ and $S_{N,M,2}$ are defined in \eqref{def:SNM1SNM2} (taking $s=1$), and
\[
S_{N,M,3}:=\sum_{k=1}^{M}\left(\frac{2^{n_{k}}}{N}\right)^2 \rho_{N,k},\qquad S_{N,M,4}:=\sum_{k=M+1}^{\tau(N)}\left(\frac{2^{n_{k}}}{N}\right)^2 \rho_{N,k}.
\]
As in \eqref{bound1} we have
\[
|\lambda_{N,k}|\leq C_{1},\qquad \mbox{for all}\,\,N\in\mathcal{N}\,\,\mbox{and}\,\,k=1,\ldots,\tau(N),
\]
for some constant $C_{1}>0$. Therefore as in \eqref{bound5} we have
\[
|S_{N,M,2}|< C_{1} \epsilon,\qquad \mbox{for all}\,\,N\in\mathcal{N}.
\]
We again write
\[
S_{N,M,1}=D_{N,M,1}+D_{N,M,2}
\]
with $D_{N,M,1}$ and $D_{N,M,2}$ given by \eqref{def:DNM1} and \eqref{def:DNM2}, respectively, taking $s=1$ in these formulas. We also have the estimate \eqref{bound6}. If we use \eqref{energydecomp:3}, the bound \eqref{bound7} and the previous estimates, we conclude that
\begin{gather}
S_{N,M,1}+S_{N,M,2}\label{estimateSNM1SNM2}\\
=\sum_{k=1}^{M}\left(\frac{2^{n_{k}}}{2^{n_{1}}+\cdots+2^{n_{M}}}\right)^2\left(\mathcal{R}_{1}(2^{n_{k}})+2^{-n_{k}+1}\,(2\,\mathcal{R}_{1}(2^{n_{k}+1})-\mathcal{R}_{1}(2^{n_{k}})) \sum_{j=k+1}^{M}2^{n_{j}}\right)+O(\epsilon).\notag
\end{gather}
The analysis for the sum $S_{N,M,3}+S_{N,M,4}$ follows the same argument, so we will not reproduce it below. Now we need to take into account the following estimates, which are easy to check: There exists an absolute constant $C>0$, independent of $N$ and $M$, such that
\begin{align*}
\left|\frac{2^{n_{k}}}{N}\,\rho_{N,k}\right| & <C,\qquad \mbox{for all}\,\,N\in\mathcal{N}\,\,\mbox{and}\,\,k=1,\ldots,\tau(N), \\
\sum_{k=1}^{M}\frac{2^{n_{k}}}{N}\,\left|\log\left(\frac{2^{n_{k}}}{N}\right)\right| & <C,\qquad \mbox{for all}\,\,N\in\mathcal{N}\,\,\mbox{and}\,\,M< \tau(N).
\end{align*}
Using these estimates we find similarly that
\begin{gather}
S_{N,M,3}+S_{N,M,4}\label{estimateSNM3SNM4}\\
=\sum_{k=1}^{M}\left(\frac{2^{n_{k}}}{2^{n_{1}}+\cdots+2^{n_{M}}}\right)^2\left(\widetilde{r}(2^{n_{k}})+2^{-n_{k}+1}\,(2\,\widetilde{r}(2^{n_{k}+1})-\widetilde{r}(2^{n_{k}})) \sum_{j=k+1}^{M}2^{n_{j}}\right)+O(\epsilon),\notag
\end{gather}
where we use the notation
\[
\widetilde{r}(2^{n_{k}})=\frac{1}{\pi}\log\left(\frac{2^{n_{k}}}{2^{n_{1}}+\cdots+2^{n_{M}}}\right),\qquad \widetilde{r}(2^{n_{k}+1})=\frac{1}{\pi}\log\left(\frac{2^{n_{k}+1}}{2^{n_{1}}+\cdots+2^{n_{M}}}\right).
\]
Finally, we let $\widetilde{\mathcal{N}}$ be a subsequence of $\mathcal{N}$ such that the limits \eqref{cond:subseqNtilde} hold. Then, as in \eqref{limitfinitetau} we see that along the subsequence $\widetilde{\mathcal{N}}$, the first expression on the right-hand side of \eqref{estimateSNM1SNM2} converges to $L$, and the first expression on the right-hand side of \eqref{estimateSNM3SNM4} converges to $(1/\pi)\,K(\theta_{1},\ldots,\theta_{M})$. Therefore, applying \eqref{eq:decompE1norm:2}, \eqref{estimateSNM1SNM2} and \eqref{estimateSNM3SNM4}, we conclude that
\begin{gather*}
\lim_{N\in\mathcal{N}}\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}N^{2}\log N}{N^{2}}
=\lim_{N\in\widetilde{\mathcal{N}}}\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}N^{2}\log N}{N^{2}}\\=\lim_{N\in\widetilde{\mathcal{N}}}\left(S_{N,M,1}+S_{N,M,2}+S_{N,M,3}+S_{N,M,4}\right)
\leq L+\frac{1}{\pi} K(\theta_{1},\ldots,\theta_{M})+O(\epsilon)\leq L+\frac{\kappa}{\pi}+O(\epsilon).
\end{gather*}
This proves \eqref{eq:ineqtobeproved2} since $\epsilon$ is arbitrary.
The formula \eqref{secondorderasympcritcase:2} follows immediately from \eqref{eq:secorderFeketecrit} and the inequality $E_{1}(\alpha_{N})\geq \mathcal{L}_{1}(N)$, which is an equality for all $N$ of the form $N=2^{n}$. \hfill$\Box$
\bigskip
The following figure shows the first $4200$ values of the sequence $\left(\frac{E_{1}(\alpha_{N})-\frac{1}{\pi}N^{2}\log N}{N^2}\right)_{N}$.
\begin{figure}[h!]
\centering
\includegraphics[totalheight=3in,keepaspectratio]{sequences1for4200.eps}
\label{fig:sequences1first4200}
\end{figure}
\bigskip
\noindent\textbf{Acknowledgments}
\smallskip
\noindent We thank the referee for the many valuable comments provided.
|
\section*{Supplementary Material}
\section{Threshold singularities in Equation (4) of the main text (MT) for
generic spin-orbit coupling (SOC)}
For a generic (but still weak compared to the Fermi energy SOC) the spin-orbit band splitting, $\Delta_{\bk k_F}$, is anisotropic. Suppose that $\Delta_{\bk k_F}$
varies from $\Delta_{\text{min}}$ to $\Delta_{\text{max}}$ along the Fermi surface (FS). Quite generally,
the angular dependence of $\Delta_{\bk k_F}$ near the extremal points is $\Delta_{\text{min}}\approx \Delta_{-}+\beta^2_{+}\theta^2$ and $\Delta_{\text{max}}\approx \Delta_{+}-\beta^2_{-}\theta^2$. Near each of these extremal points, the diverging part of the angular integral in Eq.~(4) of MT can be written as
\bea
\left\langle\frac{\Delta_{\bk_F}^2}{\Omega^2-\Delta_{\bk_F}^2}\right\rangle_{\mathrm{FS}} &\approx &
\left\{
\begin{array}{cll}
\frac 12\int\frac{d\theta}{2\pi}~
\frac{\Delta_+}{(\Omega-\Delta_++\beta_+^2\theta^2)}\propto\frac{1}{\sqrt{\Omega-\Delta_+}},~~~\text{near maximum},\nonumber\\
\frac12\int\frac{d\theta}{2\pi}~
\frac{\Delta_+}{(\Omega-\Delta_--\beta_-^2\theta^2)}\propto-\frac{1}{\sqrt{\Delta_--\Omega}},~~~\text{near minimum}.
\end{array}
\right.
\label{eq:app1}
\eea
Note that the angular average is imaginary when $\Delta_-<\Omega<\Delta_+$, which correspond to the continuum of inter-subband particle-hole excitations. The singularity in the real part of the susceptibility at the lower end of the continuum ensures that the collective modes exist even for infinitesimally weak electron-electron interaction ({\em eei}). For an isotropic SOC, {\em e.g.}, either for Rashba or Dresselhaus SOC, $\Delta_+=\Delta_-\equiv \Delta$ and
the two square-root branch cuts merge into a simple pole $1/(\Omega-\Delta)$.
\section{Static part of the self-energy for momentum-dependent interaction}
\label{sec:Sigma}
In the MT, we argued that the difference of the self-energies of chiral subbands does not contain the zeroth angular harmonic of the interaction potential. In this section, we prove this statement. The self-energy of a subband with chirality $s=\pm 1$ due to {\em eei} via a static but momentum-dependent potential, $U_{|\bq|}$, is equal to
\beq\label{eq:self energy} \Sigma_s(K)=-\frac12\int_{K'} \left
\{g_+(K')+g_-(K')+s\left[g_+(K')-g_-(K')\right]\cos\phi_{\bk\bk'}\right\}U_{|\bk-\bkp|}.
\eeq
For the case of linear Rashba SOC, $\phi_{\bk \bk'}$ is the angle between $\vec{k}$ and $\vec{k'}$ ($\equiv \theta_{\bk\bk'}$); for a general case, $\phi_{\bk \bk'}$ is some function of $\theta_{\bk\bk'}$
such that $\int\cos\phi_{\bk\bk'} d\theta_{\bk\bk'} =0$. We first consider the case of Rashba SOC in some detail and then point out the qualitative features that remain the same for arbitrary SOC.
In addition, we also assume that Rashba SOC is weak ($\lambda\ll v_F$), in which case the spin-orbit splitting can be approximated by its value projected on the FS in the absence
of SOC: $\Delta\approx 2\lambda k_F$. The difference of the self-energies of the chiral subbands, $\delta\Sigma_s\equiv \Sigma_s-\Sigma_{-s}$, is now given by
\bea\label{eq:delta sigma}
\delta\Sigma_s(K)&=&-s \int_{\vec{k}'} \left[n_F\left(\varepsilon_{\bk'}+\frac{\Delta}{2}\right)-n_F\left(\varepsilon_{\bk'}-\frac{\Delta}{2}\right) \right]\cos\theta_{\bk\bk'}U_{|\bk-\bk'|},
\eea
where $n_F(\epsilon)$ is the Fermi function.
Setting $T=0$ and expanding the Fermi functions in $\Delta$, we find
\beq
\delta\Sigma_s(K)=s\nu\Delta\int\frac{d\theta_{\bk\bk'}}{2\pi} \cos\theta_{\bk\bk'}U_{|\bk-\bk'_F|},
\eeq
where $\bk'_F=k_F \bk'/\bk$. To find the frequency of the collective mode, one needs to project $\bk$ on the FS upon which $\delta\Sigma_s$ becomes proportional to the first harmonic
of the interaction potential.
A similar argument can be made for an arbitrary (but still weak) SOC. The Green's function in Eq.~(2) of MT can be separated into symmetric (S) and asymmetric (A) parts: $G=G^{\mathrm{S}}+G^{\mathrm{A}}$. Any crystalline plane has at least a $C_2$ symmetry upon which $\bk\to -\bk$. This guarantees that $\Delta_{-\bk}=\Delta_{\bk}$ and thus the subband Green's functions,
$g_{\pm}$, are even on $\bk\to -\bk$. Then $G^{\mathrm{S}}$ is even while $G^{\mathrm{A}}$ is odd on this operation. Consider the self-energy in the spin basis (a $2\times 2$ matrix)
\beq
\Sigma(K)=-\int_{K'} U_{|\bk-\bk'|} G(K'),
\eeq
which can be also decomposed into symmetric and asymmetric parts: $\Sigma=\Sigma^{\mathrm{S}}+\Sigma^{\mathrm{A}}$.
The zeroth angular harmonic of the interaction potential, $U^{\{0\}}$, survives only in the part of the integral associated with $G^{\mathrm{S}}$ and enters the diagonal elements of $\Sigma$ as $-U^{\{0\}}n$, where $n$ is the total number density. When $\Sigma$ is transformed to the (diagonal) chiral basis, the $-U^{\{0\}}n$ terms remain on the diagonal and cancel out in the difference $\Sigma_{+}-\Sigma_{-}$.
\section{Aslamazov-Larkin (AL) diagrams}
In the presence of a dynamic and long range interaction, the AL diagrams are of the same order as the self energy and vertex diagrams (See Refs. 29-32 of MT). In the absence of SOC and for a spin-invariant interaction, however, AL diagrams for the spin susceptibility vanish for a trivial reason: Tr[$\sigma_i {G} {G}{G}$]$=0$, ($i=1,2,3$) because ${G} \propto\sigma_0$. In the presence of the magnetic field (along the $3$-axis), ${G}= a\sigma_0 + b\sigma_3$. The relevant quantity in this case in the transverse spin susceptibility, which contains the trace Tr[$\sigma_i {G}{G}{G}$] with $i=1,2$. Again, this trace is equal to zero. The vanishing of the AL diagrams
in both these cases is a consequence of conservation of either total spin (in the former case) or its component along the magnetic field (in the latter case). In the presence of SOC, this reasoning does not apply because ${G}$ now has off-diagonal elements. However, some general statements about the AL diagrams can still be made for $q=0$ case.
\begin{figure}[htp]
\includegraphics[width=0.7\columnwidth]{5}
\caption{
\label{fig:AL} Detailed AL diagrams from Fig. 2 of MT}
\end{figure}
For convenience, we present the AL diagrams from Fig. 2, {\em d} and {\em e} of MT here in Fig.~\ref{fig:AL}. It suffices to consider only the diagonal components of the spin susceptibility,
$\chi_{ii}$, with $i=1\dots 3$, in which case the diagrams read
\bea\label{eQ:AL1} \Pi^d_{ii}&=&\int_{P}\int_K \text{Tr}\left[ \sigma_i{G}(K){G}(K+P){G}(K+Q)\right] V_PV_{P-Q}\int_{K_1}\text{Tr}\left[\sigma_i{G}(K_1+Q){G}(K_1+P){G}(K_1)\right] , \nonumber\\
\Pi_{ii}^e&=&\int_{P}\int_K \text{Tr}\left[\sigma_i {G}(K+Q){G}(K+Q-P){G}(K) \right] V_PV_{P-Q}\int_{K_1}\text{Tr}\left[\sigma_i{G}(K_1+Q){G}(K_1+P){G}(K_1)\right]. \nonumber\\
\eea
Here, $Q$ has only the temporal component: $Q=(iq_0,\vec 0)$. We can define
\bea\label{eq:4func}
h_i(P,Q)\equiv\int_K\text{Tr}\left[\sigma_i{G}_{K+Q}{G}_{K+P}{G}_K\right],\nonumber\\
\tilde{h}_i(P,Q)\equiv\int_K\text{Tr}\left[\sigma_i{G}_{K}{G}_{K+P}{G}_{K+Q}\right],\nonumber\\
f_i(P,Q)\equiv\int_K\text{Tr}\left[\sigma_i{G}_{K+Q}{G}_{K+Q-P}{G}_K\right].
\eea
Using the matrix structure of ${G}(K)$, we observe that\beq\label{eq:AL2}h_{1,2}(P,Q)=+\tilde{h}_{1,2}(P,Q);~ h_3(P,Q)=-\tilde{h}_3(P,Q)\; \text{and}\;\;f_i(P,Q)=-\tilde h_i^*(P,Q)~~\text{for}~~ i=1,2,3.
\eeq
With the help of these properties, the AL diagrams can now be written as
\bea
\Pi_{ii}^d&=&\int_{P} h_i(P,Q)\tilde{h}_i(P,Q) V_PV_{P-Q}= \int_{P} h_i^2(P,Q) V_PV_{P-Q}~\text{for}~i=1,2~\text{and}~\nonumber\\
&=&-\int_{P} h_i^2(P,Q) V_PV_{P-Q}~\text{for}~i=3,\label{eQ:AL5}\\
\Pi_{ii}^e
&=&\int_{P} f(P,Q)\tilde{h}_i(P,Q) V_PV_{P-Q}=-\int_{P}|h_i(P,Q)|^2 V_PV_{Q-P},
\eea
and thus their sum is reduced to
\bea\label{eQ:AL6}
\Pi^d_{ii}+\Pi^e_{ii}
&=&\int_{P} \left(h_i^2(P,Q) - |h_i(P,Q)|^2\right) V_PV_{P-Q}, ~~~\text{for}~~i=1,2~~\text{and}\nonumber\\
&=&-\int_{P}
\left(h_i^2(P,Q) + |h_i(P,Q)|^2\right) V_PV_{P-Q}, ~~~\text{for}~~i=3.
\eea
We now study two particular cases--that of Rashba SOC and of combined Rashba-Dresselhaus SOC--
and show that the AL diagrams do not contribute to damping in either of these two cases.
\emph{Rashba SOC}: For $i=3$,
\beq
\label{h3}
h_3(P,Q)=-\frac i4\sum_{a,b,c=\pm} b(a-c)\int_K \sin(\theta_{\vec k}-\theta_{\vec k+\vec p})g_a(K+Q)g_b(K+P)g_c(K).
\eeq
Since $Q$ has no spatial component, we can choose the direction of $\bp$ as a reference. On reflecting vector $\bk$ about this direction, both $\theta_\bk$ and $\theta_{\bk+\bp}$ change signs and so does the factor of $\sin(\dots)$. On the other hand, since the energy spectrum is isotropic, each of the three subband Green's function in the formula above is a function of the magnitude of the corresponding momentum. It is then easy to see that all the three Green's function are even on reflection $k_1\to k_1$, $k_2\to -k_2$. Therefore, the integral over $\theta_{\bk}$ vanishes, and the AL diagrams give no contribution to
$\chi_{33}$.
The same reasoning applies for $i=1$. In this case,
\beq
\label{h1}
h_1(P,Q)=\frac14\sum_{a,b,c=\pm}\int_K\left[(a+c)\sin\theta_{\bk}+abc\sin(2\theta_{\bk}+\theta_{\bk+\bp})
+b\sin\theta_{\bk+\bp}\right]g_a(K+Q)g_b(K+P)g_c(K)
\eeq
contains again an angular factor, which is odd on reflection, and a combination of the Green's functions, which is even on reflection. Therefore, $h_1(P,Q)=0$ and there is no contribution to $\chi_{11}$ from the AL diagrams either.
The same argument cannot be used for $i=2$, because the angular factor in
\beq
\label{h2}
h_2(P,Q)=-\frac14\sum_{a,b,c=\pm}\int_K\left[(a+c)\cos\theta_{\bk}+abc\cos\left(2\theta_{\bk}+\theta_{\bk+\bp}\right)
+b\cos\theta_{\bk+\bp})\right]g_a(K+Q)g_b(K+P)g_c(K).
\eeq
is even on reflection. Nevertheless, in-plane rotational symmetry of the Rashba Hamiltonian ensures that $\chi_{11}=\chi_{22}$. Therefore, one does not need to consider $\chi_{22}$ separately: if there is no damping of resonance associated with $\chi_{11}$, the same is true for $\chi_{22}$.
\emph{Combined Rashba and Dresselhaus SOC}: In the presence of both Rashba and Dresselhaus SOC, the only change in Eqs.~((\ref{h3}-\ref{h2})) is that $\theta_{\bk}$ needs to replaced by $\phi_{\bk}$, as defined after Eq.~(\ref{eq:self energy}). It is convenient to choose the Cartesian system rotated by $\pi/4$ compared to the conventional one. In such a system, the combined Rashba plus Dresselhaus Hamiltonian is described by in Eq.~(1) of MT with $\lambda\vec f (k)=\left([\beta+\alpha]k_2,[\beta-\alpha]k_1,0\right)$ and
\bea\label{eq:app2}
\cos\phi_{\bk}&=&-\frac{(\alpha-\beta) \cos\theta_\bk}{\Lambda_\bk},\nonumber\\
\sin\phi_\bk&=&-\frac{(\beta+\alpha)\sin\theta_\bk}{\Lambda_\bk},\eea
where $\Lambda_\bk=\sqrt{\alpha^2+\beta^2-2\alpha\beta\cos2\theta_\bk}$. Note that $\phi_{\bk}=\tan^{-1}\left(\frac{\alpha+\beta}{\alpha-\beta}\tan\theta_{\vec k}\right)$.
The spin-orbit splitting $\Delta_{\bk_F}=2k_F\Lambda_{\bk}$ is even on reflection
about the $k_1$-axis. As discussed in MT one can approximate $\Delta_{\vec k + \vec p}$ by $\Delta_{\vec{k}_F}$ for a long-range interaction; therefore, the products of the Green's functions are still even on reflection. Because of the relation between $\phi_{\vec k}$ and $\theta_{\vec k}$, the same symmetries that led to the vanishing of $h_{1}$ and $h_{3}$ in the case with only Rashba SOC ensure that $h_{1}$ and $h_{3}$ vanish is the case when both Rashba and Dresselhaus SOC are present as well.
\section{Derivation of Equation (11) of the Main Text \label{suppl:damping}}
In this section, we present the evaluation of $\Pi_{\text{D}}$ from Eq.~(10) of the MT. For convenience, we present this equation below in an expanded form:
\bea
\label{sun1}
\Pi_{\text{D}}=-\int_K \frac{\Delta_{\vec k_F}}{q_0^2+\Delta_{\vec k_F}^2}\left\{\left[\Sigma_{--}(K+Q)-\Sigma_{+-}(K)\right]g_-(K)g_+(K+Q) - \left[\Sigma_{+-}(K+Q)-\Sigma_{--}(K)\right]g_+(K)g_-(K+Q)\right\},\nn\\
\eea
where
$\Sigma_{ab}(K)=-\int_P g_a(K+P)V_{\text{eff}}(P)\left[1+b\cos(\phi_{\bk}-\phi_{\bk+\bp})\right]$ and the angle $\phi_{\bk}$ is defined after Eq.~(\ref{eq:self energy}). Since $\Pi_{\text{D}}$ is already proportional to $\Delta_{\vec k_F}$, we can neglect the effect of SOC on the interaction potential and take it to be a dynamically screened Coulomb potential in 2D:
\beq
\label{eq:scrC}
V_{\mathrm{eff}}(P)=\frac{2\pi e^2}{p+\kappa\left(1-\frac{|p_0|}{\sqrt{p_0^2+v_F^2p^2}}\right)},
\eeq
where $\kappa= \sqrt{2}r_s k_F$ is the inverse Thomas-Fermi screening length and $r_s =\sqrt{2}e^2/v_F$. The momentum transfer $\vec p$ is decomposed into two components: along the normal to the FS at point $\vec k$ ($p_{||}$) and along the tangent ($p_\perp$). We assume and then verify that the quasiparticle interaction is determined by processes with typical $p_{||}$ on the order of $p_0/v_F$, whereas typical $p_\perp$ are on the order of $\kappa$. In turn, typical $p_0$ are on the order of the external frequency which, in our case, is fixed by spin-orbit splitting, $\bar\Delta$, averaged over the FS. We choose to work in a realistic regime of $\bar\Delta\ll v_F\kappa\ll E_F$. [Although the form of the potential in Eq.~(\ref{eq:scrC}) is, strictly speaking, valid for $r_s\ll 1$ which implies that $\kappa\ll k_F$, these strong inequalities are never satisfied in real materials. Therefore, the interval of energies in between $v_F\kappa$ and $E_F$ is never wide enough to consider a possibility of $\bar\Delta$ being within this interval.] Since \beq
|p_{||}|\sim |p_0|/v_F\ll |p_\perp|\label{ineq}
\eeq in this regime, we can simplify the Landau-damping term in Eq.~(\ref{eq:scrC}) to $|p_0|/v_F|p_\perp|$ and also expand the denominator of this equation in this parameter. Subtracting off the static screened potential which gives no contribution to damping, we obtain for the dynamical ($p_0$-dependent) part of the interaction
\beq\label{eq:dyn_2}
U^{\text{dyn}}_{\text{eff}}(P)=\left(\frac{2\pi e^2}{|p_\perp|+\kappa}\right)^2
\frac{\nu}{v_F}\frac{|p_0|}{|p_\perp|}.\eeq
Thanks to the inequality (\ref{ineq}), the integrals over the fermionic ($K$)
and bosonic ($P$) momenta become separable. Let us pick one of the terms
in Eq.~(\ref{sun1}): \bea\label{eq:step1}
\int_K \Sigma_{--}(K+Q)g_-(K)g_+(K+Q)=-\int_K \int_P U^{\text{dyn}}_{\text{eff}}g_-(K+Q+P)g_-(K)g_+(K+Q)\left[1-\cos\left(\phi_{\vec k}-\phi_{\vec k
+\vec p}\right)\right].
\eea
Since it is expected that typical $|p_\perp|\sim\kappa\ll k_F$ while $\vec k$ is expected to be near the FS, the angular factor can be expanded as $1-\cos(\phi_{\vec k}-\phi_{\vec k +\vec p}) \approx\xi^2
_{\bk}\frac{p_\perp^2}{2k_F^2}$, where $\xi_{\bk}$ is a form-factor that depends on the details of the SOC.
For example, in the case of linear Rashba SOC, $\phi_{\vec k}$ coincides with the angle $\theta_{\bk}$ that $\vec k$ makes with, e.g., the $1$-axis. Also, due to rotational symmetry, the result of the integral over $P$ does not depend on the direction of $\bk$ and thus we can choose $\theta_{\bk}=0$ in $1-\cos(\theta_{\vec k}-\theta_{\vec k +\vec p})$. An expansion of $1-\cos(\theta_{\vec k +\vec p})$ then yields $\frac{p_\perp^2}{2k_F^2}$, which means that $\xi_{\bk}=1$ for a system with linear Rashba SOC.
Notice that the condition $|p_{||}|\ll |p_{\perp}|$ implies that $\bp$ is almost perpendicular to $\bk$, i.e., that $|\theta_{\bp}-\theta_{\bk}|\approx\pi/2$.
In the presence of both Rashba and Dresselhaus SOC, $\phi_{\bk}$ is given by Eq.~(\ref{eq:app2}). Expanding in $p$ and using that $|\theta_\bp-\theta_\bk|\approx\pi/2$, we get
\bea\label{eq:app3}
1-\cos(\phi_{\vec k}-\phi_{\vec k +\vec p})&\approx& \frac{p_\perp^2}{2k_F^2}\left[1
+\left(\frac{\gamma}{1-\gamma\cos2\theta_
\bk}\right)^2\right], ~~\text{where}~~\gamma=\frac{2\alpha\beta}{\alpha^2+\beta^2}<1
\eea
Thus $\xi^2_\bk=1+\left(\frac{\gamma}{1-\gamma\cos2\theta_\bk}\right)^2$ for this system. We avoid the special case of $|\alpha|=|\beta|$, when $\gamma=1$ and $\xi_{\bk}$ has a pole: it can be shown that collective modes are completely covered by the continuum in this case.\cite{imran}
Next, we write $g_-(K)g_+(K+Q)$ as $[g_-(K)-g_+(K+Q)]/(iq_0-\Delta_{\vec k})$ and integrate the products of two Green's functions over $\e_{\vec k}$, $k_0$ and $p_{||}$; for one of the products, we obtain
\bea\label{eq:temp step}
&&\int\frac{dp_{\parallel}}{2\pi}\int\frac{dk_0}{2\pi}\int d\varepsilon_{\vec k} g_a(K+P)g_b(K)=\int\frac{dp_{\parallel}}{2\pi} \frac{ip_0}{ip_0-v_Fp_{\parallel} - \frac{a-b}{2}\Delta_{\vec k}}=\frac{|p_0|}{2v_F}.
\eea
The rest of the integrals are evaluated in the similar manner. Collecting all terms, we arrive at
\bea\label{eq:leftover}
\Pi_{\mathrm{D}}=
-\left\langle\left(\frac{\Delta_{\vec k_F}\xi_{\bk_F}}{q_0^2+\Delta^2_{\vec k_F}}\right)^2\right\rangle_{\vec k}\frac{2\nu}{v_F k_F^2}\int_0^{k_F}\frac{dp_{\perp}}{2\pi}{p^2_{\perp}}\int_0^\infty \frac{dp_0}{2\pi}
U_{\text{eff}}^{\text{dyn}}(P)\left\{|q_0+p_0|+|q_0-p_0|-2|p_0|\right\}.\eea
The factor in $\{\dots \}$ constraints the range of integration over $p_0$ to $(0,q_0)$. This leads to
\bea\label{eq:formal}
\Pi_D&=&-\nu\left\langle\left(\frac{\Delta_{\vec k}\xi_{\vec k_F}
}{q_0^2+\Delta_{\vec k}^2}\right)^2\right\rangle_{\vec k}\frac{ r_s^2}{2\pi E_F}\int_0^{q_0} dp_0(q_0-p_0)p_0\int_{0}^{k_F} d p_\perp \frac{p_\perp}{(p_\perp+\kappa)^2}
\eea
To logarithmic accuracy, the integral over $p_\perp$ gives $\ln k_F/\kappa\approx -\ln r_s$ while the integral over $p_0$ gives $q_0^3/6$. Note that typical $p_0\sim q_0$ while typical $p_{\perp}\sim\kappa$ (in the logarithmic sense), and thus our initial assumption is verified.
After analytic continuation, we get \bea\label{eq:final result22}
\Pi_D=-i\nu\left\langle\left(\frac{\Delta_{\vec k_F}\xi_{k_F}}{\Omega^2-\Delta^2_{\vec k_F}}\right)^2\right\rangle_{{\vec k}_F}\omega_C^2\left(\frac{\Omega}{E_F}\right)^3,
\eea
where $\omega^2_C=r_s^2 E_F^2 \ln r_s^{-1}/12\pi$ and $r_s=\sqrt{2}e^2/v_F$. This is the correction to $\Pi^0_{33}$ that is responsible for damping. Recalling that the physical susceptibility is given by $\chi_{33}=-\mu_B^2\Pi_{33}/(1+\frac{U_{\text{x}}}{2}\Pi_{33})$ $\Pi_{33}$ differs from $\Pi_{33}^0$ due to the damping correction. To account for this correction, we may write
\beq\label{eq:w}
\Pi_{33}=2\nu\left\langle\frac{\Delta_{\vec k_F}^2\xi^2_{\vec k_F}}{\Delta_{\vec k_F}^2-\left(\Omega+i\Gamma/2\right)^2}\right\rangle_{\text{FS}},
\eeq
where $\Gamma = \frac{\omega_C^2}{2E_F}\left(\frac{\Delta_{\vec k_F}}{E_F}\right)^2$. Using the fact that $\Omega_3$ is the pole in $\chi_{33}$ when $\Gamma=0$, we can eliminate the $U_{\text{x}}$ dependence. Writing $\Delta_{\vec k_F}^2-\left(\Omega+i\Gamma/2\right)^2$ as $\Delta_{\vec k_F}^2-\Omega_3^2+\Omega_3^2-\left(\Omega+i\Gamma/2\right)^2$ and expanding in $\Omega_3^2-\left(\Omega+i\Gamma/2\right)^2$ near the resonance, we obtain Eq.~(11) of MT.
\end{widetext}
|
\section{Introduction}\label{sec:intro}
Let $(M,d)$ be a locally compact separable metric space equipped with
a $\sigma$-finite Borel measure $\mu$ and let
$X=\bigl(\{X_{t}\}_{t\in[0,\infty]},\{\mathbb{P}_{x}\}_{x\in M_{\cemetery}}\bigr)$
be a diffusion on $M$, where $M_{\cemetery}:=M\cup\{\cemetery\}$
denotes the one-point compactification of $M$.
The themes of this paper are existence of the heat kernel
$p_{t}(x,y)$ (the transition density of $X$ with respect to $\mu$)
and off-diagonal upper bounds of $p_{t}(x,y)$ of the form
\begin{equation}\label{eq:UHK-intro}
p_{t}(x,y)\leq F_{t}(x,y)\exp\biggl(-\gamma\Bigl(\frac{d(x,y)^{\beta}}{t}\Bigr)^{\frac{1}{\beta-1}}\biggr)
\end{equation}
for some $\gamma\in(0,\infty)$, $\beta\in(1,\infty)$ and a positive function $F_{t}(x,y)$.
In most typical cases, $F_{t}(x,y)$ is given either by the power function
$F_{t}(x,y)=c_{0}t^{-\alpha}$ for some $c_{0},\alpha\in(0,\infty)$ or by the volume function
\begin{equation}\label{eq:F-vol-intro}
F_{t}(x,y)=c_{0}\mu\bigl(B(x,t^{1/\beta})\bigr)^{-1/2}\mu\bigl(B(y,t^{1/\beta})\bigr)^{-1/2},
\end{equation}
where $\beta$ is as in \eqref{eq:UHK-intro} and
$B(x,r):=\{y\in M\mid d(x,y)<r\}$ for $(x,r)\in M\times(0,\infty)$.
For $\beta=2$, \eqref{eq:UHK-intro} is called a \emph{Gaussian} upper bound and
has been extensively studied in the classical setting where $M$ is a complete
Riemannian manifold. For example, when $M$ has non-negative Ricci curvature,
the Gaussian bound \eqref{eq:UHK-intro} under \eqref{eq:F-vol-intro},
together with a matching lower bound, has been proved
for the Brownian motion on $M$ by Li and Yau \cite{LiYau:ActaMath86} and for
uniformly elliptic diffusions on $M$ by Saloff-Coste \cite{S-C:JDG92}. It is
also known by the results of Grigor'yan \cite{Gri:MatSb91,Gri:RevMatIberoam94}
and Saloff-Coste \cite{S-C:IMRN92,S-C:JDG92} that these bounds are
characterized or implied by certain scale-invariant functional inequalities,
such as Poincar\'{e}, local Sobolev and relative Faber-Krahn inequalities,
in conjunction with the volume doubling property
\begin{equation}\label{eq:VD-intro}
0<\mu(B(x,2r))\leq c_{\mathrm{vd}}\mu(B(x,r))<\infty.
\end{equation}
Saloff-Coste's proofs have developed from Moser's iteration argument in
\cite{Moser:CPAM64,Moser:CPAM71} combined with Davies' method in \cite{Dav:AJM87}
for making the constant $\gamma$ in \eqref{eq:UHK-intro} arbitrarily close to $\frac{1}{4}$,
and have been extended by Sturm \cite{Sturm:ALDS2,Sturm:ALDS3} to the framework of
a general strongly local regular Dirichlet space whose associated intrinsic metric
is non-degenerate. This last property basically means that for each relatively
compact ball $B(x,r)$ there exists a cutoff function $\varphi=\varphi_{x,r}$
satisfying $\ind{B(x,r)}\leq\varphi\leq\ind{B(x,2r)}$ and
``$|\nabla\varphi|\leq r^{-1}$" $\mu$-a.e., which makes it possible to apply
the methods developed for Riemannian manifolds to an abstract setting.
It should also be noted that such cutoff functions allow us to deduce
\emph{localized} Gaussian bounds from \emph{localized} assumptions; for example,
a Gaussian upper bound of $p_{t}(x,y)$ for \emph{given} $x,y\in M$ is implied by
a local Sobolev inequality \emph{on two balls $B(x,r_{x})$ and $B(y,r_{y})$ alone}.
See \cite{Dav:HK,Gri:HKAnalysisManifolds,S-C:AspectsSobolev,Sturm:ALDS2,Sturm:ALDS3}
and references therein for further details of Gaussian bounds.
The values of $\beta$ \emph{greater than $2$} naturally appear in the study of
diffusions on fractals. Barlow and Perkins have proved in their seminal work
\cite{BP} that the canonical diffusion on the two-dimensional Sierpi\'{n}ski
gasket satisfies \eqref{eq:UHK-intro} with \eqref{eq:F-vol-intro} and
$\beta=\log_{2}5>2$ as well as a matching lower bound, which indicate a lower
diffusion speed of the heat and are thereby called \emph{sub}-Gaussian bounds.
Such two-sided bounds with $\beta>2$ have been established also for nested fractals
by Kumagai \cite{Kum:nested}, affine nested fractals by Fitzsimmons, Hambly and
Kumagai \cite{FHK:affinenested} and Sierpi\'{n}ski carpets by Barlow and Bass
\cite{BB92,BB99} (see also \cite{BBKT}), which in turn have motivated a number
of recent studies on characterizing sub-Gaussian bounds, like
\cite{BBK:CS,BGK:PHI,GriHu:HKGreen,GriHuLau:GenCap,GT,Kig:RFQS,Kum:RF} for two-sided
and \cite{AB,Gri:HKfractal,GriHu:Upper,GriHuLau:GenCap,Kig:localNash} for upper.
A huge technical difficulty in the sub-Gaussian case is that, even though
we can construct good cutoff functions similar to the Gaussian case
\emph{a posteriori on the basis of sub-Gaussian bounds}
as has been done in \cite{AB,BBK:CS,GriHuLau:GenCap},
it is hopeless to have such functions \emph{a priori}; indeed, the natural
distance function may well even \emph{not} belong to the domain of the
Dirichlet form as proved in \cite[Proposition A.3]{K:aghSG} for
the two-dimensional Sierpi\'{n}ski gasket. Therefore in getting
sub-Gaussian bounds, practically we cannot use analytic methods developed for
Gaussian bounds, and most of the existent researches have made indispensable
use of arguments on the diffusion process instead.
While calculations with the diffusion enable us to estimate various
analytic quantities through probabilistic considerations, it is not clear
whether they admit localized implications similar to the analytic proofs
of Gaussian bounds, and there seems to be no result in the literature stating
such implications explicitly. In fact, unless the diffusion $X$ has a certain
prescribed local regularity property as in the case of Riemannian manifolds and
that of resistance forms treated in \cite{Kig:RFQS}, localizing \emph{existence} results
for the heat kernel $p_{t}(x,y)$ is already highly non-trivial, since its existence
on a given subset could be prevented by the possibly very bad behavior of the
diffusion outside the subset. These issues of localization have been carefully
avoided in the known probabilistic derivations of sub-Gaussian heat kernel bounds,
either by assuming as in \cite{Kig:localNash} the ultracontractivity of
the heat semigroup and thereby the existence and boundedness of
the heat kernel $p_{t}(x,y)$, or by assuming good situations everywhere
in every scale as in \cite{Gri:HKfractal,GriHu:Upper,GT} and
their descendants \cite{GriHu:HKGreen,GriHuLau:GenCap}.
The purpose of this paper is to provide a new probabilistic method of obtaining
\emph{localized} existence and sub-Gaussian upper bounds of the heat kernel
$p_{t}(x,y)$ of $X$ from \emph{localized} assumptions on $X$. Now we briefly
outline the statements of our main theorems.
The main localized existence theorem for the heat kernel (Theorem \ref{thm:HK-existence})
is proved for a Radon measure $\mu$ on $M$ with full support and a $\mu$-symmetric
Hunt process
$X=\bigl(\{X_{t}\}_{t\in[0,\infty]},\{\mathbb{P}_{x}\}_{x\in M_{\cemetery}}\bigr)$
on $M$ (\emph{not} necessarily with continuous sample paths) whose Dirichlet form
$(\functionspace{E},\functionspace{F})$ is regular on $L^{2}(M,\mu)$.
Let $U$ be a non-empty open subset of $M$, set
$\tau_{U}:=\inf\{t\in[0,\infty)\mid X_{t}\in M_{\cemetery}\setminus U\}$
($\inf\emptyset:=\infty$) and let
$\{T^{U}_{t}\}_{t\in(0,\infty)}$ denote the Dirichlet heat semigroup on $U$.
Then Theorem \ref{thm:HK-existence} states that for an interval
$I\subset(0,\infty)$ and open subsets $V,W$ of $M$, a
``\emph{$\mu$-almost everywhere} upper bound for $\{T^{U}_{t}\}_{t\in(0,\infty)}$ on
$I\times V\times W$ by a locally bounded upper semi-continuous kernel $H=H_{t}(x,y)$"
yields a Borel measurable function
$p^{U}=p^{U}_{t}(x,y)$ with $0\leq p^{U}_{t}(x,y)\leq H_{t}(x,y)$ such that
for \emph{$\functionspace{E}$-quasi-every} $x\in V$, for any $t\in I$,
\begin{equation}\label{eq:HK-existence-intro}
\mathbb{P}_{x}[X_{t}\in dy,\,t<\tau_{U}]=p^{U}_{t}(x,y)\,d\mu(y)\qquad\textrm{on }W.
\end{equation}
In fact, the same sort of results along with some additional regularity
properties of $p_{t}(x,y)$ have been obtained for $I=(0,\infty)$ and $U=V=W=M$
in \cite[Sections 7 and 8]{Gri:HKfractal} and \cite[Theorem 3.1]{BBCK}, but
our Theorem \ref{thm:HK-existence} should suffice for most applications
since it already guarantees the expected bound $p^{U}_{t}(x,y)\leq H_{t}(x,y)$
without requiring any regularity of the heat kernel $p^{U}_{t}(x,y)$.
The proof of Theorem \ref{thm:HK-existence} is mostly based on potential theory
for regular symmetric Dirichlet forms developed in \cite[Chapters 2 and 4]{FOT};
it should not be very difficult to generalize Theorem \ref{thm:HK-existence} to
a wider framework where the same kind of potential theory is still available.
As an intermediate step for the proof of Theorem \ref{thm:HK-existence},
we also prove in Proposition \ref{prop:HK-existence} that
``for \emph{$\functionspace{E}$-quasi-every} $x\in V$" in the above
statement can be improved to ``for \emph{any} $x\in V$" if the inequality
$\mathbb{P}_{x}[X_{t}\in dy,\,t<\tau_{U}]\leq H_{t}(x,y)\,d\mu(y)$
holds on $W$ for \emph{any} $(t,x)\in I\times V$.
Next we turn to our second main theorem on localized sub-Gaussian upper bounds
of heat kernels (Theorem \ref{thm:HKUB-localized}). For the reader's convenience,
we give here the precise statement of a simplified version of it. For $B\subset M$,
set $\tau_{B}:=\inf\{t\in[0,\infty)\mid X_{t}\in M_{\cemetery}\setminus B\}$
($\inf\emptyset:=\infty$) and let $\sigmafield{B}(B)$ denote its Borel
$\sigma$-field under the relative topology inherited from $M$.
\begin{theorem}\label{thm:HKUB-localized-intro}
Let $(M,d)$ be a locally compact separable metric space, let $\mu$ be a
$\sigma$-finite Borel measure $\mu$ on $M$ and let
$X=\bigl(\Omega,\sigmafield{M},\{X_{t}\}_{t\in[0,\infty]},
\{\mathbb{P}_{x}\}_{x\in M_{\cemetery}}\bigr)$
be a Hunt process on $(M,\sigmafield{B}(M))$ with life time $\zeta$.
Let $N\in\sigmafield{B}(M)$ and assume that for any $x\in M\setminus N$,
\begin{equation}\label{eq:properly-exceptional-X-continuous-intro}
\mathbb{P}_{x}\bigl[X_{t}\in M_{\cemetery}\setminus N\textrm{ for any }t\in[0,\infty),\,
[0,\zeta)\ni t\mapsto X_{t}\in M\textrm{ is continuous}\bigr]=1
\end{equation}
\textup{(namely, $M\setminus N$ is $X$-invariant and the restriction
$X|_{M\setminus N}$ of $X$ to $M\setminus N$ is a diffusion)}.
Let $\beta\in(1,\infty)$, let $R\in(0,\infty)$, let $U$ be a non-empty open subset of $M$
with $\diam U\leq R$ and let $F=F_{t}(x,y):(0,R^{\beta}]\times U\times U\to(0,\infty)$
be Borel measurable. Let $c_{F},\alpha_{F},c,\gamma\in(0,\infty)$
and assume that the following three conditions \textup{(DB)$_{\beta}$},
\textup{(DU)$_{F}^{U,R}$} and \textup{(P)$_{\beta}^{U,R}$} hold:
\begin{itemize}[label=\textup{(DU)$_{F}^{U,R}$},align=left,leftmargin=*]
\item[\textup{(DB)$_{\beta}$}]
For any $(t,x,y),(s,z,w)\in(0,R^{\beta}]\times U\times U$ with $s\leq t$,
\begin{equation}\label{eq:upper-bound-function-F-DBbeta}
\frac{F_{s}(z,w)}{F_{t}(x,y)}
\leq c_{F}\biggl(\frac{t\vee d(x,z)^{\beta}\vee d(y,w)^{\beta}}{s}\biggr)^{\alpha_{F}}.
\end{equation}
\item[\textup{(DU)$_{F}^{U,R}$}]
For any $(t,x)\in(0,R^{\beta})\times(U\setminus N)$ and any $A\in\sigmafield{B}(U)$,
\begin{equation}\label{eq:HKUB-localized-on-diag-intro}
\mathbb{P}_{x}[X_{t}\in A,\,t<\tau_{U}]\leq\int_{A}F_{t}(x,y)\,d\mu(y).
\end{equation}
\item[\textup{(P)$_{\beta}^{U,R}$}]
For any $(x,r)\in(U\setminus N)\times(0,R)$ with $B(x,r)\subset U$
and any $t\in(0,\infty)$,
\begin{equation}\label{eq:exit-probability-intro}
\mathbb{P}_{x}[\tau_{B(x,r)}\leq t]
\leq c\exp\bigl(-\gamma(r^{\beta}/t)^{\frac{1}{\beta-1}}\bigr).
\end{equation}
\end{itemize}
Let $\varepsilon\in(0,1)$ and set
$U^{\circ}_{\varepsilon R}:=\{x\in M\mid\inf_{y\in M\setminus U}d(x,y)>\varepsilon R\}$
\textup{(note that $U^{\circ}_{\varepsilon R}$ is an open subset of $U$)}.
Then there exists a Borel measurable function
$p=p_{t}(x,y):(0,\infty)\times(M\setminus N)\times U^{\circ}_{\varepsilon R}\to[0,\infty)$
such that for any $(t,x)\in(0,\infty)\times(M\setminus N)$ the following hold:
\begin{equation}\label{eq:HKUB-localized-existence-intro}
\mathbb{P}_{x}[X_{t}\in A]=\int_{A}p_{t}(x,y)\,d\mu(y)
\qquad\textrm{for any }A\in\sigmafield{B}(U^{\circ}_{\varepsilon R}),
\end{equation}
and furthermore for any $y\in U^{\circ}_{\varepsilon R}$,
\begin{equation}\label{eq:HKUB-localized-off-diag-U-intro}
p_{t}(x,y)\leq
\begin{cases}
c_{\varepsilon}F_{t}(x,y)\exp\bigl(-\gamma_{\varepsilon}(d(x,y)^{\beta}/t)^{\frac{1}{\beta-1}}\bigr) &\textrm{if }t<R^{\beta}\textrm{ and }x\in U,\\
c_{\varepsilon}(\inf_{U\times U}F_{(2t)\wedge R^{\beta}})\exp\bigl(-\gamma_{\varepsilon}(R^{\beta}/t)^{\frac{1}{\beta-1}}\bigr) &\textrm{if }t<R^{\beta}\textrm{ and }x\not\in U,\\
c_{\varepsilon}(\inf_{U\times U}F_{R^{\beta}}) &\textrm{if }t\geq R^{\beta}
\end{cases}
\end{equation}
for some $c_{\varepsilon}\in(0,\infty)$ explicit in
$\beta,c_{F},\alpha_{F},c,\gamma,\varepsilon$ and
$\gamma_{\varepsilon}:=(\frac{1}{5}\varepsilon)^{\frac{\beta}{\beta-1}}\gamma$.
\end{theorem}
The strength of Theorem \ref{thm:HKUB-localized-intro} is that
\emph{the conditions \textup{(DU)$_{F}^{U,R}$} and \textup{(P)$_{\beta}^{U,R}$}
are independent of the behavior of $X$ after exiting $U$ and thereby completely
localized within $U$} but assure nevertheless the existence and an upper bound
of the heat kernel $p=p_{t}(x,y)$ for the \emph{global} transition function
$\mathbb{P}_{x}[X_{t}\in dy]$.
The power function $F_{t}(x,y)=c_{0}t^{-\alpha}$ clearly satisfies \textup{(DB)$_{\beta}$},
and it is easy to see that \textup{(DB)$_{\beta}$} holds also for the volume function
\eqref{eq:F-vol-intro} provided \eqref{eq:VD-intro} is satisfied for any
$(x,r)\in U\times(0,R)$; see Example \ref{exmp:upper-bound-function} for some more details.
In view of these examples of $F=F_{t}(x,y)$, \textup{(DU)$_{F}^{U,R}$} amounts
to an \emph{on-diagonal} upper bound of the heat kernel $p^{U}=p^{U}_{t}(x,y)$ for
$\{T^{U}_{t}\}_{t\in(0,\infty)}$, which is known to be implied in the setting of
a regular symmetric Dirichlet form by the \emph{local Nash inequality} as shown
in \cite[Lemma 4.3]{Kig:localNash} and by the \emph{Faber-Krahn inequality}
as treated in \cite[Subsection 5.2 and (5.48)]{GriHu:Upper}.
The proof of Theorem \ref{thm:HKUB-localized-intro} relies essentially only on
two probabilistic iteration arguments based on the strong Markov property of $X$,
where the series in the resulting upper estimates are shown to converge to the
desired bounds by making heavy use of the condition \textup{(P)$_{\beta}^{U,R}$}.
In this sense, \emph{\textup{(P)$_{\beta}^{U,R}$} could be considered as the
probabilistic replacement for cutoff functions with well-controlled energy}.
One iteration argument involves the behavior of $X$ within $U$ alone and is
used in the first step of the proof of Theorem \ref{thm:HKUB-localized-intro} to
obtain an off-diagonal sub-Gaussian type upper bound of the Dirichlet heat kernel
$p^{U}=p^{U}_{t}(x,y)$ on $U$ \emph{without assuming the symmetry of $X$}
(Proposition \ref{prop:HKUB-localized}). The other iteration is
formulated as an equality, which we call a \emph{multiple Dynkin-Hunt formula},
expressing the global transition function $\mathbb{P}_{x}[X_{t}\in A]$
in terms of $\mathbb{P}_{y}[X_{s}\in A,\,s<\tau_{U}]$, $(s,y)\in[0,t]\times U$,
for each Borel subset $A$ of $M$ with $\overline{A}\subset U$
(Theorem \ref{thm:multiple-DH}) and thus enabling us to deduce upper bounds for
the former from those for the latter together with \textup{(P)$_{\beta}^{U,R}$}
(Proposition \ref{prop:HKUB-localized-difference}).
Note that the case of bounded $(M,d)$ has been excluded from the main results
of \cite{AB,Gri:HKfractal,GriHu:HKGreen,GriHu:Upper,GriHuLau:GenCap,GT},
mainly due to their construction of the global heat kernel $p_{t}(x,y)$ as
the limit as $U\uparrow M$ of the Dirichlet heat kernel $p^{U}_{t}(x,y)$ on $U$;
indeed, taking the limit as $U\uparrow M$ is not allowed for bounded $(M,d)$
since part of their conditions \textup{(FK)$_{\Psi}$} (Faber-Krahn inequality)
and \textup{(E)$_{\Psi}$} (mean exit time estimate, see
\eqref{eq:exit-time-upper} and \eqref{eq:exit-time-lower}
in Theorem \ref{thm:exit-time-probability} below)
must fail when the ball $B(x,r)$ coincides with $M$.
We expect that this difficulty can be overcome by applying the main results of this
paper, so that their results should be easily extended to the case of bounded $(M,d)$.
In fact, Barlow, Bass, Kumagai and Teplyaev \cite{BBKT:supplement} have used
an argument very similar to our proof of Theorem \ref{thm:multiple-DH}
and Proposition \ref{prop:HKUB-localized-difference}
in \cite[Proof of Proposition 2.12]{BBKT:supplement}
for the resolvent of the diffusion to extend part of
the main results of \cite{GriHu:HKGreen,GT} to the case of bounded $(M,d)$.
Our proof of Theorem \ref{thm:HKUB-localized-intro} has successfully localized
their idea by working directly with the transition function (semigroup)
rather than the resolvent.
Finally, we remark that Theorem \ref{thm:HKUB-localized-intro} has been recently
applied in \cite{AK} to prove the continuity and sub-Gaussian off-diagonal
upper bounds of the heat kernel of the \emph{Liouville Brownian motion}, the
canonical diffusion in a certain random geometry of $\mathbb{R}^{2}$ induced by
a (massive) Gaussian free field. These results in \cite{AK} have had to rely strongly
on Theorem \ref{thm:HKUB-localized-intro} due to the fact that the unboundedness
of $\mathbb{R}^{2}$ precludes any uniform estimates of volumes and exit times
over the whole $\mathbb{R}^{2}$ valid for almost every environment, as opposed
to the case of the two-dimensional torus, where the same kind of results have
been obtained independently and simultaneously in \cite{MRVZ}.
The rest of this paper is organized as follows.
In Section \ref{sec:Hunt-processes}, we collect basic definitions and facts
concerning Hunt processes. Section \ref{sec:multiple-DH} formulates one of our
two iteration arguments as a multiple Dynkin-Hunt formula and proves it for an
\emph{arbitrary} Hunt process (Theorem \ref{thm:multiple-DH}).
In Section \ref{sec:Dirichlet-form}, we recall the notions of the symmetry of
a Hunt process, the associated symmetric Dirichlet form and its regularity,
together with some basic potential theory that is needed in
Section \ref{sec:HK-existence} to state and prove our main localized
existence theorem for the heat kernel (Theorem \ref{thm:HK-existence}).
In Section \ref{sec:HKUB} we state our main theorem on localized sub-Gaussian
upper bounds of heat kernels (Theorem \ref{thm:HKUB-localized}) and a global version
of it (Theorem \ref{thm:HKUB-localized-global}) and prove them on the basis of
our other probabilistic iteration (Proposition \ref{prop:HKUB-localized}) and
the multiple Dynkin-Hunt formula combined with the condition
\textup{(P)$_{\beta}^{U,R}$} (Proposition \ref{prop:HKUB-localized-difference}).
Lastly, Section \ref{sec:exit-probability} is devoted to providing sufficient
conditions for \textup{(P)$_{\beta}^{U,R}$}
(Theorems \ref{thm:exit-probability} and \ref{thm:exit-time-probability}) as a
localized version of the (well-)known results in \cite{Bar,Gri:HKfractal,GriHu:Upper}.
\begin{notation}
In this paper, we adopt the following notation and conventions.
\begin{itemize}[label=\textup{(0)},align=left,leftmargin=*]
\item[\textup{(0)}]The symbols $\subset$ and $\supset$ for set inclusion
\emph{allow} the case of the equality.
\item[\textup{(1)}]$\mathbb{N}=\{n\in\mathbb{Z}\mid n>0\}$, i.e., $0\not\in\mathbb{N}$.
\item[\textup{(2)}]We set $\sup\emptyset:=0$ and $\inf\emptyset:=\infty$. We write
$a\vee b:=\max\{a,b\}$, $a\wedge b:=\min\{a,b\}$, $a^{+}:=a\vee 0$ and
$a^{-}:=-(a\wedge 0)$ for $a,b\in[-\infty,\infty]$, and
we use the same notation also for $[-\infty,\infty]$-valued functions
and equivalence classes of them. All numerical functions treated in this paper
are assumed to be $[-\infty,\infty]$-valued.
\item[\textup{(3)}]Let $E$ be a topological space.
The Borel $\sigma$-field of $E$ is denoted by $\sigmafield{B}(E)$.
We set
\begin{align*}
C(E)&:=\{u\mid u:E\to\mathbb{R},\,u\textrm{ is continuous}\},\\
C_{\mathrm{c}}(E)&:=\{u\in C(E)
\mid\textrm{the closure of }u^{-1}(\mathbb{R}\setminus\{0\})\textrm{ in }E\textrm{ is compact}\},\\
\functionspace{B}(E)&:=\{u\mid u:E\to[-\infty,\infty],\,u\textrm{ is Borel measurable (i.e.,\ }\sigmafield{B}(E)\textrm{-measurable)}\},\\
\functionspace{B}^{+}(E)&:=\{u\in\functionspace{B}(E)\mid u\textrm{ is }[0,\infty]\textrm{-valued}\},\\
\functionspace{B}_{\mathrm{b}}(E)&:=\{u\in\functionspace{B}(E)\mid\|u\|_{\mathrm{sup}}<\infty\},
\end{align*}
where $\|u\|_{\mathrm{sup}}:=\|u\|_{\mathrm{sup},E}:=\sup_{x\in E}|u(x)|$ for $u:E\to[-\infty,\infty]$.
\end{itemize}
\end{notation}
\section{Basics on Hunt processes}\label{sec:Hunt-processes}
In this section, we introduce our framework of a Hunt process. To keep the main
results of this paper accessible to those who are not familiar with the theory
of Markov processes, we explain basic definitions and facts in some detail. See
\cite[Section A.2]{FOT} and \cite[Section A.1]{CF} for further details on Hunt processes.
Let $M$ be a locally compact separable metrizable topological space.
The interior, closure and boundary of $A\subset M$ in $M$ are denoted by
$\interior A$, $\overline{A}$ and $\partial A$, respectively.
Each $A\subset M$ is equipped with the relative topology inherited from $M$,
so that its Borel $\sigma$-field $\sigmafield{B}(A)$ can be expressed as
$\sigmafield{B}(A)=\{B\cap A\mid B\in\sigmafield{B}(M)\}$. Let
$M_{\cemetery}:=M\cup\{\cemetery\}$ denote the one-point compactification of $M$,
which satisfies
$\sigmafield{B}(M_{\cemetery})=\sigmafield{B}(M)\cup\{A\cup\{\cemetery\}\mid A\in\sigmafield{B}(M)\}$.
In what follows, $[-\infty,\infty]$-valued functions on $M$ are always set
to be $0$ at $\cemetery$ unless their values at $\Delta$ are already defined:
$u(\cemetery):=0$ for $u:M\to[-\infty,\infty]$.
Let
$X=\bigl(\Omega,\sigmafield{M},\{X_{t}\}_{t\in[0,\infty]},
\{\mathbb{P}_{x}\}_{x\in M_{\cemetery}}\bigr)$
be a Hunt process on $(M,\sigmafield{B}(M))$ with life time $\zeta$ and
shift operators $\{\theta_{t}\}_{t\in[0,\infty]}$.
By definition, $(\Omega,\sigmafield{M})$ is a measurable space,
$\{X_{t}\}_{t\in[0,\infty]}$ is a family of
$\sigmafield{M}/\sigmafield{B}(M_{\cemetery})$-measurable maps
$X_{t}:\Omega\to M_{\cemetery}$ such that
$X_{t}(\omega)=\cemetery$ for any $t\in[\zeta(\omega),\infty]$
for each $\omega\in\Omega$,
where $\zeta(\omega):=\inf\{t\in[0,\infty)\mid X_{t}(\omega)=\cemetery\}$,
and $\{\theta_{t}\}_{t\in[0,\infty]}$ is a family of maps
$\theta_{t}:\Omega\to\Omega$ satisfying
$X_{s}\circ\theta_{t}=X_{s+t}$ for any $s,t\in[0,\infty]$.
It is further assumed that for each $\omega\in\Omega$,
$[0,\infty)\ni t\mapsto X_{t}(\omega)\in M_{\cemetery}$ is right-continuous and
the limit $X_{t-}(\omega):=\lim_{s\to t,\,s<t}X_{s}(\omega)$ exists in $M_{\cemetery}$
for any $t\in(0,\infty)$; see \cite[Section A.2, (M.6)]{FOT}. The pair $X$
of such a stochastic process $\bigl(\Omega,\sigmafield{M},\{X_{t}\}_{t\in[0,\infty]}\bigr)$
and a family $\{\mathbb{P}_{x}\}_{x\in M_{\cemetery}}$ of probability measures
on $(\Omega,\sigmafield{M})$ is then called a
\emph{Hunt process on $(M,\sigmafield{B}(M))$}
if and only if it is a normal Markov process on
$(M,\sigmafield{B}(M))$ whose minimum completed admissible filtration
$\sigmafield{F}_{*}=\{\sigmafield{F}_{t}\}_{t\in[0,\infty]}$ is \emph{right-continuous}
and it is \emph{strong Markov} and \emph{quasi-left-continuous} with respect to
$\sigmafield{F}_{*}$; see
\cite[Section A.2, (M.2)--(M.5), the paragraph before Lemma A.2.2, (A.2.3) and (A.2.4)]{FOT}
for the precise definitions of these notions.
For $x\in M_{\cemetery}$, the expectation (that is, the integration on $\Omega$)
under the measure $\mathbb{P}_{x}$ is denoted by $\mathbb{E}_{x}[(\cdot)]$.
We remark that by \cite[Exercise A.1.20-(i)]{CF}, for each
$\sigmafield{F}_{\infty}$-measurable random variable $Y:\Omega\to[0,\infty]$
the function $M_{\cemetery}\ni x\mapsto\mathbb{E}_{x}[Y]\in[0,\infty]$ is
\emph{universally measurable}, i.e., measurable with respect to the
\emph{universal $\sigma$-field $\sigmafield{B}^{*}(M_{\cemetery})$ of $M_{\cemetery}$} defined as
$\sigmafield{B}^{*}(M_{\cemetery}):=\bigcap_{\nu}\sigmafield{B}^{\nu}(M_{\cemetery})$;
here $\nu$ runs through the set of probability (or equivalently, $\sigma$-finite)
measures on $(M_{\cemetery},\sigmafield{B}(M_{\cemetery}))$ and
$\sigmafield{B}^{\nu}(M_{\cemetery})$ denotes the $\nu$-completion
of $\sigmafield{B}(M_{\cemetery})$.
The Hunt process $X$ gives rise to a family $\{\functionspace{P}_{t}\}_{t\in[0,\infty)}$
of Markovian kernels on $(M,\sigmafield{B}(M))$ called the
\emph{transition function of $X$}, which is defined by
\begin{equation}\label{eq:transition-function}
\functionspace{P}_{t}(x,A):=\mathbb{P}_{x}[X_{t}\in A]
\qquad t\in[0,\infty),\ x\in M,\ A\in\sigmafield{B}(M).
\end{equation}
Then for $t\in[0,\infty)$ and $u\in\functionspace{B}(M)$, we define
\begin{equation}\label{eq:PtEx}
\functionspace{P}_{t}u(x):=\int_{M}u(y)\functionspace{P}_{t}(x,dy)=\mathbb{E}_{x}[u(X_{t})]
\end{equation
for $x\in M$ satisfying $\mathbb{E}_{x}[u^{+}(X_{t})]\wedge\mathbb{E}_{x}[u^{-}(X_{t})]<\infty$,
so that $\functionspace{P}_{t}(\functionspace{B}^{+}(M))\subset\functionspace{B}^{+}(M)$ and
$\functionspace{P}_{t}(\functionspace{B}_{\mathrm{b}}(M))\subset\functionspace{B}_{\mathrm{b}}(M)$.
Note that our convention of setting $\functionspace{P}_{t}u(\cemetery):=0$ is
consistent with \eqref{eq:PtEx} for $x=\cemetery$ since
$\mathbb{E}_{\cemetery}[u(X_{t})]=\mathbb{E}_{\cemetery}[u(\cemetery)]=0$
by $\mathbb{P}_{\cemetery}[X_{t}=\cemetery]=1$.
Obviously, if $u\in\functionspace{B}(M)$ is $[0,1]$-valued then so is
$\functionspace{P}_{t}u$, and the Markov property of $X$
(see \cite[(A.2.2)]{FOT} or \cite[(A.1.3)]{CF}) easily implies the
semigroup property
\begin{equation}\label{eq:Pt-semigroup}
\functionspace{P}_{t}\functionspace{P}_{s}u=\functionspace{P}_{t+s}u,
\qquad t,s\in[0,\infty),\ u\in\functionspace{B}^{+}(M)\cup\functionspace{B}_{\mathrm{b}}(M).
\end{equation}
Moreover,
it easily follows from the sample path right-continuity of $X$ and
the Dynkin class theorem \cite[Chapter 0, Theorem 2.2]{BG} that
\begin{equation}\label{eq:Pt-measurable}
[0,\infty)\times M\ni(t,x)\mapsto\functionspace{P}_{t}u(x)
\textrm{ is Borel measurable for any }u\in\functionspace{B}^{+}(M)\cup\functionspace{B}_{\mathrm{b}}(M).
\end{equation}
Recall that $\sigma:\Omega\to[0,\infty]$ is called an
\emph{$\sigmafield{F}_{*}$-stopping time} if and only if
$\{\sigma\leq t\}\in\sigmafield{F}_{t}$ for any $t\in[0,\infty)$.
For $B\subset M_{\cemetery}$, we define its \emph{entrance time $\dot{\sigma}_{B}$}
and \emph{exit time $\tau_{B}$ for $X$} by
\begin{equation}\label{eq:entrance-exit}
\dot{\sigma}_{B}(\omega):=\inf\{t\in[0,\infty)\mid X_{t}(\omega)\in B\},
\quad\omega\in\Omega\qquad\textrm{and}\qquad
\tau_{B}:=\dot{\sigma}_{M_{\cemetery}\setminus B},
\end{equation}
and we also set $\hat{\sigma}_{B}(\omega):=\inf\{t\in(0,\infty)\mid X_{t-}(\omega)\in B\}$
for $\omega\in\Omega$. If $B\in\sigmafield{B}(M_{\cemetery})$, then
$\dot{\sigma}_{B},\tau_{B},\hat{\sigma}_{B}$ are
$\sigmafield{F}_{*}$-stopping times and
\begin{equation}\label{eq:entrance-leq-entrance-left-limit}
\mathbb{P}_{x}[\dot{\sigma}_{B}\leq\hat{\sigma}_{B}]=1
\qquad\textrm{for any }x\in M_{\cemetery}
\end{equation}
by \cite[Theorem A.2.3]{FOT}, where the case of $\cemetery\in B$ is easily
deduced from that of $B\in\sigmafield{B}(M)$ by using the equalities
$\dot{\sigma}_{B\cup\{\cemetery\}}=\dot{\sigma}_{B}\wedge\zeta$ and
$\hat{\sigma}_{B\cup\{\cemetery\}}=\hat{\sigma}_{B}\wedge\hat{\sigma}_{\{\cemetery\}}$
for $B\subset M$ and the quasi-left-continuity \cite[(A.2.4)]{FOT} of $X$
(see also \cite[Theorem A.1.19 and Exercise A.1.26-(ii)]{CF}).
Note that if $B\subset M_{\cemetery}$, $t\in[0,\infty]$ and
$\omega\in\{\dot{\sigma}_{B}\geq t\}$ then
$\dot{\sigma}_{B}(\omega)=t+\dot{\sigma}_{B}(\theta_{t}(\omega))$.
Next we introduce the part of $X$ on open sets. Let $U$ be a non-empty open
subset of $M$, let $U_{\cemetery}:=U\cup\{\cemetery_{U}\}$ denote its one-point
compactification and define
\begin{equation}\label{eq:part-process}
X^{U}_{t}(\omega):=
\begin{cases}
X_{t}(\omega)&\textrm{if }t<\tau_{U}(\omega),\\
\cemetery_{U}&\textrm{if }t\geq\tau_{U}(\omega),\\
\end{cases}
\qquad(t,\omega)\in[0,\infty]\times\Omega
\end{equation}
and $\mathbb{P}_{\cemetery_{U}}:=\mathbb{P}_{\cemetery}$. Then
$X^{U}:=\bigl(\Omega,\sigmafield{M},\{X^{U}_{t}\}_{t\in[0,\infty]},\{\mathbb{P}_{x}\}_{x\in U_{\cemetery}}\bigr)$,
called the \emph{part of $X$ on $U$}, is a Hunt process on $(U,\sigmafield{B}(U))$
by \cite[Theorem A.2.10]{FOT}. Its transition function is naturally extended to
$(M,\sigmafield{B}(M))$ as a family $\{\functionspace{P}^{U}_{t}\}_{t\in[0,\infty)}$
of Markovian kernels on $(M,\sigmafield{B}(M))$ given by
(with the obvious convention that $\cemetery_{U}\not\in M$)
\begin{equation}\label{eq:transition-function-part}
\functionspace{P}^{U}_{t}(x,A):=\mathbb{P}_{x}[X^{U}_{t}\in A]
=\mathbb{P}_{x}[X_{t}\in A,\,t<\tau_{U}],
\quad t\in[0,\infty),\ x\in M,\ A\in\sigmafield{B}(M).
\end{equation}
Also for $t\in[0,\infty)$ and $u\in\functionspace{B}(M)$,
similarly to \eqref{eq:PtEx} we further define
\begin{equation}\label{eq:PtEx-part}
\functionspace{P}^{U}_{t}u(x):=\int_{M}u(y)\functionspace{P}^{U}_{t}(x,dy)
=\int_{U}u(y)\functionspace{P}^{U}_{t}(x,dy)
=\mathbb{E}_{x}[u(X_{t})\ind{\{t<\tau_{U}\}}]
\end{equation}
for $x\in M$ satisfying
$\mathbb{E}_{x}[u^{+}(X_{t})\ind{\{t<\tau_{U}\}}]\wedge\mathbb{E}_{x}[u^{-}(X_{t})\ind{\{t<\tau_{U}\}}]<\infty$,
where $\ind{A}:\Omega\to\{0,1\}$ denotes the indicator function of $A\subset\Omega$
given by $\ind{A}|_{A}:=1$ and $\ind{A}|_{\Omega\setminus A}:=0$. Then
$\functionspace{P}^{U}_{t}u(x)=0$ for $x\in M\setminus U$,
$\functionspace{P}^{U}_{t}(\functionspace{B}^{+}(M))\subset\functionspace{B}^{+}(M)$,
$\functionspace{P}^{U}_{t}(\functionspace{B}_{\mathrm{b}}(M))\subset\functionspace{B}_{\mathrm{b}}(M)$,
and \eqref{eq:Pt-semigroup} and \eqref{eq:Pt-measurable} hold
with $\{\functionspace{P}^{U}_{t}\}_{t\in[0,\infty)}$
in place of $\{\functionspace{P}_{t}\}_{t\in[0,\infty)}$.
\section{A multiple Dynkin-Hunt formula for Hunt processes}\label{sec:multiple-DH}
As in Section \ref{sec:Hunt-processes}, let $M$ be a locally compact separable
metrizable topological space and let $X$ be a Hunt process on $(M,\sigmafield{B}(M))$
with life time $\zeta$ and shift operators $\{\theta_{t}\}_{t\in[0,\infty]}$.
\emph{Throughout the rest of this paper, we fix this setting and follow
the notation introduced in Section \textup{\ref{sec:Hunt-processes}}}.
In this section, we state and prove a \emph{multiple Dynkin-Hunt formula}
(Theorem \ref{thm:multiple-DH} below) which gives an expression of
$\functionspace{P}_{t}u$ in terms of $\functionspace{P}^{U}_{s}u$, $s\in[0,t]$,
for a non-empty open subset $U$ of $M$ and functions
$u\in\functionspace{B}^{+}(M)\cup\functionspace{B}_{\mathrm{b}}(M)$
supported in $U$. It will be used later in Section \ref{sec:HKUB}
to deduce upper bounds for $\{\functionspace{P}_{t}\}_{t\in(0,\infty)}$
from those for $\{\functionspace{P}^{U}_{t}\}_{t\in(0,\infty)}$.
The statement of Theorem \ref{thm:multiple-DH} requires the following definition and proposition.
\begin{definition}\label{dfn:entrance-exit-after-sigma}
For $\sigma:\Omega\to[0,\infty]$ and $B\subset M_{\cemetery}$,
the \emph{entrance time $\dot{\sigma}_{B,\sigma}$} and
\emph{exit time $\tau_{B,\sigma}$ of $B$ after $\sigma$ for $X$}
are defined by (with the convention that $[\infty,\infty):=\emptyset$)
\begin{equation}\label{eq:entrance-exit-after-sigma}
\dot{\sigma}_{B,\sigma}(\omega):=\inf\{t\in[\sigma(\omega),\infty)\mid X_{t}(\omega)\in B\},
\quad\omega\in\Omega\quad\textrm{and}\quad
\tau_{B,\sigma}:=\dot{\sigma}_{M_{\cemetery}\setminus B,\sigma},
\end{equation}
so that
$\dot{\sigma}_{B,\sigma}(\omega)=\sigma(\omega)+\dot{\sigma}_{B}(\theta_{\sigma(\omega)}(\omega))$
and
$\tau_{B,\sigma}(\omega)=\sigma(\omega)+\tau_{B}(\theta_{\sigma(\omega)}(\omega))$
for any $\omega\in\Omega$.
\end{definition}
\begin{proposition}\label{prop:entrance-exit-after-sigma}
For any $\sigmafield{F}_{*}$-stopping time $\sigma$ and any
$B\in\sigmafield{B}(M_{\cemetery})$,
the entrance time $\dot{\sigma}_{B,\sigma}$ and exit time $\tau_{B,\sigma}$ of
$B$ after $\sigma$ for $X$ are $\sigmafield{F}_{*}$-stopping times.
\end{proposition}
\begin{proof}
This proposition should be well-known, but we give an explicit proof for completeness.
We follow \cite[Proof of Theorem A.1.19]{CF}. For each $t\in(0,\infty)$, the set
$\{\dot{\sigma}_{B,\sigma}<t\}=\{\omega\in\Omega\mid\dot{\sigma}_{B,\sigma}(\omega)<t\}$
is equal to the projection on $\Omega$ of
\begin{equation*}
\{(s,\omega)\in[0,t)\times\Omega\mid\sigma(\omega)\leq s,\,X_{s}(\omega)\in B\},
\end{equation*}
which is easily shown to belong to the product $\sigma$-field
$\sigmafield{B}([0,t])\otimes\sigmafield{F}_{t}$ by the sample path
right-continuity of $X$ and the assumption that
$\sigma$ is an $\sigmafield{F}_{*}$-stopping time.
Therefore \cite[Chapter III, 13 and 33]{DM} imply that
$\{\dot{\sigma}_{B,\sigma}<t\}\in\sigmafield{F}_{t}$,
which means that $\dot{\sigma}_{B,\sigma}$, and hence also
$\tau_{B,\sigma}$, are $\sigmafield{F}_{*}$-stopping times
since $\sigmafield{F}_{*}$ is right-continuous.
\end{proof}
Now we state the main theorem of this section.
Recall for $\sigma:\Omega\to[0,\infty]$ that the map
$X_{\sigma}:\Omega\to M_{\cemetery}$ is defined as
$X_{\sigma}(\omega):=X_{\sigma(\omega)}(\omega)$ and that $X_{\sigma}$ is
$\sigmafield{F}_{\infty}/\sigmafield{B}(M_{\cemetery})$-measurable
if $\sigma$ is $\sigmafield{F}_{\infty}$-measurable
by the sample path right-continuity of $X$.
\begin{theorem}[A multiple Dynkin-Hunt formula]\label{thm:multiple-DH}
Let $U$ be a non-empty open subset of $M$, let $B\in\sigmafield{B}(M)$ satisfy
$\overline{B}\subset U$ and define $\sigmafield{F}_{*}$-stopping times
$\tau_{n}$ and $\sigma_{n}$, $n\in\mathbb{N}$, by
\begin{equation}\label{eq:multiple-DH-stopping-times}
\tau_{1}:=\tau_{U}\qquad\textrm{and inductively}\qquad
\sigma_{n}:=\dot{\sigma}_{B,\tau_{n}}\quad\textrm{and}\quad
\tau_{n+1}:=\tau_{U,\sigma_{n}},\quad n\in\mathbb{N}.
\end{equation}
Then for any $u\in\functionspace{B}^{+}(M)\cup\functionspace{B}_{\mathrm{b}}(M)$
with $u|_{M\setminus B}=0$ and any $(t,x)\in[0,\infty)\times M$,
\begin{equation}\label{eq:multiple-DH}
\functionspace{P}_{t}u(x)=\functionspace{P}^{U}_{t}u(x)
+\sum_{n\in\mathbb{N}}\mathbb{E}_{x}\bigl[\ind{\{\sigma_{n}\leq t\}}
\functionspace{P}^{U}_{t-\sigma_{n}}u(X_{\sigma_{n}})\bigr].
\end{equation}
\end{theorem}
Note that by \eqref{eq:Pt-measurable} for
$\{\functionspace{P}^{U}_{t}\}_{t\in[0,\infty)}$, the random variable
$\ind{\{\sigma_{n}\leq t\}}\functionspace{P}^{U}_{t-\sigma_{n}}u(X_{\sigma_{n}})$
in \eqref{eq:multiple-DH} is $\sigmafield{F}_{\infty}$-measurable
for any $u\in\functionspace{B}^{+}(M)\cup\functionspace{B}_{\mathrm{b}}(M)$,
any $t\in[0,\infty)$ and any $n\in\mathbb{N}$.
Recall that the \emph{Dynkin-Hunt formula} refers to
(the heat kernel version of) the following equality,
which is an easy consequence of Proposition \ref{prop:strong-Markov} below:
for any non-empty open subset $U$ of $M$,
any $u\in\functionspace{B}^{+}(M)\cup\functionspace{B}_{\mathrm{b}}(M)$
and any $(t,x)\in[0,\infty)\times M$,
\begin{equation}\label{eq:DH}
\functionspace{P}_{t}u(x)=\functionspace{P}^{U}_{t}u(x)
+\mathbb{E}_{x}\bigl[\ind{\{\tau_{U}\leq t\}}
\functionspace{P}_{t-\tau_{U}}u(X_{\tau_{U}})\bigr].
\end{equation}
\eqref{eq:multiple-DH} can be regarded as an indefinite iteration of
\eqref{eq:DH} through restarting $X$ at the entrance time
$\dot{\sigma}_{B,\tau_{U}}$ of $B$ after $\tau_{U}$, which is why we call
\eqref{eq:multiple-DH} a \emph{multiple Dynkin-Hunt formula}.
For the proof of Theorem \ref{thm:multiple-DH} we need a variation of
the strong Markov property of $X$ as in the following proposition.
Recall for each $\sigmafield{F}_{*}$-stopping time $\sigma$ that the collection
\begin{equation}\label{eq:Fsigma}
\sigmafield{F}_{\sigma}:=\{A\in\sigmafield{F}_{\infty}\mid
A\cap\{\sigma\leq t\}\in\sigmafield{F}_{t}\textrm{ for any }t\in[0,\infty)\}
\end{equation}
is a $\sigma$-field in $\Omega$ with respect to which $\sigma$ is measurable,
that $X_{\sigma}$ is $\sigmafield{F}_{\sigma}/\sigmafield{B}^{*}(M_{\cemetery})$-measurable
by \cite[Exercise A.1.20-(ii)]{CF}, and that the map $\theta_{\sigma}:\Omega\to\Omega$,
$\theta_{\sigma}(\omega):=\theta_{\sigma(\omega)}(\omega)$, is
$\sigmafield{F}_{\infty}/\sigmafield{F}_{\infty}$-measurable by \cite[Theorem A.1.21]{CF}.
\begin{proposition}\label{prop:strong-Markov}
Let $\sigma$ be an $\sigmafield{F}_{*}$-stopping time, let $\tau:\Omega\to[0,\infty]$
be $\sigmafield{F}_{\infty}$-measurable and let $T:\Omega\to[0,\infty]$ be
$\sigmafield{F}_{\sigma}$-measurable and satisfy $\sigma(\omega)\leq T(\omega)$
for any $\omega\in\Omega$. Then for any $x\in M_{\cemetery}$ and any
$u\in\functionspace{B}_{\mathrm{b}}(M_{\cemetery})$,
it holds that for $\mathbb{P}_{x}$-a.e.\ $\omega\in\{\sigma<\infty\}$,
\begin{equation}\label{eq:strong-Markov}
\mathbb{E}_{x}\bigl[u(X_{T})\ind{\{T<\sigma+\tau\circ\theta_{\sigma}\}}\bigm|\sigmafield{F}_{\sigma}\bigr](\omega)
=\mathbb{E}_{X_{\sigma}(\omega)}\bigl[u(X_{T(\omega)-\sigma(\omega)})\ind{\{T(\omega)-\sigma(\omega)<\tau\}}\bigr].
\end{equation}
\end{proposition}
\begin{proof}
We follow \cite[Proofs of Proposition 2.6.17 and Corollary 2.6.18]{KS}.
For $u\in\sigmafield{B}_{\mathrm{b}}(M_{\cemetery})$, let
$Y_{u}(\omega)$ denote the right-hand side of \eqref{eq:strong-Markov} for
$\omega\in\{\sigma<\infty\}$ and set $Y_{u}(\omega):=0$ for $\omega\in\{\sigma=\infty\}$.
Let $x\in M_{\cemetery}$. For the proof of \eqref{eq:strong-Markov} it suffices
to show that $Y_{u}:\Omega\to\mathbb{R}$ possesses the following properties:
\begin{equation}\label{eq:strong-Markov-proof}
Y_{u}\textrm{ is }\sigmafield{F}_{\sigma}\textrm{-measurable}\mspace{15mu}\textrm{and}\mspace{15mu}
\mathbb{E}_{x}\bigl[u(X_{T})\ind{\{T<\sigma+\tau\circ\theta_{\sigma}\}}\ind{A}\bigr]
=\mathbb{E}_{x}[Y_{u}\ind{A}]
\textrm{ for any }A\in\sigmafield{F}_{\sigma}.
\end{equation}
We first prove \eqref{eq:strong-Markov-proof} for $u\in C(M_{\cemetery})$.
Let $n\in\mathbb{N}$ and define $T_{n}:\Omega\to[0,\infty]$ by
\begin{equation}\label{eq:strong-Markov-Tn}
T_{n}|_{\{\sigma+(k-1)2^{-n}\leq T<\sigma+k2^{-n}\}}:=\sigma+k2^{-n},
\quad k\in\mathbb{N}\qquad\textrm{and}\qquad T_{n}|_{\{T=\infty\}}:=\infty,
\end{equation}
so that $T_{n}$ is $\sigmafield{F}_{\sigma}$-measurable and
$T_{n}-2^{-n}\leq T\leq T_{n}$. Also define $Y_{u,n}$
in the same way as $Y_{u}$ with $T_{n}$ in place of $T$.
Then $Y_{u,n}|_{\{T=\infty\}}=0=Y_{u}|_{\{T=\infty\}}$, and
$\lim_{n\to\infty}Y_{u,n}=Y_{u}$ on $\{T<\infty\}$ by
$T_{n}-2^{-n}\leq T\leq T_{n}$, the sample path right-continuity of $X$
and dominated convergence. Also for $k\in\mathbb{N}$,
on $\{\sigma+(k-1)2^{-n}\leq T<\sigma+k2^{-n}\}\in\sigmafield{F}_{\sigma}$ we have
$Y_{u,n}=\mathbb{E}_{X_{\sigma}}[u(X_{k2^{-n}})\ind{\{k2^{-n}<\tau\}}]$,
and since the latter is $\sigmafield{F}_{\sigma}$-measurable by
\cite[Exercise A.1.20]{CF} so are $Y_{u,n}$ and $Y_{u}=\lim_{n\to\infty}Y_{u,n}$.
Now for $A\in\sigmafield{F}_{\sigma}$, thanks to dominated convergence,
\begin{align*}
\mathbb{E}_{x}\bigl[u(X_{T_{n}})\ind{\{T_{n}<\sigma+\tau\circ\theta_{\sigma}\}}\ind{A}\bigr]
&=\sum_{k\in\mathbb{N}}\mathbb{E}_{x}\bigl[\ind{A\cap\{T_{n}=\sigma+k2^{-n}<\infty\}}\bigl((u(X_{k2^{-n}})\ind{\{k2^{-n}<\tau\}})\circ\theta_{\sigma}\bigr)\bigr]\\
&=\sum_{k\in\mathbb{N}}\mathbb{E}_{x}\bigl[\ind{A\cap\{T_{n}=\sigma+k2^{-n}<\infty\}}\mathbb{E}_{X_{\sigma}}[u(X_{k2^{-n}})\ind{\{k2^{-n}<\tau\}}]\bigr]\\
&=\mathbb{E}_{x}[Y_{u,n}\ind{A}]
\end{align*}
by the strong Markov property \cite[Theorem A.1.21]{CF} of $X$ at time $\sigma$,
and we conclude \eqref{eq:strong-Markov-proof} by using $T_{n}-2^{-n}\leq T\leq T_{n}$
and the sample path right-continuity of $X$ to let $n\to\infty$.
Note that for $u\in\functionspace{B}_{\mathrm{b}}(M_{\cemetery})$ and
$\{u_{n}\}_{n\in\mathbb{N}}\subset\functionspace{B}_{\mathrm{b}}(M_{\cemetery})$
such that $\sup_{n\in\mathbb{N}}\|u_{n}\|_{\mathrm{sup}}<\infty$ and
$\lim_{n\to\infty}u_{n}(y)=u(y)$ for any $y\in M_{\cemetery}$,
if $u_{n}$ satisfies \eqref{eq:strong-Markov-proof} for any $n\in\mathbb{N}$
then so does $u$ by dominated convergence. Therefore it follows from
the previous paragraph that \eqref{eq:strong-Markov-proof} holds for
$u=\ind{B}$ with $B\subset M_{\cemetery}$ closed in $M_{\cemetery}$, hence
also with $B\in\sigmafield{B}(M_{\cemetery})$ by the Dynkin class theorem
\cite[Chapter 0, Theorem 2.2]{BG}, and thus for any
$u\in\functionspace{B}_{\mathrm{b}}(M_{\cemetery})$.
\end{proof}
\begin{proof}[Proof of Theorem \textup{\ref{thm:multiple-DH}}]
For $n\in\mathbb{N}$, $\tau_{n}\leq\sigma_{n}\leq\tau_{n+1}$ by
\eqref{eq:entrance-exit-after-sigma} and \eqref{eq:multiple-DH-stopping-times},
and the sample path right-continuity of $X$ implies that
$X_{\tau_{n}}\in M\setminus U$ and $\tau_{n}<\sigma_{n}$ on $\{\tau_{n}<\zeta\}$ and that
$X_{\sigma_{n}}\in\overline{B}$ and $\sigma_{n}<\tau_{n+1}\wedge\zeta$ on $\{\sigma_{n}<\infty\}$.
Moreover, setting $\tau:=\lim_{n\to\infty}\tau_{n}=\lim_{n\to\infty}\sigma_{n}$,
we see from the quasi-left-continuity \cite[(A.2.4)]{FOT} of $X$ that
for any $x\in M$,
\begin{equation}\label{eq:tau-sigma-lim}
\mathbb{P}_{x}[\tau<\zeta]
=\mathbb{P}_{x}[\tau<\zeta,\,\lim\nolimits_{n\to\infty}X_{\tau_{n}}=X_{\tau}=\lim\nolimits_{n\to\infty}X_{\sigma_{n}}]
=\mathbb{P}_{x}[\emptyset]=0.
\end{equation}
Let $(t,x)\in[0,\infty)\times M$.
Then for each $\omega\in\{\textrm{$X_{t}\in B$, $\zeta\leq\tau$}\}$,
$t<\zeta(\omega)\leq\tau(\omega)$ and hence either $t<\tau_{1}(\omega)$,
or $\tau_{n}(\omega)\leq t<\tau_{n+1}(\omega)$ for some $n\in\mathbb{N}$, whence
$\sigma_{n}(\omega)\leq t<\tau_{n+1}(\omega)$ by $X_{t}(\omega)\in B$; namely
$\{\textrm{$X_{t}\in B$, $\zeta\leq\tau$}\}\subset
\{t<\tau_{1}\}\cup\bigcup_{n\in\mathbb{N}}\{\sigma_{n}\leq t<\tau_{n+1}\}$,
and this union is disjoint. Therefore for any
$u\in\functionspace{B}_{\mathrm{b}}(M)$ with $u|_{M\setminus B}=0$, noting that
$\tau_{n+1}=\sigma_{n}+\tau_{U}\circ\theta_{\sigma_{n}}$ for any $n\in\mathbb{N}$
and using \eqref{eq:tau-sigma-lim}, dominated convergence and
Proposition \ref{prop:strong-Markov}, we obtain
\begin{align*}
\functionspace{P}_{t}u(x)&=\mathbb{E}_{x}[u(X_{t})]
=\mathbb{E}_{x}[u(X_{t})\ind{\{X_{t}\in B,\,\zeta\leq\tau\}}]\\
&=\mathbb{E}_{x}\biggl[u(X_{t})\ind{\{X_{t}\in B,\,\zeta\leq\tau\}}\biggl(\ind{\{t<\tau_{1}\}}
+\sum_{n\in\mathbb{N}}\ind{\{\sigma_{n}\leq t<\tau_{n+1}\}}\biggr)\biggr]\\
&=\mathbb{E}_{x}[u(X_{t})\ind{\{t<\tau_{U}\}}]
+\sum_{n\in\mathbb{N}}\mathbb{E}_{x}[u(X_{t})\ind{\{\sigma_{n}\leq t<\sigma_{n}+\tau_{U}\circ\theta_{\sigma_{n}}\}}]\\
&=\functionspace{P}^{U}_{t}u(x)
+\sum_{n\in\mathbb{N}}\mathbb{E}_{x}\bigl[\ind{\{\sigma_{n}\leq t\}}\mathbb{E}_{x}\bigl[u(X_{t})\ind{\{t<\sigma_{n}\wedge t+\tau_{U}\circ\theta_{\sigma_{n}\wedge t}\}}\bigm|\sigmafield{F}_{\sigma_{n}\wedge t}\bigr]\bigr]\\
&=\functionspace{P}^{U}_{t}u(x)
+\sum_{n\in\mathbb{N}}\int_{\{\sigma_{n}\leq t\}}\mathbb{E}_{X_{\sigma_{n}\wedge t}(\omega)}\bigl[u(X_{t-\sigma_{n}(\omega)\wedge t})\ind{\{t-\sigma_{n}(\omega)\wedge t<\tau_{U}\}}\bigr]\,d\mathbb{P}_{x}(\omega)\\
&=\functionspace{P}^{U}_{t}u(x)
+\sum_{n\in\mathbb{N}}\mathbb{E}_{x}\bigl[\ind{\{\sigma_{n}\leq t\}}\functionspace{P}^{U}_{t-\sigma_{n}}u(X_{\sigma_{n}})\bigr],
\end{align*}
where the equality in the fourth line holds since $\{\sigma_{n}\leq t\}\in\sigmafield{F}_{\sigma_{n}\wedge t}$
by \cite[Lemma 1.2.16]{KS}. Thus we have proved \eqref{eq:multiple-DH} for
$u\in\functionspace{B}_{\mathrm{b}}(M)$ with $u|_{M\setminus B}=0$,
which easily implies \eqref{eq:multiple-DH} for $u\in\functionspace{B}^{+}(M)$
with $u|_{M\setminus B}=0$ by monotone convergence.
\end{proof}
\section{Symmetry of a Hunt process and the associated Dirichlet form}\label{sec:Dirichlet-form}
In this section, assuming the symmetry of our Hunt process $X$, we first recall
that such $X$ naturally gives rise to a symmetric Dirichlet form, and then
introduce related potential theoretic notions.
We refer the reader to \cite{FOT,CF} for further details.
\subsection{The Dirichlet form of a symmetric Hunt process}\label{ssec:Dirichlet-form}
In the rest of this paper, we fix a metric $d$ on $M$ compatible with
the topology of $M$, and a Radon measure $\mu$ on $M$ with full support,
i.e., a Borel measure on $M$ such that $\mu(K)<\infty$ for any
$K\subset M$ compact and $\mu(U)>0$ for any $U\subset M$ non-empty open.
We set $B(x,r):=\{y\in M\mid d(x,y)<r\}$ for $(x,r)\in M\times(0,\infty)$
and $\diam A:=\sup_{x,y\in A}d(x,y)$ for $A\subset M$.
For $q\in[1,\infty)$, we set
$\|u\|_{q}:=(\int_{M}|u|^{q}d\mu)^{1/q}$ for $u\in\functionspace{B}(M)$ and
$\functionspace{B}L^{q}(M,\mu):=\{u\in\functionspace{B}(M)\mid\|u\|_{q}<\infty\}$,
and we also set $\langle u,v\rangle:=\int_{M}uv\,d\mu$ for
$u,v\in\functionspace{B}^{+}(M)$ and for $u,v\in\functionspace{B}(M)$
with $\|uv\|_{1}<\infty$. For $\|\cdot\|_{q}$ and $\langle\cdot,\cdot\rangle$,
we use the same notation for $\mu$-equivalence classes of functions as well.
Now we assume that $X$ is \emph{$\mu$-symmetric},
i.e., $\langle\functionspace{P}_{t}u,v\rangle=\langle u,\functionspace{P}_{t}v\rangle$
for any $t\in(0,\infty)$ and any $u,v\in\functionspace{B}^{+}(M)$.
Then for each $t\in(0,\infty)$, as in \cite[(1.4.13)]{FOT} we can easily verify that
$\|\functionspace{P}_{t}u\|_{2}\leq\|u\|_{2}$ for any $u\in\functionspace{B}^{+}(M)$,
so that $\functionspace{P}_{t}u$ is defined $\mu$-a.e.\ and determines an
element $T_{t}u$ of $L^{2}(M,\mu)$ for each $u\in L^{2}(M,\mu)$ independently of a
particular choice of a $\mu$-version of $u$. Thus the transition function
$\{\functionspace{P}_{t}\}_{t\in[0,\infty)}$ of $X$ canonically induces a
symmetric contraction semigroup $\{T_{t}\}_{t\in(0,\infty)}$ on $L^{2}(M,\mu)$
which is also \emph{Markovian}, i.e., $0\leq T_{t}u\leq 1$ $\mu$-a.e.\ for any
$t\in(0,\infty)$ and any $u\in L^{2}(M,\mu)$ with $0\leq u\leq 1$ $\mu$-a.e.
This semigroup $\{T_{t}\}_{t\in(0,\infty)}$ is
in fact strongly continuous thanks to the sample path right-continuity of $X$
as shown in \cite[Lemma 1.4.3-(i)]{FOT} and hence determines a symmetric
Dirichlet form $(\functionspace{E},\functionspace{F})$ on $L^{2}(M,\mu)$
by \cite[Lemma 1.3.4-(i) and Theorem 1.4.1]{FOT}. Namely, we have a dense linear
subspace $\functionspace{F}$ of $L^{2}(M,\mu)$ and a non-negative definite symmetric
bilinear form $\functionspace{E}:\functionspace{F}\times\functionspace{F}\to\mathbb{R}$
given by
\begin{equation}\label{eq:Dirichlet-form}
\begin{split}
\functionspace{F}&:=\Bigl\{u\in L^{2}(M,\mu)\Bigm|\lim_{t\downarrow 0}t^{-1}\langle u-T_{t}u,u\rangle<\infty\Bigr\},\\
&\functionspace{E}(u,v):=\lim_{t\downarrow 0}t^{-1}\langle u-T_{t}u,v\rangle,\qquad u,v\in\functionspace{F},
\end{split}
\end{equation}
respectively, and $(\functionspace{E},\functionspace{F})$ is \emph{closed}
(i.e., $\functionspace{F}$ forms a Hilbert space with inner product
$\functionspace{E}_{1}:=\functionspace{E}+\langle\cdot,\cdot\rangle$) and
\emph{Markovian} (i.e., $u^{+}\wedge 1\in\functionspace{F}$ and
$\functionspace{E}(u^{+}\wedge 1,u^{+}\wedge 1)\leq\functionspace{E}(u,u)$
for any $u\in\functionspace{F}$). $(\functionspace{E},\functionspace{F})$
is called the \emph{Dirichlet form of the $\mu$-symmetric Hunt process $X$}.
Note that by \cite[Lemma 1.3.3-(i)]{FOT},
\begin{equation}\label{eq:TtL2-F}
T_{t}(L^{2}(M,\mu))\subset\functionspace{F}\qquad\textrm{for any }t\in(0,\infty).
\end{equation}
In what follows we further assume that the Dirichlet form
$(\functionspace{E},\functionspace{F})$ of $X$ is \emph{regular on $L^{2}(M,\mu)$},
i.e., that $\functionspace{F}\cap C_{\mathrm{c}}(M)$ is dense
both in $(\functionspace{F},\functionspace{E}_{1})$
and in $(C_{\mathrm{c}}(M),\|\cdot\|_{\mathrm{sup}})$.
Note that this framework actually contains any regular symmetric Dirichlet form
on any locally compact separable metric space $(M,d)$ equipped with a Radon measure
$\mu$ with full support, since any such form can be realized as the Dirichlet
form of some $\mu$-symmetric Hunt process on $(M,\sigmafield{B}(M))$
by the fundamental result \cite[Theorem 7.2.1]{FOT} from Dirichlet form theory.
\subsection{Capacity, quasi-continuity and exceptional sets}\label{ssec:capacity}
The following potential theoretic notions are adopted from
\cite[Section 2.1]{FOT} and \cite[Sections 1.2 and 1.3]{CF}.
\begin{definition}\label{dfn:capacity}
\begin{itemize}[label=\textup{(1)},align=left,leftmargin=*]
\item[\textup{(1)}]We define the \emph{$1$-capacity $\Capa_{1}$ associated with $(M,\mu,\functionspace{E},\functionspace{F})$} by
\begin{align}
\capa_{1}(U)
&:=\inf\{\functionspace{E}_{1}(u,u)\mid u\in\functionspace{F},\ u\geq 1\ \mu\textrm{-a.e.\ on }U\}
&&\textrm{for }U\subset M\textrm{ open in }M,\notag\\
\Capa_{1}(A)
&:=\inf\{\capa_{1}(U)\mid U\subset M\textrm{ open in }M,\ A\subset U\}
&&\textrm{for }A\subset M
\label{eq:capacity}
\end{align}
(recall $\functionspace{E}_{1}:=\functionspace{E}+\langle\cdot,\cdot\rangle$).
Clearly, $\Capa_{1}$ extends $\capa_{1}$ and
$\mu(A)\leq \Capa_{1}(A)$ for $A\in\sigmafield{B}(M)$.
\item[\textup{(2)}]A subset $N$ of $M$ is called \emph{$\functionspace{E}$-polar} if and only if
$\Capa_{1}(N)=0$. Moreover, if $A\subset M$ and $\functionspace{S}(x)$ is
a statement in $x\in A$, then we say that
\emph{$\functionspace{S}$ holds $\functionspace{E}$-q.e.\ on $A$} if and only if
$\{x\in A\mid\textrm{$\functionspace{S}(x)$ fails}\}$ is $\functionspace{E}$-polar.
When $A=M$ we simply say \emph{``$\functionspace{S}$ holds $\functionspace{E}$-q.e."}\ instead.
\item[\textup{(3)}]Let $U\subset M$ be open in $M$. A function $u:U\setminus N\to[-\infty,\infty]$,
with $N\subset M$ $\functionspace{E}$-polar, is called
\emph{$\functionspace{E}$-quasi-continuous on $U$} if and only if
for any $\varepsilon\in(0,\infty)$ there exists an open subset
$V$ of $M$ with $U\cap N\subset V$ and $\Capa_{1}(V)<\varepsilon$ such that
$u|_{U\setminus V}$ is $\mathbb{R}$-valued and continuous.
When $U=M$, such $u$ is simply called
\emph{$\functionspace{E}$-quasi-continuous} instead.
\end{itemize}
\end{definition}
\begin{remark}\label{rmk:capacity}
There are several equivalent ways of defining the notions of
$\functionspace{E}$-polar sets and $\functionspace{E}$-quasi-continuous functions.
See \cite[Section 1.2 and Theorem 1.3.14]{CF} in this connection.
\end{remark}
Note that $\Capa_{1}$ is countably subadditive by \cite[Lemma 2.1.2 and Theorem A.1.2]{FOT}.
Let $U\subset M$ be open in $M$. By \cite[Lemma 2.1.4]{FOT}, if $u,v$ are
$\functionspace{E}$-quasi-continuous functions
on $U$ and $u\leq v$ $\mu$-a.e.\ on $U$,
then $u\leq v$ $\functionspace{E}$-q.e.\ on $U$.
In particular, for each $u\in L^{2}(M,\mu)$, an $\functionspace{E}$-quasi-continuous
$\mu$-version of $u$, if it exists, is unique up to $\functionspace{E}$-q.e.
By \cite[Theorem 2.1.3]{FOT}, each $u\in\functionspace{F}$
admits an $\functionspace{E}$-quasi-continuous $\mu$-version,
which is denoted as $\widetilde{u}$.
For each $t\in(0,\infty)$, while $T_{t}u=\functionspace{P}_{t}u$ $\mu$-a.e.\ for
any $u\in L^{2}(M,\mu)$ by the definition of $T_{t}$, more strongly it actually
holds by \cite[Theorem 4.2.3-(i)]{FOT} that for any $u\in\functionspace{B}L^{2}(M,\mu)$,
\begin{equation}\label{eq:Ptu-quasi-continuous}
\functionspace{P}_{t}u\textrm{ is an }\functionspace{E}\textrm{-quasi-continuous }\mu\textrm{-version of }T_{t}u.
\end{equation}
The following definition gives a probabilistic counterpart of the notion of
$\functionspace{E}$-polar sets.
\begin{definition}\label{dfn:properly-exceptional}
A Borel set $N\in\sigmafield{B}(M)$ is called
\emph{properly exceptional for $X$} if and only if $\mu(N)=0$ and
for any $x\in M\setminus N$,
$\mathbb{P}_{x}[\dot{\sigma}_{N}\wedge\hat{\sigma}_{N}=\infty]=1$
or, by \eqref{eq:entrance-leq-entrance-left-limit}, equivalently
\begin{equation}\label{eq:properly-exceptional}
\mathbb{P}_{x}[\dot{\sigma}_{N}=\infty]=1.
\end{equation}
\end{definition}
Note that
$\{\dot{\sigma}_{N}\wedge\hat{\sigma}_{N}=\infty\}
=\{\textrm{$X_{0},X_{t},X_{t-}\in M_{\cemetery}\setminus N$ for any $t\in(0,\infty)$}\}
\in\sigmafield{F}_{\infty}$
and that $\{\dot{\sigma}_{N}=\infty\}
=\{\textrm{$X_{t}\in M_{\cemetery}\setminus N$ for any $t\in[0,\infty)$}\}
\in\sigmafield{F}_{\infty}$.
Every properly exceptional set for $X$ is $\functionspace{E}$-polar by
\cite[Theorem 4.2.1-(ii)]{FOT}, and conversely any $\functionspace{E}$-polar
set is included in a Borel properly exceptional set for $X$ by \cite[Theorem 4.1.1]{FOT}.
\subsection{The Dirichlet form of the part process on open sets}\label{ssec:part-form}
Let $U$ be a non-empty open subset of $M$ and set $\mu|_{U}:=\mu|_{\sigmafield{B}(U)}$.
Recall that the part $X^{U}$ of $X$ on $U$ is a Hunt process on
$(U,\sigmafield{B}(U))$ defined in \eqref{eq:part-process} and that
its transition function naturally extends to $(M,\sigmafield{B}(M))$ as
a family $\{\functionspace{P}^{U}_{t}\}_{t\in[0,\infty)}$ of Markovian kernels
on $(M,\sigmafield{B}(M))$ given by \eqref{eq:transition-function-part}.
In the present situation, the assumed $\mu$-symmetry of $X$ implies that
$X^{U}$ is $\mu|_{U}$-symmetric. More precisely,
for any $t\in(0,\infty)$ and any $u,v\in\functionspace{B}^{+}(M)$, we have
$\langle\functionspace{P}^{U}_{t}u,v\rangle=\langle u,\functionspace{P}^{U}_{t}v\rangle$
by \cite[Lemma 4.1.3]{FOT} and hence also
$\|\functionspace{P}^{U}_{t}u\|_{2}\leq\|u\|_{2}$ as in \cite[(1.4.13)]{FOT}.
Thus we obtain a Markovian symmetric contraction semigroup $\{T^{U}_{t}\}_{t\in(0,\infty)}$
on $L^{2}(M,\mu)$ canonically induced by $\{\functionspace{P}^{U}_{t}\}_{t\in(0,\infty)}$
in the same way as for $\{\functionspace{P}_{t}\}_{t\in(0,\infty)}$. Moreover,
under the natural identification of $L^{2}(U,\mu|_{U})$ with the closed linear
subspace $\{u\in L^{2}(M,\mu)\mid\textrm{$u=0$ $\mu$-a.e.\ on $M\setminus U$}\}$
of $L^{2}(M,\mu)$, the strongly continuous Markovian semigroup on $L^{2}(U,\mu|_{U})$
induced by the transition function of $X^{U}$ is easily shown to be given by
$\{T^{U}_{t}|_{L^{2}(U,\mu|_{U})}\}_{t\in(0,\infty)}$, and hence
\eqref{eq:Dirichlet-form} with $T^{U}_{t}$ in place of $T_{t}$ gives the Dirichlet
form $(\functionspace{E}^{U},\functionspace{F}_{U})$ of $X^{U}$. In fact,
\begin{equation}\label{eq:part-form}
\functionspace{F}_{U
=\{u\in\functionspace{F}\mid\widetilde{u}=0\ \functionspace{E}\textrm{-q.e.\ on }M\setminus U\}
\qquad\textrm{and}\qquad
\functionspace{E}^{U}=\functionspace{E}|_{\functionspace{F}_{U}\times\functionspace{F}_{U}}
\end{equation}
by \cite[Theorem 4.4.2]{FOT} and $(\functionspace{E}^{U},\functionspace{F}_{U})$
is regular on $L^{2}(U,\mu|_{U})$ by \cite[Lemma 1.4.2-(ii) and Corollary 2.3.1]{FOT}.
$(\functionspace{E}^{U},\functionspace{F}_{U})$ is called the
\emph{part of the Dirichlet form $(\functionspace{E},\functionspace{F})$ on $U$}.
For $t\in(0,\infty)$ and $u\in\functionspace{B}L^{2}(M,\mu)$, while
$T^{U}_{t}u=\functionspace{P}^{U}_{t}u$ $\mu$-a.e.\ by definition, more strongly
\begin{equation}\label{eq:Ptu-quasi-continuous-part}
\functionspace{P}^{U}_{t}u\textrm{ is an }\functionspace{E}\textrm{-quasi-continuous }
\mu\textrm{-version of }T^{U}_{t}u,
\end{equation}
similarly to \eqref{eq:Ptu-quasi-continuous}.
Indeed, since $v:=T^{U}_{t}u\in\functionspace{F}_{U}\subset\functionspace{F}$
by \eqref{eq:TtL2-F} and \eqref{eq:part-form},
$v$ admits an $\functionspace{E}$-quasi-continuous $\mu$-version
$\widetilde{v}$ and then $\widetilde{v}=0=\functionspace{P}^{U}_{t}u$
$\functionspace{E}$-q.e.\ on $M\setminus U$ by \eqref{eq:part-form}.
On the other hand, $(\functionspace{P}^{U}_{t}u)|_{U}$ is a $\mu$-version
of $v|_{U}$ which is $\functionspace{E}$-quasi-continuous on $U$ by
\cite[Theorem 4.4.3]{FOT} and therefore $(\functionspace{P}^{U}_{t}u)|_{U}=\widetilde{v}|_{U}$
$\functionspace{E}$-q.e.\ on $U$ by \cite[Lemma 2.1.4]{FOT}.
Thus $\functionspace{P}^{U}_{t}u=\widetilde{v}$ $\functionspace{E}$-q.e.,
which together with the $\functionspace{E}$-quasi-continuity of $\widetilde{v}$
yields \eqref{eq:Ptu-quasi-continuous-part}.
\section{Localized quasi-everywhere existence of the heat kernel}\label{sec:HK-existence}
As in Section \textup{\ref{sec:Dirichlet-form}}, let $(M,d)$ be a locally compact
separable metric space equipped with a Radon measure $\mu$ with full support, and
let $X$ be a $\mu$-symmetric Hunt process on $(M,\sigmafield{B}(M))$ whose
Dirichlet form $(\functionspace{E},\functionspace{F})$ is regular on $L^{2}(M,\mu)$.
\emph{Throughout the rest of this paper, we fix this setting and follow the notation
introduced in Section \textup{\ref{sec:Dirichlet-form}} in addition to that from
Sections \textup{\ref{sec:Hunt-processes}} and \textup{\ref{sec:multiple-DH}}.}
The purpose of this section is to prove Theorem \ref{thm:HK-existence}
below on the existence of the heat kernel $p^{U}=p^{U}_{t}(x,y)$
for $\{\functionspace{P}^{U}_{t}\}_{t\in(0,\infty)}$
on a given subset of $(0,\infty)\times M\times M$ under a suitable
upper bound on the Markovian semigroup $\{T^{U}_{t}\}_{t\in(0,\infty)}$
\emph{which is assumed only on the given subset}.
In the case where the subset is the whole $(0,\infty)\times M\times M$,
similar results have been obtained, e.g., in
\cite[Sections 7 and 8]{Gri:HKfractal} and \cite[Theorem 3.1]{BBCK}.
\begin{remark}\label{rmk:Dirichlet-form}
The $\mu$-symmetry of $X$ and the regularity of its Dirichlet form are assumed mostly
for the sake of simplicity of the framework. In fact, we need these assumptions
\emph{only in order to use potential theoretic results from
\cite[\emph{Chapters \textup{2} and \textup{4}}]{FOT}
in the proof of Theorem \textup{\ref{thm:HK-existence}}};
it should be possible to extend Theorem \ref{thm:HK-existence} to a more
general framework where the same kind of potential theory remains available,
and the reader is referred to
Remarks \ref{rmk:HK-existence}, \ref{rmk:HKUB} and \ref{rmk:exit-probability}
for the precise settings actually required for the (other) results in
Sections \ref{sec:HK-existence}, \ref{sec:HKUB} and \ref{sec:exit-probability}, respectively.
\end{remark}
For $A\subset M$, let $\ind{A}:M\to\{0,1\}$ denote its indicator function
given by $\ind{A}|_{A}:=1$ and $\ind{A}|_{M\setminus A}:=0$.
In what follows we allow an interval $I\subset\mathbb{R}$ to be a one-point set.
\begin{definition}\label{dfn:upper-bound-function}
Let $I\subset(0,\infty)$ be an interval, $V$ an open subset of $M$ and
$W\in\sigmafield{B}(M)$. A Borel measurable function
$H=H_{t}(x,y):I\times V\times W\to[0,\infty]$ is called a
\emph{$\mu$-upper bound function on $I\times V\times W$} if and only if
the following three conditions are satisfied:
\begin{itemize}[label=\textup{(UB1)},align=left,leftmargin=*]
\item[\textup{(UB1)}]$\limsup_{s\downarrow t}H_{s}(x,y)\leq H_{t}(x,y)$
for any $(t,x,y)\in I\times V\times W$ with $t<\sup I$.
\item[\textup{(UB2)}]$H_{t}(\cdot,y):V\to[0,\infty]$ is upper semi-continuous
for any $(t,y)\in I\times W$.
\item[\textup{(UB3)}]There exist $\{h_{n}\}_{n\in\mathbb{N}}\subset\functionspace{B}^{+}(M)$
and non-decreasing sequences $\{I_{n}\}_{n\in\mathbb{N}}$ of open subsets of $I$,
$\{V_{n}\}_{n\in\mathbb{N}}$ of open subsets of $V$ and
$\{W_{n}\}_{n\in\mathbb{N}}$ of Borel subsets of $W$
with $I=\bigcup_{n\in\mathbb{N}}I_{n}$, $V=\bigcup_{n\in\mathbb{N}}V_{n}$
and $W=\bigcup_{n\in\mathbb{N}}W_{n}$ such that for any $n\in\mathbb{N}$,
\begin{equation}\label{eq:upper-bound-function}
\int_{W_{n}}h_{n}\,d\mu<\infty\quad\textrm{and}\quad H_{t}(x,y)\leq h_{n}(y)
\textrm{ for any }(t,x,y)\in I_{n}\times V_{n}\times W_{n}.
\end{equation}
\end{itemize}
\end{definition}
\begin{remark}\label{rmk:upper-bound-function}
\begin{itemize}[label=\textup{(1)},align=left,leftmargin=*]
\item[\textup{(1)}]In \textup{(UB3)} we may assume that
$\mu(V_{n}\cup W_{n})<\infty$ for any $n\in\mathbb{N}$, by taking a
non-decreasing sequence $\{M_{n}\}_{n\in\mathbb{N}}$ of open subsets of $M$
with $\overline{M_{n}}$ compact and $M=\bigcup_{n\in\mathbb{N}}M_{n}$ and replacing
$V_{n}$ and $W_{n}$ with $V_{n}\cap M_{n}$ and $W_{n}\cap M_{n}$, respectively.
\item[\textup{(2)}]It is easy to see that the condition \textup{(UB3)} in
Definition \ref{dfn:upper-bound-function} is satisfied if $W$ is open in $M$ and
$\|H\|_{\mathrm{sup},K}=\sup_{(t,x,y)\in K}H_{t}(x,y)<\infty$ for any compact subset $K$ of $I\times V\times W$.
\end{itemize}
\end{remark}
\begin{theorem}\label{thm:HK-existence}
Let $I\subset(0,\infty)$ be an interval, $V$ an open subset of $M$,
$W\in\sigmafield{B}(M)$ and let $H=H_{t}(x,y)$ be a $\mu$-upper bound function
on $I\times V\times W$. Let $U$ be a non-empty open subset of $M$.
Then for each countable dense subset $J$ of $I$ satisfying $\max I\in J$
if $\max I$ exists, the following three conditions are equivalent:
\begin{itemize}[label=\textup{(1)},align=left,leftmargin=*]
\item[\textup{(1)}]For any $t\in J$ and any $v,w\in L^{2}(M,\mu)$
with $(v\ind{V})\wedge(w\ind{W})\geq 0$ $\mu$-a.e.,
\begin{equation}\label{eq:HK-existence1}
\langle v\ind{V},T^{U}_{t}(w\ind{W})\rangle\leq\int_{V\times W}v(x)H_{t}(x,y)w(y)\,d(\mu\times\mu)(x,y).
\end{equation}
\item[\textup{(2)}]For each $t\in J$ and each $w\in L^{2}(M,\mu)$
with $w\ind{W}\geq 0$ $\mu$-a.e.,
\begin{equation}\label{eq:HK-existence2}
T^{U}_{t}(w\ind{W})(x)\leq\int_{W}H_{t}(x,y)w(y)\,d\mu(y)
\quad\textrm{for }\mu\textrm{-a.e.\ }x\in V.
\end{equation}
\item[\textup{(3)}]There exist a properly exceptional set $N\in\sigmafield{B}(M)$
for $X$ and a Borel measurable function
$p^{U}=p^{U}_{t}(x,y):I\times(V\setminus N)\times W\to[0,\infty]$
such that for any $(t,x)\in I\times(V\setminus N)$,
\begin{gather}\label{eq:HK-existence3-1}
\functionspace{P}^{U}_{t}(x,A)=\int_{A}p^{U}_{t}(x,y)\,d\mu(y)
\qquad\textrm{for any }A\in\sigmafield{B}(W),\\
p^{U}_{t}(x,y)\leq H_{t}(x,y)\qquad\textrm{for any }y\in W.
\label{eq:HK-existence3-2}
\end{gather}
\end{itemize}
\end{theorem}
We first show the following proposition, which is of independent interest and
will be used in the proof of the implication \textup{(2)}$\Rightarrow$\textup{(3)}
of Theorem \ref{thm:HK-existence} and also in the proof of
Theorems \ref{thm:HKUB-localized} and \ref{thm:HKUB-localized-global}
in the next section.
\begin{remark}\label{rmk:HK-existence}
In fact, \emph{Proposition \textup{\ref{prop:HK-existence}} below applies,
without any changes in the proof, to any locally compact separable metrizable
topological space $M$, any $\sigma$-finite Borel measure $\mu$ on $M$
and any Hunt process $X$ on $(M,\sigmafield{B}(M))$}.
\end{remark}
\begin{proposition}\label{prop:HK-existence}
Let $I\in\sigmafield{B}([0,\infty))$, let $V,W\in\sigmafield{B}(M)$ and let
$H=H_{t}(x,y):I\times V\times W\to[0,\infty]$ be Borel measurable. Let $U$ be a
non-empty open subset of $M$. Then the following two conditions are equivalent:
\begin{itemize}[label=\textup{(1)},align=left,leftmargin=*]
\item[\textup{(1)}]For any $(t,x)\in I\times V$ and any $A\in\sigmafield{B}(W)$,
\begin{equation}\label{eq:HK-existence4}
\functionspace{P}^{U}_{t}(x,A)\leq\int_{A}H_{t}(x,y)\,d\mu(y).
\end{equation}
\item[\textup{(2)}]There exists a Borel measurable function
$p^{U}=p^{U}_{t}(x,y):I\times V\times W\to[0,\infty]$ such that
\eqref{eq:HK-existence3-1} and \eqref{eq:HK-existence3-2} hold for any $(t,x)\in I\times V$.
\end{itemize}
\end{proposition}
\begin{proof}
Since the implication \textup{(2)}$\Rightarrow$\textup{(1)} is immediate,
it suffices to show the converse \textup{(1)}$\Rightarrow$\textup{(2)}.
By the $\sigma$-finiteness of $\mu$, we can choose
$\{W_{n}\}_{n\in\mathbb{N}}\subset\sigmafield{B}(W)$ with
$W=\bigcup_{n\in\mathbb{N}}W_{n}$ so that
$W_{n}\subset W_{n+1}$ and $\mu(W_{n})<\infty$ for any $n\in\mathbb{N}$.
We will construct for each $n\in\mathbb{N}$ a function
$p^{U,n}=p^{U,n}_{t}(x,y):I\times V\times W_{n}\to[0,\infty]$
possessing the required properties with $W_{n}$ in place of $W$.
If $\mu(W_{n})=0$ then it suffices to set $p^{U,n}:=0$ in view of
\eqref{eq:HK-existence4}, and therefore we may assume $\mu(W_{n})>0$.
Let $\sigmafield{U}=\{A_{k}\}_{k\in\mathbb{N}}$ be a countable open base for
the topology of $M$, set $A_{k}^{0}:=M\setminus A_{k}$ and $A_{k}^{1}:=A_{k}$
for $k\in\mathbb{N}$, and define
\begin{equation}\label{eq:HK-existence-filtration}
\sigmafield{A}_{k}:=\{{\textstyle\bigcup_{\alpha\in\functionspace{I}}}A_{k}^{\alpha}\mid
\functionspace{I}\subset\{0,1\}^{k}\},\quad k\in\mathbb{N},
\end{equation}
where $A_{k}^{\alpha}:={\textstyle\bigcap_{i=1}^{k}}A_{i}^{\alpha_{i}}$
for $\alpha=(\alpha_{i})_{i=1}^{k}\in\{0,1\}^{k}$, so that
$\{\sigmafield{A}_{k}\}_{k\in\mathbb{N}}$ is a non-decreasing sequence of
$\sigma$-fields in $M$ with $\bigcup_{k\in\mathbb{N}}\sigmafield{A}_{k}$
generating $\sigmafield{B}(M)$. For $k\in\mathbb{N}$, noting that
$M=\bigcup_{\alpha\in\{0,1\}^{k}}A_{k}^{\alpha}$ and that
$A_{k}^{\alpha}\cap A_{k}^{\beta}=\emptyset$ for $\alpha,\beta\in\{0,1\}^{k}$
with $\alpha\not=\beta$, define $p^{U,n,k}=p^{U,n,k}_{t}(x,y):I\times V\times M\to[0,\infty)$
by, for $\alpha\in\{0,1\}^{k}$ and $(t,x,y)\in I\times V\times A_{k}^{\alpha}$,
\begin{equation}\label{eq:HK-existence-conditional-expectation}
p^{U,n,k}_{t}(x,y):=
\begin{cases}
\mu(A_{k}^{\alpha}\cap W_{n})^{-1}\functionspace{P}^{U}_{t}\ind{A_{k}^{\alpha}\cap W_{n}}(x)
&\textrm{if }\mu(A_{k}^{\alpha}\cap W_{n})>0,\\
0&\textrm{if }\mu(A_{k}^{\alpha}\cap W_{n})=0.
\end{cases}
\end{equation}
Then $p^{U,n,k}$ is Borel measurable by \eqref{eq:Pt-measurable}
for $\{\functionspace{P}^{U}_{t}\}_{t\in[0,\infty)}$.
Furthermore for each $(t,x)\in I\times V$, since
$\functionspace{P}^{U}_{t}(x,(\cdot)\cap W_{n})$
is absolutely continuous with respect to $\mu((\cdot)\cap W_{n})$ and
$f^{t,x}_{n}:=\frac{d\functionspace{P}^{U}_{t}(x,(\cdot)\cap W_{n})}{d\mu((\cdot)\cap W_{n})}
\leq H_{t}(x,\cdot)$ $\mu$-a.e.\ on $W_{n}$ by \eqref{eq:HK-existence4},
$p^{U,n,k}_{t}(x,\cdot)$ is a version of the $\sigmafield{A}_{k}$-conditional
$\frac{\mu((\cdot)\cap W_{n})}{\mu(W_{n})}$-expectation of $f^{t,x}_{n}$
and hence $\lim_{k\to\infty}p^{U,n,k}_{t}(x,y)=f^{t,x}_{n}(y)\leq H_{t}(x,y)$
for $\mu$-a.e.\ $y\in W_{n}$ by the martingale convergence theorem
\cite[Theorem 10.5.1]{Dud:RAP}. Therefore the function
$p^{U,n}_{t}(x,y):=H_{t}(x,y)\wedge\liminf_{k\to\infty}p^{U,n,k}_{t}(x,y)$,
$(t,x,y)\in I\times V\times W_{n}$, has the desired properties.
Now the proof of \textup{(2)} is completed by setting $p^{U}_{t}(x,y):=p^{U,n}_{t}(x,y)$
for $n\in\mathbb{N}$ and $(t,x,y)\in I\times V\times(W_{n}\setminus W_{n-1})$
($W_{0}:=\emptyset$) and using monotone convergence.
\end{proof}
\begin{proof}[Proof of Theorem \textup{\ref{thm:HK-existence}}]
The implication \textup{(2)}$\Rightarrow$\textup{(1)} is immediate, and it is
easy to see from \textup{(UB3)} of Definition \ref{dfn:upper-bound-function}
and Remark \ref{rmk:upper-bound-function}-(1) that \textup{(1)} implies
\textup{(2)}. The implication \textup{(3)}$\Rightarrow$\textup{(2)} also
follows easily since $T^{U}_{t}u=\functionspace{P}^{U}_{t}u$ $\mu$-a.e.\ for
any $t\in(0,\infty)$ and any $u\in\functionspace{B}L^{2}(M,\mu)$.
Therefore it remains to prove \textup{(2)}$\Rightarrow$\textup{(3)}. Let
$\{h_{n}\}_{n\in\mathbb{N}},\{I_{n}\}_{n\in\mathbb{N}},
\{V_{n}\}_{n\in\mathbb{N}},\{W_{n}\}_{n\in\mathbb{N}}$
be as in (UB3) with $\mu(W_{n})<\infty$ for any
$n\in\mathbb{N}$ as noted in Remark \ref{rmk:upper-bound-function}-(1).
Let $\sigmafield{A}_{k}$ be as in \eqref{eq:HK-existence-filtration} for each
$k\in\mathbb{N}$ and set $\sigmafield{A}:=\bigcup_{k\in\mathbb{N}}\sigmafield{A}_{k}$,
so that $\sigmafield{A}$ is countable, generates $\sigmafield{B}(M)$ and satisfies
$\emptyset\in\sigmafield{A}$, $M\setminus A\in\sigmafield{A}$ for any $A\in\sigmafield{A}$
and $A\cup B\in\sigmafield{A}$ for any $A,B\in\sigmafield{A}$.
By \eqref{eq:Ptu-quasi-continuous-part} and \cite[Theorem 2.1.2-(i)]{FOT},
there exists a non-decreasing sequence $\{F_{k}\}_{k\in\mathbb{N}}$
of closed subsets of $M$ such that $\lim_{k\to\infty}\Capa_{1}(M\setminus F_{k})=0$
and for each $k\in\mathbb{N}$,
$\mu(G\cap F_{k})>0$ for any open subset $G$ of $M$ with $G\cap F_{k}\not=\emptyset$ and
$\{\functionspace{P}^{U}_{t}\ind{A\cap W_{n}}|_{F_{k}}\mid\textrm{$n\in\mathbb{N}$, $t\in J$, $A\in\sigmafield{A}$}\}
\subset C(F_{k})$.
Moreover, since
\begin{equation*}
\functionspace{P}^{U}_{t}\ind{A\cap W_{n}}(x)
=\functionspace{P}^{U}_{t-l^{-1}}(\functionspace{P}^{U}_{l^{-1}}\ind{A\cap W_{n}})(x)
=\mathbb{E}_{x}\bigl[\functionspace{P}^{U}_{l^{-1}}\ind{A\cap W_{n}}(X_{t-l^{-1}})\ind{\{t-l^{-1}<\tau_{U}\}}\bigr]
\end{equation*}
for $l\in\mathbb{N}$ and $t\in[l^{-1},\infty)$ by \eqref{eq:Pt-semigroup} for
$\{\functionspace{P}^{U}_{t}\}_{t\in[0,\infty)}$, an application of
\eqref{eq:Ptu-quasi-continuous-part} and \cite[Theorem 4.2.2]{FOT}
to $\functionspace{P}^{U}_{l^{-1}}\ind{A\cap W_{n}}$
with $l,n\in\mathbb{N}$ and $A\in\sigmafield{A}$ yields an
$\functionspace{E}$-polar set $N_{0}\in\sigmafield{B}(M)$ such that
$(0,\infty)\ni t\mapsto\functionspace{P}^{U}_{t}\ind{A\cap W_{n}}(x)\in\mathbb{R}$
is right-continuous for any $x\in M\setminus N_{0}$, any $n\in\mathbb{N}$ and
any $A\in\sigmafield{A}$. Then $(M\setminus\bigcup_{k\in\mathbb{N}}F_{k})\cup N_{0}$
is $\functionspace{E}$-polar and therefore by \cite[Theorem 4.1.1]{FOT} we can
take a properly exceptional set $N\in\sigmafield{B}(M)$ for $X$ satisfying
$(M\setminus\bigcup_{k\in\mathbb{N}}F_{k})\cup N_{0}\subset N$.
Let $n\in\mathbb{N}$ and $(t,x)\in I\times(V\setminus N)$.
We claim that for any $A\in\sigmafield{B}(M)$,
\begin{equation}\label{eq:HK-existence2-qe}
\functionspace{P}^{U}_{t}\ind{A\cap W_{n}}(x)\leq\int_{A\cap W_{n}}H_{t}(x,y)\,d\mu(y),
\end{equation}
whose limit as $n\to\infty$ results in \eqref{eq:HK-existence4} with $V\setminus N$
in place of $V$ by monotone convergence, thereby proving
\textup{(2)}$\Rightarrow$\textup{(3)} by virtue of Proposition \ref{prop:HK-existence}.
Thus it remains to show \eqref{eq:HK-existence2-qe}. To this end, let
$A\in\sigmafield{A}$ and choose $k\in\mathbb{N}$ with $k\geq n$ so that
$t\in I_{k}$ and $x\in V_{k}\cap F_{k}$.
First we assume $t\in J$. Then
$\functionspace{P}^{U}_{t}\ind{A\cap W_{n}}\leq\int_{A\cap W_{n}}H_{t}(\cdot,y)\,d\mu(y)$
$\mu$-a.e.\ on $V$ by (2), and since $\mu(G\cap V_{k}\cap F_{k})>0$
for any open subset $G$ of $M$ with $x\in G$ we can take
$\{x_{l}\}_{l\in\mathbb{N}}\subset V_{k}\cap F_{k}$ such that
$\lim_{l\to\infty}x_{l}=x$ in $M$ and
$\functionspace{P}^{U}_{t}\ind{A\cap W_{n}}(x_{l})\leq\int_{A\cap W_{n}}H_{t}(x_{l},y)\,d\mu(y)$
for any $l\in\mathbb{N}$. Now \eqref{eq:HK-existence2-qe} follows by utilizing
$\functionspace{P}^{U}_{t}\ind{A\cap W_{n}}|_{F_{k}}\in C(F_{k})$,
Fatou's lemma and \textup{(UB2)} to let $l\to\infty$, where the use of Fatou's
lemma is justified by \eqref{eq:upper-bound-function} with $k$ in place of $n$.
Next for $t\in I\setminus J$, with $k\in\mathbb{N}$ as above, we can take
a strictly decreasing sequence $\{t_{l}\}_{l\in\mathbb{N}}\subset I_{k}\cap J$
satisfying $\lim_{l\to\infty}t_{l}=t$, and then
$\functionspace{P}^{U}_{t_{l}}\ind{A\cap W_{n}}(x)\leq\int_{A\cap W_{n}}H_{t_{l}}(x,y)\,d\mu(y)$
for any $l\in\mathbb{N}$ by the previous paragraph. Now letting $l\to\infty$
yields \eqref{eq:HK-existence2-qe} for this case by the right-continuity of
$\functionspace{P}^{U}_{{\scriptscriptstyle(\cdot)}}\ind{A\cap W_{n}}(x)$,
Fatou's lemma and \textup{(UB1)}, where \eqref{eq:upper-bound-function} with
$k$ in place of $n$ is used again to verify the applicability of Fatou's lemma
to the right-hand side.
Thus \eqref{eq:HK-existence2-qe} has been proved for any $A\in\sigmafield{A}$.
Further, we easily see from \eqref{eq:upper-bound-function} with $k$
in place of $n$ and the dominated convergence theorem that
$\{A\in\sigmafield{B}(M)\mid\textrm{$A$ satisfies \eqref{eq:HK-existence2-qe}}\}$
is closed under monotone countable unions and intersections, and hence
the monotone class theorem \cite[Theorem 4.4.2]{Dud:RAP} implies that
\eqref{eq:HK-existence2-qe} holds for any $A\in\sigmafield{B}(M)$.
\end{proof}
The rest of this section is devoted to presenting examples of $\mu$-upper bound
functions. We start with a lemma which is mostly due to \cite[Subsection 3.4]{GT}.
\begin{lemma}\label{lem:upper-bound-function-Psi}
Let $\Psi:[0,\infty)\to[0,\infty)$ be a homeomorphism satisfying
\begin{equation}\label{eq:upper-bound-function-Psi}
c_{\Psi}^{-1}\Bigl(\frac{R}{r}\Bigr)^{\beta_{1}}\leq\frac{\Psi(R)}{\Psi(r)}
\leq c_{\Psi}\Bigl(\frac{R}{r}\Bigr)^{\beta_{2}}
\qquad\textrm{for any }r,R\in(0,\infty)\textrm{ with }r\leq R
\end{equation}
for some $c_{\Psi},\beta_{1},\beta_{2}\in(0,\infty)$ with
$1<\beta_{1}\leq\beta_{2}$, and for $(R,t)\in[0,\infty)\times(0,\infty)$ define
\begin{equation}\label{eq:upper-bound-function-Phi}
\Phi(R,t):=\Phi_{\Psi}(R,t)
:=\sup_{r\in(0,\infty)}\Bigl\{\frac{R}{r}-\frac{t}{\Psi(r)}\Bigr\}
=\sup_{\lambda\in(0,\infty)}\Bigl\{\frac{R}{\Psi^{-1}(\lambda^{-1})}-\lambda t\Bigr\}.
\end{equation}
Then $\Phi=\Phi_{\Psi}$ is a $[0,\infty)$-valued lower semi-continuous function
such that for any $R,t\in(0,\infty)$, $\Phi(\cdot,t)$ is non-decreasing,
$\Phi(R,\cdot)$ is non-increasing, $\Phi(0,t)=0<\Phi(R,t)$,
\begin{align}\label{eq:upper-bound-function-Phi-constant}
a\Phi(R,t)&\leq\Phi(aR,t)\quad\textrm{for any }a\in[1,\infty),\\
(c_{\Psi}2^{\beta_{1}})^{-\frac{1}{\beta_{1}-1}}\min_{k\in\{1,2\}}\Bigl(\frac{\Psi(R)}{t}\Bigr)^{\frac{1}{\beta_{k}-1}}
&\leq\Phi(R,t)\leq c_{\Psi}^{\frac{1}{\beta_{1}-1}}\max_{k\in\{1,2\}}\Bigl(\frac{\Psi(R)}{t}\Bigr)^{\frac{1}{\beta_{k}-1}}.
\label{eq:upper-bound-function-Phi-ULE}
\end{align}
\end{lemma}
\begin{proof}
The lower semi-continuity of $\Phi=\Phi_{\Psi}$ is clear from
\eqref{eq:upper-bound-function-Phi}, and the other assertions
except the upper inequality in \eqref{eq:upper-bound-function-Phi-ULE}
have been verified in \cite[Remark 3.16 and Lemma 3.19]{GT}.
To see the upper inequality in \eqref{eq:upper-bound-function-Phi-ULE},
let $R,t,r\in(0,\infty)$ and set $a:=R\Psi(r)/(rt)$. Noting that
$R/r-t/\Psi(r)=(a-1)t/\Psi(r)\leq 0$ if $a\leq 1$, we assume $a>1$,
and set $\beta:=\beta_{1}$ if $r\leq R$ and $\beta:=\beta_{2}$ if $r>R$.
Then $at/\Psi(r)=R/r\leq(c_{\Psi}\Psi(R)/\Psi(r))^{1/\beta}$
by \eqref{eq:upper-bound-function-Psi}, hence
$\Psi(r)\geq at(c_{\Psi}^{-1}at/\Psi(R))^{\frac{1}{\beta-1}}$, and therefore
\begin{equation*}
\frac{R}{r}-\frac{t}{\Psi(r)}\leq\frac{R}{r}=\frac{at}{\Psi(r)}
\leq\Bigl(\frac{c_{\Psi}\Psi(R)}{at}\Bigr)^{\frac{1}{\beta-1}}
\leq c_{\Psi}^{\frac{1}{\beta_{1}-1}}\max_{k\in\{1,2\}}\Bigl(\frac{\Psi(R)}{t}\Bigr)^{\frac{1}{\beta_{k}-1}},
\end{equation*}
where the last inequality follows by $a\geq 1$, $1<\beta_{1}\leq\beta_{2}$
and the fact that $c_{\Psi}\geq 1$ by \eqref{eq:upper-bound-function-Psi}.
Now taking the supremum in $r\in(0,\infty)$ yields the desired inequality.
\end{proof}
\begin{example}\label{exmp:upper-bound-function-Psi}
An important special case of Lemma \ref{lem:upper-bound-function-Psi}
is that of $\Psi(r)=r^{\beta}$ for some $\beta\in(1,\infty)$ treated in
\cite[Example 3.17]{GT}, where $\Phi=\Phi_{\Psi}$ is easily evaluated as
\begin{equation}\label{eq:upper-bound-function-beta}
\Phi(R,t)=\beta^{-\frac{\beta}{\beta-1}}(\beta-1)\Bigl(\frac{R^{\beta}}{t}\Bigr)^{\frac{1}{\beta-1}}.
\end{equation}
\end{example}
The following lemma provides a class of typical $\mu$-upper bound functions,
which has essentially appeared in \cite[(6.10)]{GriHuLau:comp}. Note that
\emph{Lemma \textup{\ref{lem:upper-bound-function}} and Example
\textup{\ref{exmp:upper-bound-function}} below, as well as
Remark \textup{\ref{rmk:upper-bound-function}} above, apply to any locally
compact separable metric space $(M,d)$ and any Radon measure $\mu$ on $M$}
(i.e., any Borel measure on $M$ that is finite on compact sets).
\begin{lemma}\label{lem:upper-bound-function}
Let $\Psi$ and $\Phi=\Phi_{\Psi}$ be as in Lemma \textup{\ref{lem:upper-bound-function-Psi}}.
Let $I\subset(0,\infty)$ be an interval, let $V,W$ be open subsets of $M$ and
let $F=F_{t}(x,y):I\times V\times W\to(0,\infty)$ be a Borel measurable function
satisfying \textup{(UB1)} and \textup{(UB2)} of Definition \textup{\ref{dfn:upper-bound-function}}
and the following \emph{$\Psi$-doubling condition \textup{(DB)$_{\Psi}$}}:
\begin{itemize}[label=\textup{(DB)$_{\Psi}$},align=left,leftmargin=*]
\item[\textup{(DB)$_{\Psi}$}]There exist $\alpha_{F},c_{F}\in(0,\infty)$ such that
for any $(t,x,y),(s,z,w)\in I\times V\times W$ with $s\leq t$,
\begin{equation}\label{eq:upper-bound-function-F-DBPsi}
\frac{F_{s}(z,w)}{F_{t}(x,y)}
\leq c_{F}\Bigl(\frac{t\vee\Psi(d(x,z))\vee\Psi(d(y,w))}{s}\Bigr)^{\alpha_{F}}.
\end{equation}
\end{itemize}
Also let $c_{1},c_{2}\in(0,\infty)$ and define
$H=H_{t}(x,y):I\times V\times W\to(0,\infty)$ by
\begin{equation}\label{eq:upper-bound-function-H}
H_{t}(x,y):=F_{t}(x,y)\exp\bigl(-c_{1}\Phi(c_{2}d(x,y),t)\bigr).
\end{equation}
Then $F=F_{t}(x,y)$ and $H=H_{t}(x,y)$ are $\mu$-upper bound functions on $I\times V\times W$.
\end{lemma}
\begin{proof}
It is immediate to see that $H=H_{t}(x,y)$ is Borel measurable and satisfies
\textup{(UB1)} and \textup{(UB2)}, from the corresponding properties of $F=F_{t}(x,y)$
and the lower semi-continuity of $\Phi$. Also \textup{(DB)$_{\Psi}$} easily implies
that $F=F_{t}(x,y)$ and hence $H=H_{t}(x,y)$ are bounded on each compact subset of
$I\times V\times W$, so that they satisfy \textup{(UB3)}
by Remark \ref{rmk:upper-bound-function}-(2).
\end{proof}
\begin{example}\label{exmp:upper-bound-function}
Let $\Psi$ be as in Lemma \textup{\ref{lem:upper-bound-function-Psi}}.
\begin{itemize}[label=\textup{(1)},align=left,leftmargin=*]
\item[\textup{(1)}]A continuous function
$F=F_{t}(x,y):(0,\infty)\times M\times M\to(0,\infty)$ of the form
\begin{equation}\label{eq:upper-bound-function-F-ex1}
F_{t}(x,y)=c_{3}t^{-\alpha_{1}}\bigl(\log(2+t^{-1})\bigr)^{\alpha_{2}}\bigl(\log(2+t)\bigr)^{\alpha_{3}}
\end{equation}
for some $c_{3},\alpha_{1}\in(0,\infty)$ and $\alpha_{2},\alpha_{3}\in\mathbb{R}$ clearly
satisfies \textup{(UB1)}, \textup{(UB2)} and \textup{(DB)$_{\Psi}$}.
\item[\textup{(2)}]
Let $R\in(0,\infty]$, let $V,W$ be open subsets of $M$ with
$(\diam V)\vee(\diam W)\leq R$ and let $\nu$ be a Borel measure on $M$
satisfying the \emph{volume doubling property}
\begin{equation}\label{eq:VD}
0<\nu(B(x,2r))\leq c_{\mathrm{vd}}\nu(B(x,r))<\infty
\end{equation}
for any $(x,r)\in(V\cup W)\times(0,R)$ for some $c_{\mathrm{vd}}\in(0,\infty)$.
Then for each $c_{4}\in(0,\infty)$, the function
$F=F_{t}(x,y):(0,\Psi(R)]\times V\times W\to(0,\infty)$
($(0,\infty)$ in place of $(0,\Psi(R)]$ for $R=\infty$) defined by
\begin{equation}\label{eq:upper-bound-function-F-vol}
F_{t}(x,y):=c_{4}\nu\bigl(B(x,\Psi^{-1}(t))\bigr)^{-1/2}\nu\bigl(B(y,\Psi^{-1}(t))\bigr)^{-1/2}
\end{equation}
is easily proved to be upper semi-continuous and satisfy \textup{(DB)$_{\Psi}$}
thanks to \eqref{eq:VD} and \eqref{eq:upper-bound-function-Psi}, and in particular
it is Borel measurable and satisfies \textup{(UB1)} and \textup{(UB2)}.
\end{itemize}
\end{example}
\section{Localized upper bounds of heat kernels for diffusions}\label{sec:HKUB}
In this section, we state and prove the main theorem of this paper on deducing
heat kernel upper bounds for $\{\functionspace{P}_{t}\}_{t\in(0,\infty)}$
from those for $\{\functionspace{P}^{U}_{t}\}_{t\in(0,\infty)}$
(Theorem \ref{thm:HKUB-localized} below).
The arguments heavily rely on the decay estimate \eqref{eq:exit-probability}
for the exit probabilities $\mathbb{P}_{x}[\tau_{B(x,r)}\leq t]$, for which
reasonable sufficient conditions will be presented in the next section.
\emph{In the rest of this paper, we fix a homeomorphism
$\Psi:[0,\infty)\to[0,\infty)$ and $c_{\Psi},\beta_{1},\beta_{2}\in(0,\infty)$
with $1<\beta_{1}\leq\beta_{2}$ satisfying \eqref{eq:upper-bound-function-Psi},
and $\Phi=\Phi_{\Psi}$ denotes the function given by \eqref{eq:upper-bound-function-Phi}.}
\emph{Throughout this section, we fix an arbitrary properly exceptional set
$N\in\sigmafield{B}(M)$ for $X$ such that for any
$x\in M\setminus N$,
\begin{equation}\label{eq:X-continuous}
\mathbb{P}_{x}\bigl[[0,\zeta)\ni t\mapsto X_{t}\in M\textrm{ is continuous}\bigr]=1,
\end{equation}
where
$\{\textrm{$[0,\zeta)\ni t\mapsto X_{t}\in M$ is continuous}\}\in\sigmafield{F}_{\infty}$
by \cite[Chapter III, 13 and 33]{DM}. According to \cite[Theorem 4.5.1]{FOT},
such $N$ exists if and only if
$(\functionspace{E},\functionspace{F})$ is \emph{local},
i.e., $\functionspace{E}(u,v)=0$ for any $u,v\in\functionspace{F}$ with
$\supp_{\mu}[u],\supp_{\mu}[v]$ compact and $\supp_{\mu}[u]\cap\supp_{\mu}[v]=\emptyset$.
Here for $u\in\functionspace{B}(M)$ or its $\mu$-equivalence class,
$\supp_{\mu}[u]$ denotes its \emph{$\mu$-support} defined as
the smallest closed subset of $M$ such that
$u=0$ $\mu$-a.e.\ on $M\setminus\supp_{\mu}[u]$, which exists since
$M$ has a countable open base for its topology. Note that
$\supp_{\mu}[u]=\overline{u^{-1}(\mathbb{R}\setminus\{0\})}$ for $u\in C(M)$.
\begin{remark}\label{rmk:HKUB}
In fact, \emph{Theorems \textup{\ref{thm:HKUB-localized}},
\textup{\ref{thm:HKUB-localized-global}},
Propositions \textup{\ref{prop:HKUB-localized}} and
\textup{\ref{prop:HKUB-localized-difference}} below apply, without
any changes in the proofs, to any locally compact separable metric space $(M,d)$,
any $\sigma$-finite Borel measure $\mu$ on $M$, any Hunt process $X$ on
$(M,\sigmafield{B}(M))$ and any $N\in\sigmafield{B}(M)$ satisfying
\eqref{eq:properly-exceptional} and \eqref{eq:X-continuous}
for any $x\in M\setminus N$}.
\end{remark}
\begin{theorem}\label{thm:HKUB-localized}
Let $R\in(0,\infty)$, let $U$ be a non-empty open subset of $M$ with $\diam U\leq R$
and let $F=F_{t}(x,y):(0,\Psi(R)]\times U\times U\to(0,\infty)$ be a Borel measurable
function satisfying \textup{(DB)$_{\Psi}$} of Lemma \textup{\ref{lem:upper-bound-function}}
with $I=(0,\Psi(R)]$ and $V=W=U$.
Let $c,\gamma\in(0,\infty)$ and assume that the following two conditions
\textup{(DU)$_{F}^{U,R}$} and \textup{(P)$_{\Psi}^{U,R}$} are fulfilled:
\begin{itemize}[label=\textup{(DU)$_{F}^{U,R}$},align=left,leftmargin=*]
\item[\textup{(DU)$_{F}^{U,R}$}]
For any $(t,x)\in(0,\Psi(R))\times(U\setminus N)$ and any $A\in\sigmafield{B}(U)$,
\begin{equation}\label{eq:HKUB-localized-on-diag}
\functionspace{P}^{U}_{t}(x,A)\leq\int_{A}F_{t}(x,y)\,d\mu(y).
\end{equation}
\item[\textup{(P)$_{\Psi}^{U,R}$}]
For any $(x,r)\in(U\setminus N)\times(0,R)$ with $B(x,r)\subset U$
and any $t\in(0,\infty)$,
\begin{equation}\label{eq:exit-probability}
\mathbb{P}_{x}[\tau_{B(x,r)}\leq t]\leq c\exp(-\Phi(\gamma r,t)).
\end{equation}
\end{itemize}
Let $\varepsilon\in(0,1)$ and set
$U^{\circ}_{\varepsilon R}:=\{x\in M\mid\inf_{y\in M\setminus U}d(x,y)>\varepsilon R\}$
\textup{(note that $U^{\circ}_{\varepsilon R}$ is an open subset of $U$)}.
Then there exists a Borel measurable function
$p=p_{t}(x,y):(0,\infty)\times(M\setminus N)\times U^{\circ}_{\varepsilon R}\to[0,\infty)$
such that for any $(t,x)\in(0,\infty)\times(M\setminus N)$ the following hold:
\begin{equation}\label{eq:HKUB-localized-existence}
\functionspace{P}_{t}(x,A)=\int_{A}p_{t}(x,y)\,d\mu(y)
\qquad\textrm{for any }A\in\sigmafield{B}(U^{\circ}_{\varepsilon R}),
\end{equation}
and furthermore for any $y\in U^{\circ}_{\varepsilon R}$,
\begin{equation}\label{eq:HKUB-localized-off-diag-U}
p_{t}(x,y)\leq
\begin{cases}
c_{\varepsilon}F_{t}(x,y)\exp\bigl(-\Phi(\gamma_{\varepsilon}d(x,y),t)\bigr) &\textrm{if }t<\Psi(R)\textrm{ and }x\in U,\\
c_{\varepsilon}(\inf_{U\times U}F_{(2t)\wedge\Psi(R)})\exp(-\Phi(\gamma_{\varepsilon}R,t)) &\textrm{if }t<\Psi(R)\textrm{ and }x\not\in U,\\
c_{\varepsilon}(\inf_{U\times U}F_{\Psi(R)}) &\textrm{if }t\geq\Psi(R)
\end{cases}
\end{equation}
for some $c_{\varepsilon}\in(0,\infty)$ explicit in
$c_{\Psi},\beta_{1},\beta_{2},c_{F},\alpha_{F},c,\gamma,\varepsilon$ and
$\gamma_{\varepsilon}:=\frac{1}{5}\varepsilon\gamma$.
\end{theorem}
In light of the equivalence stated in Proposition \ref{prop:HK-existence} and
the examples of Borel measurable functions $F=F_{t}(x,y)$ satisfying
\textup{(DB)$_{\Psi}$} in Example \ref{exmp:upper-bound-function},
\textup{(DU)$_{F}^{U,R}$} of Theorem \ref{thm:HKUB-localized} amounts to an
\emph{on-diagonal} upper bound of the heat kernel $p^{U}=p^{U}_{t}(x,y)$ for
$\{\functionspace{P}^{U}_{t}\}_{t\in(0,\infty)}$.
Note that \emph{the two conditions \textup{(DU)$_{F}^{U,R}$} and
\textup{(P)$_{\Psi}^{U,R}$} involve only the part $X^{U}$ of $X$ on $U$
and hence are independent of the behavior of $X$ after exiting $U$},
on account of \eqref{eq:transition-function-part} and the obvious fact that
$\tau_{B}(\omega)=\inf\{t\in[0,\infty)\mid X^{U}_{t}(\omega)\in U_{\cemetery}\setminus B\}$
for $B\subset U$ and $\omega\in\Omega$.
\begin{remark}\label{rmk:HKUB-localized}
Theorem \ref{thm:HK-existence} tells us that \textup{(DU)$_{F}^{U,R}$}
of Theorem \ref{thm:HKUB-localized} is implied,
at the price of replacing $N$ with a larger properly exceptional set for $X$,
by its \emph{``$\mu$-a.e."} counterpart for the Markovian semigroup
$\{T^{U}_{t}\}_{t\in(0,\infty)}$ provided $F=F_{t}(x,y)$ is a
$\mu$-upper bound function on $(0,\Psi(R))\times U\times U$.
Remember, though, that \emph{we have proved Theorem \textup{\ref{thm:HK-existence}}
only for a Radon measure $\mu$ on $M$ with full support and a $\mu$-symmetric
Hunt process $X$ on $(M,\sigmafield{B}(M))$ whose Dirichlet form
$(\functionspace{E},\functionspace{F})$ is regular on $L^{2}(M,\mu)$};
recall Remark \ref{rmk:Dirichlet-form} in this connection.
\end{remark}
We also have a global version of Theorem \ref{thm:HKUB-localized}
for the case where its assumptions are valid on $B(y_{0},\frac{R'}{2})$ for any
$(y_{0},R')\in M\times(0,\infty)$ with $R'\leq R$, as follows. Set $\Psi(\infty):=\infty$.
\begin{theorem}\label{thm:HKUB-localized-global}
Let $\delta\in(0,1]$, let $R\in(0,\infty]$ satisfy $R\geq\delta\diam M$
and let $F=F_{t}(x,y):(0,\Psi(R)]\times M\times M\to(0,\infty)$ be a Borel
measurable function satisfying \textup{(DB)$_{\Psi}$} of
Lemma \textup{\ref{lem:upper-bound-function}} with $I=(0,\Psi(R)]$ and $V=W=M$
\textup{($(0,\infty)$ in place of $(0,\Psi(R)]$ for $R=\infty$)}.
Let $c,\gamma\in(0,\infty)$ and assume that the two conditions
\textup{(DU)$_{F}^{B(y_{0},R'/2),R'}$} and \textup{(P)$_{\Psi}^{B(y_{0},R'/2),R'}$}
from Theorem \textup{\ref{thm:HKUB-localized}} are fulfilled for any
$(y_{0},R')\in M\times(0,\infty)$ with $R'\leq R$.
Then there exists a Borel measurable function
$p=p_{t}(x,y):(0,\infty)\times(M\setminus N)\times M\to[0,\infty)$
such that for any $(t,x)\in(0,\infty)\times(M\setminus N)$,
\eqref{eq:HKUB-localized-existence} with $\sigmafield{B}(M)$ in place of
$\sigmafield{B}(U^{\circ}_{\varepsilon R})$ holds and
\begin{equation}\label{eq:HKUB-localized-off-diag}
p_{t}(x,y)\leq
\begin{cases}
c'\delta^{-\beta_{2}\alpha_{F}}F_{t}(x,y)\exp\bigl(-\Phi(\gamma'_{\delta}d(x,y),t)\bigr) &\textrm{if }t<\Psi(R),\\
c'\delta^{-\beta_{2}\alpha_{F}}(\inf_{M\times M}F_{\Psi(R)}) &\textrm{if }t\geq\Psi(R)
\end{cases}
\end{equation}
for any $y\in M$ for some $c'\in(0,\infty)$ explicit in
$c_{\Psi},\beta_{1},\beta_{2},c_{F},\alpha_{F},c,\gamma$
and $\gamma'_{\delta}:=\frac{1}{40}\delta\gamma$.
\end{theorem}
The rest of this section is devoted to the proof of
Theorems \ref{thm:HKUB-localized} and \ref{thm:HKUB-localized-global}.
We start with the proof of the following proposition, which, in view of
Proposition \ref{prop:HK-existence}, can be considered
as a localized version of \cite[Theorem 6.3]{GriHuLau:comp}.
Its proof in \cite{GriHuLau:comp} is based on a general comparison inequality
\cite[Theorem 5.1]{GriHuLau:comp} among the heat kernels on different open sets
\emph{which heavily relies on the symmetry of the Markovian semigroups
$\{T^{U}_{t}\}_{t\in(0,\infty)}$}; see also \cite[Theorem 10.4]{Gri:HKfractal}
for an alternative probabilistic proof of the same comparison inequality.
Here we give a new proof \emph{which does not require the $\mu$-symmetry of $X$}.
\begin{proposition}\label{prop:HKUB-localized}
Under the same assumptions as those of Theorem \textup{\ref{thm:HKUB-localized}},
there exists $c'_{\varepsilon}\in(0,\infty)$ explicit in
$c_{\Psi},\beta_{1},\beta_{2},c_{F},\alpha_{F},c,\gamma,\varepsilon$ such that,
with $\gamma_{\varepsilon}:=\frac{1}{5}\varepsilon\gamma$,
for any $(t,x)\in(0,\Psi(R))\times(U\setminus N)$ and
any $A\in\sigmafield{B}(U^{\circ}_{\varepsilon R})$,
\begin{equation}\label{eq:HKUB-localized-off-diag-local}
\functionspace{P}^{U}_{t}(x,A)
\leq\int_{A}c'_{\varepsilon}F_{t}(x,y)\exp\bigl(-\Phi(\gamma_{\varepsilon}d(x,y),t)\bigr)\,d\mu(y).
\end{equation}
\end{proposition}
\begin{proof}
Let $(t,x)\in(0,\Psi(R))\times(U\setminus N)$.
Let $y_{0}\in U^{\circ}_{\varepsilon R}\setminus\{x\}$, set
$r:=\frac{1}{4}\varepsilon d(x,y_{0})\in(0,\frac{1}{4}\varepsilon R]$
and let $A\in\sigmafield{B}(B(y_{0},r))$.
We first verify \eqref{eq:HKUB-localized-off-diag-local} for such $A$.
Since $A\subset B(y_{0},r)\subset B(y_{0},4r)\subset U$ by
$y_{0}\in U^{\circ}_{\varepsilon R}$ and $r\in(0,\frac{1}{4}\varepsilon R]$, if
$t\geq\Psi((2+4/\varepsilon)r)$ then \eqref{eq:HKUB-localized-off-diag-local}
is immediate from \textup{(DU)$_{F}^{U,R}$}
and the upper inequality in \eqref{eq:upper-bound-function-Phi-ULE},
and therefore we may assume $t<\Psi((2+4/\varepsilon)r)$.
We set $r_{n}:=r+2^{-n/(2\beta_{2})}r$ and
$\sigma_{n}:=\dot{\sigma}_{B(y_{0},r_{n})}$ for $n\in\mathbb{N}$, so that
$B(y_{0},r)\subset B(y_{0},r_{n})\subset B(y_{0},r_{k})$ and hence
$\sigma_{k}\leq\sigma_{n}\leq\dot{\sigma}_{B(y_{0},r)}$ for any $k\in\{1,\dots,n\}$.
Let $\omega\in\{\textrm{$[0,\zeta)\ni s\mapsto X_{s}\in M$ is continuous}\}$.
It is easy to see that for $B\subset M$,
\begin{equation}\label{eq:X-continuous-entrance}
\textrm{if}\quad X_{0}(\omega)\not\in\interior B\quad\textrm{and}\quad
\dot{\sigma}_{B}(\omega)<\infty\qquad\textrm{then}\qquad
X_{\dot{\sigma}_{B}}(\omega)\in\partial B.
\end{equation}
Assume further that $\omega\in\{X_{t}\in B(y_{0},r),\,X_{0}=x\}$. Then since
$X_{0}(\omega)=x\not\in B(y_{0},4r)$ by $d(x,y_{0})>\varepsilon d(x,y_{0})=4r$
and $\dot{\sigma}_{B(y_{0},r)}(\omega)\leq t$ by $X_{t}(\omega)\in B(y_{0},r)$,
it follows from \eqref{eq:X-continuous-entrance} that
$X_{\dot{\sigma}_{B(y_{0},r)}}(\omega)\in\partial B(y_{0},r)$ and hence that
\begin{equation}\label{eq:sigma-n-leq-t}
\sigma_{n}(\omega)\leq\dot{\sigma}_{B(y_{0},r)}(\omega)<t
\qquad\textrm{for any }n\in\mathbb{N}.
\end{equation}
In particular, $\sigma_{n+1}(\omega)\leq\frac{1}{2}(\sigma_{n}(\omega)+t)$
for some $n\in\mathbb{N}$; indeed, otherwise for any $n\in\mathbb{N}$
we would have $\sigma_{n+1}(\omega)\geq\frac{1}{2}(\sigma_{n}(\omega)+t)$, or equivalently
$t-\sigma_{n+1}(\omega)\leq\frac{1}{2}(t-\sigma_{n}(\omega))$, and hence
$0<t-\dot{\sigma}_{B(y_{0},r)}(\omega)\leq t-\sigma_{n}(\omega)\leq 2^{1-n}(t-\sigma_{1}(\omega))$
by \eqref{eq:sigma-n-leq-t}, contradicting $\lim_{n\to\infty}2^{-n}=0$.
Thus, setting $\Omega_{1}:=\Omega$ and
\begin{equation}\label{eq:Omega-n}
\Omega_{n}:=\{\sigma_{k+1}>{\textstyle\frac{1}{2}}(\sigma_{k}+t)\textrm{ for any }k\in\{1,\dots,n-1\}\},
\quad n\in\mathbb{N}\setminus\{1\},
\end{equation}
we obtain
\begin{multline}\label{eq:sigma-n-t-gap}
\{X_{t}\in B(y_{0},r),\,X_{0}=x,\,[0,\zeta)\ni s\mapsto X_{s}\in M\textrm{ is continuous}\}\\
\subset\bigcup\nolimits_{n\in\mathbb{N}}\bigl(\Omega_{n}\cap\{\sigma_{n+1}\leq{\textstyle\frac{1}{2}}(\sigma_{n}+t)\}\bigr),
\quad\textrm{where the union is disjoint.}
\end{multline}
Note that $\Omega_{n}\in\sigmafield{F}_{\sigma_{n}}$ for any $n\in\mathbb{N}$
since $\sigmafield{F}_{\sigma_{k}}\subset\sigmafield{F}_{\sigma_{n}}$ by
$\sigma_{k}\leq\sigma_{n}$ and \cite[Lemma 1.2.15]{KS} for any $k\in\{1,\dots,n\}$.
Now by \eqref{eq:sigma-n-t-gap} along with $A\subset B(y_{0},r)$,
$\mathbb{P}_{x}[X_{0}=x]=1$ and \eqref{eq:X-continuous},
\begin{align}
\functionspace{P}^{U}_{t}(x,A)&=\mathbb{P}_{x}[X_{t}\in A,\,t<\tau_{U}]\notag\\
&=\mathbb{P}_{x}\Bigl[\{X_{t}\in A,\,t<\tau_{U}\}\cap
\bigcup\nolimits_{n\in\mathbb{N}}\bigl(\Omega_{n}\cap\{\sigma_{n+1}\leq{\textstyle\frac{1}{2}}(\sigma_{n}+t)\}\bigr)\Bigr]\notag\\
&=\sum_{n\in\mathbb{N}}\mathbb{P}_{x}\bigl[\{X_{t}\in A,\,t<\tau_{U}\}
\cap\bigl(\Omega_{n}\cap\{\sigma_{n+1}\leq{\textstyle\frac{1}{2}}(\sigma_{n}+t)\}\bigr)\bigr].
\label{eq:PUtxA-expansion}
\end{align}
Let $n\in\mathbb{N}$, set $\sigma_{n,t}:=\sigma_{n}\wedge t$ and
$\Omega'_{n}:=\Omega_{n}\cap\{\sigma_{n}\leq t,\,\sigma_{n,t}\leq\tau_{U}\}$,
so that $\Omega'_{n}\in\sigmafield{F}_{\sigma_{n,t}}$ by
$\Omega_{n}\in\sigmafield{F}_{\sigma_{n}}$ and \cite[Lemma 1.2.16]{KS}.
Then $\{X_{t}\in A\}\subset\{\sigma_{n}\leq t\}$ by $A\subset B(y_{0},r_{n})$,
clearly $\tau_{U}=\sigma_{n,t}+\tau_{U}\circ\theta_{\sigma_{n,t}}$
on $\{\sigma_{n,t}\leq\tau_{U}\}$,
and by $\sigma_{n}\leq\sigma_{n+1}$
we also have $\sigma_{n+1}=\sigma_{n}+\sigma_{n+1}\circ\theta_{\sigma_{n}}$,
which easily implies that
$\{\sigma_{n}\leq t,\,\sigma_{n+1}\leq\frac{1}{2}(\sigma_{n}+t)\}
=\{\sigma_{n}+2\sigma_{n+1}\circ\theta_{\sigma_{n}}\leq t\}$.
Therefore,
\begin{align}
&\{X_{t}\in A,\,t<\tau_{U}\}\cap\bigl(\Omega_{n}\cap\{\sigma_{n+1}\leq{\textstyle\frac{1}{2}}(\sigma_{n}+t)\}\notag\\
&=\{X_{t}\in A\}\cap\Omega_{n}\cap\{\sigma_{n}\leq t<\tau_{U},\,\sigma_{n+1}\leq{\textstyle\frac{1}{2}}(\sigma_{n}+t)\}\notag\\
&=\{X_{t}\in A\}\cap\Omega_{n}\cap\{\sigma_{n}\leq t<\tau_{U},\,\sigma_{n,t}\leq\tau_{U},\,\sigma_{n}+2\sigma_{n+1}\circ\theta_{\sigma_{n}}\leq t\}\notag\\
&=\{X_{t}\in A\}\cap\Omega'_{n}
\cap\{\sigma_{n,t}+2\sigma_{n+1}\circ\theta_{\sigma_{n,t}}
\leq t<\sigma_{n,t}+\tau_{U}\circ\theta_{\sigma_{n,t}}\}.
\label{eq:PUtxA-expansion-n-event}
\end{align}
Noting that $(\sigma_{n,t}+2\sigma_{n+1}\circ\theta_{\sigma_{n,t}})
\wedge(\sigma_{n,t}+\tau_{U}\circ\theta_{\sigma_{n,t}})
=\sigma_{n,t}+((2\sigma_{n+1})\wedge\tau_{U})\circ\theta_{\sigma_{n,t}}$,
we see from \eqref{eq:PUtxA-expansion-n-event},
$\Omega'_{n}\in\sigmafield{F}_{\sigma_{n,t}}$ and Proposition \ref{prop:strong-Markov} that
\begin{align}
&\mathbb{P}_{x}\bigl[\{X_{t}\in A,\,t<\tau_{U}\}\cap\bigl(\Omega_{n}\cap\{\sigma_{n+1}\leq{\textstyle\frac{1}{2}}(\sigma_{n}+t)\}\bigr)\bigr]\notag\\
&=\mathbb{E}_{x}\bigl[\ind{A}(X_{t})
\ind{\Omega'_{n}\cap\{\sigma_{n,t}+2\sigma_{n+1}\circ\theta_{\sigma_{n,t}}
\leq t<\sigma_{n,t}+\tau_{U}\circ\theta_{\sigma_{n,t}}\}}\bigr]\notag\\
&=\mathbb{E}_{x}\bigl[\ind{\Omega'_{n}}\ind{A}(X_{t})\bigl(\ind{\{t<\sigma_{n,t}+\tau_{U}\circ\theta_{\sigma_{n,t}}\}}
-\ind{\{t<\sigma_{n,t}+((2\sigma_{n+1})\wedge\tau_{U})\circ\theta_{\sigma_{n,t}}\}}\bigr)\bigr]\notag\\
&=\mathbb{E}_{x}\bigl[\ind{\Omega'_{n}}\mathbb{E}_{x}\bigl[\ind{A}(X_{t})\ind{\{t<\sigma_{n,t}+\tau_{U}\circ\theta_{\sigma_{n,t}}\}}\bigm|\sigmafield{F}_{\sigma_{n,t}}\bigr]\bigr]\notag\\
&\qquad\qquad-\mathbb{E}_{x}\bigl[\ind{\Omega'_{n}}\mathbb{E}_{x}\bigl[\ind{A}(X_{t})\ind{\{t<\sigma_{n,t}+((2\sigma_{n+1})\wedge\tau_{U})\circ\theta_{\sigma_{n,t}}\}}\bigm|\sigmafield{F}_{\sigma_{n,t}}\bigr]\bigr]\notag\\
&=\int_{\Omega'_{n}}\mathbb{E}_{X_{\sigma_{n,t}}(\omega)}\bigl[\ind{A}(X_{t-\sigma_{n,t}(\omega)})\ind{\{t-\sigma_{n,t}(\omega)<\tau_{U}\}}\bigr]\,d\mathbb{P}_{x}(\omega)\notag\\
&\qquad\qquad-\int_{\Omega'_{n}}\mathbb{E}_{X_{\sigma_{n,t}}(\omega)}\bigl[\ind{A}(X_{t-\sigma_{n,t}(\omega)})\ind{\{t-\sigma_{n,t}(\omega)<(2\sigma_{n+1})\wedge\tau_{U}\}}\bigr]\,d\mathbb{P}_{x}(\omega)\notag\\
&=\int_{\Omega'_{n}}\mathbb{E}_{X_{\sigma_{n,t}}(\omega)}\bigl[\ind{A}(X_{t-\sigma_{n,t}(\omega)})\bigl(\ind{\{t-\sigma_{n,t}(\omega)<\tau_{U}\}}-\ind{\{t-\sigma_{n,t}(\omega)<(2\sigma_{n+1})\wedge\tau_{U}\}}\bigr)\bigr]\,d\mathbb{P}_{x}(\omega)\notag\\
&=\int_{\Omega'_{n}}\mathbb{E}_{X_{\sigma_{n,t}}(\omega)}\bigl[\ind{A}(X_{t-\sigma_{n,t}(\omega)})\ind{\{2\sigma_{n+1}\leq t-\sigma_{n,t}(\omega)<\tau_{U}\}}\bigr]\,d\mathbb{P}_{x}(\omega)\notag\\
&=\int_{\Omega'_{n}\cap\{X_{\sigma_{n}}\in(\partial B(y_{0},r_{n}))\setminus N\}}\mathbb{E}_{X_{\sigma_{n,t}}(\omega)}\bigl[\ind{A}(X_{t-\sigma_{n,t}(\omega)})\ind{\{2\sigma_{n+1}\leq t-\sigma_{n,t}(\omega)<\tau_{U}\}}\bigr]\,d\mathbb{P}_{x}(\omega),
\label{eq:PUtxA-expansion-n-step1}
\end{align}
where the equality in the last line follows since
$\ind{\{\sigma_{n}\leq t\}}=\ind{\{\sigma_{n}\leq t,\,X_{\sigma_{n}}\in(\partial B(y_{0},r_{n}))\setminus N\}}$
$\mathbb{P}_{x}$-a.s.\ by $x\in M\setminus(N\cup B(y_{0},4r))$, $\mathbb{P}_{x}[X_{0}=x]=1$,
\eqref{eq:X-continuous}, \eqref{eq:X-continuous-entrance} and \eqref{eq:properly-exceptional}.
Let $\omega\in\Omega'_{n}\cap\{X_{\sigma_{n}}\in(\partial B(y_{0},r_{n}))\setminus N\}$,
set $s:=t-\sigma_{n,t}(\omega)$ and $z:=X_{\sigma_{n,t}}(\omega)$, so that
$\sigma_{n,t}(\omega)=\sigma_{n}(\omega)$, $s=t-\sigma_{n}(\omega)\in[0,t]$
and $z=X_{\sigma_{n}}(\omega)\in(\partial B(y_{0},r_{n}))\setminus N$ by
$\sigma_{n}(\omega)\leq t$. The integrand in \eqref{eq:PUtxA-expansion-n-step1}
is $\mathbb{E}_{z}[\ind{A}(X_{s})\ind{\{2\sigma_{n+1}\leq s<\tau_{U}\}}]$,
which is $0$ if $s=0$ by $\mathbb{P}_{z}[X_{0}=z]=1$ and
$z\not\in B(y_{0},r_{n})\supset B(y_{0},r)\supset A$.
Assume $s>0$ and set $\sigma_{n+1,s}:=\sigma_{n+1}\wedge s$.
Noting that $\tau_{U}=\sigma_{n+1,s}+\tau_{U}\circ\theta_{\sigma_{n+1,s}}$
on $\{\sigma_{n+1,s}\leq\tau_{U}\}$ and that
$\{\sigma_{n+1}\leq\frac{s}{2},\,\sigma_{n+1,s}\leq\tau_{U}\}=\{\sigma_{n+1}\leq\tau_{U}\wedge\frac{s}{2}\}\in\sigmafield{F}_{\sigma_{n+1,s}}$
by \cite[Lemma 1.2.16]{KS}, we see from Proposition \ref{prop:strong-Markov} that
\begin{align}
&\mathbb{E}_{z}[\ind{A}(X_{s})\ind{\{2\sigma_{n+1}\leq s<\tau_{U}\}}]\notag\\
&=\mathbb{E}_{z}\bigl[\ind{A}(X_{s})\ind{\{s<\tau_{U}\}}\ind{\{\sigma_{n+1}\leq s/2,\,\sigma_{n+1,s}\leq\tau_{U}\}}\bigr]\notag\\
&=\mathbb{E}_{z}\bigl[\ind{A}(X_{s})\ind{\{s<\sigma_{n+1,s}+\tau_{U}\circ\theta_{\sigma_{n+1,s}}\}}\ind{\{\sigma_{n+1}\leq\tau_{U}\wedge(s/2)\}}\bigr]\notag\\
&=\mathbb{E}_{z}\bigl[\ind{\{\sigma_{n+1}\leq\tau_{U}\wedge(s/2)\}}
\mathbb{E}_{z}\bigl[\ind{A}(X_{s})\ind{\{s<\sigma_{n+1,s}+\tau_{U}\circ\theta_{\sigma_{n+1,s}}\}}\bigm|\sigmafield{F}_{\sigma_{n+1,s}}\bigr]\bigr]\notag\\
&=\int_{\{\sigma_{n+1}\leq\tau_{U}\wedge(s/2)\}}\mathbb{E}_{X_{\sigma_{n+1,s}}(\omega')}
\bigl[\ind{A}(X_{s-\sigma_{n+1,s}(\omega')})\ind{\{s-\sigma_{n+1,s}(\omega')<\tau_{U}\}}\bigr]\,d\mathbb{P}_{z}(\omega')\notag\\
&=\mathbb{E}_{z}\bigl[\ind{\{\sigma_{n+1}\leq\tau_{U}\wedge(s/2),\,X_{\sigma_{n+1}}\in(\partial B(y_{0},r_{n+1}))\setminus N\}}
\functionspace{P}^{U}_{s-\sigma_{n+1}}(X_{\sigma_{n+1}},A)\bigr],
\label{eq:PUtxA-expansion-n-step2}
\end{align}
where again the last equality follows since
$\ind{\{\sigma_{n+1}\leq s/2\}}=\ind{\{\sigma_{n+1}\leq s/2,\,X_{\sigma_{n+1}}\in(\partial B(y_{0},r_{n+1}))\setminus N\}}$
$\mathbb{P}_{z}$-a.s.\ by $z\in M\setminus(N\cup B(y_{0},r_{n}))$, $\mathbb{P}_{z}[X_{0}=z]=1$,
\eqref{eq:X-continuous}, \eqref{eq:X-continuous-entrance} and \eqref{eq:properly-exceptional}.
Further let
$\omega'\in\{\sigma_{n+1}\leq\tau_{U}\wedge\frac{s}{2},\,X_{\sigma_{n+1}}\in(\partial B(y_{0},r_{n+1}))\setminus N\}$,
set $u:=s-\sigma_{n+1}(\omega')$ and $w:=X_{\sigma_{n+1}}(\omega')$, so that
$0<\frac{s}{2}\leq u\leq s\leq t<\Psi(R)$,
$w\in(\partial B(y_{0},r_{n+1}))\setminus N\subset U\setminus N$ and
$d(w,x)\leq d(w,y_{0})+d(y_{0},x)=r_{n+1}+4r/\varepsilon<(2+4/\varepsilon)r$.
Then by \textup{(DB)$_{\Psi}$} and the assumption that $t<\Psi((2+4/\varepsilon)r)$,
\begin{equation*}
\frac{F_{u}(w,y)}{F_{t}(x,y)}
\leq c_{F}\Bigl(\frac{t\vee\Psi(d(w,x))}{u}\Bigr)^{\alpha_{F}}
\leq c_{F}\Bigl(\frac{\Psi((2+4/\varepsilon)r)}{s/2}\Bigr)^{\alpha_{F}}
\end{equation*}
for any $y\in U$, which together with $A\subset U$ and \textup{(DU)$_{F}^{U,R}$} yields
\begin{align}
\functionspace{P}^{U}_{s-\sigma_{n+1}(\omega')}(X_{\sigma_{n+1}}(\omega'),A)
=\functionspace{P}^{U}_{u}(w,A)&\leq\int_{A}F_{u}(w,y)\,d\mu(y)\notag\\
&\leq c_{F}\Bigl(\frac{\Psi((2+4/\varepsilon)r)}{s/2}\Bigr)^{\alpha_{F}}
\int_{A}F_{t}(x,y)\,d\mu(y).
\label{eq:PUtxA-expansion-n-step3}
\end{align}
Recalling that
$z=X_{\sigma_{n}}(\omega)\in(\partial B(y_{0},r_{n}))\setminus N$, we have
$(z,r_{n}-r_{n+1})\in(U\setminus N)\times(0,R)$, $B(z,r_{n}-r_{n+1})\subset U$,
and $\tau_{B(z,r_{n}-r_{n+1})}\leq\sigma_{n+1}$ by
$B(y_{0},r_{n+1})\subset M\setminus B(z,r_{n}-r_{n+1})$.
Therefore it follows from \eqref{eq:PUtxA-expansion-n-step2},
\eqref{eq:PUtxA-expansion-n-step3} and
\textup{(P)$_{\Psi}^{U,R}$} for $(z,r_{n}-r_{n+1})$ that
\begin{align}
&\mathbb{E}_{z}[\ind{A}(X_{s})\ind{\{2\sigma_{n+1}\leq s<\tau_{U}\}}]\notag\\
&\leq c_{F}\Bigl(\frac{\Psi((2+4/\varepsilon)r)}{s/2}\Bigr)^{\alpha_{F}}
\mathbb{P}_{z}[\sigma_{n+1}\leq\tau_{U}\wedge{\textstyle\frac{s}{2}}]\int_{A}F_{t}(x,y)\,d\mu(y)\notag\\
&\leq c_{F}\Bigl(\frac{\Psi((2+4/\varepsilon)r)}{s/2}\Bigr)^{\alpha_{F}}
\mathbb{P}_{z}[\tau_{B(z,r_{n}-r_{n+1})}\leq{\textstyle\frac{s}{2}}]\int_{A}F_{t}(x,y)\,d\mu(y)\notag\\
&\leq cc_{F}\Bigl(\frac{\Psi((2+4/\varepsilon)r)}{s/2}\Bigr)^{\alpha_{F}}
\exp\bigl(-\Phi(\gamma(r_{n}-r_{n+1}),{\textstyle\frac{s}{2}})\bigr)\int_{A}F_{t}(x,y)\,d\mu(y).
\label{eq:PUtxA-expansion-n-step4}
\end{align}
We easily see from $\omega\in\Omega'_{n}\subset\Omega_{n}$ and \eqref{eq:Omega-n}
that $0<s=t-\sigma_{n}(\omega)\leq 2^{1-n}t$, and then by $t<\Psi((2+4/\varepsilon)r)$
we have $\Psi((2+4/\varepsilon)r)/(2^{n/2}s/2)\geq 2^{n/2}>1$,
which together with \eqref{eq:upper-bound-function-Phi-ULE},
$r_{n}-r_{n+1}=(1-2^{-1/(2\beta_{2})})2^{-n/(2\beta_{2})}r$,
\eqref{eq:upper-bound-function-Psi} and $1<\beta_{1}\leq\beta_{2}$ implies that
\begin{align}
\Phi(\gamma(r_{n}-r_{n+1}),{\textstyle\frac{s}{2}})
&\geq(c_{\Psi}2^{\beta_{1}})^{-\frac{1}{\beta_{1}-1}}
\min_{k\in\{1,2\}}\Bigl(\frac{\Psi(\gamma(r_{n}-r_{n+1}))}{s/2}\Bigr)^{\frac{1}{\beta_{k}-1}}\notag\\
&\geq c_{\varepsilon,1}\min_{k\in\{1,2\}}\Bigl(\frac{\Psi((2+4/\varepsilon)r)}{2^{n/2}s/2}\Bigr)^{\frac{1}{\beta_{k}-1}}\notag\\
&=c_{\varepsilon,1}\Bigl(\frac{\Psi((2+4/\varepsilon)r)}{2^{n/2}s/2}\Bigr)^{\frac{1}{\beta_{2}-1}}
\geq c_{\varepsilon,1}2^{\frac{n}{2(\beta_{2}-1)}}
\label{eq:PUtxA-expansion-n-step5}
\end{align}
for some $c_{\varepsilon,1}\in(0,\infty)$ explicit in $c_{\Psi},\beta_{1},\beta_{2},\gamma,\varepsilon$.
\eqref{eq:PUtxA-expansion-n-step5} in turn yields
\begin{align}
&\Bigl(\frac{\Psi((2+4/\varepsilon)r)}{s/2}\Bigr)^{\alpha_{F}}
\exp\bigl(-\Phi(\gamma(r_{n}-r_{n+1}),{\textstyle\frac{s}{2}})\bigr)\notag\\
&\leq\Bigl(\frac{\Psi((2+4/\varepsilon)r)}{s/2}\Bigr)^{\alpha_{F}}
\exp\biggl(-c_{\varepsilon,1}\Bigl(\frac{\Psi((2+4/\varepsilon)r)}{2^{n/2}s/2}\Bigr)^{\frac{1}{\beta_{2}-1}}\biggr)\notag\\
&\leq\Bigl(\frac{\Psi((2+4/\varepsilon)r)}{2^{n/2}s/2}\Bigr)^{\alpha_{F}}2^{\alpha_{F}n/2}
\exp\biggl(-\frac{c_{\varepsilon,1}}{2}\Bigl(\frac{\Psi((2+4/\varepsilon)r)}{2^{n/2}s/2}\Bigr)^{\frac{1}{\beta_{2}-1}}
-\frac{c_{\varepsilon,1}}{2}2^{\frac{n}{2(\beta_{2}-1)}}\biggr)\notag\\
&\leq c_{\varepsilon,2}2^{-\alpha_{F}n/2},
\label{eq:PUtxA-expansion-n-step6}
\end{align}
where
$c_{\varepsilon,2}:=2^{5\alpha_{F}(\beta_{2}-1)}
\bigl(\alpha_{F}(\beta_{2}-1)/(ec_{\varepsilon,1})\bigr)^{3\alpha_{F}(\beta_{2}-1)}$.
By \eqref{eq:PUtxA-expansion-n-step4} and \eqref{eq:PUtxA-expansion-n-step6},
\begin{equation}\label{eq:PUtxA-expansion-n-step7}
\mathbb{E}_{z}[\ind{A}(X_{s})\ind{\{2\sigma_{n+1}\leq s<\tau_{U}\}}]
\leq\frac{cc_{F}c_{\varepsilon,2}}{2^{\alpha_{F}n/2}}\int_{A}F_{t}(x,y)\,d\mu(y)
\end{equation}
for $s=t-\sigma_{n,t}(\omega)$ and $z=X_{\sigma_{n,t}}(\omega)$ for any
$\omega\in\Omega'_{n}\cap\{X_{\sigma_{n}}\in(\partial B(y_{0},r_{n}))\setminus N\}$,
and therefore from \eqref{eq:PUtxA-expansion-n-step1}, \eqref{eq:PUtxA-expansion-n-step7}
and $\Omega'_{n}\subset\{\sigma_{n}\leq t\}$ we obtain
\begin{multline}\label{eq:PUtxA-expansion-n-step8}
\mathbb{P}_{x}\bigl[\{X_{t}\in A,\,t<\tau_{U}\}\cap\bigl(\Omega_{n}\cap\{\sigma_{n+1}\leq{\textstyle\frac{1}{2}}(\sigma_{n}+t)\}\bigr)\bigr]\\
\leq\frac{cc_{F}c_{\varepsilon,2}}{2^{\alpha_{F}n/2}}\mathbb{P}_{x}[\sigma_{n}\leq t]\int_{A}F_{t}(x,y)\,d\mu(y).
\end{multline}
To conclude \eqref{eq:HKUB-localized-off-diag-local} from \eqref{eq:PUtxA-expansion}
and \eqref{eq:PUtxA-expansion-n-step8}, we show that
\begin{equation}\label{eq:PUtxA-expansion-n-step9}
\mathbb{P}_{x}[\sigma_{n}\leq t]\leq c\exp(-\Phi(\gamma r,t)).
\end{equation}
Indeed, setting $\sigma:=\dot{\sigma}_{B(y_{0},3r)}$, we have $\sigma\leq\sigma_{n}$
by $B(y_{0},r_{n})\subset B(y_{0},3r)$ and hence
$\sigma_{n}=\sigma+\sigma_{n}\circ\theta_{\sigma}$. Therefore
$\{\sigma_{n}\leq t\}\subset\{\sigma\leq t,\,\sigma_{n}\circ\theta_{\sigma}\leq t\}$,
and then by the strong Markov property \cite[Theorem A.1.21]{CF} of $X$ at time $\sigma$,
\begin{align}
\mathbb{P}_{x}[\sigma_{n}\leq t]
\leq\mathbb{P}_{x}[\sigma\leq t,\,\sigma_{n}\circ\theta_{\sigma}\leq t]
&=\mathbb{E}_{x}[\ind{\{\sigma\leq t\}}(\ind{\{\sigma_{n}\leq t\}}\circ\theta_{\sigma})]\notag\\
&=\mathbb{E}_{x}\bigl[\ind{\{\sigma\leq t\}}\mathbb{E}_{X_{\sigma}}[\ind{\{\sigma_{n}\leq t\}}]\bigr]\notag\\
&=\mathbb{E}_{x}\bigl[\ind{\{\sigma\leq t,\,X_{\sigma}\in(\partial B(y_{0},3r))\setminus N\}}
\mathbb{P}_{X_{\sigma}}[\sigma_{n}\leq t]\bigr],
\label{eq:PUtxA-expansion-n-step10}
\end{align}
where the last equality follows since
$\ind{\{\sigma\leq t\}}=\ind{\{\sigma\leq t,\,X_{\sigma}\in(\partial B(y_{0},3r))\setminus N\}}$
$\mathbb{P}_{x}$-a.s.\ by $x\in M\setminus(N\cup B(y_{0},4r))$, $\mathbb{P}_{x}[X_{0}=x]=1$,
\eqref{eq:X-continuous}, \eqref{eq:X-continuous-entrance} and \eqref{eq:properly-exceptional}.
Moreover, for $z\in(\partial B(y_{0},3r))\setminus N$,
$B(y_{0},r_{n})\subset M\setminus B(z,r)$ by $r_{n}<2r$,
hence $\sigma_{n}\geq\tau_{B(z,r)}$, and therefore noting that
$(z,r)\in(U\setminus N)\times(0,R)$ and that $B(z,r)\subset B(y_{0},4r)\subset U$,
we see from \textup{(P)$_{\Psi}^{U,R}$} for $(z,r)$ that
$\mathbb{P}_{z}[\sigma_{n}\leq t]\leq\mathbb{P}_{z}[\tau_{B(z,r)}\leq t]
\leq c\exp(-\Phi(\gamma r,t))$,
which together with \eqref{eq:PUtxA-expansion-n-step10} yields
\eqref{eq:PUtxA-expansion-n-step9}.
Now \eqref{eq:HKUB-localized-off-diag-local} with
$c'_{\varepsilon}:=c^{2}c_{F}c_{\varepsilon,2}/(2^{\alpha_{F}/2}-1)$ is
immediate from \eqref{eq:PUtxA-expansion}, \eqref{eq:PUtxA-expansion-n-step8},
\eqref{eq:PUtxA-expansion-n-step9} and the fact that
$d(x,y)\leq d(x,y_{0})+d(y_{0},y)<4r/\varepsilon+r<5r/\varepsilon$
for any $y\in A$ by $A\subset B(y_{0},r)$.
Finally, we prove \eqref{eq:HKUB-localized-off-diag-local} for general
$A\in\sigmafield{B}(U^{\circ}_{\varepsilon R})$. Note that
\eqref{eq:HKUB-localized-off-diag-local} holds for $A=\{x\}$ by
\textup{(DU)$_{F}^{U,R}$}. In particular, \eqref{eq:HKUB-localized-off-diag-local}
is valid for any $A\in\sigmafield{B}(U^{\circ}_{\varepsilon R})$ if
$U^{\circ}_{\varepsilon R}\setminus\{x\}=\emptyset$, and thus we may assume
$U^{\circ}_{\varepsilon R}\setminus\{x\}\not=\emptyset$. Let $\{y_{k}\}_{k\in\mathbb{N}}$
be a countable dense subset of $U^{\circ}_{\varepsilon R}\setminus\{x\}$
and set $B_{0}:=U^{\circ}_{\varepsilon R}\cap\{x\}$,
$B_{1}:=B(y_{1},\frac{1}{4}\varepsilon d(x,y_{1}))$ and
$B_{k}:=B(y_{k},\frac{1}{4}\varepsilon d(x,y_{k}))\setminus\bigcup_{j=1}^{k-1}B(y_{j},\frac{1}{4}\varepsilon d(x,y_{j}))$
for $k\in\mathbb{N}\setminus\{1\}$. Then
$\{B_{k}\}_{k\in\mathbb{N}\cup\{0\}}\subset\sigmafield{B}(U)$, and it is easy to
see that $U^{\circ}_{\varepsilon R}\subset\bigcup_{k\in\mathbb{N}\cup\{0\}}B_{k}$,
where the union is disjoint. Now for any $A\in\sigmafield{B}(U^{\circ}_{\varepsilon R})$,
since $A\cap B_{0}\in\{\emptyset,\{x\}\}$ and
$A\cap B_{k}\in\sigmafield{B}\bigl(B(y_{k},\frac{1}{4}\varepsilon d(x,y_{k}))\bigr)$
for $k\in\mathbb{N}$, we have already proved \eqref{eq:HKUB-localized-off-diag-local}
with $A\cap B_{k}$ in place of $A$ for any $k\in\mathbb{N}\cup\{0\}$, and therefore
\begin{align}
\functionspace{P}^{U}_{t}(x,A)
&=\functionspace{P}^{U}_{t}\Bigl(x,\bigcup\nolimits_{k\in\mathbb{N}\cup\{0\}}(A\cap B_{k})\Bigr)
=\sum_{k\in\mathbb{N}\cup\{0\}}\functionspace{P}^{U}_{t}(x,A\cap B_{k})\notag\\
&\leq\sum_{k\in\mathbb{N}\cup\{0\}}\int_{A\cap B_{k}}c'_{\varepsilon}F_{t}(x,y)\exp\bigl(-\Phi(\gamma_{\varepsilon}d(x,y),t)\bigr)\,d\mu(y)\notag\\
&=\int_{A}c'_{\varepsilon}F_{t}(x,y)\exp\bigl(-\Phi(\gamma_{\varepsilon}d(x,y),t)\bigr)\,d\mu(y)
\label{eq:HKUB-localized-off-diag-local-end}
\end{align}
by monotone convergence, completing the proof of Proposition \ref{prop:HKUB-localized}.
\end{proof}
Theorems \ref{thm:HKUB-localized} and \ref{thm:HKUB-localized-global}
are easy consequences of Propositions \ref{prop:HK-existence},
\ref{prop:HKUB-localized}, and \ref{prop:HKUB-localized-difference} below.
\begin{proposition}\label{prop:HKUB-localized-difference}
Under the same assumptions as those of Theorem \textup{\ref{thm:HKUB-localized}},
there exists $c''_{\varepsilon}\in(0,\infty)$ explicit in
$c_{\Psi},\beta_{1},\beta_{2},c_{F},\alpha_{F},c,\gamma,\varepsilon$ such that, with
$\gamma_{\varepsilon}:=\frac{1}{5}\varepsilon\gamma$,
for any $(t,x)\in(0,\Psi(R))\times(M\setminus N)$ and
any $A\in\sigmafield{B}(U^{\circ}_{\varepsilon R})$,
\begin{equation}\label{eq:HKUB-localized-difference}
\functionspace{P}_{t}(x,A)\leq\functionspace{P}^{U}_{t}(x,A)
+c''_{\varepsilon}(\inf\nolimits_{U\times U}F_{t})
\exp(-\Phi(\gamma_{\varepsilon}R,t))\mu(A).
\end{equation}
\end{proposition}
\begin{proof}
If $U=M$, then \eqref{eq:HKUB-localized-difference} is trivially valid since
$X_{t}=X^{U}_{t}$ and hence $\functionspace{P}_{t}=\functionspace{P}^{U}_{t}$
for any $t\in[0,\infty)$. Therefore we may assume $U\not=M$.
Set $B:=U^{\circ}_{(\varepsilon/2)R}$, so that $B$ is open in $M$ and
$\overline{B}\subset U$, and define $\sigmafield{F}_{*}$-stopping times $\tau_{n}$
and $\sigma_{n}$, $n\in\mathbb{N}$, by \eqref{eq:multiple-DH-stopping-times}.
For each $n\in\mathbb{N}$, as noted at the beginning of the proof of
Theorem \ref{thm:multiple-DH}, on $\{\sigma_{n}<\infty\}$ we have
$X_{\sigma_{n}}\in\overline{B}\subset U$, $\tau_{n}\leq\sigma_{n}<\zeta$,
hence $X_{\tau_{n}}\in M\setminus U$ and $\tau_{n}<\sigma_{n}$
by the sample path right-continuity of $X$, and we also easily see that
\begin{equation}\label{eq:multiple-DH-sigma-n-boundary}
X_{\sigma_{n}}\in(\partial B)\setminus N\quad\textrm{on}\quad
\{\sigma_{n}<\infty=\dot{\sigma}_{N},\,[0,\zeta)\ni t\mapsto X_{t}\in M\textrm{ is continuous}\}.
\end{equation}
Let $(t,x)\in(0,\Psi(R))\times(M\setminus N)$ and
$A\in\sigmafield{B}(U^{\circ}_{\varepsilon R})$. Since
$\ind{A}|_{M\setminus B}=0$ by $A\subset U^{\circ}_{\varepsilon R}\subset B$,
from Theorem \ref{thm:multiple-DH} with $u=\ind{A}$ we obtain
\begin{equation}\label{eq:multiple-DH-apply}
\functionspace{P}_{t}(x,A)=\functionspace{P}^{U}_{t}(x,A)
+\sum_{n\in\mathbb{N}}\mathbb{E}_{x}\bigl[\ind{\{\sigma_{n}\leq t\}}
\functionspace{P}^{U}_{t-\sigma_{n}}(X_{\sigma_{n}},A)\bigr].
\end{equation}
Noting \eqref{eq:multiple-DH-sigma-n-boundary}, to estimate each term of the series
in \eqref{eq:multiple-DH-apply} let $s\in[0,t]$, $z\in(\partial B)\setminus N$ and
let $c'_{\varepsilon}\in(0,\infty)$ and $\gamma_{\varepsilon}=\frac{1}{5}\varepsilon\gamma$
be as in Proposition \ref{prop:HKUB-localized}. We claim that
\begin{equation}\label{eq:multiple-DH-apply-remainder-proof1}
\functionspace{P}^{U}_{s}(z,A)\leq c'_{\varepsilon}c_{F}c_{\varepsilon,3}
(\inf\nolimits_{U\times U}F_{t})\mu(A)
\end{equation}
for some $c_{\varepsilon,3}\in(0,\infty)$ explicit in
$c_{\Psi},\beta_{1},\beta_{2},\alpha_{F},\gamma,\varepsilon$.
Indeed, \eqref{eq:multiple-DH-apply-remainder-proof1} trivially holds for $s=0$
since $\functionspace{P}^{U}_{0}(z,A)=\mathbb{P}_{z}[X_{0}\in A,\,0<\tau_{U}]=0$
by $\mathbb{P}_{z}[X_{0}=z]=1$ and $z\not\in B\supset A$, and thus we may assume
$s\in(0,t]$. Then $s\in(0,\Psi(R))$, $z\in U\setminus N$ by $\overline{B}\subset U$,
and hence an application of Proposition \ref{prop:HKUB-localized} yields
\eqref{eq:HKUB-localized-off-diag-local} with $(s,z)$ in place of $(t,x)$.
Let $y\in U^{\circ}_{\varepsilon R}$ and $x_{0},y_{0}\in U$.
By \textup{(DB)$_{\Psi}$}, $0<s\leq t<\Psi(R)$ and $\diam U\leq R$ we have
$F_{s}(z,y)\leq c_{F}(\Psi(R)/s)^{\alpha_{F}}F_{t}(x_{0},y_{0})$,
and furthermore we easily see from
$z\in\partial B=\partial U^{\circ}_{(\varepsilon/2)R}$ that
$d(z,y)>\frac{1}{2}\varepsilon R$, so that
$\exp\bigl(-\Phi(\gamma_{\varepsilon}d(z,y),s)\bigr)
\leq\exp\bigl(-\Phi(\frac{1}{2}\varepsilon\gamma_{\varepsilon}R,s)\bigr)$
by the monotonicity of $\Phi(\cdot,s)$.
These facts, \eqref{eq:upper-bound-function-Phi-ULE}
and \eqref{eq:upper-bound-function-Psi} together imply that
\begin{align*}
&F_{s}(z,y)\exp\bigl(-\Phi(\gamma_{\varepsilon}d(z,y),s)\bigr)\\
&\leq c_{F}\Bigl(\frac{\Psi(R)}{s}\Bigr)^{\alpha_{F}}F_{t}(x_{0},y_{0})
\exp\bigl(-\Phi({\textstyle\frac{1}{2}}\varepsilon\gamma_{\varepsilon}R,s)\bigr)\\
&\leq c_{F}\Bigl(\frac{\Psi(R)}{s}\Bigr)^{\alpha_{F}}F_{t}(x_{0},y_{0})
\exp\biggl(-(c_{\Psi}2^{\beta_{1}})^{-\frac{1}{\beta_{1}-1}}
\min_{k\in\{1,2\}}\Bigl(\frac{\Psi(\frac{1}{2}\varepsilon\gamma_{\varepsilon}R)}{s}\Bigr)^{\frac{1}{\beta_{k}-1}}\biggr)\\
&\leq c_{F}c_{\varepsilon,3}F_{t}(x_{0},y_{0}),
\end{align*}
and taking the infimum in $(x_{0},y_{0})\in U\times U$ shows that for any $y\in U^{\circ}_{\varepsilon R}$,
\begin{equation}\label{eq:multiple-DH-apply-remainder-proof2}
F_{s}(z,y)\exp\bigl(-\Phi(\gamma_{\varepsilon}d(z,y),s)\bigr)
\leq c_{F}c_{\varepsilon,3}(\inf\nolimits_{U\times U}F_{t}).
\end{equation}
Then \eqref{eq:multiple-DH-apply-remainder-proof1} is immediate from
\eqref{eq:HKUB-localized-off-diag-local} with $(s,z)$ in place of $(t,x)$,
$A\subset U^{\circ}_{\varepsilon R}$ and \eqref{eq:multiple-DH-apply-remainder-proof2}.
Let $n\in\mathbb{N}$. By \eqref{eq:properly-exceptional},
\eqref{eq:X-continuous}, \eqref{eq:multiple-DH-sigma-n-boundary} and
\eqref{eq:multiple-DH-apply-remainder-proof1},
\begin{align}
\mathbb{E}_{x}\bigl[\ind{\{\sigma_{n}\leq t\}}\functionspace{P}^{U}_{t-\sigma_{n}}(X_{\sigma_{n}},A)\bigr]
&=\mathbb{E}_{x}\bigl[\ind{\{\sigma_{n}\leq t,\,X_{\sigma_{n}}\in(\partial B)\setminus N\}}\functionspace{P}^{U}_{t-\sigma_{n}}(X_{\sigma_{n}},A)\bigr]\notag\\
&\leq c'_{\varepsilon}c_{F}c_{\varepsilon,3}(\inf\nolimits_{U\times U}F_{t})\mu(A)
\mathbb{P}_{x}[\sigma_{n}\leq t],
\label{eq:multiple-DH-apply-remainder-proof3}
\end{align}
and we need to estimate $\mathbb{P}_{x}[\sigma_{n}\leq t]$. Recall that
$\tau_{n}\leq\sigma_{n}\leq\tau_{n+1}=\sigma_{n}+\tau_{U}\circ\theta_{\sigma_{n}}$
as mentioned in the proof of Theorem \ref{thm:multiple-DH}. Assume $n\geq 2$.
For each $\omega\in\{\sigma_{n}\leq t\}$, since
$0\leq\sigma_{k}(\omega)\leq\sigma_{n}(\omega)\leq t$
for any $k\in\{1,\dots,n\}$, we have
$t\geq\sigma_{n}(\omega)\geq\sigma_{n}(\omega)-\sigma_{1}(\omega)
=\sum_{k=1}^{n-1}(\sigma_{k+1}(\omega)-\sigma_{k}(\omega))$
and therefore
$\tau_{U}\circ\theta_{\sigma_{k}}(\omega)=\tau_{k+1}(\omega)-\sigma_{k}(\omega)
\leq\sigma_{k+1}(\omega)-\sigma_{k}(\omega)\leq\frac{t}{n-1}$
for some $k\in\{1,\dots,n-1\}$. Thus
$\{\sigma_{n}\leq t\}\subset\bigcup_{k=1}^{n-1}\{\sigma_{k}\leq t,\,\tau_{U}\circ\theta_{\sigma_{k}}\leq\frac{t}{n-1}\}$
and hence
\begin{equation}\label{eq:multiple-DH-apply-remainder-proof4}
\mathbb{P}_{x}[\sigma_{n}\leq t]
\leq\mathbb{P}_{x}\Biggl[\bigcup_{k=1}^{n-1}\{\sigma_{k}\leq t,\,\tau_{U}\circ\theta_{\sigma_{k}}\leq{\textstyle\frac{t}{n-1}}\}\Biggr]
\leq\sum_{k=1}^{n-1}\mathbb{P}_{x}[\sigma_{k}\leq t,\,\tau_{U}\circ\theta_{\sigma_{k}}\leq{\textstyle\frac{t}{n-1}}].
\end{equation}
Furthermore by using first the strong Markov property \cite[Theorem A.1.21]{CF}
of $X$ at time $\sigma_{k}$ and then \eqref{eq:properly-exceptional},
\eqref{eq:X-continuous} and \eqref{eq:multiple-DH-sigma-n-boundary}
we see that for any $k\in\{1,\dots,n-1\}$,
\begin{align}
\mathbb{P}_{x}[\sigma_{k}\leq t,\,\tau_{U}\circ\theta_{\sigma_{k}}\leq{\textstyle\frac{t}{n-1}}]
&=\mathbb{E}_{x}[\ind{\{\sigma_{k}\leq t\}}(\ind{\{\tau_{U}\leq t/(n-1)\}}\circ\theta_{\sigma_{k}})]\notag\\
&=\mathbb{E}_{x}\bigl[\ind{\{\sigma_{k}\leq t\}}\mathbb{P}_{X_{\sigma_{k}}}[\tau_{U}\leq{\textstyle\frac{t}{n-1}}]\bigr]\notag\\
&=\mathbb{E}_{x}\bigl[\ind{\{\sigma_{k}\leq t,\,X_{\sigma_{k}}\in(\partial B)\setminus N\}}\mathbb{P}_{X_{\sigma_{k}}}[\tau_{U}\leq{\textstyle\frac{t}{n-1}}]\bigr]\notag\\
&\leq c\exp\bigl(-\Phi({\textstyle\frac{1}{2}}\varepsilon\gamma R,{\textstyle\frac{t}{n-1}})\bigr);
\label{eq:multiple-DH-apply-remainder-proof5}
\end{align}
here the last inequality follows from the fact that for any
$z\in(\partial B)\setminus N$, $\tau_{B(z,\varepsilon R/2)}\leq\tau_{U}$
by $B(z,\frac{1}{2}\varepsilon R)\subset U$ and hence
$\mathbb{P}_{z}[\tau_{U}\leq{\textstyle\frac{t}{n-1}}]
\leq\mathbb{P}_{z}[\tau_{B(z,\varepsilon R/2)}\leq{\textstyle\frac{t}{n-1}}]
\leq c\exp\bigl(-\Phi({\textstyle\frac{1}{2}}\varepsilon\gamma R,{\textstyle\frac{t}{n-1}})\bigr)$
by \textup{(P)$_{\Psi}^{U,R}$} for $(z,\frac{1}{2}\varepsilon R)$.
Also, by the monotonicity of $\Phi(\cdot,\frac{t}{n-1})$ and
$\Phi(\gamma_{\varepsilon}R,\cdot)$, \eqref{eq:upper-bound-function-Phi-constant},
\eqref{eq:upper-bound-function-Phi-ULE}, \eqref{eq:upper-bound-function-Psi},
$1<\beta_{1}\leq\beta_{2}$ and $t<\Psi(R)$,
for some $c_{\varepsilon,4}\in(0,\infty)$ explicit in
$c_{\Psi},\beta_{1},\beta_{2},\gamma_{\varepsilon}$,
\begin{align}
\Phi({\textstyle\frac{1}{2}}\varepsilon\gamma R,{\textstyle\frac{t}{n-1}})
&\geq\Phi(2\gamma_{\varepsilon}R,{\textstyle\frac{t}{n-1}})
\geq 2\Phi(\gamma_{\varepsilon}R,{\textstyle\frac{t}{n-1}})\notag\\
&\geq\Phi(\gamma_{\varepsilon}R,t)+(c_{\Psi}2^{\beta_{1}})^{-\frac{1}{\beta_{1}-1}}
\min_{k\in\{1,2\}}\Bigl(\frac{\Psi(\gamma_{\varepsilon}R)}{t/(n-1)}\Bigr)^{\frac{1}{\beta_{k}-1}}\notag\\
&\geq\Phi(\gamma_{\varepsilon}R,t)+c_{\varepsilon,4}(n-1)^{\frac{1}{\beta_{2}-1}}.
\label{eq:multiple-DH-apply-remainder-proof6}
\end{align}
From \eqref{eq:multiple-DH-apply-remainder-proof4},
\eqref{eq:multiple-DH-apply-remainder-proof5} and
\eqref{eq:multiple-DH-apply-remainder-proof6} we conclude that
for any $n\in\mathbb{N}\setminus\{1\}$,
\begin{align}
\mathbb{P}_{x}[\sigma_{n}\leq t]
&\leq c(n-1)e^{-c_{\varepsilon,4}(n-1)^{\frac{1}{\beta_{2}-1}}}\exp(-\Phi(\gamma_{\varepsilon}R,t))\notag\\
&\leq cc_{\varepsilon,5}(n-1)^{-2}\exp(-\Phi(\gamma_{\varepsilon}R,t)),
\label{eq:multiple-DH-apply-remainder-proof7}
\end{align}
where $c_{\varepsilon,5}:=\bigl(3(\beta_{2}-1)/(ec_{\varepsilon,4})\bigr)^{3(\beta_{2}-1)}$.
For $\mathbb{P}_{x}[\sigma_{1}\leq t]$, set $B':=U^{\circ}_{(\varepsilon/4)R}$
and $\sigma:=\dot{\sigma}_{B',\tau_{U}}=\dot{\sigma}_{B',\tau_{1}}$
(recall Definition \ref{dfn:entrance-exit-after-sigma}), so that we have
\eqref{eq:multiple-DH-sigma-n-boundary} with $B'$ and $\sigma$ in place of $B$ and
$\sigma_{n}$, respectively, by substituting $\frac{1}{2}\varepsilon$ for $\varepsilon$.
Noting that $\sigma_{1}=\sigma+\dot{\sigma}_{B}\circ\theta_{\sigma}$ by
$B\subset B'$ and thus that
$\{\sigma_{1}\leq t\}\subset\{\sigma\leq t,\,\dot{\sigma}_{B}\circ\theta_{\sigma}\leq t\}$,
from the strong Markov property \cite[Theorem A.1.21]{CF} of
$X$ at time $\sigma$, \eqref{eq:properly-exceptional}, \eqref{eq:X-continuous}
and \eqref{eq:multiple-DH-sigma-n-boundary} we obtain
\begin{align}
\mathbb{P}_{x}[\sigma_{1}\leq t]
\leq\mathbb{P}_{x}[\sigma\leq t,\,\dot{\sigma}_{B}\circ\theta_{\sigma}\leq t]
&=\mathbb{E}_{x}[\ind{\{\sigma\leq t\}}(\ind{\{\dot{\sigma}_{B}\leq t\}}\circ\theta_{\sigma})]\notag\\
&=\mathbb{E}_{x}\bigl[\ind{\{\sigma\leq t\}}\mathbb{P}_{X_{\sigma}}[\dot{\sigma}_{B}\leq t]\bigr]\notag\\
&=\mathbb{E}_{x}\bigl[\ind{\{\sigma\leq t,\,X_{\sigma}\in(\partial B')\setminus N\}}\mathbb{P}_{X_{\sigma}}[\dot{\sigma}_{B}\leq t]\bigr]\notag\\
&\leq c\exp(-\Phi(\gamma_{\varepsilon}R,t));
\label{eq:multiple-DH-apply-remainder-proof8}
\end{align}
here, similarly to \eqref{eq:multiple-DH-apply-remainder-proof5},
the last inequality holds since for any
$z\in(\partial B')\setminus N$, $\tau_{B(z,\varepsilon R/4)}\leq\dot{\sigma}_{B}$
by $B(z,\frac{1}{4}\varepsilon R)\subset U\setminus B$ and hence
$\mathbb{P}_{z}[\dot{\sigma}_{B}\leq t]
\leq\mathbb{P}_{z}[\tau_{B(z,\varepsilon R/4)}\leq t]
\leq c\exp(-\Phi(\gamma_{\varepsilon}R,t))$
by \textup{(P)$_{\Psi}^{U,R}$} for $(z,\frac{1}{4}\varepsilon R)$ and
the monotonicity of $\Phi(\cdot,t)$.
Now \eqref{eq:HKUB-localized-difference} with
$c''_{\varepsilon}:=cc'_{\varepsilon}c_{F}c_{\varepsilon,3}(2c_{\varepsilon,5}+1)$
is immediate from \eqref{eq:multiple-DH-apply},
\eqref{eq:multiple-DH-apply-remainder-proof3},
\eqref{eq:multiple-DH-apply-remainder-proof7} and
\eqref{eq:multiple-DH-apply-remainder-proof8},
completing the proof of Proposition \ref{prop:HKUB-localized-difference}.
\end{proof}
\begin{proof}[Proof of Theorem \textup{\ref{thm:HKUB-localized}}]
Let $c'_{\varepsilon},c''_{\varepsilon}\in(0,\infty)$ be as in
Propositions \ref{prop:HKUB-localized} and \ref{prop:HKUB-localized-difference},
respectively, and let $\gamma_{\varepsilon}:=\frac{1}{5}\varepsilon\gamma$.
We show that Theorem \ref{thm:HKUB-localized} can be concluded from
Proposition \ref{prop:HK-existence} applied to $I=(0,\infty)$,
$V=M\setminus N$, $W=U^{\circ}_{\varepsilon R}$, $M$ in place of $U$, and
$H=H_{t}(x,y):(0,\infty)\times(M\setminus N)\times U^{\circ}_{\varepsilon R}\to[0,\infty)$
given by
\begin{equation}\label{eq:HKUB-localized-dfn-H}
H_{t}(x,y):=
\begin{cases}
(c'_{\varepsilon}+c''_{\varepsilon})F_{t}(x,y)\exp\bigl(-\Phi(\gamma_{\varepsilon}d(x,y),t)\bigr)
&\textrm{if }t<\Psi(R)\textrm{ and }x\in U,\\
c''_{\varepsilon}c_{F}2^{\alpha_{F}}(\inf_{U\times U}F_{\Psi(R)/2^{n}})\exp(-\Phi(\gamma_{\varepsilon}R,t))
&\textrm{if }\frac{\Psi(R)}{2^{n+1}}\leq t<\frac{\Psi(R)}{2^{n}}\textrm{ and }x\not\in U,\\
c_{\varepsilon}(\inf_{U\times U}F_{\Psi(R)}) &\textrm{if }t\geq\Psi(R),
\end{cases}
\end{equation}
where $n\in\mathbb{N}\cup\{0\}$ in the second line and
$c_{\varepsilon}:=((c'_{\varepsilon}+c''_{\varepsilon})c_{F}2^{\alpha_{F}})
\vee(c''_{\varepsilon}c_{F}^{2}2^{2\alpha_{F}})$.
Obviously $H=H_{t}(x,y)$ is Borel measurable, and by using \textup{(DB)$_{\Psi}$}
it is easily seen to be less than or equal to the right-hand side of
\eqref{eq:HKUB-localized-off-diag-U}, so that it remains to verify that
\begin{equation}\label{eq:HK-existence-global}
\functionspace{P}_{t}(x,A)\leq\int_{A}H_{t}(x,y)\,d\mu(y)
\end{equation}
for any $(t,x)\in(0,\infty)\times(M\setminus N)$ and any
$A\in\sigmafield{B}(U^{\circ}_{\varepsilon R})$. Note that by
\textup{(DB)$_{\Psi}$} and $\diam U\leq R$ we also have
\begin{equation}\label{eq:HKUB-localized-proof}
H_{\Psi(R)/2}(z,y)\leq c_{\varepsilon}(\inf\nolimits_{U\times U}F_{\Psi(R)})\qquad
\textrm{for any }(z,y)\in(M\setminus N)\times U^{\circ}_{\varepsilon R}.
\end{equation}
Let $(t,x)\in(0,\infty)\times(M\setminus N)$ and $A\in\sigmafield{B}(U^{\circ}_{\varepsilon R})$.
If $t<\Psi(R)$ and $x\in U$, then \eqref{eq:HK-existence-global} easily follows from
Propositions \ref{prop:HKUB-localized} and \ref{prop:HKUB-localized-difference}
in view of the fact that $\Phi(\gamma_{\varepsilon}R,t)\geq\Phi(\gamma_{\varepsilon}d(x,y),t)$
for any $y\in U^{\circ}_{\varepsilon R}$ by $\diam U\leq R$ and the monotonicity of
$\Phi(\cdot,t)$. If $t<\Psi(R)$ and $x\not\in U$, then we see from \textup{(DB)$_{\Psi}$} that
$c''_{\varepsilon}(\inf\nolimits_{U\times U}F_{t})\exp(-\Phi(\gamma_{\varepsilon}R,t))\leq H_{t}(x,y)$
for any $y\in U^{\circ}_{\varepsilon R}$, which together with
Proposition \ref{prop:HKUB-localized-difference} and
$\functionspace{P}^{U}_{t}(x,A)=0$ immediately implies \eqref{eq:HK-existence-global}.
Now assume $t\geq\Psi(R)$. Since
$\functionspace{P}_{t-\Psi(R)/2}(x,N)=\mathbb{P}_{x}[X_{t-\Psi(R)/2}\in N]=0$
by \eqref{eq:properly-exceptional} and
\begin{equation*}
\functionspace{P}_{\Psi(R)/2}(z,A)\leq\int_{A}H_{\Psi(R)/2}(z,y)\,d\mu(y)\leq
\int_{A}H_{t}(x,y)\,d\mu(y)
\end{equation*}
for any $z\in M\setminus N$ by the previous paragraph,
\eqref{eq:HKUB-localized-proof} and \eqref{eq:HKUB-localized-dfn-H},
from \eqref{eq:Pt-semigroup} we get
\begin{align*}
\functionspace{P}_{t}(x,A)
=\functionspace{P}_{t-\Psi(R)/2}(\functionspace{P}_{\Psi(R)/2}\ind{A})(x)
&=\int_{M\setminus N}\functionspace{P}_{\Psi(R)/2}(z,A)\functionspace{P}_{t-\Psi(R)/2}(x,dz)\\
&\leq\int_{M\setminus N}\biggl(\int_{A}H_{t}(x,y)\,d\mu(y)\biggr)\functionspace{P}_{t-\Psi(R)/2}(x,dz)\\
&\leq\int_{A}H_{t}(x,y)\,d\mu(y).
\end{align*}
Thus \eqref{eq:HK-existence-global} has been proved and hence
Theorem \ref{thm:HKUB-localized} follows from Proposition \ref{prop:HK-existence}.
\end{proof}
\begin{proof}[Proof of Theorem \textup{\ref{thm:HKUB-localized-global}}]
Define $H=H_{t}(x,y):(0,\infty)\times(M\setminus N)\times M\to[0,\infty)$
by the right-hand side of \eqref{eq:HKUB-localized-off-diag}, so that it is
clearly Borel measurable. Thanks to Proposition \ref{prop:HK-existence},
it suffices to show \eqref{eq:HK-existence-global} for any
$(t,x)\in(0,\infty)\times(M\setminus N)$ and any $A\in\sigmafield{B}(M)$
for some $c'\in(0,\infty)$ explicit in
$c_{\Psi},\beta_{1},\beta_{2},c_{F},\alpha_{F},c,\gamma$.
For applications of Theorem \ref{thm:HKUB-localized} and
Proposition \ref{prop:HKUB-localized-difference},
we remark that for any $(y_{0},R')\in M\times(0,\infty)$,
\begin{equation}\label{eq:U-ball}
\textrm{if we set}\quad U:=B(y_{0},{\textstyle\frac{R'}{2}})
\quad\textrm{then}\quad\diam U\leq R'\quad\textrm{and}\quad
B(y_{0},{\textstyle\frac{R'}{4}})\subset U^{\circ}_{(1/4)R'}.
\end{equation}
Let $(t,x)\in(0,\infty)\times(M\setminus N)$. If $R=\infty$, then for any
$A\in\sigmafield{B}(M)$ and any $n\in\mathbb{N}$ with $n>\Psi^{-1}(t)$,
in view of \eqref{eq:U-ball} we can apply Theorem \ref{thm:HKUB-localized} with
$n,B(x,\frac{n}{2}),\frac{1}{4}$ in place of $R,U,\varepsilon$ respectively
and $A\cap B(x,\frac{n}{4})$ in place of $A$ in
\eqref{eq:HKUB-localized-existence} and obtain
\begin{equation*}
\functionspace{P}_{t}(x,A\cap B(x,{\textstyle\frac{n}{4}}))
\leq\int_{A\cap B(x,n/4)}H_{t}(x,y)\,d\mu(y)
\end{equation*}
with $c'=c_{1/4}$, which yields \eqref{eq:HK-existence-global}
by using monotone convergence to let $n\to\infty$.
Thus we may assume $R<\infty$.
Let $y_{0}\in M$ and $A\in\sigmafield{B}(B(y_{0},\frac{R}{4}))$.
We claim that \eqref{eq:HK-existence-global} holds for such $A$.
Indeed, setting $U:=B(y_{0},\frac{R}{2})$,
we have \eqref{eq:HK-existence-global} with $H_{t}(x,y)$ replaced by
\begin{equation}\label{eq:HKUB-localized-global-Htilde}
\widetilde{H}_{t}(x,y):=
\begin{cases}
c_{1/4}F_{t}(x,y)\exp\bigl(-\Phi(\frac{1}{20}\gamma d(x,y),t)\bigr) &\textrm{if }t<\Psi(R)\textrm{ and }x\in U,\\
c''_{1/4}(\inf\nolimits_{U\times U}F_{t})\exp(-\Phi(\frac{1}{20}\gamma R,t))
&\textrm{if }t<\Psi(R)\textrm{ and }x\not\in U,\\
c_{1/4}(\inf_{U\times U}F_{\Psi(R)}) &\textrm{if }t\geq\Psi(R)
\end{cases}
\end{equation}
since Theorem \ref{thm:HKUB-localized} and
Proposition \ref{prop:HKUB-localized-difference}
with $\varepsilon=\frac{1}{4}$ are applicable by \eqref{eq:U-ball}
and $\functionspace{P}^{U}_{t}(x,A)=0$ if $x\not\in U$.
Moreover, if $t\leq\Psi(R)$ then for any $y\in M$ and any $z,w\in U$,
\begin{align}\label{eq:HKUB-localized-global-Rfinite1}
\inf\nolimits_{U\times U}F_{t}\leq F_{t}(z,w)
&\leq c_{F}\Bigl(\frac{t\vee\Psi(d(x,z))\vee\Psi(d(y,w))}{t}\Bigr)^{\alpha_{F}}F_{t}(x,y)\notag\\
&\leq c_{F}\Bigl(\frac{\Psi(\delta^{-1}R)}{t}\Bigr)^{\alpha_{F}}F_{t}(x,y)\notag\\
&\leq c_{F}c_{\Psi}^{\alpha_{F}}\delta^{-\beta_{2}\alpha_{F}}
\Bigl(\frac{\Psi(R)}{t}\Bigr)^{\alpha_{F}}F_{t}(x,y)
\end{align}
by \textup{(DB)$_{\Psi}$}, $\diam M\leq\delta^{-1}R$ and
\eqref{eq:upper-bound-function-Psi} (even if $x\in N$), hence
\begin{equation}\label{eq:HKUB-localized-global-Rfinite2}
c_{1/4}(\inf\nolimits_{U\times U}F_{\Psi(R)})
\leq c_{1/4}c_{F}c_{\Psi}^{\alpha_{F}}\delta^{-\beta_{2}\alpha_{F}}
(\inf\nolimits_{M\times M}F_{\Psi(R)}),
\end{equation}
and we also easily see from \eqref{eq:HKUB-localized-global-Rfinite1},
\eqref{eq:upper-bound-function-Phi-constant},
\eqref{eq:upper-bound-function-Phi-ULE}, \eqref{eq:upper-bound-function-Psi}
and $R\geq\delta\diam M$ that
\begin{align}
&c''_{1/4}(\inf\nolimits_{U\times U}F_{t})\exp(-\Phi({\textstyle\frac{1}{20}}\gamma R,t))\notag\\
&\leq c''_{1/4}c_{F}c_{\Psi}^{\alpha_{F}}\delta^{-\beta_{2}\alpha_{F}}
\Bigl(\frac{\Psi(R)}{t}\Bigr)^{\alpha_{F}}F_{t}(x,y)\exp(-2\Phi({\textstyle\frac{1}{40}}\gamma R,t))\notag\\
&\leq c''\delta^{-\beta_{2}\alpha_{F}}F_{t}(x,y)\exp\bigl(-\Phi(\gamma'_{\delta}d(x,y),t)\bigr)
\label{eq:HKUB-localized-global-Rfinite3}
\end{align}
for any $y\in M$ for some $c''\in(0,\infty)$ explicit in
$c_{\Psi},\beta_{1},\beta_{2},c_{F},\alpha_{F},c,\gamma$.
Therefore putting $c':=c''\vee(c_{1/4}c_{F}c_{\Psi}^{\alpha_{F}})$,
we have $\widetilde{H}_{t}(x,y)\leq H_{t}(x,y)$ for any $y\in M$
by \eqref{eq:HKUB-localized-global-Htilde},
\eqref{eq:HKUB-localized-global-Rfinite2} and
\eqref{eq:HKUB-localized-global-Rfinite3}, and thus the inequality
\eqref{eq:HK-existence-global} follows from that with
$\widetilde{H}_{t}(x,y)$ in place of $H_{t}(x,y)$.
Now let $\{y_{k}\}_{k\in\mathbb{N}}$ be a countable dense
subset of $M$ and set $B_{1}:=B(y_{1},\frac{R}{4})$ and
$B_{k}:=B(y_{k},\frac{R}{4})\setminus\bigcup_{j=1}^{k-1}B(y_{j},\frac{R}{4})$
for $k\in\mathbb{N}\setminus\{1\}$, so that
$\{B_{k}\}_{k\in\mathbb{N}}\subset\sigmafield{B}(M)$ and
$M=\bigcup_{k\in\mathbb{N}}B_{k}$, where the union is disjoint.
Then for any $A\in\sigmafield{B}(M)$, for each $k\in\mathbb{N}$ we have
\eqref{eq:HK-existence-global} with $A\cap B_{k}$ in place of $A$ by the
previous paragraph and $A\cap B_{k}\in\sigmafield{B}(B(y_{k},\frac{R}{4}))$,
from which \eqref{eq:HK-existence-global} follows
in exactly the same way as \eqref{eq:HKUB-localized-off-diag-local-end},
completing the proof of Theorem \ref{thm:HKUB-localized-global}.
\end{proof}
\section{Exit probability estimates for diffusions}\label{sec:exit-probability}
As already mentioned at the beginning of Section \ref{sec:HKUB},
the purpose of this section is to provide reasonable sufficient conditions
for the exit probability estimate \textup{(P)$_{\Psi}^{U,R}$} of
Theorem \ref{thm:HKUB-localized}. Recall that since Section \ref{sec:HKUB}
we have fixed $\Psi,c_{\Psi},\beta_{1},\beta_{2}$ and
$\Phi=\Phi_{\Psi}$ as in Lemma \ref{lem:upper-bound-function-Psi}.
\emph{In this section, we fix an arbitrary properly exceptional set
$N\in\sigmafield{B}(M)$ for $X$ satisfying
\begin{equation}\label{eq:X-continuous-no-killing-inside}
\textrm{both}\quad\eqref{eq:X-continuous}\quad\textrm{and}\quad
\mathbb{P}_{x}[\zeta<\infty,\,X_{\zeta-}\in M]=0
\end{equation}
for any $x\in M\setminus N$,}
where $X_{\zeta-}(\omega):=X_{\zeta(\omega)-}(\omega)$
($X_{0-}:=X_{0}$, $X_{\infty-}:=\cemetery=X_{\infty}$), so that
$X_{\zeta-}:\Omega\to M_{\cemetery}$ is
$\sigmafield{F}_{\infty}/\sigmafield{B}(M_{\cemetery})$-measurable by the
left-continuity of $[0,\infty)\ni t\mapsto X_{t-}(\omega)\in M_{\cemetery}$.
By \cite[Theorem 4.5.3]{FOT}, such $N$ exists if and only if
$(\functionspace{E},\functionspace{F})$ is \emph{strongly local},
i.e., $\functionspace{E}(u,v)=0$ for any $u,v\in\functionspace{F}$
with $\supp_{\mu}[u],\supp_{\mu}[v]$ compact and $u=c$ $\mu$-a.e.\ on a
neighborhood of $\supp_{\mu}[v]$ for some $c\in\mathbb{R}$.
Below we will also consider the situation where the set $N$
fixed above satisfies
\begin{equation}\label{eq:X-continuous-conservative}
\textrm{both}\quad\eqref{eq:X-continuous}\quad\textrm{and}\quad
\mathbb{P}_{x}[\zeta<\infty]=0
\end{equation}
for any $x\in M\setminus N$,
more strongly than \eqref{eq:X-continuous-no-killing-inside}.
By \cite[Theorem 4.5.1 and Exercise 4.5.1]{FOT},
such a properly exceptional set $N\in\sigmafield{B}(M)$ for $X$ exists if and
only if $(\functionspace{E},\functionspace{F})$ is local and \emph{conservative},
i.e., $T_{t}\ind{}=\ind{}$ $\mu$-a.e.\ for any (or equivalently, for some)
$t\in(0,\infty)$, where $\ind{}:=\ind{M}$; recall
(see, e.g., \cite[(1.1.9) and (1.1.11)]{CF}) that
for a Markovian bounded linear operator $T:L^{2}(M,\mu)\to L^{2}(M,\mu)$,
$T|_{L^{2}(M,\mu)\cap L^{\infty}(M,\mu)}$ can be uniquely extended to
a linear operator $T:L^{\infty}(M,\mu)\to L^{\infty}(M,\mu)$ such that
$\lim_{n\to\infty}Tu_{n}=Tu$ $\mu$-a.e.\ for any $u\in L^{\infty}(M,\mu)$
and any $\{u_{n}\}_{n\in\mathbb{N}}\subset L^{\infty}(M,\mu)$ with
$u_{n}\leq u_{n+1}$ $\mu$-a.e.\ for any $n\in\mathbb{N}$ and
$\lim_{n\to\infty}u_{n}=u$ $\mu$-a.e.
\begin{remark}\label{rmk:exit-probability}
In fact, \emph{Theorems \textup{\ref{thm:exit-probability}} and
\textup{\ref{thm:exit-time-probability}} below apply, without any changes in
the proofs, to any locally compact separable metric space $(M,d)$,
any Hunt process $X$ on $(M,\sigmafield{B}(M))$ and any
$N\in\sigmafield{B}(M)$ satisfying \eqref{eq:properly-exceptional} and
\eqref{eq:X-continuous-no-killing-inside} for any $x\in M\setminus N$}.
\end{remark}
The main result of this section is the following theorem, which is a localized version
of an unpublished result \cite[Theorem 9.1]{Gri:HKfractal} by the first named author.
We refer the reader to \cite[Subsection 5.4]{GriHu:Upper} for an alternative
analytic approach. We set $e^{-\infty}:=0$.
\begin{theorem}\label{thm:exit-probability}
Let $U$ be a non-empty open subset of $M$ and let $R\in(0,\infty]$.
Then among the following seven conditions, the latter six \textup{(2)--(7)}
are equivalent and imply \textup{(1)}:
\begin{itemize}[label=\textup{(1)},align=left,leftmargin=*]
\item[\textup{(1)}]
There exist $\varepsilon\in(0,\frac{1}{2})$ and $\delta\in(0,\infty)$ such that
for any $(x,r)\in(U\setminus N)\times(0,R)$ with $B(x,r)\subset U$
and $\overline{B(x,r)}$ compact and for any $t\in(0,\delta\Psi(r)]$,
\begin{equation}\label{eq:exit-probability1}
\mathbb{P}_{x}[X_{t}\in M_{\cemetery}\setminus B(x,r)]\leq\varepsilon.
\end{equation}
\item[\textup{(2)}]
There exist $\varepsilon\in(0,1)$ and $\delta\in(0,\infty)$ such that
for any $(x,r)\in(U\setminus N)\times(0,R)$ with $B(x,r)\subset U$
and $\overline{B(x,r)}$ compact,
\begin{equation}\label{eq:exit-probability2}
\mathbb{P}_{x}[\tau_{B(x,r)}\leq\delta\Psi(r)]\leq\varepsilon.
\end{equation}
\item[\textup{(3)}]
There exists $\varepsilon\in(0,\infty)$ such that
for any $(x,r)\in(U\setminus N)\times(0,R)$ with $B(x,r)\subset U$
and $\overline{B(x,r)}$ compact,
\begin{equation}\label{eq:exit-probability3}
\mathbb{E}_{x}[\tau_{B(x,r)}\wedge\Psi(r)]\geq\varepsilon\Psi(r).
\end{equation}
\item[\textup{(4)}]
There exist $\varepsilon\in(0,1)$ and $\delta\in(0,\infty)$ such that
for any $(x,r)\in(U\setminus N)\times(0,R)$ with $B(x,r)\subset U$
and $\overline{B(x,r)}$ compact,
\begin{equation}\label{eq:exit-probability4}
\mathbb{E}_{x}\Bigl[\exp\Bigl(-\frac{\tau_{B(x,r)}}{\delta\Psi(r)}\Bigr)\Bigr]\leq\varepsilon.
\end{equation}
\item[\textup{(5)}]
There exist $c,\gamma\in(0,\infty)$ such that for any
$(x,r)\in(U\setminus N)\times(0,R)$ with $B(x,r)\subset U$
and $\overline{B(x,r)}$ compact and for any $\lambda\in(0,\infty)$,
\begin{equation}\label{eq:exit-probability5}
\mathbb{E}_{x}[e^{-\lambda\tau_{B(x,r)}}]\leq c\exp\Bigl(-\frac{\gamma r}{\Psi^{-1}(\lambda^{-1})}\Bigr).
\end{equation}
\item[\textup{(6)}]
There exist $c,\gamma\in(0,\infty)$ such that for any
$(x,r)\in(U\setminus N)\times(0,R)$ with $B(x,r)\subset U$
and $\overline{B(x,r)}$ compact and for any $t\in(0,\infty)$,
\begin{equation}\label{eq:exit-probability6}
\mathbb{P}_{x}[\tau_{B(x,r)}\leq t]\leq c\exp(-\Phi(\gamma r,t)).
\end{equation}
\item[\textup{(7)}]
There exist $c,\gamma\in(0,\infty)$ such that for any
$(x,r)\in(U\setminus N)\times(0,R)$ with $B(x,r)\subset U$
and $\overline{B(x,r)}$ compact and for any $t\in(0,\infty)$,
\begin{equation}\label{eq:exit-probability7}
\mathbb{P}_{x}[\tau_{B(x,r)}\leq t]\leq
c\exp\biggl(-\gamma\Bigl(\frac{\Psi(r)}{t}\Bigr)^{\frac{1}{\beta_{2}-1}}\biggr).
\end{equation}
\end{itemize}
Moreover, if $N$ satisfies \eqref{eq:X-continuous-conservative} for any
$x\in M\setminus N$, then with ``and $\overline{B(x,r)}$ compact" all removed,
still the conditions \textup{(2)--(7)} are equivalent, imply \textup{(1)}
and are implied by the following condition \textup{(1)$'$}:
\begin{itemize}[label=\textup{(1)$'$},align=left,leftmargin=*]
\item[\textup{(1)$'$}]
There exist $\varepsilon\in(0,\frac{1}{2})$ and $\delta\in(0,\infty)$ such that
for any $(x,r)\in(\overline{U}\setminus N)\times(0,\frac{R}{2})$
and any $t\in(0,\delta\Psi(r)]$,
\begin{equation}\label{eq:exit-probability1'}
\mathbb{P}_{x}[X_{t}\in M\setminus B(x,r)]\leq\varepsilon.
\end{equation}
\end{itemize}
\end{theorem}
\begin{proof}
We follow \cite[Proof of Theorem 9.1]{Gri:HKfractal}; for the implications
\textup{(4)}$\Rightarrow$\textup{(5)}$\Rightarrow$\textup{(6)}$\Rightarrow$\textup{(7)}
see also \cite[Proofs of Lemma 3.14, Theorem 3.15 and Corollary 3.20]{GT}.
We treat the two cases simultaneously, one with ``and $\overline{B(x,r)}$ compact"
kept and without \eqref{eq:X-continuous-conservative} and the other with
``and $\overline{B(x,r)}$ compact" removed and \eqref{eq:X-continuous-conservative} assumed.
Let $(x,r)\in(U\setminus N)\times(0,R)$ satisfy $B(x,r)\subset U$
and set $\tau:=\tau_{B(x,r)}$. We assume in the former case that
$\overline{B(x,r)}$ is compact, while not in the latter case.
It easily follows either from \eqref{eq:X-continuous-no-killing-inside} and the
compactness of $\overline{B(x,r)}$ or from \eqref{eq:X-continuous-conservative},
together with $\mathbb{P}_{x}[X_{0}=x]=1$ and \eqref{eq:properly-exceptional}, that
\begin{equation}\label{eq:exit-ball-boundary}
\mathbb{P}_{x}\bigl[\tau_{B(x,\rho)}<\infty,\,X_{\tau_{B(x,\rho)}}\not\in(\partial B(x,\rho))\setminus N\bigr]=0
\qquad\textrm{for any }\rho\in(0,r].
\end{equation}
\textup{(2)}$\Rightarrow$\textup{(3)}:
Since
$\mathbb{P}_{x}[\tau>\delta\Psi(r)]=1-\mathbb{P}_{x}[\tau\leq\delta\Psi(r)]\geq 1-\varepsilon$
by \eqref{eq:exit-probability2},
\begin{equation*}
\mathbb{E}_{x}[\tau\wedge\Psi(r)]
\geq\mathbb{E}_{x}[(\tau\wedge\Psi(r))\ind{\{\tau>\delta\Psi(r)\}}]
\geq(\delta\wedge 1)\Psi(r)\mathbb{P}_{x}[\tau>\delta\Psi(r)]
\geq(1-\varepsilon)(\delta\wedge 1)\Psi(r).
\end{equation*}
\textup{(3)}$\Rightarrow$\textup{(4)}:
For $\lambda,t\in(0,\infty)$, by considering the case of $\tau\geq t$ and
that of $\tau\leq t$ separately we easily see that
$e^{-\lambda\tau}\leq 1-\lambda e^{-\lambda t}(\tau\wedge t)$,
and therefore for \emph{any} $\delta\in(0,\infty)$, setting $\lambda:=(\delta\Psi(r))^{-1}$
and $t:=\Psi(r)$, taking $\mathbb{E}_{x}[(\cdot)]$ and applying \eqref{eq:exit-probability3}, we obtain
\begin{equation*}
\mathbb{E}_{x}\Bigl[\exp\Bigl(-\frac{\tau}{\delta\Psi(r)}\Bigr)\Bigr]
\leq 1-\frac{1}{\delta\Psi(r)}e^{-1/\delta}\mathbb{E}_{x}[\tau\wedge\Psi(r)]
\leq 1-\frac{\varepsilon}{\delta}e^{-1/\delta}.
\end{equation*}
\textup{(4)}$\Rightarrow$\textup{(5)}:
Let $\lambda\in[(\delta\Psi(r))^{-1},\infty)$, set
$n:=\max\{k\in\mathbb{N}\mid\lambda\delta\Psi(r/k)\geq 1\}$ and $\rho:=r/n$.
Also set $B_{k}:=B(x,k\rho)$ and $\tau_{k}:=\tau_{B_{k}}$ for $k\in\{1,\dots,n\}$.
We claim that
\begin{equation}\label{eq:exit-probability-iteration}
\mathbb{E}_{x}[e^{-\lambda\tau_{k+1}}]\leq\varepsilon\mathbb{E}_{x}[e^{-\lambda\tau_{k}}]
\qquad\textrm{for any }k\in\{1,\dots,n\}\textrm{ with }k<n.
\end{equation}
To see \eqref{eq:exit-probability-iteration}, let $k\in\{1,\dots,n\}$ satisfy $k<n$
and let $y\in(\partial B_{k})\setminus N$. Then obviously
$(y,\rho)\in(U\setminus N)\times(0,R)$,
$B(y,\rho)\subset B_{k+1}\subset B(x,r)\subset U$,
hence $\tau_{B(y,\rho)}\leq\tau_{k+1}$, and
$\overline{B(y,\rho)}$ is compact if $\overline{B(x,r)}$ is.
Thus \textup{(4)} applies to $(y,\rho)$, so that from
\eqref{eq:exit-probability4} we obtain
\begin{equation}\label{eq:exit-probability4-apply}
\mathbb{E}_{y}[e^{-\lambda\tau_{k+1}}]
\leq\mathbb{E}_{y}[e^{-\lambda\tau_{B(y,\rho)}}]
\leq\mathbb{E}_{y}\Bigl[\exp\Bigl(-\frac{\tau_{B(y,\rho)}}{\delta\Psi(\rho)}\Bigr)\Bigr]
\leq\varepsilon,
\end{equation}
noting that $\lambda\geq(\delta\Psi(\rho))^{-1}$ by the choice of $n$.
Now since $\tau_{k+1}=\tau_{k}+\tau_{k+1}\circ\theta_{\tau_{k}}$,
it follows by the strong Markov property \cite[Theorem A.1.21]{CF} of $X$,
\eqref{eq:exit-ball-boundary} and \eqref{eq:exit-probability4-apply} that
\begin{align*}
\mathbb{E}_{x}[e^{-\lambda\tau_{k+1}}]
=\mathbb{E}_{x}[e^{-\lambda\tau_{k}}(e^{-\lambda\tau_{k+1}}\circ\theta_{\tau_{k}})]
&=\mathbb{E}_{x}\bigl[e^{-\lambda\tau_{k}}\mathbb{E}_{x}[e^{-\lambda\tau_{k+1}}\circ\theta_{\tau_{k}}\mid\sigmafield{F}_{\tau_{k}}]\bigr]\\
&=\mathbb{E}_{x}\bigl[e^{-\lambda\tau_{k}}\mathbb{E}_{X_{\tau_{k}}}[e^{-\lambda\tau_{k+1}}]\bigr]\\
&=\mathbb{E}_{x}\bigl[\ind{\{X_{\tau_{k}}\in(\partial B_{k})\setminus N\}}e^{-\lambda\tau_{k}}\mathbb{E}_{X_{\tau_{k}}}[e^{-\lambda\tau_{k+1}}]\bigr]\\
&\leq\varepsilon\mathbb{E}_{x}[e^{-\lambda\tau_{k}}].
\end{align*}
\begin{figure}[t]\centering
\includegraphics[height=180pt]{fpic5.eps}
\caption{Proof of \textup{(4)}$\Rightarrow$\textup{(5)}: the exit times $\tau_{k}$, $\tau_{k+1}$ and $\tau_{B(y,\rho)}$}
\label{fig:exit-probability-iteration}
\end{figure}
Thus we have proved \eqref{eq:exit-probability-iteration}, which together with
$\tau=\tau_{n}$ and \eqref{eq:exit-probability4} for $(x,\rho)$ yields
\begin{align*}
\mathbb{E}_{x}[e^{-\lambda\tau}]\leq\varepsilon^{n-1}\mathbb{E}_{x}[e^{-\lambda\tau_{1}}]
\leq\varepsilon^{n-1}\mathbb{E}_{x}\Bigl[\exp\Bigl(-\frac{\tau_{1}}{\delta\Psi(\rho)}\Bigr)\Bigr]
\leq\varepsilon^{n}
<\varepsilon^{-1}\exp\Bigl(-\frac{\gamma r}{\Psi^{-1}(\lambda^{-1})}\Bigr),
\end{align*}
where $\gamma:=\eta\log(\varepsilon^{-1})$ with
$\eta:=(\delta/c_{\Psi})^{1/\beta_{1}}\wedge 1$ and the last inequality follows since
$1>\lambda\delta\Psi(\frac{r}{n+1})\geq\lambda\Psi(\frac{\eta r}{n+1})$
by the choice of $n$ and \eqref{eq:upper-bound-function-Psi} and hence
$n+1>\eta r/\Psi^{-1}(\lambda^{-1})$.
\eqref{eq:exit-probability5} therefore holds for $\lambda\in[(\delta\Psi(r))^{-1},\infty)$.
On the other hand, for $\lambda\in(0,(\delta\Psi(r))^{-1})$, since
$\lambda^{-1}>\delta\Psi(r)\geq\Psi(\eta r)$ by \eqref{eq:upper-bound-function-Psi}
and hence $\Psi^{-1}(\lambda^{-1})>\eta r$, we have
$\mathbb{E}_{x}[e^{-\lambda\tau}]\leq 1
<e^{\gamma/\eta}\exp\bigl(-\gamma r/\Psi^{-1}(\lambda^{-1})\bigr)$,
completing the proof of \textup{(4)}$\Rightarrow$\textup{(5)}.
\textup{(5)}$\Rightarrow$\textup{(6)}:
For any $t,\lambda\in(0,\infty)$, we see from \eqref{eq:exit-probability5} that
\begin{equation*}
\mathbb{P}_{x}[\tau\leq t]=\mathbb{P}_{x}[e^{-\lambda t}\leq e^{-\lambda\tau}]
\leq e^{\lambda t}\mathbb{E}_{x}[e^{-\lambda\tau}]
\leq c\exp\Bigl(\lambda t-\frac{\gamma r}{\Psi^{-1}(\lambda^{-1})}\Bigr),
\end{equation*}
and taking the infimum of the right-hand side in $\lambda\in(0,\infty)$ shows
\eqref{eq:exit-probability6} in view of \eqref{eq:upper-bound-function-Phi}.
\textup{(6)}$\Rightarrow$\textup{(7)}:
Since $\Psi(\gamma r)\geq c_{\Psi}^{-1}(\gamma^{\beta_{2}}\wedge 1)\Psi(r)$ with
$\gamma\in(0,\infty)$ as in \textup{(6)} by \eqref{eq:upper-bound-function-Psi},
if $\Psi(\gamma r)\geq t$ then \eqref{eq:exit-probability7} is immediate from
\eqref{eq:exit-probability6} and the lower inequality in
\eqref{eq:upper-bound-function-Phi-ULE}, whereas if $\Psi(\gamma r)<t$ then we have
$\mathbb{P}_{x}[\tau\leq t]\leq 1\leq c'\exp\bigl(-\gamma'(\Psi(r)/t)^{\frac{1}{\beta_{2}-1}}\bigr)$
for some $c',\gamma'\in(0,\infty)$ explicit in $c_{\Psi},\beta_{1},\beta_{2},\gamma$.
\textup{(7)}$\Rightarrow$\textup{(2)},\textup{(1)}:
For \emph{any} $\varepsilon\in(0,c\wedge\frac{1}{2})$, setting
$\delta:=\bigl(\gamma/\log(c/\varepsilon)\bigr)^{\beta_{2}-1}\in(0,\infty)$,
for any $t\in(0,\delta\Psi(r)]$ we see from
$\{X_{t}\in M_{\cemetery}\setminus B(x,r)\}\subset\{\tau\leq\delta\Psi(r)\}$
and \eqref{eq:exit-probability7} that
\begin{align*}
\mathbb{P}_{x}[X_{t}\in M_{\cemetery}\setminus B(x,r)]
\leq\mathbb{P}_{x}[\tau\leq\delta\Psi(r)]
\leq c\exp\bigl(-\gamma\delta^{-\frac{1}{\beta_{2}-1}}\bigr)
=\varepsilon.
\end{align*}
\emph{\textup{(1)$'$}$\Rightarrow$\textup{(2)} under \eqref{eq:X-continuous-conservative}}:
Note that \eqref{eq:exit-probability1'} is valid also for $t=0$ since $\mathbb{P}_{y}[X_{0}=y]=1$
for $y\in M$. Let $t:=c_{\Psi}^{-1}2^{-\beta_{2}}\delta\Psi(r)$, so that
$t\in(0,\delta\Psi(\frac{r}{2})]$ by \eqref{eq:upper-bound-function-Psi}.
We first show that
\begin{equation}\label{eq:exit-probability1'-proof1}
\mathbb{P}_{x}[\tau\leq t,\,X_{t}\in B(x,{\textstyle\frac{r}{2}})]\leq\varepsilon.
\end{equation}
Indeed, if $y\in(\partial B(x,r))\setminus N$, then
clearly $B(x,\frac{r}{2})\subset M\setminus B(y,\frac{r}{2})$,
$(y,\frac{r}{2})\in(\overline{U}\setminus N)\times(0,\frac{R}{2})$
by $B(x,r)\subset U$ and hence \textup{(1)$'$} applies to $(y,\frac{r}{2})$,
so that \eqref{eq:exit-probability1'} yields
\begin{equation}\label{eq:exit-probability1'-proof2}
\mathbb{P}_{y}[X_{s}\in B(x,{\textstyle\frac{r}{2}})]
\leq\mathbb{P}_{y}[X_{s}\in M\setminus B(y,{\textstyle\frac{r}{2}})]
\leq\varepsilon
\end{equation}
for any $y\in(\partial B(x,r))\setminus N$
and any $s\in[0,t]\subset[0,\delta\Psi(\frac{r}{2})]$.
Then since $\{\tau\leq t\}\in\sigmafield{F}_{\tau\wedge t}$ by \cite[Lemma 1.2.16]{KS},
it follows from Proposition \ref{prop:strong-Markov}, \eqref{eq:exit-ball-boundary}
and \eqref{eq:exit-probability1'-proof2} that
\begin{align*}
\mathbb{P}_{x}[\tau\leq t,\,X_{t}\in B(x,{\textstyle\frac{r}{2}})]
&=\mathbb{E}_{x}[\ind{\{\tau\leq t\}}\ind{B(x,r/2)}(X_{t})]\\
&=\mathbb{E}_{x}\bigl[\ind{\{\tau\leq t\}}\mathbb{E}_{x}[\ind{B(x,r/2)}(X_{t})\mid\sigmafield{F}_{\tau\wedge t}]\bigr]\\
&=\int_{\{\tau\leq t\}}\mathbb{E}_{X_{\tau\wedge t}(\omega)}[\ind{B(x,r/2)}(X_{t-\tau(\omega)\wedge t})]\,d\mathbb{P}_{x}(\omega)\\
&=\int_{\{\tau\leq t,\,X_{\tau}\in(\partial B(x,r))\setminus N\}}
\mathbb{P}_{X_{\tau}(\omega)}[X_{t-\tau(\omega)}\in B(x,{\textstyle\frac{r}{2}})]\,d\mathbb{P}_{x}(\omega)\\
&\leq\varepsilon\mathbb{P}_{x}[\tau\leq t,\,X_{\tau}\in(\partial B(x,r))\setminus N]\leq\varepsilon.
\end{align*}
Now noting that $\mathbb{P}_{x}[X_{t}=\cemetery]=0$ by \eqref{eq:X-continuous-conservative},
from \eqref{eq:exit-probability1'} for $(x,\frac{r}{2})$ and
\eqref{eq:exit-probability1'-proof1} we obtain
\begin{align*}
\mathbb{P}_{x}[\tau\leq t]
&=\mathbb{P}_{x}[\tau\leq t,\,X_{t}=\cemetery]
+\mathbb{P}_{x}[\tau\leq t,\,X_{t}\in M\setminus B(x,{\textstyle\frac{r}{2}})]
+\mathbb{P}_{x}[\tau\leq t,\,X_{t}\in B(x,{\textstyle\frac{r}{2}})]\\
&\leq\mathbb{P}_{x}[X_{t}=\cemetery]+\mathbb{P}_{x}[X_{t}\in M\setminus B(x,{\textstyle\frac{r}{2}})]+\varepsilon\\
&\leq 2\varepsilon<1,
\end{align*}
which, in view of $t=c_{\Psi}^{-1}2^{-\beta_{2}}\delta\Psi(r)$,
completes the proof of \textup{(1)$'$}$\Rightarrow$\textup{(2)}
under \eqref{eq:X-continuous-conservative}.
\end{proof}
At the last of this paper, as an application of Theorem \ref{thm:exit-probability}
we state and prove a localized version of the well-known fact that the comparability
of the mean exit time $\mathbb{E}_{x}[\tau_{B(x,r)}]$ to $\Psi(r)$ implies the exit
probability estimate \eqref{eq:exit-probability6}. This fact was first observed by
M.\ T.\ Barlow as treated in \cite[Proof of Theorem 3.11]{Bar}, and
the proof below is also based on an idea of his in \cite{Bar}.
\begin{theorem}\label{thm:exit-time-probability}
Let $U$ be a non-empty open subset of $M$, let $R\in(0,\infty]$,
and assume that there exists $c_{\mathrm{E}}\in(0,\infty)$ such that
for any $(x,r)\in(U\setminus N)\times(0,2R)$,
\begin{equation}\label{eq:exit-time-upper}
\mathbb{E}_{x}[\tau_{B(x,r)}]\leq c_{\mathrm{E}}\Psi(r),
\end{equation}
and for any $(x,r)\in(U\setminus N)\times(0,R)$ with
$B(x,r)\subset U$ and $\overline{B(x,r)}$ compact,
\begin{equation}\label{eq:exit-time-lower}
\mathbb{E}_{x}[\tau_{B(x,r)}]\geq c_{\mathrm{E}}^{-1}\Psi(r).
\end{equation}
Then Theorem \textup{\ref{thm:exit-probability}-(6)} holds.
Moreover, additionally if $N$ satisfies \eqref{eq:X-continuous-conservative}
for any $x\in M\setminus N$ and if \eqref{eq:exit-time-lower} holds for
any $(x,r)\in(U\setminus N)\times(0,R)$ with $B(x,r)\subset U$, then
Theorem \textup{\ref{thm:exit-probability}-(6)} with ``and $\overline{B(x,r)}$ compact" removed holds.
\end{theorem}
\begin{proof}
As in the proof of Theorem \ref{thm:exit-probability}, we treat the two cases
simultaneously, one with ``and $\overline{B(x,r)}$ compact" kept and without
\eqref{eq:X-continuous-conservative} and the other with ``and $\overline{B(x,r)}$ compact"
removed and \eqref{eq:X-continuous-conservative} assumed.
Let $(x,r)\in(U\setminus N)\times(0,R)$ satisfy $B(x,r)\subset U$.
We assume in the former case that
$\overline{B(x,r)}$ is compact, while not in the latter case. We claim that
\begin{equation}\label{eq:exit-time-probability-claim}
\mathbb{P}_{x}[\tau_{B(x,r)}\leq{\textstyle\frac{1}{2}}c_{\mathrm{E}}^{-1}\Psi(r)]
\leq 1-(c_{\mathrm{E}}^{2}c_{\Psi}2^{\beta_{2}+1})^{-1},
\end{equation}
which together with the implication \textup{(2)}$\Rightarrow$\textup{(6)}
of Theorem \ref{thm:exit-probability} shows the assertions.
To see \eqref{eq:exit-time-probability-claim} we follow \cite[Proof of Lemma 3.16]{Bar}.
Set $\tau:=\tau_{B(x,r)}$ and $t:=\frac{1}{2}c_{\mathrm{E}}^{-1}\Psi(r)$.
By using \eqref{eq:exit-time-lower}, the obvious relation
$\tau\leq t+(\tau-t)\ind{\{t<\tau\}}=t+(\tau\circ\theta_{t})\ind{\{t<\tau\}}$
and the Markov property \cite[Theorem A.1.21]{CF} of $X$ at time $t$, we have
\begin{equation}\label{eq:exit-time-probability-proof}
2t=c_{\mathrm{E}}^{-1}\Psi(r)\leq\mathbb{E}_{x}[\tau]
\leq t+\mathbb{E}_{x}[(\tau\circ\theta_{t})\ind{\{t<\tau\}}]
=t+\mathbb{E}_{x}\bigl[\ind{\{t<\tau\}}\mathbb{E}_{X_{t}}[\tau]\bigr].
\end{equation}
Note that $X_{t}\in B(x,r)$ on $\{t<\tau\}$, that
$\mathbb{P}_{x}[X_{t}\in N]=0$ by \eqref{eq:properly-exceptional},
and that for any $y\in B(x,r)\setminus N$,
$\tau\leq\tau_{B(y,2r)}$ by $B(x,r)\subset B(y,2r)$ and hence
$\mathbb{E}_{y}[\tau]\leq\mathbb{E}_{y}[\tau_{B(y,2r)}]
\leq c_{\mathrm{E}}\Psi(2r)\leq c_{\mathrm{E}}c_{\Psi}2^{\beta_{2}}\Psi(r)$
by \eqref{eq:exit-time-upper} and \eqref{eq:upper-bound-function-Psi}.
It follows from \eqref{eq:exit-time-probability-proof} and these facts that
\begin{equation*}
t\leq\mathbb{E}_{x}\bigl[\ind{\{t<\tau\}}\mathbb{E}_{X_{t}}[\tau]\bigr]
=\mathbb{E}_{x}\bigl[\ind{\{t<\tau,\,X_{t}\in B(x,r)\setminus N\}}\mathbb{E}_{X_{t}}[\tau]\bigr]
\leq c_{\mathrm{E}}c_{\Psi}2^{\beta_{2}}\Psi(r)(1-\mathbb{P}_{x}[\tau\leq t]),
\end{equation*}
which immediately implies \eqref{eq:exit-time-probability-claim}.
\end{proof}
\begin{acknowledgements}
Essential parts of this paper were written while the second named author was
visiting the University of Bonn from March to September 2014. He would like to
thank Kobe University for its financial and administrative supports for his visit.
He also would like to express his deepest gratitude toward the members of the
stochastics research groups of the University of Bonn for their heartfelt hospitality.
The authors would like to thank Dr.\ Sebastian Andres and Dr.\ Kouji Yano
for their valuable comments on earlier versions of the manuscript.
\end{acknowledgements}
|
\section{Introduction}\label{sec:int}
\lettrine[lines=2]{C}{ellular} networks, flooded by an enormous demand for mobile data services, are expected to undergo a fundamental transformation. In order to significantly increase the data capacity, coverage performance, and energy efficiency, next generation mobile networks ($5$G) \cite{546843} are expected to move from the traditional single, high-powered base-station (BS) to the deployments of multiple overlaying access points of diverse sizes, i.e., microcell, picocell, femtocell, etc., using different radio access technologies. To efficiently manage the resulting high levels of interference, connecting BSs through efficient backhauling becomes a critical component in the network planning. The big increase in small cell deployment, however, necessitates a considerable amount of backhaul communications in order to share the data streams between all BSs across the network \cite{6736746}. Giving that the links are capacity limited, upgrading the backhaul and increasing its ability to support the tremendous amount of data is a necessity \cite{13050958}. Therefore, choosing the suitable technology(ies) and design of the backhaul network is of great interest, especially that its deployment cost is a dominant cost driver for many operators, e.g., approximately between $30\%$ and $50\%$ of the total operating costs for $4$G systems \cite{591842}. With the deployment of multiple small cells expected in $5$G, the implementation costs are believed to get even higher \cite{13050958}. This paper proposes a cost efficient backhaul solution for next generation backhaul systems using techniques from graph theory.
\subsection{Backhaul Technologies}
Traditional technologies for the backhaul network design include copper, microwave radio links (RF), and optical fibers (OF). The leased T$1$/E$1$ copper lines is the most widely used backhaul technology with approximately $90\%$ of the total backhaul deployment in the US \cite{5473878}. With a provided data rate of $1.544$ Mbit/s for T$1$ and $2.048$ Mbit/s for E$1$ \cite{6777766}, copper lines provide satisfying data rates for voice traffic for $2$G networks. However, to achieve the data rate demand of $3$G traffic and beyond, multiple parallel connections are required which results in a price growing linearly with the provided capacity. For high data rates, copper lines become expensive and hence not a suitable solution for the backhaul upgrade of next generation systems, i.e., $5$G.
Microwave radio is the second most used technology for the backhaul network design as it represents $6\%$ of the total used transport media \cite{5473878} in the US. The RF technology represents a reasonable alternative to copper, especially in locations in which the deployment of wired connections is challenging. However, such solution requires an initial investment in the licensed part of the spectrum \cite{5185525}. Moreover, low frequency radio (radio waves below $6$ GHz) is limited in terms of data rates due to interference problems and high frequency radio (microwave \& millimetre wave (mmwave) from $6$ to $300$ GHz) are limited in the transmission coverage area.
Optical fiber (OF) backhaul links provide the highest rates over long distances, e.g., $155.52$ Mbit/s for STM-$1$, $622$ Mbit/s for STM-$4$, $2.4$ Gbit/s STM-$16$, and $9.9$ Gbit/s for STM-$64$ \cite{5473878}. However, as they are expensive to be deployed and require a considerable initial investment \cite{6226966}, they represent $4\%$ of the total backhaul deployment in the US \cite{5473878}. They, further, suffer from the drawbacks of wired connections, i.e., deployment is not always feasible, which restricts their utilization in particular applications.
Recently, the free-space optics technology (FSO) emerges as a substitute \cite{5771213} for next generation backhauls. An FSO link refers to a laser beam between a pair of photo-detector transceivers using the free-space as medium of transportation. Giving that its wavelength is in the micrometer range, which is an unlicensed band, FSO links are not only free to use but also immune to electromagnetic interference generated by the RF links. The high bandwidth and interference immunity features make an FSO link up to $25$ fold more efficient than an RF link in terms of capacity \cite{1495057}. FSO particularly represents a cost-efficient solution compared to OF.
In contrast with the omni-reliability of the OF and RF links, FSO links are sensitive to weather conditions, e.g., fog, snow, and rain \cite{65841515}. Therefore, reliability becomes an important factor to address for the design of FSO-based backhaul networks. To cope with such varying reliability, and combine the advantages of RF (reliability) and FSO (capacity), the hybrid RF/FSO backhaul becomes an attractive cost-effective and reliable solution. Hybrid RF/FSO transmits, when possible, simultaneously on both the RF and FSO links. In harsh weather conditions that affect the FSO link, the data is sent solely on the RF link \cite{6876609}. Moreover, hybrid RF/FSO transceivers can be quickly deployed over several kilometers \cite{16546169} and can also be easily combined with OF links \cite{5545666}. For all previously mentioned benefits, hybrid RF/FSO is a graceful complementary option for upgrading the existing backhaul network \cite{254515158}, as further shown in our paper.
\subsection{Related Work}
In the past few years, hybrid RF/FSO attracted a significant amount of research. Most of the current work \cite{6844864,168714992012} focus on the determination of the factors affecting the FSO link performance, e.g, weather conditions, scintillation, ect., and finding solutions to improve the quality, e.g., use of multiple lasers and multiple apertures, etc. However, fundamental problems of hybrid RF/FSO architecture optimization for the backhaul network topology design are only at their beginning.
The authors in \cite{1495122} design an efficient and scalable algorithm to optimize a given physical layer objective for $2$ and $3$ optical transceivers per node with a minimum number of links. The authors in \cite{Smadi:09,4609027} propose upgrading an RF network by optimally deploying the minimal number of FSO transceivers so as to achieve a given throughput. While Kashyap {\it{et al}} \cite{1495057} design a routing algorithm for hybrid RF/FSO networks that backs up the traffic to the FSO routes when the RF links could not carry it, Rak {\it{et al}} \cite{6876609} introduce a linear integer programming model to determine routing in hybrid RF/FSO network in which the FSO link availability is varying with the weather conditions. In \cite{4357553}, the authors consider a hybrid RF/FSO system in which the RF and FSO links operate at different data rates. They derive an upper bound for the capacity per node that is asymptotically achievable for random networks.
Numerous mixed integer programming model are proposed to formulate the problem of backhaul network design using hybrid RF/FSO technology. In particular, Son {\it{et al}} \cite{5462107} present an algebraic connectivity-based formulation for the design of the backbone of wireless mesh networks with FSO links and solve it using a greedy approach that iteratively inserts nodes to maximize the algebraic connectivity. The authors in \cite{4746591} propose to maximize the network throughput by installing as many FSO links as possible under the constraint that the number of FSO links in a node is bounded. In \cite{6134071}, Ahdi {\it{et al}} introduce a mixed integer programming model to find the optimal placement of FSO links in order to upgrade an existing RF backhaul network. Similarly, reference \cite{4609027} propose to improve an existing RF backhaul network with FSO links using the minimum number of FSO links to guarantee a target network throughput when RF links are non-available due to interference.
This paper suggests upgrading a pre-deployed OF backhaul network using hybrid RF/FSO links, and, hence, is related to the concept developed in \cite{6777766,6844494,2514014}. The authors in \cite{6777766} consider the upgrade of a pre-deployed OF backhaul network using FSO links and mirrors for nodes not in line-of-sight of each other. For two link-disjoint paths networks, they formulate the problem as a mixed integer programming and extend the study in \cite{6844494} to $K$ link-disjoint paths. In \cite{2514014}, the same group of authors analyze the impact of the parameter $K$ on the design. This paper extends the concept by suggesting using hybrid RF/FSO links, considering a minimum reliability constraint, and proposing a close-to-optimal explicit solution using graph theory techniques.
\subsection{Contributions}
This paper examines the problem of upgrading a pre-deployed OF backhaul network. It considers the problem of minimizing the cost of the backhaul network by choosing either OF or hybrid RF/FSO backhaul links between the base-stations (BS) so as to satisfy data rate, connectivity, and reliability constraints. Unlike our recent works which focus on connecting BSs through $K$ link-disjoint paths in order to cope with possible link failures (see \cite{Dahrouj_backnet_magazine} for the business case of the RF backhaul, and \cite{Hybrid_Douik_ICC16} for the technical details of the proposed resilient solution), a primary concern of the current paper is to guarantee network connectivity achieved by connecting each pair of nodes in the network, possibly via multiple hops. While the deployment cost of hybrid RF/FSO links depends mainly on the expense of the hybrid RF/FSO transceivers, the implementation cost of OF links depends mostly on the distance between the two end nodes. On the other hand, OF links always satisfy the data rate and reliability constraint. The performance of hybrid RF/FSO, however, degrades with the distance and the number of installed links. The paper solves the problem using graph theory techniques by introducing the corresponding planning graph. The paper's main contribution is to provide a close-to-optimal explicit solution to the problem. The paper shows that under a specified realistic assumption about the cost of OF and hybrid RF/FSO links, the problem can be reformulated as a maximum weight clique problem, that can be globally solved using efficient algorithms \cite{9874286,16513519,13265492,6607889}.
The rest of this paper is organized as follows: \sref{sec:sys} presents the considered system model and the problem formulation. In \sref{sec:prob}, the planning problem is approximated by a more tractable one. \sref{sec:prop} illustrates the proposed solution. Before concluding in \sref{sec:conc}, simulation results are presented in \sref{sec:sim}.
\section{System Model and Problem Formulation}\label{sec:sys}
\subsection{System Model and Parameters}
\begin{figure}[t]
\centering
\includegraphics[width=.6\linewidth]{figs/network.eps}\\
\caption{Network containing $5$ base-stations connected with OF and hybrid RF/FSO links.}\label{fig:network}
\end{figure}
This paper considers a backhaul network connecting a set $\mathcal{B}=\{b_1,\ \cdots,\ b_M\}$ of $M$ base-stations with pre-deployed OF connections. All nodes (interchangeably denoting base-stations) are assumed to have a line-of-sight connections. Therefore, each node can be connected to any other node with either an OF or a hybrid RF/FSO connection, as in \fref{fig:network} which shows a network containing $5$ base-stations.
Let $P_{ij}, 1 \leq i,j \leq M$ be a binary variable indicating the existence, i.e., $P_{ij}=1$, of a pre-deployed OF link between base-stations $b_i$ and $b_j$ and $P_{ij}=0$ otherwise. For simplicity of notations, $P_{ij}$ may be also written as $P(b_i,b_j)$.
Let $\pi^{(O)}_{ij}$ and $\pi^{(h)}_{ij}$ the cost an OF and a hybrid RF/FSO link between nodes $b_i$ and $b_j$, respectively. Obviously, the functions $\pi^{(O)}_{ij}$ and $\pi^{(h)}_{ij}$ are positive and symmetric, e.g., $\pi^{(O)}_{ij}=\pi^{(O)}_{ji}$. Finally, as hybrid RF/FSO is a cost effective solution, the paper assumes that $\pi^{(h)}_{ij} \leq \pi^{(O)}_{ij} , \forall \ i,j \in \mathcal{B}$.
Let $D^{(O)}_{ij}$ and $D^{(h)}_{ij}$ be the provided data rates of an OF and a hybrid RF/FSO links between nodes $b_i$ and $b_j$, respectively. Let $D_t$ be the targeted data rate. Since the OF provides high data rates, without loss of generality, this work assumes that $D^{(O)}_{ij} \geq D_t, \forall \ i \neq j \in \mathcal{B}$.
Similarly, let $R^{(O)}_{ij}$ and $R^{(h)}_{ij}$ be the reliability of an OF and a hybrid RF/FSO links between nodes $b_i$ and $b_j$, respectively. Let $\alpha$ be the targeted reliability. As OF links are always reliable, this paper assumes that $R^{(O)}_{ij} \geq \alpha, \forall \ i \neq j \in \mathcal{B}$.
\subsection{Problem Formulation}
Let $X_{ij}, 1 \leq i,j \leq M$ be a binary variable indicating if base-stations $b_i$ and $b_j$ are connected with an OF connection. Similarly, let $Y_{ij}, 1 \leq i,j \leq M$ indicate if they are connected with a hybrid RF/FSO link. To simplify the problem formulation and constraints, this paper takes as a convention that $X_{ii}=Y_{ii}=0, \forall \ i \in \mathcal{B}$ in all the rest of the equations.
This paper considers the problems of minimizing the network deployment cost under the following constraints:
\begin{enumerate}
\item C1: Some nodes have pre-deployed OF links.
\item C2: Connections between nodes can be either OF or hybrid RF/FSO.
\item C3: Each node has a data rate that exceeds the targeted data rate.
\item C4: The reliability of each node exceeds the targeted reliability.
\item C5: Each node can communicate with any other node through single or multiple hop links.
\end{enumerate}
Let the Laplacian matrix $L$ be defined as $\mathbf{L} = \mathbf{D}-\mathbf{C}$, where $\mathbf{D}=\text{diag}(d_1,\ \cdots,\ d_M)$ is a diagonal matrix with $d_i = \sum_{j=1}^M X_{ij}+Y_{ij}$ and $c_{ij} = X_{ij}+Y_{ij}$. The diagonalization of the Laplacian matrix is given by $\mathbf{L}= \mathbf{Q}\mathbf{\Lambda} \mathbf{Q}^{-1}$, where $\mathbf{\Lambda}= \text{diag}(\lambda_1,\ \cdots,\ \lambda_M)$ with $\lambda_1 \leq \lambda_2 \leq \cdots \leq \lambda_M$. The connectivity condition C5 of the matrix can be written using the algebraic formulation proposed in \cite{25181258} as $\lambda_2 > 0$.
The following lemma introduces the cost-efficient backhaul design problem formulation:
\begin{lemma}
The problem of minimizing the cost of the backhaul network planning can be formulated as:
\begin{subequations}
\label{Original_optimization_problem}
\begin{align}
\min & \ \cfrac{1}{2} \sum_{i=1}^M \sum_{j=1}^M X_{ij}\pi^{(O)}_{ij} + Y_{ij}\pi^{(h)}_{ij} \\
{\rm s.t.\ } & X_{ij} = X_{ji} \label{eq:1} \\
&Y_{ij} = Y_{ji} \label{eq:2} \\
&X_{ij}P_{ij} = P_{ij} \label{eq:np} \\
&X_{ij}Y_{ij} = 0 \label{eq:3} \\
&\sum_{j=1}^M X_{ij}D_t + Y_{ij}D^{(h)}_{ij} \geq D_t\label{eq:5} \\
&1 - \prod_{j=1}^M (1 - X_{ij} \alpha)(1 - Y_{ij} R^{(h)}_{ij}) \geq \alpha \label{eq:nr} \\
&\lambda_2 > 0 \label{eq:6} \\
&X_{ij},Y_{ij} \in \{0,1\},\ 1 \leq i,j\leq M \label{eq:4},
\end{align}
\end{subequations}
where the optimization is over both binary variables $X_{ij}$ and $Y_{ij}$.
\label{l1}
\end{lemma}
\begin{proof}
To formulate the problem, the objective function and the system constraints C1 to C5 are expressed in terms of the variables $X_{ij}$ and $Y_{ij}$. Combining all the expressions yields the optimization problem \eref{Original_optimization_problem}. The complete proof can be found in \appref{ap1}.
\end{proof}
The optimization problem \eref{Original_optimization_problem} is equivalent to a weighted Steiner tree problem which is NP-hard with a complexity of order $2^{M^2}$. The optimal solution to such problem is referred to as the optimal planning. The rest of this paper proposes an efficient method to solve the problem \eref{Original_optimization_problem}, under the assumption that the hybrid RF/FSO connection between two nodes that are not neighbors is always more expensive than the cost of the OF connections between each node and its closest neighbour. The rationale for such assumption is that, for short distances, OF links are much cheaper than hybrid RF/FSO ones. Under this assumption, the next section shows that the solution for backhaul network design becomes mathematically tractable with a complexity of order $2^{M}$.
\section{Problem Approximation} \label{sec:prob}
As highlighted above, the original optimization problem \eref{Original_optimization_problem} is an NP-hard problem. The difficulty in solving the problem lies particularly in the structure of constraint \eref{eq:6} and in simultaneously optimizing \eref{Original_optimization_problem} over both binary variables $X_{ij}$ and $Y_{ij}$. This section presents an efficient heuristic to solve the problem under the assumption that the hybrid RF/FSO connection between two nodes that are not neighbors is always more expensive than the OF connections between each node and its closest neighbour. The assumption is motivated by the fact that, for short distances, OF links are much cheaper than hybrid RF/FSO ones. The heuristic is based on first finding the solution to the problem when only OF links can be used. Afterwards, it solves an approximate of the backhaul network planning problem via relating problem \eref{Original_optimization_problem} to solution reached by the planning problem when only OF links are allowed.
\subsection{Optimal Planning Using Optical Fiber Only}
The following lemma introduces the reduced problem when only OF links are allowed.
\begin{lemma}
The problem of backhaul design with minimum cost, when only OF links are allowed, is the following:
\begin{subequations}
\label{eq:11}
\begin{align}
\min& \ \cfrac{1}{2} \sum_{i=1}^M \sum_{j=1}^M X_{ij} \pi^{(O)}_{ij} \\
{\rm s.t.\ }& X_{ij} = X_{ji} \label{eq:9} \\
&X_{ij}P_{ij} = P_{ij} \label{eq:np2} \\
&\lambda_2 > 0 \label{eq:8} \\
&X_{ij} \in \{0,1\},\ 1 \leq i,j\leq M \label{eq:10}.
\end{align}
\end{subequations}
\label{l2}
\end{lemma}
\begin{proof}
To show this lemma, it is sufficient to show that, for backhaul network using only OF links, i.e., $Y_{ij}=0,\forall \ i,j$, constraints \eref{eq:2}, \eref{eq:3}, \eref{eq:5}, and \eref{eq:nr} become redundant. This can easily be done by noting that for a connected graph, each node is connected to, at least, another node. Since only a single OF connection is sufficient to ensure throughput and reliability, the constraints become redundant.
\end{proof}
To solve the problem mentioned above, the paper proposes to cluster BSs, according to the minimal price. First, a cluster $\mathcal{Z}$ containing all BSs is formed. For each connected nodes with pre-deployed OF links, the base-stations are merged into a single group (belonging to the big cluster $\mathcal{Z}$) and the corresponding $X_{ij}$ set to $1$. Afterwards, find the two minimum-price clusters and merge them into a single group. The cost between two clusters is defined as the minimum cost between all BS in each cluster. When two clusters are merged, the two minimum-price BSs in each cluster are connected through an OF link. The process is repeated until only one group remains in the system. In other words, the process terminates when all nodes are merged into a single cluster, i.e., $|\mathcal{Z}|=1$. The steps of the algorithm are summarized in \algref{alg1}. The following theorem characterizes the solution produced by \algref{alg1} with respect to the problem defined in \lref{l2}:
\begin{algorithm}[t]
\begin{algorithmic}
\REQUIRE $\mathcal{B}$, $P_{ij}$, and $\pi^{(O)}$.
\STATE Initialize $X_{ij}=P_{ij},\ 1 \leq i,j\leq M $.
\STATE Initialize $\mathcal{Z} = \varnothing$.
\FORALL {$b \in \mathcal{B}$}
\STATE Initialize $t=0$.
\FORALL {$Z \in \mathcal{Z}$}
\FORALL {$b^{\prime} \in Z$}
\IF{$P(b,b^{\prime}) = 1$}
\STATE $\mathcal{Z} = \mathcal{Z} \setminus \{Z\}$.
\STATE $Z = \{Z,b\}$.
\STATE $\mathcal{Z} = \{\mathcal{Z},\{Z\}\}$.
\STATE $t=1$.
\ENDIF
\ENDFOR
\ENDFOR
\IF{$t=0$}
\STATE $\mathcal{Z} = \{\mathcal{Z},\{b\}\}$.
\ENDIF
\ENDFOR
\WHILE {$|\mathcal{Z}| > 1$}
\STATE $(Z_i,Z_j) = \arg \min\limits_{\substack{Z,Z^{\prime} \in \mathcal{Z} \\ Z \neq Z^{\prime}}} \left[ \min\limits_{\substack{b \in Z \\ b^{\prime} \in Z^{\prime}}} \pi^{(O)}(b , b^{\prime}) \right] $.
\STATE $(b_i,b_j) = \arg \min\limits_{\substack{b \in Z_i \\ b^{\prime} \in Z_j}} \pi^{(O)}(b , b^{\prime})$.
\STATE $X_{ij} = X_{ji} = 1$.
\STATE $\mathcal{Z} = \mathcal{Z} \setminus \{Z_i\}$.
\STATE $\mathcal{Z} = \mathcal{Z} \setminus \{Z_j\}$.
\STATE $\mathcal{Z} = \{\mathcal{Z},\{Z_i,Z_j\}\}$.
\ENDWHILE
\end{algorithmic}
\caption{Optimal planning using only OF links}
\label{alg1}
\end{algorithm}
\begin{theorem}
The solution reached by \algref{alg1} is the optimal solution to the problem proposed in \lref{l2}. Such solution is referred to, in this paper, as the optimal OF only planning.
\label{th1}
\end{theorem}
\begin{proof}
To prove this theorem, we employ a two-stage proof. The first part of the proof shows that the solution reached by \algref{alg1} is the optimal solution to the problem proposed in \lref{l2} for a network without pre-deployed OF links. The second part of the proof extends the result for networks with pre-deployed OF connections. Therefore, we first show that \algref{alg1} produces a feasible solution to the problem. Afterwards, we show that any graph that can be reduced, using an algorithm similar to\algref{alg1}, to a single cluster includes the graph designed by \algref{alg1}. Finally, we show that any solution that cannot be reduced to a single group is not optimal. The complete proof can be found in \appref{ap3}.
\end{proof}
\subsection{Problem Approximation}
This subsection approximates the backhaul network planning problem \eref{Original_optimization_problem} under the assumption that a hybrid RF/FSO connection between two nodes that are not neighbours is more expensive than the OF links between each node and its closest neighbour. We first define $b_{i^*}$ as the closest node to base-station $b_i$ as follows:
\begin{align}
b_{i^*} = \arg \min_{\substack{b \in \mathcal{B} \\ b \neq b_i } } \pi^{(O)}(b_i,b).
\end{align}
The set of neighbours $\mathcal{N}_i$ of base-station $b_i$ is defined as the set of base-stations that are closest to base-station $b_i$, and that satisfy the connectivity condition. Mathematically, the condition can be written as:
\begin{align}
\mathcal{N}_i = \left\{b \in \mathcal{B} \setminus b_i \text{ such that } \pi^{(h)}(b_i,b) \leq \max_{b_j \in \mathcal{B}}\overline{X}_{ij}\pi^{(h)}_{ij} \right\},
\label{eq:nei}
\end{align}
where $\overline{X}_{ij},\ 1 \leq i \neq j \leq M$ is the optimal solution found in solving the OF only planning problem \eref{eq:11}.
\begin{remark}
The results presented in this paper do not depend on the definition of the set of neighbours $\overline{\mathcal{N}}_i$ of node $b_i$ as long as $\mathcal{N}_i \subset \overline{\mathcal{N}}_i$. Intuitively, as the set $\overline{\mathcal{N}}_i$ gets bigger and bigger, the approximation of the solution is more tight. For $\overline{\mathcal{N}}_i = \mathcal{B} \setminus \{b_i\}$, the proposed algorithm reduces to an exhaustive search.
\end{remark}
The assumption that two nodes that are far away from each others (i.e., not neighbours) connected with hybrid RF/FSO link generate a cost greater that the expense of the same nodes connected with OF links with their closest neighbours can be written $\forall\ (b_i,b_j) \notin \mathcal{N}_j \times \mathcal{N}_i$ as follows:
\begin{align}
\pi^{(O)}_{ii^*}+\pi^{(O)}_{jj^*} \leq \pi^{(h)}_{ij}.
\label{eq:12}
\end{align}
Let $\mathcal{R}_i = \{ b_j \in \mathcal{B} \setminus \{b_i\} \ | \ R^{(h)}_{ij} \geq \alpha\}$ be the set of nodes that satisfy, by their own, the reliability condition for node $b_i$. Based on the above assumption and definitions, the following lemma approximates the optimization problem \eref{Original_optimization_problem} under the assumption \eref{eq:12}.
\begin{lemma}
The problem of backhaul network cost minimization design using OF and hybrid RF/FSO connections can be approximated by the following problem:
\begin{subequations}
\label{Approximate_optimization_problem}
\begin{align}
\min& \ \cfrac{1}{2} \sum_{i=1}^M \sum_{j=1}^M X_{ij}\pi^{(O)}_{ij}+Y_{ij}\pi^{(h)}_{ij} \label{eq:25} \\
{\rm s.t.\ } & X_{ij} = X_{ji} \label{eq:20} \\
&Y_{ij} = Y_{ji} \label{eq:21}\\
&X_{ij}P_{ij} = P_{ij} \label{eq:np3} \\
&X_{ij}Y_{ij} = 0 \label{eq:22} \\
&\sum_{j=1}^M X_{ij}D_t + Y_{ij}D^{(h)}_{ij} \geq D_t \label{eq:23}\\
&\sum_{j=1}^M X_{ij}\tilde{\alpha} + \sum_{j \in \mathcal{R}_i}Y_{ij}\tilde{\alpha} + \sum_{j \in \overline{\mathcal{R}}_i}Y_{ij} R^{(h)}_{ij} \geq \tilde{\alpha} \label{eq:nr2} \\
&(X_{ij} + Y_{ij})\overline{X}_{ij} = \overline{X}_{ij} \label{eq:14} \\
&X_{ij},Y_{ij} \in \{0,1\},\ 1 \leq i,j\leq M,
\end{align}
\end{subequations}
where $\overline{\mathcal{R}}_i= \mathcal{B} \setminus \mathcal{R}_i$ is the complementary set of $\mathcal{R}_i$ and $\tilde{\alpha} = -\log(1-\alpha)$.
\label{l4}
\end{lemma}
\begin{proof}
To show that the original problem, it is sufficient to show that any solution to \eref{Approximate_optimization_problem} is a feasible solution to \eref{Original_optimization_problem}. Therefore, we show that constraint \eref{eq:nr2} is equivalent to constraint \eref{eq:nr} and that constraint \eref{eq:14} is included in constraint \eref{eq:6}. The complete proof can be found in \appref{ap4}.
\end{proof}
\section{Proposed Solution}\label{sec:prop}
This section proposes the solution for the approximate problem \eref{Approximate_optimization_problem}. The solution is based first on constructing the network planning graph, and then on formulating the problem \eref{Approximate_optimization_problem} as a graph theory problem that can be optimally solved with moderate complexity.
\subsection{Planning Graph}
In this section, we introduce the undirected \emph{planning} graph $\mathcal{G}(\mathcal{V},\mathcal{E})$, where $\mathcal{V}$ is the set of vertices and $\mathcal{E}$ the set of edges. Before stating the vertices construction and the edge connection, we first define the cluster $\mathcal{C}_i$ for each node $b_i, 1 \leq i \leq M$ as follows:
\begin{align}
\mathcal{C}_i = \{((X_{ij_1},Y_{ij_1}),\ \cdots,\ &(X_{ij_{|\mathcal{N}_i|}},Y_{ij_{|\mathcal{N}_i|}})), \text{ such that } \nonumber \\
\bigcup_{j \in \mathcal{N}_i}b_{j} &= \mathcal{N}_i \nonumber \\
X_{ij}P_{ij} &= P_{ij}, \forall \ j \in \mathcal{N}_i \nonumber \\
X_{ij}Y_{ij} &= 0, \forall \ j \in \mathcal{N}_i \nonumber \\
\sum_{j \in \mathcal{N}_i} X_{ij}D_t + Y_{ij}D^{(h)}_{ij} &\geq D_t \\
\sum_{j \in \mathcal{N}_i} X_{ij}\tilde{\alpha} + \sum_{j \in \mathcal{N}_i \cap \mathcal{R}_i}Y_{ij}\tilde{\alpha} &+ \sum_{j \in \mathcal{N}_i \cap \overline{\mathcal{R}}_i}Y_{ij} R^{(h)}_{ij} \geq \tilde{\alpha} \nonumber \\
(X_{ij}+Y_{ij})\overline{X}_{ij} &= \overline{X}_{ij}, \forall \ j \in \mathcal{N}_i \}.\nonumber
\end{align}
Define the weight of each element $\gamma_i \in \mathcal{C}_i$, ($\gamma_i = \{(X_{ij_1},Y_{ij_1}),\ \cdots,\ (X_{ij_{|\mathcal{N}_i|}},Y_{ij_{|\mathcal{N}_i|}})\}$), as follows:
\begin{align}
w(\gamma_i) = -\cfrac{1}{2}\sum_{j \in \mathcal{N}_i} X_{ij}\pi^{(O)}_{ij}+Y_{ij}\pi^{(h)}_{ij}.
\label{eq:15}
\end{align}
For each cluster $\gamma_i \in \mathcal{C}_i$, a vertex $v_{ij}, 1 \leq j \leq |\mathcal{C}_i|$ is generated. Two distinct vertices $v_{ij}$ and $v_{kl}$ are connected with an edge in $\mathcal{E}$ if the two following conditions are satisfied:
\begin{enumerate}
\item C1: $i \neq k$: The vertices represents different nodes in the network.
\item C2: $(X_{ik},Y_{ik}) = (X_{ki},Y_{ki})$ if $(b_i,b_k) \in (\mathcal{N}_k,\mathcal{N}_i)$: The vertices are non conflicting.
\end{enumerate}
\subsection{Proposed Algorithm}
The following theorem characterizes the solution of the approximated backhaul network planning problem \eref{Approximate_optimization_problem}.
\begin{theorem}
Let $(X^*_{ij},Y^*_{ij}), 1 \leq i,j \leq M$ be the optimal solution to the planning problem \eref{Approximate_optimization_problem} then we have $X^*_{i,j}+Y^*_{i,j}=1$ only if $(i,j) \in \mathcal{N}_j \times \mathcal{N}_i$.
\label{th2}
\end{theorem}
\begin{proof}
To show this theorem, the scenarios that can result in a violation of the desired property are identified. Using the cost optimality, connectivity constraint, and the assumption about the relative value, all such scenarios are shown to be sub-optimal to \eref{Approximate_optimization_problem}. Therefore, the optimal solution satisfies the property. The complete proof can be found in \appref{ap5}.
\end{proof}
The following theorem links the solution of problem \eref{Approximate_optimization_problem} to the planning graph.
\begin{theorem}
The solution of the approximation of the backhaul network problem \eref{Approximate_optimization_problem} using hybrid RF/FSO can be formulated as a maximum weight clique, among the cliques of size $M$ in the planning graph, in which the weight of each vertex $v_{ij}$ is the weight of the corresponding cluster $\gamma_i$ defined in \eref{eq:15}.
\label{th3}
\end{theorem}
\begin{proof}
To prove this theorem, we first show that there is a one to one mapping between the set of feasible solutions of the problem \eref{Approximate_optimization_problem} and the set of cliques of degree $M$ in the planning graph $\mathcal{G}(\mathcal{V},\mathcal{E})$. To conclude the proof, we show that the weight of the clique is equivalent to the merit function of the optimization problem \eref{Approximate_optimization_problem}. The complete proof can be found in \appref{ap6}.
\end{proof}
\subsection{Complexity Analysis}
This subsection characterizes the complexity of solving the original $0-1$ integer program proposed in \eref{Original_optimization_problem} and its relaxed version proposed in \eref{Approximate_optimization_problem}.
In order to characterize the complexity of \eref{Original_optimization_problem}, we first compute and reduce the number of variables. Initially, the number of $X_{ij}$ and $Y_{ij}$ is $M^2$ each. However, as $X_{ii}$ and $Y_{ii}$ can take arbitrary values, the number reduces to $M^2-M$ variables each. Furthermore, from constraint \eref{eq:1} and \eref{eq:2} which translate the symmetry of the problem, only half of the variables are independent. Hence, the number of free variables is $\frac{M^2-M}{2}$. Finally, the pre-deployed OF links, i.e., constraint \eref{eq:np}, limits the number of variables. In fact, the constraint $P_{ij}=1$ translates to $X_{ij}=1$ and $Y_{ij}=0$. Let $\overline{P}=\sum_{i=1}^M \sum_{j=1}^M P_{ij}$ be the number of pre-deployed links. It can clearly be seen that the number of variables of $X_{ij}$ and $Y_{ij}$ is $\frac{M^2-M-\overline{P}}{2}$. Let $\mathcal{P}=\{(i,j) \ | \ i<j, P_{ij}=0\}$ be the set of nodes that do not have a pre-deployed links. From constraint \eref{eq:3}, the variables $X_{ij}$ and $Y_{ij}$ are not independent. As only $3$ combinations are possible, they can be represented by $\frac{M^2-M-\overline{P}}{2}$ ternary variable $Z_{ij}, (i,j) \in \mathcal{P}$ defined as follows:
\begin{align}
Z_{ij} =
\begin{cases}
0 \hspace{0.5cm} &\text{if } X_{ij}=0 \text{ and } Y_{ij}=0 \\
1 \hspace{0.5cm} &\text{if } X_{ij}=1 \text{ and } Y_{ij}=0 \\
2 \hspace{0.5cm} &\text{if } X_{ij}=0 \text{ and } Y_{ij}=1.
\end{cases}
\end{align}
Therefore, the optimization problem \eref{Original_optimization_problem} can be written as follows:
\begin{align*}
\min & \ \sum_{(i,j) \in \mathcal{P}} \delta(Z_{ij}-1)\pi^{(O)}_{ij} + \delta(Z_{ij}-2)\pi^{(h)}_{ij} \nonumber \\
{\rm s.t.\ } & \sum_{j\in \mathcal{P}} \delta(Z_{ij}-1)D_t + \delta(Z_{ij}-2)D^{(h)}_{ij} \geq D_t \nonumber \\
&1 - \prod_{j\in \mathcal{P}} (1 - \delta(Z_{ij}-1)\alpha)(1 - \delta(Z_{ij}-2) R^{(h)}_{ij}) \geq \alpha \nonumber \\
\end{align*}
\begin{align}
&\lambda_2 > 0 \nonumber \\
&Z_{ij} \in \{0,1,2\},\ (i,j) \in \mathcal{P},
\label{eq:zz}
\end{align}
where $\delta(x)$ is the discrete Delta function equal to $1$ if and only if its argument is equal to $0$. The formulation in \eref{eq:zz} allows to derive the complexity of the optimal solution as proportional to $\eta^{\frac{M^2-M-\overline{P}}{2}}$, where $1 < \eta \leq 3$ is the complexity constant that depends on the algorithm used to solving the weighted Steiner tree problem and where the extreme case $\eta=3$ reduces to the exhaustive search.
In order to characterize the complexity of the relaxed optimization problem \eref{Approximate_optimization_problem}, we first identify the number of vertices in the planning graph. As the number of vertices depends on the relative position of the nodes, the data rates, and the reliability functions, this subsection characterizes the worst case complexity. Let $n$ be the maximum number of neighbours of the nodes. From constraint \eref{eq:3}, in each cluster, the number of vertices is bounded by $3^n$. Therefore, the number of vertices of the whole planning graph is bounded by $3^nM$. The formulation of the problem as a graph theory problem allows to derive the complexity of the approximate solution as proportional to $\xi^{3^nM}$, where $1 < \xi \leq 2$ is the complexity constant that depends on the algorithm used to solving the maximum weight clique problem and where the extreme case $\xi=2$ reduces to the exhaustive search.
For the minimal number of neighbours as defined in \eref{eq:nei}, the number of neighbours does not depend on the number of nodes $M$ in the network. Therefore, for a large enough number of base-station, the realistic assumption about the cost of the OF and hybrid RF/FSO links the backhaul network design problem become more mathematically tractable with a reduction in complexity from an order of $2^{M^2}$ to a complexity of order $2^{M}$.
\section{Simulation Results}\label{sec:sim}
This section shows the performance of the proposed solution to the backhaul network planning problem using hybrid RF/FSO technology. The base-stations are randomly placed on a $5$ Km long square. The pre-deployed OF links are randomly placed between the base-stations. The ratio of pre-deployed links by the total number of possible connections is of $1/5$. These simulation assume that the price, the provided data rate, and the reliability are sole function of the distance separating the two end nodes. The cost of a multi-mode OM$3$ $(50/125)$ OF link is, according to various constructors (Asahi Kasei, Chromis, Eska, OFS HCS) between $3$ \$ and $30$ \$ per meter depending on the number of cores. In these simulations, the cost of the optical transceivers, being negligible, is ignored, and a medium price $\pi^{(O)}=13.5$ \$ per meter is adopted. The cost of a hybrid RF/FSO link is taken to be independent of the distance. Given the prices offered by the different constructors (fSONA, LightPointe, and RedLine), two types of costs are considered: $\pi^{(h)}=10$ k\$ and $20$ k\$. The price $\pi^{(h)}=40$ k\$ is proposed as a cut-off price for which hybrid RF/FSO do not represent any advantage.
The data rate of a hybrid RF/FSO links is taken to be $D_t$ over a distance $d_D$ after which it decays exponentially. In other words, $D^{(h)}(x)= D_t$ if $x<d_D$ and $D^{(h)}(x) = D_t\exp^{-(x-d_D)}$ otherwise. The reliability follows a similar model. The maximal distance satisfying the reliability condition is $d_R$. For illustration purposes, the length $d_D$ and $d_R$ are assumed to be $3$ and $2$ Km unless indicated otherwise.
The numbers of base-stations, the price of the hybrid RF/FSO transceivers, and the distances $d_D$ and $d_R$ vary in the simulations so as to study the methods performance for various scenarios. The planning simulated in this section are the optimal planning (solution of \eref{Original_optimization_problem}), the OF only planning (solution of \eref{eq:11}), and our proposed heuristic hybrid RF/FSO-OF planning (solution of \eref{Approximate_optimization_problem}).
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{figs/MC.eps}\\
\caption{Mean cost of the network versus the number of base-stations $M$. The solid lines are obtained for a price of a hybrid RF/FSO links of $\pi^{(h)}=10$ k\$, the dashed for a cost $\pi^{(h)}=20$ k\$ and the dotted for the cutoff price $\pi^{(h)}=40$ k\$ at which the different planning coincide.}\label{fig:MC}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{figs/CC.eps}\\
\caption{Mean cost of the network versus the cost of hybrid RF/FSO links $\pi^{(h)}$. The solid lines are obtained for some base-station $M=6$, and the dashed for $M=7$.}\label{fig:CC}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{figs/CR.eps}\\
\caption{Average percentage of OF connections versus the cost of hybrid RF/FSO links $\pi^{(h)}$ for a network containing $7$ nodes.}\label{fig:CR}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{figs/DDR.eps}\\
\caption{Average percentage of OF connections versus the distance $d_D$ satisfying the data rate. The solid lines are obtained for a perfect reliability distance $d_R=2Km$ and the dashed one for $d_R=4Km$.}\label{fig:DDR}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{figs/DRC.eps}\\
\caption{Mean cost of the network versus the maximum distance $d_R$ satisfying the target reliability $\alpha$ for a system containing $7$ nodes.}\label{fig:DRC}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{figs/DRR.eps}\\
\caption{Average percentage of OF connections versus the maximum distance $d_R$ satisfying the target reliability $\alpha$ for a network containing $7$ nodes.}\label{fig:DRR}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{figs/AC.eps}\\
\caption{Mean cost of the network versus the targeted reliability $\alpha$. The solid lines are obtained for a price of a hybrid RF/FSO links of $\pi^{(h)}=10$ k\$, the dashed for a cost $\pi^{(h)}=20$ k\$ and the dotted for the cutoff price $\pi^{(h)}=40$ k\$ at which the different planning coincide.}\label{fig:AC}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{figs/AR.eps}\\
\caption{Average percentage of OF connections versus the targeted reliability $\alpha$. The solid lines are obtained for a price of a hybrid RF/FSO links of $\pi^{(h)}=10$ k\$, the dashed for a cost $\pi^{(h)}=20$ k\$ and the dotted for the cutoff price $\pi^{(h)}=40$ k\$ at which the different planning coincide.}\label{fig:AR}
\end{figure}
\fref{fig:MC} plots the cost of the network versus the number of BSs, for various costs of the hybrid RF/FSO transceivers. We clearly see that the degradation of our proposed solution against the optimal solution becomes less severe when first the number of base-stations increases, and secondly when the hybrid RF/FSO transceivers become more expensive. The increase in performance in the first case can be explained by the fact that the connectivity opportunities of nodes increase as the number of base-stations increases, due to the rise in the neighbours sets $\mathcal{N}_i$. The gain in performance when the price of the hybrid RF/FSO transceivers increases can be explained by the fact that our assumption \eref{eq:12} becomes more valid as the price of the hybrid RF/FSO transceivers increases.
\fref{fig:CC} and \fref{fig:CR} illustrate the cost of the network and the ratio of the OF link, respectively, against the expense of the hybrid RF/FSO transceivers. As shown in \fref{fig:MC}, the performance of our proposed algorithm is more close to the one of the optimal planning as the cost of the hybrid RF/FSO transceivers increases. From \fref{fig:CR}, we clearly see that if the hybrid RF/FSO transceivers are expensive enough, both the optimal and our proposed solution contain only OF links. In fact, for expensive hybrid RF/FSO transceivers, the OF links offers a noticeable rate advantage that explain their use. It is worth mentioning that for a cost $\pi^{(h)} \geq 30$ k\$ in \fref{fig:CR}, even though the link's nature utilized in the optimal solution and our proposed solution are different, the total cost of the network is almost the same (\fref{fig:CC} for $M=7$).
To quantify the performance of the proposed algorithms with respect to the distance $d_D$, \fref{fig:DDR} plots the percentage of the OF link used against the distance $d_D$ for different reliability $d_R$ and a price $\pi_{(h)}=20$ k\$. \fref{fig:DDR} depicts that for a small $d_D$, our proposed solution uses more OF links than the optimal solution. Whereas for a $d_D \geq 2$ the ratio is almost the same. This can be explained by our choice of neighbours $\mathcal{N}_i$. The connectivity opportunities of our proposed solution are less than the one of the optimal solution. Hence, for small $d_D$, to satisfy the rate constraint our proposed solution connects to the neighbours using OF links since the hybrid RF/FSO links do not satisfy the constraint. The optimal solution connects to more nodes (outside the neighbours sets) to meet the rate constraint. We further note that for $d_R=2$ Km, the improvement in the provided data rate of the hybrid RF/FSO link does not decrease the total cost of the network as the used connections are the same. This can be explained by the fact that the solution is limited by the reliability constraint. Therefore, there is no gain in improving the provided data rate of the hybrid RF/FSO link unless the reliability of the connection is improved simultaneously.
\fref{fig:DRC} and \fref{fig:DRR} show the total cost of the network and the percentage of used OF links, respectively, against the maximum distance satisfying the reliability constraint $d_R$ for a system composed of $M=7$ base-stations and for different prices of hybrid RF/FSO of $\pi^{(h)}=20$ k\$. From \fref{fig:DRC}, as the reliability of the hybrid RF/FSO link increases, the proposed solution provides a cost similar to the optimal one. This can be explained by the fact that the reliability condition can be satisfied by connecting to base-stations inside the set of neighbours using exclusively hybrid RF/FSO links. The analysis is more corroborated by \fref{fig:DRR} that shows that the proposed solution uses more and more hybrid RF/FSO links as the reliability of such links increases. Hence, even if the optimal and heuristic hybrid FR/FSO plans are different for high reliability, the total deployment cost of the network is very similar.
Finally, to quantify the performance of the proposed solution against the target reliability,\fref{fig:AC} and \fref{fig:AR} plot the total cost of the network and the percentage of used OF links, respectively, against the reliability threshold $\alpha$ for a system composed of $M=7$ base-stations and for different prices of hybrid RF/FSO links. Again, from \fref{fig:AR} we clearly see that for expensive hybrid RF/FSO transceivers both the optimal solution and our proposed solution use exclusively OF links. \fref{fig:AC} shows that, even for cheap hybrid RF/FSO transceivers, our proposed solution performs as good as the optimal solution for a reliability $\alpha \geq 0.8$ even if the nature of the used links is not the same for that price as displayed in \fref{fig:AR}.
\section{Conclusion}\label{sec:conc}
In this paper, we consider the problem of backhaul network design using the OF and hybrid RF/FSO technologies. We first formulate the planning problem under connectivity, rate constraints, and reliability. We, then, solve the problem optimally when only OF links are allowed. Using the solution of the OF deployment, we formulate an approximation of the general planning problem and show that under a realistic assumption about the relative cost of the OF links and the hybrid RF/FSO transceivers, the solution can be expressed as a maximum weight clique in the planning graph. Simulation results show that our approach shows a close-to-optimal performance, especially for practical prices of the hybrid RF/FSO. As a future research direction, network design can be investigated while taking into account the varying reliability the hybrid RF/FSO links.
\appendices
\numberwithin{equation}{section}
\section{Proof of \lref{l1}}\label{ap1}
To show this lemma, this section express the objective function and the system constraints C1 to C5. The objective function can be written as:
\begin{align}
\cfrac{1}{2} \sum_{i=1}^M \sum_{j=1}^M X_{ij}\pi^{(O)}_{ij} + Y_{ij}\pi^{(h)}_{ij}.
\label{eqobj}
\end{align}
\begin{remark}
Naturally, the objective function should not include the price of the pre-deployed OF links. In other words, it should be written as:
\begin{align}
\cfrac{1}{2} \sum_{i=1}^M \sum_{j=1}^M (X_{ij}-P_{ij})\pi^{(O)}_{ij} + Y_{ij}\pi^{(h)}_{ij}.
\end{align}
However, as the term $- 1/2\sum_{i=1}^M \sum_{j=1}^M P_{ij}\pi^{(O)}_{ij}$ is constant with respect to the optimization variables $X_{ij}$ and $Y_{ij}$, then it is removed throughout this paper (including the simulations).
\end{remark}
The pre-deployed OF connections condition, i.e., constraint C1, states that the planning solution should include the pre-deployed OF links. In other words, if nodes $b_i$ and $b_j$ have a pre-deployed OF link $P_{ij}=1$, then the solution should have the same connection $X_{ij}=1$. However, the non-existence of a pre-deployed OF link does not add extra constraints to the system. Therefore, constraint C1 can be written, for arbitrary nodes $b_i$ and $b_j$, as follows:
\begin{align}
X_{ij}P_{ij}=P_{ij}.
\label{eqc1}
\end{align}
The system constraint C2 implies that, at maximum, only one type of connection may exist between any arbitrary nodes $b_i$ and $b_j$. Hence, constraint C2 can be mathematically written as follows:
\begin{align}
X_{ij}Y_{ij} = 0.
\label{eqc2}
\end{align}
For an arbitrary BS $b_i$, the data rate constraint C3 is satisfied if the sum of the data rate provided by all adjacent nodes exceeds the targeted data rate. Therefore, the constraint can be formulated for all BS $b_i$ as follows:
\begin{align}
&\sum_{j=1}^M X_{ij}D^{(O)}_{ij}+ Y_{ij}D^{(h)}_{ij} \geq D_t.
\end{align}
As OF links always satisfy the targeted data rate, i.e., $D^{(O)}_{ij} \geq D_t, \ \forall \ i \neq j$, then the data rate constraint C3 for node $b_i$ can be reformulated as follows:
\begin{align}
&\sum_{j=1}^M X_{ij}D_t + Y_{ij}D^{(h)}_{ij} \geq D_t.
\label{eqc3}
\end{align}
The reliability constraint C4 implies that each node should be connected to the network, at all time, with probability $\alpha$. As the reliability of each link is independent of the other links, such constraint can be formulated for node $b_i$ using the complementary event as follows:
\begin{align}
1 - \prod_{j=1}^M (1 - X_{ij} R^{(O)}_{ij})(1 - Y_{ij} R^{(h)}_{ij}) \geq \alpha.
\end{align}
As OF links are always reliable, i.e., $R^{(O)}(d_{ij}) \geq \alpha \ \forall i \neq j$, the reliability constraint can be simplified as follows:
\begin{align}
1 - \prod_{j=1}^M (1 - X_{ij}\alpha)(1 - Y_{ij} R^{(h)}_{ij}) \geq \alpha.
\label{eqc4}
\end{align}
Define $\mathbf{C} = [c_{ij}]$ as the adjacency matrix by $c_{ij} = X_{ij}+Y_{ij}$. Since only one type of connections exists between the same BSs then $c_{ij}$ is a binary variable (i.e., $c_{ij} \in \{0,1\}$). The connectivity constraint C5 implies that the graph representing the base-stations is connected. From a graph theory perspective \cite{25181258}, such graph connectivity constraint can be expressed as a function of the Laplacian matrix $L$ defined as $\mathbf{L} = \mathbf{D}-\mathbf{C}$, where $\mathbf{D}=\text{diag}(d_1,\ \cdots,\ d_M)$ is a diagonal matrix with $d_i = \sum_{j=1}^M c_{ij}$. The diagonalization of the Laplacian matrix is given by $\mathbf{L}= \mathbf{Q}\mathbf{\Lambda} \mathbf{Q}^{-1}$, where $\mathbf{\Lambda}= \text{diag}(\lambda_1,\ \cdots,\ \lambda_M)$ with $\lambda_1 \leq \lambda_2 \leq \cdots \leq \lambda_M$. The connectivity condition C5 of the matrix can be written using the algebraic formulation proposed in \cite{25181258} as:
\begin{align}
\lambda_2 > 0.
\label{eqc5}
\end{align}
Combining the objective function \eref{eqobj} and the constraints \eref{eqc1}, \eref{eqc2}, \eref{eqc3}, \eref{eqc4}, and \eref{eqc5}, gives the optimization problem proposed in \eref{Original_optimization_problem}.
\section{Proof of \thref{th1}}\label{ap3}
To show the theorem, this section proposes first to demonstrate that \algref{alg1} outputs the optimal solution to the optimization problem stated in \lref{l2} for a network without pre-deployed links. The second part of the section extends the result to network with pre-deployed OF connections.
\subsection{Network Without Pre-deployed OF Links}
To proof this theorem for a network without pre-deployed links, we first prove that \algref{alg1} produces a feasible solution to the problem. Afterward, we show that any graph that can be reduced, using \algref{alg2}, to a single cluster includes the graph outputted by \algref{alg1}. Finally, we show that any solution that cannot be reduced to a single cluster is not optimal.
\algref{alg2} can be seen as a complement of \algref{alg1}. As for \algref{alg1}, in \algref{alg2} begins by generating a cluster of each BS in the system. Afterward, two groups at the minimum price of each other and whose BSs at the minimum price are connected are merged into a single cluster. The process is repeated until no further connection can be found. The steps of the algorithm are summarized in \algref{alg2}.
Let $\overline{\mathcal{Z}}= \{\overline{Z}_1,\ \cdots,\ \overline{Z}_{|\overline{\mathcal{Z}}|}\}$ be the clustering at any step of \algref{alg1}. First note that $\overline{\mathcal{Z}}$ is a partition of $\mathcal{B}$. We proof by induction that nodes inside any cluster $\overline{Z}_i,\ 1 \leq i \leq |\overline{\mathcal{Z}}|$ are connected. Clearly, for a cluster $\overline{Z}_i$ with $|\overline{Z}_i|=1$ (the cluster contains a single node), all nodes inside the cluster are connected. Assume that all clusters $\overline{Z}_i$ of size $|\overline{Z}_i| \leq n$ are connected. From the last step of \algref{alg1} clusters of size $n+1$ can be generated only by merging two clusters $\overline{Z}_j$ and $\overline{Z}_k$ with $|\overline{Z}_j|,|\overline{Z}_k| < n$ and $|\overline{Z}_j|+|\overline{Z}_k|=n+1$. Since by construction such clusters are connected ($\overline{X}_{jk}=\overline{X}_{kj}=1$ with $b_j \in \overline{Z}_j$ and $b_k \in \overline{Z}_k$), then the resulting cluster $\overline{Z}_i$ from merging $\overline{Z}_j$ and $\overline{Z}_k$ is also connected. Therefore, all nodes within any arbitrary cluster $\overline{Z}_i,\ 1 \leq i \leq |\overline{\mathcal{Z}}|$ are connected. Finally, since $\overline{\mathcal{Z}}$ contains a single cluster at the end of \algref{alg1} (i.e., $|\overline{\mathcal{Z}}|=1$) and it is a partition of $\mathcal{B}$ (contains all the nodes in the network), then all the nodes are connected. Hence the outputted solution satisfy constraint \eref{eq:8}. By construction, we can easily see that the connections are binary and symmetric. In other words, the outputted solution satisfy constraints \eref{eq:9} and \eref{eq:10} which conclude that it is a feasible solution.
\begin{algorithm}[t]
\begin{algorithmic}
\REQUIRE $\mathcal{B}, \pi^{(O)},$ and $X_{ij},\ 1 \leq i,j\leq M$.
\STATE Initialize $\mathcal{Z} = \varnothing$.
\FORALL {$b \in \mathcal{B}$}
\STATE $\mathcal{Z} = \{\mathcal{Z},\{b\}\}$.
\ENDFOR
\STATE Initialize $t=$TRUE.
\WHILE {$t=$TRUE}
\STATE $t=$FALSE.
\FORALL {$Z \neq Z^{\prime} \in \mathcal{Z}$}
\IF {$\sum\limits_{\substack{b_i \in Z \\ b_j \in Z^{\prime}}}X_{ij} = 1$}
\STATE $Z^* = \arg \min\limits_{\substack{X \in \mathcal{Z} \\ X \neq Z^{\prime}}} \left[ \min\limits_{\substack{b \in X \\ b^{\prime} \in Z^{\prime}}} \pi^{(O)}(b , b^{\prime})\right]$
\STATE $Z^{\prime *} = \arg \min\limits_{\substack{X \in \mathcal{Z} \\ X \neq Z}} \left[ \min\limits_{\substack{b \in X \\ b^{\prime} \in Z}} \pi^{(O)}(b^{\prime} ,b)\right]$.
\IF {$Z = Z^*$ and $Z^{\prime}=Z^{\prime*}$}
\STATE $(b_i,b_j) = \arg \min\limits_{\substack{b \in Z \\ b^{\prime} \in Z^{\prime}}} \pi^{(O)}(b , b^{\prime})$.
\IF {$X_{ij}=1$}
\STATE $\mathcal{Z} = \mathcal{Z} \setminus \{Z\}$
\STATE $\mathcal{Z} = \mathcal{Z} \setminus \{Z^{\prime}\}$
\STATE $\mathcal{Z} = \{\mathcal{Z},\{Z_i,Z_j\}\}$.
\STATE $t=$TRUE
\ENDIF
\ENDIF
\ENDIF
\ENDFOR
\ENDWHILE
\end{algorithmic}
\caption{Clustering Algorithm}
\label{alg2}
\end{algorithm}
Let $X_{ij}, 1 \leq i,j \leq M$ be a feasible solution and let $\mathcal{Z}$ be the outputted clustering by \algref{alg2} when inputted $X_{ij}$. We can clearly see that $\mathcal{Z}$ is a partition of $\mathcal{B}$. Therefore, if $|\mathcal{Z}|=1$, then $\mathcal{Z}=\overline{\mathcal{Z}}$ since there exist only a unique partition containing a single element (the set $\mathcal{B}$ itself). This concludes that any graph that can be reduced, using \algref{alg2}, to a single cluster includes all the connections that are created in the graph outputted by \algref{alg1}. Since $\pi^{(O)}$ is a strictly positive function and that the graph outputted by \algref{alg1} have the minimum number of connections among all the graphs that can be reduced to a single cluster using \algref{alg2}, then the solution of \algref{alg1} is the best solution among the solutions that can be reduced to a single cluster using \algref{alg2}.
Now assume that $|\mathcal{Z}| \neq 1$. We can clearly see that $|\mathcal{Z}| \geq 3$. Otherwise, if there exist only two clusters (i.e., $|\mathcal{Z}| =2$) and since the solution is feasible, then they are connected. Due to the fact they are only two, then two cases can be distinguished:
\begin{itemize}
\item The BSs at the minimum price of each others are connected and hence they can be reduced to a single cluster. Therefore, clustering $|\mathcal{Z}| =2$ cannot be outputted by \algref{alg2}.
\item The BSs at the minimum price of each others are not connected. Then the solution having the same connections except for the link between the two cluster being replaced with the connection of BSs at the minimum price of each others produces a feasible solution at a lower cost. Therefore, the initial solution is not optimal.
\end{itemize}
Let $\mathcal{Z}= \{Z_1,\ \cdots,\ Z_{|\mathcal{Z}|}\}, |\mathcal{Z}| \geq 3$ be the outputted clustering by \algref{alg2}. Define the reduced clustering as $\tilde{\mathcal{Z}}= \{Z_1,\ Z_2,\ \{Z_3 \cup \cdots \cup \ Z_{|\mathcal{Z}|}\}\ \} = \{\tilde{Z_1},\ \tilde{Z_2},\ \tilde{Z_3}\}$. We can clearly see that $|\tilde{\mathcal{Z}}| = 3$ with none of the clusters connected and at minimum price of each other.
\begin{lemma}
For any three points in the plane, there must exist two points at the minimum distance from each others.
\label{l3}
\end{lemma}
\begin{proof}
Let $a,b,c$ be the three points in the plane and assume that there do not exist two points at minimum distance of each other. The only possible configuration (up to a permutation of the points) is that $a$ at minimum distance from $b$ and $b$ is not. Hence $b$ is at minimum distance from $c$ which is at its turn at minimum distance from $a$. These conditions yield $d(a,b)<d(a,c)$, $d(b,c)<d(b,a)$ and $d(c,a)<d(c,b)$. Since the distance operator is symmetric then, $d(c,a)<d(c,b)=d(b,c)<d(b,a)=d(a,b)<d(a,c)$. In other words, $d(a,c)<d(a,c)$, which is impossible. Therefore for any three points in the plane, there must exist two points at minimum distance of each others.
\end{proof}
From \lref{l3}, there must exist two clusters at the minimum price of each other. Since they have not been reduced to a single cluster, then they are not connected. For simplicity, assume $\tilde{Z_1}$ and $\tilde{Z_2}$ are at minimum price of each other and since the graph is connected then $\tilde{Z_1}$ and $\tilde{Z_3}$ are connected and similarly for $\tilde{Z_2}$ and $\tilde{Z_3}$. Moreover, it can be easily concluded that all nodes inside the clusters are connected. Otherwise, assume $\tilde{Z_3}$ can be split into two non-connected clusters $\tilde{Z}$ and $\tilde{Z}^{\prime}$ with $\tilde{Z_1}$ connected only to $\tilde{Z}$ and $\tilde{Z_2}$ connected only to $\tilde{Z}^{\prime}$. Then, since $\tilde{Z_1}$ and $\tilde{Z_2}$ are not connected, the whole graph is not connected and the solution is not feasible.
The clustering connecting $\tilde{Z_1}$ with $\tilde{Z_2}$ and $\tilde{Z_1}$ with $\tilde{Z_3}$ (or $\tilde{Z_2}$ with $\tilde{Z_3}$) produces also a feasible solution at a lower cost since the sum of the prices is minimized ($\pi^{(O)}(\tilde{Z_1},\tilde{Z_2}) < \min(\pi^{(O)}(\tilde{Z_1},\tilde{Z_3}),\pi^{(O)}(\tilde{Z_2},\tilde{Z_3}))$). Therefore, $\tilde{\mathcal{Z}}$ is not optimal and by extension $\mathcal{Z}$ is also not optimal. Finally, we can conclude that the optimal solution is the solution containing a single cluster. Therefore, the solution outputted by \algref{alg1} is the optimal solution to the problem proposed in \lref{l2}.
\subsection{Network With Pre-deployed OF Links}
It can explicitly be noted that for a network with pre-deployed OF links, \algref{alg1} produces a feasible solution. In fact, as for the previous subsection, the solution outputted by the algorithm satisfy the constraints \eref{eq:8}, \eref{eq:9}, and \eref{eq:10}. Furthermore, from the initialization of the variables $X_{ij}, 1 \leq i,j \leq M$, the solution satisfies constraint \eref{eq:np2}.
From the initialization of the clusters, all the base-stations that are connected with pre-deployed OF links are merged in the same cluster. Such initialization implies the following two properties:
\begin{enumerate}
\item Only the pre-deployed OF links connect such BSs.
\item Any node connected to the cluster is connected to the BS with minimal cost inside that cluster.
\end{enumerate}
Therefore, to show that the solution of \algref{alg1} is the optimal planning for the problem proposed in \lref{l2}, it is sufficient to show that any solution that violates the properties above is not optimal. Finally, as clusters can be seen as new nodes in a network without pre-deployed OF links, the result of the first part of the section guarantee the optimality of the solution.
Assume that in the optimal solution, a connection exists between two nodes in the same cluster. As these nodes are connected via pre-deployed OF link through single or multi-hop connection, then removing the extra connection produces a feasible solution at a lower cost. Hence, the optimal solution contains only the pre-deployed OF links connecting BSs in each cluster.
Similarly, assume there exists a node connected to a cluster in the optimal solution that is not linked to the BS with minimal cost in that cluster. It can be readily seen that as the nodes in the cluster are all connected, then removing the connection and replacing it with the minimal cost one produces a feasible solution at a lower cost. Therefore, such connection does not exist in the optimal solution. Finally, the solution generated by \algref{alg1} is the optimal planning for the problem proposed in \lref{l2}.
\section{Proof of \lref{l4}}\label{ap4}
To prove this lemma, we prove that any solution to \eref{Approximate_optimization_problem} is a feasible solution to \eref{Original_optimization_problem}. We first show that constraint \eref{eq:nr2} is equivalent to constraint \eref{eq:nr}. In the original problem formulation \eref{Original_optimization_problem}, the reliability constraint is:
\begin{align}
1 - \prod_{j=1}^M (1 - X_{ij}\alpha)(1 - Y_{ij} R^{(h)}_{ij}) \geq \alpha.
\label{eq:ap41}
\end{align}
It can easily be seen that \eref{eq:ap41} is satisfied, for node $b_i$ if and only if at least one of the following options is correct:
\begin{enumerate}
\item $\sum_{j=1}^M X_{ij} > 0 \Rightarrow$ Node $b_i$ is connected to another node with an OF link which provides full reliability.
\item $\sum_{j \in \mathcal{R}_i}^M Y_{ij} > 0 \Rightarrow$ Node $b_i$ is connected to another node with a hybrid RF/FSO link to a nearby node that provides full reliability.
\item $1 - \prod_{j \in \overline{\mathcal{R}}_i}(1 - Y_{ij} R^{(h)}_{ij}) \geq \alpha \Rightarrow$ Node $b_i$ is connected to a sufficiently large number of nodes to satisfy the reliability constraint.
\end{enumerate}
In the rest part of the proof, we show that constraint \eref{eq:nr2} regroups the three scenarios mentioned above. Applying a logarithmic transformation to the third option and rearranging the terms yields the following expression:
\begin{align}
\sum_{j \in \overline{\mathcal{R}}_i}\log (1-Y_{ij} R^{(h)}_{ij}) \leq \log(1-\alpha).
\end{align}
Given that the reliability of nodes $j \notin \mathcal{R}_i$ is small, i.e., $R^{(h)}_{ij} \lll 1$, then $Y_{ij} R^{(h)}_{ij} \lll 1$, $\forall \ j \notin \mathcal{R}_i$. Applying a first order Taylor expansion of the logarithm yield the condition:
\begin{align}
\sum_{j \in \overline{\mathcal{R}}_i} -Y_{ij} R^{(h)}_{ij} \leq \log(1-\alpha).
\label{eq:ap42}
\end{align}
Let $\tilde{\alpha} = \cfrac{1}{\log(1-\alpha)}$. Rearranging the terms of \eref{eq:ap42} and including the first and second option discussed in the previous paragraph gives the following constraint:
\begin{align}
\sum_{j=1}^M X_{ij}\tilde{\alpha} + \sum_{j \in \mathcal{R}_i}Y_{ij}\tilde{\alpha} + \sum_{j \in \overline{\mathcal{R}}_i}Y_{ij} R^{(h)}_{ij} \geq \tilde{\alpha}.
\label{eq:ap43}
\end{align}
It can clearly be seen that setting $\sum_{j=1}^M X_{ij} > 0$, and/or $\sum_{j \in \mathcal{R}_i}^M Y_{ij} > 0$, and/or $1 - \prod_{j \in \overline{\mathcal{R}}_i}(1 - Y_{ij} R^{(h)}_{ij}) \geq \alpha$ automatically satisfy \eref{eq:ap43} which proves that constraint \eref{eq:nr2} is equivalent to constraint \eref{eq:nr}.
We, now, show that constraint \eref{eq:14} is included in constraint \eref{eq:6} as it is the second and last constraint changing from one formulation to the other. Constraint \eref{eq:14} ensures that, for all connections $\overline{X}_{ij}=1$ that are generated by \algref{alg1}, a similar connections (OF or hybrid RF/FSO link) between nodes $b_i$ and $b_j$ must exist. For connections $\overline{X}_{ij}=0$, the constraint is always satisfied and connection may or may not exist. From \thref{th1}, \algref{alg1} produces a connected graph. In other words, $\lambda_2 > 0$. Therefore, constraint \eref{eq:14} is included in constraint \eref{eq:6}. A feasible solution to \eref{Approximate_optimization_problem} is, therefore, a feasible solution to \eref{Original_optimization_problem}. In \thref{th2}, we show that the optimal solution to \eref{Approximate_optimization_problem} is the optimal solution to \eref{Original_optimization_problem} in many scenarios (but not all). Therefore, the approximation of problem \eref{Original_optimization_problem} by the problem \eref{Approximate_optimization_problem} is tight.
\section{Proof of \thref{th2}}\label{ap5}
In this theorem, we show that the optimal solution $X^*_{ij},Y^*_{ij},\ 1 \leq i,j \leq M$ to \eref{Approximate_optimization_problem} should satisfy $X^*_{ij}+Y^*_{ij}=1$ only if $(i,j) \in \mathcal{N}_j \times \mathcal{N}_i$.
\begin{remark}
Note that if any feasible solution $X_{ij},Y_{ij}$ to the general problem \eref{Original_optimization_problem} that verify \thref{th2}, then the solution is feasible to \eref{Approximate_optimization_problem}. In other words, if $X_{ij},Y_{ij}$ feasible to \eref{Original_optimization_problem} and $X_{ij}+Y_{ij}=1$ only if $(i,j) \in \mathcal{N}_j \times \mathcal{N}_i$, then $(X_{ij}+Y_{ij})\overline{X}_{ij}=\overline{X}_{ij}$. This can be easily concluded given the construction of $\mathcal{N}_i, \leq i \leq M$ as the minimum set of nodes that can generate a connected graph. Since all $(x,y)$ such that $\overline{X}_{xy}=1$ are at the edge of at least one of the $\mathcal{N}_i$, then the only connected solution that satisfies \thref{th2} is feasible to \eref{Approximate_optimization_problem}. In that scenario, the optimal solution of \eref{Original_optimization_problem} and \eref{Approximate_optimization_problem} are the same.
\end{remark}
Assume $\exists (x,y)$ such that $X^*_{xy}+Y^*_{xy}=1$ and $(x,y) \notin \mathcal{N}_y \times \mathcal{N}_x$. Two scenarios can be distinguished:
\begin{itemize}
\item $\mathcal{B} \setminus \{b_x,b_y\}$ represents a connected subgraph.
\item $\mathcal{B} \setminus \{b_x,b_y\}$ is not a connected subgraph.
\end{itemize}
For the first scenario, consider the reduced network $\{b_x,b_y,\tilde{b}_k\}$. Clearly, we have $b_{x^*},b_{y^*} \in \tilde{b}_k$. Define the following planning:
\begin{align}
\tilde{X}_{ij} &=
\begin{cases}
1 \hspace{1cm} &\text{if } i=x \text{ and } j=x^* \\
1 \hspace{1cm} &\text{if } j=y \text{ and } j=y^* \\
0 \hspace{1cm} &\text{if } i=x \text{ and } j=y \\
X^*_{ij} \hspace{1cm} &\text{otherwise}
\end{cases} \nonumber \\
\tilde{Y}_{ij} &=
\begin{cases}
0 \hspace{1cm} &\text{if } i=x \text{ and } j=x^* \\
0 \hspace{1cm} &\text{if } j=y \text{ and } j=y^* \\
0 \hspace{1cm} &\text{if } i=x \text{ and } j=y \\
Y^*_{ij} \hspace{1cm} &\text{otherwise},
\end{cases}
\end{align}
in which the connection between $b_x$ and $b_y$ is replaced by two connections between $b_x$ and $b_{x^*}$ and between $b_y$ and $b_{y^*}$. We can clearly see that the network is connected. Moreover, since $b_x$ and $b_y$ are connected with an OF link, then the data rate constraint is satisfied for both nodes. Therefore, $\tilde{X}_{ij},\tilde{Y}_{ij},\ 1 \leq i,j \leq M$ represents a feasible solution. Moreover, The difference in cost between the optimal planning $X^*_{ij},Y^*_{ij}$ and the planning $\tilde{X}_{ij},\tilde{Y}_{ij}$ is lower bounded by:
\begin{align}
&\pi(b_x,b_y) - (\pi^{(O)}(b_x,b_{x^*}) + \pi^{(O)}(b_y,b_{y^*})) \\
&\geq \pi^{(h)}(b_x,b_y) - (\pi^{(O)}(b_x,b_{x^*}) + \pi^{(O)}(b_y,b_{y^*})) \geq 0.\nonumber
\end{align}
From assumption \eref{eq:12}, the difference is positive. This concludes that $X^*_{i,j},Y^*_{i,j},\ 1 \leq i,j \leq M$ is not the optimal solution.
\begin{remark}
For scenario 1, $X^*_{i,j},Y^*_{i,j},\ 1 \leq i,j \leq M$ can be the optimal solution to the original problem \eref{Original_optimization_problem}. Hence, for this configuration, the optimal solution of \eref{Original_optimization_problem} and \eref{Approximate_optimization_problem} are the same.
\end{remark}
For scenario 2, let the network be reduced to $\{b_x,b_y,\tilde{b}_k,\tilde{b}_l\}$ with $b_x$ connected to $\tilde{b}_k$, which is a connected subgraph, $b_y$ connected $\tilde{b}_l$, which is a connected subgraph, and $\tilde{b}_k$ and $\tilde{b}_l$ are not connected. Since $X^*_{i,j},Y^*_{i,j},\ 1 \leq i,j \leq M$ is a feasible solution to \eref{Approximate_optimization_problem}, then it satisfies constraint \eref{eq:14}. In other words, $X^*_{k,l} + Y^*_{k,l}=1, \ \forall \ k,l$ such that $\overline{X}_{kl} =1$. Note that $\overline{X}_{xy} = 0$. Otherwise, by construction of the neighbours sets, we have $(x,y) \in \mathcal{N}_y \times \mathcal{N}_x$. Define the planning $\tilde{X}_{ij},\tilde{Y}_{ij}$ such that
\begin{align}
\tilde{X}_{ij} &=
\begin{cases}
0 \hspace{1cm} &\text{if } i=x \text{ and } j=y \\
X^*_{ij} \hspace{1cm} &\text{otherwise}
\end{cases} \nonumber \\
\tilde{Y}_{ij} &=
\begin{cases}
0 \hspace{1cm} &\text{if } i=x \text{ and } j=y \\
Y^*_{ij} \hspace{1cm} &\text{otherwise}.
\end{cases}
\end{align}
The planning $\tilde{X}_{ij},\tilde{Y}_{ij}$ satisfy \eref{eq:14}. However, we can clearly see that the graph is not connected. Therefore, $X^*_{i,j},Y^*_{i,j},\ 1 \leq i,j \leq M$ is not a feasible solution. This concludes that scenario 2 is not feasible. Finally, we conclude that the optimal solution $X^*_{i,j},Y^*_{i,j},\ 1 \leq i,j \leq M$ to \eref{Approximate_optimization_problem} should satisfy $X^*_{xy}+Y^*_{xy}=1$ only if $(x,y) \in \mathcal{N}_y \times \mathcal{N}_x$.
\begin{remark}
Scenario 2 can be a feasible scenario if $X^*_{i,j},Y^*_{i,j},\ 1 \leq i,j \leq M$ is the optimal solution to the original problem \eref{Original_optimization_problem}. In that case, two scenarios can be distinguished:
\begin{itemize}
\item $b_{x^*} \in \mathcal{N}_y$ and $b_{y^*} \in \mathcal{N}_x$. In that case $\tilde{X}_{ij},\tilde{Y}_{ij}$ presented for scenario 1 produces a feasible solution with lower cost. Therefore, the optimal solution of \eref{Original_optimization_problem} and \eref{Approximate_optimization_problem} are the same.
\item $b_{x^*} \notin \mathcal{N}_y$ or $b_{y^*} \notin \mathcal{N}_x$. In this configuration, no conclusion can be reached about the optimal solution of \eref{Original_optimization_problem} and \eref{Approximate_optimization_problem} is an upper bound of the minimum of \eref{Original_optimization_problem}.
\end{itemize}
Therefore, the approximation of the problem \eref{Original_optimization_problem} by the problem \eref{Approximate_optimization_problem} is a tight approximation.
\end{remark}
\section{Proof of \thref{th3}}\label{ap6}
To proof this theorem, we first prove that there is a one to one mapping between the set of feasible solution of a modified version of problem \eref{Approximate_optimization_problem} and the set of cliques of degree $M$ in the \emph{planning} graph $\mathcal{G}(\mathcal{V},\mathcal{E})$. To conclude the proof, we show that the weight of the clique is equivalent to the merit function of the optimization problem \eref{Approximate_optimization_problem}.
From \thref{th2}, we have $X_{ij}=0$ and $Y_{ij}=0,\ \forall \ (b_i,b_j) \notin \mathcal{N}_j \times \mathcal{N}_i $. Hence the objective function \eref{eq:25} and constraint \eref{eq:23} can be replaced by:
\begin{align}
&\max -\cfrac{1}{2} \sum_{i=1}^M \sum_{j \in \mathcal{N}_i} X_{ij}\pi^{(O)}_{ij} + Y_{ij}\pi^{h}_{ij} .
\end{align}
Similarly, since $X_{ij}=0$ and $Y_{ij}=0,\ \forall \ (b_i,b_j) \notin \mathcal{N}_j \times \mathcal{N}_i $, then $(X_{ij},Y_{ij}) = (X_{ji},Y_{ji})$ is always verified for $(b_i,b_j) \notin \mathcal{N}_j \times \mathcal{N}_i $. Hence the constraints \eref{eq:20}, \eref{eq:21}, and \eref{eq:22} can be replaced by:
\begin{align}
&(X_{ij},Y_{ij}) = (X_{ji},Y_{ji}) ,\ \forall \ (b_i,b_j) \in \mathcal{N}_j \times \mathcal{N}_i \nonumber \\
&X_{ij}Y_{ij} = 0,\ \forall \ (b_i,b_j) \in \mathcal{N}_j \times \mathcal{N}_i \nonumber \\
&\sum_{j \in \mathcal{N}_i} X_{ij}D_t + Y_{ij}D^{(h)}_{ij} \geq 1,\ D_t \leq i \leq M.
\end{align}
By definition of the set $\mathcal{N}_i$, we have $\overline{X}_{ij}=0, \forall \ j \notin \mathcal{N}_i$. Therefore, constraint \eref{eq:14} may be written as:
\begin{align}
&(X_{ij} + Y_{ij})\overline{X}_{ij} = \overline{X}_{ij}, \forall \ j \in \mathcal{N}_i, 1 \leq i \leq M.
\end{align}
The problem \eref{Approximate_optimization_problem} can be reformulated as:
\begin{align}
\max &-\cfrac{1}{2} \sum_{i=1}^M \sum_{j \in \mathcal{N}_i} X_{ij}\pi^{(O)}_{ij} + Y_{ij}\pi^{h}_{ij} \nonumber \\
\text{subject to } &(X_{ij},Y_{ij}) = (X_{ji},Y_{ji}) ,\ \forall \ (b_i,b_j) \in \mathcal{N}_j \times \mathcal{N}_i \nonumber \\
&X_{ij}P_{ij} = P_{ij} ,\ \forall \ (b_i,b_j) \in \mathcal{N}_j \times \mathcal{N}_i \nonumber \\
&X_{ij}Y_{ij} = 0,\ \forall \ (b_i,b_j) \in \mathcal{N}_j \times \mathcal{N}_i \nonumber \\
&\sum_{j \in \mathcal{N}_i} X_{ij}D_t + Y_{ij}D^{(h)}_{ij} \geq D_t,\ 1 \leq i \leq M \nonumber \\
&\sum_{j \in \mathcal{N}_i \cap \mathcal{R}_i} (X_{ij}+Y_{ij})\tilde{\alpha} + \sum_{j \in \in \mathcal{N}_i \cap\overline{\mathcal{R}}_i}Y_{ij} R^{(h)}_{ij} \geq \tilde{\alpha} \nonumber \\
&(X_{ij} + Y_{ij})\overline{X}_{ij} = \overline{X}_{ij}, \forall \ j \in \mathcal{N}_i, 1 \leq i \leq M \nonumber \\
&X_{ij},Y_{ij} \in \{0,1\},\ 1 \leq i,j\leq M.
\end{align}
Let $\gamma_i = \{(X_{ij_1},Y_{ij_1}),\ \cdots,\ (X_{ij_{|\mathcal{N}_i|}},Y_{ij_{|\mathcal{N}_i|}})\}$ be the new variable. Using the variables $\gamma_i$ and the definition of the sets $\mathcal{C}_i$, the problem can be written as:
\begin{subequations}
\begin{align}
\max &\sum_{i=1}^M w(\gamma_i) \nonumber \\
\text{subject to } &\gamma_i \in \mathcal{C}_i, 1 \leq i \leq M \label{eq:31} \\
&(X_{ij},Y_{ij}) = (X_{ji},Y_{ji}) ,\ \forall\ 1 \leq i \neq j \leq M. \label{eq:30}
\end{align}
\label{eq:32}
\end{subequations}
Let $\mathbb{C}$ be the set of cliques of degree $M$ in the \emph{planning} graph and let $\mathbb{F}$ be the set of feasible solutions to the optimization problem \eref{eq:32}. We first proof that any clique $C=\{v_1,\ \cdots,\ v_M\} \in \mathbb{C}$ satisfy constraints \eref{eq:31}, and \eref{eq:30}. Then, we prove the converse. In other words, for element in $\mathbb{F}$, there exists a clique in $\mathbb{C}$.
Let $C=\{v_1,\ \cdots,\ v_M\} \in \mathbb{C}$. Assume $\exists \ k,i,j$ such that $v_i,v_j \in \mathcal{C}_k$. Since all the vertices in a clique are connected, then from the connectivity condition C1, vertices $v_i$ and $v_j$ are not connected. Hence $\nexists \ k,i,j$ such that $v_i,v_j \in \mathcal{C}_k$. Given that the clique contain $M$ elements, then constraint \eref{eq:31} is satisfied. The connectivity condition C2 ensures that $(X_{ij},Y_{ij}) = (X_{ji},Y_{ji})$ for all vertices. Therefore, $C$ is a feasible solution to \eref{eq:32}. Similarly, let $\{c_1,\ \cdots,\ c_M\}$ be a feasible solution to \eref{eq:32}, then clearly the vertices corresponding to each cluster are connected. Finally, there a one to one mapping between $\mathbb{C}$ and $\mathbb{F}$. Moreover, the weight of the clique is $w(C) = \sum_{i=1}^M w(\gamma_i) $ which conclude that the solution of \eref{eq:32} is the maximum weight clique, among the clique of size $M$, in the \emph{planning} graph.
\bibliographystyle{IEEEtran}
|
\subsection{{Estimates of expectations}}
By definition, the moment estimation in \eqref{eq:force} requires small fragments of trajectories
passing through each point of the membrane surface, which is precisely the massive data generated by
the sptPALM method \cite{Manley,Giannone} on biological samples. The empirical estimator defined by
\eqref{eq:force} has been used over the past 60 years in signal processing \cite{Schuss3} and
recently applied in several cell biology contexts \cite{Verstergaard}. {The empirical estimator used in \cite{Hoze2012} is optimal as its variance is minimal and equals to the Cram\'er-Rao lower bound, as shown in section IIIE of \cite{Verstergaard}, using numerical simulations. This approach is clearly different from the Bayesian method (see formula 6 in \cite{Masson} and previous publications).}
A positive point of the manuscript \cite{Masson} is the attempt to include in the analysis the influence of the Gaussian instrumental noise on the localization of moving receptors and thus on their trajectories. { However, the variance term in formula 6 of \cite{Masson} is inapplicable to sort out the biophysical parameters when there is a force or a drift component that can vary significantly in space. This is indeed the case at potential wells. The additional term, needed in formula 6 of \cite{Masson} appears in the direct estimation of the variance $E\left( (\mbox{\boldmath$X$}_{n+1}-\mbox{\boldmath$X$}_n)^2|\mbox{\boldmath$X$}_n=\mbox{\boldmath$X$} \right)$ and is proportional to the Laplacian of a potential well $\Delta V$, multiplied by the variance of the localization precision.} Finally, the motion blur due to the open camera shutter during acquisition can also be included in the analysis, as recently discussed in \cite{Verstergaard}. \\[1mm]
\textbf{The potential energy} \\[1mm]
The fluctuation-dissipation principle invoked in \cite{Masson} expresses the relation between the local friction acting on the diffusing receptor and the diffusion coefficient. This relation is microscopic and is a part of the collision model of diffusion. The observed coarse-grained dynamics, which depends on obstacle density and may vary spatially, while friction remains constant, should not be expected to obey the fluctuation-dissipation principle \cite{hoze}. Formula 1 of \cite{Masson} postulates that the friction coefficient is an average with respect to the obstacle density, which leads to a coarse-graining approximation, incompatible with the local interpretation that the potential creates by local molecular interactions (which can extend only to few nanometers). This interpretation is thus inconsistent with the hundreds-of-nanometers size of the effective potential well described in Fig. 3a of \cite{Masson}. As described above, the effective field $\mb{b}(\mbox{\boldmath$X$})$ that accounts for local traps \cite{taflia} may have no potential at all. The relation between the local molecular energy and the size and depth of the potential well remains unclear. \\[1mm]
\textbf{Energy maps} \\[1mm]
The purpose of estimating the drift term is to clarify whether forces other than diffusion contribute to the dynamics. In this respect, although a clear large locally confining potential well was presented in \cite{Masson}, it is not clear what is the meaning of the energy map presented in Fig. 2c and f. Potential interactions are local and the force is due to the gradient of the potential, thus a large region with given constant energy indicates pure Brownian motion in this region, without interactions with any field. Is the energy actually calibrated or extracted from data?\\[1mm]
\textbf{Simulations} \\[1mm]
Another unclear point of the paper \cite{Masson} is the role, goal, and the construction of the simulations. The statement "Fokker-Planck equations can always be approximated by master equations" seems to indicate that the observed drift vector and diffusion tensor, which may be space and time-dependent, are known in the entire domain for all relevant times, which may be quite long. The Fokker-Planck equation (FPE) can be extracted from its global empirical solution, which is not contained in the empirical data. Thus the only way to extract a FPE from the data is to estimate the coefficients from the data, as described in \cite{HozePNAS}. Simulations of long trajectories of the assumed diffusion process can be run once the coefficients have been estimated in the entire domain. The approximations of the solution of the FPE proposed in \cite{Masson} is a path integral that requires entire trajectories, not only disjoint fragments. The statement "these simulated trajectories [...] characteristics match those of the experimental ones" should be clarified, because theoretically, all diffusion processes may have the same trajectories, but with different probabilities. Thus simulated trajectories do not "match" empirical trajectories. What can be matched are various statistics of the trajectories, though it is not enough to match moments, because statistics, such as first passage times, which may be rare events, cannot be extracted from short simulations and certainly not from short fragments of trajectories. { The computation of the MFPT of a given diffusion process to a boundary requires either solving the Pontryagin-Andronov-Vitt boundary value problem \cite{Schuss2} or running Brownian simulations, once the drift and diffusion tensor have been estimated. But it cannot be computed from the solution of the Fokker-Planck equation in free space, because absorbing boundary conditions are necessary. A new approach based on stochastic simulations in the entire dendrite is developed in \cite{HozeBJ2014}.}
{Finally, estimating biophysical quantities, such as the first and second moments from the path-integral approach { used in \cite{Masson} may} introduce another computational bias: {using entire trajectories instead of two consecutive points from the same trajectory to compute these moments results in an estimation procedure that smoothes out the values of the diffusion coefficient.} The diffusion coefficients and energy extracted in \cite{Masson} are significantly lower and smoother than the ones presented in \cite{HozePNAS}. It is not clear whether this is a consequence of the method of \cite{Masson} or it captures the biophysical reality of inhibitory receptor trafficking.}
|
\section{Introduction}
Lie bialgebras, having
close relations with Yang-Baxter equations \cite{2}, are important ingredients in quantum groups, which have drawn more and more attentions in literature (e.g., \cite{1,1-2,D,G,M1,3,4,5,6,7,8,9,10,12,13,14}).
Michaelis \cite{4} investigated structures of Witt type Lie bialgebras. Ng and Taft \cite{9} gave a classification of this
type Lie bialgebras, and obtained that all structures of Lie bialgebras on the one sided
Witt algebra, the Witt algebra and the Virasoro algebra are
coboundary triangular (cf.~\cite{8}).
For the cases of generalized Witt type Lie algebras and generalized Virasoro-like Lie algebras, the authors of \cite{10, 13} proved that
all structures of Lie bialgebras on them are
coboundary triangular. Similar results hold
for some other kinds of Lie algebras (cf., e.g., \cite{13, 14}).
From the examples of infinite dimensional Lie bialgebras constructed in \cite{4}, many infinite
dimensional Lie bialgebra structures we know are coboundary
triangular. It may sound that coboundary triangular Lie bialgebras
are relatively simple. However, they are not trivial in the sense that many
natural problems associated with them remain open (see, also Remark \ref{Raa22}). For
example, even for the (two-sided) Witt algebra and the Virasoro
algebra, a complete classification of coboundary triangular Lie bialgebra
structures on them is still an open problem. Nevertheless, not much on representations of infinite dimensional Lie bialgebras
is known.
From the viewpoint of
Lie bialgebras, considering dual Lie bialgebra structures may help us
understand more on infinite dimensional Lie bialgebra structures. For instance, by considering structures of dual Lie
bialgebras of Witt and Virasoro types, the authors of \cite{16}
surprisingly obtained some new series of infinite
dimensional Lie algebras.
In the present paper, we study structures of dual Lie bialgebras of Poisson
type.
One may have noticed that the dual of a finite dimensional Lie
bialgebra is naturally a Lie bialgebra, and so one would not predict
anything new in this case. However, for the cases of infinite
dimensional Lie bialgebras, the situations become quite different, which can be seen in the following contents.
Let us recall the definition of Poisson algebras here: a {\it Poisson algebra} is a triple $({\cal P},[\cdot,\cdot],\cdot)$ such that $({\cal P},[\cdot,\cdot])$ is a Lie algebra, $({\cal P},\cdot)$ is an associative algebra, and the following {\it Leibniz rule} holds:
\begin{equation}\label{Poi1}
[a,bc]=[a,b]c+b[a,c]\mbox{ \ for \ }a,b,c\in {\cal P}.
\end{equation}
In particular, for any commutative associative algebra $({\cal A},\cdot)$, and any commutative derivations $\ptl_1,\ptl_2$ of ${\cal A}$, we obtain a
{\it Poisson algebra} $({\cal A},[\cdot,\cdot],\cdot)$ with Lie bracket $[\cdot,\cdot]$ defined as follows.
\begin{equation}\label{Poi2}
[a,b]=\ptl_1(a)\ptl_2(b)-\ptl_2(a)\ptl_1(b)\mbox{ \ for \ }a,b\in {\cal A}.
\end{equation}
If we take ${\cal A}=\F[x^{\pm1},y^{\pm1}]$ (where $\F$ is an algebraically closed field of characteristic zero) and $\ptl_1\!=\!x\frac{\ptl}{\ptl x},\,\ptl_2\!=\!y\frac{\ptl}{\ptl y}$, then we obtain the {\it Virasoro-like algebra} $({\cal A},[\cdot,\cdot])$
with basis $\{x^iy^j\,|\,i,j\!\in\!\Z\}$ and Lie bracket
\begin{equation}\label{Poi3}
[x^iy^j,x^ky^\ell]=(i\ell-jk)x^{i+k}y^{j+\ell}\mbox{ \ for \ }i,j,k,\ell\in \Z.
\end{equation}
The Virasoro-like algebra \eqref{Poi3} can be generalized as follows:
For any {\it nondegenerate} additive subgroup $\Gamma$ of $\F^2$ (i.e., $\Gamma$ contains an $\F$-basis of $\F^2$), we have the group algebra ${\cal A}=\F[\Gamma]$ with basis $\{ L_{\a}\, |\, \a \in \Gamma \}$ and
multiplication defined by $\mu(L_{\a}, L_{\b}) = L_{\a+\b}$ for $\a, \b \in \Gamma$.
Then we have the (generalized) Virasoro-like algebra $(\cal{A}, \v )$ with Lie bracket $\v$ defined by
\begin{equation}\label{VL-1}
\v(L_{\a}, L_{\b}) = (\a_1\b_2- \b_1\a_2)L_{\a+\b}
\mbox{ \ for \ $\a=(\a_1, \a_2), \,\b=(\b_1, \b_2) \in \Gamma.$}
\end{equation}
Furthermore, if we take ${\cal A}=\F[x,y]$ and $\ptl_1=\frac{\ptl}{\ptl x},\,\ptl_2=\frac{\ptl}{\ptl y}$, then we obtain the {\it classical Poisson algebra} $(\F[x,y],[\cdot,\cdot],\cdot)$, whose
Lie bracket is given by
\begin{equation}\label{Poi4}
[f,g]=J(f,g)\mbox{ \ for \ }f,g\in\F[x,y],
\end{equation}
where $J(f,g):=\Big|{}^{\frac{\ptl f}{\ptl x}\ \frac{\ptl f}{\ptl y}}_{\frac{\ptl g}{\ptl x}\ \frac{\ptl g}{\ptl y}}\Big|$ is the {\it Jacobian determinant} of $f$ and $g$.
The reason we have a special interest in the classical Poisson algebra also lies in the fact that this algebra is closely related to the distinguished Jacobian conjecture (e.g., \cite{V2,Y}), which can be stated as ``any non-zero endomorphism of $(\F[x,y],[\cdot,\cdot],\cdot)$ is an isomorphism''. One observes that a Jacobi pair $(f,g)$ (i.e., $f,g\in\F[x,y]$ satisfying $J(f,g)\in\F\backslash\{0\}$) corresponds to a solution $r=f\otimes fg-fg\otimes f$ of the classical Yang-Baxter Equation (cf.~\eqref{CYBE}), thus gives rise to a Lie bialgebra structure on $\F[x,y]$.
The paper is organized as follows. Some
definitions and preliminary results are briefly recalled in Section 2. Then in Section 3, structures of dual coalgebras
of $\F[x, y]$ are addressed. Finally in Section 4, structures of
dual Lie bialgebras of Poisson type are investigated.
The main results of the present paper are summarized in Theorems \ref{theo-3-1}, \ref{theorem-2}, \ref{th-2-2}, \ref{theorem-3-1} and \ref{last-theo}.
\section{Definitions and preliminary
results}
Throughout the paper,
all vector spaces are assumed to be over an algebraically closed field $\F$ of characteristic zero. As usual,
we use $\Z_+$ to denote the set of nonnegative
integers.
We briefly recall some notions on Lie bialgebras, for details, we
refer readers to, e.g., \cite{2,10}.
\begin{defi}\adddot{\label{2.1}}\rm\begin{enumerate}\parskip-1pt\item A {\it Lie bialgebra}
is a triple $(L, [\cdot, \cdot], \d)$ such that
$(L, [\cdot, \cdot])$ is a Lie algebra,
$(L, \d)$ is a Lie coalgebra, and $\d:L\to L\otimes L$ is a derivation, namely,
$\d[x, y] = x \cdot \d(y) - y\cdot\d(x)$ for $x,y \in L$,
where $ x \cdot (y\otimes z) = [x, y]\otimes z + y
\otimes [x, z]$ for $x, y, z \in L.$
\item A Lie bialgebra
$(L, [\cdot, \cdot], \d)$ is {\it coboundary} if $\d$ is coboundary in the sense that there exists $r \in L \otimes L$ written as $r= \sum r^{[1]} \otimes r^{[2]} $, such that $\d(x) = x\cdot r$
for $x\in L$.
\item A coboundary Lie bialgebra
$(L, [\cdot, \cdot], \d)$ is {\it triangular} if $r$ satisfies the following {\it classical Yang-Baxter Equation} (CYBE),
\begin{equation}\label{CYBE}C(r) = [r_{12}, r_{13}] + [r_{12}, r_{23}] + [r_{13}, r_{23}]
=0,\end{equation}
\noindent where
$r_{12} = \sum r^{[1]} \otimes r^{[2]} \otimes 1,$
$r_{13}= \sum r^{[1]} \otimes 1 \otimes r^{[2]},$
$r_{23} = \sum r^{[1]} \otimes 1 \otimes r^{[2]}$
are elements in $ \UU(L)\!
\otimes\! \UU(L) \!\otimes\! \UU(L)$, and $\UU(L)$ is the universal enveloping algebra of $L$.\end{enumerate}
\end{defi}
Two Lie bialgebras $(\gg,[\cdot,\cdot],\d)$ and $(\gg',[\cdot,\cdot]',\d')$ are said to be {\it dually paired}
if their bialgebra structures are related via
\begin{equation}\label{Dual-paired}
\langle[f, h ]', \xi\rangle = \langle f \otimes h, \d \xi \rangle ,\ \ \ \langle \d' f, \xi \otimes
\eta \rangle = \langle f, [\xi, \eta]\rangle
\mbox{ \ for $f, h \in \gg',\ \xi, \eta \in \gg,$ }\end{equation}
where $\langle\cdot,\cdot\rangle$ is a nondegenerate
bilinear form on $\gg'\times\gg$, which is naturally extended to a nondegenerate
bilinear form on \mbox{$(\gg'\otimes\gg')\times(\gg\otimes\gg)$.}
In particular, if $\gg'=\gg$ as a vector space, then $\gg$ is called a {\it
self-dual Lie bialgebra}.
The following result whose proof is straightforward can be found in \cite{3}.
\begin{prop}\adddot
Let $(\gg, [\cdot, \cdot], \d)$ be a finite dimensional Lie
bialgebra, then so is the linear dual space $\gg ^*:={\rm Hom}_\F(\gg ,\F)$ by dualisation, namely $(\gg^*,[\cdot,\cdot]',\d')$ is the Lie bialgebra
defined by \eqref{Dual-paired} with $\gg'=\gg^*$. In particular, $\gg$ and $\gg^*$ are dually paired.
\end{prop}
Thus a finite dimensional Lie biallgebra $(\gg,[\cdot,\cdot],\d)$ is always self-dual as there exists a vector space isomorphism $\gg\to\gg^*$ which pulls back the bialgebra structure on $\gg^*$ to $\gg$ to obtain another bialgebra structure on $\gg$ to make it to be self-dual.
However, in sharp contrast to the finite dimensional case, infinite
dimensional Lie bialgebras are not self-dual in general.
For convenience, we denote by $\v$ the Lie bracket of Lie algebra $(\gg ,
[\cdot, \cdot])$, which can be regarded as a linear map $\v:\gg\otimes\gg\to\gg$. Let $\v^{\ast}:\gg^*\to(\gg\otimes\gg)^*$ be the dual of $\v$.
\begin{defi}\adddot\rm \cite{M1}
Let $(\gg , \v)$ be a Lie algebra over $\F$. A subspace $V$ of $\gg ^*$ is called a {\it good
subspace} if $\v^*(V) \subset V \otimes V.$
Denote $\Re = \{ V\, |\, V {\rm\ is \ a \ good \ subspace \ of \ }\gg ^*\}$. Then
\gg ^{\circ} = \mbox{$\su
_{V\in \Re}$} V,$
is also a good subspace
of $\gg ^*$, which is obviously the maximal good subspace
of $\gg ^*$.
\end{defi}
It is clear that
if $\gg $ is a finite dimensional Lie algebra, then
$\gg ^{\circ} = \gg ^*$.
\begin{prop}\adddot{\rm\cite{M1}}\label{PPPP} For any good subspace $V$ of $\gg ^*$, the pair $(V,
\v^{\ast})$ is a Lie coalgebra. In particular, $(\gg ^{\circ}, \v^*)$
is a Lie coalgebra.
\end{prop}
For any Lie algebra $\gg $, the dual space $\gg ^*$ has a natural right $\gg $-module
structure defined for $f\in \gg ^*$ and $ x\in \gg $ by $$(f\cdot x)(y) = f([x,
y])\mbox{ \ for \ }y\in \gg .$$ We denote $f\cdot \gg = {\rm span} \{ f\cdot x\, |\, x \in
\gg \}$, the {\it space of translates} of $f$ by elements of $\gg $.
We summarize some results of \cite{1,1-2,D, G} as follows.
\begin{prop}\adddot\label{Th111}
Let $\gg $ be a Lie algebra.
Then\begin{enumerate}\item $\gg ^{\circ}=\{f\in\gg^*\,|\,f\cdot\gg\mbox{ is finite dimensional}{\sc\,}\}.$
\item
$\gg ^{\circ}=(\v^*)^{-1}(\gg^*\otimes\gg^*)$, the preimage of $\gg^*\otimes\gg^*$ in $\gg^*$.
\end{enumerate}
\end{prop}
The notion of good subspaces of an associative algebra can be defined analogously.
In the next two sections, we shall investigate $\gg^\circ$ for some associative or
Lie algebras $\gg$.
\section{The structure of $\F[x,y]^{\circ}$
}
Let $(\cal{A}, \mu, \eta)$ be an associative $\F$-algebra with unit, where $\mu$ and $\eta$ are respectively the multiplication $\mu: \cal{A} \otimes
\cal{A} \rar \cal{A}$ and the unit $\eta: \F \rar \cal{A}$,
satisfying $$\begin{array}{ll}
\mu \circ (id \otimes \mu)= \mu \circ (\mu \otimes id)):&
\cal{A} \otimes \cal{A} \otimes \cal{A} \rar \cal{A},\\[2pt]
(\eta\otimes id)(k\otimes a) = (id \otimes \eta)(a \otimes k): &\F \otimes
\cal{A}\cong \cal{A}\otimes \F \cong \cal{A},\end{array}$$
for $k \in \F,\, a \in \cal{A}$.
Then a coassociative coalgebra is a triple $(C, \D, \e)$, which is obtained
by conversing arrows in the definition of an associative algebra.
Namely, $\D: C \rar C\otimes C$ and $\e: \F \rar C$ are respectively
comultiplication and counit of $C$, satisfying
$$\begin{array}{ll}(\D \otimes id)\circ \D = (id \otimes \D)\circ \D: &C \rar
C\otimes C \otimes C,\\[2pt]
(\e \otimes id)\circ \D = (id \otimes
\e)\circ \D: &C \rar C\otimes C \rar \F\otimes C \cong C\otimes \F
\cong C.\end{array}$$
For any vector space $\cal{A}$, there exists a natural injection $\rho:
\cal{A^{\ast}} \otimes \cal{A^{\ast}} \rar (\cal{A} \otimes
\cal{A})^{\ast}$ defined by $\rho(f, g )(a, b)= \langle f, a\rangle \langle g, b\rangle $ for
$f, g \in \cal{A^{ \ast}}$ and $ a, b, \in \cal{A}$. In case $\cal A$ is finite dimensional,
$\rho$ is an isomorphism.
If $(\cal{A}, \mu)$ is associative, the multiplication $\m
$ induces the map $\mu^{\ast}: \cal{A^{\ast}}
\rar (\cal{A} \otimes \cal{A})^{\ast}$. If $\cal{A}$ is finite
dimensional, then the isomorphism
$\rho$ insures that $(\cal{A^{\ast}}, \mu^{ \ast},
\e^{\ast})$ is a coalgebra, \vspace*{-4pt}where for simplicity, $\mu^{\ast}$ denotes the
composition of the maps: $ \cal{A^{\ast}}
\stackrel{\mu^{\ast}}{\rar}(\cal{A} \otimes \cal{A})^{\ast}
\stackrel{(\rho)^{{\rm -1}}}{\rar} \cal{A^{\ast}} \otimes
\cal{A^{\ast}}$.
Now let $(\cal{A}, \mu)$ be a commutative associative algebra.
Then $\cal{A^{\circ}} =
(\mu^{\ast})^{-1}(\cal{A^*} \otimes \cal{A^*})$ (cf.~\cite{12} and Proposition \ref{Th111}). For
$\ptl \in {\rm Der}(\cal{A})$ and $f \in \cal{A^{\circ}}$, using
$$\ptl \mu
=\mu( id \otimes \ptl + \ptl \otimes id),\ \ \ \mu^*\ptl^*(f) = (id
\otimes \ptl^* + \ptl^* \otimes id )\mu^*(f) \in \cal{A^*} \otimes
\cal{A^*},$$ we obtain $\ptl^*(\cal{A^{\circ}}) \subset
\cal{A^{\circ}}$. Thus, we observe that there are two natural approaches to produce Lie
coalgebras from some subspaces of $\cal{A^*}$. One is induced
from the associative structure of $\cal{A}$ as
follows: First we have the
cocommutative coassociative
coalgebra $(\cal{A^{\circ}}, \mu^{\circ})$ with $\mu^{\circ}:= \mu^*|_{\cal{A^{\circ}}}$. Then we obtion
the Lie coalgebra $\cal{A}^{\circ}_{\mu}:= (\cal{A^{\circ}}, \D)$ with cobracket, induced from
cocommutative coassociative coalgebra structure, defined by
\begin{equation}\label{co-b0} \D(f) = (\ptl_1^{\circ} \otimes \ptl_2^{\circ} -\ptl_2^{\circ} \otimes \ptl_1^{\circ} )
\mu^{\circ} (f) {\rm \ \ for\ }\ f \in \cal{A^{\circ}},
\end{equation}
where $\ptl_1, \ptl_2
\in {\rm Der}{\cal{A}}$ are two fixed derivations satisfying $\ptl_1\ptl_2= \ptl_2\ptl_1.$ Here and below, for any $\ptl \in {\rm Der}(\cal{A}),$
we denote $\ptl^{\circ}=\ptl^*|_{\cal{A^{\circ}}}$.
Another approach is as follows: Let $\cal{A}_{\v} = (\cal{A}, [\cdot, \cdot])$ be the Lie algebra
defined in \eqref{Poi2} (where $\v = [\cdot, \cdot]$). The Lie coalgebra induced from $\cal{A}_{\v}$ is
$\cal{A}_{\v}^{\circ}= (\cal{A}_{\v}^{\circ}, \v^{\circ})$, where the subspace
$\AA_{\v}^{\circ}$ of $\AA^*$ is determined by Proposition \ref{Th111} with
cobracket
defined by
\begin{equation}\label{cob-1} \v^{\circ}(f) = (\mu(\ptl_1\otimes \ptl_2 -
\ptl_2\otimes \ptl_1))^{\ast}(f)=(\ptl_1\otimes \ptl_2 -
\ptl_2\otimes \ptl_1)^{\ast}\mu^*(f){\rm \ \ for \ }\ f \in
\cal{A}_{\v}^{\circ}.
\end{equation}
\begin{prop}\adddot\label{p2} Let $(\cal{A}, \mu)$ be a commutative
associative algebra with unit,
and $\ptl_1, \ptl_2 \in
{\rm Der}(\cal{A})$ are commutative. Then the Lie coalgebra $\cal{A}^{\circ}_{\mu}$
is a Lie subcoalgebra of $\cal{A}_{\v}^{\circ}$.
\end{prop}
\ni{\it Proof.~}~ For $f \in \cal{A}^{\circ}_\mu$, we have
$\v^{\circ}(f)
(\ptl_1\otimes \ptl_2 - \ptl_2\otimes \ptl_1)^{\ast}\mu^*(f)
(\ptl_1\otimes \ptl_2 - \ptl_2\otimes
\ptl_1)^{\ast} \mu^{\circ}(f)
=(\ptl_1^{\circ}\otimes \ptl_2^{\circ} -
\ptl_2^{\circ}\ptl_1^{\circ})\mu^{\circ}(f)= \D(f),
$
where the last equality follows from \eqref{co-b0}. Thus, $\cal{A}_{\mu}^{\circ}$ is a Lie subcoalgebra of $\cal{A}^{\circ}_{\v}$.
\QED\vskip4pt
\begin{theo}\adddot\label{theo-3-1} Let $(\AA, \mu)$ be a commutative associative algebra, and $(\cal{A}, \v)$ the Poisson Lie algebra
defined in \eqref{Poi2}. If there exists $h \in \AA$ such that the ideal $I$ of $(\AA,\v)$ generated by $h$ has finite codimension, then $\cal{A}_{\mu}^{\circ} = \cal{A}_{\v}^{\circ}.$ In particularly, if $\AA= \F[x, y]$, then $\cal{A}_{\mu}^{\circ} = \cal{A}_{\v}^{\circ}.$
\end{theo}
\ni{\it Proof. }
Denote $\cdot$ and $\star$ the actions of
$(\cal{A}, \mu)$ and $(\cal{A}, \v)$ on $\cal{A}^*$ respectively, i.e.,
$(f\cdot a)(b) = f(\mu(a, b))$, and $(f\star a)(b)= f(\v(a, b))$
for $a, b \in \cal{A}, f \in \cal{A}^*$. From the relation $\v(a, bc) = \v(a, c)b + \v(a, b)c$ for $ a, b, c \in \AA$, we have
\begin{equation*
(f\star a)\cdot b - (f\cdot b)\star a = f\cdot \v(a, b), \ \forall\,a, b \in \AA.
\end{equation*}
If $f \in {\AA}_{\v}^{\circ}$, i.e. $f\star \AA$ is finite dimensional, then $f\cdot \v(b, \AA)$ is finite dimensional. Thus if the ideal $I$ has finite codimension, and $f\cdot I$ is finite dimensional, it follows that $f\cdot \AA$ is finite dimensional. From Proposition \ref{Th111}, we have $f\in \AA_{\mu}^{\circ}$.
\QED
\begin{rema} \rm
The difference between $\cal{A}_{\mu}^{\circ}$ and $\cal{A}_{\v}^{\circ}$ is that
${\cal{A}}_{\mu}^{\circ}$, as a Lie coalgebra, is induced from coassociative coalgebra $(\AA^{\circ}, \mu^{\circ})$,
and $\AA^{\circ}_{\mu}$ is determined by $(\mu^*)^{-1}(\AA^*\otimes \AA^*)$ as a vector subspace of $\AA^*$
$($a good subspace of the dual of $(\AA, \mu)\,)$, but $\cal{A}_{\v}^{\circ}$ is the dual of the Lie algebra $(\AA, \v)$ (determined by Proposition \ref{Th111}).
\end{rema}
\section{Dual Lie bialgebras of Poisson type
}\setcounter{section}{4}\setcounter{theo}{0}\setcounter{equation}{0}
Poisson algebras (cf.~\eqref{Poi1}) have important algebra structures, which have close relations with the Virasoro
algebra and vertex operator (super)algebras (e.g., \cite{AH}). They can be also regarded as special cases of Lie algebras of Block type.
Therefore, some attentions have been paid on them and some related Lie algebras
(e.g., \cite{CL,13,Su04,17,GGS,SXX13,SXX12,SXZ,SXY,11, X}).
In this section,
we consider the dual structures of
Poisson type Lie bialgebras.
The following result can be found in \cite{12}.
\begin{prop} \adddot\label{S} Let $A, B$ be commutative associative algebras,
regarding $A^* \otimes B^* \subset (A \otimes B)^*$, then $A^{\circ}
\otimes B^{\circ} = (A \otimes B)^{\circ}.$
\end{prop}
Recall from \cite{10} that the dual space of $\F[x]$ can be identified with the space $\F[[\vp]]$. From \cite{7, 16}, and Proposition \ref{S}, we have
\begin{prop}\adddot\label{F[x]}\begin{enumerate}
\item Let $f= \sum_{i=0}^{\infty} f_i \vp^i \in
\F[[\vp]]$ with $f_i\in\F$. Then
\begin{eqnarray*}
f \!\in\! \F[x]^{\circ}&\!\!\! \Longleftrightarrow\!\!\!& f_n \!=\! h_1f_{n-1}\! +\! h_2
f_{n-2}
\! +\! \cdots
\! +\! h_r f_{n-r}\mbox{ for some $r\in\N,\,h_i\!\in\!\F$ and all $n\!>\!r$}\\[-4pt]
&\!\!\!\Longleftrightarrow \!\!\!&f \in \Big\{\frac{g(\vp)}{h(\vp)}\,\Big|\,g(\vp),h(\vp)\in\F[\vp],\,h(0)\ne0\Big\}.
\end{eqnarray*}
\item Denote $\cal{A}= \F[x, y],$ the polynomial algebra on tow variables $x,
y$. Then
$$\cal{A}^{\circ} = \F[x, y]^{\circ} \cong \F[x]^{\circ}
\otimes \F[y]^{\circ}.$$
\end{enumerate}
\end{prop}
Let $(\gg , \v, \d)$ be a Lie bialgebra. The map
$\v^*$ induces a map $\v^{\circ}:=\v^*|_{\gg^{\circ}}:\gg ^{\circ}
\rar \gg ^{\circ} \otimes \gg ^{\circ}$, making
$(\gg^{\circ},\v^\circ)$ to be a Lie coalgebra. By
\cite[Proposition 3]{8}, the map $\d^*:\gg ^* \otimes \gg ^*
\hookrightarrow (\gg \otimes \gg )^* \stackrel{\d^*}{\rar} \gg ^*$
induces a map
$\d^{\circ}:=\d^*_{\gg ^{\circ} \otimes \gg ^{\circ}}:
\gg ^{\circ} \otimes \gg ^{\circ} \rar \gg ^{\circ}$, making
$(\gg^{\circ},\d^\circ)$ to be a Lie algebra. Thus we obtain a Lie bialgebra
$(\gg ^{\circ}, \d^{\circ}, \v^{\circ} )$, the {dual Lie
bialgebra} of $(\gg , \v, \d)$.
\def\ell{l}Now take $\AA=\F[x, y]$.
Let $\vp^i, \e^i$ be duals of $x^i, y^i \in \AA$ respectively,
namely, $\langle\vp^i\e^j,x^ky^\ell\rangle=\vp^i\e^j(x^ky^\ell)=\d_{i,k}\d_{j,\ell}$
for $i,j,k,\ell\in\Z_+$.
Any
element $u\in{\cal A}^*$ can be written as $u=\sum_{i,j}u_{i,j}\vp^i\e^j$ (possibly an infinite
sum). Let $g=\sum_{k,l}g_{k,l}x^ky^l \in{\cal A}$ (a finite sum).
Then
\begin{equation}\la u,g\ra=
u(g) = \mbox{$\sum\limits_{i,j, k,l}$}u_{i,j}g_{k,l}\langle \vp^i\e^j, x^ky^l
\rangle =\mbox{$\sum\limits_{i,j, k,l}$}u_{i,j}g_{k,l}\d_{i,k}\d_{j, l}
\mbox{ (a finite sum).}
\end{equation}
Let $\ptl_1=\frac{\partial}{\partial x},\, \ptl_2=\frac{\partial}{\partial y}$.
Then we have the Poisson Lie algebra $(\cal{A}, \v)$ defined by \eqref{Poi4}.
From Theorem \ref{theo-3-1},
it is easy to check that $\AA^{\circ}_{\v} = \AA^{\circ}_{\mu}$.
\begin{conv}\rm\label{Con1} \begin{itemize}\item[(1)]
If an undefined notation appears in an expression, we treat it zero; for instance $\vp^i\e^j=0$ if $i<0\ \mbox{or}\ j<0.$
\item[(2)]
When there is no confusion, we use $[\cdot,\cdot]$ to denote the bracket in $\gg$ or $\gg^\circ$, i.e., $[\cdot,\cdot]=\v$ or $\d^{\circ}$.
We also use $\D$ to denote the cobracket in $\gg$ or $\gg^\circ$, i.e., $\D=\d$ or
$\v^\circ$.
\end{itemize}\end{conv}
Let $m,n\in\Z_+$, and take $ a = x^my^n, \, b = xy \in \cal{A}$.
Then $[a, b ]
(m-n)a$. Thus by \cite{4}, the triple $(\cal{A}, [\cdot, \cdot],
\D_r)$ with $r = a\otimes b
- b \otimes a$ is a coboundary triangular Lie bialgebra whose cobracket is
defined by
\begin{equation}\D_r (f) = f\cdot r= [f, a] \otimes b + a \otimes [f, b] - [f, b]\otimes a - b\otimes [f, a]\ {\rm
for \ }\ f \in \cal{A}.
\end{equation}
\begin{theo}\adddot \label{theorem-2}Let
$(\AA, [\cdot, \cdot], \D_r)$ be
the coboundary triangular Lie bialgebra defined above.
The dual Lie bialgebra of $\AA$
is $(\AA^{\circ}, [\cdot, \cdot], \D)$, where $\AA^{\circ}$ is described by Proposition $\ref{F[x]}(2)$ with cobracket $\D$ defined by
\begin{equation} \label{dual-1}\D(\vp^m\e^n) =\mbox{$\sum\limits_{k+s=m+1,\ l+t=n+1}(kt-ls)$}\vp^k\e^l\otimes
\vp^s\e^t,
\end{equation}
and bracket $[\cdot,\cdot]$ uniquely determined by the skew-symmetry and the following
\begin{equation}\label{bracket-4-2}
[\vp^i\e^j,\vp^s\e^t]=\left\{
\begin{array}{ll}
(m(t\!+\!1)\!-\!n(s\!+\!1))\vp^{s+1-m}\e^{t+1-n}\!\!\!&\mbox{if \ }(i,j)=(1,1),\,(s,t) \neq (1,1),\\[4pt]
(s-t)\vp^s\e^t&\mbox{if \ }(i,j)=(m, n) \neq (s, t)\neq(1, 1), \\[4pt]
0&\mbox{otherwise}.\end{array}\right.
\end{equation}
\end{theo}
\ni{\it Proof.~}~
Assume
$\mu^{\circ}
(\vp^m\e^n) = \mbox{$\sum_{k,l,.s,t \in
\Z_+}$}c_{k,l,s,t}\vp^k\e^l\otimes \vp^s\e^t\mbox{ for some }c_{k,l,s,t}
\in \F.$
\noindent Then
\begin{eqnarray*}\begin{array}{lll}
c_{i,j,p,q} =
\mu^{\circ}(\vp^m\e^n)(x^iy^j\otimes x^py^q)
\langle\vp^m\e^n,
\mu(x^iy^j\otimes x^py^q)\rangle
\langle\vp^m\e^n,
x^{i+p}y^{j+q}\rangle
\d_{m, i+p}\d_{n, j+q}.
\end{array}
\end{eqnarray*}
Thus,
\mu^{\circ}(\vp^m\e^n) =\mbox{$\su
_{k+s=m,\,
l+t=n}$}\vp^k\e^l\otimes \vp^s\e^t.
$
Assume $\ptl_1^{\circ}(\vp^i\e^j) =
\sum_{s,t}c_{s,t}\vp^s\e^t$. Then
$$c_{k,l} =
\ptl_1^{\circ}(\vp^i\e^j)(x^ky^l)=
\vp^i\e^j(\ptl_1(x^ky^l))=k\d_{i,k-1}\d_{j,l}=(i+1)\d_{i+1, k}\d_{j, l},$$
\noindent i.e., $\ptl_1^{\circ}(\vp^i\e^j)
=(i+1)\vp^{i+1}\e^j.$ Similarly, $\ptl_2^{\circ}(\vp^i\e^j)=
(j+1)\vp^i\e^{j+1}.$ From (\ref{co-b0}), we obtain
\begin{eqnarray*}\begin{array}{lll}
\D(\vp^m\e^n) &=& (\ptl_1^{\circ} \otimes \ptl_2^{\circ}-
\ptl_2^{\circ}\otimes\ptl_1^{\circ})\mu^{\circ}(\vp^m\e^j)
\sum\limits_{k+s=m+1,\, l+t=n+1}(kt-ls)\vp^k\e^l\otimes \vp^s\e^t.
\end{array}
\end{eqnarray*}
Therefore, (\ref{dual-1}) holds. Next, we verify \eqref{bracket-4-2}.
We have
\begin{eqnarray
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\langle[\vp^i\e^j, \vp^s\e^t], x^ky^l \rangle
=\langle \vp^i\e^j\otimes \vp^s\e^t, (kn-lm)x^{k+m-1}y^{l+n-1}\otimes xy
+(k-l)x^my^n\otimes x^ky^l
\nonumber\\\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\phantom{\langle[\vp^i\e^j, \vp^s\e^t], x^ky^l \rangle=}
- (kn-lm)xy\otimes x^{k+m-1}y^{l+n-1}
-(k-l)x^ky^l\otimes x^my^n\rangle
\nonumber
\\
\label{f-3}\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\phantom{\langle[\vp^i\e^j, \vp^s\e^t], x^ky^l \rangle}=\langle(n(i+1)-m(j+1))\d_{s,1}\d_{t,1}\vp^{i+1-m}\e^{j+1-n}
+(s-t)\d_{i,m}\d_{j,n}\vp^s\e^t
\nonumber\\\!\!\!\!\!\!\!\!\!\!
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\phantom{\langle[\vp^i\e^j, \vp^s\e^t], x^ky^l \rangle=}
-(n(s\!+\!1)\!-\!m(t\!+\!1))\d_{i,1}\d_{j,1}\vp^{s+1-m}\e^{t+1-n}
\!-\!(i\!-\!j)\d_{s,m}\d_{t,n}\vp^i\e^j,x^ky^l\rangle.
\end{eqnarray}
If $(i,j)=(1, 1)$ and $ (s, t)\neq (1, 1)$, then (\ref{f-3}) gives
(note that $(m, n)\neq (1, 1)$)
$$[\vp\e, \vp^s\e^t ] = \left\{\begin{array}{lll}(m(t+1)-n(s+1))\vp^{s+1-m}\e^{t+1-n}&\mbox{if}\ s+1-m\geq 0,\ t+1-m\geq 0,\\
0&\mbox{otherwise}.\end{array}\right.$$
Thus we obtain the first case of (\ref{bracket-4-2}) (cf.~Convention \ref{Con1}\,(1)).
If $(i, j)= (m, n)\neq (s, t)\neq (1, 1)$,
then (\ref{f-3}) gives $[\vp^m\e^n, \vp^s\e^t ]=(s-t)\vp^s\e^t,$ which is the second case of (\ref{bracket-4-2}).
It remains to verify the last case of
(\ref{bracket-4-2}). By (\ref{f-3}), we have $[\vp^i\e^j, \vp^s\e^t] =0$ if $i, s \neq 1, m$
or $j, t \neq 1, n$.
We
discuss the situations in two subcases.\vskip4pt
\noindent{\it Subcase 1.} Assume $i=1$ (thus $j\ne1$). Then (\ref{f-3}) becomes
\begin{equation}\label{f-1-1}\begin{array}{lll}
[\vp\e^j, \vp^s\e^t]&\!\!\!\!=\!\!\!\!&\big(2n-m(j\!+\!1)\big)\d_{s,1}\d_{t,1}\vp^{2-m}\e^{j+1-n}+(s\!-\!t)\d_{1,m}\d_{j,n}\vp^s\e^t
-(1\!-\!j)\d_{s,m}\d_{t,n}\vp\e^j.
\end{array}
\end{equation}
Note that
$\vp^{2-m}=0$ if $m>2$,
in this case, (\ref{f-1-1}) becomes
$[\vp\e^j, \vp^s\e^t]
-(1-j)\d_{s,m}\d_{t,n}\vp\e^j$,
and we have (\ref{bracket-4-2}).
Now assume $m=0$. Then (\ref{f-1-1}) gives
$[\vp\e^j, \vp^s\e^t]=2n\d_{s,1}\d_{t,1}\vp^{2}\e^{j+1-n}
-(1-j)\d_{s,0}\d_{t,n}\vp\e^j,$
and we see that
the last case of (\ref{bracket-4-2}) holds in this case.
Next assume $ m=1$. Then (\ref{f-1-1}) becomes
[\vp\e^j, \vp^s\e^t]=(2n-(j+1))\d_{s,1}\d_{t,1}\vp\e^{j+1-n}+(s-t)\d_{j,n}\vp^s\e^t
-(1-j)\d_{s,1}\d_{t,n}\vp\e^j.
$
Hence the last case of (\ref{bracket-4-2}) holds.
Finally assume $m=2$. By (\ref{f-1-1}), we have
\begin{equation}\label{f-1-3}\begin{array}{lll}
[\vp\e^j, \vp^s\e^t]&=&(2n-2(j+1))\d_{s,1}\d_{t,1}\e^{j+1-n}
-(1-j)\d_{s,2}\d_{t,n}\vp\e^j,\end{array}
\end{equation}
and the last case of (\ref{bracket-4-2}) holds again.
\vskip4pt
\noindent{\it Subcase 2.} Assume $i=m\neq 1 \neq s. $ We have
$[\vp^m\e^j, \vp^s\e^t]=(s-t)\d_{j,n}\vp^s\e^t-(m-j)\d_{s,m}\d_{t,n}\vp^m\e^j$ by (\ref{f-3}), i.e.,
$[\vp^m\e^j, \vp^s\e^t]=(s-t)\vp^s\e^t$ if $j=n$, or
$(j-m)\vp^m\e^j$ if $(s, t)=(m, n)$, or
$0$ otherwise.
This completes the proof of the theorem.
\QED
\begin{prop}\label{b-6-0} Let $f(x,y)= \su
_{i=0}^{m}\su
_{j=0}^n a_{i,j}x^iy^j \in \F[x,y]$ with $a_{m,n} \neq 0$, and
$k,l\in\Z_+,$ $c\in\F$.
Denote ${\rm Supp\,}f=\{(i,j)\in\Z_+^2\,|\,a_{ij}\ne0\}$.
Then
\begin{eqnarray}\label{eq111}
{}\!\!\!\![x^ky^l, f(x, y)]= cf(x, y)\neq 0&\Longleftrightarrow&k=l=1,\ \ j-i=c,\ \forall\,(i,j)\in{\rm Supp\,}f,\\
\label{eq111+}
{}[x^ky^l, f(x, y)] =0\ \ \ \ \ \ \ \ \ \ \ \ \ &\Longleftrightarrow&kj-li=0, \ \forall\,(i,j)\in{\rm Supp\,}f.
\end{eqnarray}
In particular, if denote $r=A\otimes B- B\otimes A,$ where either $A=xy,$ $B= \su
_{i=0}^n a_ix^iy^{c+i}$ for some $c \in \Z_+$ and $a_i\in\F$, or $ A=x^ky^l$, $B=f(x, y)=\su
_{i=0}^{m}\su
_{j=0}^n a_{i,j}x^iy^j$ such that $kj-li=0$ for all $(i, j)\in{\rm Supp\,}f $, then $r$ is a solution
of classical Yang-Baxter equation.
\end{prop}
\ni{\it Proof.~}~ We have
$[x^ky^l, f(x, y)
\!=\!\sum_{i,j}a_{i,j}(kj-li)x^{k+i-1}y^{l+j-1}.$
By comparing the coefficients of the highest term (i.e., $x^my^n$) and $x^iy^j$ for all $i,j$, one immediately obtains \eqref{eq111} and \eqref{eq111+}.
\QED\vskip4pt
First consider $A=xy,\, B=\su
_{i=0}^na_ix^iy^{i+m}$ with $a_n \neq 0$ and $m\in\Z_+\backslash\{0\}$.
Then $[A, B]= mB\neq0$.
The triple $(\cal{A}, [\cdot, \cdot], \D_r)$ with $r= A\otimes B - B\otimes A$ is a coboundary triangular Lie bialgebra of Poisson type with bracket defined by
(\ref{Poi4}) and cobracket
defined by $$\D_r (g)
= [g, A] \otimes B + A \otimes [g, B] - B \otimes [g, A] - [g, B] \otimes A,\ \forall\, g\in \AA.$$ The following is one of the main results of the present paper.
\begin{theo}\label{th-2-2} Let $(\cal{A}, [\cdot, \cdot], \D_r)$ be the coboundary triangular Lie bialgebra defined as above.
The dual Lie bialgebra of $(\cal{A}, [\cdot, \cdot], \D_r)$ is $(\cal{A}^{\circ}, [\cdot, \cdot], \D)$, where $\AA^{\circ}$ is described by Proposition $\ref{F[x]}(2),$
with cobracket $\D$ defined by \eqref{dual-1} and bracket uniquely determined by the skew-symmetry and the following.
\begin{itemize}\item[\rm(1)]
In case $m\neq 0$, we have $($cf.~Convention $\ref{Con1}\,(1)$\,$)$, for $(s,t)\ne(1,1)$ and $p\ne q$,
\begin{eqnarray}
\label{eq2.2-1}
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
[\vp^p\e^p, \vp^s\e^t]=\d_{p,1} \mbox{$\sum\limits_{i=0}^n$} a_i(si\!+\!sm\!+\!m\!-\!ti)\vp^{s-i+1} \e^{t-i-m+1}
,\\
\label{eq2.2-2}
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
[\vp^p\e^q, \vp^s\e^t]=(p\!-\!q)\mbox{$\sum\limits_{i=0}^n$} a_i\d_{s,i}\d_{t, i+m}\vp^p\e^q -(s\!-\!t)\mbox{$\sum\limits_{i=0}^n$} a_i\d_{p,i}\d_{q, i+m}\vp^s\e^t
.
\end{eqnarray}
\item[\rm(2)] In case $m=0$, we have, for $(s,t)\ne(1,1)$,
\begin{eqnarray}
\label{+eq2.2-1}
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
[\vp\e, \vp^s\e^t] =\mbox{$\sum\limits_{i=2}^n$}a_i (s-t)i\vp^{s-i+1} \e^{t-i+1}
,\\
\label{+eq2.2-2}
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
[\vp^p\e^p, \vp^s\e^t] =(t-s)a_p\vp^s\e^t \mbox{ \ if \ }\ p\in \{0, 2, 3, \cdots, n\},\\
\label{+eq2.2-3}
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
[\vp^p\e^q, \vp^s\e^t] = 0 \mbox{ \ if \ }\ p\neq q,\, s\neq t.
\end{eqnarray}
\end{itemize}
\end{theo}
\ni{\it Proof.~}~ Denote $C\!=\![x^ky^l, xy]\!=\!(k\!-\!l)x^ky^l$ and
$D\! =\![x^ky^l, B
\!=\!\mbox{$ \su
_{i=0}^n$}a_i(ki\!+\!km\!-\!li)x^{i+k-1}y^{i+l+m-1}.$
We have
\begin{eqnarray}\label{theorem-2-3}
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\langle[\vp^p\e^q, \vp^s\e^t], x^ky^l\rangle
=\langle\vp^p\e^q\otimes \vp^s\e^t, C \otimes B + A\otimes D- B\otimes C - D \otimes A\rangle
=P^{p,q}_{s,t} - P^{s,t}_{p,q}
,\end{eqnarray}
where (regarding $k,l$ as fixed)
$$\begin{array}{lll} P^{p,q}_{s,t}&\!\!\!=\!\!\!&(k-l)\d_{p,k}\d_{q,l}\sum\limits_{i=0}^n a_i\d_{s,i}\d_{t, i+m}
+ \d_{p,1}\d_{q,1}\sum\limits_{i=0}^n a_i(ki+km-li)\d_{s, i+k-1}\d_{t, i+m+l-1}\\[12pt]
&\!\!\!=\!\!\!& (p-q)\d_{p,k}\d_{q,l}\sum\limits_{i=0}^n a_i\d_{s,i}\d_{t, i+m} +
\d_{p,1}\d_{q,1}\sum\limits_{i=0}^na_i(si+sm+m-ti)\d_{s-i+1, k}\d_{t-i-m+1,l}
\\[12pt]
&\!\!\!=\!\!\!&\langle H^{p,q}_{s,t}, x^ky^l\rangle,\mbox{ \ and where,}\end{array}$$
\begin{equation}\label{theorem-2-4}H^{p,q}_{s,t} = (p-q)\mbox{$\sum\limits_{i=0}^n$} a_i\d_{s,i}\d_{t, i+m}\vp^p\e^q + \d_{p,1}\d_{q,1}
\mbox{$\sum\limits_{i=0}^n$}a_i (si+sm+m-ti)\vp^{s-i+1} \e^{t-i-m+1}.\end{equation}
Thus
\begin{equation}\label{theorem-2-6}
[\vp^p\e^q, \vp^s\e^t] = H^{p,q}_{s,t} - H^{s,t}_{p,q}.
\end{equation}Assume $(s,t)\ne(1,1)$.
First suppose $m\ne0$.
Then \eqref{theorem-2-3}--\eqref{theorem-2-6} give $[\vp^p\e^p, \vp^s\e^t] =0$ for $p=q\ne1$, and
$[\vp\e, \vp^s\e^t]=\sum_{i=0}^n a_i(si+sm+m-ti)\vp^{s-i+1} \e^{t-i-m+1}
$ for $p=q=1$.
We have \eqref{eq2.2-1}.
If $p\neq q$, we have
$
[\vp^p\e^q, \vp^s\e^t]
(p-q)\su
_{i=0}^n a_i\d_{s,i}\d_{t, i+m}\vp^p\e^q -(s-t)\su
_{i=0}^n a_i\d_{p,i}\d_{q, i+m}\vp^s\e^t
,$
by \eqref{theorem-2-3}--\eqref{theorem-2-6},
and we have \eqref{eq2.2-2}.
Now suppose $m=0$.
Then
\eqref{theorem-2-4} becomes
\begin{equation}\label{theorem-2-7}H^{p,q}_{s,t} = (p-q)\mbox{$\sum\limits_{i=0}^n$} a_i\d_{s,i}\d_{t, i}\vp^p\e^q +
\d_{p,1}\d_{q,1}\mbox{$\sum\limits_{i=0}^n$}a_i (s-t)i\vp^{s-i+1} \e^{t-i+1}.\end{equation}
We have
$[\vp\e, \vp^s\e^t] =\su
_{i=2}^na_i (s-t)i\vp^{s-i+1} \e^{t-i+1}$ for $(p,q)=(1, 1)$,
i.e., we have \eqref{+eq2.2-1}.
If $p=q\in\{0,2,3,...,n\}$,
it is easy to see from \eqref{theorem-2-7} and \eqref{theorem-2-6} that we have \eqref{+eq2.2-2}.
Finally assume $p\neq q$. One can easily obtain \eqref{+eq2.2-3}.
\QED
\begin{rema}\rm\label{Raa22} Theorems \ref{theorem-2} and \ref{th-2-2} and the following theorem provide us some examples of
the nontriviality of coboundary triangular Lie bialgebras, even in the case of very trivial solution $r=A\otimes B-B\otimes A$ of CYBE
with $[A, B]=0$.
\end{rema}
\begin{theo}\label{theorem-3-1} Let $(k,l)\in\Z_+^2$ be fixed and $A=x^ky^l, B=f(x, y)=\sum_{(i, j) \in S}a_{i, j}x^iy^j \in \AA$ with $a_{i,j}\in\F$, where
$S={\rm Supp\,}f$ is some subset of $\Z_+^2$ such that $kj-li=0$ for $(i,j)\in S$. Denote $r=A\otimes B- B\otimes A$.
Then $(\cal{A}, [\cdot, \cdot], \D_r)$ is a coboundary triangular Lie bialgebra
of Poisson type. The dual Lie bialgebra of $(\cal{A}, [\cdot, \cdot], \D_r)$ is $(\cal{A^{\circ}}, [\cdot, \cdot], \D)$ with cobracket $\D$ defined as in Theorem $\ref{theorem-2}$ and bracket uniquely determined by the skew-symmetry and the following.\vskip4pt
$(1)$ If $(s, t) \neq (k, l)$, then
\begin{eqnarray}\label{caseqq1}
[\vp^k\e^l, \vp^s\e^t]=
\left\{\begin{array}{ll}\!\!
(l\!-\!k)a_{s, t}\vp \e\!+\!\sum\limits_{(i,j)\in S}a_{i, j}(j\!-\!i)\vp^{s-i+1}\e^{t-j+1}
\!-\!(l\!-\!k)a_{k,l}\vp^{s-k+1}\e^{t-l+1
,\\[12pt]\!\!\!
\sum\limits_{(i,j)\in S}\!\!a_{i, j}(j(s\!+\!1)\!-\!i(t\!+\!1))\vp^{s-i+1}\e^{t-j+1}\!-\!(l(s\!+\!1)\!-\!k(t\!+\!1))a_{k,l}\vp^{s-k+1}\e^{t-l+1}
,\\[12pt]
(l\!-\!k)a_{s, t}\vp \e \!+\!\sum\limits_{(i,j)\in S}a_{i, j}(j\!-\!i)\vp^{s-i+1}\e^{t-j+1
,\\[12pt]
\sum\limits_{(i,j)\in S}a_{i, j}(j(s+1)-i(t+1))\vp^{s-i+1}\e^{t-j+1
,\end{array}\right.
\end{eqnarray
according to the following four cases
$$\mbox{{\rm(i)} $(s, t), (k, l) \!\in\! S$, {\rm(ii)} $(s, t) \!\notin \!S,(k,l)\! \in\! S$,
{\rm(iii)} $(s, t)\!\in\! S, (k,l)\! \notin\! S$ or {\rm (iv)} $(s, t),(k, l)\notin S$.}$
$(2)$ If $(p, q)\neq (k, l),\, (s, t) \neq (k, l)$, then
\begin{eqnarray}\label{caseqq2}
[\vp^p\e^q, \vp^s\e^t ]=\left\{\begin{array}{ll}
(l\!-\!k)a_{s, t}\vp^{p-k+1}\e^{q-l+1}-(l\!-\!k)a_{p, q}\vp^{s-k+1}\e^{t-l+1}\!\!\!&\mbox{if}\ (p, q), (s, t) \in S,\\[4pt]
(l(p\!+\!1)\!-\!k(q\!+\!1))a_{s, t}\vp^{p-k+1}\e^{q-l+1}& \mbox{if}\ (p, q)\!\notin\! S, (s, t)\!\in\! S,\\[4pt]
0&\mbox{if}\ (p, q)\notin S, (s, t) \notin S.\end{array}\right.\end{eqnarray}
\end{theo}
\ni{\it Proof.~}~ Since $[A, B] =0,$ the triple $(\cal{A}, [\cdot, \cdot], \D_r)$ is obviously a coboundary triangular Lie bialgebra. We only need to determine the bracket relations.
Note that $\langle [\vp^p\e^q, \vp^s\e^t], x^{m^{\prime}}y^{n^{\prime}} \rangle$ equals
\begin{eqnarray}\label{theorem-3-4}
&\!\!\!\!\!\!\!\!\!\!\!\!&
\phantom{X=}\Big\langle \vp^p\e^q\otimes\vp^s\e^t, [x^{m^{\prime}}y^{n^{\prime}}, A] \otimes B + A\otimes [x^{m^{\prime}}y^{n^{\prime}}, B]
-B\otimes[x^{m^{\prime}}y^{n^{\prime}}, A] -[x^{m^{\prime}}y^{n^{\prime}}, B]\otimes A \Big\rangle
\nonumber\\&\!\!\!\!\!\!\!\!\!\!\!\!&
\phantom{X}
=
(m^{\prime}l-n^{\prime}k)\d_{p,{m^{\prime}+k-1}}\d_{q,{n^{\prime}+l-1}}\mbox{$\sum\limits_{(i, j)\in S}$}a_{i, j}\d_{s,i}\d_{t,j}
+\d_{p,k}\d_{q, l}\mbox{$\sum\limits_{(i,j)\in S}$}a_{i, j}(m^{\prime}j - n^{\prime} i)\d_{s,{i+m^{\prime}-1}}\d_{t, {j+n^{\prime}-1}}
\nonumber\\&\!\!\!\!\!\!\!\!\!\!\!\!&
\phantom{X=}
-(m^{\prime}l\!-\!n^{\prime}k)\d_{s, {m^{\prime}+k-1}}\d_{t,{n^{\prime}+l-1}}\mbox{$\sum\limits_{(i, j)\in S}$}a_{i, j}\d_{p,i}\d_{q,j}
\!-\!\d_{s,k}\d_{t, l}\mbox{$\sum\limits_{(i, j)\in S}$}a_{i, j}(m^{\prime}j\! -\! n^{\prime} i)\d_{p,{i+m^{\prime}-1}}\d_{q, {j+n^{\prime}-1}}
\nonumber\\&\!\!\!\!\!\!\!\!\!\!\!\!&
\phantom{X}
=\langle H^{p, q}_{s, t}- H^{s, t}_{p, q}, x^{m^{\prime}}y^{n^{\prime}}\rangle ,
\end{eqnarray}
where (noting that $a_{s,t}=0$ if $(s,t)\notin S$)
\begin{equation*
\begin{array}{lll}H^{p, q}_{s, t}\!=\!\big(l(p\!+\!1)\!-\!k(q\!+\!1)\big)a_{s,t}\vp^{p-k+1}\e^{q-l+1
\!+\!\d_{p,k}\d_{q, l}\sum\limits_{(i,j)\in S}a_{i, j}\big(j(s\!+\!1)\!-\!i(t\!+\!1)\big)\vp^{s-i+1}\e^{t-j+1}.\end{array}\end{equation*}
If $(p, q)= (k, l)$ and $(s, t) \neq (k, l)$, then
\eqref{theorem-3-4} gives (noting that $a_{k,l}=0$ if $(k,l)\notin S$)
$$\begin{array}{lll}[\vp^k\e^l, \vp^s\e^t]& =&(l-k)a_{s,t}\vp\e
+\sum\limits_{(i,j)\in S}a_{i, j}\big(j(s+1)-i(t+1)\big)\vp^{s-i+1}\e^{t-j+1}\\[12pt]
&&- \big(l(s+1)-k(t+1)\big)a_{k,l}\vp^{s-k+1}\e^{t-l+1
.\end{array}$$
In particular, if $(s, t), (k, l) \in S$, then (using the fact that $kj-li=0$ for $(i,j)\in S$)
$$[\vp^k\e^l, \vp^s\e^t]=(l-k)a_{s, t}\vp \e +\mbox{$\sum\limits_{(i,j)\in S}$}a_{i, j}(j-i)\vp^{s-i+1}\e^{t-j+1}
-(l-k)a_{k,l}\vp^{s-k+1}\e^{t-l+1},$$
which gives the first case of \eqref{caseqq1}.
If $(s, t) \notin S, (k,l) \in S$, then
$$[\vp^k\e^l, \vp^s\e^t]=\mbox{$\sum\limits_{(i,j)\in S}$}a_{i, j}\big(j(s\!+\!1)-i(t\!+\!1)\big)\vp^{s-i+1}\e^{t-j+1}-\big(l(s\!+\!1)-k(t\!+\!1)\big)a_{k,l}\vp^{s-k+1}\e^{t-l+1},$$
which gives the second case of \eqref{caseqq1}.
If $(s, t)\in S, (k,l) \notin S$, then
$[\vp^k\e^l, \vp^s\e^t]=(l-k)a_{s, t}\vp \e +\mbox{$\su
_{(i,j)\in S}$}a_{i, j}(j-i)\vp^{s-i+1}\e^{t-j+1},$
which gives the third case of \eqref{caseqq1}.
If $(s, t) \notin S, (k, l)\notin S$, then
$[\vp^k\e^l, \vp^s\e^t]=\mbox{$\su
_{(i,j)\in S}$}a_{i, j}(j(s+1)-i(t+1))\vp^{s-i+1}\e^{t-j+1},$
which completes the proof of \eqref{caseqq1}.
Now assume $(p, q)\neq (k, l), (s, t) \neq (k, l)$. Then
\eqref{theorem-3-4} gives
$$\begin{array}{lll}[\vp^p\e^q, \vp^s\e^t ]
\big(l(p+1)-k(q+1)\big)a_{s,t}\vp^{p-k+1}\e^{q-l+1
-\big(l(s+1)-k(t+1)\big)a_{p,q}\vp^{s-k+1}\e^{t-l+1
.\end{array}$$
If $(p, q), (s, t) \!\in\! S$, then
$[\vp^p\e^q, \vp^s\e^t ]\!=\!(l\!-\!k)a_{s, t}\vp^{p-k+1}\e^{q-l+1}\!-\!(l\!-\!k)a_{p, q}\vp^{s-k+1}\e^{t-l+1},$
giving the first case of \eqref{caseqq2}. If
$(p, q)\!\notin\! S, (s, t)\!\in\! S$, then
$[\vp^p\e^q, \vp^s\e^t ]\!=\!\big(l(p\!+\!1)\!-\!k(q\!+\!1)\big)a_{s, t}\vp^{p-k+1}\e^{q-l+1},$
giving the second case of \eqref{caseqq2}. In case
$(p, q)\notin S, (s, t) \notin S$, we have
$[\vp^p\e^q, \vp^s\e^t] =0$, which completes the proof of the theorem.
\QED\vskip4pt
In the final part of the paper, we will present an example of a dual Lie bialgebra which has different feature from the previous dual Lie bialgebras (Theorems \ref{theorem-2}, \ref{th-2-2} and \ref{theorem-3-1}).
Denote $A=x(1+y)(2+y),$ $B=x^2(1+y)^3(2+y) \in \AA$. Then \eqref{Poi4} shows $[A, B] =B$. Take
$r= A\otimes B- B\otimes A$. We obtain a coboundary triangular Lie bialgebra
$(\AA, [\cdot, \cdot], \D)$ with bracket defined by (\ref{Poi4}), and cobracket defined by
$$\D(f) = f\cdot r=[f,A]\otimes B+A\otimes [f,B]-[f,B]\otimes A-B\otimes [f,A].$$
\begin{theo}\adddot \label{last-theo}Let $(\AA, [\cdot, \cdot], \D)$ be the Lie bialgebra defined above. The dual Lie bialgebra
of $(\AA, [\cdot, \cdot], \D)$ is $(\AA^{\circ}, [\cdot, \cdot], \D)$, where $\AA^{\circ}$ and $\D$ are
defined in Theorem $\ref{theorem-2}$, and the bracket is uniquely determined by the skew-symmetry and the following.
First, denote
$$\begin{array}{ll} c_0=2,\ c_1=3, \ c_2=1, \mbox{ \ and \ } c_j=0\mbox{ for }j\neq 0,1,2,\\[2pt]
k_0=2,\ k_1=7,\ k_2=9,\ k_3=5,\ k_4=1, \mbox{ \ and \ } k_t=0 \mbox{ for }\ 4<t \in \Z_+.\end{array}
$$
Then
\begin{eqnarray}
\label{b-5-1}&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\begin{array}{lll} [\vp\e^j, \vp^s\e^t]&\!\!\!=\!\!\!&c_j\Big((4s\!-\!2t\!+\!2)\vp^{s-1}\e^{t-3}\!+\!(15s\!-\!10t\!+\!5)\vp^{s-1}\e^{t-2}\\[4pt] &&
+18(s\!-\!t)\vp^{s-1}\e^{t-1}
\!+\!(7s\!-\!14t\!-\!7)\vp^{s-1}\e^
\!+\!4(t\!+\!1)\vp^{s-1}\e^{t+1}\Big)\mbox{ if }s\ne1,2,
\end{array}
\end{eqnarray}
\begin{eqnarray}
\label{b-5-2}&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\begin{array}{lll}
[\vp\e^j, \vp\e^t] &\!\!\!=\!\!\!&c_j\Big((6\!-\!2t)\e^{t-3}
\!+\!(20\!-\!10t)\e^{t-2}\! +\!18(1\!-\!t)\e^{t-1
\!-\!14t\e^t \!+\!4(t\!+\!1)\e^{t+1}\Big)\\
&&-c_t\Big((6\!-\!2j)\e^{j-3}\!+\!(20\!-\!10j)\e^{j-2}
\!+\!18(1\!-\!j)\e^{j-1}\!-\!14j\e^j\! +\!4(j\!+\!1)\e^{j+1}\Big),
\end{array}\\[-5pt]
\label{b-5-3}
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\begin{array}{lll}
[\vp\e^j, \vp^2\e^t] &\!\!\!=\!\!\!&k_t\Big((3-j)\vp\e^{j-1}+3(1-j)\vp\e^j-2(j+1)\vp\e^{j+1}\Big)\\
&&+ c_j\Big((10\!-\!2t)\vp\e^{t-3}
\\
&&\phantom{XX} \!+\!(35\!-\!10t)\vp\e^{t-2} \!+\!18(2\!-\!t)\vp\e^{t-1}\!+\!(7\!-\!14t)\vp\e^t \!+\!4(t\!+\!1)\vp\e^{t+1}\Big),
\end{array}
\\[-5pt
\label{b-5-4}&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\begin{array}{lll}
[\vp^2\e^j, \vp^s\e^t] = -k_j\Big((2s-t+1)\vp^s\e^{t-1}+3(s-t)\vp^s\e^t-2(t+1)\vp^s\e^{t+1}\Big)\mbox{ if }s\neq 1,2,
\end{array}
\\[-5pt]
\label{b-5-5}
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\begin{array}{lll}
[\vp^2\e^j, \vp^2\e^t]&\!\!\!=\!\!\!&k_t\Big((5-j)\vp^2\e^{j-1}+3(2-j)\vp^2\e^j-2(j+1)\vp^2\e^{j+1}\Big) \\
&&-k_j\Big((5-t)\vp^2\e^{t-1}+3(2-t)\vp^2\e^t-2(t+1)\vp^2\e^{t+1}\Big),\end{array}
\\[-5pt]
\label{b-5-6}
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\begin{array}{lll}
[\vp^i\e^j, \vp^s\e^t ]= 0 \mbox{ \ if \ }\ i\neq 1,2, s\neq 1,2.
\end{array}
\end{eqnarray}
\end{theo}
\ni{\it Proof.~}~ Denote $C= [x^my^n, A], D=[x^my^n, B]$. By (\ref{Poi4}), we \vspace*{-5pt}have
\begin{eqnarray*}
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
C= [x^my^n, x(y+1)(y+2)]= (2m-n)x^my^{n+1} + 3(m-n)x^my^n -2nx^my^{n-1},\\
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
D=[x^my^n, x^2(y+1)^3(y+2)]
\\[-2pt]
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\phantom{D}
=(4m-2n)x^{m+1}y^{n+3} + (15m-10n)x^{m+1}y^{n+2}
+18(m-n)x^{m+1}y^{n+1}
\\[-2pt]
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!&
\phantom{D=}
+(7m-14n)x^{m+1}y^n -4nx^{m+1}y^{n-1}.\end{eqnarray*}
For $i, j, s, t \in \Z_+$, note that
$\langle [\vp^i\e^j, \vp^s\e^t], x^my^n\rangle$ is equal to
\begin{eqnarray}
&\!\!\!\!\!\!\!\!\!\!\!\!&
\phantom{XX=}
\langle \vp^i\e^j \otimes \vp^s\e^t, [x^my^n, A]\otimes B + A\otimes [x^my^n, B]
-[x^my^n, B] \otimes A -B \otimes [x^my^n, A] \rangle
\nonumber\\
&\!\!\!\!\!\!\!\!\!\!\!\!&
\phantom{XX}
=\langle \vp^i\e^j \otimes \vp^s\e^t, C\otimes B + A\otimes D- D\otimes A-B\otimes C \rangle
=
P^{i,j}_{s,t} - P^{s,t}_{i,j},
\end{eqnarray}
where
$ P^{i,j}_{s,t}= \langle \vp^i\e^j, C \rangle \langle \vp^s\e^t, B \rangle + \langle \vp^i\e^j, A \rangle \langle \vp^s\e^t, D \rangle$, which is equal \vspace*{-5pt}to
$$\begin{array}{lll}
&\!\!\!\phantom{=}\!\!\!& \Big((2m\!-\!n)\d_{i,m}\d_{j, n+1}\!+\!3(m\!-\!n)\d_{i,m}\d_{j,n}\!-\!2n\d_{i,m}\d_{j,n-1}\Big)\\[6pt]
&\!\!\!\!\!\!&\phantom{\Big(}\times\Big(2\d_{s,2}\d_{t,0}+7\d_{s,2}\d_{t,1}+9\d_{s,2}\d_{t,2}+5\d_{s,2}\d_{t,3}+\d_{s,2}\d_{t,4}\Big)
\\[4pt]
&\!\!\!\!\!\!&+\Big(2\d_{i,1}\d_{j,0} + 3\d_{i,1}\d_{j,1}+\d_{i,1}\d_{j,2}\Big)\\[4pt]
&\!\!\!\!\!\!&\phantom{\Big(}\times\Big((4m\!-\!2n)\d_{s,m+1}\d_{t,n+3}\!+\! (15m\!-\!10n)\d_{s,m+1}\d_{t,n+2} \!+\!18(m\!-\!n)\d_{s,m+1}\d_{t,n+1}\\[4pt]
&\!\!\!\!\!\!&\phantom{\Big(\times\Big(}+ (7m-14n)\d_{s, m+1}\d_{t,n}-4n\d_{s,m+1}\d_{t,n-1}\Big)
\\[6pt]
&\!\!\!=\!\!\!&\Big\langle \Big(2\d_{s,2}\d_{t,0}+7\d_{s,2}\d_{t,1}+9\d_{s,2}\d_{t,2}+5\d_{s,2}\d_{t,3}+\d_{s,2}\d_{t,4}\Big)\\[4pt]
&\!\!\!\!\!\!&\phantom{\Big\langle+}\times\Big((2i-j+1)\vp^i\e^{j-1}+3(i-j)\vp^i\e^j-2(j+1)\vp^i\e^{j+1}) \\[4pt]
&\!\!\!\!\!\!&\phantom{\Big\langle}+\Big(2\d_{i,1}\d_{j,0} + 3\d_{i,1}\d_{j,1}+\d_{i,1}\d_{j,2}\Big)\\[4pt]
&\!\!\!\!\!\!&\phantom{\Big\langle+}\times\Big((4s-2t+2)\vp^{s-1}\e^{t-3}
+(15s-10t+5)\vp^{s-1}\e^{t-2} +18(s-t)\vp^{s-1}\e^{t-1}\\[4pt]
&\!\!\!\!\!\!&\phantom{\Big\langle+\times}+(7s-14t-7)\vp^{s-1}\e^t +4(t+1)\vp^{s-1}\e^{t+1}\Big),x^my^n\Big\rangle.\end{array}$$
We obtain
\begin{equation}\label{H-12}[\vp^i\e^j, \vp^s\e^t ]= H^{i,j}_{s,t} - H^{s, t}_{i, j},\end{equation}
where $H^{i,j}_{s,t}$ is equal to \begin{eqnarray*
&\!\!\!\!\!\!&
\Big(2\d_{s,2}\d_{t,0}\!+\!7\d_{s,2}\d_{t,1}\!+\!9\d_{s,2}\d_{t,2}\!+\!5\d_{s,2}\d_{t,3}\!+\!\d_{s,2}\d_{t,4}\Big)
\Big((2i\!-\!j\!+\!1)\vp^i\e^{j-1}\!+\!3(i\!-\!j)\vp^i\e^j\!-\!2(j\!+\!1)\vp^i\e^{j+1}\Big) \nonumber\\[-4pt]
&\!\!\!\!\!\!&+\Big(2\d_{i,1}\d_{j,0} + 3\d_{i,1}\d_{j,1}+\d_{i,1}\d_{j,2}\Big)\nonumber\\[-4pt]
&\!\!\!\!\!\!&\phantom{\Big(}\times\Big((4s\!-\!2t\!+\!2)\vp^{s-1}\e^{t-3}\!+\!(15s\!-\!10t\!+\!5)\vp^{s-1}\e^{t-2} \!+\!18(s\!-\!t)\vp^{s-1}\e^{t-1}
\nonumber\\[-4pt]
&\!\!\!\!\!\!&\phantom{\Big(\times\Big(}+(7s-14t-7)\vp^{s-1}\e^t +4(t+1)\vp^{s-1}\e^{t+1}\Big).\nonumber\end{eqnarray*}
If $i \!\neq\! 1,2$, $s\!\neq\! 1,2$,
then
\eqref{H-12} gives that $[\vp^i\e^j, \vp^s\e^t ]=0$, which is \eqref{b-5-6}.
Assume $i=1,\, s\neq 1,2$. Then
$[\vp\e^j, \vp^s\e^t] = H^{1, j}_{s,t}$, which gives
\eqref{b-5-1}.
If $i=1, s=1,$ then
$[\vp\e^j, \vp\e^t] $ is equal to
$$\begin{array}{lll}\Big(2\d_{j,0} + 3\d_{j,1}+\d_{j,2}\Big)\Big((6-2t)\e^{t-3}
+(20-10t)\e^{t-2}
+18(1-t)\e^{t-1}-14t\e^t +4(t+1)\e^{t+1}\Big)\\[4pt]
-\Big(2\d_{t,0} \!+\! 3\d_{t,1}\!+\!\d_{t,2}\Big)\Big((6\!-\!2j)\e^{j-3}\!+\!(20\!-\!10j)\e^{j-2}
\!+\!18(1\!-\!j)\e^{j-1}\!-\!14j\e^j \!+\!4(j\!+\!1)\e^{j+1}\Big),\end{array}$$
which implies (\ref{b-5-2}).
Assume $i=1, s=2$. Then
(\ref{H-12}) implies that $[\vp\e^j, \vp^2\e^t]$ is equal to
$$\begin{array}{lll}\Big(2\d_{t,0}+7\d_{t,1}+9\d_{t,2}
+5\d_{t,3}+\d_{t,4}\Big)\Big((3-j)\vp\e^{j-1}+3(1-j)\vp\e^j-2(j+1)\vp\e^{j+1}\Big) \\[4pt]
+\Big(2\d_{j,0}\! +\! 3\d_{j,1}\!+\!\d_{j,2}\Big)\Big((10\!-\!2t)\vp\e^{t-3}\!+\!(35\!-\!10t)\vp\e^{t-2}\! +\!18(2\!-\!t)\vp\e^{t-1}\\[4pt]
\phantom{+\Big(2\d_{j,0}\! +\! 3\d_{j,1}\!+\!\d_{j,2}\Big)\Big(}+(7-14t)\vp\e^t +4(t+1)\vp\e^{t+1}\Big),\end{array}$$
which gives (\ref{b-5-3}).
If $i=2, s\neq 1,2$, then
(\ref{H-12}) shows that
$[\vp^2\e^j, \vp^s\e^t]$ is equal to
$$\begin{array}{lll}\!\!\!\!-(2\d_{j,0}+7\d_{j,1}+9\d_{j,2}
+5\d_{j,3}+\d_{j,4})\Big((2s-t+1)\vp^s\e^{t-1}+3(s-t)\vp^s\e^t-2(t+1)\vp^s\e^{t+1}\Big),\end{array}$$
which is (\ref{b-5-4}).
Finally if $i=2, s=2$, we see from
(\ref{H-12}) that $[\vp^2\e^j, \vp^2\e^t]$ is equal to
$$\begin{array}{lll}\Big(2\d_{t,0}+7\d_{t,1}+9\d_{t,2}+5\d_{t,3}+\d_{t,4}\Big)
\Big((5-j)\vp^2\e^{j-1}+3(2-j)\vp^2\e^j-2(j+1)\vp^2\e^{j+1}\Big) \\[4pt]
-\Big(2\d_{j,0}+7\d_{j,1}+9\d_{j,2}+5\d_{j,3}+\d_{j,4}\Big)\Big((5-t)\vp^2\e^{t-1}+3(2-t)\vp^2\e^t-2(t+1)\vp^2\e^{t+1}\Big),\end{array}$$
and we obtain (\ref{b-5-5}). The proof of the theorem is completed.
\QED
\section{Conclusion remark}
Theorems \ref{theorem-2}, \ref{th-2-2}, \ref{theorem-3-1} and \ref{last-theo} provide
five classes of infinite dimensional Lie Bialgebras $(\AA^{\circ}, [\cdot, \cdot], \D)$ of Poisson type.
As by-products, we obtain five new classes of infinite dimensional Lie algebras $({\cal L}, [\cdot, \cdot])$ with the underlining space ${\cal L}=\F[\vp,\e]$
(the polynomial algebra on two variables $\vp,\,\e$)
and brackets defined respectively by (\ref{bracket-4-2}), (\ref{eq2.2-1})--(\ref{eq2.2-2}), (\ref{+eq2.2-1})--(\ref{+eq2.2-3}), (\ref{caseqq1})--(\ref{caseqq2}) and (\ref{b-5-1})--(\ref{b-5-6}). We close the paper by proposing the following questions: In which conditions will
Lie bialgebras $(\AA^{\circ}, [\cdot, \cdot], \D)$ defined in Theorems \ref{theorem-2}, \ref{th-2-2}, \ref{theorem-3-1} and \ref{last-theo} be coboundary triangular? What
kinds of structure and representation theories will these Lie algebras $({\cal L}, [\cdot, \cdot])$ have?
\Acknowledgements{ Supported by NSF grant 11071147, 11431010, 11371278
of China,
NSF grant ZR2010AM003, ZR2013AL013 of Shandong Province,
SMSTC grant 12XD1405000, Fundamental Research Funds for the Central Universities.}
|
\section{Introduction and overview}
Galaxies are three-dimensional (3D) structures moving under the
dictates of gravity in a 3D Universe. From our position on the Earth,
astronomers have only the opportunity to observe their properties
projected onto a two-dimensional (2D) plane, usually called the plane
of the sky. Since we can neither circumnavigate galaxies nor wait
until they spin around, our knowledge of the intrinsic shape of
galaxies is still limited, relying on sensible, but sometimes not
accurate, physical and geometrical hypotheses.
Despite the obvious difficulties inherent to measure the intrinsic 3D
shape of galaxies, it is doubtless that it keeps an invaluable piece
of information about their formation and evolution. In fact,
astronomers have acknowledged this since galaxies were established to
be {\it island universes} and the topic has produced an outstanding
amount of literature during the last century.
In this paper I discuss the main developments and results in the quest
to better understand the 3D shape of galaxy bulges. Given the limited
space available in this chapter, I have not elaborated on the concept
and definition of a bulge, leaving this discussion to another chapter
in this volume. In the same way, I have deliberately not included the
intrinsic shape of boxy/peanut (B/P) structures located in the centre
of disc galaxies which some authors associate to galaxy bulges
\citep{lutticke00}. Currently it is well established that these
structures are actually part of the bar and intimately related to
their secular evolution \citep{combessanders81,chungbureau04}. As
bars evolve, stars can be moved perpendicular to the disc plane due to
a coherent bending of the bar producing its characteristic shape
\citep{debattista04, martinezvalpuesta06}. B/P structures share the
same photometric and kinematic properties of bars
\citep{mendezabreu08b,erwindebattista13}.
On the other hand, I have included a historical review of the
evolution of our knowledge of the intrinsic shape of elliptical
galaxies. The properties of elliptical galaxies and those of
intermediate/massive galaxy bulges have been often considered to be
similar \citep{wyse97}. This is particularly true when referring to
their surface-brightness distributions and shapes. Indeed, it has
been common in the literature to rely on both simulations and
observations of elliptical galaxies to interpret the observational
properties of bulges \citep[e.g.,][]{kormendybender12}.
This paper is structured as follows. In Section \ref{sec:scene} I
describe the basic geometric considerations of the problem and set up
the notation used throughout the chapter. In Section \ref{sec:back} I
review our current knowledge on the intrinsic shape of both elliptical
and disc galaxies. Section \ref{sec:bulges} introduces the advantages
and drawbacks of studying galaxy bulges with respect to ellipticals
and a historical perspective of their 3D shape measurements. In
Section \ref{sec:MW} I summarize the evolution of the concept of the
Milky Way bulge and its intrinsic 3D shape. Section
\ref{sec:simulations} addresses the importance of numerical
simulations to understand the physical processes that shape galaxy
ellipsoids. Finally, in Section \ref{sec:conclusions} I sketch out
the current view on the intrinsic shape of bulges and explore future
prospects.
\section{Setting up the Scene}
\label{sec:scene}
This section briefly summarizes the basic notation and geometrical
considerations to be used during this chapter.
Let ($x, y, z$) be the Cartesian coordinates on the reference system
of the galaxy with the origin in the galaxy centre, the $x-$axis and
$y-$axis corresponding to the principal equatorial axes of the
ellipsoidal component, and the $z-$axis corresponding to the polar
axis. Therefore, if $A$, $B$, and $C$ are the intrinsic lengths of the
ellipsoid semi-axes, the corresponding equation of the bulge on its
own reference system is given by
\begin{equation}
\frac{x^2}{A^2} + \frac{y^2}{B^2} + \frac{z^2}{C^2} = 1
\label{eqn:ellipsoid}
\end{equation}
Let $(x',y',z')$ now be the Cartesian coordinates on the observer
reference system. It has its origin in the galaxy centre, the polar
$z'-$axis is along the line of sight (LOS) and points toward the
galaxy. $(x',y')$ represents the plane of the sky.
The equatorial plane $(x, y)$ of the ellipsoid and the plane of the
sky $(x',y')$ intersect in the so-called line of nodes (LON). The
angle between both planes, i.e., the angle subtended between $z$ and
$z'$ is defined as the inclination $\theta$ of the ellipsoid. The
remaining two Euler angles which allow for the transformation from the
reference system of the galaxy to that of the sky are defined as: i)
$\phi$ is the angle subtended between the $x-$axis and the LON in the
ellipsoid equatorial plane, and ii) $\psi$ is the angle subtended
between the $x'-$axis and the LON in the plane of the sky. It is
often useful to choose the $x'-$axis to be along the LON, consequently
it holds that $\psi=0$ (see Figure \ref{fig:geometry}).
\begin{figure}[t]
\includegraphics[width=\textwidth]{mendezabreu_fig1.ps}
\caption{Schematic three-dimensional view of the ellipsoid geometry.
The bulge ellipsoid, the disc plane, and the sky plane are shown
in red, blue, and green, respectively. The reference systems of
both the ellipsoid and the observer as well as the LON are plotted
with thin solid lines, thin dashed lines, and a thick solid line,
respectively. The bulge ellipsoid is shown as seen from an
arbitrary viewing angle (left panel), along the LOS (central
panel), and along the polar axis (i.e., the $z-$axis; right
panel). Extracted from \citep{mendezabreu10}. Reproduced with
permission from Astronomy \& Astrophysics, \textcircled{c} ESO.}
\label{fig:geometry}
\end{figure}
It is well known that the projection of a triaxial ellipsoid onto the
plane of the sky describes an ellipse
\citep{contopoulos56,stark77,binney85,franx91}, which is usually
written as
\begin{equation}
\frac{x_{\rm e}^2}{a^2} + \frac{y_{\rm e}^2}{b^2} = 1
\label{eqn:ellipse}
\end{equation}
where $x_{\rm e}$ and $y_{\rm e}$ represent the axes of symmetry of
the projected ellipse, $a$ and $b$ are the corresponding semi-major
and semi-minor axes of the ellipse. The observed ellipticity of the
ellipse can be easily derived from the apparent axis ratio as
$\epsilon=1 - b/a$. The $x_{\rm e}$− axis forms an angle $\delta$
with the LON (twist angle), which for convenience is usually made to
correspond with the $x'$-axis. It is worth noting that both the
apparent axis ratio ($q=b/a$) and the orientation of the ellipses
($\delta$) depend only, and unambiguously, on the direction of the
LOS, i.e., on $\theta$, $\phi$, and $\psi$, and on the intrinsic shape
of the ellipsoid, i.e., $A, B,$ and $C$, see \citet{simonneau98} for
the full derivation.
Based on this simple geometric representation, if we assume a galaxy
is composed of a set of triaxial emitting ellipsoidal shells, which
are concentric and coaxial (same axes of symmetry) but non-homologous
(intrinsic semi-axes vary with the distance to the centre), their
projections onto the plane of the sky are concentric ellipses, but
non-homologous and non-coaxial. Therefore, the twisting of the galaxy
isophotes can be explained just as an effect of the projection of
non-homologous triaxial ellipsoids \citep{williamsschwarzchild79}.
\section{Historical background on the intrinsic shape of galaxies}
\label{sec:back}
Elliptical galaxies are structurally the simplest stellar systems
where mathematical techniques can be applied to recover their
intrinsic 3D shape. Thus, the huge amount of literature on the
subject is not surprising. In fact, the continuously increasing
availability of better measurements of the apparent axis ratios of
elliptical galaxies have motivated great debate over the years. On
the other hand, the similarities between the photometric properties of
intermediate/massive bulges and ellipticals \citep[e.g.,][]{gadotti09}
have usually motivated an extrapolation of the results on the
intrinsic 3D shape of ellipticals and their implications on galaxy
formation and evolution onto the bulges of disc galaxies. In this
section I revisit our current knowledge on the intrinsic shape of
elliptical galaxies (Sect. \ref{sec:ellipticals}) and, for the sake
of completeness, of disc galaxies (Sect. \ref{sec:discs}) to put in
context the historical background on the intrinsic shape of bulges.
\subsection{Intrinsic shape of elliptical galaxies}
\label{sec:ellipticals}
\subsubsection{Photometric approach}
The first attempt to derive the intrinsic shape of elliptical galaxies
was done by \citet{hubble26}. At that time, it was already realized
the importance of relying on statistical methods to recover the 3D
shape of galaxies. In fact, Hubble obtained the frequency of
intrinsic short-to-long axis ratio under the assumption that
elliptical galaxies were oblate ellipsoids with random orientations
with respect to the LOS.
Since then, this statistical approach based on the measurement of the
apparent axis ratio distribution (AARD) and the assumption that the 3D
intrinsic shape is an ellipsoids of revolution, either oblate or
prolate, has been extensively used in the literature. For the sake of
clarity I briefly outline here the basic statistical concepts.
Let us assume the basic geometry proposed in Sect. \ref{sec:scene}
and define both the intrinsic ellipticity, $Q=B/A$, and intrinsic
flattening, $F=C/A$, of the ellipsoid as the corresponding intrinsic
axis ratios in the $(x,y)$ and $(x,z)$ planes,
respectively. Therefore, in the case of either a pure oblate ($Q=1$)
or pure prolate ($Q=F$) ellipsoid the Eq. \ref{eqn:ellipsoid} can be
described by one single parameter. If the polar axis of the ellipsoid
forms an angle ($\theta$) with respect to the LOS then the apparent
axis ratio of the projected ellipse can be written as
\begin{equation}
F^2\,\sin^2{\theta}\,+\,\cos^2{\theta}= \left\{
\begin{array}{l l}
q^2 & \quad {\rm if\,\, oblate}\\
q^{-2} & \quad {\rm if\,\, prolate}\\
\end{array} \right.
\label{eqn:q}
\end{equation}
Under the realistic assumption of randomly distributed orientations
and using Eq. \ref{eqn:q} where $q=q(\theta)$, the probability
$P(q|F)dq$ that a galaxy with intrinsic axis ratio $F$ is observed
with an apparent axis ratio in the range ($q, q+dq$) is
\begin{equation}
P(q|F)dq=\frac{\sin{\theta}\,dq}{|dq/d\theta|}
\label{eqn:probqF}
\end{equation}
At this point, the AARD $\zeta(q)$, can be related to the intrinsic
probability distribution $\xi(F)$ by
\begin{equation}
\zeta(q) = \int_0^1 P(q|F)\,\xi(F)\,dF
\label{eqn:probq}
\end{equation}
The relation between the known (observed) frequency of galaxies of
apparent axis ratio $\zeta(q)$ to the unknown frequency $\xi(F)$ of
galaxies with intrinsic axis ratio $F$ can be written such as
\begin{equation}
\zeta(q)= \left\{
\begin{array}{l l}
q\,\displaystyle\int_0^q \frac{\xi(F)dF}{\sqrt{(1-F^2)(q^2-F^2)}} & \quad {\rm if\,\, oblate} \\
q^{-2}\,\displaystyle\int_0^q \frac{\xi(F)F^2\,dF}{\sqrt{(1-F^2)(q^2-F^2)}} & \quad {\rm if\,\, prolate}\\
\end{array} \right.
\label{eqn:q2}
\end{equation}
Based on this approach and using the hypothesis of oblateness,
\citet{sandage70} derived the intrinsic distribution of flattening
$\xi(F)$ for different Hubble types ranging from ellipticals to Sc.
They found that the observed axis ratios of 168 elliptical galaxies
present in the Reference Catalog of Bright Galaxies (RC1)
\citep{devaucouleurs64} were well reproduced using a skewed binomial
distribution of oblate ellipsoids given by
\begin{equation}
\xi(F)\propto\left(1+\frac{F-F_{0}}{\beta}\right)^{\alpha}\,exp\left[-\alpha\,(F-F_{0})\right]
\label{eqn:probF}
\end{equation}
with main parameters $F_0=0.58$ and $\beta=0.31$ (Figure
\ref{fig:ellipticals}, left panels).
\citet{binney78} used the same sample but introducing the prolate
approach. Adopting the same functional form for $\xi(F)$ he found
values of $F_0=0.40$ and $\beta=0.71$. However, even if using
arbitrary analytical representations of $\xi(F)$ can turn out in a
good fit of the AARD, in principle they do not have a physical
motivation. This approximation was improved by \citet{noerdlinger79}
by solving Eq. \ref{eqn:q2} using the non-parametric inversion
technique proposed by \citet{lucy74}. His results show how under the
hypothesis of oblateness the $\xi(F)$ distribution of
\citet{sandage70} was correct, but he also noticed that a prolate
distribution peaking at around $F\sim0.7$ would produce a good
representation of the data as well.
At the same time, some kinematic findings led to the suggestion that
the structure of elliptical galaxies can be represented by neither
oblate nor prolate ellipsoids of revolution. In fact, the low ratio
between rotational velocity and velocity dispersion found in flat
systems \citep{bertolacapaccioli75, illingworth77, peterson78} or the
rotation measured along the minor axis of some elliptical galaxies
\citep{schechtergunn79} were interpreted as resulting from a triaxial
structure. From the photometric point of view, the twisting of the
inner isophotes of elliptical galaxies was known since the early work
of \citet{evans51} and it was latter confirmed in several works
\citep{liller60, carter78, bertolagalletta79}.
As a consequence, \citet{benacchiogalletta80} and
\citet{binneydevaucouleurs81} showed that the AARD could be
satisfactorily accounted for also in terms of a distribution of
triaxial ellipsoids. Nevertheless, these works still presented
significant differences in the predicted number of spherical galaxies
mainly due to the differences in the original samples. Other groups
reached similar conclusions analysing higher quality data coming from
new CCD detectors \citep{fasanovio91}.
A new step forward in the methodology to recover the intrinsic 3D
shape of galaxies was done by \citet{fallfrenk83}. They showed how
the inversion of the integral equations for oblate and prolate
ellipsoids (Eq. \ref{eqn:q2}) can be performed analytically,
resulting in
\begin{equation}
\xi(F)= \frac{2}{\pi}\,\sqrt{1-F^2}\left\{
\begin{array}{l l}
\frac{1}{F}\,\int_0^F \frac{qdq}{\sqrt{F^2 - q^2}}\,\frac{d\zeta}{dq} & \quad {\rm if\,\, oblate}\\
\frac{1}{F^3}\,\int_0^F \frac{qdq}{\sqrt{F^2 - q^2}}\,\frac{d(q^3\zeta)}{dq} & \quad {\rm if\,\, prolate}\\
\end{array} \right.
\label{eqn:inversion}
\end{equation}
Using this analytical inversion and the largest sample of galaxies to
that date (2135 elliptical galaxies), \citet{lambas92} demonstrated how
neither oblate nor prolate models could adequately reproduce the data.
Contrarily, triaxial ellipsoids with intrinsic axis ratios selected
from 1D Gaussians provided an adequate fit to the data. They found a
best fit with $Q=0.95$ and $F=0.55$.
A similar approach was used by \citet{ryden92} on a smaller sample of
171 elliptical galaxies. She used a 2D Gaussian combining both
intrinsic axis ratios obtaining $Q=0.98$ and $F=0.69$. The same
sample was later analysed by \citet{tremblaymerritt95} using a
non-parametric technique to test the triaxial hypothesis. They
confirmed previous results that discarded a distribution of intrinsic
shapes compatible with axisymmetric ellipsoids thus favouring triaxial
distributions. Similar conclusions were reached by \citet{ryden96} on
a larger sample using the same non-parametric approach.
During these years it became increasingly clear that the distribution
of intrinsic flattenings of elliptical galaxies was broad and possibly
bimodal \citep{fasanovio91,ryden92,tremblaymerritt95,ryden96}. In
fact, combining the galaxy sample described in \citet{ryden92} with a
new sample of brightest cluster galaxies (BCGs) from
\citet{lauerpostman94}, \citet{tremblaymerritt96} found that the AARD
of galaxies brighter than $M_B\simeq-20$ was different from that of
the less luminous ones. This reflected a difference in the shape of
low-luminosity and high-luminosity ellipticals: fainter ellipticals
are moderately flattened and oblate, while brighter ellipticals are
rounder and triaxial. Recently, \citet{fasano10} also found that even
if both normal ellipticals and BCGs are triaxial, the latter tend to
have a more prolate shape, and the tendency to prolateness is mainly
driven by the central dominant (cD) galaxies present in their sample.
The next qualitative leap in studies of the intrinsic shape of
elliptical galaxies happened with the advent of the Sloan Digital Sky
Survey (SDSS). With respect to previous statistical analyses, SDSS
improved not only the number of galaxies under study (an order of
magnitude larger) but also the quality and homogeneity of the
photometry. All these improvements allowed to study the dependence of
the intrinsic shape with other galaxy properties such as the
luminosity, colour, physical size, and environment. Using data from
the SDSS-DR3 \citep{abazajian05} \citet{vincentryden05} found that
bright galaxies ($M_{r}\leq-21.84$) with a de Vaucouleurs profile have
an AARD consistent with a triaxiality parameter in the range
$0.4<T<0.8$, where $T=(1-Q^2)/(1-F^2)$, and mean flattening
$0.66<F<0.69$. The faintest de Vaucouleurs galaxies are best fit with
prolate ellipsoids ($T=1$) with mean flattening $F=0.51$. Using the
SDSS-DR5 \citep{adelmanmccarthy07}, \citet{kimmyi07} were able to
reproduce the AARD by using a combination of oblate, prolate, and
triaxial galaxy populations. Following the early work of
\citet{tremblaymerritt96}, they assumed each population having a
Gaussian distribution of their intrinsic axis ratios. The best fit to
the AARD was found using a fraction of O:P:T=0.29:0.26:0.45
(Oblate:Prolate:Triaxial) with a best triaxial distribution with axis
ratios $Q=0.92$ and $F=0.78$. In 2008, \citet{padillastrauss08} used
the SDSS-DR6 \citep{adelmanmccarthy08} to derive the intrinsic shape
of ellipticals with the main improvement of taking into account the
effects of dust extinction. They found that the AARD of elliptical
galaxies shows no dependence on colour, suggesting that dust
extinction is not important for this sample. The full population of
elliptical galaxies was well characterized by a Gaussian distribution
in the equatorial ellipticity with mean $Q=0.89$ and a lognormal
distribution of the flattening with mean $F=0.43$, which corresponds
to slightly oblate ellipsoids in agreement with
\citet{vincentryden05}. In a recent paper, \citet{rodriguezpadilla13}
have used the SDSS-DR8 \citep{aihara11} and the morphological
information from Galaxy Zoo \citep{lintott11} finding that elliptical
galaxies have a mean value of $F=0.58$ (Figure \ref{fig:ellipticals},
right panels). They concluded that the increase in $F$ is mainly due
to the removal of the spiral galaxy contamination thanks to the Galaxy
Zoo morphologies. A historical summary in tabular form of all these
measurements is shown in Table \ref{tab:historical}.
Owing to the ill-posed problem of deriving the 3D intrinsic shape of
elliptical galaxies, its historical perspective is mainly weighted
toward statistical methods. As previously showed in this section, the
inventiveness of astronomers, the development of statistical methods,
and the advent of large surveys have significantly improved our
knowledge of the intrinsic shape of elliptical galaxies. Other
methods based on the photometric study of individual galaxies have
also been developed but to a smaller extent. One of the pioneering
works to derive the intrinsic shape of an individual elliptical using
its observed ellipticity and isophotal twist was done by
\citet{williams81}. They modelled the elliptical galaxy NGC~0523
assuming a given intrinsic density distribution and finding that the
preferred models were prolate in the external regions but increasingly
mixed (oblate and prolate) towards the centre. This idea was further
developed by other authors using more complex models of the density
distribution \citep{fasano95, thakurchakraborty01}. In 2008,
\citet{chakraborty08} estimated the shapes of 10 elliptical galaxies
with apparent ellipticities $\epsilon\leq 0.3$, finding that radial
differences in the triaxiality parameter can be tightly constrained to
values $0.29<\Delta T<0.54$. \citet{chakraborty11} extended this
analysis to 3 very flat galaxies with ellipticity $\epsilon\sim0.3$ or
more. They found values of the intrinsic flattening of these galaxies
around $F\sim0.5$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.8\textwidth]{mendezabreu_fig2a.ps}\\
\includegraphics[width=0.75\textwidth]{mendezabreu_fig2b.ps}
\caption{Composite figure showing the evolution of the galaxy
samples used in the derivation of the intrinsic shape of
ellipticals and discs. Upper panels: histograms of the AARD for
ellipticals and spiral galaxies. The overplotted curves are
predicted ratios for various assumptions of the distribution of
intrinsic flattening. On the right, the assumed intrinsic
distribution corresponding to the curves on the left. Extracted
from \citet{sandage70}. Reproduced with permission,
\textcircled{c} AAS. Bottom panels: best fit models to AARD
compared to the observations. Top: spirals. Bottom: ellipticals.
Left: galaxies selected only by fracDeV, see \citet{abazajian05}
for definition. Right: galaxies selected by Galaxy Zoo morphology
and fracDeV. Extracted from
\citet{rodriguezpadilla13}. Reproduced by permission of Oxford
University Press.}
\label{fig:ellipticals}
\end{center}
\end{figure}
\subsubsection{Kinematic approach}
Determining the distribution of the 3D intrinsic shape of elliptical
galaxies is also possible by combining photometric and kinematic
information. In a first attempt, \citet{binney85} used simple
kinematical models to understand the ratio of rotational motion along
both the major and minor isophotal axes of the galaxy. Using a sample
of 10 ellipticals he found that elliptical galaxies were not well
represented by axisymmetric oblate or prolate models. \citet{franx91}
revisited this approach by using a larger sample of 38 elliptical
galaxies and studying the probability distribution of photometric
ellipticities and kinematics misalignments. In particular, they
explored the possibility that the angular momentum could not be
aligned with the polar axis of the galaxy but it may have any
orientation within the plane containing the short and the long axis
($x,z$). They found that a variety of models was able to reproduce the
observations. Models with all galaxies being triaxial with
well-aligned angular momentum were indistinguishable from models with
all galaxies being oblate with nonaligned angular momentum.
A different standpoint to statistical studies implies an investigation
into the intrinsic shape of elliptical galaxies using detailed
individual dynamical modelling of the galaxy kinematics.
\citet{tenjes93} modelled the photometric and stellar kinematic
measurements of three elliptical galaxies adopting a specific form for
the intrinsic density and streaming motions. They found tightly
constrained geometries with $0.7<Q<0.8$ and $0.4<F<0.6$. This
methodology was further improved in a series of papers by Statler
\citep{statler94a,statler94b,statler94c}. He showed how using not only
their apparent shapes and velocity field misalignments, but also the
velocity field asymmetry, it is possible to place tighter constraints
on the intrinsic shape of ellipticals. Using this approach
\citet{bakstatler00} derived the intrinsic shape of 13 elliptical
galaxies finding that although photometric studies give similar
results for the flattening, none is able to put real constraints on
triaxiality even when large samples are studied, hence demonstrating
the need to include kinematic data in the models. Figure
\ref{fig:ellipticals_kin} show the probability distribution of
intrinsic axis ratio for 9 galaxies with significant rotation in their
sample. It is clear that most of the galaxies can be well described by
nearly oblate models but some of them present significant triaxiality
or even prolateness. \citet{vandenboschvandeven09} investigated how
well the intrinsic shape of elliptical galaxies can be recovered by
fitting realistic triaxial dynamical models to simulated photometric
and kinematic observations. They found that for axisymmetric
galaxies, the models are able to exclude triaxiality but the intrinsic
flattening is nearly unconstrained. On the other hand, the shape of
triaxial galaxies can be accurately determined when additional
photometric and kinematic complexity, such as the presence of
isophotal twist or a kinematically decoupled core is observed.
Recently, \citet{weijmans14} studied the intrinsic shape of the
early-type galaxies described in the ATLAS3D survey
\citep{cappellari11}. Using a purely photometric approach and assuming
axisymmetry, they found that the fast rotator population was much
flatter than the slow rotator population, as expected from their
dynamical status. Moreover, when the kinematic misalignment is
included as a constraint in the analysis, they demonstrated that fast
rotators are still better represented to oblate ellipsoids.
\begin{figure}[t]
\includegraphics[width=\textwidth]{mendezabreu_fig3.ps}
\caption{Posterior probability densities in the plane of intrinsic
triaxiality, T, and flattening, $c_{L}$ ($F$ in this chapter), for
each of the nine galaxies that show significant rotation in
\citet{bakstatler00}. Contours indicate the 68\% and 95\% highest
posterior density regions. In each panel, round prolate galaxies
are at the top left, flattened oblate galaxies at bottom right,
and objects in between are triaxial. Most galaxies are well
represented by oblate models but prolate and triaxial are also
allowed in many galaxies, e.g., NGC~741, NGC~4486, or NGC~7626.
Extracted from \citet{bakstatler00}. Reproduced with permission,
\textcircled{c} AAS.}
\label{fig:ellipticals_kin}
\end{figure}
\begin{table}[t]
\caption{Historical summary of the intrinsic shapes of elliptical galaxies.}
\label{tab:historical}
\begin{tabular}{p{1cm}p{1.5cm}p{3cm}p{2cm}p{2cm}p{1.5cm}}
\hline\noalign{\smallskip}
Year & N. Galaxies & Hypothesis & $Q$ & $F$ & Reference \\
(1) & (2) & (3) & (4) & (5) & (6) \\
\hline
1970 & 168 & Oblate & 1 & 0.58 & [1]\\
1978 & 168 & Prolate & 0.4 & 0.4 & [2]\\
1979 & 168 & Oblate/Prolate & 1/0.7 & 0.55/0.7 & [3]\\
1980 & 348 & Triaxial & 0.81 & 0.62 & [4]\\
1981 & 196 & Oblate/Prolate/Triaxial & 1/0.62/0.79 & 0.62/0.62/0.57 & [5]\\
1992 & 2135 & Triaxial & 0.95 & 0.55 & [6]\\
1992 & 171 & Triaxial & 0.98 & 0.69 & [7]\\
2005 & 26994 & Triaxial & 0.66-0.85 & 0.66-0.69 & [8]\\
2007 & 3922 & Oblate/Prolate/Triaxial & 1/0.72/0.92 & 0.44/0.72/0.78 & [9]\\
2008 & 303390 & Triaxial & 0.89 & 0.38 & [10]\\
2013 & 112100 & Triaxial & 0.88 & 0.58 & [11]\\
\noalign{\smallskip}\hline\noalign{\smallskip}
\end{tabular}
Notes. (1) Year of publication of the paper. (2) Number of elliptical
galaxies in each sample. (3) Hypothesis used to derive the intrinsic
shape of the ellipticals. (4) Mean value of the intrinsic
ellipticity. (5) Mean value of the intrinsic flattening. (6) Reference
of the corresponding paper: [1] \cite{sandage70}, [2] \cite{binney78},
[3] \cite{noerdlinger79}, [4] \cite{benacchiogalletta80}, [5]
\cite{binneydevaucouleurs81}, [6] \cite{lambas92}, [7] \cite{ryden92},
[8] \cite{vincentryden05}, [9] \cite{kimmyi07}, [10]
\cite{padillastrauss08}, [11] \cite{rodriguezpadilla13}.
\end{table}
\subsection{Intrinsic shape of disc galaxies}
\label{sec:discs}
In this section I briefly summarise our current understanding about
the intrinsic 3D shape of discs. Bulges are embedded into the disc
light and axisymmetry is usually a requirement to derive the bulge
intrinsic shape. However, although the discs of lenticular and spiral
galaxies are often considered to be infinitesimally thin and perfectly
circular, their intrinsic shape is better approximated by flattened
triaxial ellipsoids.
The disc flattening, defined analogously as for ellipticals
(Sect. \ref{sec:ellipticals}), can be directly determined from edge-on
galaxies. It depends both on the wavelength at which discs are
observed and on galaxy morphological type. Indeed, galactic discs
become thicker at longer wavelengths \citep{dalcantonbernstein02,
mitronova04} and late-type spirals have thicker discs than
early-type spirals \citep{bottinelli83, guthrie92}.
Determining the distribution of both the intrinsic flattening and
ellipticity of discs is possible by a statistical analysis of the AARD
of randomly oriented spiral galaxies.
Similarly for elliptical galaxies, \citet{sandage70} analysed the
spiral galaxies listed in the RC1. They concluded that discs are
circular with a mean flattening of $\langle F \rangle = 0.25$.
However, the lack of nearly circular spiral galaxies ($q\simeq1$)
rules out that discs have a perfectly axisymmetric shape. Indeed,
\citet{binggeli80}, \citet{benacchiogalletta80}, and
\citet{binneydevaucouleurs81} have shown that discs are slightly
elliptical with a mean intrinsic ellipticity $\langle 1-Q \rangle =
0.1$.
These early findings were based on the analysis of photographic plates
of a few hundreds of galaxies. They were later confirmed by measuring
ellipticities of several thousands of objects in CCD images and
digital scans of plates obtained in wide-field surveys.
\citet{lambas92} found that pure oblate models failed to reproduce the
AARD of spiral galaxies, whereas nearly oblate models with $F\sim0.2$
and $Q\sim0.9$ produce a good fit with values similar to those of
\citet{sandage70}. These values were confirmed later on by different
authors \citep{fasano93, alamryden02, ryden04}.
Like the flattening, the intrinsic ellipticity depends on the
morphological type and wavelength. The discs of early-type spirals
are more elliptical than those of late-type spirals and their median
ellipticity increases with observed wavelength \citep{ryden06}.
Furthermore, luminous spiral galaxies tend to have thicker and rounder
discs than low-luminosity spiral galaxies \cite{padillastrauss08}. In
\citet{sanchezjanssen10} they studied the role of stellar mass in
shaping the thickness of galaxy discs. They found that the intrinsic
thickness distribution of discs has a characteristic {\it U-shape} and
identify a limiting mass $M_{\star} \approx 2\times 10^{9}M_{\odot}$
below which low-mass galaxies start to be systematically
thicker. Recently, \citet{rodriguezpadilla13} analyse a sample of
92923 spiral galaxies extracted from the SDSS-DR8, and taking into
account the effects of dust in their analysis, they found a
distribution of flattening with mean $F=0.27$ and ellipticity
$Q=0.22$, i.e., disc are less round than in previous studies (Figure
\ref{fig:ellipticals}, right panels).
Despite the large effort made to understand the intrinsic 3D shape of
galaxy discs, it is still unclear whether the inferred slight
triaxiality could be due to the presence of substructure in galaxy
discs or if it really reflects truly triaxial potential in spirals.
\section{The intrinsic shape of extragalactic bulges.}
\label{sec:bulges}
The study of the intrinsic shape of bulges presents similarities,
advantages, and drawbacks with respect to that of elliptical galaxies.
Bulges are ellipsoidal systems located in the centre of disc galaxies,
thus, the main drawback with respect to elliptical galaxies is that
their analysis requires the isolation of their light distributions
from other structural galaxy components. However, it is worth noting
that a similar problem is faced in elliptical galaxies when defining a
characteristic radius to measure the global axis ratio of the galaxy
\citep{fasanovio91}. The most common approach to identifying a global
axis ratio for the bulge is by performing a photometric decomposition
of the galaxy surface-brightness distribution. In this method, the
galaxy light is usually modelled as the sum of the contributions from
the different structural components, i.e., bulge and disc, and
eventually lenses, bars, spiral arms, and rings
\citep{prieto01,laurikainen05}. A number of two-dimensional
parametric decomposition techniques have been developed to this aim,
such as: GIM2D \citep{simard98}, GALFIT \citep{peng02}, BUDDA
\citep{desouza04}, GASP2D \citep{mendezabreu08}, GALPHAT
\citep{yoon11}, or IMFIT \citep{erwin15}.
On the other hand, the main drawback on the study of galaxy bulges,
i.e., the presence of other components such as the main disc,
represents in turn the main advantage. The presence of the galactic
disc allows for accurately constraining the inclination of the galaxy.
Hence, under the assumption that the two components share the same
polar axis (i.e., the equatorial plane of the disc coincides with that
of the bulge) it allows for the determination of the inclination of
the bulge. This is crucial to solve one of the main concerns when
dealing with elliptical galaxies.
\subsection{Photometric approach}
\label{sec:bulphot}
Galaxy bulges were initially thought as axisymmetric ellipsoids placed
at the centre of disc galaxies. The first piece of photometric
evidence against this idea was given by \citet{lindblad56}. He showed
a misalignment between the major axes of the disc and bulge in M31,
realising that this would be impossible if both the disc and bulge
were oblate. This photometric misalignment is similar to the isophote
twist observed in elliptical galaxies and used as an indication of
triaxiality in these systems \citep{williamsschwarzchild79}. The
extensive study undergone by \citet{kent84} showed that the twisting
isophotes between the central and outer parts of disc galaxies are
quite common, but it was not until 1986 when \citet{zaritskylo86}
properly studied the deviations from axisymmetry in the bulges of
spiral galaxies. They found bulge-to-disc misalignments in their
sample of 11 spiral galaxies hence confirming the high incidence of
non-axisymmetric bulges in ordinary spirals and placing some
parallelisms with elliptical galaxies. \citet{beckman91} also found
compelling photometric evidence for triaxiality in the bulge of
NGC~4736.
The first quantitative estimation of the intrinsic 3D shape of galaxy
bulges using a statistical approach was performed by
\citet{bertola91}. They measured the bulge AARD and the misalignments
between the major axes of the bulge and disc in a sample 32 S0--Sb
galaxies. Under the hypothesis that discs are circular, they found
that these bulges are triaxial with mean axial ratios $\langle Q
\rangle=0.86$ and $\langle F \rangle=0.65$. Interestingly, they also
demonstrated that a random projection of the probability distribution
function of the bulges axis ratios fit sufficiently well to the AARD
of the elliptical galaxies presented in \citet{binneydevaucouleurs81}.
The results were interpreted as both populations of objects having the
same origin.
\citet{fathipeletier03} derived the intrinsic ellipticity of bulges by
analysing the deprojected apparent axis ratio of the galaxy isophotes
within the bulge radius. This work did not assume any geometrical
model for the galaxy but only that the disc be circular. They found
$\langle Q\rangle = 0.79$ and $\langle Q\rangle = 0.71$ for the bulges
of 35 early-type and 35 late-type disc galaxies, respectively. Despite
the different methodologies, these results were in good agreement with
previous results by \citet{bertola91}. Along the same lines, none of
the 21 disc galaxies with morphological types between S0 and Sab
studied by \citet{noordermeervanderhulst07} harbours a truly spherical
bulge. They reach this conclusion by assuming bulges to be oblate
ellipsoids and comparing the isophotal axis ratio in the
bulge-dominated region to that measured in the disc-dominated
region. A mean flattening $\langle F \rangle=0.55$ was obtained which
is slightly lower than the value found by \citet{bertola91}.
The number of galaxy bulges under study increased by an order of
magnitude with the work of \citet{mendezabreu08}. They measured the
structural parameters of bulges and discs of a sample of 148
early-to-intermediate spiral galaxies using a 2D photometric
decomposition. They computed the probability distribution function of
the intrinsic ellipticity from the bulges AARD, disc ellipticities,
and misalignments between bulges and discs position angles. They
suggested that about 80\% of the sample bulges are triaxial ellipsoids
with a mean axial ratio $\langle B/A \rangle$ = 0.85, confirming that
bulges are slightly triaxial structures.
The vertical extension of galaxy bulges remains usually hidden from
observations except for edge-on galaxies. \citet{mosenkov10} obtained
a median value of the flattening $\langle F \rangle=0.63$ for a sample
of both early- and late-type edge-on galaxies using near infrared
photometry. These results match well with the early findings by
\citet{bertola91}.
As well as for elliptical galaxies a number of works have attempted to
quantify the intrinsic shape of individual bulges using only
photometric data. The pioneering work of \citet{varela96} used a
combination of geometrical deprojection and photometric inversion to
work out the actual shape of the galaxy bulge in NGC~2841. They found
that a family of triaxial ellipsoids with variable axis ratios is
necessary to explain the photometric properties of its bulge. In 1998,
\citet{simonneau98} derived a set of equations defining the three
intrinsic axes of a triaxial ellipsoid as a function of the measured
geometry of a galaxy bulge and disc (axis ratios and position angles)
and the unknown Euler angle $ \phi$ (see Sect. \ref{sec:scene} for
definition). This seminal paper promoted the work of
\citet{mendezabreu10}. They introduced a new method to derive the
intrinsic shape of bulges based upon the analytical relations between
the observed and intrinsic shapes of bulges and their surrounding
discs. Using the equations derived in \citet{simonneau98} and
introducing physical constraints on the accessible viewing angles,
they found the following relation between the intrinsic semi-axes of
the bulge and their observed properties
\begin{small}
\begin{equation}
\frac{2\,\sin{\left(2\phi_C\right)}}{F_{\rm \theta}}\,F^2 = \sin{\left(2\phi_C-\phi_B\right)} \sqrt{\left(1-Q^2\right)^2 - \sin^2{\phi_B}\left(1+Q^2\right)^2}-\sin{\phi_B}\cos{\left(2\phi_C - \phi_B\right)}\left(1+Q^2\right)^2
\label{eqn:vartie}
\end{equation}
\end{small}
where $\phi_B$, $\phi_C$ and $F_{\rm \theta}$ are functions of the
observed quantities $a$, $b$, $\delta$, and $\theta$, see equations
12, 13, and 43 of \citet{mendezabreu10}. Therefore,
Eq. \ref{eqn:vartie} directly relates the intrinsic 3D shape of the
bulge with its observed properties. Unfortunately, the relation
between the intrinsic and projected variables also depends on the
spatial position of the bulge with respect to the disc on its own
reference system (i.e., on the $\phi$ angle) and therefore, as well as
for ellipticals, a deterministic solution of the problem cannot be
given. However, the statistical analysis provided in
\citet{mendezabreu10} allows us to obtain the probability distribution
function of both semi-axis ratios, $Q$ and $F$, for every single
bulge, thus imposing tight constraints on its actual shape. Applying
this technique to the sample of bulges presented in
\citet{mendezabreu08} they found a bimodal distribution of the
triaxiality parameter (Figure \ref{fig:bulgestriax}, left panel). In
particular, bulges with S\'ersic index $n \leq 2$ exhibit a larger
fraction of oblate axisymmetric (or nearly axisymmetric) bulges, a
smaller fraction of triaxial bulges, and fewer prolate axisymmetric
(or nearly axisymmetric) bulges with respect to bulges with $n > 2$.
Despite no correlations being found between the intrinsic shape of
bulges and other properties such as bulge luminosity or velocity
dispersion, the differences with the bulge surface-brightness
distribution hint towards the presence of different bulge populations
as suggested by \citet{kormendykennicutt04}.
\begin{figure*}
\includegraphics[width=0.49\textwidth]{mendezabreu_fig4a.ps}
\includegraphics[width=0.49\textwidth]{mendezabreu_fig4b.ps}
\caption{Composite figure showing the similar bimodal distribution
of triaxiality parameters from observations (left panel) and
simulations (right panel). Left panel: distribution of the
triaxiality parameter $T$ obtained from the sample of
\citet{mendezabreu10} (continuous line) and for a simulated sample
with both 30\% and 100\% of bulges hosting a nuclear bar (dashed
and dotted lines), respectively. Extracted from
\citet{mendezabreu10}. Reproduced with permission from Astronomy
\& Astrophysics, \textcircled{c} ESO. Right panel: distribution
of both dissipational (hatched histogram) and dissipationless
(solid line) mergers remnant triaxiality parameter from
\cite{cox06}. In both panels oblate galaxies have $T=0$, prolate
galaxies have $T=1$, and all values in between are
triaxial. Extracted from \citet{cox06}. Reproduced with
permission, \textcircled{c} AAS.}
\label{fig:bulgestriax}
\end{figure*}
\subsection{Evidences of triaxiality from kinematic measurements}
\label{sec:bulkin}
Early kinematic studies of galaxy bulges were shown to rotate more
rapidly than elliptical galaxies \citep{kormendyillingworth82}. In
fact, the kinematic properties of many bulges are well described by
dynamical models of oblate ellipsoids which are flattened by rotation
with little or no anisotropy \citep{daviesillingworth83,
jarvisfreeman85, fillmore86, corsini99, pignatelli01}. However,
there are also kinematic evidences supporting a triaxial shape in a
non-negligible fraction of these bulges. In 1989, two independent
works of \citet{bertola89} and \citet{gerhard89} reached the same
conclusion about the triaxial bulge of the Sa galaxy NGC~4845. Using a
combination of photometric and kinematic measurements they restrict
the intrinsic axis ratio of its bulge to $Q=0.74$ and $F=0.6$. Their
works were mainly supported by the presence of non-circular
gas-motions in the galaxy centre. In a non-axisymmetric potential, the
shape of the rotation curve will depend on the position of the LOS and
the major axis of the non-axisymmetric component. A slowly rising
rotation curve or one in which a bump of extreme velocities is seen
near the centre are indications of triaxiality
\citep{gerhard89}. Based on these considerations, and building on the
early work of \citet{lindblad56}, \citet{berman01} demonstrated the
presence of a triaxial bulge in the Andromeda galaxy (M31) by using a
hydrodynamical simulation to match the observed properties of the
galaxy. Further evidences for non-circular gas motion in galaxy
centres can be found in \citet{falconbarroso06} and
\citet{pizzella08}. Other kinematic evidence for the existence of
triaxial bulges comes from the presence of velocity gradients along
the galaxy minor axis. \citet{corsini03} found minor axis rotation in
80\% of their early-type spiral sample. In a series of papers,
\citet{coccato04, coccato05} found that 60\% of the unbarred galaxies
show a remarkable gas velocity gradient along their optical minor
axis. This was achieved by combining their own data with that present
in the literature (Revised Shapley-Ames Catalog of Bright Galaxies)
\citep{sandagetammann81}.
Despite the importance of adding kinematic information to determine
the intrinsic shape of the bulges, and contrary to the works on
elliptical galaxies \citep[e.g.,][]{statler94a}, there is not a
well-established methodology to quantify the degree of triaxiality of
bulges using the combined photometric and kinematic information, yet.
\subsection{Polar bulges}
Polar bulges, as well as their analogous polar rings
\citep{whitmore90}, are elongated structures perpendicular to the
plane of the galaxy disc. A common signature of both the orthogonally
decoupled bulge systems and the polar ring galaxies is that both
contain a structural component whose angular momentum vector is
roughly parallel to the major axis of the host galaxy.
Vertical elongation is not a common feature of bulges. Indeed, most
bulges can be assumed to be flattened by rotation (see
Sect. \ref{sec:bulkin}). Furthermore, orthogonally decoupled bulges
are usually not even {\it allowed} in most statistical works since the
condition $A>B>C$ is commonly used, see \citet{bertola91}.
\citet{mendezabreu10} relaxed this condition and found that only 18\%
of the observed bulges have a probability $>50$\% of being elongated
along the polar axis with no bulges reaching a probability $>90$\%.
In fact, to date NGC 4698 \citep{bertola99}, NGC 4672 \citep{sarzi00},
and UGC 10043 \citep{matthewsdegrijs04} are the only spiral galaxies
known to host a prominent bulge sticking out from the plane of the
disc.
The case of NGC~4698 is particularly intriguing since it hosts also a
polar nuclear stellar disc aligned with its polar bulge and thus
perpendicular to the main disc. This galaxy was recently revisited by
\citet{corsini12} and its intrinsic shape was derived using the
methodology proposed by \citet{mendezabreu10}. They found a slightly
triaxial polar bulge elongated along the vertical direction with axis
ratios $Q = 0.95$ and $F$ = 1.60. This result agrees well with the
observed kinematics presented in \citet{bertola99} and with a model
where the nuclear disc is the end result of the acquisition of
external gas by the pre-existing triaxial bulge on the principal plane
perpendicular to its shortest axis and perpendicular to the main disc
of the galaxy.
\section{The Intrinsic Shape of the Milky Way Bulge}
\label{sec:MW}
Owing to its vicinity, the Galactic bulge has always been targeted as
the ideal benchmark for structure, kinematic, and stellar populations
studies of bulges. In fact, it can be studied at a unique level of
detail, in comparison to external galaxies, thanks to the possibility
of measuring the properties of individual stars. However, our {\it
inside view} of the Galaxy generally restricts our knowledge to
pencil beam areas around the Galactic centre due to either the high
extinction, the crowding, or the superposition of multiple structures
along the LOS, making studies of the inner Galactic regions
challenging.
The structure of the Galaxy has accounted for a significant amount of
literature in the past and the topic has come back in the limelight in
recent years. In this section I briefly review the Galactic bulge
topic focusing on its intrinsic shape heading the readers to other
chapters in this volume for more information about its stellar content
and kinematics.
In recent decades it has become clear that the Galaxy is a barred
system \citep{blitzspergel91,lopezcorredoira05} and that most likely
its central regions are dominated by a boxy bulge created by vertical
instabilities within the Galactic bar
\citep{dwek95,martinezvalpuestagerhard11, ness13}. The historical
evolution of our knowledge of the intrinsic structure of the Galactic
bulge has been written by a succession of progressively larger scale,
deeper sensitivity photometric and spectroscopic surveys.
The first attempt to understand the shape of the Galactic bulge was
made by \citet{devaucouleurspence78}. They found that models ranging
from spherical to $F=0.6$ were able to represent well both the
distribution of globular clusters around the Galactic centre and the
infrared isophotes observed at 2.4$\mu$m \citep{maihara78}. The
flattening of the Galactic bulge was then further constrained with the
arrival of the Infrared Astronomical Satellite (IRAS). Using IRAS
data, \citet{harmongilmore88} and \citet{whitelock91} found values of
the intrinsic flattening spanning $0.6<F<0.8$ using $JHK$
near-infrared bands. Similarly, \citet{kent91} found that, at first
order, the Galactic bulge can be represented by an oblate ellipsoid
with $F=0.61$ using data from the Infrared Telescope (IRT).
The picture changed drastically with the advent of the COBE satellite
\citep{hauser90}. The new striking image of the Milky Way (Figure
\ref{fig:mwcobe}) provided by the DIRBE experiment on board of COBE
allowed \citet{blitzspergel91}, and later on \citet{blitz93}, to find
the first direct evidence for a bar at the Galactic centre.
Interestingly, they also found the presence of a triaxial bulge
structurally distinct from the main bar. The modelling of this
triaxial bulge was performed by different teams with different sets of
data in the subsequent years. Consequently, different axis ratios
represented as 1:Q:F were found: 1:0.33:0.22 \citep{dwek95}, 1:0.6:0.4
\citep{binney97},1:0.43:0.29 \citep{stanek97}, 1:0.38:0.26
\citep{freudenreich98}, 1:0.54:0.33 \citep{lopezcorredoira00},
1:(0.3--0.4):0.3 \citep{bissantzgerhard02}, 1:0.5:0.4
\citep{lopezcorredoira05}. In general, these values implied the
Galactic bulge to be a triaxial structure with a tendency to
prolateness, thus not in agreement with the triaxial/oblate picture
outlined in Section \ref{sec:bulges} for extragalactic bulges.
\begin{figure}[t]
\includegraphics[width=\textwidth]{mendezabreu_fig5.ps}
\caption{False-colour image of the near-infrared sky as seen by the
DIRBE. Data at 1.25, 2.2, and 3.5 $\mu$m wavelengths are
represented respectively as blue, green and red colours. The
image is presented in Galactic coordinates, with the plane of the
Milky Way Galaxy horizontal across the middle and the Galactic
centre at the centre. Credits: E. L. Wright (UCLA), The COBE
Project, DIRBE, NASA.}
\label{fig:mwcobe}
\end{figure}
Although the idea of a triaxial bulge worked well at first order, the
boxy shape noticed earlier by \citet{kent91} and \citet{kent92} and
confirmed by \citet{dwek95} was not recovered by a triaxial
ellipsoid. In the meanwhile, different scenarios came up to explain
these differences and account for the continuously increasing
kinematic and stellar populations information. \citet{alard01}
suggested the presence of two different bars in the Galaxy by
analysing data from the Two Micron All Sky Survey (2MASS)
\citep{skrutskie06}. Another possible scenario was worked out by
\citet{babusiaux10} suggesting a model composed by a classical bulge
in the centre and a boxy bulge in the outer parts.
\citet{shen10} proposed a simple model yet backed up by the high
quality stellar kinematics provided by the Bulge Radial Velocity Assay
(BRAVA) \citep{rich07}. Using N-body simulations they found no
evidence for a classical bulge in the Galaxy but the bulge appears to
be only part of the bar and therefore not a separated component.
Figure \ref{fig:mwshen} shows that the inclusion of a classical bulge
greatly worsens the model fit to the data. Models from \citet{shen10}
rule out that the Milky Way has a significant classical bulge with
mass $>$15\% of the disc mass.
\begin{figure}[t]
\includegraphics[width=\textwidth]{mendezabreu_fig6.ps}
\caption{Best models fits to the BRAVA stellar kinematics using
different hypothesis on the classical bulge mass. Mean velocity
(top panels) and velocity dispersion (lower panels) profiles of
all available kinematic observations presented in \citet{shen10}.
The left two panels are for the Galactic latitude b =
−4$^{\degree}$ strip; the middle two panels are for the b =
−8$^{\degree}$; and the right two panels are for the l =
0$^{\degree}$ minor axis. The heavy black lines represent the
model without a classical bulge. The red, green, and blue lines
are for models whose classical bulges have masses of 8\%, 15\%,
and 30\%, respectively, of the disk mass. Including a classical
bulge significantly worsens the model fits to the data, especially
along the minor axis. Extracted from \citet{shen10}. Reproduced
with permission, \textcircled{c} AAS.}
\label{fig:mwshen}
\end{figure}
Following this line, \citet{martinezvalpuestagerhard11} demonstrated
how the star counts measurements by \citet{cabreralavers07} agrees
with a scenario composed by a single bar and a boxy bulge. More
recent measurements of star counts from the VISTA Variables in The Via
Lactea (VVV) \citep{gonzalez11}, metallicity gradients from the
Abundances and Radial velocity Galactic Origins Survey (ARGOS)
\citep{ness13}, or stellar kinematics from BRAVA have also been
reconciled within this picture
\citep{gerhardmartinezvalpuesta12,martinezvalpuestagerhard13}.
\section{The 3D shape of bulges in numerical simulations.}
\label{sec:simulations}
The intrinsic shape of bulges keeps important information about their
formation history, with different merger, accretion and assembly
scenarios resulting in different shapes. Hence, the comparison of
measured intrinsic shapes with the output from numerical simulations
represents an intrinsic way to gain insights on their formation.
However, numerical resolution problems have often hampered these
studies and our interpretation of the shapes of bulges is usually
restricted to the analysis of simulated elliptical galaxies.
\citet{cox06} studied the structure of ellipsoidal remnants formed by
either major (equal-mass) dissipationless or dissipational mergers of
disc galaxies. They found a bimodal distribution of the triaxiality
parameter in their remnant ellipticals (see right panel in Figure
\ref{fig:bulgestriax}). Thus, dissipationless remnants are triaxial
with a tendency to be more prolate and with a mean triaxiality
parameter $T=0.55$, whereas dissipational remnants are triaxial and
tend to be much closer to oblate with triaxiality $T=0.28$. This
simulated bimodal distribution was compared by \citet{mendezabreu10}
to the triaxiality measured in their sample of 115 galaxy bulges
(Figure \ref{fig:bulgestriax}). They concluded that both major
dissipational and dissipationless mergers are required to explain the
variety of shapes found for bulges. The detailed study presented by
\citet{cox06} is consistent with previous studies of dissipationless
and dissipational mergers
\citep[e.g.,][]{barnes92,hernquist92,springel00}. However, the study
of \citet{gonzalezgarciabalcells05} they found how the degree of
triaxiality of the elliptical remnants in dissipationless mergers also
depends on the morphology of the progenitor spirals. The presence of
central bulges on the progenitor galaxies produce remnants which tend
to be more oblate whereas bulgeless progenitors lead to highly
triaxial remnants which seems inconsistent with observations.
Therefore, the comparison between simulations and observations are
still subject to the range of initial conditions explored by numerical
simulations.
On the other hand, even if the similarities between bulges and
ellipticals have prompted observers to compare the measured properties
of bulges to the properties of simulated elliptical galaxies, the
formation path of bulges is likely a more complex process involving
the interaction with other galaxy structural components
\citep{kormendykennicutt04,athanassoula05}. The recent work by
\citet{tapia14} has started to fill the gap on studies about the
intrinsic shape of galaxy bulges from numerical simulations. They
analysed a set of $N-$body simulations of intermediate and minor dry
mergers onto S0s to understand the structural and kinematic evolution
induced by the encounters. In their experiments, the progenitor
bulges are nearly spherical. The remnant bulges remain spherical as
well ($Q \sim F >$ 0.9), but exhibiting a wide range of triaxialities
($0.20 < T < 1.00$), remarking how the definition of this shape
parameter is too sensitive to nearly spherical systems. Figure
\ref{fig:bulgestapia} (second panel) shows how the axis ratios derived
from these simulations (open stars) are hardly reconcilable with the
observations (black diamonds) by \citet{mendezabreu10}. Still, the
strong triaxiality agrees with the structure of elliptical remnants
resulting from major-to-intermediate mergers \citep{cox06}.
\begin{figure}[t]
\includegraphics[width=\textwidth]{mendezabreu_fig7.ps}
\caption{Intrinsic shape of bulges and elliptical galaxies obtained
from numerical simulations. A comparison with observed bulges is
shown in the second panel. The blue and green stars in all panels
represent the bulge remnants after suffering intermediate/minor
mergers. The location of the progenitor bulges is shown with
orange stars. The elliptical remnants of major mergers with pure
exponential stellar discs (black circles) and containing 40\% of
gas (red circles) are also shown. First panel: intrinsic
ellipticity $b$ ($Q$ in this chapter) versus the intrinsic
flattening $c$ ($F$ in this chapter) Second panel: as panel 1 but
adding the observed distribution of bulges in
\citet{mendezabreu10} (black diamonds). Third and fourth panels:
triaxiality parameter as a function of the intrinsic ellipticity
and flattening. Extracted from \citet{tapia14}. Reproduced with
permission from Astronomy \& Astrophysics, \textcircled{c} ESO.}
\label{fig:bulgestapia}
\end{figure}
\section{Concluding remarks and future prospects}
\label{sec:conclusions}
I present here a review of our current understanding of the intrinsic
3D shape of galaxy bulges. The approach taken in this review is
largely observational and follows the historical development of the
field. Thus, a journey through the past and present of our knowledge
on the intrinsic shape of other galaxy ellipsoids such as elliptical
galaxies or galaxy discs was needed to put the problem in context. The
major conclusions of this review are:
\begin{itemize}
\item The observational data representing the whole population of
elliptical galaxies is consistent with a mixed model, combining
partly oblate and partly prolate galaxies, although a more likely
alternative point towards at least some fraction of the ellipticals
being triaxial ellipsoids. Triaxiality is also supported by several
photometric and kinematics properties, as well as for detailed
modelling of individual galaxies.
\item The intrinsic shape of ellipticals shows a dependence on galaxy
luminosity. Bright ellipticals are in general triaxial with a
tendency to be rounder whereas faint ellipticals are more flattened
with a tendency to be oblate ellipsoids.
\item Even if uncertainties due to the lack of number statistics have
been overcome with the advent of recent surveys, the data can still
be reproduced by a wide variety of intrinsic shape distributions.
Furthermore, a proper interpretation of the data is complicated by
the fact that the AARD and kinematic misalignments are often a
function of the radius. Therefore it is generally impossible to
characterize the full shape of a single elliptical galaxy with only
one or two parameters.
\item Galaxy discs are, in general, well represented by nearly oblate
models with $Q\sim0.9$. Their intrinsic flattening is also well
constrained to values spanning $0.2<F<0.3$.
\item The population of galaxy bulges can be modelled as slightly
triaxial ellipsoids with a tendency to be oblate. This population
has typical intrinsic flattenings of $F\sim0.65$. However,
individual galaxies can have a variety of intrinsic flattenings with
some extreme cases sticking out the plane of the disc, these are
called polar bulges.
\item The distribution of the triaxiality parameter of galaxy bulges
is strongly bimodal. This bimodality is driven by bulges with
S\'ersic index $n>2$. According to numerical simulations they can
be explained assuming a combination of major dissipational and
dissipationless mergers during their formation.
\item Despite previous findings showing a triaxial bulge in the Milky
Way, more recent studies have found that is more likely a boxy bulge
produced by the vertical instabilities of the Galactic bar. Owing to
recent kinematic measurements a classical bulge with mass $>15\%$ of
the disc mass can be ruled out.
\end{itemize}
Despite the study of the intrinsic shape of elliptical galaxies has a
long track record, our knowledge of the 3D shape of bulges is still in
its infancy. Therefore, further work on the topic is needed to fully
exploit its possibilities. A few guidelines to this future prospects
are outlined in the following:
\begin{itemize}
\item From a photometric point of view, even if new methodologies have
been developed they need to be applied to larger samples of galaxy
bulges. The number of elliptical galaxies recently analysed to
recover their intrinsic shape is several orders of magnitude larger
than the current samples of galaxy bulges. Large number statistics
have led to the discovery of important relations for ellipticals
galaxies, such as the different shapes of bright and faint
ellipticals, and similar studies can be crucial for galaxy bulges.
This is particularly relevant in the current picture of bulge
formation with a different population of classical and pseudobulges
dependent of the galaxy mass \citep{fisherdrory11}.
\item An even more promising path, already explored in elliptical
galaxies, is the use of combined information from photometric and
kinematic data. In particular, the common use of integral field
spectroscopy is now providing an exquisite detail of the stellar and
gaseous kinematics on large sample of galaxies. This wealth of
information together with the development of galaxy dynamical
modelling can provide a proper understanding of the intrinsic shape
of galaxy bulges.
\item It is doubtless that the comparison of the derived intrinsic
shape of bulges with the state-of-the-art numerical simulations is a
promising way to gain insights on the formation and evolution of
bulges. However, there is still a lack of simulations with a large
variety of initial and physical conditions interested on a
structural analysis of the different galaxy components, and in
particular, in the intrinsic shape evolution of galaxy bulges.
\item Historically, galaxy bulges were thought as single-component
objects at the centre of galaxies. This picture is now questioned
since different bulge types with different formation paths have been
found coexisting within the same galaxy \citep[see][and references
therein]{mendezabreu14}. A proper separation of different bulges
types, as well as the identification of possible unresolved nuclear
structures such as bars, rings, etc, must be accounted for to
improve our knowledge on bulge formation and evolution.
\item The study of the intrinsic shape of elliptical galaxies at high
redshift has recently suffered a boost thanks to the arrival of high
spatial resolution surveys on large fields of view \citep[see][and
references therein]{chang13}. This kind of studies can provide an
in-situ view of galaxy evolution and their application to the
intrinsic shape of bulges will be key to further progress on this
topic.
\end{itemize}
\begin{acknowledgement}
I would like to thank the editors E. Laurikainen, R.F. Peletier, and
D. Gadotti for their invitation to take part in this volume. I would
also like to thank A. de Lorenzo-C\'aceres and J. Argyle for a
careful reading of this manuscript. JMA acknowledges support from the
European Research Council Starting Grant (SEDmorph; P.I. V. Wild).
\end{acknowledgement}
\newpage
\bibliographystyle{mn2e}
|
\section{Abstract}
For $\mg$ a simple Lie algebra and $G$ its adjoint group, the Chevalley map and work of Coxeter gives a concrete description of the algebra of $G$-invariant polynomials on $\mg$ in terms of traces over various representations. Here we provide an extension of this description to $G$-invariant tensors on $\mg$, although restricted to only providing generators and only for the classical Lie algebras.
\section{Introduction}
For a group $G$ and a $\mC$-linear representation $V$ of $G$, we can look at the polynomial functions on $V$, here denoted $\mC(V)$, and then at the $G$-invariant polynomials $\mC(V)^G$. Chevalley showed that for a simple Lie algebra $\mg$ with adjoint group $G$, Cartan subalgebra $\mh$ and Weyl group $W$, we have that
\begin{equation}
\mC(\mg)^G \cong \mC(\mh)^W
\end{equation}
This map was extended by Harish-Chandra to a relation on the universal enveloping algebra of $\mg$, and was used by Kostant to provide part of his theorem on the decomposition of $\mC(\mg)$ into $G$-representations [Ko63].\\
By work of Coxeter, we know that $\mC(\mh)^W$ is a finitely-generated polynomial algebra with with a minimal generating set of $r = \dim(\mh) = \text{rk}(\mg)$ generators. The corresponding generators of $\mC(\mg)^G$ can be easily described for all $\mg$. For a representation $(V,\pi)$ of a simple Lie algebra $\mg$, we define
\begin{equation}M_V:= \pi(X_\alpha)\otimes K^{\alpha\beta}X_\beta \in End(V)\otimes \mC(\mg)\end{equation}
where the $X_\alpha$ provide a basis for $\mg$ and $K$ is the Killing form in this basis. Then $\mC(\mg)^G$ can be generated by $r$ elements of the form
\begin{equation}tr_V(M_V^k) \in \mC(\mg)\end{equation}
for various $V$ and $k$. The representations $V$ can be almost arbitrary but the set of numbers $k-1$ are fixed by $G$ and are called the exponents $e_i$ of $G$.\\
Here we provide a similar result for the free algebra on $\mg$, which we view as the tensor algebra $T(\mg) = \bigoplus_{k=0}^\infty \mg^{\otimes k}$. We show that the subalgebra $T(\mg)^G$ is generated by elements of a certain form. The set of elements we give is neither finite nor minimal, and we do not provide enough relations to furnish a presentation.
\section{Main Theorem}
For each classical Lie algebra we take the defining representation $(V,\pi)$. For $A_r$, this is one of the two $r+1$-dimensional representations; for $B_r$ this is the $2r+1$-dimensional representation; for $C_r$ and $D_r$ this is the $2r$-dimensional representation. We assume each representation comes with a basis so that $\pi$ yields matrices of the appropriate dimension.\\
Given a degree $k$ tensor $T$ in $T(\mg)^G$, we get an action of the symmetric group $\mathfrak{S}_k$ on $T$ via
\begin{equation}(\sigma.T)(X_1,\ldots,X_k) = T(X_{\sigma^{-1}(1)},\ldots,X_{\sigma^{-1}(k)})\end{equation}
We call this action a permutation of the indices.\\
We define a trace to be a tensor of the form
\begin{equation}T_{V,k}(X_1,\ldots,X_k) = tr_V(\pi(X_1)\pi(X_2)\cdots\pi(X_k))\end{equation}
Define $[-]^a_b:End(V)\rar \mC$ to take an element of $End(V)$ to the $(a,b)$ entry of the matrix corresponding to that element. We define $\pi^k$ to be a degree-$k$ $End(V)$-valued tensor that such that $[\pi^k(-)]_b^a$ takes $(X_1,X_2,\ldots,X_k)$ to
\begin{equation}[\pi(X_1)]_b^{c_1}[\pi(X_2)]_{c_1}^{c_2}\cdots[\pi(X_k)]_{c_{k-1}}^a\end{equation}
Hence $T_{V,k} = tr_V(\pi^k)$.
\begin{theorem}
For $A_r, B_r,$ and $C_r$, $T(\mg)$ is generated as a tensor algebra by traces $T_V$, with $V$ being the defining representation listed above, allowing for permutation of the indices of tensors. For $D_r$, $T(\mg)$ is generated as tensor algebra by traces $T_V$, with $V$ being the defining representation listed above, and by tensors of the form
\begin{equation}\epsilon_{a_1,\ldots,a_{2r}}\prod_{i=1}^r g^{a_i,b_i}[\pi^{k_i}]_{b_i}^{a_{r+i}}\end{equation}
allowing for permutation of the indices, where $\epsilon_{a_1,\ldots,a_{2r}}$ is the Levi-Civita tensor.
\end{theorem}
Note that in the $D_r$ case, for $k_1 = k_2 = \ldots = k_r = 1$, the extra generating term becomes the Pfaffian up to some phase convention.\\
Remark: For $A_r, C_r$ and $D_r$, the representations chosen here are not actually representations of the adjoint group $G$. Rather they are representations of some cover of $G$ in terms of which all representations of $G$ can be expressed. Given a cover $\tilde{G}$ of $G$, every representation of $G$ lifts to a representation of $\tilde{G}$, so we can pick a cover $\tilde{G}$ that is easier to work with than $G$ and consider all representations of $\tilde{G}$. For $B_r$, the adjoint group is $SO(2r+1,\mC)$, so we can use the $2r+1$-dimensional representation and discuss invariants of $G$ directly.
\section{Proof of the main theorem}
Elements of $T(\mg)^G$ are all formal sums of elements of the form
\begin{equation}N_{\alpha_1,\ldots,\alpha_n} = P_{b_1,\ldots,b_n}^{a_1,\ldots,a_n}[\pi(X_{\alpha_1})]_{a_1}^{b_1}\cdots[\pi(X_{\alpha_n})]_{a_n}^{b_n}\end{equation}
where the $a_i$ and $b_i$ are indices in some representations of $G$ and $P$ is an invariant tensor. Thus the problem comes down to determining what $P$ could be. These involve the invariants on $G$-invariant tensors on $V$. For a proof that the invariants mentioned are the only ones, see [GW09].
\subsection{$A_r$}
For $A_r$, the adjoint group is $PSL(r+1,\mC)$, and all of the representations of $PSL(r+1,\mC)$ can be written as representations of $SL(r+1,\mC)$. In turn, these representations can all be written as symmetric powers or alternating powers or some combination thereof of the $r+1$-dimensional representation $V$, so we only need to consider tensors written in terms of $V$. Thus we only need to consider $a_i$ and $b_i$ to be in $V$ or its dual $V^\vee$. We interpret all of the upper indices as being in $V$ and the lower ones as being in $V^\vee$, with the indices $\alpha_i$ being refered to only as adjoint indices.\\
$V$ contains as invariants only the Kronecker delta, the Levi-Civita tensor and its dual. Hence any invariant tensor $P$ must be built out of Kronecker deltas and Levi-Civita tensors. Each instance of the Levi-Civita tensor introduces $r+1$ lower indices; since there cannot be any free lower $V$ indices in an element of $T(\mg)^G$, each Levi-Civita tensor must be matched by an appropriate source of upper indices, i.e. a dual Levi-Civita tensor.\\
Then we have the identity:
\begin{equation}\epsilon_{a_1,\ldots,a_{r+1}}\epsilon^{b_1,\ldots,b_{r+1}} = \delta_{[a_1}^{b_1}\delta_{a_2}^{b_2}\cdots\delta_{a_{r+1}]}^{b_{r+1}}\end{equation}
which allows us to replace a Levi-Civita and dual Levi-Civita pair with Kronecker deltas. Hence we can reduce any invariant $P$ to a sum of products of Kronecker deltas.\\
Thus
\begin{equation}N_{\alpha_1,\ldots,\alpha_n} = \dsum_{\sigma \in S_n}c_\sigma \delta_{a_{\sigma(1)}}^{b_1}\ldots\delta_{a_{\sigma(n)}}^{b_n}[\pi(X_{\alpha_1})]_{a_1}^{b_1}\cdots[\pi(X_{\alpha_n})]_{a_n}^{b_n}\end{equation}
for some set of coefficients $c_\sigma$. For a given $\sigma$, the contraction of Kronecker deltas and $[\pi(X_{\alpha_i})]$ yields a trace, as desired.
\subsection{$B_r$}
For $B_r$, the adjoint group is $SO(2r+1,\mC)$, and the representations of $SO(2r+1,\mC)$ can be written in terms of the $2r+1$-dimensional representation $V$. Note that while $B_r$ has spin representations which cannot be written as symmetric or alternating powers of $V$, $SO(2r+1,\mC)$ does not have such representations; the spin representations are only representations of $Spin(2r+1,\mC)$.\\
By the same argument as for $A_r$, the invariant $P$ for $B_r$ can be written in terms of the invariants of $V$. $V$ has the Kronecker delta and the Levi-Civita tensors as $A_r$ does, but also comes with the metric $g_{a,b}$ and dual metric $g^{a,b}$. Of these, we only want the Kronecker delta to be necessary for $P$.\\
Unlike in the $A_r$ case, the dual Levi-Civita tensor is not the only source of upper indices that don't come with lower indices. We can use the dual metric $g^{a,b}$ for that. But $g^{a,b}$ gives pairs of upper indices, while each Levi-Civita tensor comes with an odd number of lower indices, so not all of the lower indices from a Levi-Civita tensor can be contracted to indices from dual metrics $g^{a,b}$. Hence there must be at least one other Levi-Civita tensor, which we can turn into a dual Levi-Civita tensor by contracting with copies of $g^{a,b}$, and then the Levi-Civita and dual Levi-Civita pair can be replaced by Kronecker deltas, as in the $A_r$ case.\\
Hence we only have Kronecker deltas and metric/dual metric tensors.\\
Each copy of the metric tensor gives us two lower indices that must be matched by upper indices coming from a dual metric tensor. Unlike in the case of Levi-Civita tensors, the only way to take a metric tensor and a dual metric tensor and replace them by Kronecker deltas is to contract them:
\begin{equation}g_{a,b}g^{b,c} = \delta_a^c = g^{c,b}g_{b,a}\end{equation}
But while $P$ must have a dual metric tensor for each metric tensor it has, the metric tensor might be contracted to a Lie algebra element rather than directly to the dual metric tensor. Consider the following:
\begin{equation}g_{a,b}[\pi(X_1)]_{c_1}^b[\pi(X_2)]_{c_2}^{c_1}\cdots[\pi(X_k)]_{c}^{c_{k-1}}g^{c,d}\cdots\end{equation}
where we know that $g^{c,d}$ must appear somewhere in this expression as otherwise we would have more lower indices than upper indices.\\
The statement that $g_{a,b}$ is invariant under the action of $B_r$ is equivalent to the following:
\begin{equation}[\pi(X)]_a^bg_{b,c} = -g_{a,b}[\pi(X)]_c^b\end{equation}
So we can swap the metric and a Lie algebra element at the cost of a sign change. Hence we can move the metric along the chain until it is contracted directly to a copy of the dual metric, and then replace both by a Kronecker delta. Since every metric and dual metric pair can be eliminated this way, we are left with only Kronecker deltas, which again leaves us with the trace.
\subsection{$C_r$}
For $C_r$, the adjoint group is $PSp(r,\mC)$, but the representations of $PSp(r,\mC)$ can be written as symmetric and antisymmetric powers of the $2r$-dimensional representation of $Sp(r,\mC)$, so we shall do so. The $2r$-dimensional representation has as invariants the Kronecker delta and the Levi-Civita tensors, as well as the symplectic form $f_{a,b}$ and its dual $f^{a,b}$. In fact, the Levi-Civita tensor can be written in terms of the symplectic form by taking $r$ copies of the symplectic form and antisymmetrizing over all of the indices. This is why there is no "special symplectic group"; keeping the symplectic form invariant automatically makes the determinant invariant.\\
So we only need to eliminate the symplectic form. The symplectic form obeys the same invariant equation as the metric did for $B_r$:
\begin{equation}[\pi(X)]_a^bf_{b,c} = -f_{a,b}[\pi(X)]_c^b\end{equation}
so we can swap the symplectic and Lie algebra elements at the cost of a sign change. Hence we can move symplectic forms along the chain until they are contracted to dual symplectic forms, and then convert such pairs to Kronecker deltas. So again we get traces.
\subsection{$D_r$}
The $D_r$ case is somewhat more involved. The adjoint group of $D_r$ is $PSO(2r,\mC)$; unlike the $B_r$ case, $SO(2r,\mC)$ has nontrivial center. The representations of $PSO(2r,\mC)$ can still be written in terms of the $2r$-dimensional representation of $SO(2r,\mC)$, and thus we write everything in terms of the $2r$-dimensional representation. The invariants are the Kronecker deltas, Levi-Civita tensors, and metric tensors, as in the $B_r$ case. Here, though, we cannot rule out the need for Levi-Civita tensors, because here each Levi-Civita tensor gives $2r$ lower indices, and hence the lower indices can be matched entirely by upper indices from the dual metric. However pairs of Levi-Civita tensors can still be replaced by Kronecker deltas, and chains involving only the metric and dual metric but not the Levi-Civita tensors can also be turned into traces.\\
So we have traces as in the $A_r, B_r$ and $C_r$ cases, as well as invariants built from single Levi-Civita tensors and metric tensors.\\
For an invariant involving a single instance of the Levi-Civita tensor, since there are no free lower indices we must again get chains leading from an index on the Levi-Civita tensor, through several Lie algebra elements and copies of the metric and dual metric tensors, and then back to the Levi-Civita tensor by a different lower index. On each chain there must be one more dual metric tensor than metric tensor since the Levi-Civita tensor provides two lower indices. We can use the invariance equation for $g_{a,b}$ to move it past Lie algebra elements until it is contracted with a dual metric tensor and then replace both by a Kronecker delta. So we are left with chains leading from an index on the Levi-Civita tensor, through several Lie algebra elements, and then through exactly one dual metric tensor before connecting to the Levi-Civita tensor again, as described in the theorem.\\
Finally, we need to show that the elements listed in the theorem are not just in $T(\mg)^{\tilde{G}}$ but in $T(\mg)^G$ proper. Here we simply note that the action of $\tilde{G}$ on $\mg$ factors through the action of $G$, and hence any element of $T(\mg)$ that is invariant under the action of $\tilde{G}$ is necessarily invariant under the action of $G$. Hence all of the elements listed in the theorem are in $T(\mg)^G$, despite being written in terms of representations of $\tilde{G}$ rather than representations of $G$.
While the proof is done in terms of representations of $\tilde{G}$, we can also consider traces over representations of $G$. Here we have the issue that the invariants of representations of $G$ are of higher order, cubic or quartic or higher, and do not have the nice relation that the Levi-Civita tensor does. For instance, the adjoint representation has the Kronecker delta, a symmetric bilinear form from the Killing form, the Levi-Civita tensor, and at least one other invariant in the form of the structure constants $f_{\alpha,\beta}^\gamma$, from $[x_\alpha,x_\beta] = f_{\alpha,\beta}^\gamma$. The structure constants obey the Jacobi identity, but the Jacobi identity only involves a specific combination of structure constants, unlike for example the Levi-Civita identity which eliminates a pair regardless of how they are related.\\
For $D_{2k}$, the use of an additional generator type is necessary regardless of what representation is used. Taking a trace and symmetrizing yields an element of $\mC(\mg)$, which by Kostant's result can be written as a polynomial in some set of generators. Given a tensor that is a tensor product of multiple traces, symmetrizing yields the product of the symmetrization of the individual traces. Hence for each generator we only have to consider elements obtained by symmetrizing a single trace rather than symmetrizing tensor products of traces. For a single trace, there is only one symmetrization, so we can distinguish these symmetrizations of single traces by degree. However for $D_{2k}$, the polynomial algebra has two generators of degree $2k$, not one, and so we would need two traces of degree $2k$ with distinct symmetrizations, which cannot occur.\\
For $r = 2k+1$, one could consider one of the spin representations of $D_r$; at least in terms of symmetrization, traces in the spin representation do yield all of the generators of the symmetric algebra, since the Pfaffian in $D_{2k+1}$ is degree $2k+1$ and thus does not overlap with any of the other generators, which all have even degree. However the spin representations yield higher-order invariants, and thus the argument used above does not necessarily hold. Hence we stay with the current description in terms of the $2r$-dimensional representation and with the additional generators.
\section{$GL(n,\mC)$ and $O(n,\mC)$}
The groups $GL(n,\mC)$ and $O(n,\mC)$ also act on $sl(n,\mC)$ and $so(n,\mC)$. $GL(n,\mC)$ and $O(2r+1,\mC)$ are just direct products of $SL(n,\mC)$ and $SO(2r+1,\mC)$ with $\mC^\times$ and $\mZ_2$, so the conjugation action doesn't change.\\
For the case of $SO(2r,\mC)$, $O(2r,\mC)$ is not a central extension, so the conjugation action by elements of the other component does change some behavior. It is still an automorphism of the Lie algebra, though.\\
Due to being an automorphism of the Lie algebra, we get that while $O(2r,\mC)$ doesn't preserve the Pfaffian, which is linear in each basis vector and thus gets a sign change when we change the sign of a single basis vector, the adjoint tensors defined using the Pfaffian are still invariant under $O(2r,\mC)$, as all of the indices in $V$ are contracted over. Hence we get that the adjoint tensor invariants of $GL(n,\mC)$ and $O(n,\mC)$ are identical to those of $SL(n,\mC)$ and $SO(n,\mC)$; despite not being an invariant of $O(2r,\mC)$, the Pfaffian is still an invariant of the Lie algebra $so(2r,\mC)$ and hence gives adjoint tensor invariants.
\section{Some relations}
The Jacobi Identity for the structure constants provide a relation for the traces as listed. In particular, for each Lie algebra the structure constant can be written as the antisymmetrization of a degree $3$ trace and the contraction of two structure constants can be written as a combination of traces in degrees $2$ and $4$. Thus the Jacobi identity, which states that a linear combination of three contractions of pairs of structure constants vanishes, can be written as the vanishing of a linear combination of traces of degree 4 and tensor products of traces in degree 2. The exact statement varies based on $\mg$.\\
Other relations come from the absence of generators of the symmetric algebra in various degrees. If a trace invariant $N$ of $\mg$ has degree $k$ and the polynomial algebra $\mC(\mg)$ has no generator in degree $k$, then the symmetrization of $N$ can be written as a product of symmetrizations of tensor invariants of lower order. For instance for $\mg$ being $B_r$ and $C_r$, there are no elements of $\mC(\mg)$ with odd degree, so the symmetrizations of odd degree traces all vanish identically.\\
Another source of relations is what one might call the Cayley-Hamilton identity for $M$. $M$ is a $\dim(V)$-dimensional square matrix, and hence if we interpret $M$ as living in $End(V)\otimes \mC(\mg)$ rather than in $End(V)\otimes T(\mg)$ then $M$ obeys the Cayley-Hamilton identity, and thus $M^{\dim(V)}$ can be written as a linear combination of $P_kM^k$ for $k < \dim(V)$, where $P_k$ is a polynomial in symmetrized traces of powers of $M$. Thus symmetrized instances of $M^{\dim(V)}$ can be replaced by lower powers of $M$ times other traces. See [RSV] for details of how symmetric traces of other degrees decompose.\\
The author does not claim that all of the relations can be generated from the ones mentioned here.
\section{Diagrams}
We can write the main theorem diagrammatically. Consider the form $V\otimes V^\vee\otimes \mg \rar \mC$ given by the adjoint of the representation $\pi: \mg \rar V\otimes V^\vee$. We denote this form by an triangle, with the thin line corresponding to the adjoint index, the point toward the index in $V^\vee$ and the flat side toward the index in $V$:
\btk
\draw[E2] (0,0)to(0,-1);
\draw[E1] (1, 0) to (-0.1,0);
\draw[E1] (0,0)to (-1,0);
\filldraw[fill = white] (-0.14,0)to(0.1,0.1)to(0.1,-0.1)to(-0.14,0);
\etk
In this notation, a contraction of a $V$ index with a $V^\vee$ index corresponds to joining the ends of the corresponding lines, so that a product of two elements in $\pi(\mg)$ as elements of $End(V)$ is written as:
\btk
\draw[E2] (0,0)to(0,-1);
\draw[E2] (-1,0)to(-1,-1);
\draw[E1] (1,0)to (-2,0);
\filldraw[fill = white] (-0.14,0)to(0.1,0.1)to(0.1,-0.1)to(-0.14,0);
\filldraw[fill = white] (-1.14,0)to(-0.9,0.1)to(-0.9,-0.1)to(-1.14,0);
\etk
A trace then becomes
\btk
\draw[E2] (0,0)to(0,-1);
\draw[E2] (-0.5,0)to(-0.5,-1);
\draw[E2] (-2,0)to(-2,-1);
\draw[E2] (-2.5,0)to(-2.5,-1);
\draw[E1] (0,0)to (-0.8,0);
\draw[E1] (-1.7,0) to (-2.5,0);
\draw[E1] (0,0) arc[radius = .5, start angle = -90, end angle = 90];
\draw[E1] (0,1) to (-2.5,1);
\draw[E1] (-2.5,0) arc[radius = 0.5, start angle = 270, end angle = 90];
\node at(-1.25,0){$\ldots$};
\node at(-1.25,-0.5){$\ldots$};
\filldraw[fill = white] (-0.14,0)to(0.1,0.1)to(0.1,-0.1)to(-0.14,0);
\filldraw[fill = white] (-0.5-0.14,0)to(-0.5+0.1,0.1)to(-0.5+0.1,-0.1)to(-0.5-0.14,0);
\filldraw[fill = white] (-2.14,0)to(-1.9,0.1)to(-1.9,-0.1)to(-2.14,0);
\filldraw[fill = white] (-0.5-2.14,0)to(-0.5-1.9,0.1)to(-0.5-1.9,-0.1)to(-0.5-2.14,0);
\etk
and the tensor product of two traces is then two such diagrams placed side by side.\\
The operation of permuting the adjoint indices is then rearranging the free ends of the thin lines. The main theorem then becomes the statement that for $\mg = A_r, B_r,C_r$, any element of $(T(\mg))^G$ can be written as a formal sum of such diagrams.\\
Note that the process of simplifying tensors to traces, we could end up with the following picture:
\btk
\draw[E1] (-1,0)to(1,0) arc[radius=0.5,start angle = 90, end angle = -90];
\draw[E1](1,-1)to(-1,-1);
\draw[E1] (2.5,0.5)to(0.5,0.5) arc[radius=0.5,start angle = 90, end angle = 270];
\draw[E1] (0.5,-0.5) to (2.5,-0.5);
\draw[E2] (0.5,-0.5) to (0.5,-2);
\draw[E2] (1,-1)to (1,-2);
\filldraw[fill = white] (1-0.14,-1)to(1.1,-1+0.1)to(1.1,-1-0.1)to(1-0.14,-1);
\filldraw[fill = white] (0.5-0.14,-0.5)to(0.5+.1,-0.5+0.1)to(0.5+.1,-0.5-0.1)to(0.5-0.14,-0.5);
\etk
which we can rewrite as
\btk
\draw[E1] (-1,0)to(0,0) arc[radius=0.5,start angle = 90, end angle = -90];
\draw[E1](0,-1)to(-1,-1);
\draw[E1] (2.5,0)to(1.5,0) arc[radius = 0.5, start angle = 90, end angle = 270];
\draw[E1](1.5,-1)to(2.5,-1);
\draw[E2](0,-1)to(1.5,-2);
\draw[E2](1.5,-1)to(0,-2);
\filldraw[fill = white] (0-0.14,-1)to(0+.1,-1+0.1)to(0+.1,-1-0.1)to(0-0.14,-1);
\filldraw[fill = white] (1.5-0.14,-1)to(1.5+.1,-1+0.1)to(1.5+.1,-1-0.1)to(1.5-0.14,-1);
\etk
So any crossings can be transferred to the thin lines, i.e. can be turned into permutation of the adjoint indices.\\
For the case of $D_r$, we use the following notation for the Levi-Civita tensor:
\btk
\filldraw[fill = black] (0,-0.1)to(0,0.1)to(3,0.1)to(3,-0.1);
\draw[E1](0.5,0)to(0.5,1);
\draw[E1](1.0,0)to(1.0,1);
\node at(1.5,.5){$\ldots$};
\draw[E1](2.0,0)to(2.0,1);
\draw[E1](2.5,0)to(2.5,1);
\etk
in accordance to [CV08]. The corresponding adjoint tensors are of the form:
\btk
\filldraw[fill = black] (0,-0.1)to(0,0.1)to(4,0.1)to(4,-0.1);
\draw[E1] (1.5,0) arc[radius= .5,start angle = 0, end angle = 180];
\node[VS]at(1.4,.3){};
\draw[E2](1,0.5)to(1,1);
\filldraw[fill = white] (1-0.14,0.5)to(1+.1,0.5+0.1)to(1+.1,0.5-0.1)to(1-0.14,0.5);
\node at(2.0,.5){$\ldots$};
\draw[E1] (3.5,0) arc[radius= .5,start angle = 0, end angle = 180];
\node[VS]at(3.4,.3){};
\draw[E2](3,0.5)to(3,1);
\filldraw[fill = white] (3-0.14,0.5)to(3+.1,0.5+0.1)to(3+.1,0.5-0.1)to(3-0.14,0.5);
\etk
where the white circles indicate the metric and the thin lines are shorthand for multiple adjoint lines.
\section{Trees}
Note that, up to a scaling, the Killing form on $\mg$ can be written as a trace:
\begin{equation}K_{\alpha\beta} = K(X_\alpha,X_\beta) = tr_V(\pi(X_\alpha)\pi(X_\beta))\end{equation}
The rule for $\pi$ being a representation is that
\begin{equation}\pi(X)\pi(Y) - \pi(Y)\pi(X) = \pi([X,Y])\end{equation}
This allows us to write the structure constants in terms of traces as well:
\begin{eqnarray}c_{\alpha\beta}^\gamma &=& K([X_\alpha,X_\beta],X_\delta)K^{\delta\gamma}\\
&=& tr_V(\pi(X_\alpha)\pi(X_\beta)\pi(X_\delta)-\pi(X_\beta)\pi(X_\alpha)\pi(X_\delta))K^{\delta\gamma}
\end{eqnarray}
It also allows us to write
\begin{equation}
\pi(X_\alpha)\pi(X_\beta) = \frac{1}{2}(\pi(X_\alpha)\pi(X_\beta)+\pi(X_\beta)\pi(X_\alpha)) + \frac{1}{2}c_{\alpha\beta}^\gamma \pi(X_\gamma)
\end{equation}
and thus any trace can be written as a symmetric trace plus lower degree traces contracted to structure constants. the previous section notes that we only get irreducible symmetric traces in certain dimensions.\\
Borrowing some of the notation from the previous section, if we write a symmetric trace with $k$ adjoint indices as a white vertex with $k$ thin lines coming out of it and a structure constant as a black vertex with $3$ thin lines coming out of it, the diagrams described above are equivalent to forests where each tree has one white vertex and some number of black vertices. By the discussion in the section on relations, the white vertices in turn must have degree $k = e_i+1$ for the exponents $e_i$ of $G$.\\
Note that for $B_r$ and $C_r$, the exponents are all odd, so the corresponding $k$ are all even. The white vertices can thus be replaced by traces over the adjoint representation rather than the defining representation $V$ using the relations described in [RSV], and the branches of black vertices can then be read as permutations of indices of traces in the adjoint representation, rather than permutations of indices of traces in the defining representation. Thus for $B_r$ and $C_r$, the adjoint tensor invariants can be expressed as tensor products of traces over the adjoint representation, with permutation of the indices. Because of the Killing form, the discussion in the section on $B_r$ also applies to the adjoint representation, so that traces of odd degree over the adjoint representation vanish identically. Hence for $A_r$ and $D_{2k+1}$, which both have even exponents, we cannot use the adjoint representation for everything, and we already noted that we need at least two representations for $D_{2k}$.
\section{Kirillov's Family Algebras and the Exceptional Cases}
In 2001 Kirillov defined a set of algebras that he calls family algebras. He considers a representation $V$ of $G$ and defines
\begin{equation}C_V(\mg) = (End(V)\otimes S(\mg))^G\end{equation}
The original intent was to compute the behavior of the $G$-harmonic polynomials of Kostant, applying algebraic methods in place of the combinatorial methods already known.\\
For the case where $G$ is classical, $V$ is a symmetric, antisymmetric, or tensor power of the adjoint representation, we can use the above theorem to describe the family algebra $C_V(\mg)$, noting that
\begin{equation}C_V(\mg) = (V\otimes V^\vee \otimes S(\mg))^G \subset T(\mg)^G\end{equation}
Hence the generators of the tensor algebra describe the possible elements of the family algebra, which can then be described in terms of generators within the family algebra (using the family algebra's multiplication structure rather than tensor product). For instance, the author has used the main theorem to compute the structure of $C_{adj}(\mg)$ for $\mg$ classical and $adj$ the adjoint representation.\\
The author has also examined $C_{adj}(\mg)$ for $\mg$ exceptional. Just as the elements of the family algebras in the classical cases could be described as tensor products of traces, or alternatively as trees in the sense of the previous section, so could the elements of the family algebras in the exceptional cases. Thus the author conjectures that the main theorem may hold for the exceptional Lie algebras, with specific choices for the representation used to define $M$:
\begin{conjecture}
For $G_2$, let $(V,\pi)$ be the $7$-dimensional representation. For $F_4$, let $(V,\pi)$ be the $26$-dimensional representation. For $E_6$ let $(V,\pi)$ be the $27$-dimensional representation. For $E_7$ let $(V,\pi)$ be the $56$-dimensional representation. For $E_8$ let $(V,\pi)$ be the $248$-dimensional representation. Defining traces as above, $T(\mg)^G$ is generated by traces, allowing for permutation of the indices.
\end{conjecture}
Unfortunately, the argument used for the classical Lie algebras does not work here. The representations of the exceptional groups carry irreducible invariants of degree $3$ and $4$ that don't reduce or cancel. In particular, $G_2$'s and $E_8$'s defining representations carry antisymmetric cubic invariants, $F_4$'s and $E_6$'s defining representations carry symmetric cubic invariants, and $E_7$'s defining representation carries a symmetric quartic invariant. Thus, although all of the representations of a given exceptional group can be embedded into the tensor powers of the defining representation of that group, it is not clear that the invariants in the defining representation can be reduced to just Kronecker deltas.\\
Note that except for $E_6$, the exceptional Lie algebras have only odd exponents, and hence, again by [RSV], we can replace symmetric traces over $V$ with symmetric traces over the adjoint representation. Like $A_r$ and $D_r$, $E_6$ requires the use of at least one other representation.
|
\section{Introduction}
Differential K-theory is an enhanced version of topological K-theory constructed
by incorporating connections and differential forms. It was developed in order to
refine the families of Atiyah-Singer
index theorem and also to classify the Ramond-Ramond field strengths in string theory
\cite{bun}, \cite{free}. One model of the first group in differential K-theory
$\widehat{K}^0(X)$ for a manifold $X$ is a Grothendieck group of vector
bundles equipped with connections and odd differential forms \cite{kar}. In \cite{SS}, Simons
and Sullivan constructed another model of $\widehat{K}^0(X)$ using just vector
bundles and connections. They got rid of the differential form at the cost of
introducing an equivalence relation among the connections on the vector bundles.
More precisely,
they constructed $\widehat{K}^0(X)$ as the Grothendieck group of ``structured''
vector bundles; the definition of a structured vector bundle is recalled in
Section \ref{sec2}.
In addition to constructing a model of $\widehat{K}^0(X)$, Simons
and Sullivan in \cite{SS} proved an interesting result about
the existence of stable inverses of hermitian structured bundles. (It is Theorem 1.15 in
\cite{SS} which is essentially Theorem \ref{main} below for the unitary group.) However,
their proof works only for the unitary group (and also to the more general
case of compact Lie
groups) because they used the existence of universal connections \emph{\'a la}
Narasimhan-Ramanan \cite{NR}. In \cite{PT} a slightly different proof of the theorem
of Simons and Sullivan was given that did not involve universal connections. In fact,
the proof in \cite{PT} is valid even for connections that are not compatible with
the metric.
All the flat connections considered here will have trivial monodromy representation.
Note that if $\nabla$ is a flat connection with trivial monodromy on a vector bundle
$V$ over $X$, and $x_0$ is a point of $X$, then there is a unique isomorphism $f$ of
$V$ with the trivial vector bundle $X\times V_{x_0}\,\longrightarrow\, X$, equipped with
the trivial connection, such that $f$ is connection preserving and coincides with
the identity map over $x_0$ (here $V_{x_0}$ denotes the fiber of $V$ over $x_0$).
In this paper we prove the following theorem.
\begin{theorem}
Let $G$ be one of the groups ${\rm GL}(N,\mathbb{C})$,~
${\rm SO}(N,\mathbb{C})$,~ ${\rm Sp}(2N, \mathbb{C})$. Given a structured vector bundle
$\mathcal{V} \,=\, [V\, , \{ \nabla \} ]$ on a smooth manifold $X$ such that the holonomy
of some equivalent connection $\nabla$ is in $G$, there exists a structured inverse
$\mathcal{W} \,=\, [W\, , \{ \widetilde{\nabla} \} ]$ with the property that
the holonomy of $\widetilde{\nabla}$ is in the same $G$, satisfying
$$\mathcal{V} \oplus \mathcal {W} \,=\, [X \times \mathbb{C}^M \, ,
\{ \nabla _F \}]$$
where $\nabla_F$ is a flat connection with trivial monodromy on the trivial
vector bundle $X \times \mathbb{C}^M$ over $X$.
\label{main}
\end{theorem}
As an immediate consequence of Theorem \ref{main} we have the following corollary.
\begin{coro}
Let $\widehat{K^0}_G(X)$ be the Grothendieck group of structured vector bundles
$\mathcal{V} \,=\, [V,\{ \nabla \}]$ satisfying the
condition that the connection $\nabla$ has holonomy in $G$
(both $G$ and $X$ are as in Theorem \ref{main}). Let $d$ denote the trivial flat connection on a trivial bundle $X\times \mathbb{C}^k$. Then the following two hold:
\begin{enumerate}
\item Every element of $\widehat{K^0}_G(X)$ is of the form $\mathcal{V} - [k]$ where $[k]
\,=\, [X \times \mathbb{C}^k, \{d\}]$ (so $k\,=\, N$ if $G\,=\, {\rm GL}(N,\mathbb{C})$
or ${\rm SO}(N,\mathbb{C})$ and $k\,=\, 2N$ if $G\,=\, {\rm Sp}(2N, \mathbb{C})$), and
\item $\mathcal{V} \,= \,\mathcal{W}$ in $\widehat{K^0}_G(X)$ if and only if $\mathcal{V} \oplus [\mathcal{N}]
\,=\, \mathcal{W} \oplus [\mathcal{N}]$ as structured bundles for some flat bundle
$[\mathcal{N}]$ with trivial monodromy.
\end{enumerate}
\label{group}
\end{coro}
In \cite{SS}, Simons and Sullivan proved what they called the Venice lemma which
essentially says that every exact form arises out of Chern character forms of trivial
bundles (see also \cite{PT}). Using the same ideas as in the proof of Theorem
\ref{main} we give a proof of the following holonomy version of the Venice lemma.
\begin{prop}\label{holovenice}
Fix $G$ to be one the groups ${\rm GL}(N,\mathbb{C})$,
${\rm SO}(N,\mathbb{C})$ and ${\rm Sp}(2N, \mathbb{C})$.
If $\eta$ is any odd smooth form on $X$, then there exists a trivial bundle $T\,=\,
X \times \mathbb{C}^k$ ($k$ is as in Corollary \ref{group})
and a connection $\nabla$ on it whose holonomy is in $G$ such that
\begin{equation}\label{veniceeq}
\mathrm{ch}(T, \nabla) - \mathrm{ch}(T, d) \,=\, d\eta\, .
\end{equation}
\end{prop}
\section{Preliminaries}\label{sec2}
As mentioned in the introduction, in order to define structured vector bundles we
need to define an equivalence relation between connections on vector bundles
on a smooth manifold. To do so, we recall the
definition of the Chern-Simons forms. Throughout, $V$ and $W$ are smooth complex
vector bundles on a smooth manifold $X$. For a connection $\nabla$ on a vector
bundle $V$, let $F_{\nabla}\,\in\, C^\infty(X,\, End(V)\otimes \bigwedge^2 T^*X)$
be the curvature of $\nabla$, and let $$\mathrm{ch} (\nabla) \,=\,
\mathrm{Tr} \exp \left(\frac{\sqrt{-1}}{2\pi} F_{\nabla}\right)$$ be the corresponding
Chern character form on $X$.
\begin{defi}
If $\nabla _1$ and $\nabla _2$ are smooth connections on $V$, then the Chern-Simons form
between them is defined as a sum of odd differential forms $\mathrm{CS}(\nabla _1, \nabla _2)$
modulo exact forms satisfying the following two conditions:
\begin{enumerate}
\item (\emph{Transgression})~\ ~$d\mathrm{CS}(\nabla _1\, , \nabla _2) \,=\,
\mathrm{ch}(\nabla _1) - \mathrm{ch}(\nabla_2)$.
\item (\emph{Functoriality})~\ ~ If $f \,:\, Y \,\longrightarrow\, X$ is a smooth map between
smooth manifolds $Y$ and $X$, then $\mathrm{CS}(f^{*} \nabla _1\, , f^{*} \nabla _2) \,=\,
f^{*} \mathrm{CS}(\nabla _1\, , \nabla _2)$ modulo exact forms.
\end{enumerate}
\label{cs}
\end{defi}
In \cite{SS} an equivalence relation between connections was defined, which we now
recall.
\begin{defi}
If $\nabla _1$ and $\nabla_2$ are two smooth connections on a vector bundle
$V$ on $X$, then $$\nabla _1 \,\thicksim\, \nabla _2$$ if
$\mathrm{CS} (\nabla _1, \nabla _2)\,=\, 0$ modulo exact forms
on $X$. The equivalence class of $\nabla$ is denoted by $\{\nabla \}$.
\label{eq}
\end{defi}
An isomorphism class $\mathcal{V} \,=\, [V\, , \{ \nabla \}]$ as in Definition
\ref{eq} is called a
\emph{structured vector bundle}. The direct sum of $\mathcal{V}\,=\,[V\, , \{ \nabla
_V \}]$ and $\mathcal{W} \,=\, [W\, , \{ \nabla_W \}]$ is defined as
$$
\mathcal{V}\oplus\mathcal{W}\,=\, [V\oplus W\, , \{ \nabla_V \oplus \nabla _W \}]\, .
$$
A \textit{symplectic} (respectively, \textit{orthogonal}) bundle on $X$ is a pair
$(E \, ,\varphi)$, where $E$ is a $C^\infty$ vector bundle on $X$ and $\varphi$
is a smooth section of $E^* \otimes E^*$, such that
\begin{enumerate}
\item{} The bilinear form on $E$ defined by $\varphi$ is anti-symmetric (respectively, symmetric), and
\item{} the homomorphism
\begin{equation}\label{e1}
E\,\longrightarrow\, E^*
\end{equation}
defined by contraction of $\varphi$ is an isomorphism, equivalently, the form
$\varphi$ is fiber-wise non-degenerate.
\end{enumerate}
A connection on a vector bundle $E$ induces a connection on $E^* \otimes E^*$. A
\textit{symplectic} (respectively, \textit{orthogonal}) connection on a symplectic
(respectively, orthogonal) bundle $(E \, ,\varphi)$ is a $C^\infty$ connection
$\nabla$ on the vector bundle $E$ such that the section $\varphi$ is parallel with
respect to the connection on $E^* \otimes E^*$ induced by $\nabla$.
If $(E \, ,\varphi)$ is a symplectic (respectively, orthogonal) bundle, then the
inverse of the isomorphism in \eqref{e1} produces a symplectic (respectively,
orthogonal) structure $\varphi'$ on $E^*$, because $(E^*)^*\,=\, E$. Let
$\nabla$ be a symplectic (respectively, orthogonal) connection on the
symplectic (respectively, orthogonal) bundle $(E \, ,\varphi)$. Then the
connection $\nabla'$ on $E^*$ induced by $\nabla$ is a symplectic
(respectively, orthogonal) connection on $(E^*, \varphi')$, where $\varphi'$
is defined above. We note that the isomorphism in \eqref{e1} takes $\varphi$ and
$\nabla$ to $\varphi'$ and $\nabla'$ respectively.
We define the holonomy version of differential K-theory next.
\begin{defi}
Let $V$ be a vector bundle with a connection $\nabla _V$ whose holonomy
is in a group $G$. Let $\{ \nabla _V \}_G$ denote the
equivalence class of all such connections and denote the corresponding
structured bundles by $\mathcal{V}$. Also, let $T_G$ denote the free group of such structured
bundles. The following group is the holonomy version of differential K-theory on a smooth
manifold $X$:
\begin{gather}
\widehat{K^0}_G (X) = \frac{T_G}{\mathcal{V}+\mathcal{W} - \mathcal{V} \oplus \mathcal{W}}
\end{gather}
\label{holodiffK}
\end{defi}
\begin{rema}
Notice that we require all the equivalent connections in $\{ \nabla_V \}_G$ to have holonomy in $G$ as a part of the definition
of equivalence. In particular, equivalent connections in the sense of \cite{SS} do not necessarily have their holonomy in the same group.
\end{rema}
From now onwards we drop the subscript $G$ whenever it is clear from the context.
\section{Proof of Theorem \ref{main}}
We divide the proof of Theorem \ref{main} into two cases.
\subsection{The case of $\mathbf{G=GL(N,\mathbb{C})}$}\label{s2.1}
\textbf{}\\
This case has already been covered in \cite{PT}. However here we provide a different proof.
Our approach relies on Lemma \ref{trivsublemma} proved below. We believe that
Lemma \ref{trivsublemma} maybe of interest in its own right. The geometrical content
of the above mentioned lemma is that on $\mathbb{R}^n$ every trivial bundle with
a connection is a subbundle, equipped with the induced connection, of a trivial bundle
equipped with a flat connection.
\begin{lemma}
Let $V$ be a trivial complex vector bundle of rank $r$ on $\mathbb{R}^n$, and let
$A$ be a connection on $V$. Then there
exists an invertible, smooth $(2n+2)r \times (2n+2)r$ complex matrix valued function $g$
such that $A_{ij}\,=\,[dg g^{-1} ] _{ij}$, where $1\,\leq\, i\, , j \,\leq\, r$.
\label{trivsublemma}
\end{lemma}
\begin{proof}
Notice that $A \,=\, \displaystyle \sum _{k=1} ^n A_k dx^k$, where $A_k$ are smooth
$r\times r$ complex matrix valued functions and $x^k$ are coordinates on $\mathbb{R}^n$.
We may write $A_k$ as
$$
A_k \,=\, 2I+A_k ^{\dag} A_k+A_k - (2I+A_k ^{\dag} A_k)\, .
$$
Using this it can be deduced that $A_k$ is a difference of two smooth functions with
values in $r\times r$ positive definite
matrices (an $r\times r$ matrix $B$ is called positive definite if
$v^{\dag} (B+B^{\dag}) v \,>\, 0 \ \forall \ v \neq 0$). Indeed, we have
$$
I+A_k ^{\dag} A_k+\frac{A_k + A_k ^{\dag}}{2} \,=\, \left(I+\frac{A_k}{2}\right)^{\dag}
\left(I+\frac{A_k}{2}\right) +\frac{3}{4} A_k ^{\dag} A_k \,\geq\, 0\, .
$$ Also, $dx^k \,=\, e^{-x^k} d(e^{x^k})$ and
$-dx^k\,=\, e^{x^k} d(e^{-x^k})$. Hence
$$A\,=\, \displaystyle \sum _{k=1} ^{2n} f_k dh_k$$ where the $h_i$ are
positive smooth functions and the $f_i$ are $r\times r$ positive-definite smooth matrix-valued functions.
We may attempt to find $g$ by forcing the first $r\times (2n+2)r$ sub-matrix of $dg$ to be
$$\left [ \begin{array}{ccccc}
dh_1 Id_{r\times r} &\ldots & dh_{2n} Id_{r\times r} & 0 & 0
\end{array} \right ]$$
and the first $(2n+2)r\times r$ sub-matrix of $g^{-1}$ to be $$\left [ \begin{array}{c} f_1 \\ f_2 \\ \vdots \\ f_{2n} \\ -\sum h_k f_k \\ Id_{r\times r} \end{array} \right ].$$ If we manage to find such a $g$, then $A_{ij} \,=\, [dg g^{-1}]_{ij}$. \\
Indeed, we claim that the matrix $g$ defined by
\begin{gather}
g=\left [ \begin{array}{cccccc}
h_1 Id_{r\times r} & h_2 Id_{r\times r} & \ldots & h_{2n}Id_{r\times r} & Id_{r\times r} & Id_{r\times r} \\
Id_{r\times r} & 0 & \ldots & 0 & f_1 \left(\displaystyle \sum_k h_k f_k\right) ^{-1} & 0 \\
0 & Id_{r\times r} & \ldots & 0 & f_2\left(\displaystyle \sum_k h_k f_k \right)^{-1} & 0 \\
\vdots & \vdots & \vdots & \vdots & \vdots & \vdots \\
0 & 0 & \ldots & Id_{r\times r} & f_{2n}\left(\displaystyle \sum_k h_k f_k\right)^{-1} & 0 \\
0 & 0 & \ldots & 0 & \left(\displaystyle \sum _k h_k f_k\right)^{-1} & Id_{r\times r} \\
\end{array} \right ]\nonumber
\end{gather}
does the job. (Here $Id_{r\times r}$ is the $r\times r$ identity matrix.) This can be verified by a
straightforward computation. Note that $\displaystyle \sum _k h_k f_k$ is invertible because $h_k>0$ and $f_k + f_k ^{\dag} >0$.
\end{proof}
Since $\mathbb{R}^n$ is simply connected, any flat bundle on it has trivial monodromy.
Lemma \ref{trivsublemma} implies the following generalization:
\begin{prop}
Let $(V\, ,\nabla \,=\, d+A)$ be a complex rank $r$ vector bundle equipped with a
connection on a smooth manifold $X$ of
dimension $n$. Then there exists a trivial
complex vector bundle $T$ of rank $(4n+8r+2)(n+2r)$ on $X$, and a
smooth flat connection with trivial monodromy
$\widetilde{\nabla}\,=\,d+\widetilde{A}$ on $T$, such that
\begin{itemize}
\item $V\oplus W \,=\, T$ for some
smooth complex vector bundle $W$, and
\item{} the connection $A$ is induced from $\widetilde{A}$.
\end{itemize}
\label{trivsub}
\end{prop}
\begin{proof}
Using the Whitney embedding theorem, there is an embedding of the total space of $V$ in $\mathbb{R}^{2n+4r}$. The zero
section of $V$ is diffeomorphic to $X$ (and hence $X$ also sits in $\mathbb{R}^{2n+4r}$). The tangent
bundle $TV$ of $V$ is a subbundle of $T\mathbb{R}^{2n+4r}\vert_V$.
By endowing $V$ with the metric induced from the Euclidean metric on $\mathbb{R}^{2n+4r}$, we
may find the orthogonal
complement $TV^{\perp}$ of $TV$.
It satisfies $$TV\vert_X\oplus TV^{\perp} \,=\, T\mathbb{R}^{2(n+2r)}\vert_X\, .$$ The vector
bundle $V$ itself may be identified with a subbundle
of $TV$. Using the induced metric we may find the orthogonal complement $V^{\perp}$ of $V$ in $TV$. Therefore, there
exists a vector bundle $U = V^{\perp} \oplus TV^{\perp}$ on $X$ such that $V\oplus U\, =\,X \times
\mathbb{C}^{n+2r}\,=\, Q$.
We may endow $U$ with some arbitrary connection $\nabla _U$. This induces the connection
$\nabla_Q \,=\, \nabla \oplus \nabla _U$ on the bundle $Q$. Using a tubular neighborhood and a
partition of unity we may extend $\nabla_Q$ from $X$ to a connection $\nabla_{\widetilde{Q}}$
on the trivial vector bundle of rank $(n+2r)$
defined on all of $\mathbb{R}^{2n+4r}$. Now we may use Lemma \ref{trivsublemma} to come up with
a vector bundle $\widetilde{T}$ of rank $(4n+8r+2)(n+2r)$
on $\mathbb{R}^{2n+4r}$, equipped with a flat connection $\nabla _{\widetilde{T}}$, such that
$\nabla_{\widetilde{Q}}$ is induced from it. Restricting our attention to $X$ we see that
the vector bundle $T\,=\,\widetilde{T}\vert_X$ equipped with the connection
$\nabla _{\widetilde{T}} \vert_X$ satisfies the conditions in the proposition.
\end{proof}
Proposition \ref{trivsub} maybe viewed as a vector bundle version of the Nash
embedding theorem because it states that every connection arises out of a flat
connection with trivial monodromy. We now state a useful lemma \cite[Lemma 1.16]{SS}.
\begin{lemma}[Simons-Sullivan]
Let $V$ and $W$ be smooth vector bundles on a smooth manifold $X$. Let $\nabla$ be a
smooth connection on the direct sum $V\oplus W$ with curvature $R$. Let $\nabla _V$ and
$\nabla _W$ be the connections on $V$ and $W$ respectively constructed from $\nabla$ using
the decomposition of $V\oplus W$. Suppose that $R _{r,s} (V) \,\subseteq\, V$ and $R_{r,s}
(W)\,\subseteq\, W$ for all tangent vectors $r,s$ at any point of $X$. Then $$CS(\nabla _V
\oplus \nabla _W , \nabla ) \,=\, 0 \ \ \ \mathrm{modulo} \ \ \mathrm{exact}\ \
\mathrm{forms}\ \ \mathrm{on}\ \ X\, .$$
\label{sul}
\end{lemma}
Lemma \ref{sul} in conjunction with lemma \ref{trivsub} implies Theorem \ref{main} in the case $G\,=\,
{\rm GL}(N, \mathbb{C})$. Indeed, given a structured bundle $\mathcal{V} \,=\,
[V, \{ \nabla _V \}$, lemma \ref{trivsub} furnishes a flat
bundle $\mathcal{T}\,=\, [T, \{ \nabla _T \}]$ such that $V$ is a subbundle of $T$ with
the connection $\nabla _V$ being induced from $\nabla _T$.
Using the natural metric on $T$ we may find an orthogonal complement $W$ to $V$ so that
$V \oplus W \,=\, T$. Endowing $W$ with the
induced connection $\nabla _W$ from $\nabla _T$ (which is flat), it is straight-forward to check
that the conditions of lemma \ref{sul} are satisfied. Let
$$\mathcal{W} \,=\, [W\, , \{ \nabla _W \}]\, .$$ Using lemma \ref{sul} we see that
$$[T, \{ \nabla_{T} \}] \,=\, [V\oplus W, \{ \nabla_{T} \}] \,=\, [V\oplus
W, \{ \nabla _V \oplus \nabla _W \}] \,=\, \mathcal{V} \oplus \mathcal{W}\, .$$
\subsection{$\mathbf{G={\mathrm {Sp}}(2N,\mathbb{C})}$ or $\mathbf{G={\mathrm {SO}}(N,\mathbb{C})}$}
\textbf{}\label{se3.2}\\
{}From Section \ref{s2.1}
we know that there is a structured inverse $\mathcal{W} \,=\, [W\, , \{ \widetilde{\nabla} \} ]$
of $(E \, ,\nabla)$. We clarify that $W$ does not necessarily have a $G$--structure. Let
$\widetilde{\nabla}'$ denote the connection on $W^*$ induced by $\widetilde{\nabla}$.
Using the natural pairing of $W$ with $W^*$, the vector bundle $W\oplus W^*$ has a
canonical $G$--structure $\varphi_0$. We note that $\widetilde{\nabla}\oplus\widetilde{\nabla}'$ is a
$G$--connection on $(W\oplus W^*\, ,\varphi_0)$.
Clearly, $(W^*\, ,\widetilde{\nabla}')$ is a structured inverse of $(E^*\, , \nabla')$.
Therefore,
$$
(E^* \oplus W\oplus W^*\, , \nabla'\oplus \widetilde{\nabla}\oplus\widetilde{\nabla}')
$$
is a structured inverse of $(E \, ,\nabla)$. The connection
$\nabla'\oplus \widetilde{\nabla}\oplus\widetilde{\nabla}'$ preserves the $G$--structure
$\varphi'\oplus \varphi_0$ on the vector bundle $E^* \oplus W\oplus W^*$.
\section{Applications}
In this section we prove Corollary \ref{group} and Proposition \ref{holovenice}.
\subsection{Proof of Corollary \ref{group}}
\begin{enumerate}
\item Any element of $\widehat{K^0}_G(X)$ is of the form $[\mathcal{V}] - [\mathcal{W}]$ by definition. Since
there exists an inverse $\mathcal{Q}$ to $\mathcal{W}$ such that $\mathcal{Q} \oplus \mathcal{W}\,=\, [k]$,
where $[k]$ is flat with trivial monodromy, we see that $[\mathcal{V}] - [\mathcal{W}]\,=\, [\mathcal{V} \oplus
\mathcal{Q}] - [k]$.
\item If $[\mathcal{V}] \,=\, [\mathcal{W}]$ in $\widehat{K^0}_G(X)$, then $\mathcal{V} \oplus \mathcal{P} \,=
\,\mathcal{W} \oplus \mathcal{P}$ for some structured
bundle $\mathcal{P}$. Let $\mathcal{E}$ be an inverse of $\mathcal{P}$. Adding $\mathcal{E}$
to both sides we see that $\mathcal{V}
\oplus [N ] \,=\, \mathcal{W} \oplus [N]$ for some flat vector bundle $N$ with trivial monodromy.
\end{enumerate}
\subsection{Proof of Proposition \ref{holovenice}}
Using the Venice lemma in \cite{PT} we see that there exists a trivial bundle
$\widetilde{T}$ with a connection $\nabla_{\widetilde{T}} \,=\, d+A$ such that
$$\frac{d\eta}{2} \,=\,
\mathrm{ch}(\widetilde{T},\nabla _{\widetilde{T}}) - \mathrm{ch} (\widetilde{T},d)\, .$$ It is
not necessarily the case that the holonomy of $\nabla_{\widetilde{T}}$ lies in $G$. However, we
know that $$\frac{d\eta}{2} \,=\,\mathrm{ch}(\widetilde{T}^{*},\nabla _{\widetilde{T}^{*}})
- \mathrm{ch} (\widetilde{T}^{*},d)$$ because $d\eta$ is an even form. Therefore,
$$d\eta \,=\,\mathrm{ch}(\widetilde{T}^{*}\oplus \widetilde{T},\nabla _{\widetilde{T}^{*}}
\oplus \nabla _{\widetilde{T}}) - \mathrm{ch}
(\widetilde{T^{*}}\oplus \widetilde{T},d)\, .$$ Using the same reasoning as in Section
\ref{se3.2} we obtain the desired result.
\section*{Acknowledgements}
We are grateful to the referee for detailed comments to improve the exposition.
The first--named author acknowledges the support of a J. C. Bose Fellowship.
|
\section{Introduction}
In this paper we consider the Cauchy problem of the focusing fractional nonlinear Schr\"odinger equations:
\begin{align}\label{main eqn}\left\{\begin{array}{l}
i\partial_tu = |\nabla|^\alpha u - V(u) u,\;\;
\mbox{in}\;\;\mathbb{R}^{1+d} \;\;d \ge 2,\\
u(x,0) = \varphi(x)\;\; \mbox{in}\;\;\mathbb{R}^d,
\end{array}\right.
\end{align}
where $$V(u) = \left\{\begin{array}{l} |u|^\frac{2\alpha}{d-\alpha}\;\; (1 < \alpha < 2),\\(|x|^{-2\alpha}*|u|^2)\;\; (1 < \alpha < \min(2, \frac d2)).\end{array}\right.$$ The equation \eqref{main eqn} is of $\dot{H}^\frac\al2$-scaling invariance (so-called energy-critical). That is, if $u$ is a solution
of \eqref{main eqn}, then for any $\lambda > 0$ the scaled function $u_\lambda$, given by
$$
u_\lambda (t, x) = \lambda^{\frac d2-\frac\al2} u(\lambda^\alpha \,t, \lambda x ),
$$
is also a solution to \eqref{main eqn}.
The problem \eqref{main eqn} can be easily shown to be well-posedness in $C((-T_*, T^*); H_{rad}^\frac\al2)$ for $\alpha \in [\frac{2d}{2d-1}, 2)$ and $d > \alpha$ in the case of power type ($d > 2\alpha$ in the case of Hartree type) through the radial Strichartz estimate. See Lemma \ref{str} below for Strichartz estimate, and also see Theorem 4.10 of \cite{guwa} and Theorem 5.2 of \cite{chho} for LWP and small data GWP. Here $-T_*, T^* \in [-\infty, +\infty]$ are the maximal existence times and $X_{rad}$ denotes the Banach space $X$ of radially symmetric functions. The solution $u$ satisfies the mass and energy conservation laws: for $t \in (-T_*, T^*)$
\begin{align}\begin{aligned}\label{consv}
&\qquad m(u(t)) = \|u(t)\|^2_{L^2} = m(\varphi), \\
&E(u(t)) = \mathcal{K}(u(t)) + \mathcal{V}(u(t)) = E(\varphi),
\end{aligned}\end{align}
where $$\mathcal K(u) = \frac12 \int | |\nabla|^\frac\al2 u(x)|^2\,dx, \;\;\mathcal V(u) = - \frac1{\mu} \int V(u)|u|^2 dx,$$
$\mu = \frac{2d}{d-\alpha}$ for power type and $\mu = 4$ for Hartree type.
At this point due to the dependency on the profile which is the critical nature of \eqref{main eqn} we do not know that $\limsup_{t \to T^*}\||\nabla|^\frac\al2 u(t)\|_{L^2}$ is infinity or not. In this paper we want to address some energy concentration phenomena for both cases. We first consider the concentration in the case of unconfined kinetic energy.
\begin{theorem}[Unconfined case]\label{conc-infty}
Let $d > \alpha+1$ in the case of power type ($d > 2\alpha$ in the case of Hartree type). Assume that $\varphi \in H^\frac\al2_{rad}$ and $u$ is the unique solution to \eqref{main eqn} in $C([0, T^*); H_{rad}^\frac\al2)$ with
$$
\limsup_{t \to T^*}\||\nabla|^\frac\al2 u(t)\|_{L^2} = +\infty.
$$
Then for any $R > 0$ we have
$$
\limsup_{t\to T^*}\||\nabla|^\frac\al2 u(t)\|_{L^2(|x| \le R)} = +\infty.
$$
Moreover, if for $t < T^*$ $u(t) \in L^\infty$, then
$$
\limsup_{t \to T^*}\|u(t)\|_{L^\infty(|x|\le R)} = +\infty.
$$
The same result holds near $-T_*$.
\end{theorem}
Next we deal with the confined case for which it is necessary to implement subtle estimate associated with ground state. The ground state of \eqref{main eqn} plays an important role. It is a unique positive radial solution of
\begin{align}\label{ell eqn}
|\nabla|^\alpha W - V(W) W.
\end{align}
In \cite{chelo, lieb, dass} the authors showed that any solution of the elliptic equation \eqref{ell eqn} is a constant multiple, dilation and translation of the function $W_\alpha(x) = C_1(1 + C_2|x|^2)^{-\frac{d-\alpha}2}$ which is in $\dot H^\frac\al2$ for $0 < \alpha < \frac d2$, where $C_1, C_2$ depend on $d, \alpha$. The solution $W_\alpha$ is closely related to the best constant problem of the inequality
$$
\int V(u)|u|^2\,dx \le C_{d, \alpha} \||\nabla|^\frac\al2 u\|_{L^2}^\mu.
$$
Indeed, the maximizer $u \neq 0$ of the above inequality, that is,
\begin{align}\label{bc}
\int V(u)|u|^2\,dx = C_{d, \alpha} \||\nabla|^\frac\al2 u\|_{L^2}^\mu,
\end{align}
is characterized as $u = e^{i\theta}\lambda^\frac{d-\alpha}{2}W_\alpha(\lambda(x-x_0))$ for some $\theta \in [-\pi, \pi]$, $\lambda > 0$ and $x_0 \in \mathbb R^d$. See \cite{au, tal, cota} for power type. We will treat this problem for Hartree type in the appendix. Since $W_\alpha$ is a solution of \eqref{ell eqn}, $\||\nabla|^\frac\al2 W_\alpha\|_{L^2}^2 = \int V(W_\alpha)|W_\alpha|^2\,dx$. Thus $\||\nabla|^\frac\al2 W_\alpha\|_{L^2}^2 = C_{d,\alpha}^{-\frac2{\mu-2}}$.
Let us denote by $S_\al(I)$ for an interval $I$ the spaces $L_{I, x}^\frac{2(d+\alpha)}{d-\alpha}$ for power type and $L_{I}^6L_x^\frac{2d}{d-\frac{4\alpha}{3}}$ for Hartree type.
Then Lemma \ref{loc} below states that LWP of \eqref{main eqn} in $\dot H^\frac\al2$ holds for some $\alpha$ and $d$ and also shows the blowup criterion that $\|u\|_{S_\alpha((-T_*, 0])} = +\infty$ and $\|u\|_{S_\alpha([0, T^*))} = +\infty$ when $T_*, T^* < +\infty$.
Furthermore by following the arguments in \cite{keme, kv, mgz, gswz} with profile decomposition developed in Section 3 below one can readily get the following theorem.
\begin{theorem}\label{gwp}
Let $d \ge 2$, $\frac{2d}{2d-1} < \alpha < 2$, $\alpha < d \le 2\alpha$ for power type ($d > 2\alpha$ for Hartree type) and let $\varphi \in \dot H^\frac\al2_{rad}$. Assume that
$$
\sup_{t \in (-T_*, T^*)}\||\nabla|^\frac\al2 u(t)\|_{L^2} < \||\nabla|^\frac\al2 W_\alpha\|_{L^2}.
$$
Then $T_*, T^* = +\infty$ and $\|u\|_{S_\alpha(\mathbb R)} < +\infty$.
\end{theorem}
As a corollary one can show that $T_* = T^* = +\infty$ and $\|u\|_{S_\alpha(\mathbb R)} < +\infty$ if $E(\varphi) < E(W_\alpha)$ and $\||\nabla|^\frac\al2 \varphi\|_{L^2} < \||\nabla|^\frac\al2 W_\alpha\|_{L^2}$. The same result also holds for the defocusing case. The restriction $\alpha \in (\frac{2d}{2d-1}, 2)$ comes from the optimal range of Strichartz estimates (see Lemma \ref{str}). The condition $\alpha \le 2\alpha$ for power type is necessary to estimate perturbation like $\||\sum_j^J f_j|^\frac{2\alpha}{d-\alpha}(\sum_j^J f_j) - \sum_j^J |f_j|^\frac{2\alpha}{d-\alpha}f_j\|_{\dot H^\frac\al2}$. For this see the arguments below \eqref{app-norm}.
At this point one may expect the sharpness of Theorem \ref{gwp} and the blowup ($\|u\|_{S_\alpha((-T_*, T^*))} = +\infty$) when $E(\varphi) < E(W_\alpha)$ and $\||\nabla|^\frac\al2 \varphi\|_{L^2} \ge \||\nabla|^\frac\al2 W_\alpha\|_{L^2}$. Unfortunately we do not know the complete answers. We think this is just a technical problem due to non-locality arising when treating $|\nabla|^\alpha$. However, in case when kinetic energy is confined we can show the energy concentration near the maximal existence time and also find some class of initial data guaranteeing the finite time blowup. We first introduce the energy concentration.
\begin{theorem}[Confined case]\label{energy-conc}
Let $d \ge 2$, $\frac{2d}{2d-1} < \alpha < 2$, $\alpha < d \le 2\alpha$ for power type ($d > 2\alpha$ for Hartree type) and let $\varphi \in \dot H^\frac\al2_{rad}$. Assume that
$$
\|u\|_{S_\alpha([0, T^*))} = +\infty,\quad \sup_{t \in [0, T^*)}\||\nabla|^\frac\al2 u(t)\|_{L^2} < +\infty.
$$
If $T^*$ is finite, then there exists a sequence $t_n \to T^*$ such that for any sequence $R_n \in (0, \infty)$ obeying $(T^*-t_n)^{-\frac1\alpha}R_n \to \infty$,
$$
\limsup_{n \to \infty} \int_{|x| \le R_n} ||\nabla|^\frac\al2 u(t_n, x)|^2\,dx \ge \||\nabla|^\frac\al2W_\alpha\|_{L^2}^2.
$$
The same result also holds near $-T_*$ if $T_* < +\infty$.
\end{theorem}
The Sch\"{o}dinger case was treated by Killip and Visan in \cite{kv}. Here we adapt their arguments to fractional case with nonlinear profile approximation.
We want to mention that due to the lack of pseudo-conformal symmetry of the equation \eqref{main eqn} we could not get the similar result when the solution blowup at time infinity.
From now on we try to demonstrate some evidence of the finite time blowup. Based on the virial argument the finite time blowup was shown for mass-critical Hartree type fractional Schr\"{o}dinger equations in \cite{chkl} and for fourth order power type NLS \cite{cow}, where the mass-critical nature and radial symmetry are playing a crucial role in the proof of blowup. Those arguments cannot be applied to the power type mass-critical fractional NLS because of the lack of enough cancelation property of nonlinearity for virial argument to proceed. Since we do not know whether the kinetic energy is confined, it is hard to apply them to energy subcritical and mass supercritical or energy critical problem. However, if we are involved in energy critical problem and the energy is confined, then by using Sobolev inequality for radial functions \cite{chooz-ccm} it is plausible to establish the virial argument to get finite time blowup for both power type and Hartree type. The following theorem leads us off the finite time blowup.
\begin{theorem}\label{blow}
Let $\varphi \in H_{rad}^\frac\al2$ and $u$ be the unique solution of \eqref{main eqn} in $C([0,T^*); H_{rad}^\frac\al2)$ for the maximal existence time $T^* \in (0, +\infty]$. Suppose that $d \ge 2$, $\frac{4}{3} \le \alpha < 2$, $\alpha < d < 3\alpha$ for power type and $d > 2\alpha+2$, $\frac{2d}{2d-1} < \alpha < 2$ for Hartree type. Then for any $\varphi$ satisfying that
\begin{align}
&\quad\; \||x|\sqrt{1-\Delta} \varphi\|_{L^2} + \||x|^2\varphi\|_{L^2} < + \infty,\label{ass-mo}\\
&E(\varphi) < E(W_\alpha),\quad \||\nabla|^\frac\al2\varphi\|_{L^2} \ge \||\nabla|^\frac\al2 W_\alpha\|_{L^2},\label{ass}
\end{align}
if $\sup\limits_{0 \le t < T^*}\||\nabla|^\frac\al2u(t)\|_{L^2} < +\infty$, then $T^* < \infty$.
\end{theorem}
The rest of paper is organized as follows: In Section 2 we gather some preliminary lemmas necessary for the proof of confined energy concentration including the profile decomposition in energy space. In Section 3 we show the energy concentration, Theorems \ref{conc-infty} and \ref{energy-conc}. Section 4 is devoted to proving finite time blowup under energy confinement. In the last section we consider the best constant problem \eqref{bc} for Hartree equation.
\subsection*{Notations}
We will use the notations:\\
$\bullet$ $|\nabla| = \sqrt{-\Delta}$, $\dot H_r^s =
|\nabla|^{-s}L^r$, $\dot H^s=\dot H_2^s$, $H_r^s = (1 -
\Delta)^{-s/2} L^r$, $H^s = H_2^s$, $L^r = L_x^r(\mathbb R^d)$ for some $s \in \mathbb R$ and $1 \le r \le \infty$.\\
$\bullet$ We use the following mixed-norm notations $L_I^qL^r = L_t^q(I; L_x^r(\mathbb R^d))$, $L_{I, x}^q = L_I^qL^q$ and $L_t^qL^r = L_{\mathbb R}^qL^r$.\\
$\bullet$ $\widehat f(\xi) = \int_{\mathbb R^d} e^{-ix\cdot \xi}f(x)\,dx$.\\
$\bullet$ For any dyadic number $N$ we denote frequency localization of function $f$ by $f_N$, which is defined by $\widehat{f_N}(\xi) = \widehat{P_Nf}(\xi) = \beta(\xi/N)\widehat{f}$ for a fixed Littlewood-Paley function $\beta \in C_{0, rad}^\infty$ with $\beta \widetilde \beta = \beta$ and $P_N\widetilde P_N = P_N$, where $\widetilde \beta(\xi) = \beta(\xi/2) + \beta(\xi) + \beta(2\xi)$ and $\widetilde P_N = P_{N/2} + P_N + P_{2N}$.\\
$\bullet$ As usual different positive
constants are denoted by the same letter $C$, if not specified. \\
$\bullet$ $[A, B]$ denotes the commutator $AB - BA$ for any operators $A$ and $B$ defined on suitable Banach spaces.\\
$\bullet$ $\big<u, v\big> = \int_{\mathbb R^d} u \,\overline v\, dx$ and $\big<f\,;\, g \big> = \sum_{1 \le j \le d}\big<f_j, g_j \big>$ for $f = (f_1,\cdots, f_d), g = (g_1, \cdots, g_d)$.
\section{Preliminary lemmas}
We define the linear propagator $U(t)$ of the linear equation $iu_t = |\nabla|^\alpha u$ with initial datum $f$. Then it is
formally given by
\begin{align}\label{int eqn}
U (t)f = e^{-it|\nabla|^\alpha} f = \frac1{(2\pi)^d} \int_{\mathbb{R}^d} e^{i(x\cdot \xi - t|\xi|^\alpha)}\widehat{f}(\xi)\,d\xi.
\end{align}
We have Strichartz estimates for radial functions (see \cite{cholee} and \cite{guwa, ke}) as follows.
\begin{lemma}\label{str}
Suppose that $d \ge 2$, $\frac{2d}{2d-1} \le \alpha < 2$ and $f, F$ are radial. Then there hold
$$
\|U(t)f\|_{L_I^qL_x^r} \le C\|f\|_{L^2}, \quad \|\int_0^t U(t-t')F(t')\,dt'\|_{L_I^qL_x^r} \le C\|F\|_{L_I^{\widetilde q'}L_x^{\widetilde r'}}
$$
for the pairs $(q, r)$ and $(\widetilde q, \widetilde r)$ such that
$$
\frac\alpha{q} + \frac{d}r = \frac d2,\;\; 2 \le q, r \le \infty,\;\; (q, r) \neq (2 , \frac{4d-2}{2d-3}).
$$
\end{lemma}
\noindent Such pairs are said to be $\alpha$-admissible.
Then we have the following inverse Strichartz estimate.
\begin{lemma}\label{inverse str}
Fix $d \ge 2$ for power type and $d > 2\alpha$ for Hartree type. Let $f \in \dot H_{rad}^\frac\al2$ and $\eta > 0$ such that
$$
\|U(t)f\|_{S_\al(I)} \ge \eta
$$
for some interval $I \subset \mathbb R$. Then there exists $\widetilde C = \widetilde C(\||\nabla|^\frac\al2 f\|_{L^2}, \eta)$, and $J \subset I$ so that
$$
\int_{|x| \le \widetilde C|J|^\frac1\alpha} \Big|U(t)|\nabla|^\frac\al2 f \Big|^2\,dx \ge \widetilde{C}^{-1}\;\;\mbox{for all}\;\;t \in J.
$$
Here $\widetilde C$ does not depend on $I$ or $J$.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{inverse str}]
For simplicity we only consider the Hartree type, the power type can be treated similarly to \cite{kv}. We will show that
\begin{align}\label{choice m}
\|U(t)f_M\|_{S_\al(I)} \ge C^{-1}\eta^\frac{3d}{4\alpha}\||\nabla|^{\frac\al2}f\|_{L^2}^{1-\frac{3d}{4\alpha}}
\end{align}
for some dyadic $M \ge A|I|^{-\frac1\alpha}$ and some (depending on $d$ and $\alpha$). Here $A = C\||\nabla|^\frac\al2 f\|_{L^2}^{-D_1}\eta^{D_2}$ and the constants $D_1, D_2$ will be specified later.
We assume that \eqref{choice m} is true. By Strichartz estimate we have
$$
\|U(t)f_M\|_{L_I^\frac{6(d-\alpha)}{d} L^\frac{2(d-\alpha)}{d-\frac{4\alpha}{3}}} \le C M^{-\frac\al2}\||\nabla|^\frac\al2 f\|_{L^2}.
$$
Combining this with \eqref{choice m}, we get by H\"{o}lder's inequality that
$$
\|U(t)f_M\|_{L_{I,x}^\infty} \ge C^{-1}\||\nabla|^\frac\al2 f\|_{L^2}^{1-\frac{3d^2}{4\alpha^2}} \eta^{\frac{3d^2}{4\alpha^2}} M^{\frac{d-\alpha}{2}}.
$$
From this with the fact that the kernel of $M^\frac\al2 |\nabla|^{-\frac\al2} \widetilde P_M$ is integrable and its value is independent of $M$ we deduce that
$$
C^{-1}\||\nabla|^\frac\al2 f\|_{L^2}^{1-\frac{3d^2}{4\alpha^2}}\eta^{\frac{3d^2}{4\alpha^2}} M^{\frac{d}{2}} \le \|(M^\frac\al2|\nabla|^{-\frac\al2}\widetilde P_M)(U(t)|\nabla|^\frac\al2 f_M)\|_{L_{I, x}^\infty} \le C\|U(t)|\nabla|^\frac\al2 f_M\|_{L_{I, x}^\infty}.
$$
Thus there exist $(t_0, x_0) \in I \times \mathbb R^d$ so that
\begin{align}\label{lower}
|(U(t_0)|\nabla|^\frac\al2f_M(x_0)| \ge A_0M^\frac d2,
\end{align}
where $A_0 = C^{-1}\||\nabla|^\frac\al2 f\|_{L^2}^{1-\frac{3d^2}{4\alpha^2}}\eta^{\frac{3d^2}{4\alpha^2}}$. Let $A_1 = \frac{A_0}{2C\||\nabla|^\frac\al2 f\|_{L^2}}$. Then
for $|x-x_0| \le A_1M^{-1} $ and $|t-t_0| \le A_1M^{-\alpha}$ we have
\begin{align*}
|U(t_0)|\nabla|^\frac\al2 f_M(x_0) - U(t) |\nabla|^\frac\al2 f_M(x)| \le \frac12A_0 M^\frac d2
\end{align*}
and thus
\begin{align*}
|U(t) |\nabla|^\frac\al2 f_M(x)| \ge \frac12A_0 M^\frac d2.
\end{align*}
This yields for all $t \in J = \{t \in I : |t-t_0| \le A_1M^{-\alpha}\}$
$$
\int_{|x-x_0| \le A_1M^{-1}} |U(t) |\nabla|^\frac\al2 f_M(x)|^2 \,dx \ge \frac{s_d}4 A_0^2A_1^d,
$$
where $s_d$ is the measure of the unit ball.
By convexity we have $$|U(t)|\nabla|^\frac\al2 f_M|^2 \le C_0|U(t)|\nabla|^\frac\al2 f|^2 * \beta^*_M,$$ where $\beta^*_M(x) = M^d |\widehat \beta(Mx)|$ and $C_0 = \int \beta^* \,dx$. And also
\begin{align*}
\int_{|x-x_0| \le A_1M^{-1}} |U(t) |\nabla|^\frac\al2 f_M(x)|^2 \,dx \le C_0\big\langle |U(t)|\nabla|^\frac\al2 f|^2, \beta_M^* * \chi_{\{|x-x_0| \le A_1M^{-1}\}} \big\rangle.
\end{align*}
We divide inner product into two parts as follows:
$$
\big\langle |U(t)|\nabla|^\frac\al2 f|^2, \beta_M^* * \chi_{\{|x-x_0| \le A_1M^{-1}\}} \big\rangle \le I + II,
$$
where
\begin{align*}
&I = \big\langle |U(t)|\nabla|^\frac\al2 f|^2, \chi_{\{|x-x_0| \le A_2A_1M^{-1}\}}\beta_M^* * \chi_{\{|x-x_0| \le A_1M^{-1}\}} \big\rangle,\\
&II = \big\langle |U(t)|\nabla|^\frac\al2 f|^2, \chi_{\{|x-x_0| > A_2A_1M^{-1}\}}\beta_M^* * \chi_{\{|x-x_0| \le A_1M^{-1}\}} \big\rangle.
\end{align*}
Now we can find a constant $A_2 = A_2(\||\nabla|^\frac\al2f\|_{L^2}, \eta) > 1$\footnote{We may choose $A_2$ as $A_2 > 1 + \max(A_1^{-1}, \frac{C_0C_\beta A_1^{d-2}}{8A_0^2})$, where $C_\beta$ is the constant satisfying $\beta^*(x) \le C_\beta(1 + |x|)^{-1}$.} such that $C_0II \le \frac{s_d}{8}A_0^2A_1^d$. Then $I \le C_0\int_{|x-x_0| \le A_2 A_1M^{-1}} |U(t) |\nabla|^\frac\al2 f(x)|^2 \,dx$ and thus
we obtain
\begin{align}\label{lower1}
\int_{|x-x_0| \le A_2 A_1M^{-1}} |U(t) |\nabla|^\frac\al2 f(x)|^2 \,dx \ge \frac{s_d}{8C_0^2}A_0^2A_1^d.
\end{align}
On the other hand, since $f$ is radial, we use the Sobolev inequality \cite{chooz-ccm} that
\begin{align}\label{sobo-radial}
\sup_{x\in \mathbb R^d}|x|^{\frac{d-\alpha}2}|f(x)| \le C\||\nabla|^\frac\al2 f\|_{L^2}\;\;\mbox{a.e.}
\end{align}
together with \eqref{lower} to get
$$
A_0(M|x_0|)^\frac{d-\alpha}2 \le |x_0|^\frac{d-\alpha}2M^{-\frac\al2}|U(t_0)|\nabla|^{\frac\al2}f_M(x_0)| \le C\||\nabla|^{\frac\al2} f\|_{L^2},
$$
which means that
\begin{align}\label{radius}
|x_0| \le A_0^{-\frac{2}{d-\alpha}}M^{-1}.
\end{align}
Combining \eqref{radius} with \eqref{lower1}, we have
$$
\int_{|x| \le (A_0^{-\frac2{d-\alpha}} + A_2A_1)M^{-1}} |U(t) |\nabla|^\frac\al2 f(x)|^2 \,dx \ge \frac{s_d}{8C_0^2}A_0^2A_1^d.
$$
Since $|I| \ge A^\alpha M^{-\alpha}$, we deduce that $|J| \ge \frac1{2}\min(A^{\alpha}, A_1)M^{-\alpha}$. By letting $\widetilde{C} = \max(2(A_0^{-\frac2{d-\alpha}} + A_1A_2)\max(A^{-1}, A_1^{-\frac1\alpha}), \frac{8C_0^2}{s_d A_0^2A_1^d})$ we get the desired result.
Now we show \eqref{choice m}.
By Littlewood-Paley theory and H\"older's inequality we have
\begin{align*}
\eta^6 &\le \|U(t)f\|_{S_\al(I)}^6\\
& \le C\int_I \left(\int(\sum_N |U(t)f_N|^2)^\frac{d}{d-4\alpha/3}\,dx\right)^\frac{3d-4\alpha}{d}\,dt \\
&\le C\sum_{M\le N}\int_I\left( \int|U(t)f_M|^\frac{d}{d-4\alpha/3} |U(t)f_N|^\frac{d}{d-4\alpha/3}\,dx\right)^\frac{3d-4\alpha}{d}\,dt\\
&\le C\sum_{M\le N}\int_I \|U(t)f_M\|_{L^\frac{2d}{d-2\alpha}}^\frac{3d-4\alpha}{d}\|U(t)f_M\|_{L^\frac{2d}{d-4\alpha/3}}^\frac{4\alpha}{d}
\|U(t)f_N\|_{L^\frac{2d}{d-4\alpha/3}}^\frac{4\alpha}{d}\|U(t)f_N\|_{L^\frac{2d}{d-2\alpha/3}}^\frac{3d-4\alpha}{d}\,dt\\
&\le C\sum_{M\le N} (\frac{M}{N})^\frac{\alpha(3d-4\alpha)}{3d}(\||\nabla|^\frac\al2 f_M\|_{L^2}\||\nabla|^\frac\al2f_N\|_{L^2})^\frac{3d-4\alpha}{d}(\|U(t)f_M\|_{S_\al(I)} \|U(t)f_N\|_{S_\al(I)})^\frac{4\alpha}{d}\\
&\le C\sup_{M}\|U(t)f_M\|_{S_\al(I)}^\frac{8\alpha}{d}\left(\sum_{M\le N} (\frac{M}{N})^\frac\al3 \||\nabla|^\frac\al2f_M\|_{L^2}\||\nabla|^\frac\al2f_N\|_{L^2}\right)^\frac{3d-4\alpha}{d}\\
&\le C\sup_{M}\|U(t)f_M\|_{S_\al(I)}^\frac{8\alpha}{d}\||\nabla|^\frac\al2f\|_{L^2}^\frac{6d-8\alpha}{d}.
\end{align*}
From the Sobolev inequality it follows that
$$
\|U(t)f_M\|_{S_\al(I)} \le C(|I|M^\alpha)^\frac16\||\nabla|^\frac\al2f\|_{L^2}.
$$
Thus we conclude that there exists $M \ge A |I|^{-\frac1\alpha}$ so that \eqref{choice m} holds with $A = C\||\nabla|^\frac\al2 f\|_{L^2}^{-\frac{9d}{2\alpha^2}}\eta^\frac{9d}{2\alpha^2}$
\end{proof}
Next we introduce the tightness of trajectories of solution. The proof is almost same as the one of Proposition 2.13 in \cite{kv} and so we omit it.
\begin{lemma}\label{tightness}
Let $v : I \times \mathbb R^d \to \mathbb C$ be a radial solution to \eqref{main eqn} with $\|v\|_{S_\al(I)} < \infty$. Suppose that
$$
\int_{|x| \le r_k} |U(t_k)|\nabla|^\frac\al2 v(\tau_k)|^2\,dx \ge \varepsilon
$$
for some $\varepsilon > 0$, $r_k > 0$, and bounded sequences $t_k \in \mathbb R$ and $\tau_k \in I$. Then
$$
\Big| \||\nabla|^\frac\al2 v(\tau_k)\|_{L^2}^2 - \int_{|x| \le a_k r_k}|U(t_k)|\nabla|^\frac\al2v(\tau_k)|^2\,dx \Big| \to 0
$$
for any sequence $a_k \to +\infty$.
\end{lemma}
We close this section by introducing local well-posedness and stability. Since the proof is quite standard, we omit the details (for instance see \cite{keme, chho}).
\begin{lemma}\label{loc}
Let $\alpha \in (\frac{2d}{2d-1}, 2)$ and $\alpha < d < 3\alpha$ for power type ($d > 2\alpha$ for Hartree type), and let $\varphi \in \dot H_{rad}^\frac\al2$, $\|\varphi\|_{\dot H^\frac\al2} \le A$. Then there exists $\delta = \delta(A)$ such that if $\|U(t-t_0)\varphi\|_{S_\al(I)} \le \delta$, $t_0 \in I$, there exists a unique solution $u \in C(I; \dot H_{rad}^\frac\al2)$ to \eqref{main eqn} with
$$
\sup_I\|u(t)\|_{\dot H^\frac\al2} + \||\nabla|^\frac\al2 u\|_{X_\al(I)} \le C(A),\quad \|u\|_{S_\al(I)} \le 2\delta.
$$
Here $X_\al(I) = L_I^\frac{2(d+\alpha)}{d-\alpha}L^\frac{2d(d+\alpha)}{d^2+\alpha^2}$ for power type and $L_I^6L^\frac{2d}{d-\alpha/3}$ for Hartree type. Moreover, $\varphi \mapsto u \in C(I; \dot H^\frac\al2)$ is Lipschitz. If $A$ is sufficiently small, then $I = \mathbb R$.
\end{lemma}
\begin{lemma}\label{stab}
Assume that $\alpha \in (\frac{2d}{2d-1}, 2)$ and $\alpha < d < 3\alpha$ for power type ($d > 2\alpha$ for Hartree type).
Let $I = [0, L), L \le +\infty$, and let $\widetilde u$ be radial and defined on $\mathbb R^d \times I$ be such that
$$
\sup_{t \in I}\|\widetilde u(t)\|_{\dot H^\frac\al2} \le A, \;\;\|\widetilde u\|_{S_\al(I)} \le M,\;\; \||\nabla|^\frac\al2 \widetilde u\|_{X_\al(I)} < \infty
$$
for some constants $A$ and $M$, and $\widetilde u$ verifies in the sense of integral equation
$$
i\widetilde u_t = |\nabla|^\alpha \widetilde u - V(\widetilde u)\widetilde u + e
$$
for some function $e$. Let $\varphi \in \dot H_{rad}^\frac\al2$ be such that $\|\varphi - \widetilde u(0)\|_{\dot H^\frac\al2} \le A'$. Suppose there exists $\varepsilon_0 = \varepsilon_0(M, A, A')$ such that if $0 < \varepsilon \le \varepsilon_0$ and
$$
\||\nabla|^\frac\al2 e\|_{Y_\alpha(I)} \le \varepsilon, \;\;\|U(t)(\varphi - \widetilde u(0))\|_{S_\al(I)} \le \varepsilon,
$$
then there exists a unique radial solution $u$ on $\mathbb R^d \times I$ to \eqref{main eqn} such that
$$
\|u\|_{S_\al(I)} + \sup_I\|u(t) - \widetilde u(t)\|_{\dot H^\frac\al2} \le C(M, A, A').
$$
Here $Y_\alpha(I) = L_{I}^2L_x^\frac{2d}{d+\alpha}$ for both power type and for Hartree type.
\end{lemma}
Now we consider the profile decomposition in energy space. Most of them are standard and thus we only show the energy decoupling of Hartree case.
\begin{lemma}[see Theorem 1 of \cite{cox}]\label{dispest}
Let $\{t_n\}$ be sequence in $\mathbb{R}$. Suppose $\lim_{n \rightarrow \infty} |t_n| = \infty$, then for any $f \in C^\infty_0$
\[\|U(t_n)f\|_{L^p} \rightarrow 0 \text{\ as\ }n \rightarrow \infty,\]
when $p >2$.
\end{lemma}
The profile decomposition of $U(t)$ for mass critical case was already verified for radial data in \cite{chkl2} (see also \cite{chkl3}). From that decomposition, one can easily prove the following profile decomposition for the energy critical case:
\begin{lemma}\label{main}
Let $d \ge 2$, $\frac{2d}{2d-1} < \alpha < 2$, and $(q, r)$ be $\alpha$-admissible pairs with $2 < q, r <
\infty$. Suppose that
$\{u_n\}_{n \geq 1}$ is a sequence of complex-valued radial functions bounded in $\dot{H}^\frac \al2$. Then up to a subsequence, for
any $J \geq 1$, there exist a sequence of radial functions
$\{\phi^j\}_{1\leq j \leq J}\subset \dot{H}^{\frac \al2}$, $\omega_n^J \in \dot{H}^{\frac \al2}$ and a
family of parameters $(h_n^j, t_n^j)_{1 \leq j \leq J, n \geq 1}$
such that
\begin{align}\label{profileid}
u_n(x) = \sum_{1 \leq j \leq J} U(t^j_n)[(h^j_n)^{-d/2+\alpha/2}\phi^j(\cdot/{h^j_n})](x) + \omega^J_n(x)
\end{align}
and the following properties are satisfied:
\begin{align}\label{profilerem}
\lim_{J \rightarrow \infty} \limsup_{n \rightarrow \infty}
\||\nabla|^\frac \al2U(\cdot)
\omega^J_n\|_{L^q_tL^r_x} = 0,
\end{align} and for $j \neq k$, $(h^j_n, t^j_n)_{n \geq
1}$ and $(h^k_n, t^k_n)_{n \geq 1}$ are asymptotically orthogonal in
the sense that
\begin{align}\begin{aligned}\label{profileparaortho}
&\mbox{either}\;\; \limsup_{n \rightarrow
\infty}\left(\frac{h^j_n}{h^k_n} +
\frac{h^k_n}{h^j_n}\right) = \infty,\\
&\mbox{or}\;\; (h^j_n) = (h^k_n)\;\;\mbox{and}\;\;\limsup_{n
\rightarrow \infty}\frac{|t^j_n - t^k_n|}{(h^j_n)^\alpha} = \infty,
\end{aligned}\end{align}
and for each $J$
\begin{align}\label{profilenormortho}
\lim_{n \rightarrow \infty} \Big[\|u_n\|_{\dot{H}^\frac \al2}^2 - (\sum_{1 \leq j \leq J}\|\phi^j\|_{\dot{H}^\frac \al2}^2
+ \|\omega^J_n\|_{\dot{H}^\frac \al2}^2) \Big] = 0.
\end{align}
\end{lemma}
\begin{remark}
Since the space and frequency translations do not appear in the profile decomposition due to the radial symmetry, it is possible to get the strong convergence of remainder term in $L^q_t\dot{H}_r^{\frac \al2}$ as in \eqref{profilerem} not in $L^q_tL^{\frac {rd}{d-\frac {r\alpha}{2}}}_x$ norm. It plays a crucial role in the proof of Theorem \ref{energy-conc}.
\end{remark}
From energy critical profile decomposition, we prove some useful corollaries.
\begin{cor}\label{strong dec}
Suppose that $\{u_n\}_{n \geq 1}$ is a sequence of complex-valued radial functions bounded in $\dot{H}^\frac \al2$. Let $\{\phi^j\}_{1\leq j \leq J}\subset \dot{H}^{\frac \alpha 2}$, $\omega_n^J \in \dot{H}^{\frac \alpha 2}$ and a family of parameters $(h_n^j, t_n^j)_{1 \leq j \leq J, n \geq 1}$ from Lemma \ref{main}. Define group operator $G_n^j$ as $G_n^j(f) = U(t^j_n)[(h^j_n)^{-d/2+\alpha/2}f(\cdot/{h^j_n})](x)$. Then we have
\begin{align}\begin{aligned}\label{weak-limit-profile}
&(G_n^j)^{-1}(\omega_n^J) \rightharpoonup 0 \text{\ weakly\ in\ }\dot{H}^{\frac \al2}\text{\ as\ }n\ \rightarrow \infty,\\
&(G_n^j)^{-1}(u_n) \rightharpoonup \phi^j \text{\ weakly\ in\ }\dot{H}^{\frac \al2}\text{\ as\ }n\ \rightarrow \infty.
\end{aligned}\end{align}
\end{cor}
\begin{proof}
We first prove
\[(G_n^j)^{-1}U(t)(|\nabla|^\frac \al2u_n) \rightharpoonup U(t)|\nabla|^\frac \al2\phi^j \text{\ weakly\ in\ }L^{\frac {2(2+\alpha)}{d}}_{t,x}\text{\ as\ }n\ \rightarrow \infty.\]
Applying $(G_n^j)^{-1}U(t)$ to \eqref{profileid}, we obtain
\[(G_n^j)^{-1}(U(t)|\nabla|^\frac \al2u_n) = U(t)|\nabla|^{\frac \alpha 2}\phi^j + \sum_{j'\neq j}^J (G_n^j)^{-1}(G_n^{j'})(U(t)|\nabla|^\frac \al2\phi^i) + (G_n^j)^{-1}U(t)|\nabla|^\frac \al2\omega^J_n.\]
From the pairwise orthogonality of the family of parameters, we have
\[(G_n^j)^{-1}(G_n^{j'})(U(t)|\nabla|^\frac \al2\phi^j) \rightharpoonup 0 \text{\ weakly\ in\ }L^{\frac {2(2+\alpha)}{d}}_{t,x}\text{\ as\ }n\ \rightarrow \infty\]
for every $j' \neq j$. Let $\omega^J$ be the weak limit of $\{(G_n^j)^{-1}U(t)|\nabla|^\frac \al2\omega^J_n\}$. Then
\[(G_n^j)^{-1}(U(t)|\nabla|^\frac \al2u_n) \rightharpoonup U(t)|\nabla|^\frac \al2\phi^j + \omega^J.\]
Since the weak limit is unique, $\omega^J$ does not depend on $J$.
And from
\[\|(G_n^j)^{-1}U(t)|\nabla|^\frac \al2\omega^J_n\|_{L^{\frac {2(2+\alpha)}{d}}_{t,x}} \leq \limsup_{n \rightarrow \infty}
\|U(t)|\nabla|^\frac \al2\omega^J_n\|_{L^{\frac {2(2+\alpha)}{d}}_{t,x}} \longrightarrow 0 \text{ as }J\rightarrow \infty,\]
we have $\omega^J = 0$ for every $J \geq 1$.
So we have
\[(G_n^j)^{-1}U(t)(|\nabla|^\frac \al2u_n) \rightharpoonup U(t)|\nabla|^\frac \al2\phi_j \text{\ weakly\ in\ }L^{\frac {2(2+\alpha)}{d}}_{t,x}\text{\ as\ }n\ \rightarrow \infty.\]
Then following lemma gives the conclusion.
\begin{lemma}[Lemma 3.63 in \cite{meve}]
Let $\{v_n\}$ and $v$ be in $L^2$. The following statements are equivalent.
\begin{enumerate}
\item $v_n \rightharpoonup v$ weakly in $L^2$.
\item $U(t)v_n \rightharpoonup U(t)v$ weakly in $L^{\frac {2(2+\alpha)}{d}}_{t,x}$.
\end{enumerate}
\end{lemma}
\end{proof}
\newcommand{S_\al(\mathbb R)}{S_\alpha(\mathbb R)}
\newcommand{X_\al(\mathbb R)}{X_\alpha(\mathbb R)}
\begin{proposition}
Let $\{u_n\}_{n \geq 1}$ be a sequence of complex-valued radial functions
satisfying
\[\|u_n\|_{\dot{H}^\frac \al2} \leq A \text{\quad and \quad} \|U(t)u_n\|_{S_\al(\mathbb R)} \geq \delta.\]
Suppose $\{\phi^j\}_{1\leq j \leq J}\subset \dot{H}^{\frac \alpha 2}$ be linear profiles obtained in Lemma \ref{main}.
Then there exist at least one linear profile $\phi^{j_0}$ such that
\[\|U(t)\phi^{j_0}\|_{S_\al(\mathbb R)} \geq C(A,\delta).\]
\end{proposition}
\begin{proof}
By Lemma \ref{main}, we have
\[U(t)(u_n)(x) = \sum_{1 \leq j \leq J} U(t - t^j_n)[(h^j_n)^{-d/2+\alpha/2}\phi^j(\cdot/{h^j_n})](x) + U(t)\omega^J_n(x)\]
with
\[\lim_{J \rightarrow \infty} \limsup_{n \rightarrow \infty}
\||\nabla|^\frac \al2U(t)
\omega^l_n\|_{X_\al(\mathbb R)} = 0,\]
and for each $J$
\[\lim_{n \rightarrow \infty} \Big[\|u_n\|_{\dot{H}^\frac \al2}^2 - (\sum_{1 \leq j \leq J}\|\phi^j\|_{\dot{H}^\frac \al2}^2
+ \|\omega^J_n\|_{\dot{H}^\frac \al2}^2) \Big] = 0.\]
From the orthogonality(see Lemma 3.3 in \cite{chkl2}), we get
\[\lim_{n \rightarrow \infty} \|\sum_{j=1}^J U(t - t^j_n)[(h^j_n)^{-d/2+\alpha/2}\phi^j(\cdot/{h^j_n})](x)\|_{S_\al(\mathbb R)}^4 = \sum_{j=1}^J \|U(t)(\phi^j)(x)\|_{S_\al(\mathbb R)}^4\]
for every $J \geq 1$.
However,
\begin{align*}
&\limsup_{n \rightarrow \infty} \|U(t)(u_n)(x) - \sum^J_{j=1}U(t - t^j_n)[(h^j_n)^{-d/2+\alpha/2}\phi^j(\cdot/{h^j_n})](x)\|_{S_\al(\mathbb R)} \\
&\qquad \leq \limsup_{n \rightarrow \infty}\|U(t)
\omega^J_n\|_{S_\al(\mathbb R)} \leq \limsup_{n \rightarrow \infty}\||\nabla|^\frac \al2U(t)
\omega^J_n\|_{X_\al(\mathbb R)} \to 0\;\;\mbox{as}\;\;J \to \infty.
\end{align*}
So we obtain
\[\limsup_{n \rightarrow \infty} \|U(t)(u_n)(x)\|_{S_\al(\mathbb R)}^4 = \sum_{j=1}^J \|U(t)(\phi^j)(x)\|_{S_\al(\mathbb R)}^4.\]
And Strichartz estimate gives
\[\sum_{j=1}^J \|U(t)(\phi^j)(x)\|_{S_\al(\mathbb R)}^4 \le C \big(\sup_{j \geq J}\|U(t)(\phi^j)(x)\|_{S_\al(\mathbb R)}^2\big)\sum_{j \geq 1}\|\phi^j\|_{\dot{H}^\frac \al2}^2.\]
Since $\sum_{j \geq 1}\|\phi^j\|_{\dot{H}^\frac \al2}^2 \leq \limsup_{n \rightarrow \infty} \|u_n\|_{\dot{H}^\frac \al2}^2 \leq A^2$, we have
\[\sup_{j \geq 1}\|U(t)(\phi^j)(x)\|_{S_\al(\mathbb R)}^2 \geq \frac {\delta^4}{A^2}.\]
In particular, we can find $j_0$ such that
\[\|U(t)(\phi^{j_0})(x)\|_{S_\al(\mathbb R)}^2 \geq \frac {\delta^4}{A^2}.\]
\end{proof}
\begin{proposition}
Let $\{u_n\}_{n \geq 1}$ be a sequence of complex-valued radial functions bounded in $\dot H^\frac\al2$. Suppose $\{\phi^j\}_{1\leq j \leq J}\subset \dot{H}^\frac\al2$ be linear profiles obtained in Lemma \ref{main}. Then for each $J$,
\[\lim_{n \rightarrow \infty} \left(E(u_n) - \sum_{1 \leq j \leq J}E(U(t^j_n)[(h^j_n)^{-d/2+\alpha/2}\phi^j(\cdot/{h^j_n})](x)) - E(\omega^J_n)\right) = 0.\]
\end{proposition}
\begin{proof} For the power type we refer the readers to the proof in \cite{ker}. We only consider the Hartree case. Also see \cite{mgz} for NLS with Hartree nonlinearity.
Thanks to the kinetic energy decoupling \eqref{profilenormortho}, it suffices to show
\begin{align*}
\lim_{n \rightarrow \infty} \Big(\int |u_n|^2(|x|^{-2\alpha} * &|u_n|^2)dx - \sum_{1 \leq j \leq l}\int |G_n^j(\phi^j)|^2(|x|^{-2\alpha} * |G_n^j(\phi^j)|^2)dx\\
&- \int |\omega^J_n|^2(|x|^{-2\alpha} * |\omega^J_n|^2)dx\Big) = 0.
\end{align*}
We first prove
\begin{align*}
\lim_{n \rightarrow \infty} \Big(\int |u_n|^2(|x|^{-2\alpha} * |u_n|^2)dx &- \int |u_n - G_n^1(\phi^1)|^2(|x|^{-2\alpha} * |u_n - G_n^1(\phi^1)|^2)dx\\
-&\int |G_n^1(\phi^1)|^2(|x|^{-2\alpha} * |G_n^1(\phi^1)|^2)dx \Big)= 0.
\end{align*}
Then repeated arguments give the conclusion.
When $\lim_{n \rightarrow \infty} \big|\frac {t^1_n}{(h^1_n)^\alpha}\big| = \infty$, we obtain
\begin{align*}
&\quad \lim_{n \rightarrow \infty} \int |G_n^1(\phi^1)|^2(|x|^{-2\alpha} * |G_n^1(\phi^1)|^2)dx = \lim_{n \rightarrow \infty} \|G_n^1(\phi^1)\|^4_{L^{\frac {2d}{d-\alpha}}} = \lim_{n \rightarrow \infty} \Big\|U\Big(\frac {t^1_n}{(h^1_n)^\alpha}\Big)(\phi^1)\Big\|^4_{L^{\frac {2d}{d-\alpha}}} = 0
\end{align*}
by using H\"older inequality, fractional integration, scaling and Lemma \ref{dispest}. Similarly, one can prove \[\lim_{n\to \infty}\left(\int |u_n|^2(|x|^{-2\alpha} * |u_n|^2)dx - \int |u_n - G_n^1(\phi^1)|^2(|x|^{-2\alpha} * |u_n - G_n^1(\phi^1)|^2)dx\right) = 0.\]
Now we handle the case $\lim_{n \rightarrow \infty} \big|\frac {t^1_n}{(h^1_n)^\alpha}\big| < \infty$. By taking subsequence we may assume that $\lim_{n \rightarrow \infty} \frac {t^1_n}{(h^1_n)^\alpha} = t_\infty.$ Let $S^1_n(u_n) := (h^1_n)^{\frac d2 - \frac \alpha 2}u_n(h^1_n \cdot)$. Then we have
\begin{align*}
S^1_n(u_n) \rightharpoonup U(t_\infty)\phi^1 &\text{\ weakly\ in\ }\dot{H}^{\frac \al2} \text{\ as\ }n \rightarrow \infty\\
\text{\ and\ }G_n^1(S^1_n(\phi_1)) \rightarrow U(t_\infty)\phi^1 &\text{\ strongly\ in\ }L^{\frac {2d}{d-\alpha}} \text{ as }n \rightarrow \infty.
\end{align*}
The scaling symmetry yields
\begin{align*}
&\int |u_n|^2(|x|^{-2\alpha} * |u_n|^2)dx - \int |u_n - G_n^1(\phi^1)|^2(|x|^{-2\alpha} * |u_n - G_n^1(\phi^1)|^2)dx\\
&\qquad -\int |G_n^1(\phi^1)|^2(|x|^{-2\alpha} * |G_n^1(\phi^1)|^2)dx\\
&=\int |S^1_n(u_n)|^2(|x|^{-2\alpha} * |S^1_n(u_n)|^2)dx\\
&\qquad-\int |S^1_n(u_n) - G_n^1(S^1_n(\phi^1))|^2(|x|^{-2\alpha} * |S^1_n(u_n) - G_n^1(S^1_n(\phi^1))|^2)dx\\
&\qquad -\int |G_n^1(S^1_n(\phi^1))|^2(|x|^{-2\alpha} * |G_n^1(S^1_n(\phi^1))|^2)dx\\
&=: I_n + II_n + III_n,
\end{align*}
where
\begin{align*}
I_n &= \int |S^1_n(u_n)|^2(|x|^{-2\alpha} * |S^1_n(u_n)|^2)dx\\
&\quad- \int |S^1_n(u_n) - U(t_\infty)\phi^1|^2(|x|^{-2\alpha} * |S^1_n(u_n) - U(t_\infty)\phi^1|^2)dx\\
&\quad- \int |U(t_\infty)\phi^1|^2(|x|^{-2\alpha} * |U(t_\infty)\phi^1|^2)dx,\\
II_n &= \int |S^1_n(u_n) - G_n^1(S^1_n(\phi_1))|^2(|x|^{-2\alpha} * |S^1_n(u_n) - G_n^1(S^1_n(\phi_1))|^2)dx\\
&\quad - \int |S^1_n(u_n) - U(t_\infty)\phi^1|^2(|x|^{-2\alpha} * |S^1_n(u_n) - U(t_\infty)\phi^1|^2)dx,\\
III_n &= \int |G_n^1(S^1_n(\phi_1))|^2(|x|^{-2\alpha} * |G_n^1(S^1_n(\phi_1))|^2)dx\\
&\quad- \int |U(t_\infty)\phi^1|^2(|x|^{-2\alpha} * |U(t_\infty)\phi^1|^2)dx.
\end{align*}
$I_n$ goes to $0$ by Lemma \ref{refatou} below. And by using H\"older inequality, fractional integration and Lemma \ref{dispest} again, we also obtain
\begin{align*}
\lim_{n\to \infty}(II_n + III_n) = 0.
\end{align*}
\end{proof}
\begin{lemma}\label{refatou}
Let $\{f_n\}$ be bounded sequence in $\dot{H}^{\frac \al2}$. If $f_n$ weakly converges to $f$, then for some subsequence $\{f_n\}$,
\begin{align*}
\int |f_n|^2(|x|^{-2\alpha} * |f_n|^2) - &|f_n - f|^2(|x|^{-2\alpha} * |f_n - f|^2) - |f|^2(|x|^{-2\alpha} * |f|^2)dx\\
& \longrightarrow 0 \text{ as } n \rightarrow \infty.
\end{align*}
\end{lemma}
\begin{proof}
Assume that $\||\nabla|^\frac\al2 f_n\|_{L^2} \le M$ for all $n \ge 1$. Since $C^\infty_{0}$ is dense in $\dot{H}^{\frac \al2}$, one can find $\beta \in C^\infty_{0}$ such that $\|\beta - f\|_{\dot{H}^{\frac \al2}} < \frac{\varepsilon}{12C(1+M)^2}$ for some constant $C$. And since the multiplication operator $T_{\beta} : \dot{H}^{\frac \al2} \rightarrow L^p_x, T_{\beta}(f) = \beta f$ is compact when $\beta \in C^\infty_0, 1 \leq p < \frac {2d}{d-\alpha}$, there exists subsequence of $f_n$ such that $\|\beta(f_n-f)\|_{L^{\frac {d}{d-\alpha}}} < \frac{\varepsilon}{12C(1+M)^2}$ if $n \ge N$.
On the other hand, one can easily check that
\begin{align*}
\qquad&\int |f_n|^2(|x|^{-2\alpha} * |f_n|^2) - |f_n - f|^2(|x|^{-2\alpha} * |f_n - f|^2) - |f|^2(|x|^{-2\alpha} * |f|^2)dx\\
=&\int (f\overline{f_n} + \overline{f}f_n - 2|f|^2)(|x|^{-2\alpha} * |f_n|^2) - |f_n - f|^2(|x|^{-2\alpha} *(f\overline{f_n} + \overline{f}f_n - 2|f|^2))\\
& +(f\overline{f_n} + \overline{f}f_n - 2|f|^2)(|x|^{-2\alpha} *|f|^2 )-|f_n|^2(|x|^{-2\alpha} * |f|^2)+ |f|^2(|x|^{-2\alpha} * |f_n|^2)dx.
\end{align*}
Let us observe that
\[\int -|f_n|^2(|x|^{-2\alpha} * |f|^2)+ |f|^2(|x|^{-2\alpha} * |f_n|^2)\,dx = 0.\]
Then by using H\"older's inequality and Sobolev embedding, we have
\begin{align*}
&\qquad \int(f\overline{(f_n-f)})(|x|^{-2\alpha} * |f_n|^2)dx\\
& = \int((f - \beta)\overline{(f_n-f)})(|x|^{-2\alpha} * |f_n|^2) + (\beta\overline{(f_n-f)})(|x|^{-2\alpha} * |f_n|^2)dx\\
&\le C \|f - \beta\|_{\dot H^\frac\al2}\|f_n-f\|_{L^{\frac {2d}{d-\alpha}}}\|f_n\|_{L^{\frac {2d}{d-\alpha}}} + C \|\beta(f_n-f)\|_{L^{\frac {d}{d-\alpha}}}\|f_n\|_{L^{\frac {2d}{d-\alpha}}}\\
&\le CM^2\|f - \beta\|_{\dot H^\frac\al2} + CM\|\beta(f_n-f)\|_{L^{\frac {d}{d-\alpha}}} < \frac\ep6\;\;\mbox{if}\;\;n \ge N.
\end{align*}
We need to treat remaining 5 terms. But they can be done by exactly the same way as above.
\end{proof}
Using the local well-posedness theorem with initial data at $t=0$ or $t=\pm \infty $, we define the nonlinear profile by the maximal nonlinear solution for each linear profile.
\begin{defn}
Let $\{(h_n,t_n)\} $ be a family of parameters and $\{t_n\}$ have a limit in $[-\infty, \infty]$. Given a linear profile $\phi \in \dot{H}^\frac \al2$ with $\{(h_n,t_n)\}$, we define the nonlinear profile associated with it to be the maximal solution $v$ to \eqref{main eqn} which is in $C((-T_{\min},T_{\max}); \dot{H}^\frac \al2)$ satisfying an asymptotic condition: For the sequence $\{t_n\}$,
\[\lim_{n \to \infty}\|U(t_n)\phi - v(t_n)\|_{\dot{H}^\frac \al2} = 0.\]
\end{defn}
\begin{remark}\label{nonlinear prof}
Let $\{u_n\}_{n \geq 1}$ be a sequence of complex-valued radial functions bounded in $\dot H^\frac\al2$ and $\{\phi^j\}_{1\leq j \leq J}\subset \dot{H}^\frac\al2$ be the corresponding linear profiles obtained in Lemma \ref{main}. Then by refining subsequence and using diagonal argument we may assume that for each $j$ the sequence $\{t_n^j\}$ converges to $t^j \in [-\infty, +\infty]$. By using the standard time-translation and absorbing error we may assume that $t^j := 0$ and either $t_n^j := 0 $ or $t_n^j \to \pm \infty$.
As stated in \cite{keme} the nonlinear profiles $v^j : I^j \times \mathbb R^d \to \mathbb C$ associated with $\phi^j$ and $t_n^j$ always exist and they can be summarized as follows:
If $t_n^j = 0$, then $v^j$ is the maximal solution to \eqref{main eqn} with initial data $v^j(0) = \phi^j$. If $t_n^j \to \pm\infty$, then $v^j$ is the maximal solution to \eqref{main eqn} that scatters forward/backward in time to $U(t)\phi^j$.
\end{remark}
\section{Energy concentration}
In this section we show Theorems \ref{conc-infty} and \ref{energy-conc} by following the arguments as in \cite{mets} and \cite{kv}, respectively.
\subsection{Unconfined kinetic energy: Proof of Theorem \ref{conc-infty}}
Let $\beta$ be a $C_0^\infty$-bump function which is $1$ for $|x| \le 1$ and $0$ for $|x| > 1$. Then we have from Lemma \ref{com-lem} and mass conservation that
\begin{align*}
\|\beta(\cdot/R)u\|_{L^\frac{2d}{d-\alpha}}^\frac{2d}{d-\alpha} &\le C\||\nabla|^\frac\al2(\beta(\cdot/R)u)\|_{L^2}^\frac{2d}{d-\alpha}\\
&\le C\|\big[ |\nabla|^\frac\al2, \beta(\cdot/R)\big] u\|_{L^2}^\frac{2d}{d-\alpha} + C\||\nabla|^\frac\al2 u\|_{L^2(|x| \le 2R)}^\frac{2d}{d-\alpha}\\
& \le CR^{-\frac\al2}\|u\|_{L^2}^\frac{2d}{d-\alpha} + C\||\nabla|^\frac\al2 u\|_{L^2(|x| \le 2R)}^\frac{2d}{d-\alpha}\\
& \le C(R^{-\frac\al2}\|\varphi\|_{L^2}^\frac{2d}{d-\alpha}) + C\||\nabla|^\frac\al2 u\|_{L^2(|x| \le 2R)}^\frac{2d}{d-\alpha}\\
& \le A + \||\nabla|^\frac\al2 u\|_{L^2(|x| \le 2R)}^\frac{2d}{d-\alpha}
\end{align*}
for some $A = A(R, \|\varphi\|_{L^2})$. Using the endpoint Sobolev inequality (Proposition 2 of \cite{chooz-ccm}) and real interpolation \cite{bl} that
$$
|x|^\frac{d-1}2 |f(x)| \le C \|f\|_{\dot B_{2,1}^\frac12} \le C\|f\|_{L^2}^\frac{\alpha-1}\alpha\||\nabla|^\frac\al2 f\|_{L^2}^\frac1\alpha,
$$
we have
\begin{align*}
\|(1-\beta(\cdot/R))u\|_{L^\frac{2d}{d-\alpha}}^\frac{2d}{d-\alpha} &= \int (1-\beta(\cdot/R))|u|^{\frac{2\alpha}{d-\alpha}}|u|^2dx &\le CR^{-\frac{\alpha(d-1)}{d-\alpha}}\|u\|_{L^2}^\frac{2(d-1)}{d-\alpha}\||\nabla|^\frac\al2u\|_{L^2}^{\frac{2}{d-\alpha}} \le A \||\nabla|^\frac\al2u\|_{L^2}^{\frac{2}{d-\alpha}},
\end{align*}
where $\|f\|_{\dot B_{2,1}^\frac12} := \sum_N N^\frac12 \|f_N\|_{L^2}$ is the homogeneous Besov norm.
On the other hand, for Hartree type we have that
\begin{align*}
\int V(u)&|\beta(x/R)u|^2\,dx\\
&\le C\|u\|_{L^\frac{2d}{d-\alpha}}^2\|\beta(\cdot/R)u\|_{L^\frac{2d}{d-\alpha}}^2\\
&\le C\|\beta(\cdot/R)u\|_{L^\frac{2d}{d-\alpha}}^4 + C\|(1 - \beta(\cdot/R))u\|_{L^\frac{2d}{d-\alpha}}^2\|\beta(\cdot/R)u\|_{L^\frac{2d}{d-\alpha}}^2\\
&\le A + \||\nabla|^\frac\al2 u\|_{L^2(|x| \le 2R)}^4 + A\||\nabla|^\frac\al2 u\|_{L^2}^\frac{2}{d}(A + \||\nabla|^\frac\al2 u\|_{L^2(|x| \le 2R)}^2)
\end{align*}
and
\begin{align*}
\int V(u)&|(1-\beta(x/R))u|^2\,dx\\
&\le C\|u\|_{L^\frac{2d}{d-\alpha}}^2\|(1-\beta(\cdot/R))u\|_{L^\frac{2d}{d-\alpha}}^2\\
&\le C\|(1-\beta(\cdot/R))u\|_{L^\frac{2d}{d-\alpha}}^4 + C\|(1 - \beta(\cdot/R))u\|_{L^\frac{2d}{d-\alpha}}^2\|\beta(\cdot/R)u\|_{L^\frac{2d}{d-\alpha}}^2\\
&\le A\||\nabla|^\frac\al2 u\|_{L^2}^\frac{4}{d} + A\||\nabla|^\frac\al2 u\|_{L^2}^\frac{2}{d}(A + \||\nabla|^\frac\al2 u\|_{L^2(|x| \le 2R)}^2).
\end{align*}
From the energy conservation it follows that
\begin{align*}
\||\nabla|^\frac\al2u(t)\|_{L^2}^2 &= \frac{2}{\mu}\left\{\begin{array}{l}\|\beta(\cdot/R)u(t)\|_{L^\frac{2d}{d-\alpha}}^\frac{2d}{d-\alpha} + \|(1-\beta(\cdot/R))u(t)\|_{L^\frac{2d}{d-\alpha}}^\frac{2d}{d-\alpha}\\
\int V(u)|\beta(\cdot/R)u(t)|^2\,dx + \int V(u)|(1-\beta(\cdot/R))u(t)|^2\,dx\end{array}\right\} + E(\varphi).
\end{align*}
Let $y(t) = \||\nabla|^\frac\al2u(t)\|_{L^2}^2$ and $z(t) = \||\nabla|^\frac\al2 u(t)\|_{L^2(|x|\le 2R)}^2$. Then from the above estimates we have
$$
y(t) \le C\left\{\begin{array}{l} A + z^\frac{d}{d-\alpha} + y^\frac{1}{d-\alpha}\\ A + Ay^\frac1d(A+z)+Ay^\frac2d+z^2\end{array}\right\} + E(\varphi).
$$
Since $\limsup_{t\to T^*}y(t) = +\infty$ and $d > \alpha+1$ for power type ($d > 2\alpha$ for Hartree type), we conclude that
$\limsup_{t\to T^*} z(t) = +\infty$.
If $u(t) \in L^\infty$ for all $t < T^*$, then since $\|\beta(\cdot/R)u(t)\|_{L^\frac{2d}{d-\alpha}} \le CR^\frac{d-\alpha}2 \|u\|_{L^\infty(|x| \le 2R)}$, by replacing $\||\nabla|^\frac\al2u(t)\|_{L^2(|x|\le 2R)}$ with $\|u\|_{L^\infty(|x| \le 2R)}$ in the above estimates we get the desired.
\subsection{Confined kinetic energy: Proof of Theorem \ref{energy-conc}}
Choose a sequence $t_n \to T^*$ and let $u_n$ be the solution on $[0, T^* - t_n)$ to \eqref{main eqn} with initial data $u(t_n)$. Then since $\sup_{0 < t < T^*}\||\nabla|^\frac\al2u(t)\|_{L^2} =: M < +\infty$, by Lemma \ref{main} we can decompose each $u_n(0)$ by
$$
u_n(0) = \sum_{j = 1}^J G_n^j \phi^j + \omega_n^J.
$$
We denote the symmetry operator $g_n^j$ by $g_n^jf(t, x) = (h_n^j)^{-\frac{d-\alpha}{2}}f(t/(h_n^j)^\alpha, x/h_n^j)$. Then $G_n^j \phi^j = g_n^j U(t_n^j)\phi^j$. Let $v^j : I^j \times \mathbb R^d \to \mathbb C$ be nonlinear profile associated with $\phi^j$ and $(h_n^j, t_n^j)$ as stated in Remark \ref{nonlinear prof}. For each $j, n \ge 1$, we define $v_n^j : I_n^j \times \mathbb R^d \to \mathbb C$ by
$$
v_n^j(t) := g_n^jv^j(\cdot + t_n^j)(t),
$$
where $I_n^j = \{t \in \mathbb R : (h_n^j)^{-\alpha}t + t_n^j \in I^j\}$. Then $v_n^j$ is also a solution to \eqref{main eqn} with initial data $v_n^j(0) = g_n^jv^j(t_n^j)$ and maximal time interval $I_n^j = (-T_{n,j}^-, T_{n,j}^+)$ for $0 < T_{n,j}^-, T_{n,j}^+ < +\infty$. By the kinetic energy decoupling \eqref{profilenormortho} there exists $J_0 = J_0(\delta_0) \ge 1$ such that
$$
\||\nabla|^\frac\al2 \phi^j\|_{L^2} \le \delta_0\;\;\mbox{for all}\;\;j \ge J_0.
$$
For sufficiently small $\delta_0$, Lemma \ref{loc} yields that $v_n^j$ are global and satisfy that
\begin{align}\label{energy bound}
\sup_{t\in\mathbb R}\||\nabla|^\frac\al2 v_n^j(t)\|_{L^2} + \|v_n^j\|_{S_\al(\mathbb R)} \le C\||\nabla|^\frac\al2 \phi^j\|_{L^2}.
\end{align}
Now we can find a so-called bad profile $\phi^{j_0}$, $1 \le j_0 < J_0$ such that
\begin{align}\label{bad}
\limsup_{n \to \infty}\|v_n^{j_0}\|_{S_\alpha([0, T^*-t_n)} = +\infty,
\end{align}
\newcommand{S_\al([0, T_n^*))}{S_\alpha([0, T_n^*))}
\newcommand{X_\al([0, T_n^*))}{X_\alpha([0, T_n^*))}
\begin{proof}[Proof of \eqref{bad}]
We will actually show that
\begin{align}
\limsup_{n \to \infty}\|v_n^{j_0}\|_{S_\alpha([0, T_n^*))} = +\infty,
\end{align}
where $T_n^* = \min_{1 \le j < J_0}(T^*-t_n, T_{n,j}^+)$. Suppose that $\limsup_{n \to \infty}\|v_n^j\|_{S_\alpha([0, T_n^*))} < +\infty$ for all $1 \le j < J_0$. Then this implies that $T^*-t_n \le T_{n, j}^+$ for all $1 \le j < J_0$ if $n$ is large. If $T_{n, j}^+ \le T^*-t_n$ for some $j$, then since $\limsup_{n \to \infty}\|v_n^j\|_{S_\alpha([0, T_{n,j}^+))} < +\infty$, the maximality means that $T_{n,j}^+ = +\infty$ for sufficiently large $n$. This contradicts the fact $T^* < +\infty$.
Then from this together with \eqref{energy bound} and \eqref{profilenormortho} it follows that
\begin{align}\label{s-bound}
\sum_{j \ge 1}^J \|v_n^j\|_{S_\alpha([0, T_n^*))}^2 \le C (1 + \sum_{j \ge J_0}^J\||\nabla|^\frac\al2 \phi^j\|_{L^2}^2) \le C(1 + M^2)
\end{align}
for any $J$ and for sufficiently large $n$. We now define functions $u_n^J$ on $[0, T_n^{m_0}]$ approximating $u_n$ by
$$
u_n^J = \sum_{j = 1}^Jv_n^j + U(t)\omega_n^J.
$$
Since $v^j$ are nonlinear profile associated with $(\phi^j, t_n^j)$, we have
\begin{align*}
\|u_n^J(0) - u_n(0)\|_{\dot H^\frac\al2} = \|\sum_{j = 1}^J(g_n^j v^j(t_n^j) - g_n^j U(t_n^j)\phi^j)\|_{\dot H^\frac\al2} \le \sum_{j=1}^J\|v^j(t_n^j) - U(t_n^j)\phi^j\|_{\dot H^\frac\al2} \to 0
\end{align*}
as $n \to \infty$. By \eqref{profilerem} and \eqref{s-bound} we also have
\begin{align}\begin{aligned}\label{app-s-bound}
\lim_J\limsup_{n \to \infty} \|u_n^J\|_{S_\alpha([0, T_n^*))} &\le \lim_J\limsup_{n \to \infty}(\|\sum_j v_n^j\|_{S_\al([0, T_n^*))} + \|U(t)\omega_n^J\|_{S_\al([0, T_n^*))})\\
& \le C(1 + M^2).
\end{aligned}\end{align}
By the local well-posedness we deduce that
\begin{align}\label{app-norm}
\lim_J\limsup_{n \to \infty} (\|u_n^J\|_{L^\infty_{[0, T_n^*)} \dot H^{\frac \al2}} + \||\nabla|^\frac\al2 u_n^J\|_{X_\al([0, T_n^*))}) \le C(M).
\end{align}
On the other hand, $u_n^J$ satisfy that
$$
i\partial_t u_n^J = |\nabla|^\alpha u_n^J - V(u_n^J)u_n^J + e,
$$
where $e = e_1 + e_2$, $$e_1 = V(u_n^J) - V(\sum_{j=1}^J v_n^j)(\sum_{j=1}^J v_n^j)$$ and $$e_2 = V(\sum_{j=1}^Jv_n^j)(\sum_{j=1}^Jv_n^j)- \sum_{j = 1}^J V(v_n^j)v_n^j.$$
We first show that $\limsup_{n \to \infty}\||\nabla|^\frac\al2 e_2\|_{Y_\alpha([0, T_n^*))} = 0$.
In fact, from direct calculation we get that for power type
\begin{align*}
e_2 = \frac{2\alpha}{d-\alpha}{\rm Re}\sum_{j\neq j'}v_n^{j'}v_n^j\int_0^1| s \sum_{j'\neq j}v_n^{j'} + v_n^j|^\frac{4\alpha-2d}{d-\alpha}(s \sum_{j'\neq j}v_n^{j'} + v_n^j)\,ds
\end{align*}
and for Hartree type
\begin{align*}
e_2 = \sum_{j'\neq j}(|x|^{-2\alpha} * |v_n^{j'}|^2)v_n^j + \sum_j\sum_{j_1' \neq j_2'} (|x|^{-2\alpha} * (v_n^{j_1'}\overline{v_n^{j_2'}})) v_n^j.
\end{align*}
Since $\alpha < d \le 2\alpha$ for power type, we have
\begin{align*}
\||\nabla|^\frac\al2 e_2\|_{Y_{\alpha}([0,T^*))} \le C\sum_{j\neq j'}\Big(&\||\nabla|^\frac\al2(v_n^{j'}v_n^j)\|_{L^\frac{d+\alpha}{d-\alpha}_{[0,T_n^*)}L^\frac{2d(d+\alpha)}{2d^2-\alpha d+\alpha^2}}(\sum_{j=1}^J\|v_n^j\|_{S_\al([0, T_n^*))}^\frac{3\alpha-d}{d-\alpha})\\
& + (\sum_{j=1}^J\||\nabla|^\frac\al2 v_n^j\|_{X_\al([0, T_n^*))})(\sum_{j=1}^J\|v_n^j\|_{S_\al([0, T_n^*))}^\frac{4\alpha-2d}{d-\alpha})\|v_n^{j'}v_n^j\|_{L_{[0,T_n^*),x}^\frac{d+\alpha}{d-\alpha}}\Big).
\end{align*}
Thus the orthogonality \eqref{profileparaortho} gives
$$
\limsup_{n\to\infty}\||\nabla|^\frac\al2 e_2\|_{L^2_{[0, T_n^*)} L^\frac{2d}{d+\alpha}} = 0.
$$
For Hartree type by the orthogonality \eqref{profileparaortho} and the argument used for the proof of Lemma 3.3 in \cite{chkl2} one can easily get
\begin{align*}
\limsup_{n\to \infty}\||\nabla|^\frac\al2 e_2\|_{L^2_{[0,T_n^*)}L^\frac{2d}{d+\alpha}} = 0.
\end{align*}
Now let us consider $e_1$. Let $V_n^J = \sum_{j= 1}^Jv_n^j$ and let us invoke that $\mu = \frac{2d}{d-\alpha}$ for power type and $\mu = 4$ for Hartree type. Then we have
\begin{align*}
&\||\nabla|^\frac\al2 e_1\|_{Y_\alpha([0, T_n^*))}\\
&\le C\Big(\||\nabla^\frac\al2 u_n^J\|_{X_\al([0, T_n^*))} + \||\nabla|^\frac\al2V_n^J\|_{X_\al([0, T_n^*))}\Big)\Big(\|u_n^J\|_{S_\al([0, T_n^*))}^{\mu-3} + \|V_n^J\|_{S_\al([0, T_n^*))}^{\mu-3}\Big)\|U(t)\omega_n^J\|_{S_\al([0, T_n^*))}\\
&\quad + C\Big(\|u_n^J\|_{S_\al([0, T_n^*))}^{\mu-2} + \|V_n^J\|_{S_\al([0, T_n^*))}^{\mu-2}\Big)\||\nabla|^\frac\al2 U(t)\omega_n^J\|_{X_\al([0, T_n^*))}.
\end{align*}
By \eqref{profilerem} we get
$$
\lim_{J\to \infty}\limsup_{n\to \infty}\||\nabla|^\frac\al2 e_1\|_{Y_\alpha([0, T_n^*))} = 0.
$$
We apply Lemma \ref{stab} with $\widetilde u = u_n^J$ and $u = u_n$ to conclude that
$$
\|u_n\|_{S_\alpha([0,T^*-t_n))} < +\infty \;\;\mbox{for sufficiently large}\;\;n.
$$
This contradicts that $u$ blows up within finite time $T^*$.
\end{proof}
By reordering we may assume that $\limsup_{n\to\infty}\|v_n^1\|_{S_\alpha([0,T^*-t_n))} = +\infty$ and that there exists $1 \le J_1 < J_0$ such that
$$
\limsup_{n\to\infty}\|v_n^j\|_{S_\alpha([0, T^*-t_n))} = \infty\;\;(j \le J_1)\;\;\mbox{and}\;\;\limsup_{n\to\infty}\|v_n^j\|_{S_\alpha([0, T^*-t_n))} < \infty\;\;(j > J_1).
$$
Then for each $m, n \ge 1$, there exist $1 \le j(m, n) \le J_1$ and $0 < T_n^m < T^*-t_n$ such that
\begin{align}\label{pigeonhole}
\sup_{1 \le j \le J_1}\|v_n^j\|_{S_\alpha([0, T_n^m])} = \|v_n^{j(m,n)}\|_{S_\alpha([0,T_n^m])} = m.
\end{align}
By using the pigeonhole principle and then reordering, we may assume that $j(m,n) = 1$ for infinitely many $m, n$. Then by Theorem \ref{gwp} there exists $0 \le \tau_n^m \le T_n^m$ such that
$$
\limsup_{m\to\infty}\limsup_{n\to\infty}\||\nabla|^\frac\al2 v_n^1(\tau_n^m)\|_{L^2} \ge \||\nabla|^\frac\al2 W_\alpha\|_{L^2}.
$$
For any $\varepsilon > 0$ we can find $m_0 = m_0(\varepsilon)$ such that
$$
\||\nabla|^\frac\al2 v_n^1(\tau_n^{m_0})\|_{L^2} \ge \||\nabla|^\frac\al2 W_\alpha\|_{L^2} - \varepsilon\;\;\mbox{for infinitely many}\;\;n.
$$
\newcommand{\tau_n^{m_0}}{\tau_n^{m_0}}
\newcommand{\tau_n^-}{\tau_n^-}
\newcommand{\tau_n^+}{\tau_n^+}
Passing to a subsequence we may have that
\begin{align}\label{energy-ep}
\||\nabla|^\frac\al2 v_n^1(\tau_n^{m_0})\|_{L^2} \ge \||\nabla|^\frac\al2 W_\alpha\|_{L^2} - \varepsilon\;\;\mbox{for all}\;\;n \;\;\mbox{and}\;\;\lim_{n\to \infty}\||\nabla|^\frac\al2 v_n^1(\tau_n^{m_0})\|_{L^2}\;\;\mbox{exists}.
\end{align}
Now we choose a small $\eta$ to be specified later and fix $n$. Then since $\|v_n^1\|_{S_\alpha([0, T_n^{m_0}])} = m_0$, we can find $\tau_n^-, \tau_n^+$ with $0 \le \tau_n^- \le \tau_n^{m_0} \le \tau_n^+\le T_n^{m_0}$ such that
\begin{align}\label{s-eta}
\|v_n^1\|_{S_\alpha([\tau_n^-, \tau_n^+])} = \eta.
\end{align}
Using local well-posedness (Lemma \ref{loc}) we get
$$
\|U(t)v_n^1(\tau_n^{m_0})\|_{S_\alpha([\tau_n^--\tau_n^{m_0}, \tau_n^+-\tau_n^{m_0}])} \ge C\eta^{\widetilde{D}}
$$
for some dimension-dependent constant $\widetilde D$. By Lemma \ref{inverse str} there exists $\tau_n^- - \tau_n^{m_0} \le s_n \le \tau_n^+-\tau_n^{m_0}$ such that
\begin{align}\label{tightness0}
\int_{|x| \le \widetilde C|T^*-t_n'|^\frac1\alpha} |U(s_n)|\nabla|^\frac\al2 v_n^1(\tau_n^{m_0})|^2\,dx \ge \widetilde C^{-1},
\end{align}
where $\widetilde C = \widetilde C(d, M, \eta)$ and $t_n' = t_n + s_n + \tau_n^{m_0}$.
From the definition of $v_n^1$ and \eqref{tightness0} we deduce that
$$
\int_{|y| \le \widetilde C(h_n^1)^{-1}|T^*-t_n'|^\frac1\alpha} |U(s_n(h_n^1)^{-\alpha})\big(|\nabla|^\frac\al2 v^1((h_n^1)^{-\alpha} \tau_n^{m_0} + t_n^1, y)\big)|^2\,dy \ge \widetilde C^{-1}.
$$
By applying Lemma \ref{tightness} and rescaling we have
\begin{align}\label{tightness1}
\Big|\||\nabla|^\frac\al2 v_n^1(\tau_n^{m_0})\|_{L^2}^2 - \int_{|x| \le R_n} |U(s_n)v_n^1(\tau_n^{m_0})|^2\,dx\Big| \to 0
\end{align}
for any sequence $R_n \in (0, \infty)$ such that $(T^*-t_n')^{-\frac1\alpha}R_n \to \infty$ as $n \to \infty$. Let $u_n^J$ be the approximate functions defined on $[0, T_n^{m_0}]$ as above. Then in view of the proof of \eqref{bad} and \eqref{pigeonhole} we can deduce that
$$
\lim_{J \to \infty}\limsup_{n \to \infty}\||\nabla|^\frac\al2(u_n^J(s_n + \tau_n^{m_0}) - u(t_n'))\|_{L^2} = 0.
$$
\newcommand{\big\langle}{\big\langle}
\newcommand{\big\rangle}{\big\rangle}
Using \eqref{profileparaortho} and Corollary \ref{strong dec} we have
$$
\limsup_{n\to\infty}\big\langle |\nabla|^\frac\al2 u_n^J(s_n + \tau_n^{m_0}), |\nabla|^\frac\al2v_n^1(s_n +\tau_n^{m_0}) \big\rangle = \limsup_{n\to\infty}\||\nabla|^\frac\al2 v_n^1(s_n+\tau_n^{m_0})\|_{L^2}^2
$$
for all $J \ge 1$. Thus we obtain
$$
\limsup_{n\to\infty} |\big\langle |\nabla|^\frac\al2 u_n(t_n'), |\nabla|^\frac\al2 v_n^1(s_n + \tau_n^{m_0}) \big\rangle| = \limsup_{n\to\infty} \||\nabla|^\frac\al2 v_n^1(s_n+\tau_n^{m_0})\|_{L^2}^2.
$$
From \eqref{s-eta} and Strichartz estimate it follows that
$$
\||\nabla|^\frac\al2(v_n^1(s_n + \tau_n^{m_0}) - U(s_n)v_n^1(\tau_n^{m_0}))\|_{L^2} \le C\eta^{\mu-2}.
$$
So, if $\eta$ is sufficiently small, then we get
$$
\limsup_{n\to\infty}|\big\langle|\nabla|^\frac\al2 u_n(t_n'), U(s_n)|\nabla|^\frac\al2 v_n^1(\tau_n^{m_0}) \big\rangle| \ge \lim_{n\to\infty}\||\nabla|^\frac\al2 v_n^1(\tau_n^{m_0})\|_{L^2}^2 - \eta^{D'},
$$
for some $D' < \mu-2$.
Therefore by Cauchy-Schwarz inequality and \eqref{energy-ep} we obtain that
\begin{align*}
\limsup_{n\to \infty}\int_{|x| \le R_n} ||\nabla|^\frac\al2 u(t_n')|^2\,dx &\ge \frac{(\lim_{n\to\infty}\||\nabla|^\frac\al2 v_n^1(\tau_n^{m_0})\|_{L^2}^2 - \eta^{D'})^2}{\lim_{n\to\infty} \||\nabla|^\frac\al2 v_n^1(\tau_n^{m_0})\|_{L^2}^2} \\
&\ge \||\nabla|^\frac\al2 W_\alpha\|_{L^2}^2 - \varepsilon - 2\eta^{D'} + \eta^{2D'}/M^2.
\end{align*}
Since $\varepsilon$ and $\eta$ can be taken arbitrarily small, we get the desired result.
\section{Proof of finite time Blowup}
Let us denote $\sup_{0 \le t < T^*}\||\nabla|^\frac\al2 u(t)\|_{L^2}$ by $M$ and $\|\varphi\|_{L^2}$ by $m$. We will show that $T^* = T^*(\varphi, M) < +\infty$. From the regularity persistence it follow that if $\varphi \in H^2$, then $u \in C([0, T^*); H^2)$ (this is the case for the power type since $\alpha < d < 3\alpha$ and thus $\frac{2\alpha}{d-\alpha} > 1$). Since the maximal existence time $T^* = T^*(\varphi)$ is lower semi-continuous, that is, if $\varphi_k \to \varphi$ in $H^\frac\al2$, then $T^*(\varphi) \le \liminf_{k\to \infty} T^*(\varphi_k)$, we may assume that $u \in C([0,T^*); H^{2})$ and $\varphi$ satisfies the condition \eqref{ass}.
\newcommand{\mathcal A}{\mathcal A}
\newcommand{\big[}{\big[}
\newcommand{\big]}{\big]}
\newcommand{\mbox{Re}}{\mbox{Re}}
\newcommand{\mathbf{m}_{1,\ld}}{\mathbf{m}_{1,\lambda}}
\newcommand{\widetilde{\mathbf{m}}_{1, \ld}}{\widetilde{\mathbf{m}}_{1, \lambda}}
\newcommand{\mathbf{m}_{2, \ld}}{\mathbf{m}_{2, \lambda}}
\newcommand{\mbox{Im}}{\mbox{Im}}
\subsection{Moment estimates}
\begin{proposition}\label{moment}
If $\varphi$ satisfies the condition \eqref{ass-mo}, then the solution $u \in C([0,T^*); H^2)$ satisfies that for each $t \in (0, T^*)$
\begin{align*}
\||x|u(t)\|_{L^2} \le CMt + \||x|\varphi\|_{L^2},\;\;\||x||\nabla|^{\alpha-1}u(t)\|_{L^2} + \||x|^2u(t)\|_{L^2} < +\infty.
\end{align*}
\end{proposition}
\begin{proof}
For a fixed radial bump function $\psi \in C_0^\infty$ with $\psi(x) = 1$ when $|x| \le 1$ and $\psi(x) = 0$ when $|x| \ge 2$ we denote $\psi(\frac{x}{\lambda})$ by $\psi_\lambda$ for $\lambda \ge 1$. Then we can define moments $\mathbf{m}_{1,\ld}, \widetilde{\mathbf{m}}_{1, \ld}, \mathbf{m}_{2, \ld}$ by
\begin{align*}
&\mathbf{m}_{1,\ld}^2 : = \big\langle x\psi_\lambda u ; x\psi_\lambda u \big\rangle,\\
&\widetilde{\mathbf{m}}_{1, \ld}^2 : = \big\langle x \psi_\lambda|\nabla|u; x\psi_\lambda |\nabla|u \big\rangle,\\
&\mathbf{m}_{2, \ld}^2 := \big\langle |x|^2\psi_\lambda u, |x|^2\psi_\lambda u\big\rangle.
\end{align*}
Differentiating $\mathbf{m}_{1,\ld}^2$ w.r.t $t$, we have
\begin{align*}
\frac{d}{dt}\mathbf{m}_{1,\ld}^2 &= 2\mbox{Im}\big\langle |\nabla|^{-\frac\al2}x \psi_\lambda (|\nabla|^\alpha u - V(u)u) ; |\nabla|^\frac\al2x \psi_\lambda u\big\rangle = 2\mbox{Im} \big\langle |\nabla|^{-\frac\al2}x\psi_\lambda |\nabla|^\alpha u ; |\nabla|^\frac\al2 x\psi_\lambda u\big\rangle\\
&= 2 \sum_{j=1}^d\mbox{Im} \big\langle \big[ x_j\psi_\lambda, |\nabla|^\alpha \big] u, x_j\psi_\lambda u\big\rangle \le 2\left(\sum_j\|\big[ x_j\psi_\lambda, |\nabla|^\alpha \big] u\|_{L^2}^2\right)^\frac12\mathbf{m}_{1,\ld}.
\end{align*}
In order to estimate the last term we use the following lemma.
\begin{lemma}\label{com-lem}
Let $\beta_\lambda(x) = \beta(\frac{x}{\lambda})$ for $\beta \in C_0^\infty$. If $s \ge 1$ for any $f \in H^{s-1}$ we have
$$
\|\big[ \beta_\lambda, |\nabla|^s \big] f\|_{L^2} \le C_\beta\lambda^{-1}\|f\|_{H^{s-1}}.
$$
If $0 < s < 1$, then for any $f \in L^2$ we have
$$
\|\big[ \beta_\lambda, |\nabla|^s \big] f\|_{L^2} \le C_\beta\lambda^{-s}\|f\|_{L^2}.
$$
\end{lemma}
From the above lemma it follows that
$$
\frac{d}{dt}\mathbf{m}_{1,\ld}^2 \le 2\lambda\left(\sum_j\|\big[ \frac{x_j}\lambda\psi_\lambda, |\nabla|^\alpha \big] u\|_{L^2}^2\right)^\frac12\mathbf{m}_{1,\ld} \le C\|u\|_{H^{\alpha-1}}\mathbf{m}_{1,\ld} \le CM\mathbf{m}_{1,\ld}
$$
and thus $\frac{d}{dt}\mathbf{m}_{1,\ld} \le CM$. Integrating over $[0, t]$, we have
$$
\mathbf{m}_{1,\ld}(t) \le CMt + \||x|\psi_\lambda \varphi\|_{L^2}.
$$
Letting $\lambda \to +\infty$, by Fatou's lemma we get the desired result.
Next we estimate $\widetilde{\mathbf{m}}_{1, \ld}$ as follows.
\begin{align*}
\frac{d}{dt}\widetilde{\mathbf{m}}_{1, \ld}^2 &= 2\mbox{Im} \big\langle |\nabla|^{-\frac\al2} x\psi_\lambda |\nabla|(|\nabla|^\alpha - V(u)u) ; |\nabla|^\frac\al2x \psi_\lambda |\nabla|u\big\rangle\\
&= 2\mbox{Im} \big\langle |\nabla|^{-\frac\al2}x\psi_\lambda |\nabla|^{\alpha} |\nabla|u ; |\nabla|^\frac\al2 x\psi_\lambda |\nabla| u\big\rangle - 2\mbox{Im} \big\langle x\psi_\lambda |\nabla|(V(u)u) ; x\psi_\lambda |\nabla|u\big\rangle\\
&= 2\sum_j\mbox{Im} \big\langle \big[ x_j \psi_\lambda, |\nabla|^\alpha \big] |\nabla|u, x_j\psi_\lambda |\nabla|u\big\rangle\\
& \quad + 2\sum_j\mbox{Im} \big\langle x_j\psi_\lambda \nabla|\nabla|^{-1}\cdot((\nabla V(u))u + V(u)\nabla u), x_j\psi_\lambda |\nabla|u\big\rangle\\
&\le C\|u\|_{H^\alpha}\widetilde{\mathbf{m}}_{1, \ld} + C\||x|((\nabla V(u))u + V(u)\nabla u)\|_{L^2}\widetilde{\mathbf{m}}_{1, \ld}.
\end{align*}
For the last term we used the weight estimate of the singular integral operator $\nabla|\nabla|^{-1}$ with $A_2$-weight $|x|$.
If $V(u) = |u|^\frac{2\alpha}{d-\alpha}$, then by Sobolev inequality \eqref{sobo-radial} we have
\begin{align*}
\||x|((\nabla V(u))u + V(u)\nabla u)\|_{L^2} &\le C\||u|^\frac{2\alpha}{d-\alpha}|\nabla u|\|_{L^2} + \||x|^\alpha |u|^\frac{2\alpha}{d-\alpha}|\nabla u|\|_{L^2}\\
&\le C\||\nabla|^\frac\al2 u\|_{L^2}^\frac{2\alpha}{d-\alpha}\|u\|_{H^{1+\frac\al2}} + \||\nabla|^\frac\al2 u\|_{L^2}^\frac{2\alpha}{d-\alpha}\||\nabla|u\|_{L^2}
\end{align*}
and thus by integrating over $[0, t]$
$$
\widetilde{\mathbf{m}}_{1, \ld}(t) \le \||x||\nabla|\varphi\|_{L^2} + C(1+M)^\frac{2\alpha}{d-\alpha}\int_0^t \|u(t')\|_{H^{1+\frac\al2}}\,dt'.
$$
If $V(u) = |x|^{-2\alpha}*|u|^2$, then from the fractional integration for radial function that \begin{align}\label{frac-int} |x|^\delta(|x|^{-\gamma}*|f|) \le C\||x|^{\delta-\gamma}f\|_{L^1}\;\; (0 <\delta \le \gamma < d-1)\end{align} we get
\begin{align*}
\||x||\nabla V(u)|u\|_{L^2} &\le \||\nabla V(u)|u\|_{L^2} + \||x|^{2\alpha}|\nabla V(u)|u\|_{L^2}\\
& \le \|\nabla V(u)\|_{L^\frac{2d}{\alpha}}\|u\|_{L^\frac{2d}{d-\alpha}} + C\|u\|_{L^2}^2\|\nabla u\|_{L^2}\\
&\le C\|u|\nabla u|\|_{L^\frac{2d}{2d-3\alpha}}\||\nabla|^\frac\al2 u\|_{L^2} + C\|\varphi\|_{L^2}^2\|\nabla u\|_{L^2}\\
&\le C\|u\|_{H^{1+\frac\al2}}^3
\end{align*}
and
\begin{align*}
\||x| V(u)|\nabla u|\|_{L^2} \le \|V(u)\|_{L^\infty}\|\nabla u\|_{L^2} + \|u\|_{L^2}^2\|\nabla u\|_{L^2} \le C\|u\|_{H^\alpha}^3.
\end{align*}
Thus $$\widetilde{\mathbf{m}}_{1, \ld}(t) \le \||x||\nabla|\varphi\|_{L^2} + C\int_0^t \|u(t')\|_{H^{1+\frac\al2}}^3\,dt'.$$
Fatou's lemma yields the desired results.
Similarly to the estimate of $\mathbf{m}_{1,\ld}$ we have for $\mathbf{m}_{2, \ld}$ that
\begin{align*}
\frac{d}{dt}\mathbf{m}_{2, \ld} &= 2\mbox{Im} \big\langle |x|^2\psi_\lambda (|\nabla|^\alpha u - V(u)u), |x|^2\psi_\lambda u\big\rangle\\
&= 2\mbox{Im} \big\langle |x|^2\psi_\lambda |\nabla|^\alpha u, |x|^2 \psi_\lambda u\big\rangle = 2\mbox{Im} \big\langle \psi_\lambda x \cdot (|\nabla|^\alpha x + \alpha |\nabla|^{\alpha-2}\nabla) u, |x|^2 \psi_\lambda u\big\rangle\\
&= 2\mbox{Im} \big\langle x\psi_\lambda \cdot |\nabla |^\alpha x u, |x|^2 \psi_\lambda u\big\rangle + 2\alpha \mbox{Im} \big\langle \psi_\lambda x \cdot \nabla/|\nabla||\nabla|^{\alpha-1} u, |x|^2 \psi_\lambda u\big\rangle\\
&= 2\sum_{j}\mbox{Im} \big\langle \big[ x_j\psi_\lambda, |\nabla|^\alpha\big] x_ju, |x|^2 \psi_\lambda u \big\rangle + 2\alpha \mbox{Im} \big\langle \psi_\lambda x \cdot \nabla/|\nabla||\nabla|^{\alpha-1} u, |x|^2 \psi_\lambda u\big\rangle.
\end{align*}
Lemma \ref{com-lem} shows that
$$
\frac{d}{dt}\mathbf{m}_{2, \ld} \le C(\||x|u\|_{H^{\alpha-1}} + \||x||\nabla|^{\alpha-1}u\|_{L^2})\mathbf{m}_{2, \ld} \le C(\|u\|_{H^{\alpha-1}} + \||x||\nabla|^{\alpha-1}u\|_{L^2})\mathbf{m}_{2, \ld},
$$
which implies that $$\||x|^2u(t)\|_{L^2} \le \||x|^2\varphi\|_{L^2} + C\int_0^t (\|u\|_{H^{\alpha-1}} + \||x||\nabla|^{\alpha-1}u\|_{L^2})\,dt'.$$
This completes the proof of Proposition \ref{moment}.
\end{proof}
\begin{proof}[Proof of Lemma \ref{com-lem}]
We show the first inequality. By Plancherel's theorem it suffices to show that $\|Tg\|_{L^2} \le C\|(1+ |\zeta|^{s-1})g\|_{L^2}$, where
$$
Tg(\xi) = \lambda^d\int \widehat\beta(\lambda(\xi-\zeta))(|\zeta|^s - |\xi|^s)g(\zeta)\,d\zeta.
$$
In fact,
\begin{align*}
|Tg(\xi)| &\le s\lambda^d \int |\widehat \beta(\lambda (\xi-\zeta))|(|\xi|^{s-1} + |\zeta|^{s-1})|\xi-\zeta||g(\zeta)|\,d\zeta\\
&\le s\lambda^d \int |\widehat \beta(\lambda (\xi-\zeta))||\xi-\zeta|^{s}|g(\zeta)|\,d\zeta + 2s\lambda^d \int |\widehat \beta(\lambda (\xi-\zeta))||\zeta|^{s-1}|\xi-\zeta||g(\zeta)|\,d\zeta\\
&= \lambda^{-s}s\lambda^d \int |\widehat \beta(\lambda (\xi-\zeta))||\lambda(\xi-\zeta)|^{s}|g(\zeta)|\,d\zeta\\
&\quad + 2s\lambda^d \int |\widehat \beta(\lambda (\xi-\zeta))|\lambda(\xi-\zeta)|||\zeta|^{s-1}g(\zeta)|\,d\zeta.
\end{align*}
Since $\widehat \beta(\xi)(|\xi|^\alpha + |\xi|)$ is integrable and $s > 1$, we get
$$
\|Tg\|_{L^2} \le C_\beta\lambda^{-1}\|(1+|\zeta|^{s-1})g\|_{L^2}.
$$
Similarly for $0 < s < 1$ we have
$$
|Tg(\xi)| \le \lambda^d \lambda^{-s} \int |\widehat \beta(\lambda (\xi-\zeta))|\lambda(\xi-\zeta)|^{s-1}|g(\zeta)|\,d\zeta
$$
and thus
$$
\|Tg\|_{L^2} \le C_\beta\lambda^{-s}\|g\|_{L^2}.
$$
This completes the proof of lemma.
\end{proof}
\subsection{Virial argument}
Here we consider the virial inequality through the moment estimates above.
Let us define two quantities associated with dilation and virial operators respectively by
$$
\mathcal A(u) := -{\rm Im}\big< u, x\cdot \nabla u \big> , \qquad \mathcal M(v) := \big<|\nabla|^{1-\frac\al2}(x u)\,;\, |\nabla|^{1-\frac\al2}(x u)\big>.
$$
From the regularity and moment estimates we can differentiate them w.r.t time.
\begin{align}\label{commut0}
\frac{d}{dt}\mathcal A(u(t)) = {\rm Re}\big\langle (|\nabla|^\alpha u - V(u)u),\, x\cdot \nabla u \big\rangle - {\rm Re}\big\langle u,\, x\cdot \nabla (|\nabla|^\alpha u - V(u)u) \big\rangle.
\end{align}
By integration by parts, we have
\begin{align*}
\frac{d}{dt}\mathcal A(u(t)) &= {\rm Re}\big\langle x |\nabla|^\alpha u\,; \nabla u \big\rangle + d \big\langle u, V(u) u \big\rangle - {\rm Re}\big\langle u, x\cdot \nabla ( |\nabla|^\alpha u) \big\rangle\\
&\qquad + 2\mbox{Re} \big\langle u, x \cdot \nabla (V(u) u) \big\rangle.
\end{align*}
Using the identity $x |\nabla|^\beta = |\nabla|^\beta x + \beta |\nabla|^{\beta-2}\nabla $ for $0 < \beta < 2$, we have
\begin{align}\begin{aligned}\label{diff}
\frac{d}{dt}\mathcal A(u(t)) &= \alpha \big\langle u, |\nabla|^\alpha u \big\rangle + 2 \mbox{Re} \big\langle u, x\cdot \nabla (V(u) u) \big\rangle + d \big\langle u, V(u) u \big\rangle.
\end{aligned}\end{align}
We first consider the power type. If $V(u) = |u|^\frac{2\alpha}{d-\alpha}$, then by direct calculation we get that
\begin{align*}
&2 \mbox{Re} \big\langle u, x\cdot \nabla (V(u) u) \big\rangle = - (d+\alpha) \mbox{Re} \big\langle u , V(u) u \big\rangle.
\end{align*}
Plugging this into \eqref{diff}, we have
\begin{align}\label{commut1}
\frac{d}{dt}\mathcal A(u(t)) = \alpha(\big[ \big\langle u, |\nabla|^\alpha u \big\rangle - \big\langle u, V(u) u \big\rangle\big]).
\end{align}
Now we consider the Hartree case $V(u) = |x|^{-2\alpha}*|u|^2$. Using integration by parts, we also get
$$
2\mbox{Re}\big\langle u, x\cdot \nabla(V(u) u)\big\rangle = \mbox{Re} \big\langle u, (x\cdot \nabla V(u)) u\big\rangle - d \big\langle u, V(u) u \big\rangle.
$$
Since $d > 2\alpha + 1$, by direct differentiation we have
\begin{align*}
\mbox{Re} \big\langle u, (x\cdot \nabla V(u)) u\big\rangle = - 2\alpha\big\langle u, V(u) u\big\rangle - 2\alpha\int\!\!\int |u(x)|^2 |x-y|^{-2\alpha-1}y\cdot \frac{x-y}{|x-y|}|u(y)|^2\,dxdy.
\end{align*}
In fact, from change of variables we deduce that
\begin{align*}
2\alpha\int\!\!\int |u(x)|^2 |x-y|^{-2\alpha-1}y\cdot \frac{x-y}{|x-y|}|u(y)|^2\,dxdy = \mbox{Re}\big\langle u, (x\cdot \nabla V(u)) u\big\rangle .
\end{align*}
So, we have
\begin{align*}
\mbox{Re} \big\langle u, (x\cdot \nabla V(u)) u\big\rangle = - \alpha\big\langle u, V(u) u\big\rangle.
\end{align*}
Putting all together, we finally have
\begin{align}\label{commut2}
\frac{d}{dt}\mathcal A(u(t)) = \alpha ( \big[\big\langle u, |\nabla|^\alpha u \big\rangle - \big\langle u, V(u) u \big\rangle \big]).
\end{align}
To deal with the RHS of \eqref{commut1} and \eqref{commut2} we introduce the following lemma to be shown in appendix.
\begin{lemma}\label{moment}
If $E(\varphi) \le (1-\delta_0) E(W_\alpha)$ and $\||\nabla|^\frac\al2 \varphi\|_{L^2} \ge \||\nabla|^\frac\al2 W_\alpha\|_{L^2}$ for some $0 < \delta_0 < 1$, then there exists a positive $\overline \delta$ such that $\||\nabla|^\frac\al2 u(t)\|_{L^2}^2 \ge (1 + \overline \delta)\||\nabla|^\frac\al2 W_\alpha\|_{L^2}^2$ for all $t \in (0, T^*)$.
\end{lemma}
From Proposition \ref{trap} it follows that
\begin{align*}
\big\langle u, |\nabla|^\alpha u \big\rangle - \big\langle u, V(u) u \big\rangle &= \mu\left(\frac12\||\nabla|^\frac\al2 u\|_{L^2}^2 - \frac1{\mu}\int V(u)|u|^2\,dx \right) - \frac{\mu-2}2\||\nabla|^\frac\al2 u\|_{L^2}^2\\
& = \mu E(\varphi) - \frac{\mu-2}2\||\nabla|^\frac\al2 u\|_{L^2}^2) \le \mu E(W_\alpha) - \frac{\mu-2}2(1 + \overline \delta)\||\nabla|^\frac\al2 W_\alpha\|_{L^2}^2\\
&= - \frac{\mu-2}2 \,\overline\delta \,C_{d, \alpha}^{-\frac2{\mu-2}} =: -\epsilon_0 < 0.
\end{align*}
Thus integrating \eqref{commut1} and \eqref{commut2} over $[0,t]$ we get
\begin{align}\label{dilation}
\mathcal A(u(t)) \le \mathcal A(\varphi) - \alpha\epsilon_0 t.
\end{align}
On the other hand by differentiating $\mathcal M$ and using the identity $x |\nabla|^\beta = |\nabla|^\beta x + \beta |\nabla|^{\beta-2}\nabla $ for $\beta = \alpha$ and $2-\alpha$, we get
\begin{align*}
\frac{d}{dt}\mathcal M(u) &= 2 \mbox{Im} \big\langle |\nabla |^{1-\alpha}x(|\nabla |^\alpha u - V(u)u) ; |\nabla |x u\big\rangle\\
&= 2\mbox{Im} \big\langle |\nabla |^{1-\alpha} x |\nabla |^\alpha u ; |\nabla |x u \big\rangle - 2\mbox{Im} \big\langle x V(u)u ; |\nabla|^{2-\alpha}x u\big\rangle\\
&= - 2\alpha\mbox{Im} \big\langle u, x\cdot \nabla u \big\rangle - 2\mbox{Im} \big\langle |x|^2 V(u)u ; |\nabla |^{2-\alpha}u\big\rangle - 2(2-\alpha)\mbox{Im} \big\langle x V(u)u ; |\nabla |^{-\alpha}\nabla u \big\rangle\\
&= 2\alpha \mathcal A(u) - 2\mbox{Im} \big\langle |x|^2 V(u)u , |\nabla |^{2-\alpha}u\big\rangle - 2(2-\alpha)\mbox{Im} \big\langle x V(u)u ; |\nabla |^{-\alpha}\nabla u \big\rangle.
\end{align*}
Since $|x|^2u \in L^2$, $V(u) \in L^\infty$ and $u \in H^{1 + \frac\al2}$, the second term of last line is at least well-defined.
Actually, it is possible to get a better estimate as below.
If $V(u) = |u|^\frac{2\alpha}{d-\alpha}$, then since $\frac 43 \le \alpha < 2$ we have
\begin{align*}
- 2\mbox{Im} \big\langle |x|^2 V(u)u ; |\nabla |^{2-\alpha}u\big\rangle &\le 2\||x|^\alpha V(u)\|_{L^\infty}\||x|^{2-\alpha}u\|_{L^2}\||\nabla|^{2-\alpha}u\|_{L^2}\\
&\le Cm^{\frac{\alpha^2 + 2\alpha -4}{\alpha}}M^{\frac{2\alpha}{d-\alpha}+\frac{4-2\alpha}{\alpha}}\||x|u\|_{L^2}^{2-\alpha}.
\end{align*}
From Lemma \ref{com-lem} it follows that
$$
- 2\mbox{Im} \big\langle |x|^2 V(u)u ; |\nabla |^{2-\alpha}u\big\rangle \le C m^{\frac{\alpha^2 + 2\alpha -4}{\alpha}}M^{\frac{2\alpha}{d-\alpha}+\frac{4-2\alpha}{\alpha}}(Mt + m_1)^{2-\alpha},
$$
where $m_1 = \||x|\varphi\|_{L^2}$.
On the other hand, the last term is bounded by
\begin{align}\label{sec}
C\int |\nabla|^{-(\alpha-1)}(|\cdot|^{-(\alpha-1)}g)(x)f(x)\,dx = C\int\!\!\int \frac{f(x)g(y)}{|x-y|^{-(d-(\alpha-1))}|y|^{\alpha-1}}\,dxdy,
\end{align}
where $f = |\nabla/|\nabla| u|$ and $g = x |x|^{\alpha-1} V(u)u$. For this we use Stein-Weiss inequality that
\begin{align}\label{s-t}
\left|\int\!\int \frac{f(x)\overline g(y)}{|x|^{\theta_1}|x-y|^{\theta}|y|^{\theta_2}}\,dxdy\right| \le C\|f\|_{L^{p_1}}\|g\|_{L^{p_2}},
\end{align}
provided that $1 < p_1, p_2 < \infty$, $\theta_1 + \theta_2 \ge 0$, $0 < \theta < d$, $\frac1{p_1} + \frac1{p_2} + \frac{\theta+\theta_1+\theta_2}{d} = 2$ and $\theta_1 < \frac d{p_1'}$, $\theta_2 < \frac{d}{p_2'}$.
Let $p_1 = p_2 = 2$ and $\theta_1 = 0, \theta_2 = \alpha-1, \theta = d-(\alpha-1)$. Then \eqref{s-t} implies that
$$
\eqref{sec} \le Cm^2M^\frac{2\alpha}{d-\alpha}.
$$
These estimates lead us to
\begin{align}\label{virial-p}
\mathcal M(u(t)) \le -\alpha^2 \epsilon_0t^2 + \left(C(m, M)(Mt + m_1)^{3-\alpha} + (Cm^2M^\frac{2\alpha}{d-\alpha} + \mathcal A(\varphi))t\right) + \mathcal M(\varphi).
\end{align}
We then consider the Hartree case. We follow the same strategy as in \cite{chkl}. To begin with
let us observe that
$$
2\mbox{Im} \big\langle |x|^2 V(u)u , |\nabla |^{2-\alpha}u\big\rangle = \mbox{Im} \big\langle \big[ |\nabla|^{2-\alpha}, g\big] u, u \big\rangle,
$$
where $g = |x|^2V(u)$. Then by the commutator estimate of \cite{chkl} one can get
$$
\|\big[ |\nabla|^{2-\alpha}, g\big] u\|_{L^2} \le C m \sup_{x\neq y}\frac{|g(x) - g(y)|}{|x-y|^{2-\alpha}}.
$$
If $x\neq y$, then
$$
|g(x) - g(y)| \le |x-y|\int_0^1 |\nabla g(z_s)|\,ds, \;\;z_s = x + s(y-x).
$$
Since $|\nabla g(z_s)| \le 2|z_s|V(u) + 2\alpha |z_s|^2\int |z_s-y|^{-(2\alpha+1)}|u(y)|^2\,dy$ and $d > 2\alpha+2$, by \eqref{frac-int} and Hardy-Sobolev inequality we have
$$
|\nabla g(z_s)| \le C|z_s|^{1-\alpha}(\||x|^{-2\alpha+\alpha}|u|^2\|_{L^1} + \||x|^{-2\alpha-1+1+\alpha}|u|^2\|_{L^1}) \le C|z_s|^{1-\alpha}\||\nabla|^\frac\al2 u\|_{L^2}^2 \le CM^2|z_s|^{1-\alpha}.
$$
Thus $|g(x) - g(y)| \le C|x-y|^{2-\alpha}M^2$, which implies that
$$
|\mbox{Im} \big\langle |x|^2 V(u)u , |\nabla |^{2-\alpha}u\big\rangle| \le CM^2m^2.
$$
Moreover, since by \eqref{frac-int} $|x|^\alpha V(u) \le C\||x|^{-\alpha}|u|^2\|_{L^2} \le CM^2$, from \eqref{sec} and \eqref{s-t} we have
$$
|\mbox{Im} \big\langle x V(u)u ; |\nabla |^{-\alpha}\nabla u \big\rangle| \le CM^2m^2.
$$
Therefore we get
\begin{align}\label{virial-h}
\mathcal M(u(t)) \le -\alpha^2\epsilon_0 t^2 + (Cm^2M^2 + \mathcal A(\varphi)) t + \mathcal M(\varphi).
\end{align}
Since $\mathcal M(u)$ is non-negative, by \eqref{virial-p} and \eqref{virial-h} we deduce that $T^* < +\infty$.
\section{Appendix}
We consider the characterization of maximizer of \eqref{bc} only for Hartree equation. For this we study a minimization problem:
\begin{align}\label{min1}
m = \inf_{u\in H^\frac\al2,\;\; \int V(u)|u|^2dx \neq 0} I(u), \quad I(u) := \frac{\||\nabla|^\frac\al2 u\|_{L^2}^4}{\int V(u)|u|^2dx}.
\end{align}
This is equivalent to the constrained minimization problem:
\begin{align}\label{min2}
m = \inf_{u\in H^\frac\al2,\;\; \int V(u)|u|^2dx = 1} J(u),\quad J(u):=\||\nabla|^\frac\al2 u\|_{L^2}^4.
\end{align}
By Sobolev embedding one can observe that $m > 0$. Suppose that $\underline u \in H^\frac\al2$ is a minimizer of \eqref{min2}. Then since $J$ is Fr\'{e}chet differentiable on $H^\frac\al2$, for any $\phi \in C_0^\infty$
$J$ should satisfy that
$$
\frac{d}{d\varepsilon}J(v_\varepsilon)\Big|_{\varepsilon=0} = 0, \;\;\mbox{where}\;\; v_\varepsilon = \frac {\underline u + \varepsilon \phi}{\left(\int V(\underline u + \varepsilon \phi)|\underline u + \varepsilon \phi|^2\,dx\right)^\frac14}.
$$
By direct calculation we conclude that
$$
\big\langle |\nabla|^\alpha \underline u - m^\frac12 V(\underline u)\underline u, \phi\big\rangle = 0.
$$
which means $\underline u$ is a solution to $|\nabla|^\alpha w - m^\frac12V(w)w$. By using a change of variables it is also a solution to \eqref{ell eqn}. Thus the minimizer $\underline u$ is $e^{i\theta}\lambda^\frac{d-\alpha}{2}W_\alpha(\lambda(x-x_0))$. Here we note that $W_\alpha \in H^\frac\al2$ because $d > 2\alpha$. Now it remains to show that $J$ attains $m$ in $H^\frac\al2$.
In fact, the minimizer can be found in $H^\frac\al2_{rad}$. Choose a minimizing sequence $u_j \in H^\frac\al2_{rad}$ with $\int V(u_j)|u_j|^2\,dx = 1$. Then it is bounded in $H^\frac\al2_{rad}$ and thus we can take a subsequence converging weakly to $\underline u$. According to Lemma 5.2 of \cite{dass} $\int V(u_j)|u_j|^2\,dx \to \int V(u)|u|^2\,dx$ due to the radial symmetry, which implies that $\int V(\underline u)|\underline u|^2\,dx = 1$. By the lower semi-continuity we deduce that $m \le \||\nabla|^\frac\al2 \underline u\|_{L^2}^4 \le \liminf_{j \to \infty}\||\nabla|^\frac\al2 u_j\|_{L^2}^4 = m$. Therefore $\underline u$ is a minimizer.
|
\section{Introduction}
The scattering amplitude for one-gluon exchange in an $SU(N_c)$
gauge theory is proportional to
\begin{eqnarray}
t^a_{ki}t^a_{lj}&=&-\frac{N_c+1}{4N_c}(\delta_{jk}\delta_{il}-\delta_{ik}\delta_{jl})
+\frac{N_c-1}{4N_c}(\delta_{jk}\delta_{il}+\delta_{ik}\delta_{jl})\, ,
\end{eqnarray}
where the $t^a$ is the generator of the gauge group, and the $i,j$
and $k,l$ are the color indexes of the two quarks in the incoming
and outgoing channels respectively. For $N_c=3$, the negative sign
in front of the antisymmetric antitriplet indicates the interaction
is attractive and favors the formation of the diquark states in the color
antitriplet, while the positive sign in front of the symmetric
sextet indicates
the interaction is repulsive and disfavors the formation of the diquark states in the color
sextet \cite{Huang-2005}.
The antitriplet diquark states have five Dirac tensor structures, scalar $C\gamma_5$,
pseudoscalar $C$, vector $C\gamma_\mu \gamma_5$, axial vector
$C\gamma_\mu $ and tensor $C\sigma_{\mu\nu}$. The structures
$C\gamma_\mu $ and $C\sigma_{\mu\nu}$ are symmetric, the structures
$C\gamma_5$, $C$ and $C\gamma_\mu \gamma_5$ are antisymmetric. The
attractive interactions of one-gluon exchange favor formation of
the diquarks in color antitriplet $\overline{3}_{ c}$, flavor
antitriplet $\overline{3}_{ f}$ and spin singlet $1_s$ (or flavor
sextet $6_{ f}$ and spin triplet $3_s$) \cite{One-gluon-1,One-gluon-2}, so the favored configurations are the scalar and axial-vector diquark states.
The scalar ($S$) and axial-vector ($A$) heavy-light diquark states have almost degenerate masses from the QCD sum rules \cite{WangDiquark-1,WangDiquark-2}.
In Refs.\cite{Wang-Scalar-1,Wang-Scalar-2}, we take the $C\gamma_5-C\gamma_5$, $C\gamma_\mu-C\gamma^\mu$, $C\gamma_\mu\gamma_5-C\gamma^\mu\gamma_5$ type interpolating currents to study the masses of the scalar tetraquark states in a systematic way using the QCD sum rules, and observe that the $S\bar{S}$ and $A\bar{A}$ type scalar tetraquark states have almost degenerate masses, about $4.36\,\rm{GeV}$, which is much larger than that from the phenomenological models \cite{EFG-2008,Maiani-2004,Maiani-2014}.
In Ref.\cite{EFG-2008}, Ebert, Faustov and Galkin calculate the masses of the excited heavy tetraquarks with hidden charm
within the relativistic diquark-antidiquark picture based on the quasipotential approach, and obtain the values $M_{J=0}=3.852\,\rm{GeV}$ and $3.812\,\rm{GeV}$ for the $A\bar{A}$ and $S\bar{S}$ type scalar tetraquark states $cq\bar{c}\bar{q}$, respectively. While L. Maiani et al obtain the values $M_{J=0}=3.832\,\rm{GeV}$ and $3.723\,\rm{GeV}$ for the $A\bar{A}$ and $S\bar{S}$ type scalar tetraquark states $cq\bar{c}\bar{q}$ respectively in the type-I diquark-antidiquark model \cite{Maiani-2004}, and
$M_{J=0}=4.000\,\rm{GeV}$ and $3.770\,\rm{GeV}$ for the $A\bar{A}$ and $S\bar{S}$ type scalar tetraquark states $cq\bar{c}\bar{q}$ respectively in the type-II diquark-antidiquark model \cite{Maiani-2014}. In those model-dependent studies, the masses of the $A\bar{A}$-type scalar tetraquark states are larger than that of the $S\bar{S}$ type scalar tetraquark states.
In Refs.\cite{WangHuangTao-1,WangHuangTao-2,WangHuangTao-3,Wang-Cu-Cu,WangHuang-molecule-1,WangHuang-molecule-2}, we explore the energy scale dependence of the hidden charmed (bottom) tetraquark states and molecular states in details for the first time, and suggest a formula
\begin{eqnarray}
\mu&=&\sqrt{M^2_{X/Y/Z}-(2{\mathbb{M}}_Q)^2} \, ,
\end{eqnarray}
with the effective heavy $Q$-quark mass ${\mathbb{M}}_Q$ to determine the energy scales of the QCD spectral densities in the QCD sum rules, which works well.
According to the formula, the energy scale $\mu=1\,\rm{GeV}$ taken in Refs.\cite{Wang-Scalar-1,Wang-Scalar-2} is too low to result in robust predictions.
In Ref.\cite{Wang-Cu-Cu}, we choose the $C\gamma_\mu-C\gamma_\nu$ type interpolating currents to study the $A\bar{A}$-type scalar, axial-vector and tensor tetraquark states in details with the QCD sum rules. The predicted masses of the axial-vector and tensor tetraquark states favor assigning the $Z_c(4020)$ and $Z_c(4025)$ as the $J^{PC}=1^{+-}$ or $2^{++}$ diquark-antidiquark type tetraquark states. While there are no experimental candidates to match the predicted mass of the scalar tetraquark state $M_{J=0}=\left(3.85^{+0.15}_{-0.09}\right)\,\rm{GeV}$. The value is consistent with the prediction $M_{J=0}=3.852\,\rm{GeV}$ based on the quasipotential approach \cite{EFG-2008}, while the upper bound reaches the prediction $M_{J=0}=4.000\,\rm{GeV}$ based on the type-II diquark-antidiquark model \cite{Maiani-2014}. According to Refs.\cite{EFG-2008,Maiani-2004,Maiani-2014}, the $S\bar{S}$-type scalar tetraquark states have smaller masses than that of the corresponding $A\bar{A}$-type scalar tetraquark states. It is interesting to see whether or not such conclusion survives when confronted with the QCD sum rules. In Refs.\cite{WangHuangTao-1,WangHuangTao-2,WangHuangTao-3}, we observe that the masses of the $S\bar{A}$ or $A\bar{S}$ type axial-vector tetraquark states are larger than that of the $A\bar{A}$ type scalar tetraquark states. So the $S\bar{S}$ scalar tetraquark state maybe the lowest tetraquark state.
In this article, we study the scalar $S\bar{S}$-type hidden charmed tetraquark state (thereafter we will denote it as $Z_c$) by calculating the contributions of the vacuum condensates up to dimension-10, and try to obtain the lowest mass based on the QCD sum rules. Furthermore, we calculate the hadronic coupling constants $G_{Z_c\eta_c\pi}$ and $G_{Z_cDD}$ with the three-point QCD sum rules, then study the strong decays $ Z_c\to \eta_c\pi\, ,\, DD$.
The article is arranged as follows: we derive the QCD sum rules for the mass and pole residue of the scalar tetraquark state $Z_c$ and for the hadronic coupling constants $G_{Z_c\eta_c\pi}$ and $G_{Z_cDD}$ in section 2; in section 3, we present the numerical results and discussions; section 4 is reserved for our conclusion.
\section{QCD sum rules for the scalar tetraquark state }
In the following, we write down the two-point correlation function $\Pi(p)$ in the QCD sum rules,
\begin{eqnarray}
\Pi(p)&=&i\int d^4x e^{ip \cdot x} \langle0|T\left\{J(x)J^{\dagger}(0)\right\}|0\rangle \, ,
\end{eqnarray}
\begin{eqnarray}
J(x)&=&\epsilon^{ijk}\epsilon^{imn}u^j(x)C\gamma_5 c^k(x) \bar{d}^m(x)\gamma_5 C \bar{c}^n(x) \, ,
\end{eqnarray}
where the $i$, $j$, $k$, $m$, $n$ are color indexes, the $C$ is the charge conjugation matrix.
At the hadronic side, we can insert a complete set of intermediate hadronic states with
the same quantum numbers as the current operator $J(x)$ into the
correlation function $\Pi(p)$ to obtain the hadronic representation
\cite{SVZ79-1,SVZ79-2,Reinders85}. After isolating the ground state
contribution of the scalar tetraquark state, we get the following result,
\begin{eqnarray}
\Pi(p)&=&\frac{\lambda_{ Z_c}^2}{M_{Z_c}^2-p^2} +\cdots \, \, ,
\end{eqnarray}
where the pole residue $\lambda_{Z_c}$ is defined by $\langle 0|J(0)|Z(p)\rangle = \lambda_{Z_c}$ .
In the following, we briefly outline the operator product expansion for the correlation function $\Pi(p)$ in perturbative QCD. We contract the $u$, $d$ and $c$ quark fields in the correlation function $\Pi(p)$ with Wick theorem, and obtain the result:
\begin{eqnarray}
\Pi(p)&=&i\epsilon^{ijk}\epsilon^{imn}\epsilon^{i^{\prime}j^{\prime}k^{\prime}}\epsilon^{i^{\prime}m^{\prime}n^{\prime}}\int d^4x e^{ip \cdot x} \nonumber\\
&&{\rm Tr}\left[ \gamma_{5}C^{kk^{\prime}}(x)\gamma_{5} CU^{jj^{\prime}T}(x)C\right] {\rm Tr}\left[ \gamma_{5} C^{n^{\prime}n}(-x)\gamma_{5} C D^{m^{\prime}mT}(-x)C\right] \, ,
\end{eqnarray}
where the $U_{ij}(x)$, $D_{ij}(x)$ and $C_{ij}(x)$ are the full $u$, $d$ and $c$ quark propagators respectively (the $U_{ij}(x)$ and $D_{ij}(x)$ can be written as $S_{ij}(x)$ for simplicity),
\begin{eqnarray}
S_{ij}(x)&=& \frac{i\delta_{ij}\!\not\!{x}}{ 2\pi^2x^4}-\frac{\delta_{ij}\langle
\bar{q}q\rangle}{12} -\frac{\delta_{ij}x^2\langle \bar{q}g_s\sigma Gq\rangle}{192} -\frac{ig_sG^{a}_{\alpha\beta}t^a_{ij}(\!\not\!{x}
\sigma^{\alpha\beta}+\sigma^{\alpha\beta} \!\not\!{x})}{32\pi^2x^2} -\frac{i\delta_{ij}x^2\!\not\!{x}g_s^2\langle \bar{q} q\rangle^2}{7776}\nonumber\\
&& -\frac{\delta_{ij}x^4\langle \bar{q}q \rangle\langle g_s^2 GG\rangle}{27648} -\frac{1}{8}\langle\bar{q}_j\sigma^{\mu\nu}q_i \rangle \sigma_{\mu\nu}-\frac{1}{4}\langle\bar{q}_j\gamma^{\mu}q_i\rangle \gamma_{\mu }+\cdots \, ,
\end{eqnarray}
\begin{eqnarray}
C_{ij}(x)&=&\frac{i}{(2\pi)^4}\int d^4k e^{-ik \cdot x} \left\{
\frac{\delta_{ij}}{\!\not\!{k}-m_c}
-\frac{g_sG^n_{\alpha\beta}t^n_{ij}}{4}\frac{\sigma^{\alpha\beta}(\!\not\!{k}+m_c)+(\!\not\!{k}+m_c)
\sigma^{\alpha\beta}}{(k^2-m_c^2)^2}\right.\nonumber\\
&&\left. +\frac{g_s D_\alpha G^n_{\beta\lambda}t^n_{ij}(f^{\lambda\beta\alpha}+f^{\lambda\alpha\beta}) }{3(k^2-m_c^2)^4}-\frac{g_s^2 (t^at^b)_{ij} G^a_{\alpha\beta}G^b_{\mu\nu}(f^{\alpha\beta\mu\nu}+f^{\alpha\mu\beta\nu}+f^{\alpha\mu\nu\beta}) }{4(k^2-m_c^2)^5}+\cdots\right\} \, ,\nonumber\\
f^{\lambda\alpha\beta}&=&(\!\not\!{k}+m_c)\gamma^\lambda(\!\not\!{k}+m_c)\gamma^\alpha(\!\not\!{k}+m_c)\gamma^\beta(\!\not\!{k}+m_c)\, ,\nonumber\\
f^{\alpha\beta\mu\nu}&=&(\!\not\!{k}+m_c)\gamma^\alpha(\!\not\!{k}+m_c)\gamma^\beta(\!\not\!{k}+m_c)\gamma^\mu(\!\not\!{k}+m_c)\gamma^\nu(\!\not\!{k}+m_c)\, ,
\end{eqnarray}
and $t^n=\frac{\lambda^n}{2}$, the $\lambda^n$ is the Gell-Mann matrix, $D_\alpha=\partial_\alpha-ig_sG^n_\alpha t^n$ \cite{Reinders85}, then compute the integrals both in the coordinate and momentum spaces to obtain the correlation function $\Pi(p)$ therefore the QCD spectral density.
In Eq.(7), we retain the terms $\langle\bar{q}_j\sigma_{\mu\nu}q_i \rangle$ and $\langle\bar{q}_j\gamma_{\mu}q_i\rangle$ originate from the Fierz re-arrangement of the $\langle q_i \bar{q}_j\rangle$ to absorb the gluons emitted from the heavy quark lines so as to extract the mixed condensate and four-quark condensate $\langle\bar{q}g_s\sigma G q\rangle$ and $g_s^2\langle\bar{q}q\rangle^2$, respectively.
Once the analytical expression is obtained, we can take the
quark-hadron duality below the continuum threshold $s_0$ and perform Borel transform with respect to
the variable $P^2=-p^2$ to obtain the following QCD sum rule:
\begin{eqnarray}
\lambda^2_{Z_c}\, \exp\left(-\frac{M^2_{Z_c}}{T^2}\right)= \int_{4m_c^2}^{s_0} ds\, \rho(s) \, \exp\left(-\frac{s}{T^2}\right) \, ,
\end{eqnarray}
where
\begin{eqnarray}
\rho(s)&=&\rho_{0}(s)+\rho_{3}(s) +\rho_{4}(s)+\rho_{5}(s)+\rho_{6}(s)+\rho_{7}(s) +\rho_{8}(s)+\rho_{10}(s)\, ,
\end{eqnarray}
\begin{eqnarray}
\rho_{0}(s)&=&\frac{1}{512\pi^6}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, yz\, (1-y-z)^3\left(s-\overline{m}_c^2\right)^2\left(7s^2-6s\overline{m}_c^2+\overline{m}_c^4 \right) \, ,
\end{eqnarray}
\begin{eqnarray}
\rho_{3}(s)&=&-\frac{m_c\langle \bar{q}q\rangle}{16\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, (y+z)(1-y-z)\left(s-\overline{m}_c^2\right)\left(2s-\overline{m}_c^2\right) \, ,
\end{eqnarray}
\begin{eqnarray}
\rho_{4}(s)&=&-\frac{m_c^2}{384\pi^4} \langle\frac{\alpha_s GG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left( \frac{z}{y^2}+\frac{y}{z^2}\right)(1-y-z)^3 \nonumber\\
&&\left\{ 2s-\overline{m}_c^2+\frac{\overline{m}_c^4}{6}\delta\left(s-\overline{m}_c^2\right)\right\} \nonumber\\
&&+\frac{1}{512\pi^4} \langle\frac{\alpha_s GG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left( y+z\right)(1-y-z)^2 \left( 10s^2-12s\overline{m}_c^2+3\overline{m}_c^4\right) \, ,
\end{eqnarray}
\begin{eqnarray}
\rho_{5}(s)&=&\frac{m_c\langle \bar{q}g_s\sigma Gq\rangle}{64\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, (y+z) \left(3s-2\overline{m}_c^2 \right) \nonumber\\
&&-\frac{m_c\langle \bar{q}g_s\sigma Gq\rangle}{64\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, \left( \frac{y}{z}+\frac{z}{y}\right) (1-y-z) \left(3s-2\overline{m}_c^2 \right) \, ,
\end{eqnarray}
\begin{eqnarray}
\rho_{6}(s)&=&\frac{m_c^2\langle\bar{q}q\rangle^2}{12\pi^2}\int_{y_i}^{y_f}dy +\frac{g_s^2\langle\bar{q}q\rangle^2}{108\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz\, yz \left\{2s-\overline{m}_c^2 +\frac{\overline{m}_c^4}{6}\delta\left(s-\overline{m}_c^2 \right)\right\}\nonumber\\
&&-\frac{g_s^2\langle\bar{q}q\rangle^2}{512\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, (1-y-z)\left\{ 2\left(\frac{z}{y}+\frac{y}{z} \right)\left(3s-2\overline{m}_c^2 \right)+\left(\frac{z}{y^2}+\frac{y}{z^2} \right)\right.\nonumber\\
&&\left.m_c^2\left[ 2+ \overline{m}_c^2\delta\left(s-\overline{m}_c^2 \right)\right] \right\} \nonumber\\
&&-\frac{g_s^2\langle\bar{q}q\rangle^2}{3888\pi^4}\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \, (1-y-z)\left\{ 3\left(\frac{z}{y}+\frac{y}{z} \right)\left(3s-2\overline{m}_c^2 \right)+\left(\frac{z}{y^2}+\frac{y}{z^2} \right)\right. \nonumber\\
&&\left.m_c^2\left[ 2+\overline{m}_c^2\delta\left(s-\overline{m}_c^2\right)\right]+(y+z)\left[12\left(2s-\overline{m}_c^2\right) +2\overline{m}_c^4\delta\left(s-\overline{m}_c^2\right)\right] \right\}\, ,
\end{eqnarray}
\begin{eqnarray}
\rho_7(s)&=&\frac{m_c^3\langle\bar{q}q\rangle}{288\pi^2 }\langle\frac{\alpha_sGG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left(\frac{y}{z^3}+\frac{z}{y^3}+\frac{1}{y^2}+\frac{1}{z^2}\right)(1-y-z)\nonumber\\
&&\left(1+\frac{ \overline{m}_c^2}{T^2}\right) \delta\left(s-\overline{m}_c^2\right)\nonumber\\
&&-\frac{m_c\langle\bar{q}q\rangle}{96\pi^2}\langle\frac{\alpha_sGG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz \left(\frac{y}{z^2}+\frac{z}{y^2}\right)(1-y-z) \left\{2+\overline{m}_c^2\delta\left(s-\overline{m}_c^2\right) \right\}\nonumber\\
&&-\frac{m_c\langle\bar{q}q\rangle}{96\pi^2}\langle\frac{\alpha_sGG}{\pi}\rangle\int_{y_i}^{y_f}dy \int_{z_i}^{1-y}dz\left\{2+ \overline{m}_c^2 \delta\left(s-\overline{m}_c^2\right) \right\} \nonumber\\
&&-\frac{m_c\langle\bar{q}q\rangle}{576\pi^2}\langle\frac{\alpha_sGG}{\pi}\rangle\int_{y_i}^{y_f}dy \left\{2+ \widetilde{m}_c^2 \, \delta \left(s-\widetilde{m}_c^2\right) \right\}\, ,
\end{eqnarray}
\begin{eqnarray}
\rho_8(s)&=&-\frac{m_c^2\langle\bar{q}q\rangle\langle\bar{q}g_s\sigma Gq\rangle}{24\pi^2}\int_0^1 dy \left(1+\frac{\widetilde{m}_c^2}{T^2} \right)\delta\left(s-\widetilde{m}_c^2\right)\nonumber \\
&&+\frac{ m_c^2\langle\bar{q}q\rangle\langle\bar{q}g_s\sigma Gq\rangle}{48\pi^2}\int_{0}^{1} dy \frac{1}{y(1-y)}\delta\left(s-\widetilde{m}_c^2\right)
\, ,
\end{eqnarray}
\begin{eqnarray}
\rho_{10}(s)&=&\frac{m_c^2\langle\bar{q}g_s\sigma Gq\rangle^2}{192\pi^2T^6}\int_0^1 dy \, \widetilde{m}_c^4 \, \delta \left( s-\widetilde{m}_c^2\right)
\nonumber \\
&&-\frac{m_c^4\langle\bar{q}q\rangle^2}{216T^4}\langle\frac{\alpha_sGG}{\pi}\rangle\int_0^1 dy \left\{ \frac{1}{y^3}+\frac{1}{(1-y)^3}\right\} \delta\left( s-\widetilde{m}_c^2\right)\nonumber\\
&&+\frac{m_c^2\langle\bar{q}q\rangle^2}{72T^2}\langle\frac{\alpha_sGG}{\pi}\rangle\int_0^1 dy \left\{ \frac{1}{y^2}+\frac{1}{(1-y)^2}\right\} \delta\left( s-\widetilde{m}_c^2\right)\nonumber\\
&&-\frac{m_c^2\langle\bar{q}g_s\sigma Gq\rangle^2}{192 \pi^2T^4} \int_0^1 dy \frac{1}{y(1-y)} \widetilde{m}_c^2 \, \delta\left( s-\widetilde{m}_c^2\right)\nonumber\\
&&+\frac{m_c^2\langle\bar{q}g_s\sigma Gq\rangle^2}{384 \pi^2T^2} \int_0^1 dy \frac{1}{y(1-y)} \delta\left( s-\widetilde{m}_c^2\right)\nonumber \\
&&+\frac{m_c^2\langle\bar{q} q\rangle^2}{216 T^6}\langle\frac{\alpha_sGG}{\pi}\rangle\int_0^1 dy \, \widetilde{m}_c^4 \, \delta \left( s-\widetilde{m}_c^2\right) \, ,
\end{eqnarray}
the subscripts $0$, $3$, $4$, $5$, $6$, $7$, $8$, $10$ denote the dimensions of the vacuum condensates, $y_{f}=\frac{1+\sqrt{1-4m_c^2/s}}{2}$,
$y_{i}=\frac{1-\sqrt{1-4m_c^2/s}}{2}$, $z_{i}=\frac{y
m_c^2}{y s -m_c^2}$, $\overline{m}_c^2=\frac{(y+z)m_c^2}{yz}$,
$ \widetilde{m}_c^2=\frac{m_c^2}{y(1-y)}$, $\int_{y_i}^{y_f}dy \to \int_{0}^{1}dy$, $\int_{z_i}^{1-y}dz \to \int_{0}^{1-y}dz$ when the $\delta$ functions $\delta\left(s-\overline{m}_c^2\right)$ and $\delta\left(s-\widetilde{m}_c^2\right)$ appear. We take into account the vacuum condensates which are
vacuum expections of the operators of the orders $\mathcal{O}( \alpha_s^{k})$ with $k\leq 1$ consistently.
Differentiate Eq.(9) with respect to $\frac{1}{T^2}$, then eliminate the
pole residues $\lambda_{Z_c}$, we obtain the QCD sum rule for
the mass of the scalar tetraquark state,
\begin{eqnarray}
M^2_{Z_c}= \frac{\int_{4m_c^2}^{s_0} ds\frac{d}{d \left(-1/T^2\right)}\rho(s)\exp\left(-\frac{s}{T^2}\right)}{\int_{4m_c^2}^{s_0} ds \rho(s)\exp\left(-\frac{s}{T^2}\right)}\, .
\end{eqnarray}
In the following, we perform Fierz re-arrangement to the current $J$ both in the color and Dirac-spinor spaces to obtain the result,
\begin{eqnarray}
J&=&\frac{1}{4}\left\{\,-\bar{c} c\,\bar{d} u+\bar{c}i\gamma_5 c\,\bar{d}i\gamma_5 u-\bar{c} \gamma^\mu c\,\bar{d}\gamma_\mu u-\bar{c} \gamma^\mu\gamma_5 c\,\bar{d}\gamma_\mu\gamma_5 u+\frac{1}{2}\bar{c}\sigma_{\mu\nu} c\,\bar{d}\sigma^{\mu\nu} u\right. \nonumber\\
&&\left.+\bar{c} u\,\bar{d} c-\bar{c}i\gamma_5 u\,\bar{d}i\gamma_5 c+\bar{c} \gamma^\mu u\,\bar{d}\gamma_\mu c+\bar{c} \gamma^\mu\gamma_5 u\,\bar{d}\gamma_\mu\gamma_5 c-\frac{1}{2}\bar{c}\sigma_{\mu\nu} u\,\bar{d}\sigma^{\mu\nu} c \,\right\} \, ,
\end{eqnarray}
the components couple to the meson pairs $\chi_{c0}a_0^{+}(980)$, $\eta_c\pi^{+}$, $J/\psi \rho^{+}$, $\chi_{c1}\pi^+$, $\chi_{c1}a_1^+(1260)$, $h_c h_1^+(1170)$, $(D_0(2400)\bar{D}_0(2400))^+$,
$(D\bar{D})^+$, $(D^*\bar{D}^*)^+$, $(D_1(2420)\bar{D}_1(2420))^+$, $(D_1(2430)\bar{D}_1(2430))^+$, respectively. The strong decays
\begin{eqnarray}
Z_c^{\pm}(0^{++}) &\to& \chi_{c0}a_0^{\pm}(980)\, , \, \eta_c\pi^{\pm}\, , \, J/\psi \rho^{\pm}\, , \, \chi_{c1}\pi^\pm\, , \, \chi_{c1}a_1^\pm(1260)\, , \, h_c h_1^\pm(1170)\, , \,(D_0(2400)\bar{D}_0(2400))^\pm \, , \, \nonumber\\
&& (D\bar{D})^\pm \, , \, (D^*\bar{D}^*)^\pm \, , \, (D_1(2420)\bar{D}_1(2420))^\pm \, , \, (D_1(2430)\bar{D}_1(2430))^\pm \, ,
\end{eqnarray}
are Okubo-Zweig-Iizuka super-allowed, if they are kinematically allowed.
The
diquark-antidiquark type tetraquark state can be taken as a special superposition of a series of meson-meson pairs, and embodies the net effects. The decays to its components (meson-meson pairs) are Okubo-Zweig-Iizuka super-allowed, but the re-arrangements in the color-space are non-trivial \cite{Nielsen3900-1,Nielsen3900-2}.
The numerical analysis indicates that the ground state mass of the $S\bar{S}$-type scalar tetraquark state is about $3.82\,\rm{GeV}$,
the strong decays
\begin{eqnarray}
Z_c^{\pm}(0^{++}) &\to& \eta_c\pi^{\pm}\, , \, \chi_{c1}\pi^\pm\, , \, (D\bar{D})^\pm \, ,
\end{eqnarray}
are kinematically allowed. The decay $Z_c^{\pm}(0^{++}) \to \chi_{c1}\pi^\pm$ takes place through relative P-wave and is kinematically suppressed.
Now we write down the three-point correlation functions
$\Pi_{1}(p,q)$ and $\Pi_{2}(p,q)$ to study the strong decays $Z_c^{\pm}(0^{++}) \to \eta_c\pi^{\pm}\, , \, (D\bar{D})^\pm $,
\begin{eqnarray}
\Pi_{1}(p,q)&=&i^2\int d^4xd^4y e^{ip \cdot x}e^{iq \cdot y}\langle 0|T\left\{J_{\eta_c}(x)J_{\pi}(y)J(0)\right\}|0\rangle\, , \nonumber \\
\Pi_{2}(p,q)&=&i^2\int d^4xd^4y e^{ip\cdot x}e^{iq\cdot y}\langle 0|T\left\{J_{D^-}(x)J_{D^0}(y)J(0)\right\}|0\rangle \, ,
\end{eqnarray}
where the currents
\begin{eqnarray}
J_{\eta_c}(x)&=&\bar{c}(x)i\gamma_5 c(x) \, ,\nonumber \\
J_\pi(y)&=&\bar{u}(y)i\gamma_5 d(y) \, ,\nonumber \\
J_{D^-}(x)&=&\bar{c}(x)i\gamma_5 d(x) \, ,\nonumber \\
J_{D^0}(y)&=&\bar{u}(y)i\gamma_5 c(y) \, ,
\end{eqnarray}
interpolate the mesons $\eta_c$, $\pi$, $D^-$, $D^0$, respectively.
We insert a complete set of intermediate hadronic states with
the same quantum numbers as the current operators into the three-point
correlation functions $\Pi_{1}(p,q)$ and $\Pi_{2}(p,q)$ and isolate the ground state
contributions to obtain the following results,
\begin{eqnarray}
\Pi_{1}(p,q)&=& \frac{f_{\pi}M_{\pi}^2f_{\eta_c}M_{\eta_c}^2\lambda_{Z_c}G_{Z_c\eta_c \pi}}{2(m_u+m_d)m_c} \frac{-q\cdot p}{(M_{Z_c}^2-p^{\prime2})(M_{\eta_c}^2-p^2)(M_{\pi}^2-q^2)} +\cdots \, , \nonumber\\
\Pi_{2}(p,q)&=& \frac{f_{D}^2M_{D}^4\lambda_{Z_c}G_{Z_cDD}}{(m_c+m_q)^2} \frac{-q\cdot p}{(M_{Z_c}^2-p^{\prime2})(M_{D}^2-p^2)(M_{D}^2-q^2)} +\cdots \, ,
\end{eqnarray}
where $p^\prime=p+q$, the $f_D$, $f_{\eta_c}$ and $f_{\pi}$ are the decay constants of the mesons $D$, $\eta_c$ and $\pi$, respectively, the $G_{Z_c\eta_c\pi}$ and $G_{Z_cDD}$ are the hadronic coupling constants. In the following, we write down the definitions,
\begin{eqnarray}
\langle0|J_{\eta_c}(0)|\eta_c(p)\rangle&=&\frac{f_{\eta_c}M_{\eta_c}^2}{2m_c} \,\, , \nonumber \\
\langle0|J_{\pi}(0)|\pi(q)\rangle&=&\frac{f_{\pi}M_{\pi}^2}{m_u+m_d} \,\, ,\nonumber \\
\langle0|J_{D}(0)|D(p/q)\rangle&=&\frac{f_{D}M_{D}^2}{m_c+m_q} \,\, , \\
\langle\eta_c(p)\pi(q)|Z_c(p^{\prime})\rangle&=&-iq\cdot p G_{Z_c\eta_c\pi}(q^2) \, , \nonumber\\
\langle D(p)D(q)|Z_c(p^{\prime})\rangle&=&-iq\cdot p G_{Z_cDD}(q^2) \, .
\end{eqnarray}
\begin{figure}
\centering
\includegraphics[totalheight=3cm,width=7cm]{qqg-decay-etapi.eps}
\caption{The connected Feynman diagram contributes to the correlation function $\Pi_{1}(p,q)$, where the dashed and solid lines denote the heavy quark and light quark lines, respectively. Other diagrams obtained by interchanging of the heavy quark lines or light quark lines are
implied. }
\end{figure}
\begin{figure}
\centering
\includegraphics[totalheight=3cm,width=7cm]{qqg-decay-DD.eps}
\caption{The connected Feynman diagram contributes to the correlation function $\Pi_{2}(p,q)$, where the dashed and solid lines denote the heavy quark and light quark lines, respectively. Other diagrams obtained by interchanging of the heavy quark lines and (or) light quark lines are
implied. }
\end{figure}
We carry out the operator product expansion and take into account the color connected Feynman diagrams \cite{Nielsen3900-1,Nielsen3900-2}, and obtain the following results,
\begin{eqnarray}
\Pi_{1}(p,q)&=&-\frac{ m_c\langle \bar{q}g_s\sigma Gq\rangle }{32\pi^2q^2}\int_0^1 dx \frac{q\cdot p}{m_c^2-x(1-x)p^2}+\cdots \, ,
\end{eqnarray}
\begin{eqnarray}
\Pi_{2}(p,q)&=&-\frac{ m_c\langle \bar{q}g_s\sigma Gq\rangle }{64\pi^2}\frac{q \cdot p}{q^2-m_c^2}\int_0^1 dx \frac{1+x}{m_c^2-(1-x)p^2} \nonumber\\
&&-\frac{ m_c\langle \bar{q}g_s\sigma Gq\rangle }{64\pi^2}\frac{q \cdot p}{m_c^2-p^2} \int_0^1 dx \frac{2-x}{xq^2-m_c^2}+\cdots \, .
\end{eqnarray}
In Fig.1 and Fig.2, we draw the connected Feynman diagrams contribute to the correlation functions $\Pi_{1}(p,q)$ and $\Pi_{2}(p,q)$, respectively.
The $\Pi_{1}(p,q)$ and $\Pi_{2}(p,q)$ can be expanded in terms of the $\cos\theta$, $\Pi_{1/2}(p,q)=\Pi^0(p^2,q^2)+\Pi^1(p^2,q^2)\cos\theta+\Pi^2(p^2,q^2)\cos^2\theta+\cdots$, at the QCD side, where the $\theta$ is the included angle of the Euclidean momenta $p$ and $q$, i.e. $\cos\theta=p\cdot{q}/\sqrt{q^2p^2}$.
There exists only one term ($\Pi^1(p^2,q^2)\cos\theta$) for the $\Pi_{1}(p,q)$, while there exist two terms ($\Pi^0(p^2,q^2)$ and $\Pi^1(p^2,q^2)\cos\theta$) for
the $\Pi_{2}(p,q)$. At the phenomenological side, the hadronic coupling constants $G_{SPP^\prime}(p,q)$ have the possible forms $G_{SPP^\prime}^0$, $G_{SPP^\prime}^1\cos\theta$, $G_{SPP^\prime}^2\cos^2\theta$, $\cdots$, where the $S$ denotes the scalar mesons, the $P$ and $P^{\prime}$ denote the pseudoscalar mesons. In the present case, it is better to choose the form $G_{SPP^\prime}^1\cos\theta$, as the correlation functions $\Pi_{1}(p,q)$ and
$\Pi_{2}(p,q)$ both have the term proportional to $\cos\theta$ at the QCD side.
The $\cos\theta$ is the pertinent tensor structure, as the correlation functions $\Pi_{1}(p,q)$ and $\Pi_{2}(p,q)$ should have the same tensor structure at the phenomenological side.
There exists some shortcoming,
if we choose the form $G_{SPP^\prime}^0$ and take the replacement $2p\cdot{q}=p^{\prime2}-p^2-q^2$, then set $p^2=p^{\prime 2}$ and perform the Borel transform with respect to the variable $P^2=-p^2$, as the $p$, $q$ and $p^{\prime}$ are not independent variables, the $\cos\theta$ cannot be replaced.
Once the analytical expressions of the correlation functions $\Pi_{1}(p,q)$ and $\Pi_{2}(p,q)$ at both the QCD side and hadron side are obtained, we perform the Borel transform with respect to the variable $P^2=-p^2$ by setting $p^2=p^{\prime 2}$, then take the quark-hadron duality and obtain the following QCD sum rules,
\begin{eqnarray}
&&\frac{f_{\pi}M_{\pi}^2f_{\eta_c}M_{\eta_c}^2\lambda_{Z_c}G_{Z_c\eta_c \pi}}{2(m_u+m_d)m_c(M_{Z_c}^2-M_{\eta_c}^2)} \left\{ \exp\left(-\frac{M_{\eta_c}^2}{T^2} \right)-\exp\left(-\frac{M_{Z_c}^2}{T^2} \right)\right\}+C_{Z_c\eta_c \pi} \exp\left(-\frac{s_0}{T^2} \right) \nonumber\\
&&=-\frac{m_c\langle\bar{q}g_s\sigma Gq\rangle}{32\pi^2}\frac{Q^2+M_{\pi}^2}{Q^2} \int_0^1 dx \frac{ 1}{x(1-x)}\exp\left( -\frac{m_c^2}{x(1-x)T^2}\right)\, ,
\end{eqnarray}
\begin{eqnarray}
&&\frac{f_{D}^2M_{D}^4\lambda_{Z_c}G_{Z_cDD}}{(m_c+m_q)^2 (M_{Z_c}^2-M_{D}^2)} \left\{ \exp\left(-\frac{M_{D}^2}{T^2} \right)-\exp\left(-\frac{M_{Z_c}^2}{T^2} \right)\right\}+C_{Z_cDD} \exp\left(-\frac{s_0}{T^2} \right) \nonumber\\
&&=-\frac{m_c\langle\bar{q}g_s\sigma Gq\rangle}{64\pi^2}(Q^2+M_{D}^2)\int_0^1 dx \left\{\frac{1}{Q^2+m_c^2}\frac{1+x}{(1-x)}\exp\left( -\frac{m_c^2}{(1-x)T^2}\right) \right. \nonumber\\
&&\left. +\frac{2-x}{xQ^2+m_c^2}\exp\left( -\frac{m_c^2}{T^2}\right) \right\}\, ,
\end{eqnarray}
where the $s_0$ is the continuum threshold parameter for the $Z_c$, and the $C_{Z_c\eta_c\pi}$ and $C_{Z_cDD}$ are unknown parameters introduced to take into account
the single-pole contributions associated with pole-continuum
transitions. In numerical analysis, we will denote the right sides of Eqs.(30-31) as $F_1(Q^2)$ and $F_2(Q^2)$ respectively. In the three-point QCD sum rules, the single-pole contributions are not suppressed if a single
Borel transform is taken.
\section{Numerical results and discussions}
The vacuum condensates are taken to be the standard values
$\langle\bar{q}q \rangle=-(0.24\pm 0.01\, \rm{GeV})^3$,
$\langle\bar{q}g_s\sigma G q \rangle=m_0^2\langle \bar{q}q \rangle$,
$m_0^2=(0.8 \pm 0.1)\,\rm{GeV}^2$, $\langle \frac{\alpha_s
GG}{\pi}\rangle=(0.33\,\rm{GeV})^4 $ at the energy scale $\mu=1\, \rm{GeV}$
\cite{SVZ79-1,SVZ79-2,Reinders85,Ioffe2005-1,Ioffe2005-2}.
The quark condensate and mixed quark condensate evolve with the renormalization group equation,
$\langle\bar{q}q \rangle(\mu)=\langle\bar{q}q \rangle(Q)\left[\frac{\alpha_{s}(Q)}{\alpha_{s}(\mu)}\right]^{\frac{4}{9}}$ and
$\langle\bar{q}g_s \sigma Gq \rangle(\mu)=\langle\bar{q}g_s \sigma Gq \rangle(Q)\left[\frac{\alpha_{s}(Q)}{\alpha_{s}(\mu)}\right]^{\frac{2}{27}}$.
The hadronic input parameters are taken as $M_{\pi}=0.13957\,\rm{GeV}$, $f_{\pi}=0.130\,\rm{GeV}$,
$M_{D^\pm}=1.8695\,\rm{GeV}$, $M_{D^0}=1.86491\,\rm{GeV}$, $f_{D}=0.208\,\rm{GeV}$, $M_{\eta_c}=2.9837\,\rm{GeV}$,
$f_{\eta_c}=0.350 \,\rm{GeV}$ \cite{PDG,WangJHEP,VANovikov}.
We take the values $m_u({\mu=\rm 1GeV})=m_d({\mu=\rm 1GeV})=m_q({\mu=\rm 1GeV})=0.006\,\rm{GeV}$ from the Gell-Mann-Oakes-Renner relation, and choose the $\overline{MS}$ mass $m_{c}(m_c)=(1.275\pm0.025)\,\rm{GeV}$
from the Particle Data Group \cite{PDG}, and take into account
the energy-scale dependence of the $\overline{MS}$ masses from the renormalization group equation,
\begin{eqnarray}
m_q(\mu)&=&m_q({\rm1GeV})\left[\frac{\alpha_{s}(\mu)}{\alpha_{s}({\rm1GeV})}\right]^{\frac{4}{9}} \, ,\nonumber\\
m_c(\mu)&=&m_c(m_c)\left[\frac{\alpha_{s}(\mu)}{\alpha_{s}(m_c)}\right]^{\frac{12}{25}} \, ,\nonumber\\
\alpha_s(\mu)&=&\frac{1}{b_0t}\left[1-\frac{b_1}{b_0^2}\frac{\log t}{t} +\frac{b_1^2(\log^2{t}-\log{t}-1)+b_0b_2}{b_0^4t^2}\right]\, ,
\end{eqnarray}
where $t=\log \frac{\mu^2}{\Lambda^2}$, $b_0=\frac{33-2n_f}{12\pi}$, $b_1=\frac{153-19n_f}{24\pi^2}$, $b_2=\frac{2857-\frac{5033}{9}n_f+\frac{325}{27}n_f^2}{128\pi^3}$, $\Lambda=213\,\rm{MeV}$, $296\,\rm{MeV}$ and $339\,\rm{MeV}$ for the flavors $n_f=5$, $4$ and $3$, respectively \cite{PDG}.
Now we study the mass and pole residue of the $S\bar{S}$ type scalar tetraquark state.
We impose
the two criteria (pole dominance and convergence of the operator product
expansion) on the hidden charmed tetraquark state to choose the Borel
parameter $T^2$ and threshold parameter $s_0$.
In the heavy quark limit, the $c$ (and $b$) quark can be taken as a static well potential,
which binds the light quark $q^{\prime}$ to form a diquark in the color antitriplet channel or binds the light antiquark $\bar{q}$ to form a meson in the color singlet channel (or a meson-like state in the color octet channel). Then the heavy tetraquark states are characterized by the effective heavy quark masses ${\mathbb{M}}_Q$ (or constituent quark masses) and the virtuality $V=\sqrt{M^2_{X/Y/Z}-(2{\mathbb{M}}_Q)^2}$ (or bound energy not as robust). It is natural to take the energy scale $\mu=V$,
the formula works well for the $X(3872)$, $Z_c(3885)$,
$Z_c(3900)$, $Z_c(4020)$, $Z_c(4025)$, $Z(4050)$, $Z(4250)$, $Y(4360)$, $Z(4430)$, $Y(4630)$, $Y(4660)$, $Z_b(10610)$ and $Z_b(10650)$ in the scenario of tetraquark states \cite{WangHuangTao-1,WangHuangTao-2,WangHuangTao-3,Wang-Cu-Cu}.
The relation
\begin{eqnarray}
M^2_{X/Y/Z}=(2{\mathbb{M}}_c)^2+\mu^2 \, ,
\end{eqnarray}
with the value ${\mathbb{M}}_c=1.8\,\rm{GeV}$ determined in previous works \cite{WangHuangTao-1,WangHuangTao-2,WangHuangTao-3,Wang-Cu-Cu} puts a strong constraint on the masses of the possible tetraquark states.
The mass gaps between the ground states and the first radial excited states are usually taken as $(0.4-0.6)\,\rm{GeV}$, for example,
the $Z(4430)$ is tentatively assigned as the first radial excitation of the $Z_c(3900)$ according to the
analogous decays,
\begin{eqnarray}
Z_c(3900)^\pm&\to&J/\psi\pi^\pm\, , \nonumber \\
Z(4430)^\pm&\to&\psi^\prime\pi^\pm\, ,
\end{eqnarray}
and the mass differences $M_{Z(4430)}-M_{Z_c(3900)}=576\,\rm{MeV}$, $M_{\psi^\prime}-M_{J/\psi}=589\,\rm{MeV}$ \cite{Maiani-2014,Nielsen-1401,Wang4430}.
The relation
\begin{eqnarray}
\sqrt{s_0}&=&M_{X/Y/Z}+(0.4-0.6)\,\rm{GeV} \, ,
\end{eqnarray}
puts another strong constraint on the masses of the possible tetraquark states.
In calculations, we observe that
\begin{eqnarray}
\mu\uparrow \, \, \, \, \, M_Z \downarrow \, ,\nonumber\\
\mu\downarrow \, \, \, \, \, M_Z \uparrow \, ,
\end{eqnarray}
from the QCD sum rule in Eq.(19). While Eq.(33) indicates that
\begin{eqnarray}
\mu\uparrow \, \, \, \, \, M_Z \uparrow \, ,\nonumber\\
\mu\downarrow \, \, \, \, \, M_Z \downarrow \, .
\end{eqnarray}
There must be a compromise, which leads to the optimal energy scale $\mu$, mass $M_Z$ and threshold parameter $s_0$.
In Fig.3, the contribution of the pole term is plotted with
variations of the threshold parameter $s_0$ and Borel parameter $T^2$ at the energy scale $\mu=1.3\,\rm{GeV}$.
From the figure, we can see that the value $\sqrt{s_0}\leq 4.1 \, \rm{GeV}$ is too small to satisfy the pole dominance condition and result in reasonable Borel window.
\begin{figure}
\centering
\includegraphics[totalheight=8cm,width=10cm]{pole.EPS}
\caption{ The pole contribution with variations of the Borel parameter $T^2$ and threshold parameter $s_0$, where the $A$, $B$, $C$, $D$, $E$, $F$ denote the threshold parameters $\sqrt{s_0}=4.0$, $4.1$, $4.2$, $4.3$, $4.4$, $4.5\,\rm{GeV}$, respectively. }
\end{figure}
In Fig.4, the contributions of different terms in the
operator product expansion are plotted with variations of the Borel parameter $T^2$ for the threshold parameter $\sqrt{s_0}= 4.3 \, \rm{GeV}$ at the energy scale $\mu=1.3\,\rm{GeV}$.
From the figure, we can see that the $D_0$, $D_3$, $D_5$, $D_6$ and $D_8$, where the $D_i$ denote the contributions of the vacuum condensates of dimensions $D=i$,
play an important role, while the $D_4$, $D_7$ and $D_{10}$ play a minor important role. At the value $T^2\leq 2.0\,\rm{GeV}$, the $D_3$, $D_5$, $D_6$ and $D_8$ decrease monotonously and quickly with increase of the $T^2$, which cannot lead to stable QCD sum rules. At the value $T^2=(2.2-2.6)\,\rm{GeV}^2$, $D_3\gg |D_5|\gg D_6\gg |D_8|$ and $D_{10}\ll 1\%$, the operator product expansion is well convergent, although $D_0\approx20\%$.
We approximate the continuum spectral density by
$\rho_{QCD}(s)\Theta(s-s_0)$; the contributions of the quark condensate $\langle\bar{q}q\rangle$ and mixed condensate $\langle\bar{q}g_s\sigma Gq\rangle$ can be very large.
\begin{figure}
\centering
\includegraphics[totalheight=8cm,width=10cm]{fractions.EPS}
\caption{ The contributions of different terms in the operator product expansion with variations of the Borel parameter $T^2$, where the 0, 3, 4, 5, 6, 7, 8, 10 denote the dimensions of the vacuum condensates. }
\end{figure}
In this article, we take the Borel parameter
$T^2=(2.2-2.6)\,\rm{GeV}^2$,
the continuum threshold parameter
$\sqrt{s_0}=(4.2-4.4)\,\rm{GeV}$ and the energy scale $\mu=1.3\,\rm{GeV}$, the pole dominance is well satisfied.
The Borel parameter, continuum threshold parameter and the pole contribution are shown explicitly in Table 1. The two criteria (pole dominance and convergence of the operator product expansion) of the QCD sum rules are fully satisfied, furthermore, the relations in Eq.(33) and Eq.(35) are also satisfied.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|}\hline\hline
$J^{PC}$ & $T^2 (\rm{GeV}^2)$ & $\sqrt{s_0} (\rm{GeV})$ & pole & $M_{Z}(\rm{GeV})$ & $\lambda_{Z}$ \\ \hline
$0^{++}$ & $2.2-2.6$ & $4.3\pm0.1$ & $(49-74)\%$ & $3.82^{+0.08}_{-0.08}$ & $1.79^{+0.29}_{-0.24}\times10^{-2}\rm{GeV}^5$ \\ \hline
\hline
\end{tabular}
\end{center}
\caption{ The Borel parameter, continuum threshold parameter, pole contribution, mass and pole residue of the scalar tetraquark state. }
\end{table}
\begin{figure}
\centering
\includegraphics[totalheight=8cm,width=10cm]{mass.EPS}
\caption{ The mass with variations of the Borel parameter $T^2$. }
\end{figure}
Taking into account all uncertainties of the input parameters,
finally we obtain the values of the mass and pole residue of
the $S\bar{S}$ type scalar tetraquark state, which are shown explicitly in Figs.5-6 and Table 1.
The central value of the present prediction $M_{Z_c}=\left(3.82^{+0.08}_{-0.08}\right)\,\rm{GeV}$ for the $S\bar{S}$ type scalar tetraquark state is smaller than
that of the $A\bar{A}$ type scalar tetraquark state $M_{J=0}=\left(3.85^{+0.15}_{-0.09}\right)\,\rm{GeV}$ obtained in Ref.\cite{Wang-Cu-Cu}.
The predictions based on the QCD sum rules are consistent with the values $M_{J=0}=3.852\,\rm{GeV}$ and $3.812\,\rm{GeV}$ for the $A\bar{A}$ and $S\bar{S}$ type scalar tetraquark states $cq\bar{c}\bar{q}$ respectively from the quasipotential approach \cite{EFG-2008}.
Now we take the mass $M_{Z_c}$ and pole residue $\lambda_{Z_c}$ as basic input parameters to study the hadronic coupling constants $G_{Z_c\eta_c\pi}$ and $G_{Z_cDD}$, and take the same threshold parameter and Borel parameter as in the QCD sum rule for the mass and pole residue. In calculations, we choose
the unknown parameters as $C_{Z_c\eta_c\pi}=0.0009\,\rm{GeV}^6 $ and $C_{Z_cDD}=0.0004\,\rm{GeV}^6 $ to obtain stable QCD sum rules with variations of the Borel parameter $T^2$ at the Borel windows $T^2=(2.2-2.6)\,\rm{GeV}^2$; the left side and right side of the QCD sum rules coincide. In fact, it is not necessary to choose the same Borel parameters both in the two-point and three-point QCD sum rules. If we take larger Borel parameter, say $T^2=(2.5-3.0)\,\rm{GeV}^2$ instead of $T^2=(2.2-2.6)\,\rm{GeV}^2$, we should alter the unknown parameters $C_{Z_c\eta_c\pi}$ and $C_{Z_cDD} $ slightly, then obtain stable QCD sum rules, the resulting values of the hadronic coupling constants change slightly.
\begin{figure}
\centering
\includegraphics[totalheight=8cm,width=10cm]{residue.EPS}
\caption{ The pole residue with variations of the Borel parameter $T^2$. }
\end{figure}
\begin{figure}
\centering
\includegraphics[totalheight=8cm,width=10cm]{F2Q2.EPS}
\caption{ The central values of the $F_2(Q^2)$ with variations of the $Q^2$. }
\end{figure}
\begin{figure}
\centering
\includegraphics[totalheight=8cm,width=10cm]{GQ2.EPS}
\caption{ The central values of the hadronic coupling constants with variations of the $Q^2$, where the $A$ and $B$ denote the $G_{Z_cDD}(Q^2)$ and $G_{Z_c\eta_c\pi}(Q^2)$, respectively. }
\end{figure}
Based on Eqs.(30-31), we can study the $Q^2$ dependence of the right side of the QCD sum rules,
\begin{eqnarray}
F_1(Q^2)&\propto&\frac{Q^2+M_{\pi}^2}{Q^2}\approx 1 \, ,
\end{eqnarray}
at the region of large (or intermediate) $Q^2$ due to the tiny mass of the $\pi$, while the $F_2(Q^2)$ has no such simple $Q^2$ dependence due to the heavy quark mass $m_c$ and heavy meson mass $M_D$. In the limit $Q^2\to \infty$,
\begin{eqnarray}
F_2(Q^2)&=&-\frac{m_c\langle\bar{q}g_s\sigma Gq\rangle}{64\pi^2}\int_0^1 dx \left\{\frac{1+x}{1-x}\exp\left( -\frac{m_c^2}{(1-x)T^2}\right) +\frac{2-x}{x}\exp\left( -\frac{m_c^2}{T^2}\right) \right\}\, ,
\end{eqnarray}
which is independence on $Q^2$.
In Fig.7, we plot the central values of the $F_2(Q^2)$ with variations of the $Q^2$ at the range $Q^2=(1-5)\,\rm{GeV}^2$ for the Borel parameters $T^2=2.2\,\rm{GeV}^2$, $2.4\,\rm{GeV}^2$ and $2.6\,\rm{GeV}^2$, respectively.
From the figure, we can see that the $Q^2$ dependence of the $F_2(Q^2)$ is rather mild and can be neglected approximately.
The left sides of the QCD sum rules in Eqs.(30-31) have no explicit $Q^2$ dependence, the $Q^2$ dependence is embodied in the right sides of the QCD sum rules ($F_1(Q^2)$ and $F_2(Q^2)$), so the hadronic coupling constants
$G_{Z_c\eta_c\pi}$ and $G_{Z_cDD}$ are independent on the $Q^2$ in the limit $Q^2\rightarrow \infty$, the conclusion survives even for much smaller $Q^2$, say $Q^2=(1-5)\,\rm{GeV}^2$ according to Eq.(38) and Fig.7.
The central values of the $G_{Z_c\eta_c\pi}(Q^2)$ and $G_{Z_cDD}(Q^2)$ can be fitted to the following constant forms,
\begin{eqnarray}
G_{Z_c\eta_c\pi}(Q^2)&=&0.43\,\rm{GeV}^{-1}\, , \nonumber \\
G_{Z_cDD}(Q^2)&=& 1.06\,\rm{GeV}^{-1} \, ,
\end{eqnarray}
at the region $Q^2=(1-5)\,\rm{GeV}^2$; the uncertainties of the $G_{Z_c\eta_c\pi}$ and $ G_{Z_cDD}$ are about $25\%$ and $18\%$, respectively.
We plot the central values of the hadronic coupling constants $G_{Z_cDD}(Q^2)$ and $G_{Z_c\eta_c\pi}(Q^2)$ with variations of the $Q^2$ at the region $Q^2=(1-5)\,\rm{GeV}^2$ for the Borel parameter $T^2=2.4\,\rm{GeV}^2$ in Fig.8. From the figure, we can see that the fitted functions in Eq.(40) are satisfactory.
We extend the coupling constants to the physical regions without difficulty, and calculate the partial decay widths,
\begin{eqnarray}
\Gamma_{Z_c \to \eta_c\pi}&=&\frac{G_{Z_c\eta_c\pi}^2(M_{Z_c}^2-M_{\eta_c}^2-M_{\pi}^2)^2\,p_{\eta_c\pi}}{32\pi M_{Z_c}^2}=(3.0\pm1.5)\,\rm{MeV}\, , \nonumber\\
\Gamma_{Z_c\to DD}&=&\frac{G_{Z_cDD}^2(M_{Z_c}^2-M_{D^+}^2-M_{D^0}^2)^2\,p_{DD}}{32\pi M_{Z_c}^2}=(17.9\pm6.4)\,\rm{MeV}\, ,
\end{eqnarray}
where
\begin{eqnarray}
p_{\eta_c\pi}&=&\frac{\sqrt{\left[M_{Z_c}^2-(M_{\eta_c}+M_{\pi})^2\right]\left[M_{Z_c}^2-(M_{\eta_c}-M_{\pi})^2\right]}}{2M_{Z_c}} \, ,\nonumber\\
p_{DD}&=&\frac{\sqrt{\left[M_{Z_c}^2-(M_{D^+}+M_{D^0})^2\right]\left[M_{Z_c}^2-(M_{D^+}-M_{D^0})^2\right]}}{2M_{Z_c}} \, .
\end{eqnarray}
The total width $\Gamma_{Z_c}$ of the $Z_c(3820)$ can be approximated by $\Gamma_{Z_c \to \eta_c\pi}+\Gamma_{Z_c\to DD}$, the numerical value is about $(20.9\pm 6.6)\,\rm{MeV}$. The radiative decay widths can be estimated by assuming vector meson dominance, for example,
$\Gamma_{Z_c^\pm \to \gamma\rho^\pm} \propto\alpha |\Gamma_{Z^\pm_c \to J/\psi^* \rho^\pm}|$ for the radiative decays $Z^\pm_c(3820) \to J/\psi^* \rho^\pm \to \gamma\rho^\pm$, the partial decay widths are of the order ${\cal{O}}(\rm {KeV})$ due to the factor $\alpha=\frac{e^2}{4\pi}=\frac{1}{137}$. The strong decays
$Z^\pm_c(3820) \to J/\psi\rho^\pm$ are kinematically forbidden, the values of the $\Gamma_{Z^\pm_c \to J/\psi^* \rho^\pm}$ are complex, so we take $|\Gamma_{Z^\pm_c \to J/\psi^* \rho^\pm}|$. The contributions of the radiative decays to the total width $\Gamma_{Z_c}$ are small and can be neglected.
\section{Conclusion}
In this article, we calculate the contributions of the vacuum condensates up to dimension-10 in the operator product expansion, study the $S\bar{S}$ type scalar tetraquark state $cq\bar{c}\bar{q}$ in details with the QCD sum rules. In calculations, we search for the optimal Borel parameter and threshold parameter to satisfy the energy scale formula $M^2_{Z}=(2{\mathbb{M}}_c)^2+\mu^2$ and the experiential threshold formula $\sqrt{s_0}=M_{Z}+(0.4-0.6)\,\rm{GeV}$, where the $\mu$ is the energy scale of the QCD spectral density, and obtain the values $M_{Z_c}=\left(3.82^{+0.08}_{-0.08}\right)\,\rm{GeV}$ and $\lambda_{Z_c}=\left(1.79^{+0.29}_{-0.24}\right)\times10^{-2}\rm{GeV}^5$. The central value of the mass of the $S\bar{S}$ type scalar tetraquark state is smaller than that of the $A\bar{A}$ type scalar tetraquark state, the $S\bar{S}$ type scalar tetraquark state $cq\bar{c}\bar{q}$ maybe the lowest hidden charmed tetraquark state. Furthermore, we calculate the hadronic coupling constants $G_{Z_c\eta_c\pi}$ and $G_{Z_cDD}$ with the three-point QCD sum rules by taking into account the color-connected diagrams, then study the strong decays $ Z_c\to \eta_c\pi\, ,\, DD$, and observe that the total width $\Gamma_{Z_c}\approx 21\,\rm{MeV}$. The present
predictions can be confronted with the experimental data in the futures at the BESIII, LHCb and Belle-II.
\section*{Acknowledgements}
This work is supported by National Natural Science Foundation,
Grant Numbers 11375063, and Natural Science Foundation of Hebei province, Grant Number A2014502017.
|
\section{Introduction}\label{INTRODUCTION}
\IEEEPARstart{V}isible light communication (VLC) refers to unguided optical transmission via the use of light emitting diodes (LEDs) \cite{komine2004fundamental,randel2010advanced,street1997indoor}. Indoor VLC is characterized by short transmission range and free from major outdoor environmental degradations such as rain, snow, building sway, and atmospheric turbulence. This technology has recently attracted significant attentions as a promising complementary technology for radio frequency (RF) in short-range communications \cite{randel2010advanced,komine2003integrated,rajagopal2012ieee,kavehrad2007broadband}. These systems offer significant technical and operational advantages such as higher bandwidth capacity, virtually unlimited reuse, unregulated spectrum and robustness to electromagnetic interference. Despite the major advantages of indoor VLC, they suffer from multipath distortion due to dispersion of the optical signal caused by reflections from various sources inside a room. This dispersion leads to inter-symbol interference (ISI) at high data rates which reduces signal to noise ratio (SNR) and severely impairs the link performance.
Multicarrier modulation which is usually implemented by orthogonal frequency division multiplexing (OFDM) has been originally introduced in RF communication to mitigate ISI and multipath dispersion. The concept of OFDM has been applied to indoor wireless optical communications (WOC) in \cite{shieh2008coherent,gonzalez2005ofdm,elgala2009indoor,armstrong2006power,armstrong2009ofdm,carruthers1996multiple} to support high data rates. However, the performance of VLC systems using intensity-modulation direct-detection (IM/DD) along with OFDM modulation is significantly affected by nonlinear characteristic of LED due to the large peak-to-average power ratio (PAPR) of OFDM signal. In particular, signal amplitudes below the LED turn-on-voltage (TOV) and above the LED saturation point are clipped. To address this issue, single-carrier frequency domain equalization (SCFDE) has been proposed in the literature reducing PAPR while achieving similar throughput as OFDM systems \cite{ciochina2010review,falconer2002frequency}. Several OFDM schemes such as DC-clipped OFDM \cite{kahn1997wireless}, asymmetrically clipped optical OFDM (ACO-OFDM) \cite{armstrong2006power} and PAM-modulated discrete multitone (PAM-DMT) \cite{lee2009pam} have been proposed in the literature. Among these schemes, ACO-OFDM has been shown to be more efficient in terms of optical power than the systems that use DC-biasing as it utilizes a large dynamic range of the LED. Therefore, it is considered in this paper.
There exist several investigations analyzing different OFDM techniques and comparing them with SCFDE \cite{mesleh2012ofdm,mesleh2011performance,dissanayake2013comparison} or single carrier modulation \cite{barros2012comparison,armstrong2008comparison}. To the best of our knowledge, these previous studies were built on the assumption of ideal additive white Gaussian noise (AWGN) channels or did not consider the nonlinear characteristics of LED. In this paper, we analyze and compare performance of the aforementioned techniques along with on-off keying (OOK) with minimum mean square error equalization (MMSE) which is commonly used in IM/DD communication systems considering an off-the-shelf LED model and a multipath channel. Moreover, bit-interleaved coded modulation (BICM) is considered for OFDM and SCFDE systems to further combat signal degradation due to LED nonlinearity and ISI.
\begin{figure*}[t]
\centering
\includegraphics[width = 13cm, height = 6cm]{ofdm.eps}
\caption{ACO-OFDM Transmitter and Receiver configuration.}
\label{aco}
\end{figure*}
The rest of the paper is organized, as follows. In Sections II and III, we briefly introduce the system model and describe ACO-OFDM and ACO-SCFDE. In Section IV, we compare the PAPR performance of ACO-OFDM and ACO-SCFDE. In Section V, we analyze and compare ACO-SCFDE, ACO-OFDM and OOK performance and investigate the impact of LED bias point on the performance. Furthermore, we show that BICM can combat signal degradation due to LED nonlinearity and ISI. Finally, Section VI concludes the paper.
\section{System Model of Asymmetrically-Clipped Optical OFDM (ACO-OFDM)}
ACO-OFDM is a form of OFDM that modulates the intensity of a LED. Because ACO-OFDM modulation employs IM/DD, the time-domain transmitted signal must be real and positive. The block diagram of an ACO-OFDM system is depicted in Fig. \ref{aco}. The information stream is first parsed into a block of $N$ complex data symbols denoted by $\mathbf{I}={{[{{I}_{0}},{{I}_{1}},\ldots ,{{I}_{N-1}}]}^{T}}$ where the symbols are drawn from constellations such as $M$-QAM or $M$-PSK where $M$ is the constellation size. To ensure a real output signal used to modulate the LED intensity, ACO-OFDM subcarriers must have Hermitian symmetry. In ACO-OFDM, only odd subcarriers are modulated, and this results in avoiding the impairment from clipping noise. Therefore, the complex symbols are mapped onto a $4N\times 1$ vector as $\mathbf{S}={{[0,{{I}_{0}},0,{{I}_{1}},\ldots ,0,{{I}_{N-1}},0,I_{N-1}^{*},0,\ldots ,I_{1}^{*},0,I_{0}^{*},0]}^{T}}$ where ${\left( . \right)}^{*}$ denotes the complex conjugate of a vector. A $4N$-point IFFT is then applied on the vector $\mathbf{S}$ to build the time domain signal $\mathbf{x}$. A cyclic prefix (CP) is added to $\mathbf{x}$ turning the linear convolution with the channel into a circular one to mitigate multipath dispersion. To make the transmitted signal unipolar, all the negative values are clipped to zero. It is proven in \cite{armstrong2006power} that since only the odd subcarriers are used to carry the data symbols, the clipping does not affect the data-carrying subcarriers, but only reduces their amplitude by a factor of two.
The unipolar signal is then converted to analog and filtered to modulate the intensity of an LED. At the receiver, the signal is converted back to digital. CP is then removed and the electrical OFDM signal is demodulated by taking a $4N$ FFT and equalized with a single-tap equalizer on each subcarrier to compensate for
channel distortion. The even subcarriers are then discarded and the transmitted data is recovered by a hard or soft decision. The extraction of odd subcarriers along with the equalization are represented by the \emph{Demapping} block in Fig. \ref{aco}.
\section{System Model of Asymmetrically-Clipped Optical SCFDE (ACO-SCFDE)}
SCFDE is a special technique which is compatible with any of OFDM techniques. In this paper, we apply asymmetrically-clipped optical to SCFDE to achieve ACO-SCFDE with low PAPR. The block diagram of an ACO-SCFDE system is depicted in Fig. \ref{sc}. ACO-SCFDE and ACO-OFDM are the same except that in ACO-SCFDE, an extra $N$-point FFT and IFFT are used at the transmitter and the receiver respectively resulting in a single carrier transmission instead of multicarrier. As it will be shown latter, the additional complexity of the extra FFT and IFFT blocks is offset by the fact that SCFDE has lower PAPR and better bit-error-rate (BER) performance than its OFDM counterpart when the signal is sent through the non-linear LED.
\begin{figure*}[t]
\centering
\includegraphics[width = 13cm, height = 6cm]{scfde2.eps}
\caption{ACO-SCFDE Transmitter and Receiver configuration.}
\label{sc}
\end{figure*}
\section{Peak-to- Average Power Ratio (PAPR)}
\begin{figure}
\centering
\includegraphics[width = 8cm, height = 7.5cm]{PAPR64.eps}
\caption{CCDF of PAPR Comparison of ACO-OFDM and ACO-SCFDE for $N=64$.}
\label{p64}
\end{figure}
\begin{figure}
\centering
\includegraphics[width = 8cm, height = 7.5cm]{PAPR256.eps}
\caption{CCDF of PAPR Comparison of ACO-OFDM and ACO-SCFDE for $N=256$.}
\label{p256}
\end{figure}
In this section, the PAPR of ACO-OFDM signals is analyzed and compared with that of ACO-SCFDE. The PAPR is defined as the maximum power of transmitted signal divided by the average power, that is
\begin{equation}
PAPR=\frac{\max {{x}^{2}}\left( n \right)}{E\left[ {{x}^{2}}\left( n \right) \right]}
\end{equation}
where $E\left[.\right]$ denotes expectation. Due to the large number of subcarriers and occasional constructive combining of them, OFDM systems have a large dynamic signal range and exhibit a very high PAPR.
Thus, the OFDM signal will be clipped when passed through a nonlinear LED at the transmitter end which results in degrading the BER performance. SCFDE can be used as a promising alternative technology for OFDM to reduce the PAPR and combat the effect of nonlinear characteristics of the LED.
PAPR is usually presented in terms of a complementary cumulative distribution function (CCDF) which is the probability that PAPR is higher than a certain PAPR value $PAP{{R}_{0}}$, i.e. $\Pr \left( PAPR>PAP{{R}_{0}} \right)$. Figs \ref{p64}-\ref{p256} demonstrate the CCDF of PAPR for $N=64$ and 256 subcarriers respectively, calculated by Monte Carlo simulation for different modulation constellations. We notice that ACO-SCFDE has a lower PAPR as compared to ACO-OFDM system for the same number of subcarriers. We also observe that the PAPR increases with increasing $N$ for all of the constellations.
\section{Performance Analysis}
\begin{table}
\begin{center}
\begin{quote}
\caption{Room configuration under consideration.}
\label{table}
\end{quote}
{\small{
\begin{tabular}{|c|c|c|}
\hline
Room &Length &6 m \\ \cline{2-3}
&Width &5 m \\ \cline{2-3}
&Height &3 m\\ \hline
Reflectivity &${\rho }_{\text{North}}$ &0.8\\ \cline{2-3}
&${\rho }_{\text{South}}$ &0.8\\ \cline{2-3}
&${\rho }_{\text{East}}$ &0.8\\ \cline{2-3}
&${\rho }_{\text{West}}$ &0.8\\ \cline{2-3}
&${\rho }_{\text{Ceiling}}$ &0.8\\ \cline{2-3}
&${\rho }_{\text{Floor}}$ &0.3\\ \hline
Source &Mode &1 \\ \cline{2-3}
&Azimuth &${{0}^{\circ }}$ \\ \cline{2-3}
&Elevation &${{-90}^{\circ }}$ \\ \cline{2-3}
&$x$, $y$, $z$ &0.1 m, 0.2 m, 3 m\\ \hline
Receiver &Area &$\text{CM}^{2}$\\ \cline{2-3}
&FOV &${{0}^{\circ }}$ \\ \cline{2-3}
&$x$, $y$, $z$ &2.5 m, 2.5 m, 1 m\\ \hline
\end{tabular}}}
\end{center}
\end{table}
Simulations are conducted assuming indoor optical multipath channel where the transmitter and receiver are placed in a room whose configuration is summarized in Table \ref{table}. The methodology developed by Barry et al \cite{barry1993simulation} is employed to simulate the impulse response of the channel where 10 reflections are taken into account. Fig. \ref{channel} presents the impulse respond of a diffuse channel.
\begin{figure}
\centering
\includegraphics[width = 7.5cm, height = 6cm]{Channel.eps}
\caption{Impulse response of the indoor diffuse channel.}
\label{channel}
\end{figure}
We assume an OFDM signal whose average electrical power before modulating the LED is varied from -10 dBm to 30 dBm, and the power of AWGN is -10 dBm. Thus, the simulated electrical SNR ranges from 0 dB to 40 dB matching the reported SNR values for indoor WOC systems \cite{o2007optical,grubor2008bandwidth}. A number of subcarriers of $N=64$ with $M$-QAM modulation are also assumed. Furthermore, OPTEK, OVSPxBCR4 1-Watt white LED is considered in simulations whose optical and electrical characteristics are given in Table \ref{table2}. A polynomial order of five is used to realistically model measured transfer function. Fig. \ref{led} demonstrates the non-linear Transfer characteristics of the LED from the data sheet and using the polynomial function.
\begin{table}
\begin{center}
\begin{quote}
\caption{Optical and Electrical Characteristics of OPTEK, OVSPxBCR4 1-Watt white LED.}
\label{table2}
\end{quote}
{\small{
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\textbf{Symbol} &\textbf{Parameter} &\textbf{MIN} &\textbf{TYP} &\textbf{MAX} &\textbf{Units}\\ \hline
$V_F$ &Forward Voltage &3.0 &3.5 &4 &$V$\\ \hline
$\Phi$ &Luminous Flux &67 &90 &113 &lm\\ \hline
${{\Theta }^{1/2}}$ &50\% Power Angle &--- &120 &--- °\\ \hline
\end{tabular}}}
\end{center}
\end{table}
\begin{figure} \centering
\subfigure[]{\includegraphics[width = 7cm, height = 5cm]{LED.eps}}
\subfigure[]{\includegraphics[width = 7cm, height = 5cm]{leddata.eps}}
\caption{Transfer characteristics of OPTEK, OVSPxBCR4 1-Watt white LED. (a) Fifth-order polynomial fit to the data. (b) The curve from the data sheet.}
\label{led}
\end{figure}
We first compare the bit error performance of ACO-SCFDE and ACO-OFDM. Fig. \ref{aco1} presents the BER performance of ACO-OFDM and ACO-SCFDE for different modulation orders and LED bias point of 3.2V. As the results indicate, SCFDE exhibits better BER performance in the optical multipath channel.
\begin{figure}
\centering
\includegraphics[width = 8cm, height = 7.5cm]{ACO1.eps}
\caption{BER Comparison of ACO-OFDM and ACO-SCFDE for bias point of 3.2V.}
\label{aco1}
\end{figure}
Furthermore, we investigate the impact of LED bias point on the performance of ACO-OFDM systems. According to the data sheet of the LED used in the simulations, three different bias points (3V, 3.2V and 3.5V) are considered. Fig. \ref{bias} demonstrates BER performance of an ACO-OFDM system with $M=16$ and different LED bias points. As it can be clearly seen, nonlinearity of LED has a significant impact on the performance of optical OFDM systems. It is also observed that there is an optimum LED bias point which is 3.2V for the case under consideration from which deviation can significantly deteriorate the system performance.
\begin{figure}
\centering
\includegraphics[width = 8cm, height = 7.5cm]{bias.eps}
\caption{BER of ACO-OFDM for $M = 16$ for different bias points.}
\label{bias}
\end{figure}
Bit-interleaved coded modulation (BICM) \cite{caire1998bit} is also considered for OFDM and SCFDE systems to further compensate signal degradation due to ISI and LED nonlinearity. To demonstrate the usefulness of BICM, we assume that the information sequence is first encoded by a rate 1/2 convolutional encoder with generator matrix $\mathbf{g} = (5,7)$, constraint length of 3 and minimum Hamming distance of 5. The coded information is then interleaved by a bitwise interleaver. At the receiver, the Viterbi soft-decoder \cite{viterbi1971convolutional} and the de-interleaver are used. Fig. \ref{code} shows the BER of uncoded and coded ACO-OFDM and ACO-SCFDE for the indoor visible light communication under consideration. As it can be clearly observed, BICM can significantly enhance the system performance. However, the achieved gains come at the cost of significant reduction in the data rate due to the insertion of coded bits.
\begin{figure}
\centering
\includegraphics[width = 8cm, height = 7.5cm]{Coding.eps}
\caption{BER Comparison of uncoded and coded ACO-OFDM and ACO-SCFDE for $M = 16$.}
\label{code}
\end{figure}
Finally, we compare the performance of ACO-OFDM, ACO-SCFDE and OOK modulation with MMSE over an indoor VLC medium. The performance comparison is done in terms of \emph{normalized SNR} and \emph{normalized bandwidth/bit-rate} relative to OOK \cite{kahn1997wireless}. According to \cite{kahn1997wireless} and \cite{armstrong2008comparison}, we define the modulation bandwidth as the position of the first spectral null. To make a fair comparison between different modulation schemes, the normalized bandwidth of the signal is calculated as the modulation bandwidth which is normalized relative to OOK of the same transmitted data rate. For ACO-OFDM and ACO-SCFDE, first null occurs at a normalized frequency of $1+2/N$. Thus, the normalized bandwidth/bit-rate is obtained as $2\left( 1+2/N \right)/\log _{2}^{M}$ for ACO-OFDM and ACO-SCFDE. Fig. \ref{comp} shows normalized SNR required for a BER of ${{10}^{-9}}$ as a function of normalized bandwidth/bit-rate for OOK, ACO-OFDM and ACO-SCFDE. We observe while ACO-OFDM and ACO-SCFDE with 4-QAM modulation of each subcarrier require approximately the same bandwidth as OOK, they are more efficient in terms of power. Particularly, ACO-OFDM and ACO-SCFDE are 2.7 dB and 3.7 dB more efficient than OOK, respectively. For the higher orders of $M$, OOK outperforms ACO-OFDM and ACO-SCFDE but it requires greater bandwidth.
\begin{figure}
\centering
\includegraphics[width = 8cm, height = 7.5cm]{Comp.eps}
\caption{Normalized SNR versus normalized bandwidth/bit-rate required to achieve BER of ${{10}^{-9}}$. }
\label{comp}
\end{figure}
\section{Conclusions}
We have evaluated and compared the performance of IM/DD single- and multi-carrier modulation schemes for indoor visible light communication systems taking into account both nonlinear characteristics of LED and dispersive nature of optical wireless channel. We have shown through the use of simulation that SCFDE system has a lower PAPR than its counterpart OFDM system and outperforms OOK and OFDM systems and therefore is a promising modulation technique for indoor VLC systems. We have also investigated the performance of OFDM systems for different LED bias points and shown that significant gain can be achieved by biasing LED with the optimum value. BICM technique has been further considered to combat signal degradation due to LED nonlinearity and dispersive nature of the channel.
\section{Acknowledgement}
The authors would like to thank the National Science Foundation (NSF) ECCS directorate for their support of this work under Award \# 1201636, as well as Award \# 1160924, on the NSF “Center on Optical Wireless Applications (COWA– \href{http://cowa.psu.edu}{http://cowa.psu.edu})”
\balance
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Sequential measurements play an important role in quantum measurement theory, e.g., \cite{DL}--\cite{SQ}. In particular, sequential measurements of two
incompatible observables $A$ and $B$ induce the probability order effect which gives the probabilistic representation of the
principle of complementarity, in the form of nonexistence of the joint probability distribution of such observables. In other words,
sequential probability distributions $p_{A-B}(x,y)$ and $p_{B-A}(y,x)$ do not coincide.
We remark that if $A$ and $B$ are quantum measurements of
the von Neumann-L\"uders type (measurements of the first kind), i.e., they are mathematically represented by Hermitian
operators and the state transformers (quantum operations)\footnote{The terminology ``state transformer'' \cite{BL} is may be old fashioned.
In modern literature \cite{Jaeger} the terms of quantum operation and quantum channel are typically in the use. The latter notions are motivated
by quantum information theory \cite{Jaeger}. However, the problems handled in this paper have no direct relation to quantum information theory. They are
motivated by applications of the quantum calculus of probabilities outside of physics, see the last part of this
section for details. In such applications
the terminology ``state transformer'' seems to be more appropriate.}
are given by the orthogonal projectors corresponding to the eigenvalues\footnote{In the mathematical model, see, e.g., \cite{DL}--\cite{SQ},
quantum measurement is represented with the aid of two structures: a quantum observable (Hermitian operator or more generally POVM) and a state transformer.
}, then they have the property of
{\it adjacent sequential reproducibility}: for $A-A$ and $B-B$ measurements the values observed in the first measurement are reproduced
(with probability 1) in the second measurement. Thus, for quantum measurements of the first kind,
the condition of adjacent reproducibility is a redundant constraint. However, for quantum measurements of the second kind,
i.e., those which cannot be represented as measurements of the first kind, this constraint is nontrivial.
We remark that, for quantum measurements of the second kind neither
an observable nor a state transformer have to be represented by orthogonal projectors. In general, for measurements of the
second kind, observables are represented by POVMs. We also remark that a quantum measurement with an observervable given by a Hermitian operator
can also be of the second kind: if the corresponding state transformers are not given by orthogonal projectors. The latter class of quantum
measurements will play an important role in problems under the study in this paper.
For measurements of the first kind, the order effect can be approached only
in case of incompatibility: commutativity of the observables $([A,B]=0)$ implies the existence
of the joint probability distribution serving both sequential measurements, $A-B$ and $B-A.$
Intuitively, in the case of incompatible observables a measurement of the observable $B$ after a preceding measurement of the observable $A$ (with the result $A=x)$ modifies crucially the post-measurement state $\psi_A$ generated as the result of the $A$ measurement. Thus, in the post-measurement state after the sequence
of measurements $A-B$, the information about the value $A=x$ is at least partially washed out and one cannot expect that in
the sequence of measurements $A-B-A$ the value $A=x$ would be reproduced (obtained in the second $A$ measurement) with probability 1. Thus, one would expect that
the quantum order effect cannot be combined with both the adjacent and separated reproducibility.
In paper \cite{BKDB} it was shown that,
for measurements of the first kind, this is really the case.
The problem of an extension of this result
to measurements of the second kind is mathematically more complicated. Some special class of measurements of the second kind was considered
in the aforementioned paper and it was shown that the combination of $A-A, B-B,$ and $A-B-A$ reproducibilities implies disappearance
of the order effect. The class of measurements of the second kind considered in \cite{BKDB} is characterized by the following conditions:
\medskip
${\bf C0}$ Observables are represented by POVMs.
\medskip
${\bf C1}$ The state transformer corresponding to an effect $E$ has the form
\begin{equation}
\label{hum}
\psi \to \frac{M\psi}{\Vert M \psi \Vert},
\end{equation}
where
\begin{equation}
\label{hh}
E=M^*M
\end{equation}
is some representation of the effect.
\medskip
${\bf C2}$ If $E=P$ is an orthogonal projector, then its state transformer is given by this projector
\begin{equation}
\label{hum1}
\psi \to \frac{P\psi}{\Vert P \psi \Vert}.
\end{equation}
The condition ${\bf C1},$ although it restricts the class of possible state transformers, is very natural,
because the majority of state transformers used in applications really have this simple form, (\ref{hum}).
By the condition ${\bf C2}$ the class of measurements corresponding to the effects given by projectors coincides
with the class of measurements of the von Neumann-L\"uders type (measurements of the first kind). Although this condition is also quite natural,
it restricts essentially the class of state transformers. In this paper we want to proceed with the class of quantum measurements
described solely by the conditions ${\bf C0}, {\bf C1}.$ Thus, even for an effect given by an orthogonal projector, in general
the corresponding state transformer does not coincide with this
projector.
Surprisingly, this extension of the class of quantum measurements makes the situation essentially more complicated (and interesting).
The order effect can coexist with separated reproducibility of $A-B-A$ type as well as
adjacent reproducibility for both observables $A$ and $B.$ However, the additional constraint in the form of separated reproducibility of the $B-A-B$
type makes this coexistence impossible.
Finally, we remark that this study (as well as the preceding study \cite{BKDB}) was motivated by applications of the quantum formalism outside of physics, namely,
in cognitive psychology and psychophysics, see, e.g., \cite{40}-\cite{C1} for introduction. In such applications the mathematical formalism of
quantum mechanics is treated as an operational formalism, see, e.g., M. D' Ariano \cite{DARIANO}, for handling probabilistic data collected in aforementioned domains of research. This formalism
works well for a wide class of psycho-effects, e.g., the disjunction effect and the order effect. In particular, to represent
these two effects one can use
the von Neumann-L\"uders measurements. As is well known, such observables satisfy the condition of
adjacent reproducibility, i.e., $A-A$ reproducibility, which is natural for the majority of experiments of cognitive psychology.
However, as was pointed out in \cite{BKDB}, in psychophysics one can find
the experimental situations violating adjacent reproducibility. Therefore one has to proceed with generalized observables given by
POVMs.\footnote{Observables of this class appeared in the applications of the quantum mathematical formalism to humanities even by
another reason \cite{UB}:
in experiments performed, e.g., in cognitive psychology,
the matrices of transition probabilities for observables with non-degenerate spectra are typically not double-stochastic. However, for
the von Neumann-L\"uders observables, they should be double-stochastic.} And this was the natural step in development of quantum-like modeling
in humanities.
As was first understood by the authors of \cite{BKDB}, there was a hidden pitfall in the rapidly increasing stream of applications of
the quantum formalism to humanities, namely, the problem of separated reproducibility. As was pointed out in \cite{BKDB}, some experimental
contexts of cognitive psychology are characterized by separated reproducibility effect in combination with the order effect. For example, political opinion
polls often demonstrate order effects, but here, e.g., by replying ``yes'' to the first question $A,$ and ``no'' to the second question $B$
a respondent is typically ``firm in her preferences'' expressed in the form of the $A$-yes, so she will practically definitely say ``yes''
again if asked the $A$-question again, i.e., in the $A-B-A$ experiment. We also remark that in known and thinkable experimental contexts
in cognitive psychology, the same should also happen for the $B-A-B$ experiment. As was shown in \cite{BKDB}, it is impossible
to use the measurements of the first kind to describe such a situation. The following natural question was posed by the authors of \cite{BKDB}:
{\it Is the operational quantum formalism powerful enough to cover all possible experimental contexts arising in humanities?}
In this paper, we continue to work to find an answer to this question. And for the moment, the answer is negative. However, although this paper
covers a wide and natural class of state transformers, it is still not the most general. There is still a ``loophole in the proof'' that
the power of quantum methods in humanities is restricted. We remark that in principle there is no reason to expect that
the quantum-like operational formalism
serving humanities would coincide with the quantum physical formalism. It may happen that novel quantum-like models would be explored, cf. \cite{UB}.
Although the problem of combination of the order effect with the two types of reproducibility was motivated by applications of the quantum formalism
outside of physics, mainly cognitive psychology, it is also important for quantum foundations as a part of the problem about the structure of sequential
quantum measurements, cf. \cite{SQ}. There is also an important experimental dimension: {\it Can one find physical measurements exhibiting combination
of the order effect with adjacent and separated reproducibilities?}
In this paper we consider only the case of finite dimensional state spaces. (Only such state spaces are used up to now in applications to humanities.)
The situation in the infinite-dimensional case is very different, see, e.g., Proposition 8, and this case has to be studied separately.
\section{The basic consequence of adjacent reproducibility}
In this section we show that the class of effects corresponding to measurements determined by
the conditions ${\bf C0}, {\bf C1},$ and satisfying the condition of adjacent reproducibility
coincides with the class of orthogonal projectors. Hence, the class the corresponding POVMs coincides with the class
of von Neumann-L\"uders observables. But, in general, the state transformers are not reduced to projectors. Here the basic result
is presented in Theorem 1. (This statement was formulated in \cite{BKDB}. However, its proof contained a loophole which could not be closed.)
\medskip
{\bf Theorem 1.} {\it Let $E$ be a Hermitian operator such that $0 \leq E \leq I$ (an effect) and let $\langle E\phi, \phi\rangle=1$
for some pure state $\phi.$ Then this $\phi$ is an eigenvector of $E$ with the eigenvalue $\lambda=1.$}
{\bf Proof.} Consider in $H$ a basis $(e_j)$ consisting of eigenvectors of $E,$ i.e.,
$E e_j= \lambda_j e_j.$ Here $0\leq \lambda_j \leq 1.$ We have $E \phi= \sum_j \lambda_j \phi_j e_j,$
where $\phi_j = \langle \phi, e_j \rangle.$ We have $ \langle \phi, \phi \rangle = \sum_j \vert \phi_j \vert^2=1.$ We also have
$\langle E\phi, \phi \rangle = \sum_j \lambda_j \vert \phi_j \vert^2=1.$ Set $O_{\phi}=\{j: \phi_j \not=0\}$ (this set depends on $E).$ Then
we have the system of two quadratic equalities:
\begin{equation}
\label{tttt}
\sum_{j\in O_{\phi}} \vert \phi_j \vert^2=1,
\end{equation}
\begin{equation}
\label{tttt1}
\sum_{j\in O_{\phi}} \lambda_j \vert \phi_j \vert^2=1.
\end{equation}
If at least one $\lambda_j$ in (\ref{tttt1}) is strictly less than 1, then we come to contradiction with (\ref{tttt1}).
Thus, in (\ref{tttt1}) all $\lambda_j=1,$ i.e., if $j \in O_{\phi},$ then $\lambda_j=1.$
Hence $ E\phi = \sum_{j \in O_{\phi}} \phi_j e_j = \phi.$
\medskip
{\bf Corollary 1.} {\it For an effect $E$ and the corresponding state transformer of the type ${\bf C1},$
the condition of adjacent reproducibility is equivalent to the operator
equality:}
\begin{equation}
\label{tmu}
E M =M .
\end{equation}
\medskip
{\bf Theorem 2.} {\it For an effect $E$ and the corresponding state transformer of the type ${\bf C1},$ the operator equality (\ref{tmu})
is equivalent to the condition ``$E$ is an orthogonal projector''.}
{\bf Proof.} Consider the representation of all operators by matrices in the orthonormal basis $(e_j, j=1,...,n)$
consisting of eigenvectors of the operator $E: E=(E_{ij})=\rm{diag} (\lambda_1,..., \lambda_n), M=(m_{ij}).$
Suppose that this operator has an eigenvalue, e.g., $\lambda=\lambda_1,$ which is different from 0 and 1.
Consider the vector given by the first row in $M, w= (m_{11}, ...., m_{1n}).$ Then the equality (\ref{tmu}) implies
that $\lambda_1 w=w,$ and, hence, $w=0.$ Thus the first row of $M$ consists of zeros (hence, the first column of $M^*$ also
consists of zeros).
Consider the vectors $v_1=(m_{21},..., m_{n1}),..., v_n=(m_{2n},..., m_{nn}),$ the columns of $M$ without the first
zero element. Since $E=M^* M,$ its matrix elements can be represented as
\begin{equation}
\label{tmu7}
E_{ij}= \langle v_i, v_j\rangle.
\end{equation}
Since the off-diagonal elements in $E$ are zeros, we have:
$ v_1 \perp v_j, j\not=1, ..., v_n \perp v_j, j\not=n.$ Consider the first condition of orthogonality. Here
a vector in the $n-1$ dimensional space, namely, $v_1,$ is orthogonal to $n-1$ vectors. There are two possibilities: either $v_1=0$
or vectors $v_2,...,v_n$ are linearly dependent. In the first case (\ref{tmu7}) implies that
$\lambda_1= \langle v_1, v_1\rangle=0.$
Suppose now that $v_1\not=0,$ and, e.g., for $k=2,$ $v_2= \sum_{j\not=1, 2} c_j v_j.$ But $v_2 \perp v_j, j \not=2,$ hence $v_2=0.$ This means that
vectors $v_3,..., v_n$ are linearly dependent, e.g., $v_3= \sum_{j\not=1, 2, 3} c_j v_j.$ In this way we get that $v_3$ is also zero
and so on. Finally, we get that all vectors $v_2,..., v_{n-1}$ equal to zero. The vectors $v_1$ and $v_2$ have to be orthogonal.
Now let us take again the relation $E M =M$ into account.
Consider vectors $w_j= (m_{j1}, ...., m_{jn}), j=1,...,n$ (the rows of the matrix $M$); so the vector $w$ used above coincides with $w_1.$
We have $\lambda_2 w_2=w_2,$ but $\lambda_2=\langle v_2, v_2\rangle=0,$ so $w_2=0.$ In the same way we get that $w_3,..., w_{n-1} =0.$
Thus in the vectors $v_1$ and $v_2$ only the last coordinate can be nonzero, but such vectors can be orthogonal only if one of them is zero.
The worst case would be $v_1\not=0,$ but $v_n=0.$ But then $\lambda_n=0$ and hence $w_n=0,$ so the last coordinate in $v_1$ also has to be zero.
\section{The structure of state transformers for effects of the projection type}
In the previous section we showed that, for measurements satisfying the condition of adjacent reproducibility, the effects compositing
observables, POVMs, are, in fact, orthogonal projectors. Now we plan to describe the structure of the corresponding state transformer
operators, see ${\bf C1}.$ We shall show that they are simply the compositions of the projectors-effects with unitary operators. This mathematical fact
will play the crucial role in our further studies.
We start with two simple lemmas about unitary operators which will be useful in the further considerations
\medskip
{\bf Lemma 1.} {\it Let $U: H\to H$ be a unitary operator and let $X$ be its invariant subspace, i.e.,
$U X \subset X.$ Then the orthogonal complement of $X$ is also invariant subspace of $U,$ i.e.,
$U X^\perp \subset X^\perp.$}
{\bf Proof.} Since the kernel of a unitary operator is zero, i.e., $Ux=0$ iff $x=0,$ dimensions
of the spaces $X$ and $UX$ are equal. Thus, for invariant subspace, we have that $UX=X.$ Hence, any
$x\in X,$ can be represented as $x= U x_0, x_0 \in X.$ Take any $y \in X^\perp.$ Then $\langle U y, x\rangle=
\langle U y, U x_0\rangle= \langle y, x_0\rangle=0.$
\medskip
{\bf Corollary 2.} {\it For any invariant subspace $X,$ the unitary operator $U$ can be decomposed as $U=\rm{diag} (V, W),$ where
$V:X\to X$ and $ W:X^\perp \to X^\perp$ are unitary operators.}
\medskip
Consider representation of an orthogonal projector $P$ in the form used in ${\bf C1}$ to define the
corresponding state transformer:
\begin{equation}
\label{SVD0}
P= M^* M.
\end{equation}
Let us fix some basis and represent all operators by matrices.
It is convenient to select a basis in which $P$ is diagonal $P =\rm{diag} (1,...,1, 0,...,0).$
By the singular value decomposition theorem any matrix $M$ can be represented in the form
\begin{equation}
\label{SVD}
M= W \Sigma V^*,
\end{equation}
where the matrices $W$ and $V$ are unitary and $\Sigma$ is a diagonal matrix such that its elements are square roots of
the eigenvalues of the matrix $M^* M.$ Hence, in the case $M^* M=P$ the matrix $\Sigma$ coincides with $P.$ Thus,
in the representation (\ref{SVD0}) we can always select
\begin{equation}
\label{SVD0a}
M = W P V^*.
\end{equation}
Moreover, we have $P= M^* M= V P V^*$ or $ VP=PV$ and $V^*P =P V^*.$ Thus the representation (\ref{SVD0a}) can be written
as
\begin{equation}
\label{SVD0b}
M = W V^* P.
\end{equation}
We remark that the composition of two unitary operators is a unitary operator.
Thus we obtained the following mathematically simple (but very useful for our further studies) result
\medskip
{\bf Lemma 2.} {\it All representations of the projector $P$ in the form
(\ref{SVD0}) are given by operators having the form:}
\begin{equation}
\label{SVD0d}
M = U P,
\end{equation}
where $U$ is a unitary operator.
\section{Main results on combination of adjacent and separated reproducibilities with the order effect}
Let $M= U P,$
where $P$ is a projector and $U$ is a unitary operator. Set $H_P= P H.$
Set $A= M^* M= P.$ So, the observable $A$
is given by a projector (Hermitian!), but the corresponding state transformer
is not of the L\"uders type (not measurement of the first kind):
\begin{equation}
\label{ST}
\psi \to \phi_A=\frac{U P\psi}{\Vert U P\psi\Vert}.
\end{equation}
{\bf Proposition 1.} {\it Observable $A=P$ with the state transformer (\ref{ST})
has the property of $A-A$ repeatability iff}
\begin{equation}
\label{ST0}
PUP=UP
\end{equation}
{\bf Proof.} $$p_{A-A}=\langle P \phi_A, \phi_A\rangle=1.$$
Then, as we know, $$
P \phi_A= \phi_A.
$$
Thus, $P M \psi= M\psi$ or $PUP=UP.$
\medskip
{\bf Corollary 2.} {\it Observable $A=P$ with the state transformer (\ref{ST})
has the property of $A-A$ repeatability iff the subspace $H_P$ is invariant
with respect to the action of the unitary operator $U.$ }
\medskip
Now we consider two observables:
Let $M= U_1 P_1$ and $N= U_2 P_2,$
where $P_j, j=1,2,$ are projectors and $U_j$ are unitary operators. Set $H_j= P_j H$ and $H_{12}= H_1 \cap H_2.$
Set $A= M^* M= P_1U_1^*U_1 P_1=P_1$ and $B= N^* N= P_2U_2^*U_2 P_2=P_2.$ So, the observables $A$ and $B$
are given by projectors $P_j$ (Hermitian!), but the corresponding state transformers
are not of the L\"uders type (these are not measurements of the first kind):
\medskip
{\bf Proposition 2.} {\it A pair of observables $A$ and $B$ exhibits the order effect iff}
\begin{equation}
\label{ST1}
P_1U_1^* P_2 U_1P_1 \not = P_2 U_2^* P_1 U_2P_2
\end{equation}
{\bf Proof.} For the $A-B$ measurement sequence, we have:
$$
p_{A-B}= \Vert \phi_A \Vert ^2 \langle P_2 \phi_A, \phi_A \rangle = \langle P_2 U_1P_1 \psi, U_1P_1 \psi \rangle=
\langle P_1U_1^* P_2 U_1P_1 \psi, \psi \rangle;
$$
in the same way, for the $B-A$ measurement sequence, we obtain:
$$
p_{B-A}=
\langle P_2U_2^* P_1 U_2P_2 \psi, \psi \rangle;
$$
hence, these sequential measurements can give different results iff
(\ref{ST1}) holds true.
\medskip
{\bf Proposition 3.} {\it A pair of observables $A$ and $B$ exhibits the $A-B-A$ repeatability iff}
\begin{equation}
\label{ST2}
P_1N M= N M, \mbox{or}\; P_1 U_2 P_2U_1 P_1 = U_2 P_2U_1 P_1.
\end{equation}
{\bf Proof.} For the $A-B-A$ measurement sequence, we have :
$$
p_{A-B-A}= \langle P_1 \phi_{A-B}, \phi_{A-B} \rangle,
$$
where
$$
\phi_{A-B}= \frac{NM \psi}{\Vert NM \psi \Vert}.
$$
The condition $p_{A-B-A}=1$ is equivalent to the condition $P_1 \phi_{A-B} =\phi_{A-B},$
or (\ref{ST2}).
\medskip
Thus, we are looking for observables satisfying conditions (\ref{ST0})--(\ref{ST2}).
For simplicity, set $U_2=I,$ i.e., $B$ is measurement of the first kind.
Then we have the system of relations
\begin{equation}
\label{ST0a}
P_1U_1P_1=U_1 P_1 \; \mbox{or equivalently}\; U_1: H_1 \to H_1;
\end{equation}
\begin{equation}
\label{ST1a}
P_1 U_1^* P_2 U_1P_1 \not = P_2 P_1 P_2;
\end{equation}
\begin{equation}
\label{ST2a}
P_1 P_2 U_1 P_1 = P_2 U_1 P_1.
\end{equation}
We shall show by an example that these conditions can be jointly satisfied.
This statement can be reformulated in the form of the following proposition
playing an important role in applications to cognition \cite{BKDB}-\cite{C1}.
\medskip
{\bf Proposition 4.} {\it There exist quantum observables $A$ and $B$
generating the order effect and satisfying the conditions of $A-A, B-B,$ and
$A-B-A$ repeatability.}
\medskip
{\bf Example.} Consider four dimensional Hilbert space with some orthonormal basis $(e_1,e_2,e_3, e_4).$
Let $H_{1}$ and $H_{2}$ have bases $(e_1, e_2, e_3)$ and $(e_1,e_2, e_4),$ respectively, and let $P_1$ and
$P_2$ be projectors on these subspaces. It is crucial that $H_{12} = H_1 \cap H_2$ (with the basis $(e_1, e_2)$)
is nontrivial and at least two dimensional.
Let $U=(u_{ij})$ be a unitary operator acting in $H_1.$ It is extended on $H$ as $U_1=\rm{diag} (U,I),$ where $I$ is the unit operator
in $H_1^\perp$ and $(e_4)$ is the basis in the latter space. Thus the
condition (\ref{ST0a}) is satisfied.
We check now condition (\ref{ST2a}). For any $\psi \in H,$ the vector $U_1 P_1 \psi \in H_1.$ But $P_2: H_1 \to H_{12}.$ Hence, $P_2 U_1 P_1 \psi \in H_{12}$ and
action onto this vector by $P_1$ cannot change it.
Finally, we show that the condition (\ref{ST1a}) holds. The crucial point is that the right-hand side does not contain
the unitary operator $U_1,$ but the left-hand side contains it. The tricky point is that the left-hand side
contains both $U_1$ and it inverse $U_1^*.$ Thus, in principle, they can compensate actions of each other (cf. with the proof
of Proposition 7). We shall see
that this is not the case.
Let $\psi= \sum_{j=1}^4 x_j e_j.$ Then $\phi= P_2 P_1 P_2 \psi \in H_{12}$ and $\phi= x_1 e_1+ x_2 e_2.$
For the other side of (\ref{ST1a}), we have: $U_1P_1 \psi= \sum_{j=1}^3 y_j e_j,$ where $y_j= \sum_{i=1}^3 u_{ji} x_i.$ Now consider the next step:
$P_2 U_1P_1 \psi= y_1 e_1 + y_2 e_2 .$ Then
$$
\phi^\prime= U_1^* P_2 U_1P_1 \psi=\sum_{k=1}^3 z_k e_k,$$ where $z_k = \sum_{j=1}^2 \bar{u}_{jk} y_j= \sum_{j=1}^2 \sum_{i=1}^3 \bar{u}_{jk} u_{ji} x_i=
\sum_{i=1}^3 (\sum_{j=1}^2 \bar{u}_{jk} u_{ji}) x_i .$
Finally, since $\phi^\prime \in H_{1},$ it cannot be changed by the action of $P_1.$
We remark that, for a unitary matrix $U,$
$\sum_{j=1}^3 \bar{u}_{jk} u_{ji}= \delta_{ki}.$ If the subspace $H_{12}$ is also invariant
for $U_1$ (and, hence, for its block $U$), i.e., e.g., $U=\rm{diag} (W, 1),$ where $W$ is a unitary
operator in $H_{12},$ then $\sum_{j=1}^2 \bar{u}_{jk} u_{ji}= \delta_{ki}.$ Hence, in this case $\phi^\prime= x_1 e_1+ x_2 e_2= \phi.$
However, if $H_{12}$ is not invariant, then in general $\phi^\prime \not= \phi.$
\medskip
Now about the possibility to satisfy both $A-B-A$ and $B-A-B$ repeatability in combination with the order effect. As we have seen in the above
example, for a nontrivial order effect it is crucial that the unitary operator does not leave the space $H_{12} = H_1 \cap H_2$ invariant. Now we shall take this
into account in the general case. We turn again to the general case of two unitary operators $U_1, U_2.$ We collect the list of conditions:
\begin{enumerate}
\item $P_1 U_1^* P_2 U_1 P_1 \not = P_2 U_2^* P_1 U_2 P_2;$
\item $P_1 U_2 P_2 U_1 P_1 = U_2 P_2 U_1 P_1;$
\item $P_2 U_1 P_1 U_2 P_2 = U_1 P_1 U_2 P_2.$
\end{enumerate}
\medskip
{\bf Proposition 5.} {\it The combination of $A-A, B-B,$ and $A-B-A$ conditions of sequential repeatability implies that}
\begin{equation}
\label{Z1}
P_2 H_1= H_{12}
\end{equation}
and
\begin{equation}
\label{Z2}
U_2 H_{12} =H_{12}.
\end{equation}
{\bf Proof.} We start with remarking that $P_1 H=H_1$ and $U_1 P_1 H= H_1$ as well (Corollary 2 for the condition $A-A$). Then, we have that
$P_2 U_1 P_1 H= P_2 H_1 \subset H_2$ and, hence, $U_2 P_2 U_1 P_1 H= U_2 (P_2 H_1) \subset H_2$ (Corollary 2 for the condition $B-B$).
However, the second equality in the above list
of conditions implies that $U_2 P_2 U_1 P_1 H = U_2 (P_2 H_1) \subset H_1$ as well. Thus, $U_2 (P_2 H_1) \subset H_1 \cap H_2= H_{12}.$ Now
we play by using finite dimensionality of the model. We remark that $H_{12} \subset H_1$ and, hence, $H_{12} \subset P_2 H_1.$ We remark
that the dimensions of the subspaces $P_2 H_1$ and $ U_2 P_2 H_1$ coincide. Since $H_{12} \subset P_2 H_1$ and $U_2 (P_2 H_1) \subset
H_{12},$ We obtain that $P_2 H_1 = H_{12}$ and $U_2: H_{12} \to H_{12} ,$ i.e., $H_{12} $ is invariant subspace of the unitary operator
$U_2.$
\medskip
{\bf Lemma 3.} {\it The condition (\ref{Z1}) is equivalent to commutativity of projectors:}
\begin{equation}
\label{Z3a}
[P_1, P_2]=0.
\end{equation}
{\bf Proof.} a). Let (\ref{Z1}) hold. Then, for any $\psi \in H,$ we have $P_2 P_1 \psi \in H_{12}.$ Hence, $P_1 P_2 P_1 \psi = P_2 P_1 \psi.$
Thus,
\begin{equation}
\label{Z3}
P_1 P_2 P_1 = P_2 P_1.
\end{equation}
Apply the operation of adjoint operator to the both sides of this equality. We get:
\begin{equation}
\label{Z4}
P_1 P_2 P_1 = P_1 P_2.
\end{equation}
Hence, we derived (\ref{Z3a}).
b). Let (\ref{Z3a}) hold. Then $P_2 H_1= P_2 P_1 H= P_1 P_2 H\subset H_1$ and, hence, $P_2 H_1 \subset H_{12}.$
\medskip
In the same way the $B-A-B$ condition implies that
\begin{equation}
\label{Z1a}
P_1 H_2= H_{12}
\end{equation}
and
\begin{equation}
\label{Z2a}
U_1 H_{12} =H_{12}.
\end{equation}
Condition (\ref{Z1a}) is as well equivalent to commutativity of projectors. In particular, the conditions (\ref{Z1}) and (\ref{Z1a})
are equivalent. The above results can be presented in the form of Proposition useful for our cognitive applications, see, e.g., \cite{BKDB}-\cite{C1}.
\medskip
{\bf Proposition 6.} {\it The combination of the sequential repeatability conditions $A-A$ and $B-B$ with $A-B-A$ or
$B-A-B$ implies that the projectors commute.}
\medskip
We emphasize that $A-B-A$ implies only $U_2$ invariance of $H_{12}$ and not $U_1$ invariance (as can be seen from the
Example). In the same way $B-A-B$ implies only $U_1$ invariance of $H_{12}$ and not $U_2$ invariance. Thus, commutativity
of projectors or conditions (\ref{Z1}) or (\ref{Z1a}) do not imply $A-B-A$ (nor $B-A-B)$ repeatability.
We can represent $H_j= H_{12} \oplus L_j, j=1,2,$ where $L_j$ is the orthogonal complement of $H_{12}$
in the subspace $H_j.$ In other terms $L_j= P_j H \cap (P_1 H \cap P_2 H)^\perp.$ We also can decompose
total Hilbert space:
\begin{equation}
\label{PPrr1}
H=H_{12} \oplus L_j \oplus H_j^\perp.
\end{equation}
This decomposition will play a crucial role in our further considerations.
We now show that the sequential repeatability conditions set a rigid constraint onto the geometric inter-relation between
subspaces $L_1$ and $L_2.$
\medskip
{\bf Proposition 6a.} {\it The combination of the sequential repeatability conditions $A-A$ and $B-B$ with $A-B-A$ or
$B-A-B$ implies that $L_1 \perp L_2.$}
{\bf Proof.} By Proposition 5 $A-B-A$ sequential repeatability implies that $P_2 H_1 =H_{12}.$
We show that, for any pair of vectors $v \in L_1$ and
$w \in L_2,$ they are orthogonal. We have $\langle v, w\rangle=\langle v, P_2 w\rangle= \langle P_2v, w\rangle =0,$ because
$P_2v \in H_{12}.$
\medskip
We remark that in Hilbert space geometry, subspaces $H_1$ and $H_2$ satisfying the condition $L_1 \perp L_2$ are
called {\it perpendicular}. It is well known that $P_1, P_2$ are the projectors onto perpendicular subspaces iff they commute.
By Lemma 1 the unitary operators can be represented as block-diagonal operators
\begin{equation}
\label{PPrr3}
U_j=\rm{diag} (V_j, W_j, T_j),
\end{equation}
where the diagonal blocks are unitary operators acting in spaces $H_{12}, L_j,$
and $H_j^\perp,$ respectively. This invariance decomposition is essential for our further considerations.
\medskip
{\bf Proposition 7.} {\it The combination of the sequential repeatability conditions $A-A$ and $B-B$ with
$A-B-A$ and $B-A-B$ is incompatible
with the order effect.}
{\bf Proof.} By Proposition 6a the Hilbert space $H$ can be represented as the direct sum:
\begin{equation}
\label{PPrr2}
H=H_{12} \oplus L_1 \oplus L_2 \oplus \tilde{H}, \; \tilde{H}= (H_{12} \oplus L_1 \oplus L_2)^\perp.
\end{equation}
Take any $\psi \in H;$ it can be represented as
$\psi= \psi_{12} + \phi_1 + \phi_2 + \tilde{\psi}$ with components belonging the corresponding subspaces in (\ref{PPrr2}).
First, we calculate $P_1 U_1^* P_2 U_1 P_1 \psi.$ We have $P_1 \psi = \psi_{12} + \phi_1.$ Then, $U_1 P_1 \psi =
V_1\psi_{12} + W_1 \phi_1,$ see (\ref{PPrr3}). It is crucial that $V_1: H_{12} \to H_{12}, W_1: L_1 \to L_1.$ Thus, $W_1 \phi_1 \perp L_2.$
Hence, $P_2U_1 P_1 \psi = V_1\psi_{12}.$ Then we get $U_1^* P_2 U_1 P_1 \psi = V_1^* V_1\psi_{12}= \psi_{12}.$
In the same way we obtain that $P_2 U_2^* P_1 U_2 P_2 \psi = \psi_{12},$ i.e., no order effect.
\section{Appendix: infinite-dimensional case}
The following result shows that the considerations of this paper cannot be repeated in the infinite-dimensional case:
\medskip
{\bf Proposition 8.} {\it Let $E=M^*M$. If the dimension of $H$ is infinite, then the equality $E M =M$ does not imply that
$E$ is an orthogonal projector.}
{\bf Proof.} We construct an example of the operator $E$ which is not an orthogonal projector, but the equality $EM=M$ holds.
Consider an orthonormal basis $(e_j, j=1,2,...)$ in $H.$ Take an arbitrary complex number $a.$ We define the operator $M$ as
\begin{equation}
\label{amba}
Me_1 = a e_2, \; M e_{n} = e_{n+1}, n \not=1.
\end{equation}
It is easy ro find the adjoint operator:
\begin{equation}
\label{amba1}
M^* e_1 =0, M^* e_2= \bar{a} e_1, \; M^* e_{n+1} = e_{n}, n \not=1.
\end{equation}
Then we have
\begin{equation}
\label{amba2}
E e_1 = M^* Me_1 = a M^* e_2 = \vert a\vert^2 e_1, \; E e_{n}= M^* M e_{n} = M^* e_{n+1}= e_{n}, n \not=1,
\end{equation}
and
\begin{equation}
\label{amba3}
E M e_1 = a E e_2 = a e_2 = Me_1, \; E M e_{n} = E e_{n+1}= e_{n+1} =M e_{n}, n \not=1
\end{equation}
\section*{Acknowledgments}
The authors would like to thank J. Busemeyer and E. Dzhafarov for fruitful discussions and M. D' Ariano, P. Lahti, W.M. de Muynck, and M. Ozawa
for knowledge transfer.
|
\section{Introduction} \label{secIntro}
Many-server queueing models with abandonment have attracted substantial attention because of their appealing applications to customer contact centers and healthcare; see, e.g., \cite{B05}, \cite{GKM03}, \cite{GMR02}, \cite{G06},
and references therein.
In the $G_t/GI/N+GI$ model, there are $N$ parallel servers, and customers arrive with a time-varying arrival rate, require i.i.d. service times,
and have i.i.d. patience times;
the arrival process, service and patience times are assumed to be mutually independent. The service discipline is first-come-first-served (FCFS) and non-idling, that is, no server will idle whenever there is a customer in queue.
Because of the difficulty in the exact analysis of such stochastic systems, fluid models have been recently developed to approximate the system dynamics and performance measures in a many-server heavy-traffic regime, where the arrival rate and the number of servers get large and service and patience time distributions are fixed.
The conventional approach of using total number of customers in the system to describe system dynamics is insufficient to give a complete description and study some performance measures. Thus, measure-valued and two-parameter processes that track elapsed or residual service and patience times of each customer have been recently used to study these stochastic models.
Whitt \cite{WW06} pioneered the use of two-parameter processes to describe the system dynamics (Definition \ref{def:FM2}). In particular, $Q(t,y)$ represents the number of customers in queue at time $t$ that have waited for less than or equal to $y$, and $B(t,y)$ represents the number of customers in service at time $t$ that have received service for less than or equal to $y$.
His idea is to represent these two-parameter processes as integrals of their densities $q(t,y)$ and $b(t,y)$ with respect to $y$ (if they exist), respectively, which satisfy two fundamental evolution equations ((2.14) and (2.15) in \cite{WW06}), respectively.
A queue boundary process plays an important role in determining the real fluid queue size: the two-parameter density function $q(t,y)$ becomes zero for $y$ beyond the queue boundary at each time $t$. This approach is generalized to study the $G_t/GI/N_t+GI$ model with both time-varying arrival rates and numbers of servers \cite{LW11} and \cite{LW11b}.
Kang and Ramanan \cite{KR10}, following Kaspi and Ramanan \cite{KaR11}, used two measure-valued processes to describe the service and queueing dynamics, one
tracking the amount of time each customer has been in service, and the other tracking the amount of time each customer has spent in a potential queue, where all customers enter the potential queue upon arrival, and stay there until their patience times run out. The potential queue includes customers waiting in the real queue as well as those that have entered service or even departed but whose patience times have not run out.
They also use a frontier waiting-time process to track the waiting time of the customer in front of the queue at each time. This frontier waiting-time process is used to determine the real fluid queue dynamics from the measure-valued process for the potential queue. The description of system dynamics is then completed by the balance equations for the fluid content processes associated with the queue, the service station and the entire system, as well as the non-idling condition; see Definition \ref{def:FMA}.
We summarize these two approaches of tracking elapsed service and patience times by stating that the two-parameter process approach in Whitt \cite{WW06} describes the system dynamics by the densities and rates, while the measure-valued process approach in Kang and Ramanan \cite{KR10} describes the system dynamics by the distributions and counting processes directly.
The existence and uniqueness of Whitt's two-parameter fluid model are shown in discrete time under the assumption that the service and patience times have densities in \cite{WW06}. They also follow, as a special case, from the existence and uniqueness results established in \cite{LW11, LW11b} of the two-parameter fluid model for $G_t/GI/N_t+GI$ queueing model with both time-varying arrival rates and numbers of servers under the assumptions that the system only alternates between overloaded and underloaded regimes (with a finite number of alternations in each finite time interval) and that the service and patience time distributions have piecewise continuous densities.
The existence and uniqueness of Kang-Ramanan's fluid model are established in \cite{KR10} via the fluid limits and more recently in \cite{Kang} via the characterization of fluid model solution directly under the assumptions that the service time distribution $G^s$ has density and the hazard rate function $h^r$ of patience times is a.e. locally bounded. Zu{\~n}iga \cite{Zu14} has recently extended Kang-Ramanan's fluid model for general service time distributions and continuous patience time distributions.
One would expect that the two approaches are equivalent since they are different formulations for the same $G_t/GI/N+GI$ queue.
Our first main result is to establish this equivalence in Theorem \ref{wellposed}: first, a set of two-parameter fluid equations derived from the measure-valued fluid model satisfies the fluid model equations in \cite{WW06} (see Proposition \ref{fluidtwo}), and second, a set of measure-valued fluid equations derived from the two-parameter fluid model satisfies the fluid model equations in \cite{KR10} (see Proposition \ref{fluidtwo2}).
The equivalence property we establish provides a proof for the conjecture on the existence and uniqueness of Whitt's two-parameter fluid model under the assumption that the service and patience time distributions have densities (Conjecture 2.2 in \cite{WW06}). The two-parameter process formulation depends critically on the existence of the densities of the service and patience time distributions, since the densities of the two-parameter processes may not exist for general service and patience time distributions (see Remark \ref{rmk:lem21}).
Aa a different approach, the system dynamics of $G_t/GI/N+GI$ queues can also be described by tracking residual service and patience times. It was conjectured in Section 3.3.2 of Kaspi and Ramanan \cite{KaR11} (in the case of no abandonment) that a measure-valued fluid model that tracks customers' residual service times and patience times can also be formulated in parallel to the fluid model tracking elapsed times. One advantage of considering a fluid model tracking residual times is that it enables us to easily analyze some performance measures, such as the system workload at any given time, which rely directly on the customers' residual service times; see, e.g., \cite{GW91, PW10} for infinite-server models and \cite{KaR11} for $G_t/GI/N$ queues.
Such a fluid model tracking residual times, if formulated properly, should be also equivalent to the above three fluid models tracking elapsed times, since all of them are formulated for the same $G_t/GI/N+GI$ queueing system.
Zhang \cite{Zhang} provided a fluid model tracking residual times for the $G/GI/N+GI$ model with a constant arrival rate (Definition \ref{def:FMZ}). Instead of using the potential queue as described in the fluid models tracking elapsed times, Zhang's model uses a virtual queue to describe the queueing dynamics, where all customers enter the virtual queue upon arrival and stay there until their time to enter service, which may include customers whose patience times have run out already. The existence and uniqueness of this fluid model are shown assuming continuous service time distribution and Lipschitz continuous patience time distributions \cite{Zhang}.
We study the relationship of Zhang's fluid model with the above three fluid models, in particular, focusing on Zu{\~n}iga's fluid model, and find that they are not equivalent formulations for the $G/GI/N+GI$ queue under general initial conditions; see Remarks \ref{remk:la}-\ref{remark:ari} in Section \ref{secZhanglimitation}.
The disparity lies in the initial conditions for the fluid models, in particular, the assumptions imposed on the initial contents in the virtual queue and in service in Zhang's fluid model. We identify the set of necessary and sufficient conditions on the initial contents
for the equivalence of Zhang's fluid model and the above three fluid models (Theorems \ref{lem:ZKR} and \ref{lem:KRZ} and Corollary \ref{coro:KRZ}). However, in some real-life applications where the initial conditions are not satisfied (e.g., the initial content does not have the same arrival rate as new arrivals), Zhang's fluid model cannot be used.
On the other hand, from Kang-Ramanan and Zu{\~n}iga's fluid models, we obtain measure-valued fluid processes tracking residual service and patience times, which, together with the same input data as in those two fluid models, describe the service and real queueing dynamics of the same $G_t/GI/N+GI$ systems.
These processes tracking residual times play an important bridging role in the discussion of the non-equivalence of Zhang's fluid model and the fluid models tracking elapsed times.
These equivalence properties established in the paper are significant to understand the fluid dynamics of the $G_t/GI/N+GI$ model from different perspectives.
They provide the flexibility of choosing the most convenient approach among the different formulations, tracking elapsed or residual times, and the possibility of applying results from one formulation to another.
We illustrate this by two examples.
First, an asymptotic periodic property is proved in \cite{LW11c} for the two-parameter fluid model tracking elapsed times for the $G_t/M_t/N_t+GI_t$ queueing model, and thus, should also hold for the associated measure-valued fluid models tracking elapsed and residual times (in the special case of $G_t/M/N+GI$ queues).
Second, it is important to show that for a fluid model, the fluid solutions converge uniformly to the steady state over all possible initial states. That has been a difficult task for general non-Markovian many-server models.
Thus, the equivalence property in this paper paves the way to show this with possibly any of the fluid models, whichever most convenient (see \cite{LZ} for some recent attempts in this direction).
Although these equivalence properties are established for the fluid limits of the associated fluid-scaled stochastic processes in the queueing model, it is conceivable that the proofs for the convergence to these fluid limits may also be unified. The two-parameter approach proves the convergence in the functional space $\sD_\sD = \sD([0,\infty), \sD([0,\infty), \RR))$ endowed with the Skorokhod $J_1$ topology. The measure-valued approach proves the convergence in the measure-valued functional space $\sD([0,\infty), \mathcal{M}([0,\infty)))$ where $\mathcal{M}([0,\infty))$ is the space of Radon measures on $\RR_+$ endowed with the Borel $\sigma$-algebra. Tracking elapsed times enables us to use martingale arguments \cite{KR10}, but tracking residual times uses a different approach to prove the convergence \cite{Zhang}.
So it is interesting to ask how these different approaches to establish the convergence are related and what would be the most general assumptions on the system primitives. We believe that these equivalence and coupling properties are useful in the study of other non-Markovian many-server queueing systems and networks.
\vspace{0.1in}
{\it Organization of the paper. }
The rest of the paper is organized as follows. We finish this section with some notation.
In Section \ref{secFluidE}, we first review the definitions of the three fluid models tracking elapsed times, and then show their equivalence (Theorem \ref{wellposed}), whose proof is given in Appendix \ref{secConn}.
In Section \ref{secFluidR}, we first state and discuss the fluid measure-valued processes tracking residual times derived from Kang-Ramanan and Zu{\~n}iga's fluid models in Section \ref{secMR}.
We then review Zhang's fluid model in Section \ref{secZhang} and discuss its connection with the three fluid models tracking elapsed times in Section \ref{secZhanglimitation}.
\vspace{0.1in}
{\it Notation. }
We use $\RR$ and $\RR_+$ to denote the spaces of real numbers and nonnegative real numbers, respectively. Given any metric space $S$, $\sC_b(S)$ is the space of bounded, continuous real-valued functions on $S$. Let $\sC_c(\RR_+)$ be the space of continuous real-valued functions on $\RR_+$ with compact support.
Given a Radon measure $\xi$ on $[0,H)$ and an interval $[a,b] \subset [0,H)$, we will use $\xi[a,b]$ to denote $\xi([a,b])$.
Let $\mathcal D^{abs}_{[0,\infty)}(\mathcal M[0,H))$ denote the set of measure-valued processes $\mu$ with values in $\mathcal M[0,H)$, the space of Radon measures on $[0,H)$, such that for any $t\geq 0$, the measure $\int_0^t \mu_s (\cdot) ds$ is absolutely continuous with respect to the Lebesgue measure on $[0,H)$.
Let $\mathcal D_{[0,\infty)}(\RR)$ be the space of real-valued c{\'a}dl{\'a}g functions on $[0,\infty)$.
\iffalse
For any non-decreasing function $f$ on $\RR_+$, $f^{-1}$ denotes the inverse function of $f$ in the sense that
$
f^{-1}(y) \doteq \inf\{x\ge 0: f(x) > y\}.
$
Similarly, given a non-increasing function $f$ defined on $\RR_+$, $f^{-1}$ denotes the inverse function of $f$ in the sense that $f^{-1}(y)=\inf\{t\geq 0:\ f(t)< y\}$.
\fi For each real-valued function $f$ defined on $[0,\infty)$, let $f^+$ and $f^-$ be the positive and the negative parts of $f$, respectively, that is, $f^+(t)=f(t)\vee 0$ and $f^-(t)=-(f(t)\wedge 0)$ for each $t\geq 0$.
\section{Fluid models tracking elapsed times} \label{secFluidE}
In the $G_t/GI/N+GI$ fluid models, we let $E(t)$ represent the cumulative amount of fluid content (representing customers) entering the system in the time interval $(0,t]$ for each $t>0$. Assume that
$E$ is a non-decreasing function defined on $[0,\infty)$ with the density function $\la(\cdot)\geq 0$, that is,
\beql{f0}
E(t)=\int_0^t \lambda(s)\,ds,\quad t\geq 0.
\eeq
Let $G^s$ and $G^r$ denote the service and patience time distribution functions, respectively. We assume that $G^s(0+) = G^r(0+) =0$. Let
$$
H^r \doteq \inf\{x \in \RR_+: G^r(x)=1\}, \quad H^s \doteq \inf\{x \in \RR_+: G^s(x)=1\}.
$$
Then $H^r$ and $H^s$ are right supports of $G^r$ and $G^s$, respectively.
\subsection{Whitt's two-parameter fluid model} \label{secFluidtwo}
In this section we state a modified version of the two-parameter fluid model in Whitt \cite{WW06}. We assume that the functions $G^s$ and $G^r$ have density functions $g^s$ and $g^r$ on $[0,\infty)$, respectively. Let the hazard rate functions of $G^s$ and $G^r$ be defined as $h^r \doteq g^r/\bar{G^r}$ on $[0,H^r)$ and $h^s \doteq g^s/\bar{G^s}$ on $[0,H^s)$, respectively, where $\bar{G^r} = 1- G^r$ and $\bar{G^s} = 1- G^s$.
Let the two-parameter processes $B(t,y)$ be the amount of fluid content in service at time $t$ that has been in service for less than or equal to $y$ units of time, $\tilde{Q}(t,y)$ be the amount of fluid content in the potential queue at time $t$ that has been in potential queue for less than or equal to $y$ units of time, which may include the fluid content that has entered service or even departed by time $t$, and $Q(t,y)$ be the portion of $\tilde{Q}(t,y)$ that excludes the fluid content which has entered service by time $t$.
Then it is obvious that $B(t, \infty)$ is the total fluid content in service and $Q(t, \infty)$ is the total fluid content in queue waiting for service.
It is assumed that these three processes are Lebesgue integrable on $[0,\infty)$ with densities $b(t,y)$, $\tilde{q}(t,y)$ and $q(t,y)$ with respect to the second component $y$, that is,
\beqal{w1}
&& B(t,y) = \int_0^y b(t,x)dx\leq 1, \quad \tilde{Q}(t,y) = \int_0^y \tilde{q}(t,x) dx\geq 0, \\
&& Q(t,y) = \int_0^y q(t,x) dx\geq 0. \non
\eeqa
Let $\tilde q(0,x)=q(0,x)$ as a function in $x$ have support in $[0,H^r)$ and $b(0,x)$ as a function in $x$ have support in $[0,H^s)$.
Note that in \cite{WW06}, it is not explicitly stated that the service and patience time distributions $G^s$ and $G^r$ can be of finite support.
\begin{definition} \label{def:FM2}
A pair of functions $(B(t,y),Q(t,y))$ is a two-parameter fluid model tracking elapsed times with the input data $(\lambda(\cdot), \tilde q(0,x),b(0,x))$ if it satisfies the following conditions.
$(i)$ The service density function $b(t,x)$ satisfies
\beql{w2}
b(t+u, x+u) = b(t,x) \frac{\bar{G}^s(x+u)}{\bar{G}^s(x)}, \quad x \in [0,H^s), \ t \ge 0, \ u > 0.
\eeq
$(ii)$ The potential queue density function $\tilde{q}(t,x)$ satisfies
\beql{w3}
\tilde{q}(t+u,x+u) = \tilde{q}(t,x) \frac{\bar{G}^r(x+u)}{\bar{G}^r(x)}, \quad x \in [0,H^r), \ t\ge 0, \ u >0.
\eeq
$(iii)$ There exists a queue boundary function $w(t)$ such that $\tilde{Q}(t,w(t))=Q(t,\infty)$ and then the queue density function $q(t,x)$ satisfies
\beql{w5}
q(t,x) = \begin{cases}
\tilde{q}(t,x), \quad x \le w(t), \\
0, \qquad \quad x > w(t).
\end{cases}
\eeq
$(iv)$ The density functions $b(t,x)$, $\tilde{q}(t,x)$ and $q(t,x)$ satisfy the following boundary properties:
\beql{w7}
b(t,0) = \begin{cases}
\la(t), \qifq B(t,\infty) < 1, \\
\sigma(t) \wedge \la(t), \qifq B(t,\infty) =1, \qandq Q(t,\infty) =0,\\
\sigma(t), \qifq B(t,\infty) = 1, \qandq Q(t,\infty) >0,
\end{cases}
\eeq
\beql{w9}
\tilde q(t,0) = \lambda(t),
\eeq
and
\beql{w8}
q(t,0) = \begin{cases}
\la(t), \qifq Q(t,\infty)>0 \ (w(t)>0),\\
\la(t) - (\sigma(t) \wedge \la(t)), \qifq B(t,\infty) =1, \qandq Q(t,\infty) =0,\\
0, \qifq B(t,\infty)<1,
\end{cases}
\eeq
where \beql{w6}
\sigma(t) = \int_{[0,H^s)} b(t,x) h^s(x) dx, \quad t\ge 0.
\eeq
$(v)$ The densities $\lambda(t),\ q(t,x),\ b(t,x)$ and $\alpha(t)$ satisfy the balance equation:
\beql{w10}
\int_0^t \lambda(s) ds+\int_0^\infty q(0,x)dx= \int_0^\infty q(t,x)dx+ \int_0^t b(s,0)ds+\int_0^t\alpha(s)ds, \eeq where \beql{w6.1}
\alpha(t) = \int_{[0,H^r)} q(t,x) h^r(x) dx, \quad t\ge 0.
\eeq
\end{definition}
In \cite{WW06}, equations \eqref{w2} and \eqref{w3} are called the first and second fundamental evolution equations, respectively.
Note that the first fundamental evolution equation \eqref{w2} essentially says that the fluid content in service that has not completed service remains in service. Similarly, the second fundamental evolution equation \eqref{w3} essentially says that the fluid content in the potential queue that has not reached its patience time remains in the potential queue. For each time $t$, the queue boundary quantity $w(t)$ divides the fluid content in the potential queue into two portions. The fluid content on the left side of $w(t)$ is still in queue waiting for service and the fluid content on the right side of $w(t)$ has entered service or even departed. The quantities $b(t,0)$, $\tilde q(t,0)$, $q(t,0)$ in condition (iv) above are exactly the rates at time $t$ at which the fluid content enters service, the potential queue and the queue, respectively. The quantities $\sigma(t)$ in (\ref{w6}) and $\alpha(t)$ in (\ref{w6.1}) are precisely the total service rate and the total abandonment rate at each time $t$, respectively.
At last, the balance equation \eqref{w10} is implicit in the definition of the fluid model in \cite{WW06} and stated in equation (6) in \cite{LW11}.
\begin{remark} \label{rmk:WW} $($Existence and uniqueness of Whitt's fluid model.$)$ Whitt \cite{WW06} has shown the existence and uniqueness of the two-parameter fluid model for $G_t/GI/N+GI$ queues in discrete time by proving a functional weak law of large numbers (FWLLN), and conjectured them in continuous time (cf. Conjecture 2.2 of \cite{WW06}). The existence and uniqueness of the two-parameter fluid model for $G_t/GI/N_t+GI$ queues with time-dependent staffing are shown in Liu and Whitt \cite{LW11, LW11b}, by an explicit characterization of the solution to the fluid model in \cite{LW11} and by proving an FWLLN in \cite{LW11b}, under the additional assumptions that the system only alternates between overloaded and underloaded regimes (with a finite number of alternations in each finite time interval) and that the service and patience time distributions have piecewise continuous densities. Thus, by specializing their argument to $G_t/GI/N+GI$ queues, the conjecture is established but with the previously mentioned additional assumptions.
In this paper, we establish the conjecture under the assumption that the service and patience time distributions have densities, without assuming, a priori, that the system only alternates between overloaded and underloaded regimes, by applying the equivalence between the two fluid models in Definitions \ref{def:FM2} and \ref{def:FMA} established in Theorem \ref{wellposed} below and the existence and uniqueness of Kang-Ramanan's fluid model established in \cite{KR10, Kang}. We remark that the existence of the densities of the service and patience time distributions is critical for the formulation of Whitt's two-parameter fluid model, because the densities of $B(t,y)$ and $Q(t,y)$ with respect to $y$ may not exist when the service and/or patience time distributions are general (see Remark \ref{rmk:lem21}).
\end{remark}
\subsection{Kang-Ramanan's measure-valued fluid model} \label{secFluidmeas}
In this section, we state the measure-valued fluid model in Kang and Ramanan \cite{KR10}. They use two measure-valued processes to describe the service and queueing dynamics. Let $\nu_t$ be a nonnegative finite measure on $[0,\infty)$ with support in $[0,H^s)$ such that $\nu_t(dx),\ x\in [0,H^s)$, represents the amount of fluid content of customers in service whose time spent in service by time $t$ lies in the range $[x,x+dx)$.
Let $\eta_t$ be another nonnegative finite measure on $[0,\infty)$ with support in $[0,H^r)$ such that $\eta_t(dx),\ x\in [0,H^r),$ represents the amount of fluid content in the potential queue whose time spent there by time $t$ lies in the range $[x,x+dx)$, where the potential queue is an artificial queue that includes the fluid content of customers in queue waiting for service and also the fluid content of customers that has entered service or even departed, but whose patience time has not been reached.
We assume that the functions $G^s$ and $G^r$ have density functions $g^s$ and $g^r$ on $[0,\infty)$, respectively.
Let $\mathcal S_0$ denote the set of triples $(\eta,\nu,x)$ such that $1 - \nu[0,H^s) = [1-x]^+$ and $\nu[0,H^s) + \eta[0,H^r)=x$, where $\eta$ is a non-negative finite measure on $[0,\infty)$ with support in $[0,H^r)$, $\nu$ is a non-negative finite measure on $[0,\infty)$ with support in $[0,H^s)$, and $x \in \RR_+$.
The set $\mathcal{S}_0$ represents all possible measures of $(\eta,\nu)$ and values of $x$ that the initial state of the measure-valued fluid model $(\eta,\nu,X)$ can take, satisfying the non-idling condition.
\begin{definition} \label{def:FMA}
A triple of functions $(\eta,\nu,X)$ is a measure-valued fluid model tracking elapsed times with the input data $(\lambda(\cdot),\eta_0,\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))\in \mathcal S_0$ if it satisfies the following equations. For every $\psi \in \sC_b(\RR_+)$ and $t\geq 0$,
\beql{f4}
\int_0^\infty \psi(x) \eta_t (d x) = \int_{[0,H^r)}\psi(x+t)\frac{\bar{G^r}(x+t)}{\bar{G^r}(x)}\eta_0(dx)+
\int_0^t \psi(t-s) \bar{G^r}(t-s) \la(s)ds ,
\eeq
\beql{f5}
\int_0^\infty \psi(x) \nu_t (d x) = \int_{[0,H^s)}\psi(x+t)\frac{\bar{G^s}(x+t)}{\bar{G^s}(x)}\nu_0(dx)+
\int_{[0,t]} \psi(t-s) \bar{G^s}(t-s) d K(s),
\eeq
where \begin{eqnarray}
K(t) &=& B(t) + D(t) - B(0) = \nu_t [0,H^s)+ D(t)- \nu_0 [0,H^s), \label{33} \\
D(t) &=& \int_0^t \left( \int_{[0,H^s)} h^s(x) \nu_s(dx) \right) ds , \label{f80}
\end{eqnarray}
\begin{eqnarray}
E(t) +Q(0)&=& Q(t) + K(t) + R(t), \label{31}
\end{eqnarray}
\beql{f8}
R(t) = \int_0^t \left( \int_{[0,\chi(s)]} h^r(x) \eta_s(dx) \right) ds,
\eeq
\beql{f9}
\chi(s) = \inf\{x\in [0,H^r): \eta_s[0,x] \ge Q(s)\},
\eeq
\beql{f1}
Q(t) = (X(t) - 1)^+,
\eeq
\beql{f2}
B(t) = \nu_t [0,\infty) = X(t) \wedge 1 = 1 - (1-X(t))^+,
\eeq
and
\beql{f12}
Q(t)(1-B(t)) = 0.
\eeq
\end{definition}
\iffalse
\begin{remark}
When the system starts empty and the arrival rate $\lambda(\cdot)$ is constant, the assumption on the existence of densities of $G^s$ and $G^r$ can be relaxed in the formulation of the fluid model since the processes $D$ and $R$ have the following alternative representations:
\[R(t)=\int_0^t G^r(\chi(s))ds.\] and
\[D(t) = \int_0^t \left( \int_0^s K(s-u) dG^s(u) \right) ds.\]
\end{remark}
\fi
In this fluid model, $B(t)$ represents the total fluid content of customers in service, $Q(t)$ represents the total fluid content of customers in queue waiting for service, and $X(t)$ represents the total fluid content of customers in the system at each time $t$. Then, by (\ref{f1}) and (\ref{f2}), \beql{32}
X(t)=B(t)+Q(t).
\eeq The additional quantities $K(t),\ R(t),\ D(t),\ \chi(t)$ can naturally be interpreted, respectively, as the cumulative amount of fluid content that has entered service by time $t$, the cumulative amount of fluid content that has abandoned from the queue by time $t$, the amount of fluid content that has departed the system after service completion by time $t$, and the waiting time of the fluid content at the head of the queue at time $t$, that is, the fluid content in queue with the longest waiting time.
For completeness, we now provide an intuitive explanation for these fluid equations. The equation (\ref{f4}) governs the evolution of the measure-valued process $\eta_t$. Note that when $x\leq t$, the amount of fluid content $\eta_t(dx)$ is the fraction of the amount of fluid content $\lambda(t-x)$ arriving to the system at time $t-x$ and whose time in the system since its arrival is more than $x$ by time $t$. It is easy to see that this fraction equals to $\bar{G^r}(x)$. When $x>t$, the amount of fluid content $\eta_t(dx)$ is the fraction of the amount of fluid content $\eta_0(d(x-t))$ initially in queue and whose waiting time is more than $x$ by time $t$ given that it is more than $x-t$ at time $0$. This fraction equals to $\bar{G^r}(x)/\bar{G^r}(x-t)$. This shows that (\ref{f4}) holds. A similar observation yields (\ref{f5}). The equations (\ref{33})--(\ref{31}) are simply mass conservation equations for the queue and the server station, respectively. Since $\nu_s(dx),\ x\in [0,s],$ represents the amount of fluid content in service whose time in service lies in the range $[x,x+dx)$ at time s, and $h^s(x)$ represents the fraction of the amount of fluid content with time in service $x$ (that is, with service time no less than $x$) that would depart from the system while having time in service in $[x, x + dx)$. Hence, it is natural to expect $\int_{[0,H^s)}h^s(x)\nu_s(dx)$ to represent the departure rate of fluid content from the fluid system at time $s$ and thus, expect (\ref{f80}) holds. A similar explanation can be applied to (\ref{f8}) except that, to consider the real reneging rate, we can only consider $x<\chi(s)$ since all the fluid content with the time in the system more than $\chi(s)$ has entered service by time $s$. The equation (\ref{f12}) represents the usual non-idling condition.
By adding (\ref{31}) and (\ref{33}) together and using (\ref{32}), we see that
\beql{total}
E(t) + X(0)=X(t)+R(t)+D(t).
\eeq
By the representations of $E$, $R$ and $D$ in (\ref{f0}), (\ref{f8}) and (\ref{f80}), we have from (\ref{total}) that $X$ is absolutely continuous. In turn, using the fact that $|[n- a]^+ - [n -b]^+| \leq |a - b|$, it is easy to see from (\ref{f2}) and (\ref{33}) that $B$ and then $K$ are absolutely continuous. So there exists a Lebesgue integrable function $\kappa$ such that
\beql{Kcont}
K(t)=\int_0^t \kappa(s)ds,\quad t\ge 0.
\eeq
By (\ref{33}) and (\ref{f80}), the process $K$ has the following representation:
\beql{t20}
K(t) = B(t)-B(0) + \int_0^t \left( \int_{[0,H^s)} h^s(x) \nu_s(dx)\right) ds.
\eeq
Then it follows from the same argument as in deriving (3.12) of \cite{KaR11} that the process $\kappa$ satisfies for a.e. $t\in \RR_+$,
\beql{kappa}
\kappa(t)=\left\{\begin{array}{ll} \lambda(t) & \mbox{ if } X(t)<1, \\
\lambda(t)\wedge \int_{[0,H^s)} h^s(x)\nu_t(dx) & \mbox{ if } X(t)=1, \\ \int_{[0,H^s)} h^s(x)\nu_t(dx) & \mbox{ if } X(t)>1. \end{array}\right.
\eeq
\begin{remark} \label{rem:euo} $($Existence and uniqueness of Kang-Ramanan's fluid model.$)$
Under the assumptions that the hazard rate functions $h^r$ and $h^s$ are either bounded or lower semi-continuous, Kang and Ramanan \cite{KR10} established the existence of the measure-valued fluid model in Definition \ref{def:FMA} by proving an FWLLN and also showed its uniqueness via the fluid model characterization. The existence and uniqueness of Kang-Ramanan's fluid model directly from the characterization of its solution is established in Kang \cite{Kang}, under the weaker assumptions that the service time distribution $G^s$ has density and the hazard rate function $h^r$ is a.e. locally bounded.
\end{remark}
Now we state our first result on the equivalence between the two fluid models described in Definitions \ref{def:FM2} and \ref{def:FMA}. Its proof is deferred to Appendix \ref{secConn}. As a consequence, it also gives a proof for Conjecture 2.2 of \cite{WW06} under the assumption that the service and patience time distributions have densities and $h^r$ is a.e. locally bounded.
\begin{theorem} \label{wellposed}
Existence and uniqueness of Whitt's fluid model in Definition \ref{def:FM2} is equivalent to existence and uniqueness of Kang-Ramanan's fluid model in Definition \ref{def:FMA} for the $G_t/GI/N+GI$ queue with the time-dependent arrival rate $\la(\cdot)$ and the initial data $(\eta_0,\nu_0,X(0))\in \mathcal S_0$, where $\eta_0(dx)=\tilde q(0,x)dx=q(0,x) dx$ and $\nu_0(dx)=b(0,x)dx$.
\end{theorem}
\subsection{Zu{\~n}iga's fluid model} \label{secFluidmeasE}
Recently, Zu{\~n}iga \cite{Zu14} extended Kang-Ramanan's fluid model without assuming that the patience time distribution $G^r$ and service time distribution $G^s$ have densities. In this section, we state this extended Kang-Ramanan's fluid model and establish some useful properties on certain quantities in the model, which are needed in the subsequent analysis.
Define a measure $M^r$ on $[0,H^r]$ by \[dM^r(x)\doteq \bone_{\{x<H^r\}}\bar{G}^r(x-)^{-1} d G^r(x) + \bone_{\{G^r(H^r-)<1\}}\delta_{H^r}(dx),\] and a measure $M^s$ on $[0,H^s]$ by \[dM^s(x)\doteq \bone_{\{x<H^s\}}\bar{G}^s(x-)^{-1} d G^s(x) + \bone_{\{G^s(H^s-)<1\}}\delta_{H^s}(dx).\]
\begin{definition} \label{def:EFMA}
A triple of processes $(\eta,\nu,X)\in \mathcal D^{abs}_{[0,\infty)}(\mathcal M[0,H^r))\times \mathcal D^{abs}_{[0,\infty)}(\mathcal M[0,H^s))\times \mathcal D_{[0,\infty)}(\RR)$ is a solution to an extended Kang-Ramanan's measure-valued fluid model with the input data $(\lambda(\cdot),\eta_0,\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))\in \mathcal S_0$ if $q_t$ and $p_t$, the densities of $\int_0^t \nu_s (\cdot) ds$ and $\int_0^t \eta_s (\cdot) ds$, respectively, satisfy the following conditions.
There exist $K(\cdot)$, a process of bounded variation started at $0$, $\chi(\cdot)$, $B(\cdot)$, $Q(\cdot)$, $D(\cdot)$, $R(\cdot)$ such that for every $\psi \in \sC_b(\RR_+)$ and $t\geq 0$, \eqref{f4}--\eqref{33}, \eqref{31}, \eqref{f9}--\eqref{f12} hold and \begin{eqnarray}
D(t) &=& \int_{[0,H^s]} q_t(x) dM^s(x) , \label{Ef80}
\end{eqnarray}
\beql{Ef8}
R(t) = \int_{[0,H^r]} \int_{[0,t]} \bone_{\{x\leq \chi(s)\}} d_s p_s(x) dM^r(x),
\eeq
where the integral with respect to $p_s(x)$ is defined as a Lebesgue-Stieltjes integral in $s$.
\end{definition}
\begin{remark} \label{rem:new}
Zu{\~n}iga's fluid model stated in Definition \ref{def:EFMA} is equivalent to Definition 3.4 of \cite{Zu14} due to Lemma 4.1 and Remark 4.2 of \cite{Zu14} and the given input data $(\lambda(\cdot),\eta_0,\nu_0,X(0))$. The main difference of Zu{\~n}iga's fluid model from Kang-Ramanan's fluid model in Definition \ref{def:FMA} is that the processes $D$ and $R$ satisfy \eqref{Ef80} and \eqref{Ef8} instead of \eqref{f80} and \eqref{f8} due to the lack of existence of densities of $G^s$ and $G^r$, respectively.
By Lemma 4.1 of \cite{Zu14}, the densities $q_t$ and $p_t$ can be written as
\beql{qt-Zu}
q_t(x) = \bar{G}^s(x-) K((t-x)^+) + \int_{[(x-t)^+, x)} \frac{\bar{G}^s(x-)}{\bar{G}^s(y)} \nu_0(dy),
\eeq
and
\beql{pt-Zu}
p_t(x) = \bar{G}^r(x-) E((t-x)^+) + \int_{[(x-t)^+, x)} \frac{\bar{G}^r(x-)}{\bar{G}^r(y)} \eta_0(dy).
\eeq
When $G^s$ and $G^r$ are assumed to have densities, $g^r$ and $g^s$, respectively, Zu{\~n}iga's fluid model is reduced to Kang-Ramanan's fluid model. Zu{\~n}iga's fluid model admits a unique solution (established in Theorem 3.5 via an FWLLN and Theorem 4.4 via the characterization of the fluid model in \cite{Zu14}) under the assumptions that $G^r$ is continuous, $\eta_0$ is diffuse, and $\nu_0$ is diffuse if $G^s$ is not continuous (Assumption 3.1 of \cite{Zu14}).
\end{remark}
We end this section by showing the following lemma for Zu{\~n}iga's fluid model in Definition \ref{def:EFMA}, which will be used in Section \ref{secFluidR} in discussing the relationship of the fluid models tracking elapsed times stated in this section and a fluid model tracking residual times stated in Section \ref{secZhang}.
\begin{lemma}\label{lem:chiinc}
In Definition \ref{def:EFMA}, the processes $D$ and $R$ have the following representations: for each $t\geq 0$,
\beql{mrd}
D(t) = \int_{[0,H^s)} \frac{G^s(y+t)-G^s(y)}{\bar G^s(y)} \nu_0(dy) +
\int_0^t G^s(t-s) dK(s) ,
\eeq
\beqal{mrq}
R(t) &=&
\int_{[0,H^r)} \left(\int_{(y,y+t]} \bone_{\{y\leq \chi(x-y)-(x-y)\}} dG^r(x)\right) \bar{G}^r(y)^{-1}\eta_0(dy) \non\\
&& \qquad + \int_0^t \int_{[0,H^r]} \bone_{\{x\leq s \wedge \chi(s)\}} \la(s-x) dG^r(x) ds .
\eeqa
Moreover, the process $K(t)$ is non-decreasing and the process $\chi(t)$ satisfies the following property:
\begin{eqnarray} \label{chiinc} \chi(t)-\chi(s)\leq t-s \mbox{ whenever } 0\leq s<t<\infty. \end{eqnarray}
\end{lemma}
\begin{remark} \label{rmk:lem21}
It is evident that the representation of the process $D$ in \eqref{mrd} implies that $D(t)$ is not absolute continuous when the service time distribution does not have density.
Thus, we cannot write the total service rate (departure rate) as in \eqref{w6}.
Although the two-parameter processes $B(t,y)$, $\tilde{Q}(t,y)$ and $Q(t,y)$ can be obtained as in \eqref{k2} from the Zu{\~n}iga's fluid model $(\nu_t, \eta_t, X)$ in Definition \ref{def:EFMA}, their densities with respect to $y$ may not exist and the associated two-parameter fluid model using densities $b(t,x)$ and $q(t,x)$ cannot be formulated with the densities as in Definition \ref{def:FM2}.
\end{remark}
\noindent {\it Proof of Lemma \ref{lem:chiinc}.}
By \eqref{Ef80} and \eqref{qt-Zu}, applying interchange of the order of integration and integration by parts, we easily obtain \eqref{mrd}.
To show $R(t)$ in \eqref{mrq}, from (\ref{Ef8}) and (\ref{pt-Zu}), we obtain that for each $t\geq 0$,
\begin{eqnarray} \label{Rr1}
&& \ R(t) = \int_{[0,H^r]} \int_0^t \bone_{\{x\leq \chi(s)\wedge s\}} \bar{G}^r(x-)\la(s-x)ds dM^r(x) \\ && \quad + \int_{[0,H^r]} \int_{[0,t]} \bone_{\{x\leq \chi(s)\}} d_s \left(\int_{[(x-s)^+,x)} \frac{\bar{G}^r(x-)}{\bar{G}^r(y)} \eta_0(dy)\right) dM^r(x) \non \\ &=& \int_0^t \int_{[0,H^r]} \bone_{\{x\leq \chi(s)\wedge s\}} \la(s-x)\bar{G}^r(x-)dM^r(x) ds \non \\ && \ + \int_{[0,H^r]} \int_{[0,x\wedge t]} 1_{\{x\leq \chi(s)\}} d_s \left(\int_{[x-s,H^r)} \bone_{\{y<x\}}\frac{\bar{G}^r(x-)}{\bar{G}^r(y)} \eta_0(dy)\right) dM^r(x) \non \\ &=& \int_0^t \int_{[0,H^r]} \bone_{\{x\leq \chi(s)\wedge s\}} \la(s-x)\bar{G}^r(x-)dM^r(x) ds \non \\ && \ + \int_{[0,H^r]} \int_{[[x-t]^+,x]} \bone_{\{x\leq \chi(x-s)\}} \bone_{\{s<x\}}\frac{\bar{G}^r(x-)}{\bar{G}^r(s)} \eta_0(ds) dM^r(x), \non \\
&=& \int_0^t \int_{[0,H^r]} \bone_{\{x\leq s \wedge \chi(s)\}} \la(s-x)\bar{G}^r(x-) dM^r(x) ds \non \\ & & \ + \int_{[0,H^r)} \left(\int_{(y,y+t]} \bone_{\{y\leq \chi(x-y)-(x-y)\}} \bar{G}^r(x-) dM^r(x)\right) \bar{G}^r(y)^{-1}\eta_0(dy) \non \\ &=& \int_0^t \int_{[0,H^r]} \bone_{\{x\leq s \wedge \chi(s)\}} \la(s-x) dG^r(x) ds \non \\ & & \ + \int_{[0,H^r)} \left(\int_{(y,y+t]} \bone_{\{y\leq \chi(x-y)-(x-y)\}} dG^r(x)\right) \bar{G}^r(y)^{-1}\eta_0(dy) , \non
\end{eqnarray}
where the second term in the second equality follows from Theorem 3.6.1 of \cite{Bo} with $X=[[x-t]^+,x]$ $Y=[0,x\wedge t]$, $f(s)=x-s$ and $\mu$ such that $\mu[a,b] = \int_{[a,b)}\frac{\bone_{\{y<x\}}}{\bar{G}^r(y)} \eta_0(dy)$ and the last equality follows from the interchange of the order of integrations.
We next prove the non-decreasing property of $K(t)$.
It follows from this representation of $R(t)$ in \eqref{mrq} that Lemma 4.4 of \cite{KR10} holds, that is, for any $0 \le a \le b < \infty$, if $Q(t) =0$ (equivalently, $\chi(t) =0$) for all $t \in [a,b]$, then $R(b) - R(a) =0$. Then the proof for the non-decreasing property of $K(t)$ will follow the same argument in Lemma 4.5 in \cite{KR10} using \eqref{Rr1}.
We now prove the property of $\chi(t)$ in \eqref{chiinc}.
By a similar argument as in Lemma 3.4 of \cite{KR10} on time shifts, to prove the lemma, without loss of generality, we may assume that $s=0$ in (\ref{chiinc}). Suppose that the property of $\chi(t)$ in \eqref{chiinc} does not hold, that is, there is a time $t_2> 0$ such that $\chi(t_2)>\chi(0)+t_2$. Let \[t_1 \doteq \sup\{u\leq t_2:\ \chi(u)\leq \chi(0)+u\}.\]
Then $\chi(t_1-)\leq \chi(0)+t_1$ and for each $u\in [t_1,t_2]$, \beql{chiu} \chi(u)\geq \chi(0)+u\geq \chi(t_1-)+(u-t_1) \mbox{ and } \chi(t_2)>\chi(t_1-)+(t_2-t_1). \eeq
By \eqref{Rr1}, it is clear that $R(t)-R(t-)\geq 0$ for each $t>0$. By applying the above display and time shift at $t_1$, we have
\begin{eqnarray*}
& & R(t_2)-R(t_1) \\ &= & \int_{[0,H^r)} \left(\int_{(y,y+t_2-t_1]} \bone_{\{y\leq \chi(t_1+x-y) -(x-y)\}} dG^r(x)\right) \bar{G}^r(y)^{-1}\eta_{t_1}(dy) \\ & & \qquad +
\int_0^{t_2-t_1} \left(\int_{[0,H^r]}\bone_{\{u\leq s\wedge \chi(t_1+s)\}} \la(t_1+s-u)dG^r(u) \right)ds.
\end{eqnarray*}
It follows from (\ref{chiu}) that $s\wedge \chi(t_1+s)=s$ and $\chi(t_1+s)-s\geq \chi(t_1-)$ for each $s\in (0,t_2-t_1]$. Hence the above display implies that
\begin{eqnarray*}
& & R(t_2)-R(t_1-) \\ &\geq &
\int_{[0,H^r)} \left(\int_{(y,y+t_2-t_1]} \bone_{\{y\leq \chi(t_1-)\}} dG^r(x)\right) \bar{G}^r(y)^{-1}\eta_{t_1}(dy) \\ & & \qquad +
\int_0^{t_2-t_1} \left(\int_{[0,H^r]}\bone_{\{u\leq s\}} \la(t_1+s-u)dG^r(u) \right)ds \\ &=& \int_{[0,H^r)} \bone_{\{y\leq \chi(t_1-)\}} \frac{G^r(y+t_2-t_1) - G^r(y)}{\bar{G}^r(y)}\eta_{t_1}(dy) \\ & & \quad + \int_{0}^{t_2-t_1} G^r(t_2-t_1-u) \la(t_1+u)du, \non
\end{eqnarray*}
where
and the second term on the right hand side of the last display follows from Proposition 0.4.5 of \cite{RM}.
Since (\ref{31}) holds for Zu{\~n}iga's fluid model $(\eta,\nu,X)$ and $K$ is non-decreasing, then the above three displays imply that
\begin{eqnarray*}
Q(t_2) &= & Q(t_1-) + (E(t_2)-E(t_1-)) -(R(t_2)-R(t_1-)) - (K(t_2)-K(t_1-)) \\ &\le & \eta_{t_1}[0,\chi(t_1-)] +\int_{t_1}^{t_2} \la(u)du - \int_{0}^{t_2-t_1} G^r((t_2-t_1-u)-) \la(t_1+u)du \\ & & \quad -\int_{[0,H^r)}\bone_{[0,\chi(t_1-)]}(x)\frac{G^r(x+(t_2-t_1))-G^r(x)}{\bar{G^r}(x)}\eta_{t_1}(dx) \\ &=& \int_{0}^{t_2-t_1} \bar{G^r}(t_2-t_1-u) \la(t_1+u)du \\ & & \quad + \int_{[0,H^r)}\bone_{[0,\chi(t_1-)+(t_2-t_1)]}(x+(t_2-t_1))\frac{\bar{G^r}(x+(t_2-t_1))}{\bar{G^r}(x)}\eta_{t_1}(dx) \\ &=& \eta_{t_2}[0,\chi(t_1-)+(t_2-t_1)].
\end{eqnarray*}
From this and the definition of $\chi$, we have $\chi(t_2)\leq \chi(t_1-)+(t_2-t_1)$, which contradicts (\ref{chiu}). Thus, the lemma is proved.
~~~\bsq
\iffalse
We first show some properties of quantities in Definition \ref{def:EFMA}, which will be used in the subsequent analysis. Its proof is given in Appendix \ref{secAppProofs}.
\begin{lemma}\label{lem:chiinc}
In Definition \ref{def:EFMA}, the process $K$ is non-decreasing, the process $\chi(t)$ satisfies the following property:
\begin{eqnarray} \label{chiinc} \chi(t)-\chi(s)\leq t-s \mbox{ whenever } 0\leq s<t<\infty, \end{eqnarray} and the processes $B$, $D$, $Q$ and $R$ have the following representations, respectively: for each $t\geq 0$,
\beql{mr621}
B(t)
=\int_{[0,H^s)} \frac{\bar G^s(y+t)}{\bar G^s(y)} \nu_0(dy) +
\int_0^t \bar G^s(t-s) dK(s),
\eeq
\beql{mrd}
D(t) = \int_{[0,H^s)} \frac{G^s(y+t)-G^s(y)}{\bar G^s(y)} \nu_0(dy) +
\int_0^t G^s(t-s) dK(s) ,
\eeq
\beql{mr61}
Q(t) = 1_{\{\varsigma(t)\leq 0\}}\int_{[0,-\varsigma(t)]} \frac{\bar G^r(y+t)}{\bar G^r(y)}\eta_0(dy) +
\int_{\varsigma^+(t)}^t \bar G^r(t-s) \la(s)ds ,
\eeq
\beqal{mrq}
R(t) &=& \int_{[0,H^r)} \left(\int_{(y,y+t]} 1_{\{y\leq -\varsigma(x-y)\}} \bar{G}^r(x-) dM^r(x)\right) (\bar{G}^r(y))^{-1}\eta_0(dy) \non \\ & & \qquad +
\int_0^t \int_{[0,H^r]} 1_{\{x\leq s-\varsigma^+(s)\}} \la(s-x)\bar{G}^r(x-) dM^r(x) ds,
\eeqa
where \beql{chiw}
\varsigma(t) = t - \chi(t).
\eeq
\end{lemma}
We next show that Zu{\~n}iga's fluid model naturally gives rise to two measure-valued processes that keep track of residual times of fluid content of customers in queue and in service. These two measure-valued processes play an important role in Section \ref{secFluidR}.
For each $t\geq 0$, clearly the following mapping
\begin{eqnarray*}\psi &\mapsto & \int_{[0,H^s)}\left(\int_{(y+t,\infty)} \frac{\psi(x-y-t)}{\bar G^s(y)}d G^s(x)\right)\nu_0(dy) \\
&& \qquad + \int_{[0,t]} \left(\int_{(t-s,\infty)} \psi(x-t+s)d G^s(x)\right)dK(s)\end{eqnarray*} is a positive linear functional on $\mathcal{C}_c(\RR_+)$ since $K$ is non-decreasing by Lemma \ref{lem:chiinc}. Then by Riesz-Markov-Kakutani representation theorem {\color{red} reference?}, there is a unique regular Borel measure $\nu^\ell_t$ on $\RR_+$ such that
\beqal{mr2}
\int_0^\infty \psi(x) \nu^\ell_t(dx) &=& \int_{[0,H^s)}\left(\int_{(y+t,\infty)} \frac{\psi(x-y-t)}{\bar G^s(y)}d G^s(x)\right)\nu_0(dy) \non\\
&& \qquad + \int_{[0,t]} \left(\int_{(t-s,\infty)} \psi(x-t+s)d G^s(x)\right)dK(s).
\eeqa
It is clear that $\nu^\ell_t$ has support $[0,H^s)$. Similarly, for each $t\geq 0$, there is a unique regular Borel measure $\eta^\ell_t$ with support $[0,H^r)$ such that
\beqal{mr4}
\int_0^\infty \psi(x)\eta^\ell_t(dx) & = & 1_{\{\varsigma(t)\leq 0\}}\int_{[0,-\varsigma(t)]}\left(\int_{(y+t,\infty)} \frac{\psi(x-y-t)}{\bar G^r(y)}dG^r(x)\right)\eta_0(dy) \non\\ && \qquad + \int_{\varsigma^+(t)}^{t} \left(\int_{(t-s,\infty)} \psi(x-t+s) d G^r(x) \right)\lambda(s) ds.
\eeqa
For each $z\geq 0$, by plugging $\psi(x)=1_{(z,\infty)}(x)$ into (\ref{mr2}) and (\ref{mr4}), we have
\begin{eqnarray*}
\nu^\ell_t(z,\infty) &=& \int_{[0,H^s)}\frac{\bar{G}^s(y+t+z)}{\bar G^s(y)}\nu_0(dy) + \int_{[0,t]} \bar{G}^s(t-s+z)dK(s) \\ \eta^\ell_t(z,\infty) &=& 1_{\{\varsigma(t)\leq 0\}}\int_{[0,-\varsigma(t)]}\frac{\bar{G}^r(y+t+z)}{\bar G^r(y)}\eta_0(dy) + \int_{\varsigma^+(t)}^{t} \bar{G}^r(t-s+z)\lambda(s) ds.
\end{eqnarray*}
Since $\nu$ satisfies (\ref{f5}), then
\begin{eqnarray*}
\nu_{t+z}(z,\infty) =\int_{[0,H^s)}\frac{\bar{G^s}(y+t+z)}{\bar{G^s}(y)}\nu_0(dy)+
\int_{[0,t]}\bar{G^s}(t+z-s) d K(s),
\end{eqnarray*}
and hence
\beql{mrme1}
\nu^\ell_t(z,\infty) = \nu_{t+z}(z,\infty).
\eeq
Intuitively, $\nu_{t+z}(z,\infty)$ represents the amount of fluid content in service at time $t+z$ with elapsed service time greater than $z$, which is precisely the amount of fluid content will still be in service at time $t+z$ with residual service time greater than $0$ and then is equal to the amount of fluid content in service at time $t$ that has residual service time greater than $z$. Thus, $\nu^\ell_t(z,\infty)$ also represents the amount of fluid content in service at time $t$ that has residual service time greater than $z$, that is $\nu^\ell_t$ keeps track of the residual times of customers in service at time $t$. On the other hand, consider the amount of fluid content of customers in queue whose residual patience time lies in the range $(z,\infty)$ at time $t$. This amount of fluid content comes from both the fluid content of customers initially in queue with elapsed times in $[0,\chi(t)-t]$ (if $\chi(t)\geq t$) whose residual time lies in the range $(z+t,\infty)$ and the fluid content of customers that have entered queue during $[\varsigma^+(t),t]$. For the fluid content of customers initially in queue with ages in $(\chi(t)-t,\infty)$ and the fluid content of customers arriving in the time interval $[0, \varsigma^+(t)]$, it has entered service or reneged by time $t$. For the amount of fluid content of customers initially in queue whose elapsed times lie in $[y,y+dy]$ (represented by $\eta_0(dy)$), $y\in [0,\chi(t)-t]$, the fraction of fluid content with residual service times in the range $(z+t,\infty)$ is given by $\bar{G}^s(y+t+z)/\bar G^s(y)$. So it is natural to expect that the amount of fluid content of customers initially in queue with elapsed times in $[0,\chi(t)-t]$ (if $\chi(t)\geq t$) whose residual times lie in the range $(z+t,\infty)$ is given by $1_{\{\chi(t)\geq t\}}\int_{[0,\chi(t)-t]}(\bar{G}^r(y+t+z)/\bar G^r(y))\eta_0(dy)$. Likewise, for the amount of fluid content of customers that have entered service at time $\varsigma^+(t)<s<t$ (represented by $\la(s)$), the fraction of fluid content with residual service time in the range $(z,\infty)$ at time $t$ is given by $\bar{G}^r(t -s + z)$. So $\int_{\varsigma^+(t)}^t \bar{G}^r(t -s + z) \la (s)ds$ represents the fluid content of customers that have entered service during $(\varsigma^+(t),t]$ and whose residual times lie in the range $(z,\infty)$ at time $t$. Thus, by the above representation for $\eta^\ell_t(z,\infty)$, we see that $\eta^\ell_t(z,\infty)$ is precisely the amount of fluid content of customers in queue whose residual patience time lies in the range $(z,\infty)$ at time $t$ and hence $\eta^\ell_t$ keeps track of the residual times of customers in queue at time $t$.
\begin{remark} When $G^r$ and $G^s$ have densities $g^r$ and $g^s$, respectively, \eqref{mr4} and \eqref{mr2} are equivalent to the following representations:
\beqal{mr41}
\int_0^\infty \psi(x)\eta^\ell_t(dx) & = & 1_{\{\varsigma(t)\leq 0\}}\int_{[0,-\varsigma(t)]}\left(\int_0^\infty \frac{g^r(y+t+x)}{\bar G^r(y)}\psi(x)dx\right)\eta_0(dy) \non\\ && \qquad + \int_{\varsigma^+(t)}^{t} \left(\int_0^\infty g^r(t-s+x) \psi(x)dx \right)\lambda(s) ds ,
\eeqa
\beqal{mr21}
\int_0^\infty \psi(x) \nu^\ell_t(dx) &=& \int_{[0,H^s)}\left(\int_0^\infty \frac{g^s(y+t+x)}{\bar G^s(y)}\psi(x)dx\right)\nu_0(dy) \non\\
&& \qquad + \int_{[0,t]} \left(\int_0^\infty g^s(t-s+x)\psi(x)dx\right)dK(s).
\eeqa
In this case, for each $t\geq 0$, the two measures $\eta^\ell_t$ and $\nu^\ell_t$ have
densities $b_\ell(t,x)$ and $q_t(t,x)$, respectively, which can be expressed as
\beql{tr56}
b_\ell(t,y) = \int_{[0,H^s)}\frac{g^s(x+t+y)}{\bar G^s(x)}\nu_0(dx) + \int_{[0,t]} g^s(y+t-u) dK(u),
\eeq
and
\beql{tr567} q_\ell(t,y) = 1_{\{\varsigma(t)\leq 0\}}\int_{[0,-\varsigma(t)]}\frac{g^r(x+t+y)}{\bar G^r(x)}\eta_0(dx)+ \int_{\varsigma^+(t)}^{t} g^r(y+t-u) \lambda (u) du.
\eeq
\end{remark}
\begin{remark} \label{rem:ic}
When $t=0$, \eqref{mr4} and \eqref{mr2} are equivalent to the following representations: for each $z\geq 0$,
\beqal{mr22}
\nu^\ell_0(z,\infty) = \int_{[0,H^s)}\frac{\bar{G}^s(y+z)}{\bar G^s(y)}\nu_0(dy)
\eeqa
\beqal{mr42}
\eta^\ell_0(z,\infty) = \int_{[0,H^r)}\frac{\bar{G}^r(y+z)}{\bar G^r(y)}\eta_0(dy).
\eeqa
\end{remark}
\fi
\iffalse
In the end of this section, we state the following continuity property of all the processes in Definition \ref{def:FMA}.
\begin{lemma}
The processes $\eta$, $\nu$, $K(\cdot)$, $B(\cdot)$, $D(\cdot)$, $R(\cdot)$, $\chi(\cdot)$, $Q(\cdot)$ and $X(\cdot)$ are continuous on $[0,\infty)$.
\end{lemma}
\proof Fix $t\in [0,\infty)$. It is clear that $\nu$, $D(\cdot)$ and $R(\cdot)$ are continuous at $t$ by (\ref{f4}), (\ref{f80}) and (\ref{f8}), respectively. By Theorem 3.5 of \cite{KR10}, it is also clear that $X(\cdot)$, $K(\cdot)$ and $\nu$ are continuous at $t$. Then $B(\cdot)$ and $Q(\cdot)$ are also continuous at $t$. At last, we show that $\chi(\cdot)$ is continuous at $t$.
Let $\{s_n:\ n\geq 1\}\subset [0,\infty)$ such that $s_n\rightarrow t$ as $n\rightarrow \infty$ and $\lim_{n\rightarrow \infty} \chi(s_n)=\liminf_{s\rightarrow t}\chi(s)$. It follows from (\ref{f9}) that $\eta_{s_n}[0,\chi(s_n)]\geq Q(s_n)$ for each $n\geq 1$. By taking the limit as $n\rightarrow \infty$ and the continuity of $\eta$ and $Q$, we have $\eta_{t}[0,\lim_{n\rightarrow \infty}\chi(s_n)]\geq Q(t)$ and then $\liminf_{s\rightarrow t}\chi(s)=\lim_{n\rightarrow \infty}\chi(s_n)\geq \chi(t)$. On the other hand, Let $\{u_n:\ n\geq 1\}\subset [0,\infty)$ such that $u_n\rightarrow t$ as $n\rightarrow \infty$ and $\lim_{n\rightarrow \infty} \chi(u_n)=\limsup_{s\rightarrow t}\chi(s)$. It follows from (\ref{f9}) that $\eta_{u_n}[0,\chi(u_n)-\delta]< Q(s_n)$ for each $n\geq 1$ and $\delta>0$. By taking the limit as $n\rightarrow \infty$ and the continuity of $\eta$ and $Q$, we have $\eta_{t}[0,\lim_{n\rightarrow \infty}\chi(u_n)-\delta]\leq Q(t)$.
\fi
\section{Measure-valued fluid models tracking residual times} \label{secFluidR}
We first state the two measure-valued processes tracking residual times that arise from Zu{\~n}iga's fluid model for the same $G_t/GI/N+GI$ queueing system in Section \ref{secMR}. We then state Zhang's fluid model in Section \ref{secZhang}, and discuss its connection with the three fluid models tracking elapsed times in Section \ref{secZhanglimitation}. The two measure-valued processes tracking residual times introduced in Section \ref{secMR} play an important bridging role in making the connection.
\subsection{Measure-valued processes tracking residual times from Zu{\~n}iga's fluid model}
\label{secMR}
\iffalse
Let $\nu^\ell_t$ be a nonnegative finite measure on $[0,\infty)$ with support in $[0,H^s)$ such that $\nu^\ell_t(dx)$ represents the amount of fluid content of customers in service whose residual service time at time $t$ lies in the range $[x,x+dx)$. Similarly, let $\eta^\ell_t$ be a nonnegative finite measure on $[0,\infty)$ with support in $[0,H^r)$ such that $ \eta^\ell_t(dx)$ represents the amount of fluid content of customers in queue whose residual patience time at time $t$ lies in the range $[x,x+dx)$.
\fi
Zu{\~n}iga's fluid model naturally give rise to the following two measure-valued processes $\nu^\ell_t$ and
$\eta^\ell_t$.
For each $t\geq 0$, clearly the mapping
\begin{eqnarray*}\psi &\mapsto & \int_{[0,H^s)}\left(\int_{(y+t,\infty)} \frac{\psi(x-y-t)}{\bar G^s(y)}d G^s(x)\right)\nu_0(dy) \\
&& \qquad + \int_{[0,t]} \left(\int_{(t-s,\infty)} \psi(x-t+s)d G^s(x)\right)dK(s)\end{eqnarray*} is a positive linear functional on $\mathcal{C}_c(\RR_+)$ since $K$ is non-decreasing by Lemma \ref{lem:chiinc}. Then by Riesz-Markov-Kakutani representation theorem, there is a unique regular Borel measure $\nu^\ell_t$ with support $[0,H^s)$ such that for every $\psi \in \mathcal{C}_b(\RR_+)$,
\beqal{mr2}
\int_0^\infty \psi(x) \nu^\ell_t(dx) &=& \int_{[0,H^s)}\left(\int_{(y+t,\infty)} \frac{\psi(x-y-t)}{\bar G^s(y)}d G^s(x)\right)\nu_0(dy) \non\\
&& \qquad + \int_0^{t} \left(\int_{(t-s,\infty)} \psi(x-t+s)d G^s(x)\right)dK(s).
\eeqa
Similarly, for each $t\geq 0$, there is a unique regular Borel measure $\eta^\ell_t$ with support $[0,H^r)$ such that for every $\psi \in \mathcal{C}_b(\RR_+)$,
\beqal{mr4}
\int_0^\infty \psi(x)\eta^\ell_t(dx) & = & \bone_{\{\varsigma(t)\leq 0\}}\int_{[0,-\varsigma(t)]}\left(\int_{(y+t,\infty)} \frac{\psi(x-y-t)}{\bar G^r(y)}dG^r(x)\right)\eta_0(dy) \non\\ && \qquad + \int_{\varsigma^+(t)}^{t} \left(\int_{(t-s,\infty)} \psi(x-t+s) d G^r(x) \right)\lambda(s) ds ,
\eeqa
where \beql{chiw}
\varsigma(t) = t - \chi(t).
\eeq
Since $\chi(t)$ represents the elapsed patience time of the fluid content of customers that has been in queue the longest at time $t$, then the quantity $\varsigma(t)$ can be interpreted as the arrival time of the fluid content of customers that has been in queue the longest at time $t$. It is clear that $ \varsigma(t)\leq t$ for each $t\ge 0$.
At time 0, $\varsigma(0) = - \chi(0)$ represents the arrival time of the oldest fluid content in queue initially, and thus, it follow from \eqref{f9} that
\beql{mrvs1}
\varsigma(0)=-\inf\{x\in [0,H^r):\ \eta_0[0,x) \geq X(0)-\nu_0[0,H^s)\} .
\eeq
We first argue that $\nu^\ell$ and $\eta^\ell$ are two measure-valued processes tracking residual times of fluid content of customers in service and in queue, respectively.
\iffalse
So it is clear that
\beql{mrvs2}
\varsigma(t)\leq t \quad \text{for each} \ t \ge 0,
\eeq
and intuitively, we see that $\varsigma(t)$ can be related to $\chi(t)$ as in \eqref{chiw}.
At time 0, $\varsigma(0) = - \chi(0)$ represents the arrival time of the oldest fluid content in queue initially, and thus, it follow from \eqref{f9} that
\beql{mrvs1}
\varsigma(0)=-\inf\{x\in [0,H^r):\ \eta_0[0,x) \geq X(0)-\nu_0[0,H^s)\} .
\eeq
If $\varsigma(t) <0$ at time $t> 0$, it means that the initial fluid content of customers at time 0 has not all entered service yet by time $t$, and can be interpreted as the arrival time of the initial fluid content that has been in queue the longest at time $t$.
\fi
For each $z\geq 0$, by plugging $\psi(x)=\bone_{(z,\infty)}(x)$ into \eqref{mr2} and \eqref{mr4}, we have
\begin{eqnarray}
\nu^\ell_t(z,\infty) &=& \int_{[0,H^s)}\frac{\bar{G}^s(y+t+z)}{\bar G^s(y)}\nu_0(dy) \label{nuz} \\
&& \qquad \qquad + \int_{[0,t]} \bar{G}^s(t-s+z)dK(s),\non \\
\eta^\ell_t(z,\infty) &=& \bone_{\{\varsigma(t)\leq 0\}}\int_{[0,-\varsigma(t)]}\frac{\bar{G}^r(y+t+z)}{\bar G^r(y)}\eta_0(dy) \label{etaz}\\
&& \qquad \qquad+ \int_{\varsigma^+(t)}^{t} \bar{G}^r(t-s+z)\lambda(s) ds. \non
\end{eqnarray}
By \eqref{f4} and \eqref{f5}, we obtain
\begin{eqnarray*}
\eta_{t+z}[z,\chi(t)+z] &=& \int_{[0,H^r)}\bone_{[z,\chi(t)+z]}(y+t+z)\frac{\bar{G^r}(y+t+z)}{\bar{G^r}(y)}\nu_0(dy) \\ & & \quad +
\int_0^t\bone_{[z,\chi(t)+z]}(t+z-s)\bar{G^r}(t+z-s) \la(s) ds \\ &=& \bone_{\{\varsigma(t)\leq 0\}}\int_{[0,-\varsigma(t)]}\frac{\bar{G}^r(y+t+z)}{\bar G^r(y)}\eta_0(dy) \\
&& \quad + \int_{\varsigma^+(t)}^{t} \bar{G}^r(t-s+z)\lambda(s) ds ,
\end{eqnarray*}
and
\begin{eqnarray*}
\nu_{t+z}[z,\infty) =\int_{[0,H^s)}\frac{\bar{G^s}(y+t+z)}{\bar{G^s}(y)}\nu_0(dy)+
\int_{[0,t]}\bar{G^s}(t+z-s) d K(s).
\end{eqnarray*}
Hence,
\beql{mrme1}
\nu^\ell_t(z,\infty) = \nu_{t+z}[z,\infty) \mbox{ and } \eta^\ell_t(z,\infty) = \eta_{t+z}[z,\chi(t)+z].
\eeq
Intuitively, $\nu_{t+z}[z,\infty)$ represents the amount of fluid content in service at time $t+z$ with elapsed service time at least $z$, which is precisely the amount of fluid content in service at time $t$ that will still be in service at time $t+z$ and then is equal to the amount of fluid content in service at time $t$ that has residual service time greater than $z$. (Note that the fluid content in service at time $t$ that has residual service time exactly equal to $z$ will depart from service and hence will not be in service at time $t+z$.) Thus, by the first equality in (\ref{mrme1}), $\nu^\ell_t(z,\infty)$ represents the amount of fluid content in service at time $t$ that has residual service time greater than $z$, that is, $\nu^\ell_t$ keeps track of the residual time of fluid content in service at time $t$. Similarly, $\eta_{t+z}[z,\chi(t)+z]$ represents the amount of fluid content in the potential queue at time $t+z$ with elapsed patience time between $z$ and $\chi(t)+z$, which is precisely the amount of fluid content in queue at time $t$ that will not abandon by time $t+z$. This amount of fluid content is equal to the amount of fluid content that has residual patience time more than $z$ units of time at time $t$ and then is represented by $\eta^\ell_t(z,\infty)$ by the second equality in (\ref{mrme1}). Then $\eta^\ell_t$ keeps track of the residual patience times of customers in queue at time $t$.
\iffalse
\footnote{need more thinking} On the other hand, consider the amount of fluid content of customers in queue whose residual patience time lies in the range $(z,\infty)$ at time $t$. This amount of fluid content comes from both the fluid content of customers initially in queue with elapsed times in $[0,\chi(t)-t]$ (if $\chi(t)\geq t$) whose residual time lies in the range $(z+t,\infty)$ and the fluid content of customers that have entered queue during $[\varsigma^+(t),t]$. For the fluid content of customers initially in queue with ages in $(\chi(t)-t,\infty)$ and the fluid content of customers arriving in the time interval $[0, \varsigma^+(t)]$, it has entered service or reneged by time $t$. For the amount of fluid content of customers initially in queue whose elapsed times lie in $[y,y+dy]$ (represented by $\eta_0(dy)$), $y\in [0,\chi(t)-t]$, the fraction of fluid content with residual service times in the range $(z+t,\infty)$ is given by $\bar{G}^s(y+t+z)/\bar G^s(y)$. So it is natural to expect that the amount of fluid content of customers initially in queue with elapsed times in $[0,\chi(t)-t]$ (if $\chi(t)\geq t$) whose residual times lie in the range $(z+t,\infty)$ is given by $1_{\{\chi(t)\geq t\}}\int_{[0,\chi(t)-t]}(\bar{G}^r(y+t+z)/\bar G^r(y))\eta_0(dy)$. Likewise, for the amount of fluid content of customers that have entered service at time $\varsigma^+(t)<s<t$ (represented by $\la(s)$), the fraction of fluid content with residual service time in the range $(z,\infty)$ at time $t$ is given by $\bar{G}^r(t -s + z)$. So $\int_{\varsigma^+(t)}^t \bar{G}^r(t -s + z) \la (s)ds$ represents the fluid content of customers that have entered service during $(\varsigma^+(t),t]$ and whose residual times lie in the range $(z,\infty)$ at time $t$. Thus, by the above representation for $\eta^\ell_t(z,\infty)$, we see that $\eta^\ell_t(z,\infty)$ is precisely the amount of fluid content of customers in queue whose residual patience time lies in the range $(z,\infty)$ at time $t$ and hence $\eta^\ell_t$ keeps track of the residual times of customers in queue at time $t$.
\fi
The relationship between $(\nu, \eta)$ and $(\nu^\ell, \eta^\ell)$ in \eqref{mrme1} is called a {\it coupling} property between the measure-valued processes tracking elapsed and residual times.
When $t=0$, \eqref{nuz} and \eqref{etaz} become: for each $z\geq 0$,
\beqal{mr22}
\nu^\ell_0(z,\infty) = \int_{[0,H^s)}\frac{\bar{G}^s(y+z)}{\bar G^s(y)}\nu_0(dy) = \nu_z [z,\infty),
\eeqa
\beql{mr42}
\eta^\ell_0(z,\infty) = \bone_{\{\varsigma(0)\leq 0\}}\int_{[0,-\varsigma(0)]}\frac{\bar{G}^r(y+z)}{\bar G^r(y)}\eta_0(dy) = \eta_z[z,\chi(0)+z].
\eeq
\begin{remark}
When $G^r$ and $G^s$ have densities $g^r$ and $g^s$, respectively, \eqref{mr2} and \eqref{mr4} are equivalent to the following representations:
\beqal{mr21}
\int_0^\infty \psi(x) \nu^\ell_t(dx) &=& \int_{[0,H^s)}\left(\int_0^\infty \frac{g^s(y+t+x)}{\bar G^s(y)}\psi(x)dx\right)\nu_0(dy) \non\\
&& \qquad + \int_{[0,t]} \left(\int_0^\infty g^s(t-s+x)\psi(x)dx\right)dK(s),
\eeqa
\beqal{mr41}
\int_0^\infty \psi(x)\eta^\ell_t(dx) & = & \bone_{\{\varsigma(t)\leq 0\}}\int_{[0,-\varsigma(t)]}\left(\int_0^\infty \frac{g^r(y+t+x)}{\bar G^r(y)}\psi(x)dx\right)\eta_0(dy) \non\\ && \qquad + \int_{\varsigma^+(t)}^{t} \left(\int_0^\infty g^r(t-s+x) \psi(x)dx \right)\lambda(s) ds.
\eeqa
In this case, for each $t\geq 0$, the two measures $\eta^\ell_t$ and $\nu^\ell_t$ have
densities $b_\ell(t,x)$ and $q_t(t,x)$, respectively, which can be expressed as
\beql{tr56}
b_\ell(t,y) = \int_{[0,H^s)}\frac{g^s(x+t+y)}{\bar G^s(x)}\nu_0(dx) + \int_{[0,t]} g^s(y+t-u) dK(u),
\eeq
and
\beql{tr567} q_\ell(t,y) = \bone_{\{\varsigma(t)\leq 0\}}\int_{[0,-\varsigma(t)]}\frac{g^r(x+t+y)}{\bar G^r(x)}\eta_0(dx)+ \int_{\varsigma^+(t)}^{t} g^r(y+t-u) \lambda (u) du.
\eeq
\iffalse
In addition to $(\nu^\ell,\eta^\ell)$, we also introduce a measure-valued process $\tilde{\eta}^\ell$ satisfying that for every $\psi \in \mathcal{C}_b(\RR_+)$,
\beqal{mr62}
\int_0^\infty \psi(x) \tilde{\eta}^\ell_t(dx) &=& \int_{[0,H^r)}\left(\int_0^\infty \frac{g^r(y+t+x)}{\bar G^r(y)}\psi(x)d x\right)\eta_0(dy) \non\\
&& \qquad + \int_0^{t} \left(\int_0^\infty g^r(t-u+x)\psi(x)dx\right)\lambda(u) du.
\eeqa
For each $t\geq 0$ and $x\geq 0$, $\tilde{\eta}^\ell_t(dx)$ represents the amount of fluid content of customers in the potential queue whose residual patience times at time $t$ lie in the range $[x,x+dx)$.
For each $t\geq 0$, the measure $\tilde \eta^\ell_t$ admits a density function $\tilde{q}_\ell(t,y)$ with the representation:
\beql{mr64}
\tilde{q}_\ell(t,y)=\int_{[0,H^r)}\frac{g^r(x+t+y)}{\bar G^r(x)}\eta_0(dx)+ \int_0^{t} g^r(y+t-s) \lambda(s) ds.
\eeq
\fi
\end{remark}
\iffalse
The pair of processes $(\nu^\ell,\eta^\ell)$ in \eqref{mr2} and \eqref{mr4} with the input data
$(\lambda(\cdot), \eta_0,\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))\in \mathcal S_0$, together with the auxiliary processes $Q(t), R(t), D(t), B(t)$ in Definition \ref{def:EFMA}, {\color{red} in fact this is not precise because in those equations $\nu_t$ and $\eta_t$ are used; we may not want to formally say it is a new model....} is called {\it a measure-valued fluid model tracking residual times}, resulting directly from Kang-Ramanan and Zu{\~n}iga's fluid models. The existence and uniqueness of the pair of processes $(\nu^\ell,\eta^\ell)$ follow directly from those of Zu{\~n}iga's fluid model in \cite{Zu14},
under the assumptions that $G^r$ is continuous, $\eta_0$ is diffuse, and $\nu_0$ is diffuse if $G^s$ is not continuous (Assumption 3.1 of \cite{Zu14}).
\begin{remark}
The initial condition in the input data to the measure-valued processes $(\nu^\ell, \eta^\ell)$ is $(\eta_0,\nu_0,X(0))$ rather than the associated residual initial quantities $(\nu_0^\ell,\eta_0^\ell)$. This is because $(\nu_0^\ell,\eta_0^\ell)$ only contains information on the residual patience time of fluid content of customers in queue and the residual service time of fluid content of customers in service, but this information is insufficient to determine the service order of the initial fluid content of customers in queue since we do not know the arrival order of the fluid content from $(\nu_0^\ell,\eta_0^\ell)$ and the service discipline is FCFS. On the other hand, the information encoded in $(\eta_0,\nu_0,X(0))$ contains the elapsed time $($and hence the arrival order$)$ of the fluid content of customers in queue.
Thus, it is sufficient to use $(\eta_0,\nu_0,X(0))$ together with $\lambda(\cdot)$ as the input data for
the measure-valued processes $(\nu^\ell, \eta^\ell)$.
\end{remark}
We remark that the processes $(\nu, \eta)$ tracking elapsed times in Definition \ref{def:FMA} can also be directly obtained from the pair of processes $(\nu^\ell,\eta^\ell)$ in \eqref{mr2} and \eqref{mr4}.
For each $t\geq 0$, since $K$ is non-decreasing Lemma \ref{lem:chiinc}, the following mapping
\begin{eqnarray*}\psi &\mapsto & \int_{[0,H^s)}\psi(x+t)\frac{\bar{G^s}(x+t)}{\bar{G^s}(x)}\nu_0(dx)+
\int_{[0,t]} \psi(t-s) \bar{G^s}(t-s) d K(s)\end{eqnarray*} is a positive linear functional on $\mathcal{C}_c(\RR_+)$. Then by Riesz-Markov-Kakutani representation theorem, there is a unique regular Borel measure $\nu_t$ on $\RR_+$ such that (\ref{f5}) holds. It is clear that $\nu_t$ has support $[0,H^s)$. Similarly, for each $t\geq 0$, there is a unique regular Borel measure $\eta_t$ that satisfies (\ref{f4}). This shows that the existence and uniqueness of the fluid processes tracking residual times are equivalent to those of Zu{\~n}iga's fluid model in \cite{Zu14}.
\fi
\subsection{Zhang's fluid model} \label{secZhang}
Zhang \cite{Zhang} uses a so-called {\it virtual queue} to describe the queueing dynamics, instead of the potential queue used in the three fluid models in Section \ref{secFluidE}.
In the definitions of both potential and virtual queues, all customers enter them upon arrival.
{\it The difference between them lies in how customers depart.}
Customers can leave the potential queue only when their patience expires, that is, at the instant when their remaining patience times are zeros.
Whereas, customers can only leave the virtual queue in their turns of service.
Customers in the virtual queue may have already run out of patience (i.e., the remaining patience time is negative) at their turns of service. Whenever a server becomes free, the server will check the oldest customer in the virtual queue. If the customer being checked has not abandoned yet (its remaining patience time is still positive), then the server will start serving this customer and this customer is removed from the virtual queue, and otherwise, this customer is simply removed from the virtual queue and the server will turn to check the next oldest customer.
We now state Zhang's fluid model.
\begin{definition} \label{def:FMZ}
$($Zhang's fluid model in \cite{Zhang}.$)$
Assume that the fluid arrival rate $\lambda(t) = \la $ for each $t\ge 0$, where $\lambda> 0$ is a constant. A pair of measure-valued processes $(\mathcal{R}, \mathcal{Z})$ is a solution to the fluid model if the following conditions are satisfied:
$(i)$ $(\mathcal{R}, \mathcal{Z})$ satisfies the following two equations:
\beql{r1}
\mathcal{R}_t(C_x) = \la \int_{t- Q_v(t)/\la}^t \bar{G}^r(t+x-s) ds, \quad x \in \RR,
\eeq
and
\beql{r2}
\mathcal{Z}_t(C_x) = \mathcal{Z}_0(C_x+t) + \int_0^t \bar{G}^r(Q_v(s)/\la) \bar{G}^s(t+x-s) d L_v(s), \quad x \in \RR_+,
\eeq where $C_x \doteq (x,\infty)$ for $x \in \RR$, $Q_v(t) = \mathcal{R}_t(\RR)$ is of bounded variation and $L_v(t) = \lambda t - Q_v(t)$;
$(ii)$ the non-idling conditions in \eqref{f2} and \eqref{f12} hold for $B(t)=\mathcal{Z}_t(\RR_+)$, $Q(t)=\mathcal{R}_t(\RR_+)$ and $X(t)=B(t)+Q(t)$;
$(iii)$ the initial condition $(\mathcal{R}_0,\mathcal{Z}_0)$ satisfies
\beql{r3}
\mathcal{R}_0(C_x) = \la \int_0^{ Q_v(0)/\la}\bar{G}^r(x+s) ds, \ x \in \RR, \mbox{ and }\mathcal{Z}_0(\{0\}) = 0,
\eeq
and the non-idling condition at time $0$ in \eqref{f2} and \eqref{f12}.
\end{definition}
\vspace{0.1in}
In Zhang's fluid model, $\mathcal{R}_t(C_x)$ can be interpreted as the fluid content of customers in the virtual queue with residual patience times strictly bigger than $x$ and $\mathcal{Z}_t(C_x)$ can be interpreted as the fluid content of customers in service with residual service times strictly bigger than $x$ at each time $t$.
Then $Q(t)$, $B(t)$, $Q_v(t)$ and $L_v(t)$ represent, respectively, the total fluid content of the real queue at time $t$, the total fluid content of customers in service at time $t$, the total fluid content in the virtual queue at time $t$, and the cumulative customers removed from the virtual queue by time $t$.
The existence and uniqueness of Zhang's fluid model are proved in Theorem 3.1 of \cite{Zhang} by an explicit characterization of its solution,
under the assumptions that the service time distribution $G^s$ is continuous and the patience time distribution $G^r$ is Lipschitz continuous.
\subsection{Connection between Zhang's fluid model and the three fluid models in Section \ref{secFluidE}} \label{secZhanglimitation}
Among the three fluid models in Section \ref{secFluidE}, we have showed in Theorem \ref{wellposed} that Whitt's fluid model is equivalent to Kang-Ramanan fluid model and in Remark \ref{rem:new} that Zu{\~n}iga's fluid model extends Kang-Ramanan's fluid model by relaxing the assumption on the existence of densities of $G^r$ and $G^s$. Since Zhang's fluid model keeps track of customers' residual times and does not need $G^r$ and $G^s$ to have densities (\cite{Zhang} does assume that $G^s$ is continuous and $G^r$ is Lipschitz continuous to establish existence and uniqueness), while Zu{\~n}iga's fluid model keeps track of customers' elapsed times and also does not need $G^r$ and $G^s$ to have densities, it is natural to question if Zhang's fluid model and Zu{\~n}iga's fluid model are in fact equivalent in describing system dynamics of the same $G_t/GI/N+GI$ queues. If so, this will enable researchers to borrow results from either one of the two to study the system performance of $G_t/GI/N+GI$ queues.
In this section we provide a detailed discussion on Zhang's fluid model in connection with Zu{\~n}iga's fluid model (and hence Kang-Ramanan's fluid model and Whitt's fluid model).
The three fluid models in Section \ref{secFluidE} allow time-varying arrival rate $\lambda(\cdot)$, whereas, Zhang's fluid model requires a constant arrival rate $\lambda$. Thus the discussion in this section will focus on the four formulations with a constant arrival rate.
We first show by a series of remarks that {\it Zhang's fluid model is not equivalent to the three fluid models tracking elapsed times for the same $G/GI/N+GI$ queueing system under general initial conditions, that is, Zhang's fluid model and the three fluid models tracking elapsed times may not be formulated simultaneously for the same $G/GI/N+GI$ queueing system under general initial conditions.}
\begin{remark} $($On the arrival rate.$)$ \label{remk:la}
The imposed condition on $\mathcal{R}_0$ in Zhang's fluid model requires that the initial fluid content of customers in the virtual queue depends on the arrival rate $\la$ after time $0$. This condition is rather restrictive since in real life applications, the customers' arrival patterns before time $0$ and after time $0$ are likely different.
Thus, Zhang's fluid model cannot be used for those applications.
In contrast, the three fluid models tracking elapsed times do not have this restriction.
\end{remark}
\begin{remark}\label{remk:z} $($The initial condition on $\mathcal{R}_0$.$)$
Zhang's fluid model requires that the system initial condition $\mathcal{R}_0$ satisfies \eqref{r3}, that is,
\beql{zR0x}
\mathcal{R}_0(C_x) = \la \int_0^{ Q_v(0)/\la}\bar{G}^r(x+s) ds, \ x \in \RR.
\eeq
Let $\mathcal{R}_0^+$ be the restriction of $\mathcal{R}_0$ on $[0,\infty)$.
Then $\mathcal{R}_0^+$ keeps track of the residual patience times of the fluid content of customers initially in queue. Thus, $\mathcal{R}_0^+ = \eta^\ell_0$ in \eqref{mr42}, that is,
$$
\mathcal{R}_0^+(C_x)=\bone_{\{\varsigma(0)\leq 0\}}\int_{[0,-\varsigma(0)]}\frac{\bar{G}^r(y+x)}{\bar G^r(y)}\eta_0(dy), \quad \forall x\geq 0,
$$
where $\eta_0$ is the initial condition for the $\eta$ in Definition \ref{def:EFMA}.
So if Zhang's fluid model were equivalent to Zu{\~n}iga's fluid model for the same $G/GI/N+GI$ queueing system assuming a constant arrival rate, we must have
\beql{DisLimR0}
\la \int_0^{ Q_v(0)/\la}\bar{G}^r(x+s) ds =\bone_{\{\varsigma(0)\leq 0\}}\int_{[0,-\varsigma(0)]}\frac{\bar{G}^r(y+x)}{\bar G^r(y)}\eta_0(dy),\quad \forall x\geq 0.
\eeq
We first note that there may not be a unique $\eta_0$ satisfying \eqref{DisLimR0} for the given $\mathcal{R}_0$. For example, when $G^r$ has density $g^r(x)= e^{-x},\ x\in \RR_+$,
$$
\bone_{\{\varsigma(0)\leq 0\}}\int_{[0,-\varsigma(0)]}\frac{\bar{G}^r(y+x)}{\bar G^r(y)}\eta_0(dy)
= \bone_{\{\varsigma(0)\leq 0\}} e^{-x}\eta_0[0,-\varsigma(0)],
$$
and
$$\la \int_0^{ Q_v(0)/\la}\bar{G}^r(x+s) ds = \la e^{-x} \left(1-e^{-Q_v(0)/\la} \right).
$$ Thus, any $\eta_0$ satisfying $ \bone_{\{\varsigma(0)\leq 0\}}\eta_0[0,-\varsigma(0)]= \la (1-e^{-Q_v(0)/\la})$ will satisfy \eqref{DisLimR0}.
Moreover, it is clear that the above display \eqref{DisLimR0} does not hold for an arbitrary initial condition $\eta_0$.
For example, if $\eta_0(dx)= \lambda^\dag \bar G^r(x)dx$ for some positive $\lambda^\dag \neq \lambda$, then
\begin{eqnarray*}
&& \bone_{\{\varsigma(0)\leq 0\}}\int_{[0,-\varsigma(0)]}\frac{\bar{G}^r(y+x)}{\bar G^r(y)}\eta_0(dy)
\\
&=& \lambda^\dag \bone_{\{\varsigma(0)\leq 0\}}\int_{[0,-\varsigma(0)]} \bar{G}^r(y+x)dy ,
\end{eqnarray*}
which is not equal to $\mathcal{R}_0^+(C_x)$ in \eqref{zR0x} even if $-\varsigma(0) = Q_v(0)/\la$.
Thus, for a fluid $G/GI/N+GI$ queueing system with a constant arrival rate $\lambda$ after time $0$, the initial conditions $\eta_0(dx)= \lambda^\dag \bar G^r(x)dx$ for $\lambda^\dag \neq \lambda$ and $(\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))\in \mathcal S_0$, Zu{\~n}iga's fluid model can be well formulated, but there is no corresponding Zhang's fluid model $(\mathcal{R}, \mathcal{Z})$ that describes the same system.
\end{remark}
\begin{remark} \label{remark:ari} $($The initial condition on $\mathcal{Z}_0$.$)$
Zhang's fluid model only requires that $\mathcal{Z}_0(\{0\}) = 0$. This condition is rather general. We show by an example that for a $G/GI/N+GI$ queueing system, although Zhang's fluid model can be formulated with that initial condition $\mathcal{Z}_0$, there may not exist an (unique) initial measure $\nu_0$ to formulate a corresponding Zu{\~n}iga's fluid model for the same system.
Consider the service time distribution $G^s$ being exponential with unit rate, that is, $g^s(x)= e^{-x},\ x\in \RR_+$. Let $\mathcal{Z}_0$ be the measure that tracks the residual service times of fluid content of customers initially in service and satisfies $\mathcal{Z}_0(\{0\}) = 0$, and assume that Zhang's fluid model can be formulated with $\mathcal{Z}_0$. Suppose that Zu{\~n}iga's fluid model can also be formulated for some measure $\nu_0$, which tracks the elapsed service times of fluid content of customers initially in service.
By \eqref{mr22},
if Zhang's fluid model and Zu{\~n}iga's fluid model were equivalent, $\mathcal{Z}_0$ and $\nu_0$ must satisfy the following equation:
\[ \mathcal{Z}_0(C_x) = \int_{[0,H^s)}\frac{\bar G^s(y+x)}{\bar G^s(y)}\nu_0(dy)= \nu_0[0,H^s) e^{-x}, \quad x\geq 0. \]
If the given $\mathcal{Z}_0$ satisfies $ \mathcal{Z}_0(C_x) = c e^{-x}$ for some constant $c>0$, then any such measure $\nu_0$ in Zu{\~n}iga's fluid model satisfying $c =\nu_0[0,H^s)$ will satisfy the above display. However, on the other hand, if the given $\mathcal{Z}_0$, satisfying $\mathcal{Z}_0(\{0\}) = 0$, does not have an exponential density, then this contradicts the above equation resulting from the equivalence, and implies that no corresponding measure $\nu_0$ can be found for Zu{\~n}iga's fluid model to be well formulated for the given queueing system.
\end{remark}
From the discussion in Remarks \ref{remk:la}, \ref{remk:z} and \ref{remark:ari}, it is clear that the class of fluid many-server queueing systems where Zhang's fluid model can be formulated is \emph{not} the same as the class of fluid many-server queueing systems where Zu{\~n}iga's fluid model (and hence Kang-Ramanan's fluid model and Whitt's fluid model) can be formulated.
We next look more closely into the conditions on fluid $G/GI/N+GI$ queueing systems where Zhang's fluid model and the three fluid models tracking elapsed times can all be used to describe the system dynamics. To simplify the exposition, we focus on Zhang's fluid model and Zu{\~n}iga's fluid model.
Our findings are stated in the following two theorems.
\begin{theorem} \label{lem:ZKR}
Given a Zhang's fluid model $(\mathcal{R}, \mathcal{Z})$ for a $G/GI/N+GI$ queueing system with arrival rate $\la$, there exists a Zu{\~n}iga's fluid model $(\eta,\nu,X)$ for the same queueing system with the input data $(\la,\eta_0,\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))\in \mathcal S_0$ with
\begin{equation} \label{con4}
\eta_0(dx) \doteq \la \bone_{[0,Q_v(0)/\la]}(x) \bar{G}^r(x)dx ,
\end{equation}
if and only if, for the given $\mathcal{Z}_0$, $\nu_0$ satisfying \begin{eqnarray}
\mathcal{Z}_0(C_x) = \int_{[0,H^s)}\frac{\bar{G}^s(y+x)}{\bar G^s(y)}\nu_0(dy), \ x\geq 0. \label{con2}\end{eqnarray}
\end{theorem}
\noindent {\it Proof.} The ``only if" part follows directly from the discussion in Remark \ref{remark:ari}. We now focus on ``if" part.
Let $\eta_0$ be as given in (\ref{con4}). For each $t\geq 0$, the following mapping
\begin{eqnarray*}\psi &\mapsto & \int_{[0,H^r)}\psi(x+t)\frac{\bar{G^r}(x+t)}{\bar{G^r}(x)}\eta_0(dx)+
\la \int_0^t \psi(t-s) \bar{G^r}(t-s) ds\end{eqnarray*} is a positive linear functional on $\mathcal{C}_c(\RR_+)$. Then by Riesz-Markov-Kakutani representation theorem, there is a unique regular Borel measure $\eta_t$ on $\RR_+$ such that (\ref{f4}) holds. It is clear that $\eta_t$ has support $[0,H^r)$.
For each $t\geq 0$, define
\[K(t)\doteq \int_0^t \bar{G}^r(Q_v(s)/\la) d L_v(s)\mbox{ and } R(t)\doteq \lambda
\int_0^t G^r(Q_v(s)/\la)\,ds. \]
Then, for each $t\geq 0$, with the above $K$ and the given $\nu_0$ satisfying (\ref{con2}), the mapping
\begin{eqnarray*}\psi &\mapsto & \int_{[0,H^s)}\psi(x+t)\frac{\bar{G^s}(x+t)}{\bar{G^s}(x)}\nu_0(dx)+
\int_{[0,t]} \psi(t-s) \bar{G^s}(t-s) d K(s) \end{eqnarray*} is a positive linear functional on $\mathcal{C}_c(\RR_+)$. By Riesz-Markov-Kakutani representation theorem, there is a unique regular Borel measure $\nu_t$ that satisfies (\ref{f5}). Let $B,\ Q,\ X$ be the associated processes in Zhang's fluid model and for each $t\geq 0$, define \[D(t)\doteq \int_{[0,H^s)} \frac{G^s(y+t)-G^s(y)}{\bar G^s(y)} \nu_0(dy) +
\int_0^t G^s(t-s) dK(s).\] We show that $(\eta,\nu,X)$ satisfies Definition \ref{def:EFMA}.
From (\ref{r1}), it is clear that
\beql{qqv}
Q(t)= \mathcal{R}_t(\RR_+) = \la \int_{t-Q_v(t)/\la}^t \bar{G}^r(t-s) ds =\la G^r_d(Q_v(t)/\lambda),\eeq where $G^r_d(x)=\int_0^x \bar G^r(s)ds$.
It is established in the proof of Theorem 3.1 of \cite{Zhang} that $Q(t)/\lambda < G^r_d(\infty)= G^r_d(H^r)$. Then it follows that $Q_v(t)/\la <H^r$. Since $Q_v$ is of bounded variation by (\ref{r1}), it follows that $Q$ is also of bounded variation and by the chain rule formula (Proposition 4.6 in Chapter 0 of \cite{RM}) \[Q(t) = Q(0)+ \int_0^t \bar G^r(Q_v(s)/\lambda)dQ_v(s).\]
Thus, by the definition of $K$ and the above display for $Q(t)$,\begin{eqnarray*} K(t) &=& Q(0)-\left(Q(t)- \la\int_0^t \bar G^r(Q_v(s)/\lambda) ds\right) \\ &=& Q(0)-\left(Q(t)- \la\int_0^t \bar G^r((G^r_d)^{-1}(Q(s)/\lambda)) ds\right). \end{eqnarray*}
Then it follows from Lemma A.3 of \cite{Zhang} that $K$ is non-decreasing.
Simple calculation also shows that \begin{eqnarray*} & & Q(t)+K(t)+R(t) \\ &=& Q(0)+ \int_0^t \bar G^r(Q_v(s)/\lambda)dQ_v(s) + \int_0^t \bar{G}^r(Q_v(s)/\lambda) d (\lambda s -Q_v(s)) \\ & & \qquad + \lambda
\int_0^t G^r(Q_v(s)/\la)\,ds \\ &=& Q(0)+\lambda t, \end{eqnarray*}
which establishes (\ref{31}).
For each $t\geq 0$, define $\chi(t)$ by the right hand side of (\ref{f9}). It follows from the construction of $\eta_t$ and the given $\eta_0$ in \eqref{con4} that
\begin{eqnarray} \label{QQv1}
Q(t) &=& \eta_t[0,\chi(t)] = \la \int_0^{[\chi(t)-t]^+\wedge Q_v(0)/\la}\bar{G^r}(x+t)dx \non\\
&& \qquad \qquad \qquad + \la
\int_{[t-\chi(t)]^+}^t \bar{G^r}(t-s) ds.
\end{eqnarray}
When $\chi(t)>t$, the above display is reduced to
$$
Q(t)/\la = \int_0^{\chi(t)\wedge (t+Q_v(0)/\la)} \bar{G}^r(s)ds.
$$
Comparing this with (\ref{qqv}), we have $Q_v(t)/\la = \chi(t)\wedge (t+Q_v(0)/\la).
$
When $\chi(t)\leq t$, the display in \eqref{QQv1} is reduced to $Q(t)/\la = \int_0^{\chi(t)} \bar{G}^r(s)ds$ and hence $Q_v(t)/\la = \chi(t)$. Combining the two cases, we have for each $t\geq 0$,
\beql{Qvchi1}
Q_v(t)/\la = \chi(t)\wedge (t+Q_v(0)/\la).
\eeq
For each $t\geq 0$, it follows from \eqref{con4} and the definition of $M^r$ that
\begin{eqnarray*}
&& \int_{[0, H^r]} \int_{[0,t]} \bone_{\{x \le \chi(s) \}} d_s \left( \int_{[(x-s)^+, x)} \frac{\bar{G}^r(x-)}{\bar{G}^r(y)} \eta_0(dy ) \right) d M^r(x) \\ & & \quad +\la \int_{[0, H^r]} \int_{[0,t]} \bone_{\{x \le \chi(s) \wedge s\}} \bar{G}^r(x-) ds d M^r(x) \\
&=& \la \int_{[0, H^r]} \int_{[0,x\wedge t]} \bone_{\{x \le \chi(s)\}} d_s \left( \int_{[x-s, x)} \frac{\bar{G}^r(x-)}{\bar{G}^r(y)} \bone_{[0,Q_v(0)/\la]}(y) \bar{G}^r(y)dy \right) d M^r(x) \\ & & \quad + \la \int_{[0,t]} \int_{[0, H^r]} \bone_{\{x \le \chi(s) \wedge s\}} d G^r(x) ds \\
&=& \la \int_{[0, H^r]} \int_0^{t\wedge x} \bone_{\{x \le \chi(s)\}} d_s \left( \int_{[x-s, x)} \bone_{[0,Q_v(0)/\la]}(y) dy \right) d G^r(x) \\ & & \quad + \la \int_0^t G^r(\chi(s) \wedge s) ds \\
&=& \la \int_{[0, H^r]} \int_0^{t} \bone_{\{s\leq x \le \chi(s) \}} \bone_{[0,Q_v(0)/\la]}(x-s) ds d G^r(x) \\ & & \quad + \la \int_0^t G^r(\chi(s) \wedge s) ds \\
&=& \la \int_0^t G^r(\chi(s)\wedge (s+Q_v(0)/\la)) ds = \la \int_0^t G^r(Q_v(s)/\la) ds ,
\end{eqnarray*}
where the last equality follows from \eqref{Qvchi1}. This, together with the definition of $R(t)$ and \eqref{pt-Zu}, implies that \eqref{Ef8} holds.
By using (4.5) of \cite{Zu14}, we obtain
\begin{eqnarray*}
&&\int_{[0, H^s]} \left( \bar{G}^s(x-) K([t-x]^+) + \int_{[x-t]^+}^x \frac{\bar{G}^s(x-)}{\bar{G}^s(y)} \nu_0(dy) \right) d M^s(x) \\
&=& \int_{[0,H^s)} \frac{G^s(y+t)-G^s(y)}{\bar G^s(y)} \nu_0(dy) +
\int_0^t G^s(t-s) dK(s)
\end{eqnarray*}
which is equal to the process $D(t)$ by definition, and implies that \eqref{Ef80} holds.
For each $t\geq 0$, (\ref{33}) holds by applying interchange of the order of integration to (\ref{f5}) and using the definitions of $D$ and $B$. The properties (\ref{f1})--(\ref{f12}) follow from property (ii) of Zhang's fluid model. Thus, this completes the proof that $(\eta,\nu,X)$ is a Zu{\~n}iga's fluid model satisfying Definition \ref{def:EFMA}. Clearly from the construction, both the given Zhang's fluid model and the constructed Zu{\~n}iga's fluid model describe the same $G/GI/N+GI$ queueing system. ~~~\bsq
\begin{theorem} \label{lem:KRZ}
Given a Zu{\~n}iga's fluid model $(\eta,\nu,X)$ for a $G/GI/N+GI$ queueing system with the input data $(\la,\eta_0,\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))\in \mathcal S_0$, there exists a Zhang's fluid model $(\mathcal{R}, \mathcal{Z})$ for the same queueing system
with arrival rate $\la$ if and only if $\eta_0$ satisfies the following condition: for each $t\geq 0$, there exists a solution $z_t$, independent of $x\geq 0$, to the equation in $z$:
\beql{DisLimRtCond}
\la \int^{z}_{t\wedge \chi(t)} \bar{G}^r(x+s) ds
= \bone_{\{\chi(t)\geq t\}}\int_{[0,\chi(t)-t]}\frac{\bar{G}^r(y+t+x)}{\bar G^r(y)}\eta_0(dy),
\eeq
such that \begin{eqnarray}
\la \int_0^t G^r(z_s)ds
&=& \la \int_0^t G^r(\chi(s)\wedge s) ds \non\\ & & \quad + \int_{[0,H^r)} \left(\int_{(y,y+t]} \bone_{\{x\leq \chi(x-y)\}} dG^r(x)\right) \bar{G}^r(y)^{-1}\eta_0(dy). \label{DisLimRtCond1}
\end{eqnarray}
In this case, $\mathcal{Z}_0$ can be chosen as defined by \eqref{con2} for the given $\nu_0$, and $\mathcal{R}_0$ can be chosen as defined by \eqref{r3} for $Q_v(0)=z_0 \la$, where $z_0$ is the solution, independent of $x\geq 0$, that satisfies \eqref{DisLimRtCond} for $t=0$.
\end{theorem}
\begin{remark}
When $G^r$ has a density $g^r$, the conditions \eqref{DisLimRtCond} and \eqref{DisLimRtCond1} can be replaced as follows: for each $t\geq 0$, there exists a solution $z_t$, independent of $x\geq 0$, to the equation in $z$:
\beql{DisLimRtCond2}
\la G^r(x+z)
= \la G^r(x+t\wedge \chi(t))+ \bone_{\{\chi(t)\geq t\}}\int_{[0,\chi(t)-t]}\frac{g^r(y+t+x)}{\bar G^r(y)}\eta_0(dy).
\eeq
In fact, \eqref{DisLimRtCond} follows from \eqref{DisLimRtCond2} directly by integrating both sides of \eqref{DisLimRtCond} in $x$.
It follows from \eqref{DisLimRtCond2} with $x=0$ that \begin{eqnarray*}
& & \la \int_0^t G^r(\chi(s)\wedge s) ds \non\\ & & \quad + \int_{[0,H^r)} \left(\int_{(y,y+t]} \bone_{\{x\leq \chi(x-y)\}} dG^r(x)\right) \bar{G}^r(y)^{-1}\eta_0(dy) \\ &=& \la \int_0^t G^r(\chi(s)\wedge s) ds \non\\ & & \quad + \int_{[0,H^r)} \left(\int_0^t \bone_{\{y\leq \chi(x)-x\}} g^r(x+y)dx\right) \bar{G}^r(y)^{-1}\eta_0(dy) \\ &=& \la \int_0^t G^r(\chi(s)\wedge s) ds + \int_0^t \la(\bar{G}^r(s\wedge \chi(s))-\bar{G}^r(z_s)) ds \\ &=& \la \int_0^t G^r(z_s)ds.
\end{eqnarray*}
Thus, \eqref{DisLimRtCond1} holds.
\end{remark}
\noindent {\it Proof of Theorem \ref{lem:KRZ}.} We first show the ``only if" part. Recall that in Zhang's fluid model, $\mathcal{R}_t^+$, the restriction of $\mathcal{R}_t$ on $[0,\infty)$, tracks the residual patience times of the fluid content of customers in queue at time $t$.
If there exists a Zhang's fluid model $(\mathcal{R}, \mathcal{Z})$ to describe the same $G/GI/N+GI$ queueing system together with Zu{\~n}iga's fluid model $(\eta,\nu,X)$, the measure $\mathcal{R}_t$ must satisfies (see \eqref{etaz})
$$ \mathcal{R}_t(C_x)=\bone_{\{\chi(t)\geq t\}}\int_{[0,\chi(t)-t]}\frac{\bar{G}^r(y+t+x)}{\bar G^r(y)}\eta_0(dy) + \la\int_{0}^{t\wedge \chi(t)} \bar{G}^r(s+x) ds, $$
for each $t\ge 0$ and $x\ge0$, and
hence $\eta_0$ must satisfy that for each $t\ge 0$ and $x\geq 0$,
\beql{DisLimRt}
\la \int^{Q_v(t)/\la}_{t\wedge \chi(t)} \bar{G}^r(x+s) ds
= \bone_{\{\chi(t)\geq t\}}\int_{[0,\chi(t)-t]}\frac{\bar{G}^r(y+t+x)}{\bar G^r(y)}\eta_0(dy).
\eeq
When $t=0$, (\ref{DisLimRt}) is reduced to
\beql{con1}
\mathcal{R}_0(C_x) = \bone_{\{\chi(0)\geq 0\}}\int_{[0,\chi(0)]}\frac{\bar{G}^r(y+x)}{\bar G^r(y)}\eta_0(dy),\ x\geq 0,
\eeq
which is discussed in Remark \ref{remk:z}.
Moreover, in Zhang's fluid model, since customers in queue will renege when their residual patience times reach zero, then by differentiating (\ref{r1}) in $x$ and letting $x=0$, we have the abandonment rate at time $t$ is given by \[\la \left(\bar{G}^r(x) - \bar{G}^r(x + Q_v(t)/\la)\right)\mid_{x=0}= G^r(Q_v(t)/\la).
\]
Then $R(t)$, the cumulative abandonment by time $t$, is given by $\int_0^t G^r(Q_v(s)/\la)ds$. On the other hand, by Zu{\~n}iga's fluid model, $R(t)$ is given by (\ref{mrq}). Then $\eta_0$ must also satisfy that for each $t>0$,
\begin{eqnarray}
&& \quad \la \int_0^t G^r(Q_v(s)/\la)ds = \la \int_0^t G^r(\chi(s)\wedge s) ds \label{DisLimRt4} \\ & & \qquad \qquad \qquad + \int_{[0,H^r)} \left(\int_{(y,y+t]} \bone_{\{x\leq \chi(x-y)\}} dG^r(x)\right) \bar{G}^r(y)^{-1}\eta_0(dy). \non
\end{eqnarray}
Note that for each $t\geq 0$, $Q_v(t)$ satisfies \eqref{DisLimRtCond} and \eqref{DisLimRtCond1}, independent of $x\geq 0$. Hence the ``only if" part is established.
For the ``if" part, let $\mathcal{Z}_0$ and $\mathcal{R}_0$ be defined as in the statement of the theorem. It is clear that the defined $\mathcal{R}_0$ and $\eta_0$ satisfy (\ref{con1}) and $(\mathcal{R}_0, \mathcal{Z}_0)$ satisfies property (iii) of Zhang's fluid model.
Let $\chi(t)$, $B(t)$, $Q(t)$, $K(t)$, $D(t)$, $R(t)$ be the associated auxiliary processes from Zu{\~n}iga's fluid model $(\eta,\nu,X)$.
For each $t>0$, define \[Q_v(t)\doteq \la z_t\ \mbox{ and } L_v(t) \doteq \lambda t - Q_v(t),\] where $z_t$ is the solution, independent of $x\geq 0$, that satisfies (\ref{DisLimRtCond}) and (\ref{DisLimRtCond1}).
Define $\mathcal{R}_t$ and $\mathcal{Z}_t$ by the right hand sides of (\ref{r1}) and (\ref{r2}), respectively. We show that the pair of processes $(\mathcal{R}, \mathcal{Z})$ satisfies Zhang's fluid model. In fact, it suffices to show conditions (i) and (ii) of Zhang's fluid model.
From the definition of $Q_v(t)$, we have for each $x\geq 0$,\[\la \int^{Q_v(t)/\la }_{t\wedge \chi(t)} \bar{G}^r(x+s) ds
= \bone_{\{\chi(t)\geq t\}}\int_{[0,\chi(t)-t]}\frac{\bar{G}^r(y+t+x)}{\bar G^r(y)}\eta_0(dy).\]
Combining this, the construction of $\mathcal{R}$ and (\ref{f4}), we have
\begin{eqnarray} \mathcal{R}_t(\RR_+) &=& \bone_{\{\chi(t)\geq t\}}\int_{[0,\chi(t)-t]}\frac{\bar{G}^r(y+t)}{\bar G^r(y)}\eta_0(dy) + \la \int^{t\wedge \chi(t) }_{0} \bar{G}^r(s) ds \non\\ &=& \eta_t[0,\chi(t)] = Q(t). \label{Q*Q}
\end{eqnarray}
Since $Q(t)$ is of bounded variation by (\ref{31}), the previous display implies that $Q_v$ and hence $L_v$ are also of bounded variation. Thus, condition (i) of Zhang's fluid model holds.
Next we show that condition (ii) of Zhang's fluid model holds for $B^*(t)=\mathcal{Z}_t(\RR_+)$, $Q^*(t)=\mathcal{R}_t(\RR_+)$ and $X^*(t)=B^*(t)+Q^*(t)$. Note that for each $t\geq 0$, we have showed that $Q^*(t)=Q(t)$. By using the definition of $B^*$, the construction of $\mathcal Z$ and the property of $\mathcal Z_0$, we have \begin{eqnarray}B^*(t)&=& \mathcal{Z}_t(\RR_+) \nonumber\\ &=& \mathcal{Z}_0(C_{t}) + \int_0^t \bar{G}^r(Q_v(s)/\la) \bar{G}^s(t-s) d L_v(s) \nonumber\\ &=& \int_{[0,H^s)}\frac{\bar G^s(y+t)}{\bar G^s(y)}\nu_0(dy) + \int_0^t \bar{G}^r(Q_v(s)/\la) \bar{G}^s(t-s) d L_v(s). \label{B*} \end{eqnarray}
By \eqref{mrq}, (\ref{DisLimRtCond1}) and $\la(t) = \la$ for each $t\ge 0$, we have
\begin{eqnarray*}
R(t) &=&
\int_{[0,H^r)} \left(\int_{(y,y+t]} \bone_{\{y\leq \chi(x-y)-(x-y)\}} dG^r(x)\right) \bar{G}^r(y)^{-1}\eta_0(dy) \non\\
&& \qquad + \la \int_0^t \int_{[0,H^r]} \bone_{\{x\leq s \wedge \chi(s)\}} dG^r(x) ds \non \\ &=& \int_{[0,H^r)} \left(\int_{(y,y+t]} \bone_{\{x\leq \chi(x-y)\}} dG^r(x)\right) \bar{G}^r(y)^{-1}\eta_0(dy) \non\\
&& \qquad
+ \la \int_0^t G^r(\chi(s)\wedge s) ds \\
&=& \la \int_0^t G^r(Q_v(s)/\la)ds.
\end{eqnarray*}
In addition, since $G^r_d(Q_v(t)/\la) = Q(t)/\la$ by (\ref{Q*Q}), by the chain rule formula,
$$Q(t) = Q(0)+ \int_0^t \bar G^r(Q_v(s)/\lambda)dQ_v(s).$$
These, together with (\ref{31}), imply that \[K(t)= \la t + Q(0)-Q(t)-R(t)=\int_0^t \bar{G}^r(Q_v(s)/\la) d L_v(s).\] Hence, by (\ref{B*}), $B^*(t)=B(t)$ and then $X^*(t)=X(t)$. Since $B, Q, X$ satisfy (\ref{f1})--(\ref{f12}), then $B^*,Q^*,X^*$ satisfy condition (ii) of Zhang's fluid model. This completes the proof that $(\mathcal{R}, \mathcal{Z})$ is a Zhang's fluid model. Clearly from the construction, both the given Zu{\~n}iga's fluid model and the constructed Zhang's fluid model describe the same $G/GI/N+GI$ queueing system. ~~~\bsq
\begin{coro} \label{coro:KRZ}
Given a Zu{\~n}iga's fluid model $(\eta,\nu,X)$ for a $G/GI/N+GI$ queueing system with the input data $(\la,\eta_0,\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))\in \mathcal S_0$ and $\eta_0(dx)= \la \bone_{[0,a]}(x) \bar{G}^r(x)dx$ for some $a\geq 0$, then one can construct a Zhang's fluid model $(\mathcal{R}, \mathcal{Z})$ for the same queueing system
with arrival rate $\la$, $\mathcal{Z}_0$ defined by \eqref{con2} for the given $\nu_0$, and $\mathcal{R}_0$ defined by \eqref{r3} for $Q_v(0)=a \la$.
\end{coro}
\noindent {\it Proof.} It suffices to check that the given $\eta_0$ satisfies (\ref{DisLimRtCond}) and (\ref{DisLimRtCond1}). Note that for the given $\eta_0$, the equation in (\ref{DisLimRtCond}) becomes
\begin{eqnarray*}
\int^{z}_{t\wedge \chi(t)} \bar{G}^r(x+s) ds
= \bone_{\{\chi(t)\geq t\}}\int_t^{t+(\chi(t)-t)\wedge a}\bar{G}^r(s+x)ds.
\end{eqnarray*}
For $t\geq 0$ such that $\chi(t)\geq t$, we can choose $z_t=t+ (\chi(t)-t)\wedge a$ and for $t\geq 0$ such that $\chi(t)< t$, we can choose $z_t=\chi(t)$. Clearly, in either case, $z_t$ does not depend on $x\geq 0$. Now we show that $z_t$ satisfies (\ref{DisLimRtCond1}). Note that for the given $\eta_0$, by (\ref{Rr1}),
\begin{eqnarray*}
& & \int_{[0,H^r)} \left(\int_{(y,y+t]} \bone_{\{x\leq \chi(x-y)\}} dG^r(x)\right) \bar{G}^r(y)^{-1}\eta_0(dy) \\ &=& \int_{[0,H^r]} \int_{[0,t]} \bone_{\{x\leq \chi(s)\}} d_s \left(\int_{[(x-s)^+,x)} \frac{\bar{G}^r(x-)}{\bar{G}^r(y)} \eta_0(dy)\right) dM^r(x) \\ &=& \la \int_{[0,H^r]} \int_{[0,t]} \bone_{\{x\leq \chi(s)\}} d_s \left(\int_{(x-s)^+}^x \bone_{[0,a]}(y) dy\right) dG^r(x) \\ &=& \la \int_{[0,H^r]} \int_{[0,t]} \bone_{\{x\leq \chi(s)\}} \bone_{\{s<x\leq s+a\}} ds dG^r(x) \\ &=& \la \int_0^t (G^r(\chi(s)\wedge (s+a)) - G^r(s\wedge \chi(s))) ds.
\end{eqnarray*}
It follows that
\begin{eqnarray*}
& & \la \int_0^t G^r(\chi(s)\wedge s) ds \non\\ & & \quad + \int_{[0,H^r)} \left(\int_{(y,y+t]} \bone_{\{x\leq \chi(x-y)\}} dG^r(x)\right) \bar{G}^r(y)^{-1}\eta_0(dy) \\ &=& \la \int_0^t G^r(\chi(s)\wedge s) ds + \la \int_0^t (G^r(\chi(s)\wedge (s+a)) - G^r(s\wedge \chi(s))) ds \\ &=& \la \int_0^t \bone_{\{\chi(s)\geq s\}} \left(G^r(s)+G^r(\chi(s)\wedge (s+a)) - G^r(s) \right)ds \\ & & \quad + \la \int_0^t \bone_{\{\chi(s)< s\}} G^r(\chi(s))ds \\ &=& \la \int_0^t G^r(z_s)ds.
\end{eqnarray*}
Thus, (\ref{DisLimRtCond1}) holds for the choice of $z_t$ and hence the corollary follows directly from Lemma \ref{lem:KRZ}. ~~~\bsq
\iffalse
\bigskip
\subsection{A new fluid model tracking residual times} \label{secnewmodel}
From the discussions in the previous section, in studying many-server queueing systems where arrival rate of customers is time dependent and arrival patterns of customers before time zero and after time zero are dramatically different, Zhang's fluid model is not applicable (see Remark \ref{remk:la}).
Moreover, for a fluid many-server queueing system where Kang-Ramanan's fluid model is formulated, there may not exist a Zhang's fluid model (see Remarks \ref{remk:z} and \ref{remark:ari}).
Therefore, we introduce a new measure-valued fluid model tracking residual times and show that this fluid model is equivalent to Kang-Ramanan's fluid model (and hence Whitt's fluid model) under general initial conditions and time-varying arrival rates.
Let $\nu^\ell_t$ be a nonnegative finite measure on $[0,\infty)$ with support in $[0,H^s)$ such that $\nu^\ell_t(dx)$ represents the amount of fluid content of customers in service whose residual service time at time $t$ lies in the range $[x,x+dx)$. Similarly, let $\eta^\ell_t$ be a nonnegative finite measure on $[0,\infty)$ with support in $[0,H^r)$ such that $ \eta^\ell_t(dx)$ represents the amount of fluid content of customers in queue whose residual patience time at time $t$ lies in the range $[x,x+dx)$.
\begin{definition} \label{def:FMR}
A pair of functions $(\nu^\ell,\eta^\ell)$ is a measure-valued fluid model tracking residual times with the input data $(\lambda(\cdot), \eta_0,\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))\in \mathcal S_0$ if there exists a non-decreasing c\`{a}dl\`{a}g function $\varsigma(\cdot)$ defined on $[0,\infty)$ such that
\beql{mrvs1}
\varsigma(0)=-\inf\{x\in [0,H^r):\ \eta_0[0,x) \geq X(0)-\nu_0[0,H^s)\} ,
\eeq
\beql{mrvs2}
\varsigma(t)\leq t \quad \text{for each} \ t \ge 0,
\eeq
and the following equations hold. For every $\psi \in \mathcal{C}_b(\RR_+)$,
\beqal{mr4}
\int_0^\infty \psi(x)\eta^\ell_t(dx) & = & \int_0^{\varsigma^-(t)}\left(\int_{y+t}^\infty \frac{\psi(x-y-t)}{\bar G^r(y)}dG^r(x)\right)\eta_0(dy) \non\\ && \qquad + \int_{\varsigma^+(t)}^{t} \left(\int_{t-s}^\infty \psi(x-t+s) d G^r(x) \right)\lambda(s) ds ,
\eeqa
\beqal{mr2}
\int_0^\infty \psi(x) \nu^\ell_t(dx) &=& \int_{[0,H^s)}\left(\int_{y+t}^\infty \frac{\psi(x-y-t)}{\bar G^s(y)}d G^s(x)\right)\nu_0(dy) \non\\
&& \qquad + \int_0^{t} \left(\int_{t-s}^\infty \psi(x-t+s)d G^s(x)\right)dK^\ell(s),
\eeqa
\beql{mrb1}
Q^\ell(0)+E(t) = Q^\ell(t) + K^\ell(t) + R^\ell(t),
\eeq
\beqal{mrq}
R^\ell(t) &=&
\int_{[0, H^r)} \left( \int_{(y, y+t]} \bone_{\{y\le - \varsigma(s-y)\}} d G^r(s)\right) (\bar{G}^r(y))^{-1} \eta_0(dy) \non \\
& & \qquad +
\int_0^t \left(\int_{0}^{s-\varsigma^+(s)} \la(s-u)d G^r(u) \right)ds,
\eeqa
\beql{mr61}
Q^\ell(t) = \eta_t^\ell[0,\infty) = \int_0^{\varsigma^-(t)} \frac{\bar G^r(y+t)}{\bar G^r(y)}\eta_0(dy) +
\int_{\varsigma^+(t)}^t \bar G^r(t-s) \la(s)ds ,
\eeq
\beql{mrd}
D^\ell(t) = \int_{[0,H^s)} \frac{G^s(y+t)-G^s(y)}{\bar G^s(y)} \nu_0(dy) +
\int_0^t G^s(t-s) dK^\ell(s) ,
\eeq
\beql{mr621}
B^\ell(t)
=\int_{[0,H^s)} \frac{\bar G^s(y+t)}{\bar G^s(y)} \nu_0(dy) +
\int_0^t \bar G^s(t-s) dK^\ell(s) ,
\eeq
and the non-idling condition
\beql{mrnon-idling}
B^\ell(t)=(B^\ell(t)+Q^\ell(t))\wedge 1 \quad \mbox{ and } \quad Q^\ell(t)(1-B^\ell(t))=0 .
\eeq
\end{definition}
\begin{remark}
The initial condition in the input data to the fluid model in Definition \ref{def:FMR} is $(\eta_0,\nu_0,X(0))$ rather than $(\nu_0^\ell,\eta_0^\ell)$. This is because $(\nu_0^\ell,\eta_0^\ell)$ only contains information on the residual patience time of fluid content of customers in queue and the residual service time of fluid content of customers in service, but this information is insufficient to determine the service order of the initial fluid content of customers in queue since we do not know the arrival order of the fluid content from $(\nu_0^\ell,\eta_0^\ell)$ and the service discipline is FCFS. On the other hand, the information encoded in $(\eta_0,\nu_0,X(0))$ contains the elapsed time $($and hence the arrival order$)$ of the fluid content of customers in queue.
Thus, it is sufficient to use $(\eta_0,\nu_0,X(0))$ together with $\lambda(\cdot)$ as the input data for the fluid model in Definition \ref{def:FMR}.
\end{remark}
\begin{remark}
Zu{\~n}iga's fluid model naturally gives rise to the two measure-valued processes that keep track of residual times of fluid contents of customers in queue and in service.
For each $t\geq 0$, clearly the following mapping
\begin{eqnarray*}\psi &\mapsto & \int_{[0,H^s)}\left(\int_{(y+t,\infty)} \frac{\psi(x-y-t)}{\bar G^s(y)}d G^s(x)\right)\nu_0(dy) \\
&& \qquad + \int_{[0,t]} \left(\int_{(t-s,\infty)} \psi(x-t+s)d G^s(x)\right)dK(s)\end{eqnarray*} is a positive linear functional on $\mathcal{C}_c(\RR_+)$ since $K$ is non-decreasing by Lemma \ref{lem:chiinc}. Then by Riesz-Markov-Kakutani representation theorem, there is a unique regular Borel measure $\nu^\ell_t$ with support $[0,H^s)$ such that \eqref{mr2} holds. Similarly, for each $t\geq 0$, there is a unique regular Borel measure $\eta^\ell_t$ with support $[0,H^r)$ such that \eqref{mr4} holds.
We will show in Theorem \ref{thm-mr} that Zu{\~n}iga's fluid model and the new fluid model are indeed equivalent.
\end{remark}
\begin{remark} The fluid model in Definition \ref{def:FMR} does not require the patience time distribution $G^r$ and service time distribution $G^s$ to have densities, respectively. When $G^r$ and $G^s$ have densities $g^r$ and $g^s$, respectively, (\ref{mr4}) and (\ref{mr2}) are equivalent to the following representations:
\beqal{mr41}
\int_0^\infty \psi(x)\eta^\ell_t(dx) & = & \int_0^{\varsigma^-(t)}\left(\int_0^\infty \frac{g^r(y+t+x)}{\bar G^r(y)}\psi(x)dx\right)\eta_0(dy) \non\\ && \qquad + \int_{\varsigma^+(t)}^{t} \left(\int_0^\infty g^r(t-s+x) \psi(x)dx \right)\lambda(s) ds ,
\eeqa
\beqal{mr21}
\int_0^\infty \psi(x) \nu^\ell_t(dx) &=& \int_{[0,H^s)}\left(\int_0^\infty \frac{g^s(y+t+x)}{\bar G^s(y)}\psi(x)dx\right)\nu_0(dy) \non\\
&& \qquad + \int_0^{t} \left(\int_0^\infty g^s(t-s+x)\psi(x)dx\right)dK^\ell(s).
\eeqa
In this case, for each $t\geq 0$, the two measures $\eta^\ell_t$ and $\nu^\ell_t$ have
densities $b_\ell(t,x)$ and $q_t(t,x)$, respectively, which can be expressed as
\beql{tr56}
b_\ell(t,y) = \int_{[0,H^s)}\frac{g^s(x+t+y)}{\bar G^s(x)}\nu_0(dx) + \int_0^t g^s(y+t-u) dK^\ell(u),
\eeq
and
\beql{tr567} q_\ell(t,y) = \int_0^{\varsigma^-(t)}\frac{g^r(x+t+y)}{\bar G^r(x)}\eta_0(dx)+ \int_{\varsigma^+(t)}^{t} g^r(y+t-u) \lambda (u) du.
\eeq
In addition to $(\nu^\ell,\eta^\ell)$, we also introduce a measure-valued process $\tilde{\eta}^\ell$ satisfying that for every $\psi \in \mathcal{C}_b(\RR_+)$,
\beqal{mr62}
\int_0^\infty \psi(x) \tilde{\eta}^\ell_t(dx) &=& \int_{[0,H^r)}\left(\int_0^\infty \frac{g^r(y+t+x)}{\bar G^r(y)}\psi(x)d x\right)\eta_0(dy) \non\\
&& \qquad + \int_0^{t} \left(\int_0^\infty g^r(t-u+x)\psi(x)dx\right)\lambda(u) du.
\eeqa
For each $t\geq 0$ and $x\geq 0$, $\tilde{\eta}^\ell_t(dx)$ represents the amount of fluid content of customers in the potential queue whose residual patience times at time $t$ lie in the range $[x,x+dx)$.
For each $t\geq 0$, the measure $\tilde \eta^\ell_t$ admits a density function $\tilde{q}_\ell(t,y)$ with the representation:
\beql{mr64}
\tilde{q}_\ell(t,y)=\int_{[0,H^r)}\frac{g^r(x+t+y)}{\bar G^r(x)}\eta_0(dx)+ \int_0^{t} g^r(y+t-s) \lambda(s) ds.
\eeq
Moreover, by applying interchange of the order of integration, (\ref{mrq}) is equivalent to the following.
\beqal{mrq1}
R^\ell(t) &=& \int_0^t \left(\int_0^{\varsigma^-(s)}\frac{g^r(y+s)}{\bar G^r(y)}\eta_0(dy)\right)ds \\ & & \qquad +
\int_0^t \left(\int_{\varsigma^+(s)}^s g^r(s-u) \la(u)d u \right)ds. \non
\eeqa
\end{remark}
We now provide an informal, intuitive explanation for the structure of the fluid equations in Definition \ref{def:FMR}. The quantities $B^\ell(t)$, $Q^\ell(t)$, $D^\ell(t)$, $K^\ell(t)$, $R^\ell(t)$ in Definition \ref{def:FMR} represent, respectively, the quantities at time $t$ that correspond to $B(t)$, $Q(t)$, $D(t)$, $K(t)$, $R(t)$ in the fluid model tracking elapsed times in Definition \ref{def:FMA}. The quantity $\varsigma(t)$ can be interpreted as the arrival time of the fluid content of customers that has been in queue the longest at time $t$. So it is clear that (\ref{mrvs2}) holds and intuitively, we see that $\varsigma(t)$ can be related to $\chi(t)$ as follows: \beql{chiw}
\varsigma(t) = t - \chi(t), \quad t\ge 0.
\eeq At time 0, $\varsigma(0) = - \chi(0)$ represents the arrival time of the oldest fluid content in queue initially. Thus, \eqref{mrvs1} holds by \eqref{f9}. If $\varsigma(t) <0$ at time $t> 0$, it means that the initial fluid content of customers at time 0 has not all entered service yet by time $t$, and can be interpreted as the arrival time of the initial fluid content that has been in queue the longest at time $t$.
Recall that $\nu_t^\ell(dx)$ represents the amount of fluid content of customers in service whose residual time lies in the range $[x,x+dx)$ at time $t$. This amount of fluid content comes from both the fluid content of customers initially in service whose residual times lie in the range $[x+t,x+t+dx)$ and the fluid content of customers that have entered service during $[0,t]$. For the amount of fluid content of customers initially in service whose elapsed times lie in $[y,y+dy]$ (represented by $\nu_0(dy)$), the fraction of fluid content with residual service times in the range $[x+t,x+t+dx)$ is given by $g^s(y+t+x)/\bar G^s(y)$. So it is natural to expect that the amount of fluid content of customers initially in service whose residual times lie in the range $[x+t,x+t+dx)$ is given by $\int_{[0,H^s)}(g^s(y+t+x)/\bar G^s(y))\nu_0(dy)$. Likewise, for the amount of fluid content of customers that have entered service at time $0<s<t$ (represented by $dK^\ell(s)$), the fraction of fluid content with residual service time in the range $[x,x+dx)$ is given by $g^s(t -s + x)$. So $\int_0^t g^s(t-s + x) dK^\ell(s)$ represents the fluid content of customers that have entered service during $[0,t]$ and whose residual times lie in the range $[x,x+dx)$ at time $t$. Thus, one expects (\ref{mr21}) holds. In a similar fashion, one can also expect that (\ref{mr41}) holds. Note that the first integral on the right hand side of (\ref{mr41}) is integrated up to $\varsigma^{-}(t)$ rather than $H^r$ since only the fluid content of customers initially in queue whose elapsed times lie in the range $[0,\varsigma^{-}(t)]$ can remain in queue at time $t$, and the second integral on the right hand side of \eqref{mr41} is integrated from $\varsigma^+(t)$ to $t$ since the fluid content of customers arriving in the time interval $[0, \varsigma^+(t)]$ have entered service or reneged by time $t$.
Equations (\ref{mrb1}) and (\ref{mrnon-idling}) are simply the mass conservation equation and the non-idling condition. Since only the fluid content of customers in queue with residual time $0$ can renege from queue, and hence the reneging rate is $\eta^\ell_s[0,0+dx)$, then (\ref{mrq}) (i.e, (\ref{mrq1})) holds. Equation (\ref{mr61}) holds simply by letting $\psi=\bone$ in (\ref{mr41}). Similarly, equations (\ref{mrd}) and (\ref{mr621}) hold from (\ref{mr21}).
\subsection{Equivalence of the new fluid model tracking residual times and the two fluid models tracking elapsed times} \label{secnewequiv}
In this section, we will prove the following Theorem \ref{thm-mr}, which shows that the new fluid model tracking residual times in Definition \ref{def:FMR} is equivalent to the two fluid models tracking elapsed times.
\begin{theorem} \label{thm-mr}
For the same $G_t/GI/N+GI$ queueing system, the existence and uniqueness of the Zu{\~n}iga's fluid model in Definition \ref{def:EFMA} with the input data $(\lambda(\cdot), \eta_0,\nu_0,X(0))$ of diffusive measures $\eta_0$ and $\nu_0$, are equivalent to the existence and uniqueness of the new measure-valued fluid model in Definition \ref{def:FMR} with the same input data, under the assumptions that
$G^r$ is continuous, $\eta_0$ is diffuse, and $\nu_0$ is diffuse if $G^s$ is not continuous.
\end{theorem}
As a consequence, the existence and uniqueness for the new fluid model in Definition \ref{def:FMR} can be thus established without direct proofs by an FSLLN or characterization of its solution, but instead by the equivalence in Theorem \ref{thm-mr} and the existence and uniqueness of Kang-Ramanan's fluid model in \cite{KR10, Kang} and its generalization in \cite{Zu14}.
The rest of the section proceeds as follows. We complete the proof of Theorem \ref{thm-mr} in Section \ref{secResEqv} and then establish a coupling property as a corollary of Theorem \ref{thm-mr} in Section \ref{secCouplingver}.
\subsubsection{Proof of Theorem \ref{thm-mr}} \label{secResEqv}
We now show that Theorem \ref{thm-mr} holds by proving the following two propositions, which state that a fluid model tracking elapsed times in Definition \ref{def:FMA} can give rise to a fluid model tracking residual times in Definition \ref{def:FMR}, and vice versa.
\begin{prop} \label{prop-re}
Given a Zu{\~n}iga's fluid model tracking elapsed times $(\eta,\nu,X)$ in Definition \ref{def:EFMA} with the input data $(\lambda(\cdot),\eta_0,\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))$ $\in \mathcal S_0$, one can construct a fluid model tracking residual times $(\nu^\ell,\eta^\ell)$ in Definition \ref{def:FMR} with the same input data.
\end{prop}
\noindent {\it Proof.}
Fix a fluid model tracking elapsed times $(\eta, \nu, X)$ in Definition \ref{def:EFMA} with the input data $(\lambda(\cdot),\eta_0,\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))\in \mathcal S_0$ and $\eta_0$ and $\nu_0$ are diffusive measures. Let $\chi(t)$, $B(t)$, $Q(t)$, $K(t)$, $D(t)$, $R(t)$ be the associated auxiliary processes in Definition \ref{def:EFMA}.
Define \[\varsigma(t)\doteq t-\chi(t) \mbox{ and } K^\ell(t)\doteq K(t), \mbox{ for } t\geq 0.\] Note that $\varsigma(t)\leq t$ and
$$\varsigma(0)=-\chi(0)=-\inf\{x\in [0,H^r):\ \eta_0[0,x) \geq X(0)-\nu_0[0,H^s)\}.
$$
By Lemma \ref{lem:chiinc}, $\varsigma(t)$ is non-decreasing. (Note that $\chi(t)$ itself is not necessarily monotone.)
Define the processes $\eta^{\ell}_t$ and $\nu^{\ell}_t$ as in \eqref{mr4} and \eqref{mr2}, respectively, with the given $\nu_0$, $\eta_0$ and the defined $K^{\ell}(t)$.
Define the processes $R^{\ell}(t)$, $Q^{\ell}(t)$, $D^{\ell}(t)$ and $B^{\ell}(t)$ as in \eqref{mrq}, \eqref{mr61}, \eqref{mrd} and \eqref{mr621}, respectively, with the given $\nu_0$, $\eta_0$ and the defined $K^{\ell}(t)$ and $\varsigma(t)$. Now it remains to show that the balance equation \eqref{mrb1} and the non-idling condition \eqref{mrnon-idling} are satisfied.
It suffices to show that $R^{\ell}(t) = R(t)$, $Q^{\ell}(t)= Q(t)$, $D^{\ell}(t)= D(t)$, and $B^{\ell}(t) = B(t)$.
Since $s-\varsigma^+(s)=\chi(s)\wedge s$ for each $s\geq 0$, by applying interchange of the order of integration, we obtain that the second term on the right hand side of \eqref{mrq}
\beqal{mrqp1-1}
&& \int_0^t \int_{[0,H^r]} 1_{\{x\leq s-\varsigma^+(s)\}} \la(s-x)\bar{G}^r(x) dM^r(x) ds \\
&=& \int_0^t \int_{[0,H^r]} 1_{\{x\leq \chi(s)\wedge s\}} \la(s-x)\bar{G}^r(x)dM^r(x) ds \non\\
&=& \int_{[0,H^r]} \int_0^t 1_{\{x\leq \chi(s)\wedge s\}} \bar{G}^r(x)\la(s-x)ds dM^r(x) . \non
\eeqa
By the definition of $\varsigma$, we have that
for each $s\in [0,x]$ and $x\in [0,H^r]$, \[1_{\{x\leq \chi(x-s)\}}=1_{\{s\leq -\varsigma(x-s)\}}.\]
By applying interchange of the order of integration,
we obtain that the first term on the right hand side of \eqref{mrq}
\beqal{mrqp1-2}
&& \int_{[0, H^r)} \left( \int_{(y, y+t]} \bone_{\{y\le - \varsigma(s-y)\}} d G^r(s)\right) (\bar{G}^r(y))^{-1} \eta_0(dy) \non \\
&=& \int_{[0,H^r]} \int_{[[x-t]^+,x]} 1_{\{x\leq \chi(x-s)\}}1_{\{s<x\}} \frac{\bar{G}^r(x)}{\bar{G}^r(s)} \eta_0(ds) dM^r(x) \non \\
&=& \int_{[0,H^r]} \int_{[0,x\wedge t]} 1_{\{x\leq \chi(s)\}} d \left(\int_{[x-s,x)} 1_{\{y<x\}}\frac{\bar{G}^r(x)}{\bar{G}^r(y)} \eta_0(dy)\right) dM^r(x) \non \\
&=& \int_{[0,H^r]} \int_{[0,t]} 1_{\{x\leq \chi(s)\}} d \left(\int_{[(x-s)^+,x)} \frac{\bar{G}^r(x)}{\bar{G}^r(y)} \eta_0(dy)\right) dM^r(x) , \non
\eeqa
where the second equality follows from Theorem 3.6.1 of \cite{Bo}, as in \eqref{Rr1}.
Thus, by \eqref{mrq}, \eqref{mrqp1-1} and \eqref{mrqp1-2}, we obtain that $R^{\ell} (t)=R(t)$ defined in \eqref{Ef8}.
By \eqref{mr61}, we have
\beqal{mr61p1}
Q^\ell(t) &=& \int_0^{\varsigma^-(t)} \frac{\bar G^r(y+t)}{\bar G^r(y)}\eta_0(dy) +
\int_{\varsigma^+(t)}^t \bar G^r(t-s) \la(s)ds \non\\
&=& \int_{[0, H^r)} \bone_{[0,\chi(t)]}(y+t) \frac{\bar G^r(y+t)}{\bar G^r(y)}\eta_0(dy) \non\\
&& \quad +
\int_{0}^t \bone_{[0,\chi(t)]}(t-s)\bar G^r(t-s) \lambda(s)ds \non\\
& =& \eta_t[0,\chi(t)] = Q(t). \non
\eeqa
By \eqref{mrd}, applying interchange of the order of integration and integration by parts, and noting that $K^{\ell} (t) = K(t)$, we have
\beqal{mrdp1}
D^\ell(t) &=& \int_{[0,H^s]} \frac{G^s(y+t)-G^s(y)}{\bar G^s(y)} \nu_0(dy) +
\int_0^t G^s(t-s) dK^\ell(s) \non\\
&=& \int_{[0, H^s]} \Bigg( \int_{x-t}^x \frac{\bar{G}^s(x-)}{\bar{G}^s(y)} \nu_0(dy) + \bar{G}^s(x-) K(t-x) \Bigg) d M^s(x) , \non
\eeqa
which is equal to $D(t)$ defined in \eqref{Ef80}.
By \eqref{mr621}, \eqref{f2} and \eqref{f5}, we can easily see that $B^{\ell}(t) = B(t)$.
Thus, (\ref{mrb1}) follows from (\ref{31}) and (\ref{mrnon-idling}) follows from (\ref{f2}) and (\ref{f12}). This complete the proof of the proposition. ~~~\bsq
\begin{prop} \label{prop-er}
Given a fluid model tracking residual times $(\nu^\ell,\eta^\ell)$ in Definition \ref{def:FMR} with the input data $(\lambda(\cdot),\eta_0,\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))\in \mathcal S_0$, one can construct a fluid model tracking elapsed times $(\eta,\nu,X)$ in Definition \ref{def:EFMA} with the same input data.
\end{prop}
\noindent {\it Proof.}
Let $(\nu^\ell,\eta^\ell)$ be a fluid model tracking residual times in Definition \ref{def:FMR} with the input data $(\lambda(\cdot),\eta_0,\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))\in \mathcal S_0$. Let $\varsigma(t)$, $D^\ell(t)$, $R^\ell(t)$, $Q^\ell(t)$, $B^\ell(t)$ and $K^\ell(t)$ be the associated auxiliary processes in Definition \ref{def:FMR}.
For each $t\geq 0$, define
$$
\chi(t) \doteq t-\varsigma(t), \quad K(t)\doteq K^\ell(t), \qandq X(t)\doteq B^\ell(t)+Q^\ell(t).
$$
Let $\eta_t$ be the unique measure that satisfies (\ref{f4}) and $\nu_t$ be the unique measure that satisfies (\ref{f5}).
We now show that $(\eta, \nu, X)$ is a fluid model tracking elapsed times in Definition \ref{def:FMA}$^\ddagger$.
For each $t\ge 0$, define $B(t)\doteq \nu_t[0,\infty)$, $Q(t) \doteq \eta_t[0, \chi(t)]$, and $D(t)$ and $R(t)$ as in \eqref{Ef80} and \eqref{Ef8}, respectively. It remains to show that $B(t) =B^{\ell}(t)$, $D(t) = D^{\ell}(t)$, $R(t) = R^{\ell}(t)$, $\chi(t)$ satisfies (\ref{f9}), and the balance equations \eqref{33},\eqref{31}, and non-idling conditions \eqref{f1}, \eqref{f2} and \eqref{f12} hold.
By the definition of $B(t)\doteq \nu_t[0,\infty) $ and \eqref{f5} and $K(t) = K^{\ell}(t)$, we obtain
\beql{Bp2}
B(t) = \int_{[0,H^s)} \frac{\bar G^s(y+t)}{\bar G^s(y)} \nu_0(dy) +
\int_0^t \bar G^s(t-s) dK(s) ,
\eeq
which implies that $B(t) = B^{\ell}(t)$ defined in \eqref{mr621}.
By \eqref{Ef80}, applying integration by parts, we can obtain that $D(t)$ defined in \eqref{Ef80} is equal to $D^{\ell}(t)$ in \eqref{mrd}.
By the definition of $\chi(t)$ and (\ref{mrvs2}), we have that $\chi(t)\geq 0$ for each $t\ge 0$, and
$(t-\varsigma(t))\wedge t = t-\varsigma^+(t)$.
By the interchange of the order of integration, we obtain that
\begin{eqnarray*}
&& \int_{[0,H^r]} \int_0^t 1_{\{x\leq \chi(s)\wedge s\}} \bar{G}^r(x)\la(s-x)ds dM^r(x) \\
&=&\int_{[0,H^r]} \int_0^t 1_{\{x\leq (s-\varsigma(s))\wedge s\}}\bar{G}^r(x)\la(s-x)ds dM^r(x) \\
&=& \int_{[0,H^r]} \int_0^t 1_{\{x\leq s-\varsigma^+(s)\}}\bar{G}^r(x)\la(s-x)ds dM^r(x) ,
\end{eqnarray*}
and
\begin{eqnarray*}
&& \int_{[0,H^r]} \int_{[0,t]} 1_{\{x\leq \chi(s)\}} d \left(\int_{[(x-s)^+,x)} \frac{\bar{G}^r(x)}{\bar{G}^r(y)} \eta_0(dy)\right) dM^r(x) \\
&=& \int_{[0,H^r]} \int_{[0,t]} 1_{\{s\leq x\leq s-\varsigma(s)\}} d \left(\int_{[(x-s)^+,x)} 1_{\{s<x\}}\frac{\bar{G}^r(x)}{\bar{G}^r(y)} \eta_0(dy)\right) dM^r(x) \\
&=& \int_{[0,H^r]} \int_{[x-t\wedge x,x]} 1_{\{s\leq -\varsigma(x-s)\}}1_{\{s<x\}}\frac{\bar{G}^r(x)}{\bar{G}^r(s)}\eta_0(ds) dG^r(x) \\
&=& \int_{[0,H^r)} \left(\int_{(s,s+t]} 1_{\{s\leq -\varsigma(x-s)\}} d G^r(x)\right) \bar{G}^r(s)^{-1}\eta_0(ds)
\end{eqnarray*}
where the second equality applies Theorem 3.6.1 of \cite{Bo}.
This implies that $R(t)$ defined in \eqref{Ef8} is equal to $R^{\ell}(t)$ in \eqref{mrq}.
By the definition of $Q(t) = \eta_t[0, \chi(t)]$ for $\chi(t) \doteq t-\varsigma(t)$ and $\eta_t$ satisfying \eqref{f5}, we have
\beqal{mr611}
Q(t) &=& \eta_t[0,\chi(t)] \non\\
&=& \int_0^{H^r}\bone_{[0,\chi(t)]}(x+t)\frac{\bar G^r(x+t)}{\bar G^r(x)}\eta_0(dx)\non \\
&& \quad + \int_{0}^t \bone_{[0,\chi(t)]}(t-s)\bar G^r(t-s)\lambda(s)ds \non \\
&=& \int_0^{\varsigma^-(t)} \frac{\bar G^r(x+t)}{\bar G^r(x)}\eta_0(dx) +
\int_{\varsigma^+(t)}^t \bar G^r(t-s) \lambda(s) ds,
\eeqa
which implies that $Q(t) = Q^{\ell}(t)$ defined in \eqref{mr61}. The definition of $Q(t) = \eta_t[0, \chi(t)]$ implies that $\chi(t)$ satisfies \eqref{f9}. By (\ref{mrb1}) and the definition of $K$ and the equalities $Q(t) = Q^{\ell}(t)$, and $R(t) = R^{\ell}(t)$, we have (\ref{31}) holds.
By adding the representations of $D(t)$ in \eqref{Ef80} and $B(t)$ in \eqref{Bp2}, we have
\begin{eqnarray*}
B(t)+D(t) = \nu_0[0,\infty) + K(t) = B(0) + K(t).
\end{eqnarray*}
This shows that (\ref{33}) holds.
At last, by (\ref{mrnon-idling}), we have
\[
B(t)=X(t)\wedge 1 \quad \mbox{ and } \quad Q(t)(1-B(t))=0.
\]
Then (\ref{f1})--(\ref{f12}) hold. Thus, we have shown that $(\nu,\eta, X)$ is a fluid model tracking elapsed times in Definition \ref{def:FMA}$^\ddagger$.
~~~\bsq
\bigskip
\noindent {\it Proof of Theorem \ref{thm-mr}.}
The theorem follows directly from Proposition \ref{prop-re} and Proposition \ref{prop-er}.
~~~\bsq
\fi
\iffalse
\section{Useful properties from fluid models tracking residual times}
\subsection{Workload process}
\subsection{A coupling property} \label{secCouplingver} In this section we establish a coupling property that connects $(\eta,\nu)$ and $(\tilde \eta^\ell, \nu^\ell)$ directly.
Intuitively, $\nu^\ell_t[y,\infty)$ represents the fluid content in service at time $t$ that has residual service time greater than $y$. This amount fluid content will still be in service at time $t+y$ with residual service time greater than $0$, which is precisely the amount of fluid content in service at time $t+y$ with elapsed service time greater than $y$. Combining this with the fluid content arriving after time $t$ and before time $t+y$, represented by $\nu_{t+y}[0,y]$, we have the total fluid content in service at time $t+y$, $\nu_{t+y}[0,\infty)$. This and a similar intuition for the fluid content in the potential queue indicate the following coupling property:
\beql{mrme1}
\nu^\ell_t[y,\infty) = \nu_{t+y}[0,\infty) - \nu_{t+y}[0,y],
\eeq
and
\beql{mrme2}
\tilde \eta_t^\ell [y,\infty) = \eta_{t+y} [0,\infty) - \eta_{t+y} [0,y].
\eeq
We establish the coupling property in the following proposition.
\begin{prop} \label{couple}
Let $(\nu_t, \eta_t, X)$ satisfy the fluid model tracking elapsed times in Definition \ref{def:EFMA} and $(\nu_t^\ell,\eta_t^\ell)$ be the associated processes tracking residual times with the common input data $(\lambda(\cdot),\eta_0,\nu_0,X(0))$ such that $(\eta_0,\nu_0,X(0))\in \mathcal S_0$.
Suppose that $G^r$ is continuous, $\eta_0$ is diffuse, and $\nu_0$ is diffuse if $G^s$ is not continuous.
Then $(\nu_t^\ell,\tilde \eta_t^\ell)$ and $(\nu_t, \eta_t)$ satisfy the coupling property \eqref{mrme1} and \eqref{mrme2}.
\end{prop}
{\it Proof.} Simple calculations with (\ref{mr4}) show that
\begin{eqnarray*}
\tilde \eta_t^\ell [y,\infty) &=& \int_{[0,H^r)}\frac{\bar{G^r}(u+t+y)}{\bar G^r(u)}\eta_0(du) + \int_0^{t} \bar{G^r}(t-u+y)\lambda(u) du \\ &=& \eta_{t+y} (y,\infty) \\ &=& \eta_{t+y} [0,\infty) - \eta_{t+y} [0,y],
\end{eqnarray*}
which establishes \eqref{mrme2}. For \eqref{mrme1}, by the similar calculations with (\ref{mr2}), we have
\begin{eqnarray*}
\nu^\ell_t[y,\infty) &=& \int_{[0,H^s)}\frac{\bar{G^s}(u+t+y)}{\bar G^s(u)}\nu_0(du)+ \int_0^{t} \bar{G^s}(t-u+y)dK^\ell(u),
\end{eqnarray*}
Thus, to establish \eqref{mrme1}, it suffices to show that $K^\ell=K$. It follows from Proposition \ref{prop-er} that there exists a triple $(\eta^*,\nu^*,X^*)$ satisfying Definition \ref{def:EFMA} with the same input data. In particular, it is shown in the proof of Proposition \ref{prop-er} that the ancillary process $K^*=K^\ell$. Since both $(\eta^*,\nu^*,X^*)$ and $(\nu_t, \eta_t, X)$ satisfy Definition \ref{def:EFMA} with the same initial data, then by the uniqueness of the fluid model tracking elapsed times established in \cite{Zu14}
$K^*=K$. This shows that $K^\ell=K$ and hence \eqref{mrme1}.
~~~\bsq
\bigskip
We remark that the coupling property in \eqref{mrme1}-\eqref{mrme2} should also hold for the associated prelimit processes. Moreover, we show in the following corollary that the coupling property \eqref{mrme1}-\eqref{mrme2} implies a coupling property for the corresponding density functions $(b_\ell(t,y),\tilde{q}_\ell(t,y))$ and $(b(t,x),\tilde{q}(t,x))$.
\begin{coro} \label{fluiddensity}
Under the assumptions of Proposition \ref{couple}, suppose that $G^s$ and $G^r$ have densities $g^s$ and $g^r$, respectively, the density functions $(b_\ell(t,x),\tilde{q}_\ell(t,x))$ and $(b(t,x),\tilde{q}(t,x))$ $($if exist$)$ have the following coupling property:
\beqal{twor9}
b_\ell(t,x) &=& \int_0^{H^s} b(t,u) \frac{g^s(x+u)}{\bar G^s(u)} d u ,
\eeqa
and
\beqal{twor7}
\tilde{q}_\ell(t,x) &=& \int_0^{H^r} \tilde{q}(t,u) \frac{g^r(x+u)}{\bar G^r(u)} d u.
\eeqa
\end{coro}
\noindent {\it Proof.} Suppose that $(b(t,x),\tilde{q}(t,x))$ exist and the assumptions of Proposition \ref{couple} hold.
By \eqref{w1}, \eqref{mrme1}, $ \nu^\ell_t[y,\infty)$ can be written as integrals of $b(t,y)$:
\beql{w11}
\nu^\ell_t[y,\infty) = \int_0^{\infty} b(t+y,x) dx - \int_0^y b(t+y,x) dx = \int_y^{\infty} b(t+y,x) dx .
\eeq
By the first fundamental evolution equation \eqref{w2} and \eqref{w11}, we have
\beqal{twor1}
\nu^\ell_t[y,\infty) &=& \int_y^{\infty} b(t+y,x) dx = \int_0^{\infty} b(t+y, x+y) dx \non\\
&=& \int_0^{H^s} b(t,x) \frac{\bar G^s(x+y)}{\bar G^s(x)} dx,
\eeqa
which implies (\ref{twor9}) holds.
Similar argument using Equation \eqref{mrme2} and the fundamental evolution equation \eqref{w2} implies that
\eqref{twor7} holds.
~~~\bsq
\bigskip
\iffalse
As a converse to Proposition \ref{fluiddensity}, we have the following proposition on a representation of the fluid model $(\eta_t, \nu_t, X(t))$ from the fluid model $(\eta^\ell_t, \nu^\ell_t)$.
\begin{prop} \label{fluiddensityconverse}
Assume the system starts from empty. The measure-valued fluid model tracking elapsed times $(\eta_t, \nu_t, X(t))$ can be written from the fluid model tracking residual times $(\tilde{\eta}^\ell_t, \nu_t^\ell)$ and $\eta_t^\ell$ $($or equivalently from $b_\ell(t,y)$ and $q_\ell(t,y))$ as follows:
\beql{twor9conv}
\nu_t[0,x]=\left\{\begin{array}{ll}
\nu^\ell_t[0,\infty)-\nu^\ell_{t-x}[x,\infty), & x \le t \\
\nu^\ell_t[0,\infty), & x> t,
\end{array} \right.
\eeq
\beql{twor7conv}
\eta_t[0,x]=\left\{\begin{array}{ll}
\tilde \eta^\ell_t[0,\infty)-\tilde \eta^\ell_{t-x}[x,\infty), & x \le t \\
\tilde \eta^\ell_t[0,\infty), & x> t ,
\end{array} \right.
\eeq
and
\beql{twor8conv}
X(t)= E(t)-\int_0^tb_\ell(s,0)ds -\int_0^t q_\ell(s,0)ds.
\eeq
\end{prop}
{\it Proof.} We first note from (\ref{r11}) that when the system starts from empty, a converse to the relationship (\ref{r11}) holds:
\beql{r11conv}
B(t,y) = B_\ell(t, 0) - B_\ell(t-y,y) \mbox{ for all } 0\leq y\leq t.
\eeq
Since the system starts from empty, $B(t,y)=B(t,t)$ for all $y>t$. Thus, (\ref{twor9conv}) holds.
Similarly, the following relationship holds from (\ref{r12}) for $\tilde Q(t,y)$ as
\beql{twor9convQ}
\tilde Q(t,y)=\left\{\begin{array}{ll}
\tilde{Q}^\ell(t,0)-\tilde{Q}_\ell(t-y,y), & y \le t \\
\tilde{Q}^\ell(t,0), & y> t.
\end{array} \right.
\eeq
This implies (\ref{twor7conv}) holds. The representation of $X(t)$ in (\ref{twor8conv}) follows from the natural mass balance equation $X(t) = E(t) - B(t) - Q(t)$.
~~~\bsq
\bigskip
\fi
We remark that the coupling property and the existence and uniqueness of the fluid model in Theorem \ref{thm-mr} also indicate that under the same assumptions on the primitives of the queueing model, this dual fluid model tracking residual times is the limit of the associated fluid-scaled prelimit processes. This is because the tightness of these fluid-scaled prelimit processes follows from that of the fluid-scaled prelimit processes tracking elapsed times (see Section 6 in \cite{KR10}) and the coupling property in Proposition \ref{fluiddensity} (noting that \eqref{mrme1} and \eqref{mrme2} also hold for the associated prelimit processes),
and the identification of the limit process follows from Theorem \ref{thm-mr}.
\fi
\section*{Acknowledgement}
We thank the referees for their helpful comments and suggestions, which have helped the quality and exposition of the paper. We also thank Ward Whitt and Kavita Ramanan for many helpful discussions with the paper.
|
\section{Introduction}
The study of the role of individual nuclear reactions in stellar evolution has
been an important field
of research in the last few decades. As a star evolves with time it passes
through burning in different ranges of nuclear mass. At the same time,
different nuclear
processes become important at different time periods of evolution. A
comprehensive study of these processes sheds light on various astrophysical
phenomena.
There are certain astrophysical sites which are responsible for the production
of heavier nuclei beyond iron through the rapid capture of protons on seed
nuclides. In the mass region of our interest there are certain
proton rich naturally occurring nuclei, which are not produced by the
$r$-process or the $s$-process. These are called $p$-nuclei. Proton capture reactions in
certain astrophysical sites can account for the formation of some of these
proton rich nuclides. For example x-ray bursters with a large proton
flux in the peak temperature around 1-3 GK are suitable astrophysical sites for
the production of certain nuclei. To find out the abundance of different
nuclei as well as the evolution of the process in these sites a network
calculation is necessary which involves a large number of reactions.
It is thus imperative to calculate the rates and/or cross sections of these
reactions in different mass ranges. Our group has already calculated the
cross sections and hence the astrophysical S-factors in the mass
range $A=60-100$~\cite{GG,Lahiri1,Lahiri2,Lahiri3}. Some implications of the new
rates has also been investigated in the context of rp-process~\cite{Lahiri4,Lahiri5}.
In the present work, we extend our calculation to the $A=55 - 60$ region.
The rp-process is sensitive to a number of reactions in this region.
The most challenging aspect to look at in these scenarios is that most of the
nuclei involved in those reactions are not produced in the laboratory.
For example, Parikh {\em et al.}~\cite{parikh} have identified proton capture
reactions on $^{56}$Ni and $^{57,59}$Cu targets as important in the rp-process
in certain scenarios.
However, experimental rates are not available for these reactions because
stable targets do not occur in nature.
Hence, one has to depend on theoretical calculations in this domain.
In explosive proton rich environments, such as x-ray bursts, proton capture has
to compete with its inverse, {\em i.e.} photo-disintegration. This competition
results in
waiting points and causes delay of further nucleosynthesis. With temperature,
pressure and proton mass fractions being different at different regions of
these sites as well as being time-varying quantities, incorporation of all
these physical conditions in the nuclear network is a big challenge.
Rauscher {\em et al.}~\cite{Raus1,Raus2} have calculated the
rates for various proton, neutron and $\alpha$-particle induced reactions and their reverse reactions
in Hauser-Feshbach formalism for targets with wide range of atomic numbers and masses and for
a wide range of temperature.
Theoretical
calculations in this mass region essentially utilize the Hauser-Feshbach
formalism where, the optical model potential, a key ingredient, is often
taken in a local or a global form. However, a more microscopic approach
is also possible using
an optical potential constructed utilizing nuclear densities.
If the target is stable, nuclear density is available through electron
scattering. However, in the absence of a stable target, theory remains our sole
guide to describing the density.
It is imperative to test the theoretical
calculations, where experimental data are available, to verify
its applicability. We aim to check the success of microscopic optical
potentials based on mean-field densities in explaining
the available reaction cross sections in this mass region. A good description
depending essentially on theory will allow one to extend
the present method to the critical reactions, which are beyond present day
laboratory capabilities.
A well
defined nucleon-nucleon ($NN$) interaction is of major importance for microscopic
calculation of nucleon-nucleus and nucleus-nucleus potentials used in the
theoretical analysis of different reactions as well as scattering. The optical
model potential is highly successful for explanation of different
branches of nuclear reaction. It can reliably
predict the basic observables such as total and partial cross sections, elastic
scattering angular distributions, etc, even for those target nuclei and for
those energy regions for which no experimental data exist.
We have used the density dependent M3Y interaction by
folding the potential with target radial matter densities.
This interaction has been used in many calculations and has given satisfactory results.
The paper is organized as follows. In the next section, we outline our method
of calculation.
Essentially we construct an optical model potential through
folding an $NN$ interaction with the theoretical density
profile.
For this purpose we
use the relativistic mean field (RMF) theory to obtain the density profile of the targets.
In Sec.~\ref{secresults} the results of our work are discussed in detail.
Finally we summarize our work.
\section{Model calculation}
The RMF approach has proved to be very successful in
describing various nuclear properties such as binding energy of nuclei in
ground states as well as excited states, nuclear density profile, rms charge
radii, deformation, nuclear halo, moment of inertia, etc~\cite{Ring}. It is
considered to be the relativistic generalization of the non-relativistic models
such as Gogny force or Skyrme force Hartree-Fock theory using effective mesonic
degrees of freedom rather than instantaneous forces. The model is basically
based upon two major approximations namely mean-field approximation and no-sea
approximation~\cite {Muller}. The starting point of RMF is
a suitable Lagrangian density that includes the coupling between the nucleon field
and meson field as well as meson self couplings so that the
Lagrangian can successfully describe the properties of finite nuclei as well as the
equation of state (EOS) of nuclear matter. There are
different variations of Lagrangian density as well as different
parameterizations. An accurately calibrated relativistic Lagrangian density,
FSUGold~\cite{Todd}, has been fitted to the charge radii of nuclei. It contains two
additional parameters, compared to conventional RMF models, describing self coupling of vector-isoscalar meson and
coupling between the vector-isovector meson and vector-isoscalar meson. These
two additional parameters significantly affect the softening of the EOS,
the accurate determination of which is needed for the study
of various nuclear properties such as charge radii, masses, etc.
Thus theoretical density profiles are extracted in the RMF approach considering
the FSUGold interaction. The charge density is obtained by convoluting the point
proton density considering the finite size of the nucleus.
\begin {equation}
\rho_{ch}({\mathbf r})=e \int \rho( {\mathbf r \prime})g({\mathbf r}-{\mathbf r\prime})d{\mathbf r\prime}
\end {equation}
where $g(r)$ is the Gaussian form factor given by,
\begin{equation}
g(r)=(a\sqrt{\pi})^{-3} exp(-r^{2}/a^{2})
\end{equation}
where $a$ is a constant whose value is assigned to 0.8 fm.
Using the nuclear density profile we have numerically obtained rms charge radii.
While calculating the charge density or the radius, no attempt has been made to
take the correction due to center of mass into account. Calculations on
harmonic oscillator wave functions show that the correction is small for heavier
nuclei. For example, Quentin has shown\cite{Quentin} that the effect of inclusion of center
of
mass correction in the radius is given by $\delta r/r\approx 0.9/A^{4/3}$.
Hence, we do not expect the density profile to be affected significantly
due to this approximation.
The M3Y interaction~\cite{M1,M2} is based on a realistic G-matrix which in turn is
constructed in a harmonic oscillator representation averaging over a range of energies
as well as densities. It has no explicit density dependence nor energy
dependence. Although in most cases these averages do not matter
producing satisfactory results, in few cases it becomes necessary to
incorporate explicit density dependence into M3Y interaction and then it is
named as density dependent M3Y (DDM3Y) effective interaction~\cite{DM}.
Low energy proton capture reactions are highly sensitive to nuclear radius as
well as density. In the present work we have used density dependent M3Y
Reid-Elliot effective nucleon-nucleon interaction within a folding model
prescription \cite{Lahiri1}. The density dependence is incorporated in the same way as
suggested in Refs.~\cite{DN1,DN2}.
Further, we have included a spin-orbit term into the potential considering Scheerbaum
prescription~\cite{Scheer} which has been coupled with the phenomenological complex potential
depths. These depths are functions of energy which are assigned standard values as in
Lahiri {\em et al}.~\cite{Lahiri1}. These values are kept unaltered throughout our present work.
We have incorporated the density dependent M3Y interaction within the TALYS1.4
code~\cite{Talys} and performed a Hauser-Feshbach (HF) calculation.
We have chosen Goriely's microscopic level densities and
Hartree Fock Bogolyubov model for $E1$ $\gamma$-ray strength function.
As seen in our previous calculations~\cite{Lahiri1,Lahiri2}, these choices can
explain the experimental results more accurately. All these options are
available in the code. We have also included the effect of the width fluctuation
correction which has a significant impact at low incident energies. Up to
30 discrete levels are included for both target and residual nuclei, which
are considered in Hauser-Feshbach decay and $\gamma$-ray cascade. We also include
a maximum of 30 discrete levels for the nuclei resulting from binary emission in
Hauser-Feshbach decay and $\gamma$-ray cascade. HF calculations are
done with full $j,l$ coupling.
We have incorporated the density data obtained from RMF approach to obtain the
optical model potential.
Because of rapid variation of cross-section with energy in the low energy
region, it is difficult to compare the theory and experiment. A standard
alternative way is to compare another important quantity instead of
cross-section, namely astrophysical S-factor \cite{Rolfs}.
The proton capture reactions in astrophysical sites occur within a narrow energy window~\cite{Rolfs}. This effective energy window approximately of Gaussian shape around a peak (known as Gamow peak) is known as Gamow window.
The expressions for the Gamow peak and Gamow width in a practical form are given respectively as,
\begin{equation}
E_{0}=0.1220(Z_{t}^{2}Z_{p}^{2}\mu T_{9}^{2})^{1/3} MeV
\end{equation}
\begin{equation}
\Delta = 0.2368 (Z_{p}^{2}Z_{t}^{2}\mu T_{9}^{5})^{1/6}
\end{equation}
wherei $\mu$ is the reduced mass and $T_{9}$ denotes the temperature in GK.
Thus most of the astrophysically important reactions occur within a narrow energy
window $E_{0}-{\Delta}/{2}$ to $E_{0}+{\Delta}/{2}$.
We see that for ($p,\gamma$) reactions on stable isotopes in the mass range
55-60, the Gamow window lies between 1 MeV to 3 MeV for temperature around
3 GK. Hence, we have carried out our calculation in this low energy window
and compared our results with the measured data where available. In calculating the Gamow
peak, Gamow width and hence Gamow window we have taken the masses from Audi
{\em et al}.~\cite{Audi}.
\section{Results}
\label{secresults}
\begin{figure*}[htp]
\includegraphics[scale=0.475]{dens.eps}
\caption{\footnotesize Theoretical density profile of various nuclei in mass
55-60 region from RMF approach compared with experimental data obtained by
elastic electron scattering taken from Wohlfahrt et. al.~\cite{Wohl}. The solid
lines denote the theoretical result and discrete points with error-bars
represent the experimental data. In most of the cases, especially for lower
radii values, the errors associated are smaller than the dimension of the empty
circles.}
\label{fig:den}
\end{figure*}
Because the optical model is dependent on the density profile of the nucleus, we calculate
the density and the charge radii of nuclei in this mass region using RMF formalism. The
theoretical density values are plotted as a function of radius and compared with available
experimental values in Fig.~\ref{fig:den}. As can be seen the agreement is extremely good.
The experimental data are taken from Wohlfahrt {\em et al}~\cite{Wohl}.
We also compare all the available rms charge radii values with theoretical results in
table~\ref{tab:exp2}. The experimental values are taken from Angeli {\em et al}.~\cite{Ang}.
It can be seen that the RMF calculation has an excellent predictive power,
the relative difference between theory and experiment in all cases being less than 0.5\%.
\begin{table}[htp]
\center
\caption{\footnotesize Charge radii of various nuclei extracted in the RMF approach
compared with measured values from Angeli {\em et al}.~\cite{Ang}}
\label{tab:exp2}
\begin{tabular}{ c c c c c}
\hline
Nucleus & \multicolumn{2}{c} {Charge radius(fm)} \\
& Theory & Experiment. \\
\hline
$^{55}$Mn & 3.7057 & 3.7057 \\
$^{56}$Mn & 3.7189 & 3.7146 \\
$^{56}$Fe & 3.7361 & 3.7377 \\
$^{57}$Fe & 3.7497 & 3.7532 \\
$^{58}$Fe & 3.7634 & 3.7745 \\
$^{59}$Co & 3.7924 & 3.7875 \\
$^{58}$Ni & 3.7916 & 3.7757 \\
$^{60}$Ni & 3.8193 & 3.8118 \\
\hline
\end{tabular}
\end{table}\\
\begin {figure*}[htp]
\includegraphics[scale=0.45]{Mn55.eps}
\includegraphics[scale=0.45]{Fe58.eps}
\includegraphics[scale=0.45]{Co59.eps}
\includegraphics[scale=0.45]{Ni58.eps}
\includegraphics[scale=0.45]{Ni60.eps}
\caption{\footnotesize Comparison of theoretical astrophysical S-factors
calculations in the present work with experimental data for proton capture
on (a)$^{55}$Mn, (b)$^{58}$Fe, (c)$^{59}$Co, and (d)$^{58}$Ni, (e)$^{60}$Ni.
See text for the explanation of different symbols and the Supplemental Material in
Ref.~\cite{supply} for numerical values of S-factor.}
\label{fig:ddm3y}
\end{figure*}
We have tried to set a definite normalization for the optical model potential
that fits all the reaction data in the concerned mass region. The potential
obtained by folding has been multiplied by the normalization constant 2.0 to
get the real part of the potential. The DDM3Y interaction does not have any imaginary
part. We have multiplied the folding potential by the normalization
constant 1.4 to obtain the imaginary part of the optical potential.
These final parameters have been obtained after many trials to ensure a
reasonable agreement with experimental data for all the known low energy proton
capture reactions in the mass region of our interest.
Although a single
normalization can not reproduce the experimental data excellently for all
reactions in the region, {\em i.e.} each individual reaction may have different
normalization for best matching with measurement, it is necessary to consider a
single definite normalization to extend the work to unknown nuclei in the mass
region for which no experimental data exist.
We note that the fitted parameters for the present mass range differ
from the neighbouring mass region in our earlier calculation. This is possibly
due to the fact that the mass selected in the present calculation is lighter
than our previous regions.
Possibly, the larger depths of the potential is required to adjust for the low
mass region.
The comparison of S-factors obtained after incorporating DDM3Y interaction using
the above normalization constants with experimental data are shown in fig.
\ref{fig:ddm3y}. The numerical values of the S-factors and the reaction rates are
given in Supplemental Material~\cite{supply}. The experimental data are taken from Ref~\cite{Mn,Fe,Co} for
$^{55}$Mn, $^{58}$Fe and $^{59}$Co, respectively. For $^{58}$Ni and $^{60}$Ni,
experimental data are taken from Refs.~\cite{Ni581,Ni582,Ni583} and
Refs.~\cite{Ni601,Ni602,Ni582}, respectively.
In many cases the experimental data are very old. Errors are also not
available in some cases. For $^{55}$Mn, $^{58}$Fe and $^{59}$Co,
circles represent the experimental data. For $^{58}$Ni, triangles,
squares and circles represent the data from Refs.~\cite{Ni583,Ni581,Ni582}, respectively.
For $^{60}$Ni there are three different sources of data~\cite{Ni601,Ni602,Ni582} which are denoted by
squares, circles and diamonds, respectively.
\begin{figure*}[hbtp]
\includegraphics[scale=0.458]{Ni56.eps}
\includegraphics[scale=0.458]{Cu57.eps}
\includegraphics[scale=0.458]{Cu59.eps}
\caption{\footnotesize Comparison of astrophysical rates from the present calculation
with NON-SMOKER results from Rauscher {\em at al.}~\cite{Raus1,Raus2} for the
reactions (a)$^{56}$Ni($p,\gamma)^{57}$Cu,
(b) $^{57}$Cu($p,\gamma)^{58}$Zn and (c) $^{59}$Cu($p,\gamma)^{60}$Zn.
See text for details \label{fig:rate} and the Supplemental Material in
Ref.~\cite{supply} for numerical values of rates.}
\end{figure*}
In all cases the solid line denotes the theoretical DDM3Y result.
In $^{55}$Mn, there are certain ambiguities in
experimental data, especially in the energy range between 1.3 to 1.6 MeV. The
experiment was done using Ge(Li) detector by integrated beam current method
more than three decades ago. However, errors are not associated with most of the
data points. Only four data points in the energy range of our interest have
errors associated with those.
Our calculations give an excellent description of
experimental data for $^{58}$Fe. The experiment for $^{58}$Fe was done using
Ge(Li) detector and the data was compared with statistical model predictions~\cite{Fe}.
For $^{59}$Co again there are large fluctuations in experimental
data. Butler {\em et al.}~\cite{Co} stated that they had observed several resonances
in the reaction $^{59}$Co$(p,\gamma)^{60}$Ni but the resonances were too close
to be resolved clearly.
Our calculation for $^{58}$Ni overpredicts the measurement of Tingwell {\em et al.}~\cite{Ni583}
by a factor of $\sim$ 2.5, approximately. This experiment was
carried out by both beam current integrated method and single target irradiation method using Ge(Li) detector.
Tingwell {\em et al.} also compared their data with
statistical model calculations. They found that their statistical calculation
overestimates the measurement by a factor of $\sim$ 2.5 for $^{58}$Ni, which
agrees with our results. Cheng {\em et al.}~\cite{Ni581} also measured cross section
for this reaction using the activation technique. Except in the energy range
$\sim$ 1.4-1.8 MeV, where the measurement itself has large discrepancies, the
data agree more or less well with our theoretical calculations.
For the reaction $^{60}$Ni$(p,\gamma)^{61}$Cu, Tingwell {\em et al.} themselves
compared the experimental results with the statistical model predictions and
showed that normalizing the optical model imaginary well depth for Ni isotopes
by a factor of 1.5 leads to a better agreement between theory and experiment~\cite{Ni602}.
Our calculation, in the case of $^{60}$Ni, overpredicts the
experimental data of Tingwell {\em et al.}~\cite{Ni602} by a factor $\sim$ 1.5,
whereas it underpredicts the data of Krivonosov {\em et al.}~\cite{Ni582} by a factor $\sim$ 0.35.
With the above normalization, we have calculated the rates for $(p,\gamma)$
reactions identified as important by Parikh {\em et al}.~\cite{parikh}.
The calculated rates are compared with NON-SMOKER~\cite{Raus1,Raus2,Raus3} rates.
The NON-SMOKER results are from a HF calculation based on masses from
experimental measurements and calculation in the Finite Range
Droplet Model~\cite{FRDM}. Other details of the calculation can be obtained
from the references. The results have been plotted in figure 3.
We see that in the range 1-4 GK, the NON-SMOKER results differ
from the present calculations significantly. For the $^{56}$Ni($p,\gamma)$ reaction, although
the results agree at low temperature, at higher temperature the NON-SMOKER
rates are larger. On the other hand, for the other two reactions, {\em viz.}
$^{57}$Cu($p,\gamma)$ and $^{59}$Cu($p,\gamma)$, our calculation
predicts a significantly larger rate throughout the temperature range.
It will be interesting to see the effects of these results on
astrophysical scenarios.
\section{Summary}
Low energy $(p,\gamma)$ reactions are studied in a semi-microscopic approach
in the HF formalism and compared with experiments in the mass region 55-60.
Radial density profiles are obtained using the RMF approach and are folded
with the DDM3Y NN interaction to obtain semi-microscopic optical potentials.
Both the real and imaginary depths of the potential are normalized to obtain a
good agreement between theory and experiment. The S-factors for $(p,\gamma)$
reactions are evaluated in the Gamow window corresponding to 3 GK. We have
not modified the parameters to fit individual reactions as our aim is to
construct a framework for calculation of astrophysical reactions involving
unstable nuclei. Rates for important astrophysical reactions calculated
in the present approach differ significantly from NON-SMOKER rates. The key
feature of our work is that we have taken all nuclei
in the same footing and same methodology has been used for all of them to avoid systematic error.
\section*{Acknowledgment}
The authors acknowledge the financial support provided by UGC(DRS), DST and the
University of Calcutta.
|
\section{Introduction}
Today the Universe appears spatially flat undergoing an accelerated expansion. There are many measurements proving this pictures \cite{riess, ast, clo, ber, spe, carrol, staro, sahini}. According to the successful cosmological model \cite{carrol, staro, sahini}, there are two main ingredients in this scenario, namely Dark Matter (DM) and the cosmological constant $\Lambda$ (Dark Energy). On the galactic scales, the evolution is driven by the usual Newtonian gravitational potential, but it needs hypothesizing the existence of DM to obtain a good experimental agreement. A good model for the galactic distribution of DM, in the framework of General Relativity (GR), is the Navarro-Frenk-White model (NFW model) \cite{navarro}.
However in recent years, the effort to give a physical explanation to the cosmic acceleration has attracted an amount of interest in so called Fourth Order Gravity (FOG), and particularly the $f(R)$-Gravity, where $f$ is a generic function of Ricci scalar $R$. These alternative models have been considered as a viable mechanism to explain the cosmic acceleration. Apart the cosmological dynamics, a systematic analysis of such theories were performed at short scale and in the low energy limit \cite{olmo1, olmo2, olmo3, Damour:Esposito-Farese:1992, clifton, odintsov, newtonian_limit_fR, PRD, stelle, schm, FOG, Stabile_Capozziello}.
In particular the paper \emph{Most General Fourth Order Theory of Gravity at Low Energy} \cite{FOG} analyzed the gravitational potential, induced by a $f(X,Y,Z)$-Gravity, where for sake of simplicity we set $X\,=\,R$, $Y\,=\,R^{\alpha\beta}R_{\alpha\beta}$ and $Z\,=\,R^{\alpha\beta\gamma\delta}R_{\alpha\beta\gamma\delta}$, generalizing the Hilbert Einstein lagrangian. The added quantities are the Ricci tensor $R_{\mu\nu}$ and the Riemann tensor $R_{\mu\nu\alpha\beta}$. As astrophysical application the modified potential has been used to build the rotation curves for the Milky Way and NGC 3198 \cite{StSc}. In this paper any galactic component (bulge, disk and DM component) required an onerous numerical computation since the Gauss theorem is not applicable in the FOG. The aim of the present paper is to point out the fundamental topics of the adopted strategy using the software \emph{Mathematica}$^\circledR$.
Our analysis is then organized as follows. In section \ref{theory} we report the fundamental topics of the fourth order gravity: the field equations and their newtonian approximation, the solution for the gravitational potential and the mathematical models for the galactic componets. In section \ref{computation} we build we build the code for the numerical simulation and in section \ref{data} there is the data fit between our theoretical curves and the data of the rotation curve of the Milky Way and the galaxy NGC 3190. Finally in section \ref{conclusions} we report the conclusions.
\section{The galactic rotation curves in the framework of $f(X,Y,Z)$-Gravity}\label{theory}
Let us start with a general class of FOG given by the action
\begin{eqnarray}\label{FOGaction}
\mathcal{A}\,=\,\int d^{4}x\sqrt{-g}\biggl[f(X,Y,Z)+\mathcal{X}\mathcal{L}_m\biggr]
\end{eqnarray}
where $f$ is an unspecified function of curvature invariants.
The term $\mathcal{L}_m$ is the minimally coupled ordinary matter
contribution. In the metric approach, the field equations are
obtained by varying (\ref{FOGaction}) with respect to
$g_{\mu\nu}$. We get
\begin{widetext}
\begin{eqnarray}\label{fieldequationFOG}
\begin{array}{ll}
f_XR_{\mu\nu}-\frac{f}{2}g_{\mu\nu}-f_{X;\mu\nu}+g_{\mu\nu}\Box
f_X+2f_Y{R_\mu}^\alpha
R_{\alpha\nu}-2[f_Y{R^\alpha}_{(\mu}]_{;\nu)\alpha}\\\\\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,
\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,+\Box[f_YR_{\mu\nu}]+[f_YR_{\alpha\beta}]^{;\alpha\beta}g_{\mu\nu}
+2f_ZR_{\mu\alpha
\beta\gamma}{R_{\nu}}^{\alpha\beta\gamma}-4[f_Z{{R_\mu}^{\alpha\beta}}_\nu]_{;\alpha\beta}\,=\,
\mathcal{X}\,T_{\mu\nu}
\\\\
f_XR+2f_YR_{\alpha\beta}R^{\alpha\beta}+2f_ZR_{\alpha\beta\gamma\delta}
R^{\alpha\beta\gamma\delta}-2f+\Box[3
f_X+f_YR]+2[(f_Y+2f_Z)R^{\alpha\beta}]_{;\alpha\beta}\,=\,\mathcal{X}\,T
\end{array}
\end{eqnarray}
\end{widetext}
where $f_X\,=\,\frac{\partial f}{\partial X}$, $f_Y\,=\,\frac{\partial f}{\partial Y}$,
$f_Z\,=\,\frac{\partial f}{\partial Z}$, $\Box={{}_{;\sigma}}^{;\sigma}$ and
$\mathcal{X}\,=\,8\pi G$. $T_{\mu\nu}\,=\,-\frac{1}{\sqrt{-g}}\frac{\delta(\sqrt{-g}\mathcal{L}_m)}{\delta g^{\mu\nu}}$ is the energy-momentum tensor of matter and $T$ is
its trace. The second line of (\ref{fieldequationFOG}) is the
trace of the first one.
In the case of weak field and slow motion we consider the field
equation in the so called Newtonian limit of theory. For our aim
we can consider the metric tensor approximated as follows (for
details, see \cite{newtonian_limit_fR, rew, landau, PRD})
\begin{eqnarray}\label{metric_tensor_PPN}
g_{\mu\nu}\,=\,\begin{pmatrix}
1+2\,\Phi(t,\mathbf{x})& 0 \\
\\
0 & -\delta_{ij}\end{pmatrix}
\end{eqnarray}
where $\Phi$ is the gravitational potentials and $\delta_{ij}$ is
the Kronecker delta. The set of coordinates adopted is $x^\mu\,=\,(t,x^1,x^2,x^3)\,=\,(t,\mathbf{x})$. By
introducing the quantities
\begin{eqnarray}\label{mass_definition}
\begin{array}{ll}
{m_1}^2\,\doteq\,-\frac{f_X(0)}{3f_{XX}(0)+2f_Y(0)+2f_Z(0)}\\\\
{m_2}^2\,\doteq\,\frac{f_X(0)}{f_Y(0)+4f_Z(0)}
\end{array}
\end{eqnarray}
we get three differential equations for the curvature invariant $X$ and the gravitational potentials $\Phi$, $\Psi$
\begin{eqnarray}\label{NL-field-equation_2}
\begin{array}{ll}
(\triangle-{m_2}^2)\triangle\Phi+\biggl[\frac{{m_2}^2}{2}-\frac{{m_1}^2+2{m_2}^2}{6{m_1}^2}\triangle\biggr]
X\,=\,-{m_2}^2\mathcal{X}\,\rho\\\\
(\triangle-{m_1}^2)X\,=\,{m_1}^2\mathcal{X}\,\rho
\end{array}
\end{eqnarray}
where $\triangle$ is the Laplacian in the flat space and $\rho$ is the matter density \cite{FOG}. Further we assumes $f_X(0)\,=\,1$ without loss of generality.
By choosing ${m_1}^2\,,{m_2}^2\,>0$ and introducing $\mu_{1,2}\,\doteq\,\sqrt{|{m_{1,2}}^2|}$ the gravitational
potential in the case of point-like source ($\rho\,=\,M\,\delta(\mathbf{x})$) is given by
\begin{eqnarray}\label{sol_pointlike}
\Phi_{pl}(\mathbf{x})\,=\,-\,\frac{GM}{|\textbf{x}|}\biggl[1+\frac{1}{3}\,e^{-\mu_1|\mathbf{x}|}-\frac{4}{3}\,
e^{-\mu_2|\mathbf{x}|}\biggr]
\end{eqnarray}
while in the case of generic matter source distribution we perform the change $\Phi\,\rightarrow\,\int d\Phi$. The passage from the pointlike source to extended one is correct only in the Newtonian limit since a such limit corresponds also to the linearized version of theory.
The motion of bodies is given by geodesic equation
\begin{eqnarray}\label{geodesic}
\frac{d^2\,x^\mu}{ds^2}+\Gamma^\mu_{\alpha\beta}\frac{dx^\alpha}{ds}\frac{dx^\beta}{ds}\,=\,0
\end{eqnarray}
where $ds\,=\,\sqrt{g_{\alpha\beta}dx^\alpha dx^\beta}$ is the relativistic distance and $\Gamma^\mu_{\alpha\beta}$ are the Christoffel symbols. In the Newtonian limit we obtain
\begin{eqnarray}
\frac{d^2\,\mathbf{x}}{dt^2}\,=\,-\nabla\Phi(\mathbf{x})
\end{eqnarray}
where the study of motion is very simple in particular cases of symmetry. For example the case of stationary motion on the circular orbit we get
\begin{eqnarray}\label{stationary_motion_2}
v_c(|\mathbf{x}|)\,=\,\sqrt{|\mathbf{x}|\,\frac{\partial\Phi(\mathbf{x})}{\partial|\mathbf{x}|}}
\end{eqnarray}
The distribution of mass can be modeled simply by introducing two sets of coordinates: the spherical coordinates $(r,\theta,\phi)$ and the cylindrical coordinates $(R,\theta,z)$. An useful mathematical tool is the Gauss flux theorem for Gravity. Since the Newtonian mechanics satisfies this theorem and, by thinking to a spherical system of mass distribution, we get, from (\ref{stationary_motion_2}), the equation
\begin{eqnarray}\label{circular_velocity}
{v_c(r)}\,=\,\sqrt{\frac{G\,M(r)}{r}}\,=\,\sqrt{\frac{4\pi G}{r}\int_0^rdy\,y^2\,\rho(y)}
\end{eqnarray}
where $M(r)$ is the only mass enclosed in the sphere with radius $r$. The Green function of the $f(X,Y,Z)$-Gravity ($\neq\,|\mathbf{x}-\mathbf{x}'|^{-1}$), instead, does not satisfy the theorem \cite{Stabile_Capozziello}. In this case we must consider directly the gravitational potential
\begin{eqnarray}\label{sol_gen}
\begin{array}{ll}
\Phi(\mathbf{x})\,=\,-\,G\int d^3\mathbf{x}'\frac{\rho(\mathbf{x}')}{|\mathbf{x}-\mathbf{x}'|}
\biggl[1+\frac{1}{3}\,e^{-\mu_1|\mathbf{x}-\mathbf{x}'|}\\\\
\qquad\qquad\qquad\qquad\qquad\qquad\qquad-\frac{4}{3}\,e^{-\mu_2|\mathbf{x}-\mathbf{x}'|}\biggr]
\end{array}
\end{eqnarray}
Apart the mathematical difficulties incoming from the research of gravitational potential for a given mass distribution, the non-validity of Gauss theorem implies, for example, that a sphere can not be reduced to a point. In fact the gravitational potential generated by a ball (also with constant density) is depending also on the Fourier transform of ball \cite{Stabile_Capozziello}. Only in the limit case where the radius of ball is small with respect to the distance we obtain the simple expression (\ref{sol_pointlike}).
We remember that in the potential (\ref{sol_gen}) we can distinguish the contributions of the bulge, the disk and the (eventual) Dark Matter. $r$ is the radial coordinate in the spherical system, while $R$, $z$ are respectively the radial coordinate in the plane of disc and the distance from the plane then we have the geometric relation $r\,=\,\sqrt{R^2+z^2}$. The main item is the choice of models of matter distribution. The more simple model characterizing the shape of galaxy is the following
\begin{eqnarray}\label{density_3}
\begin{array}{ll}
\rho_{bulge}(r)\,=\,\frac{M_b}{2\,\pi\,{\xi_b}^{3-\gamma}\,\Gamma(\frac{3-\gamma}{2})}\frac{e^{-\frac{r^2}{{\xi_b}^2}}}{r^\gamma}\\\\
\sigma_{disk}(R)\,=\,\frac{M_d}{2\pi\,{\xi_d}^2}\,
e^{-\frac{R}{\xi_d}}\\\\
\rho_{DM}(r)\,=\,\frac{\alpha\,M_{DM}}{\pi\,(4-\pi){\xi_{DM}}^3}\,\frac{1}{1+\frac{r^2}{{\xi_{DM}}^2}}
\end{array}
\end{eqnarray}
where $\Gamma(x)$ is the Gamma function, $0\,\leq\,\gamma\,<\,3$ is a free parameter and $0\,\leq\,\alpha\,<\,1$ is the ratio of Dark Matter inside the sphere with radius $\xi_{DM}$ with respect the total Dark Matter $M_{DM}$. Moreover the couples $\xi_b$, $M_b$ and $\xi_d$, $M_d$ are the radius and the mass of the bulge and the disc.
\section{Numerical analysis}\label{computation}
The computational analysis here described is referred to the study of the rotation curve (\ref{stationary_motion_2}) which can be replaced as follows
\begin{eqnarray}\label{stationary_motion_3}
v(r,R,z)=\sqrt{r\frac{\partial}{\partial r}\Phi(r,R,z)}
\end{eqnarray}
where $\Phi(r,R,z)$ is the gravitational potential
\begin{widetext}
\begin{eqnarray}\label{potential}
&&\Phi(r,R,z)\,=\,\frac{4\pi G}{3}\,\biggl[\frac{1}{r}\int_0^\infty
dr'\,\rho_{bulge}(r')\,r'\,\biggl(3\,\frac{|r-r'|-r-r'}{2}
-\frac{e^{-\mu_1|r-r'|}-e^{-\mu_1(r+r')}}{2\,\mu_1}
+2\,\frac{e^{-\mu_2|r-r'|}-e^{-\mu_2(r+r')}
}{\mu_2}\biggr)\biggr]
\nonumber\\\nonumber\\&&
+\frac{4\pi G}{3}\,\biggl[\frac{1}{r}\int_0^{\Xi}
dr'\,\rho_{DM}(r')\,r'\,\biggl(3\,\frac{|r-r'|-r-r'}{2}
-\frac{e^{-\mu_1|r-r'|}-e^{-\mu_1(r+r')}}{2\,\mu_1}
+2\,\frac{e^{-\mu_2|r-r'|}-e^{-\mu_2(r+r')}
}{\mu_2}\biggr)\biggr]
\nonumber\\\nonumber\\&&
-2\,G\,\biggr\{\int_0^\infty
dR'\,\sigma_{disc}(R')\,R'\,\biggl(\frac{\mathfrak{K}(\frac{4RR'}{(R+R')^2+z^2})}{\sqrt{(R+R')^2+z^2}}
+\frac{\mathfrak{K}(\frac{-4RR'}{(R-R')^2+z^2})}{\sqrt{(R-R')^2+z^2}}\biggr)+\int_0^\infty
dR'\,\sigma_{disc}(R')\,R'\,
\\\nonumber\\&&
\times\int_0^{\pi} d\theta'\frac{1}{3\,\sqrt{(R+R')^2+z^2-4RR'\cos^2\theta'}}
\biggl[e^{-\mu_1\sqrt{(R+R')^2+z^2-4RR'\cos^2\theta'}}
-4\,e^{-\mu_2\sqrt{(R+R')^2+z^2-4RR'\cos^2\theta'}}\biggr]\biggr\}\nonumber
\end{eqnarray}
\end{widetext}
and $\mathfrak{K}(x)$ is the Elliptic function. The parameters $\mu_1$ and $\mu_2$ are the free parameters in the theory and only by fitting process can be fixed. A sensible item is the choice of distance $\Xi$ on the which we are observing the rotation curve. In fact all models for the Dark Matter component are not limited and we need to cut the upper value of integration in (\ref{potential}).
A further distinction are the contributions to the potential coming from terms of General Relativity (GR) origin and terms of Forth Order Gravity (FOG) origin. Finally our aim is the numerical evaluation of the rotation curve in the galactic plane
\begin{eqnarray}\label{velocity}
v(R,R,0)=\sqrt{R\frac{\partial}{\partial R}\Phi(R,R,0)}
\end{eqnarray}
The first step, after the definition of the numerical values for the parameters, has been the building of the velocity starting from the derivative of the potential. For this, we need to build the definitions of the contributions to the density coming from the bulge, the disk and the dark matter (respectively \texttt{$\rho$b[r\_]}, \texttt{$\sigma$d[r\_],\texttt{$\rho$DM[r\_]}} in the full code present in appendix), together with the already mentioned splitting in the GR contributions and FOG contributions (respectively \texttt{TerGR[x\_,y\_]} and \texttt{TerYu[x\_,y\_]} in the code).
The derivative and integration operations commute, then we ``transport'' the derivative in the the integrand and then we make the integration. We found this computationally more rapid. We turning off the warning messages concerning the numerical integrations with the following commands:
\begin{quote}
\verb|Off[NIntegrate::inmur]|
\verb|Off[NIntegrate::slowcon]|
\verb|Off[NIntegrate::ncvb]|
\verb|Off[NIntegrate::eincr]|
\end{quote}
The first \verb|Off| is justified since all the variables definitions are made with the ``SetDelayed'' command (:=) that postpones the numerical evaluation of the integral making it not immediately numerical. The second turn off the message of slow convergence of the integration and making so, we avoid a long series of warning messages. The third avoid to the program to inform us of the need to use a larger number of recursive refinements in the computation. In effect, we increase the refinements with the command \verb|MaxRecursion| $\rightarrow$ \verb|20| for all the integrals, but this is not sufficient by itself to turn off the warnings. To do this, we need a bigger number than 20, but the computation became excessively slow and the final result remains practically unchanged. Since the computation is faced with oscillating error estimation, that explains the origin of the last warning message. We need to add to the numerical integration, the command \verb|Method -> GlobalAdaptive|. Still here, the warning message has to be manually closed since an improvement of \verb|Method| slows down the computation without an effective change in the results.
An interesting thing to note, as it is possible to see in the complete code in Appendix A and there noted with the comment (*\text{$\leftarrow $}*), is that in the definition of the derivative by means of mute variables, it need not a ``SetDelayed'' command, but a simple ``='' command, otherwise the code is unable to make the computation.
\section{Data fit}\label{data}
The next and more interesting step, is the comparison of the experimental data and what predicted by our model. From the literature cited in \citep{StSc} we can obtain the galactic speed values as function of the distance from the center and the corresponding errors. For instance, we show in some detail the manipulation of the data coming from the analysis of \citep{Stark}, concerning the external part of the Milky Way.
We start copying the data listed in the table 1 of \citep{Stark} in a table called \verb|list1|. Then we follow the prescriptions given by the authors with the introduction of new variables. As it is possible to see in the code present in Appendix B, we preserve the same notations and with the command \verb|Append| we add to the initial \verb|list1| the new variables. For instance, for the $R$ variable, with the command \verb|MapThread[Append,{list1,R}]|, we obtain a new table, here \verb|list2|, with one more column, the $R$'s valuer. And so on with the other variables. We introduce with the usual definitions, the errors on these derived quantities, here written as $\sigma x$ and added to the table. Then $\sigma$R and $\sigma\theta$ are, respectively, the error bars on the radius (the distance from the galactic center) and on the corresponding speeds and we process them together with the data in order to form the list with the values measured (or derived) and the errors on $x$ and $y$. In order to obtai a plot with the error bars, we need to load the right package for this with the command \verb|Need["ErrorBarPlot"`]| and this make us able to to make an \verb|ErrorListPlot|. With the plot of \verb|list10|, we obtain a plot with the error bars only on y. A little bit more complicate procedure is need in order to obtain the bars on $x$ too. Indeed we need to build a list of values like \verb|{{x,y},ErrorBar{err x,err y}}| and this is done with the procedure shown in the last rows of Appendix B. In figure \ref{Stark} it is shown the result.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.8]{fig06.eps}\\
\caption{ErrorListPlot of the experimental data.}
\label{Stark}
\end{figure}
Similar procedures for the others two part of the Milky Way data and for the NGC 3190 data. At the end, the three experimental data for the Milky Way, each corresponding to a given range of distance from the center of the galaxy, are put together in order to obtain the complete galactic rotation curve.
At this point we proceeded following two strategies.
The first one, the faster, has been to overlap the theoretical graphs with the experimental one using the command \verb|Show|. In this case, the values of parameters in the densities (bulge, Dark Matter and disk) and of reduced masses, $\mu_1$ and $\mu_2$, are chosen by a direct overlap of the graphs.
The second strategy, more rigorous and slower, is the fit procedure. In Appendix C is present the part of code of interest. In this case, in the code of the galactic rotation curve, we fix all other parameters except the ``masses'' $\mu_1$ and $\mu_2$. These variables are the values that must be found whit the find fit procedure.
We note that the \verb|FindFit| procedure uses the parameter constraints option. In this way, it is possible to eliminate all the solutions not physically allowed and to find the values obtained by the direct investigation, that is the first strategy, $\mu_1=10^{-2}\,a^{-1}$, $\mu_2=10^{2}\,a^{-1}$ where $a$ is the characteristic scale length fixed to the value of 1 Kpc. Obviously, an increase of the number of parameters to be found with this procedure, increase he time of the computation.
\section{Conclusions}\label{conclusions}
In this paper we present a study of the galactic rotation curve when a FOG is considered. The purely theoretical aspects have been fully exposed in \citep{StSc}. In the present work, after an obvious theoretical introduction, useful to remember the hypothesis used, we focus our attention on the salient points of the code that us permitted the computational analysis. With this program, we test the validity of our model of galactic rotation curve and the agreement of the experimental data of two galaxies, the Milky Way and the galaxy NGC 3190, with our model.
In order to make the explanation as complete as possible and to contextualize the several pieces of code examined, we show, in Appendix A, the full code corresponding to the plot of the figure \ref{plot_1_PRD}, that is the code for a galaxy whose components are the bulge, the disk and the Dark Matter. The code referring also to the study of the galaxy NGC 3190 is exactly the same with the exclusion of the part of code referring to the bulge. As it is possible to see from figure \ref{plot_2_PRD}, the agreement of our model with the experimental data of the Milky Way is very good. Only for very low values of the distance $R$ the agreement is not perfect. This suggest us that we only need an improvement of the parameters in the code, maintaining the code itself essentially unchanged.
The complete code that refers to the data analysis is omitted since, apart the obvious introduction of the data coming from the cited literature, the complete program is a mere reply of what presented in Appendix B. \\
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.7]{fig09.eps}\\
\caption{Plot of the galactic rotation curve by using the full program for Milky Way (present in Appendix A). The cases are the following: GR (dashed line), GR$+$DM (dashed and dotted line), FOG (solid line), FOG$+$DM (dotted line). The values of masses are $\mu_1\,=\,10^{-2}\,\text{Kpc}^{-1}$ and $\mu_2\,=\,10^2\,\text{Kpc}^{-1}$ \citep{StSc}}
\label{plot_1_PRD}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.7]{fig10.eps}\\
\caption{Superposition of theoretical behaviors GR (dashed line), GR$+$DM (dashed and dotted line), FOG (solid line), FOG$+$DM (dotted line) on the experimental data for Milky Way. The values of masses are $\mu_1\,=\,10^{-2}\,\text{Kpc}^{-1}$ and $\mu_2\,=\,10^2\,\text{Kpc}^{-1}$ \citep{StSc}.}
\label{plot_2_PRD}
\end{figure}
|
\part{\mathcal{P}}
\global\long\def\surface{\mathcal{S}}
\global\long\def\stress{X}
\global\long\def\strain{\chi}
\global\long\def\strech{\varepsilon}
\global\long\def\cbnd{d}
\global\long\def\cochain{X}
\global\long\def\Lip{\mathfrak{L}}
\global\long\def\SS#1{\mathcal{L}_{\mathfrak{s}}\left(#1\right)}
\global\long\def\Limb{\Lip_{\mathrm{Em}}}
\global\long\def\Lmap{\mathcal{F}}
\global\long\def\Emb#1{\text{Emb}\left(#1\right)}
\global\long\def\mol#1{\Phi_{#1}*}
\global\long\def\reals{\mathbb{R}}
\global\long\def\too{\longrightarrow}
\global\long\def\Vs{V}
\global\long\def\we{\wedge}
\global\long\def\Vr{\Vs_{r}}
\global\long\def\Vrd{\Vs^{r}}
\global\long\def\spt{\mathrm{spt}}
\global\long\def\imag#1{\mathrm{image}\left(#1\right)}
\global\long\def\emb{\varphi}
\global\long\def\emr{g}
\global\long\def\refi{\mathfrak{S}}
\global\long\def\sform#1#2{D^{#1}\left(#2\right)}
\global\long\def\currents#1#2{D_{#1}\left(#2\right)}
\global\long\def\Fmass#1{M\left(#1\right)}
\global\long\def\Flmass#1#2{M_{#1}\left(#2\right)}
\global\long\def\Fnormal#1{N\left(#1\right)}
\global\long\def\Fcomass#1{\left\Vert #1\right\Vert _{0}}
\global\long\def\var#1{\mathrm{Var}\left(#1\right)}
\global\long\def\ball#1#2{B\left(#1,#2\right)}
\global\long\def\mtb{\Gamma}
\global\long\def\dns#1#2{d\left(#1,#2\right)}
\global\long\def\compact{K}
\global\long\def\norm#1#2{\left\Vert #1\right\Vert _{#2}}
\global\long\def\cell{\sigma}
\global\long\def\simp{\sigma}
\global\long\def\oset{U}
\global\long\def\pspace{S}
\global\long\def\chain{\mathcal{A}}
\global\long\def\bnd{\partial}
\global\long\def\mass#1{|#1|}
\global\long\def\rest{\raisebox{0.4pt}{\,\mbox{\ensuremath{\llcorner}}\,}}
\global\long\def\irest{\raisebox{0.4pt}{\mbox{\,\ensuremath{\lrcorner}\,}}}
\global\long\def\ess{\mathrm{ess}}
\global\long\def\virv{v}
\global\long\def\mvirv{\xi}
\global\long\def\conf{\kappa}
\global\long\def\confs{\mathcal{Q}}
\global\long\def\gconf{\kappa}
\global\long\def\virvs{W_{\conf}}
\global\long\def\mvirvs{W}
\global\long\def\D{\mathfrak{D}}
\global\long\def\form{\phi}
\global\long\def\Fmass#1{M\left(#1\right)}
\global\long\def\measure#1{\mu_{#1}}
\global\long\def\Fnormal#1{N\left(#1\right)}
\global\long\def\Fflat#1#2{F_{#1}\left(#2\right)}
\global\long\def\Fsharp#1#2{S_{#1}\left(#2\right)}
\global\long\def\lusb{L}
\global\long\def\wcbd{\widetilde{\cbnd}}
\global\long\def\fform#1{D_{#1}}
\global\long\def\strech{\varepsilon}
\global\long\def\ivirv{\chi}
\global\long\def\stress{\tau}
\global\long\def\ti{\mathcal{I}}
\global\long\def\endtime{\varepsilon}
\global\long\def\motion{m}
\global\long\def\Lie{\mathcal{L}}
\global\long\def\Reyo{\mathcal{R}}
\global\long\def\drivt#1{\frac{\partial#1}{\partial t}}
\global\long\def\eps{\varepsilon}
\global\long\def\compm{K_{\motion}}
\global\long\def\bodym{\body_{\motion}}
\global\long\def\ve{\hat{v}}
\newcommand{\note}[1]{**<[#1]>**}
\title{On the Role of Sharp Chains in the Transport Theorem}
\author{L. Falach and R. Segev}
\begin{abstract}
A generalized transport theorem for convecting irregular domains is
presented in the setting of Federer's geometric measure theory. A
prototypical $r$-dimensional domain is viewed as a flat $r$-chain
of finite mass in an open set of an $n$-dimensional Euclidean space.
The evolution of such a generalized domain in time is assumed to be
in accordance to a bi-Lipschitz type map. The induced curve is shown
to be continuous with respect to the flat norm and differential with
respect to the sharp norm on currents in $\reals^{n}$. A time dependent
property is naturally assigned to the evolving region via the action
of an $r$-cochain on the current associated with the domain. Applying
a representation theorem for cochains the properties are shown to
be locally represented by an $r$-form. Using these notions a generalized
transport theorm is presented.
\end{abstract}
\maketitle
\section{Introduction}
Reynolds' transport theorem \cite{Reynolds}, offers a general form
for the formulation of basic conservation laws in continuum mechanics
and in particular in fluid dynamics. The traditional formulation of
Reynolds theorem (or Leibniz-Reynolds theorem) deals with the time
derivative of the integral of $\omega(t)$, a time dependent scalar
field, over a time evolving spatial region $\part(t)$ in a Euclidean
physical space. The region $\part(t)$ is assumed to be the image
of a domain under a smooth motion, where the domain is assumed to
be sufficiently regular such that the classical divergence theorem
is applicable (for example a Lipschitz domain). The transport theorem
states that
\[
\frac{d}{dt}\left(\int_{\part(t)}\omega(t)d\lusb^{n}\right)\mid_{t=\tau}=\int_{\part(\tau)}\frac{\partial\omega}{\partial t}\mid_{t=\tau}d\lusb^{n}+\int_{\bnd\part(\tau)}\omega(\tau)V\cdot\nu dH^{n-1},
\]
with $\nu$, the unit exterior normal to the boundary $\bnd\part$
and $V$ the velocity associated with the smooth motion. It is noted
that the proof of the Reynolds' transport theorem is attributed by
Truesdell and Toupin \cite[p.~347]{Truesdell1960} to Spielrein (1916).
In the study of deforming thin films or evolving phase boundary a
transport relation for surface integrals is of interest. Such a theorem
is usually refereed to as \emph{Surface Transport Theorem}. It seems
that the basic notions of surface transport theorem in the setting
of continuum mechanics were first introduced in \cite{Gurtin_Murdoch1975}
with the introduction of the \emph{surface divergence operator}.
Betounes \cite{Betounes1986}, examined the kinematics of an $r$-dimentional
submanifold embedded in an $n$-dimentional semi-Riemannian manifold.
Betounes's formulation brings to light the strong dependence of the
formulation of the surface transport theorem on the availability of
the mean curvature normal. In \cite{Gurtin1989} Gurtin et al. formulated
a surface transport theorem for moving interfaces while additional
study and applications were presented in \cite{Gurtin1988,Angenent1989,Gurtin1990,Gurtin2000,Gurtin2002},
to name a few.
In the aforementioned versions of the transport theorem, the regularity
of the evolving domain is tacitly assumed. The inclusion of irregular
domains, where such notions as the exterior normal and mean curvature
normal not applicable, in the formulation of the transport theorem
has been presented recently in \cite{Seguin2013,Seguin2013a}. Seguin
\& Fried construct a generalized transport theorem using the setting
of Harrison's theory of differential chains (see \cite{Harrison2012,Harrison2013}).
The proposed formulation allows for singularities to evolve in the
domains, \textit{e.g.}, the domains may develop holes, split into
pieces and the fractal dimensions associated with the domain considered
may change. As their formulation of a transport theorem relies of
a dual relation between the domains and the properties considered
the resulting representation for theses properties is fairly regular.
In \cite{Falach2014}, a transport theorem is presented in the setting
of general manifolds. The domain of integration considered is viewed
as a de-Rham current of compact support thus including highly irregular
domains. The domain is assumed to evolve under a smooth map and integration
of a given property in the classical theory is replaced with the action
of the evolving current on a smooth differential form.
In the present work we wish to present a version of the transport
theorem in the setting of Federer's geometric measure theory. A generalized
domain, or a control volume, is viewed as a flat $r$-chain of finite
mass $T$. The current $T$ is assumed to evolve under the action
of $\conf$, a time dependent bi-Lipschitz homeomorphisms. With $\ti\subset\reals$
representing a time interval the control volume at time $t\in\ti$
is represented by the flat $r$-chain $\conf_{t\#}T$, given by the
pushforward of $T$ by $\conf_{t}=\conf(t)$. A considerable portion
of this work is dedicated to the study of the properties of the induced
curve $t\mapsto\conf_{t\#}T$ which is shown to be continuous with
respect to the flat topology of currents and differentiable with respect
to the sharp topology of currents. The integration of a given property
over the domain is generalized to the action of an $r$-cochain on
the current.
It is observed that Lipschitz continuity arises naturally as a characteristic
of the proposed setting both is the resulting representation of properties,
by sharp forms, and in the regularity of the motion, independently.
\section{Notation and Preliminaries\label{sec:Notation-and-Preliminaries}}
In this section we review some of the fundamental concepts of the
theory of currents in an $n$-dimensional Euclidean space. Throughout,
the notation is in the same spirit of \cite[Chapter 4]{Federer1969}.
Let $\oset$ be an open set in $\reals^{n}$, the notation $\D^{r}\left(\oset\right)$
is used for the vector space of smooth, compactly supported real valued
differential $r$-forms defined on $\oset$. The vector space $\D^{r}\left(\oset\right)$
is endowed with a family of semi-norms $\norm{\cdot}{i,\compact}$
such that for a compact $\compact\subset\oset$ and $i\in\mathbb{N}$,
\[
\norm{\form}{i,\compact}=\sup\left\{ \norm{D^{j}\form(x)}{}\mid x\in\compact,\;0\leq j\leq i\right\} .
\]
(Here, the norm $\norm{D^{j}\form(x)}{}$ is induced by some norm
on tensors in $\reals^{n}$.) This family of seminorms endows $\D^{r}\left(\oset\right)$
with a locally convex topology. For $\form\in\D^{r}(\oset)$ we use
$\cbnd\form$ to denote the \textit{exterior derivative }\textit{\emph{of}}
$\form$, an element of $\D^{r+1}(\oset)$. A linear functional $T:\D^{r}(U)\to\reals$
continuous with respect to the topology on $\D^{r}\left(\oset\right)$
is referred to as an $r$-\emph{dimensional}\textit{\emph{ }}\textit{de
Rham current} in $\oset$. The collection of all $r$-dimensional
currents defined on $\oset$ forms the vector space $\D_{r}(U)=\left[\D^{r}(\oset)\right]^{*}$,
the dual vector space of $\D^{r}(\oset)$. Let $T\in\D_{r}(\oset)$
with $r\geq1$ then, $\bnd T$, the\textit{ boundary} of $T$, is
the element of $\D_{r-1}(U)$ defined by
\begin{equation}
\bnd T(\form)=T(\cbnd\form),\quad\text{for all}\quad\form\in\D^{r-1}(\oset).
\end{equation}
Thus, we have the boundary operator $\bnd=\cbnd^{*}$, the adjoint
operator of the exterior derivative. The support of a current $T\in\D_{r}(\oset)$
is defined by
\begin{equation}
\spt\left(T\right)=\oset\setminus{\textstyle \bigcup}W,
\end{equation}
where each $W$ is an open subset of $U$ such that $T(\phi)=0$ for
all $\phi\in\D^{r}(\oset)$, with $\spt(\phi)\subset W$. Generally
speaking, the support of a current $T\in\D_{r}\left(\oset\right)$
need not be compact, however, in this work all the currents considered
will be of compact support.
The inner product in $\reals^{n}$ induces an inner product in $\bigwedge_{r}\reals^{n}$,
the vector space of $r$-vector in $\reals^{n}$, and $\mass{\xi}$
will denote the resulting norm of an $r$-vector $\xi$. An $r$-vector
$\xi$ may be written by $\xi=\sum_{\lambda\in\Lambda(r,n)}\xi^{\lambda}e_{\lambda}$
where $\Lambda(r,n)$ is the collection of increasing maps from $\{1,\dots,r\}$
to $\left\{ 1,\dots,n\right\} $, and $\left\{ e_{\lambda}\right\} $
is the standard basis for $\bigwedge_{r}\reals^{n}$ defined by $e_{\lambda}=e_{\lambda(1)}\wedge\dots\wedge e_{\lambda(r)}$.
Thus, with the above notation, $\mass{\xi}=\sqrt{\left\langle \xi,\xi\right\rangle }=\sqrt{\left(\xi^{\lambda}\right)^{2}}.$
Given $\form\in\D^{r}(\oset)$, for every $x\in\oset$, $\form(x)$
is an $r$-covector. For $\form(x)$, as well as any other covector,
one defines
\begin{equation}
\norm{\form(x)}0=\sup\left\{ \form(x)(\xi)\mid\mass{\xi}\leq1,\,\,\xi\text{ is a simple }r\text{-vector}\right\} .
\end{equation}
For a compact subset $\compact\subset\oset$, define the $\compact$-Lipschitz
constant of $\form\in\D^{r}(\oset)$ by
\begin{equation}
\Lip_{\phi,\compact}=\sup_{x,\, y\in\compact}\frac{\norm{\form(y)-\form(x)}0}{\mass{y-x}}.\label{eq:Lip_form_K}
\end{equation}
In addition to the topology of test functions, three additional topologies
will be examined on $\D^{r}(\oset)$ each of which is induced by a
corresponding family of semi-norms. For $\compact\subset\oset$ the
\textit{$\compact$-comass semi-norm}\textit{\emph{ of}}\textit{ $\form\in\D^{r}\left(\oset\right)$}
is defined by
\begin{equation}
\Flmass{\compact}{\form}=\ess\sup_{x\in\compact}\left\{ \norm{\form(x)}0\right\} ,\label{eq:Mass_form}
\end{equation}
the $K$-\textit{flat semi-norm} on $\D^{r}\left(\oset\right)$ is
defined by
\begin{equation}
\Fflat{\compact}{\form}=\ess\sup_{x\in\compact}\left\{ \norm{\form(x)}0,\norm{\cbnd\form(x)}0\right\} ,\label{eq:Flat_norm_forms}
\end{equation}
and the $\compact$-\emph{sharp semi-norm} on $\D^{r}\left(\oset\right)$
is defined by
\begin{equation}
\Fsharp{\compact}{\form}=\sup\left\{ \sup_{x\in\compact}\norm{\form(x)}0,(r+1)\Lip_{\form,\compact}\right\} .\label{eq:Sharp_norm_forms}
\end{equation}
The factor $(r+1)$ in the above definition is introduced so that
\begin{equation}
\Fflat{\compact}{\form}\leq\Fsharp{\compact}{\form},\quad\text{for all }\form\in\D^{r}(\oset).\label{eq:F<S}
\end{equation}
(The essential supremum is used above in spite of the fact that we
consider smooth functions because we are going to apply below these
definitions to essentially bounded functions.) Later on, when the
$\ess\sup$ in each of the above terms is evaluated over $\oset$,
we shall write $\Fmass{\form},\,\Fflat{}{\form},\,\Fsharp{}{\form}$
for the mass, flat and Sharp norms, respectively.
For $T\in\D_{r}(\oset)$, the \textit{mass }\textit{\emph{of}}\textit{
$T$} is dually defined by
\begin{equation}
\Fmass T=\sup\left\{ T\left(\form\right)\mid\phi\in\D^{r}\left(\oset\right),\,\,\Fmass{\form}\leq1\right\} ,\label{eq:Mass_current}
\end{equation}
and
\[
\Flmass{\compact}T=\sup\left\{ T\left(\form\right)\mid\phi\in\D^{r}\left(\oset\right),\,\,\Flmass{\compact}{\form}\leq1\right\} .
\]
An $r$-dimensional current $T$ is said to be \textit{represented
by integration} if there exists a Radon measure $\measure T$ and
an $r$-vector valued, $\measure T$-measurable function, $\overrightarrow{T}$,
with $\mass{\overrightarrow{T}(x)}=1$ for $\measure T$-almost all
$x\in\oset$, such that
\begin{equation}
T\left(\form\right)=\int_{\oset}\form(\overrightarrow{T})d\measure T,\quad\text{for all}\quad\form\in\D^{r}(\oset).\label{eq:current_by_integration}
\end{equation}
A sufficient condition for an $r$-dimensional current, $T$, to be
represented by integration is that $T$ is a current of locally finite
mass, \textit{i.e.}, $\Flmass{\compact}T<\infty$ for all compact
subsets $\compact\subset\oset$. An $r$-current $T$ of compact support
is said to be \textit{a normal current} if both $T$ and $\bnd T$
are represented by integration. The notion of normal currents leads
to the following definition
\begin{equation}
\Fnormal T=\Fmass T+\Fmass{\bnd T},\label{eq:N_norm}
\end{equation}
and clearly, every $T\in\D_{r}(\oset)$ such that $\Fnormal T<\infty$
is a normal $r$-current. The vector space of all $r$-dimensional
normal currents in $\oset$ is denoted by $N_{r}\left(\oset\right)$
and for a compact set $\compact$ of $\oset$,
\begin{equation}
N_{r,\compact}\left(\oset\right)=N_{r}(\oset)\cap\left\{ T\mid\spt\left(T\right)\subset\compact\right\} .\label{eq:N_km}
\end{equation}
The \textit{$K$-flat norm }on $\D_{r}\left(\oset\right)$ is given
by
\begin{equation}
\Fflat{\compact}T=\sup\left\{ T\left(\form\right)\mid F_{\compact}\left(\form\right)\leq1\right\} .\label{eq:Flat_norm_currents}
\end{equation}
It follows naturally from the foregoing definition that
\begin{equation}
\Fflat{\compact}{\bnd T}\leq\Fflat{\compact}T.\label{eq:F(bndT)<F(T)}
\end{equation}
Note that if $T\in\D_{r}\left(\oset\right)$ such that $\Fflat{\compact}T<\infty$,
then, $\spt(T)\subset\compact$. For a given compact subset $\compact\subset\oset$,
the set $F_{r,\compact}(U)$ is defined as the $F_{\compact}$-closure
of $N_{r,\compact}(U)$ in $\D_{r}(U)$. In addition, set
\begin{equation}
F_{r}(\oset)=\bigcup_{\compact}F_{r,\compact}(U),\label{eq:F_m(U)}
\end{equation}
where the union is taken over all compact subsets $\compact$ of
$\oset$. An element in $F_{r}(\oset)$ is referred to as a \textit{flat
$r$-chain in $\oset$}.
For $T\in F_{r,\compact}(\oset)$ it can be shown that $F_{\compact}(T)$
is given by
\begin{equation}
F_{\compact}(T)=\inf\left\{ \Fmass{T-\bnd S}+\Fmass S\mid S\in\D_{r+1}(\oset),\,\spt(S)\subset\compact\right\} ,\label{eq:Flat_norm_Whitney}
\end{equation}
and by taking $S=0$ it follows that
\begin{equation}
F_{\compact}(T)\leq\Fmass T.\label{eq:F<M}
\end{equation}
In addition, any element $T\in F_{r,\compact}\left(\oset\right)$
may be represented by $T=R+\bnd S$ where $R\in\D_{r}(\oset)$, $S\in\D_{r+1}(\oset)$,
such that $\spt(R)\subset\compact$, $\spt(S)\subset\compact$, and
\begin{equation}
F_{\compact}(T)=\Fmass R+\Fmass S,\label{eq:T=00003DR+bndS}
\end{equation}
so that $R$ and $S$ are of finite mass. By Equation (\ref{eq:F(bndT)<F(T)})
we note that the boundary of a flat $r$-chain is a flat $(r-1)$-chain.
The following representation theorem for flat chains is given in \cite[Section 4.1.18]{Federer1969}.
Let $T$ be a flat $r$-chain in $\oset$, then, $T$ is represented
by
\begin{equation}
T=\lusb^{n}\wedge\eta+\bnd\left(\lusb^{n}\wedge\xi\right),\label{eq:flat_chain_representation}
\end{equation}
with $\eta$ an $\lusb^{n}\rest\oset$-summable $r$-vector field
and $\xi$ an $\lusb^{n}\rest\oset$-summable $\left(r+1\right)$-vector
field $\xi$. Here, $\lusb^{n}\rest\oset$ denotes the restriction
of the $n$-dimensional Lebesgue measure to $U$ and for any $p$-current
$T$ and any $r$-vector field $\eta$, the $(p+r)$-current $T\we\eta$
is given by
\[
T\we\eta(\psi)=T(\psi\rest\eta).
\]
For $\form\in\D_{r}\left(\oset\right)$, the action $T(\form)$ is
given therefore by
\begin{equation}
T(\form)=\int_{U}(\form(\eta)+d\form(\xi))d\lusb^{n}.\label{eq:flat_chain_repr_a}
\end{equation}
A real valued linear functional $\cochain$ defined on $F_{r}\left(\oset\right)$
is said to be a \emph{flat $r$-cochain in $\oset$} if there exists
a number $c<\infty$ such that for any $\compact\subset\oset$ compact
subset
\[
\cochain\left(T\right)\leq c\Fflat{\compact}T,\quad\text{for all }T\in\Fflat{r,\compact}{\oset}.
\]
The infimum of all bounds $c$ is the norm of $\cochain$. An $r$-form
$\omega$ in $\oset$ is said to be a \textit{flat $r$-form in} $\oset$
if $\omega$ and $\cbnd\omega$, taken in the distributional sense,
are $\lusb^{n}$-measurable and essentially bounded (see \cite[p.~38]{Heinonen2000}).
That is, a measurable $r$-form $\omega$ is a flat $r$-form in $\oset$
if and only if $\Fflat{}{\omega}<\infty$. An important result, \textit{Wolfe's
representation theorem} (see \cite[ch.~7]{Whitney1957}, \cite[sec.~4.1.19]{Federer1969}),
states that each flat $r$-cochains $\cochain$ can be isometrically
identified with a flat $r$-form in $\oset$.
The notion of sharp chains was initially introduced in Whitney's classical
monograph \cite{Whitney1957}. In the following we present a construction
in the spirit of the formulation of the theory of flat chains in \cite{Federer1969}
(who does not consider sharp chains).
The $\compact$-\emph{sharp norm} on $\D_{r}\left(\oset\right)$ is
given by
\begin{equation}
\Fsharp{\compact}T=\sup\left\{ T\left(\form\right)\mid\Fsharp{\compact}{\form}\leq1\right\} .\label{eq:sahrp_norm_currents}
\end{equation}
For a given compact subset $\compact\subset\oset$, the set $\Fsharp{r,\compact}{\oset}$
is defined as the $S_{\compact}$-closure of $N_{r,\compact}(U)$
in $\D_{r}(U)$. In addition, we set
\begin{equation}
S_{r}(\oset)=\bigcup_{\compact}S_{r,\compact}(U),\label{eq:S_r(U)}
\end{equation}
where the union is taken over all compact subsets $\compact$ of $\oset$.
An element in $S_{r}(\oset)$ is referred to as a \textit{sharp $r$-chain
in $\oset$}. Note that as $F_{r,\compact}(U)\subset S_{r,\compact}(U)$
and the set $N_{r,\compact}\left(\oset\right)$ is a dense set in
both $F_{r,\compact}\left(\oset\right)$ and $S_{r,\compact}\left(\oset\right)$,
it follows that every sharp $r$-chain may be viewed a the limit,
in the sharp topology, of a sequence of flat $r$-chains. A representation
theorem for general sharp chains is beyond the scope of this work,
however, for sharp chains of finite mass such a representation theorem
may be found in \cite[Chapter.~XI]{Whitney1957}.
Let $\cochain$ be a real valued linear functional defined on $\Fsharp r{\oset}$,
then, $\cochain$ is said to be a\emph{ sharp $r$-cochain} \emph{in
$\oset$} provided there exists a number $b<\infty$ such that for
any $\compact\subset\oset$ compact subset
\[
\cochain\left(T\right)<b\Fsharp{\compact}T,\quad\text{for all}\; T\in\Fsharp{r,\compact}{\oset}.
\]
The infimum of all bounds $b$ is the norm of $\cochain$. An $r$-form
$\omega$ in $\oset$ is said to be a \textit{sharp $r$-form in}
$\oset$ if the coefficients of $\omega$ are bounded and Lipschitz
continuous, \textit{i.e.}, $\omega$ is a sharp $r$-from if and only
if
\[
\Fsharp{}{\omega}=\sup\left\{ \sup_{x\in\oset}\norm{\omega(x)}0,(r+1)\Lip_{\omega}\right\} <\infty.
\]
\begin{thm}
Let $\cochain$ be a sharp $r$-cochain in $\reals^{n}$. Then, $\cochain$
can be isometrically identified with $\fform{\cochain}$, a sharp
$r$-form in $\reals^{n}$ . That is, for any flat chain given by
$T=\lusb^{n}\wedge\eta+\bnd\left(\lusb^{n}\wedge\xi\right)$
\[
\cochain\left(T\right)=\int\left[\fform{\cochain}\left(\eta\right)+\cbnd\fform{\cochain}\left(\xi\right)\right]d\lusb^{n}.
\]
For a sharp $r$-chain, $T$, as flat chains are dense in the space
of sharp chains we have $T=\lim_{i\to\infty}^{S}T_{i}$ and
\[
\cochain\left(T\right)=\lim_{i\to\infty}\cochain\left(T_{i}\right).
\]
\end{thm}
For the proof see \cite[Section V.10]{Whitney1957}
Let $\oset\subset\reals^{n}$, $V\subset\reals^{m}$ be open sets
and let $T\in\D_{k}(\oset)$, $S\in\D_{l}(V)$. Then, the Cartesian
product of $T$ and $S$ is an element of $\D_{k+l}\left(\oset\times V\right)$
denoted by $T\times S$ and defined as follows. Let $\omega\in\D^{k+l}(U\times V)$
be given by
\[
\omega=\sum_{\mass{\alpha}=k,\mass{\beta}=l}\omega_{\alpha\beta}(x,y)dx^{\alpha}\wedge dy^{\beta},
\]
where $x\in U$, $y\in V$, and $\alpha,\;\beta$ are multi-indices.
Then,
\begin{equation}
T\times S(\omega)=T\left(\sum_{\mass{\alpha}=k}S\left(\sum_{\mass{\beta}=l}\omega_{\alpha\beta}(x,y)dy^{\beta}\right)dx^{\alpha}\right).\label{eq:curtesian_currents_def}
\end{equation}
For the properties of the Cartesian products of currents see \cite[Section 2.3]{Giaquinta1998}.
Let$U\subset\reals$, then, the line segment $[a,b]\subset U$ defines
naturally an element of $\D_{1}\left(U\right)$ such that for $\form\in\D_{1}\left(V\right)=C_{0}^{\infty}\left(V\right)$the
action $[a,b](\form)$ is $[a,b](\form)=\int_{a}^{b}\form(t)d\lusb_{t}^{1}$.
Consider the case of $T\in\D_{r}\left(U\right)$, then, a given $\omega\in\D^{r+1}\left(V\times\oset\right)$
may be split into a horizontal and a vertical component in the form
$\omega=\omega_{H}+\omega_{V}$ by
\begin{equation}
\omega_{H}(t,x)=\sum_{\mass{\alpha}=r}\omega_{H\alpha}(t,x)dt\wedge dx^{\alpha},\quad\omega_{V}(t,x)=\sum_{\mass{\alpha}=r+1}\omega_{V\alpha}(t,x)dx^{\alpha},
\end{equation}
where, $e_{t}$ is the pre-dual of $dt$. It is observed that
\[
\sum_{\mass{\alpha}=r}\omega_{H\alpha}(t,x)dx^{\alpha}=\omega\rest e_{t}.
\]
Applying Equation (\ref{eq:curtesian_currents_def}) to $\left([a,b]\times T\right)(\omega)$,
\begin{equation}
\left([a,b]\times T\right)(\omega)=\int_{a}^{b}T\left(\omega_{H}\right)d\lusb_{t}^{1}=\int_{a}^{b}T\left(e_{t}\irest\omega\right)d\lusb_{t}^{1}=\int_{a}^{b}e_{t}\wedge T\left(\omega\right)d\lusb_{t}^{1}.\label{eq:product_comput}
\end{equation}
Generally speaking, the Cartesian product of two flat chains is not
a flat chain. However the Cartesian product of a flat chain and a
normal current is a flat chain (see \cite[Sec. 4.1.12]{Federer1969}.
\section{Lipschitz maps\label{sec:Lipschitz-maps}}
In this section we briefly review some of the relevant properties
of Lipschitz mappings. From the point of view of kinematics, Lipschitz
mappings will be used to model the evolution of body-like regions
in space.
A map $\Lmap:\oset\to V$ from an open set $\oset\subset\reals^{n}$
to an open set $V\subset\reals^{m}$, is said to be a\textit{ (globally)
Lipschitz map} if there exists a number $c<\infty$ such that $\mass{\Lmap(x)-\Lmap(y)}\leq c\mass{x-y}$
for all $x,\, y\in\oset$. The \textit{Lipschitz constant }of $\Lmap$
is defined by
\begin{equation}
\Lip_{\Lmap}=\sup_{x,y\in\oset}\frac{|\Lmap(y)-\Lmap(x)|}{|y-x|}.\label{eq:Lipschitz-constant-1}
\end{equation}
The map $\Lmap:\oset\to V$ is said to be \textit{locally Lipschitz}
if for every $x\in\oset$ there is a neighborhood $\oset_{x}\subset\oset$
of $x$ such that the restricted map $\Lmap\mid_{\oset_{x}}$ is a
Lipschitz map. For a locally Lipschitz map, $\Lmap:\oset\to\reals^{m}$,
defined on the open set $\oset\subset\reals^{n}$ and a compact subset
$\compact\subset\oset$, the restricted map $\Lmap\mid_{\compact}$
is globally Lipschitz in the sense that $\Lip_{\Lmap,\compact}$,
the $\compact$-Lipschitz constant of the map $\Lmap\mid_{\compact}$,
is given by
\begin{equation}
\Lip_{\Lmap,\compact}=\sup_{x,y\in\compact}\frac{\mass{\Lmap(x)-\Lmap(y)}}{\mass{x-y}}.
\end{equation}
The vector space of locally Lipschitz mappings from the open set $\oset\subset\reals^{n}$
to $\reals^{m}$ is denoted by $\Lip\left(\oset,\reals^{m}\right)$.
For a compact subset $\compact\subset\oset$, define the semi-norm
\begin{equation}
\norm{\Lmap}{\Lip,\compact}=\max\left\{ \norm{\Lmap\mid_{\compact}}{\infty},\Lip_{\Lmap,\compact}\right\} ,\label{eq:Lipschit_semi_norm}
\end{equation}
on $\Lip\left(\oset,\reals^{m}\right)$, where,
\begin{equation}
\|\Lmap\mid_{\compact}\|_{\infty}=\sup_{x\in\compact}\mass{\Lmap(x)}.
\end{equation}
The vector space $\Lip\left(\oset,\reals^{m}\right)$ is endowed with
the strong Lipschitz topology (see \cite{Fukui2005} for the definition
on Riemannian manifolds). It is the analogue of Whitney's topology
(strong topology) for the space of differentiable mappings between
open sets (see \cite[p.~35]{Hirsch}) and is defined as follows.
\begin{defn}
\label{def:Lipschitz_strong_topology}Given $\Lmap\in\Lip\left(\oset,\reals^{m}\right)$,
for some index set $\Lambda$, let $\mathcal{O}=\left\{ O_{\lambda}\right\} _{\lambda\in\Lambda}$
be an open, locally finite cover of $\oset\subset\reals^{n}$, and
$\mathcal{K}=\left\{ \compact_{\lambda}\right\} _{\lambda\in\Lambda}$
a family of compact subsets covering $\oset$ such that $\compact_{\lambda}\subset O_{\lambda}$
and $\delta=\left\{ \delta_{\lambda}\right\} _{\lambda\in\Lambda}$
a family of positive numbers. A neighborhood $B^{\Lip}\left(\Lmap,\mathcal{O},\delta,\mathcal{K}\right)$
of $\Lmap$ in the strong topology is defined as the collection of
all $g\in\Lip\left(\oset,\reals^{m}\right)$ such that $\norm{\Lmap-g}{\Lip,\compact_{\lambda}}<\delta_{\lambda}$,
\emph{i.e.},
\begin{equation}
B^{\Lip}\left(\Lmap,\mathcal{O},\mathcal{K},\delta\right)=\left\{ g\in\Lip\left(\oset,\reals^{n}\right)\mid\:\norm{\Lmap-g}{\Lip,\compact_{\lambda}}<\delta_{\lambda},\,\lambda\in\Lambda\right\} .\label{eq:Lipschitz_strong_topology}
\end{equation}
For an illustrative description for the strong topology in the case
of $C^{0}$-functions, see \cite[p.~59]{Hirsch}.
\end{defn}
The following lemma shows the strong character of convergence in the
strong topology. For the proof in the case of differentiable mappings,
see \cite[p.~27]{Michor1980} or \cite[p.~43]{Golubitsky1973}.
\begin{lem}
\label{lem:compact_strong_topology}Let $\left\{ \Lmap_{\alpha}\right\} _{\alpha=1}^{\infty}$
be a sequence in $\Lip\left(\oset,\reals^{m}\right)$. Then, the sequence
converges to $\Lmap\in\Lip\left(\oset,\reals^{m}\right)$ in the strong
Lipschitz topology, if and only if there exists a compact subset $\compact\subset\oset$
such that $\Lmap_{\alpha}$ equals $\Lmap$ on $\oset\backslash\compact$
for all but finitely many $\alpha$'s and $\Lmap_{\alpha}\mid_{\compact}$
converges to $\Lmap\mid_{\compact}$ uniformly.
\end{lem}
A map $\emb:\oset\too V$, for open sets $\oset\subset\reals^{n}$,
$V\subset\reals^{m}$ and $m\geq n$, is said to be a \textit{bi-Lipschitz
}map if there are numbers $0<c\leq d<\infty$, such that
\begin{equation}
c\leq\frac{|\emb(x)-\emb(y)|}{\mass{x-y}}\leq d,\qquad\text{for all}\quad x,\, y\in\oset,\; x\not=y.\label{eq:bi-Lipschitz_constant-1}
\end{equation}
(See \cite[p.~78]{Heinonen2000} for further discussion.) Setting
$L=\max\left\{ \frac{1}{c},d\right\} $, it follows that
\begin{equation}
\frac{1}{L}\leq\frac{|\emb(x)-\emb(y)|}{\mass{x-y}}\leq L,\qquad\text{for all}\quad x,\, y\in\oset,\; x\not=y,
\end{equation}
and in such a case $\emb$ is said to be $L$-bi-Lipschitz. (See \cite[p.~78]{Heinonen2000}
for further discussion.)
\begin{defn}
The map $\Lmap:\oset\to V$, where $\oset\subset\reals^{n}$ and $V\subset\reals^{m}$
are open sets such that $m\geq n$, is a \textit{Lipschitz immersion}
if for every $x\in\oset$ there is a neighborhood $U_{x}\subset\oset$
of $x$ such that $\Lmap\mid_{U_{x}}$ is a bi-Lipschitz map, \textit{i.e.},
there are $0<c_{x}\leq d_{x}<\infty$, and
\begin{equation}
c_{x}\leq\frac{|\emb(y)-\emb(z)|}{\mass{y-z}}\leq d_{x},\quad\text{for all}\quad y,\, z\in U_{x},\; y\not=z.
\end{equation}
\end{defn}
\begin{defn}
A Lipschitz map $\emb:\oset\to V$ is said to be a\textit{ Lipschitz
embedding} if it is a Lipschitz immersion and a homeomorphism of $\oset$
onto $\emb(\oset)$.
\end{defn}
The following theorems pertaining to the set of Lipschitz immersions
and Lipschitz embeddings are given in \cite{Fukui2005} for the setting
of Lipschitz manifolds. Their proofs are analogous to the case of
differentiable mappings as in \cite[p.~36--38]{Hirsch}.
\begin{thm}
The set of Lipschitz immersions is an open subset of $\Lip\left(\oset,\reals^{m}\right)$
with respect to the strong Lipschitz topology\label{thm:Immersion_open}.
\end{thm}
\begin{thm}
The set $\Limb\left(\oset,\reals^{m}\right)$ is open in $\Lip\left(\oset,\reals^{m}\right)$
with respect to the strong Lipschitz topology \label{thm:Lipschitz_embedding_open_set}.
\end{thm}
In the following, a smooth embedding $\emb$ will be an element of
the set
\begin{equation}
\Emb{\oset,\reals^{m}}=C^{\infty}\left(\oset,\reals^{m}\right)\cap\Limb\left(\oset,\reals^{m}\right),\label{eq:DefineEmb}
\end{equation}
that is, a Lipschitz embedding whose components are smooth.
\section{The Image of Currents and Homotopy\label{sec:fushforward_homotopy}}
Let $\oset\subset\reals^{n}$ and $V\subset\reals^{m}$ be open sets,
, and let $f:\oset\to V$ be a map of class $C^{\infty}$. We recall
that for any $\omega\in\D^{r}(V)$, the pullback of $\omega$ by $f$,
is the $r$-form in $\oset$ denoted by $f^{\#}(\omega)$ such that
\begin{equation}
\begin{split}f^{\#}(\omega)(x)\left(v_{1}\wedge\dots\wedge v_{r}\right) & =\omega\left(f(x)\right)\left(Df(x)v_{1}\wedge\dots\wedge Df(x)v_{r}\right),\\
& =\omega\left(f(x)\right)\left(\bigwedge_{r}Df(x)\left(v_{1}\wedge\dots\wedge v_{r}\right)\right),
\end{split}
\label{eq:pullback_form_smooth}
\end{equation}
for any collection of vectors $v_{1},\dots,v_{r}\in\reals^{n}$. Note
that $f^{\#}\left(\omega\right)$ need not be an element of $\D^{r}(\oset)$.
For example, let $f:\oset\to\reals^{n}$ be the inclusion then for
$\omega\in\D^{r}\left(\reals^{n}\right)$ such that $\spt(\omega)\cap\oset\not=\varnothing$
and $\spt(\omega)\cap\left(\reals^{n}\backslash\oset\right)\not=\varnothing$
then $f^{\#}\left(\omega\right)$ is not compactly supported in $\oset$
an thus $f^{\#}\left(\omega\right)\not\in\D^{r}(\oset)$.
Let $T\in\D_{r}\left(\oset\right)$ such that $f\mid_{\spt\left(T\right)}$
is a proper map. The \emph{pushforward} of $T$ by $f$ is denoted
by $f_{\#}(T)\in\D_{r}(V)$ and is defined by
\begin{equation}
f_{\#}(T)\left(\omega\right)=T\left(\gamma\wedge f^{\#}(\omega)\right),\quad\text{for all }\omega\in\D^{r}(V),\label{eq:pushforward_current_smooth}
\end{equation}
where $\gamma\in\D^{0}(\oset)$ is any cutoff function satisfying
\[
\spt(T)\cap f^{-1}\left(\spt(\omega)\right)\subset\mathrm{Int}\left\{ x\mid\gamma(x)=1\right\} .
\]
For $T\in\D_{r}(\oset)$ with $\spt(T)\subset\compact$, where $K$
is a compact subset of $\oset$, we have the following bounds
\begin{equation}
\begin{split}\Fmass{f^{\#}(T)} & \leq\sup_{x\in\compact}\|Df(x)\|^{r}\Fmass T,\\
\Fnormal{f^{\#}(T)} & \leq\sup_{x\in\compact}\left\{ \|Df(x)\|^{r-1},\|Df(x)\|^{r}\right\} \Fnormal T,\\
F_{f\left\{ \compact\right\} }\left(f^{\#}(T)\right) & \leq\sup_{x\in\compact}\left\{ \|Df(x)\|^{r},\|Df(x)\|^{r+1}\right\} F_{\compact}(T).
\end{split}
\label{eq:f(T)_bounds_smooth}
\end{equation}
Let $\oset\subset\reals^{n}$ be an open set and let $f$ and $g$
be smooth maps of $\oset$ into $\reals^{m}$. For an open set $A$
of $\reals$ such that $\left[a,b\right]\subset A$, a smooth homotopy
between the maps $f$ and $g$ is a map
\begin{equation}
h:A\times\oset\to\reals^{m},
\end{equation}
such that
\begin{equation}
h\left(a,x\right)=f(x),\;\text{and }h(b,x)=g(x),\quad\text{for all }x\in\oset.
\end{equation}
Henceforth, the following notation will be used
\begin{equation}
h_{\tau}(x)=h(\tau,x),\;\text{for all}\, x\in\oset,
\end{equation}
and
\[
\dot{h}_{\tau}:\oset\to\reals^{m},\qquad\dot{h}_{\tau}(x)=Dh(\tau,x)\left(1,0\right)=\frac{\partial h}{\partial\tau}(\tau,x),\;\text{for all }x\in\oset.
\]
For $T\in\D_{r}(\oset)$ and a homotopy $h$ between $f$ and $g$,
the \textit{$h$-deformation chain of $T$} is defined as the current
\begin{equation}
h_{\#}\left(\left[a,b\right]\times T\right)\in\D_{r+1}\left(\reals^{m}\right).
\end{equation}
Traditionally, the interval $[a,b]$ is taken as the unit interval
$[0,1]$. The properties $h_{\#}\left(\left[a,b\right]\times T\right)$
are further investigated in \cite[Section 4.1.9]{Federer1969} and
\cite[sec.~2.3]{Giaquinta1998}. A fundamental tool is the following
formula
\begin{equation}
g_{\#}\left(T\right)-f_{\#}\left(T\right)=\bnd h_{\#}\left(\left[a,b\right]\times T\right)+h_{\#}\left(\left[a,b\right]\times\bnd T\right),\label{eq:homotopy_formula_for_currents}
\end{equation}
which is referred to as \emph{the homotopy} \emph{formula for currents.}
Let $\Lmap:\oset\to V$ be a locally Lipschitz map, the image of a
general $r$-current under a locally Lipschitz map is generally undefined
as for any $\omega\in D^{r}\left(V\right)$ the fullback $\Lmap^{\#}\left(\omega\right)$
need not be a smooth differential $r$-form in $\oset$, moreover,
the coefficients of $\Lmap^{\#}(\omega)$ are not necessary Borel
functions. For a normal current $T\in N_{r}\left(\oset\right)$ one
can define, see \cite[Sec.~2.3]{Giaquinta1998}, the pushforward of
$T$ by the locally Lipschitz map $\Lmap$ as the following weak limit
\[
\Lmap_{\#}\left(T\right)(\omega)=\lim_{\rho\to0}\left\{ \left(\left(\mol{\rho}\Lmap\right)_{\#}T\right)(\omega)\right\} ,
\]
where $\left\{ \mol{\rho}\Lmap\right\} _{\rho}$ is a sequence of
smooth approximations obtained by mollification of $\Lmap$ (see \cite[Section 4.1.2]{Federer1969}).
The strong convergence of the sequence is proven by Equation (\ref{eq:homotopy_formula_for_currents}),
as the sequence $\left\{ \left(\mol{\rho}\Lmap\right)_{\#}T\right\} _{\rho}$
is shown to be a Cauchy sequence in the Banach space of flat $r$-chains.
(See \cite[Section 4.1.14]{Federer1969}.)
The operator $\Lmap_{\#}:N_{r,\compact}\left(\oset\right)\to N_{r,\Lmap\left\{ \compact\right\} }\left(V\right)$
is continuous with respect to the flat norm and thus extends (we keep
the same notation) to $\Lmap_{\#}:F_{r,\compact}\left(\oset\right)\to F_{r,\Lmap\left\{ \compact\right\} }\left(V\right)$.
For $T\in\D_{r}\left(\oset\right)$ with $\spt\left(T\right)\subset\compact$,
the bounds presented in Equation (\ref{eq:f(T)_bounds_smooth}) are
replaced with
\begin{equation}
\begin{split}\Fmass{\Lmap_{\#}(T)} & \leq\sup_{x\in\compact}\left(\Lip_{\Lmap,\compact}\right)^{r}\Fmass T,\\
\Fnormal{\Lmap_{\#}(T)} & \leq\sup_{x\in\compact}\left\{ \left(\Lip_{\Lmap,\compact}\right)^{r-1},\left(\Lip_{\Lmap,\compact}\right)^{r}\right\} \Fnormal T,\\
F_{f\left\{ \compact\right\} }\left(\Lmap_{\#}(T)\right) & \leq\sup_{x\in\compact}\left\{ \left(\Lip_{\Lmap,\compact}\right)^{r},\left(\Lip_{\Lmap,\compact}\right)^{r+1}\right\} F_{\compact}(T).
\end{split}
\label{eq:f(T)_bounds_Lipschitz}
\end{equation}
For $T\in N_{r,\compact}\left(\oset\right)$ the existence of $\Lmap_{\#}(T)\in F_{r,\Lmap\left\{ \compact\right\} }\left(V\right)$,
and the second bound in Equation (\ref{eq:f(T)_bounds_Lipschitz}),
imply that $\Lmap_{\#}(T)\in N_{r,\Lmap\left\{ \compact\right\} }\left(V\right)$.
Alternatively, one may define the pushforward $\Lmap_{\#}(T)$ by
utilizing the duality of flat chains and flat forms and setting
\[
\Lmap_{\#}(T)(\omega)=\cochain_{\Lmap^{\#}(\omega)}(T),\quad\text{for all }\omega\in\D^{r}\left(\omega\right).
\]
By Rademacher\textquoteright s theorem the derivative a Lipschitz
mapping exists for $\lusb^{n}$-almost every $x\in\oset$. Thus, Equation
(\ref{eq:pullback_form_smooth}) is meaningful for $\lusb^{n}$-almost
every $x\in\oset$ and $\Lmap^{\#}(\omega)$ is a flat $r$-form in
$\oset$. It follows that $\cochain_{\Lmap^{\#}(\omega)}$ is a flat
$r$-cochain and the action $\cochain_{\Lmap^{\#}(\omega)}(T)$ is
well defined. The homotopy theorem for currents, and in particular,
the homotopy formula given in Equation (\ref{eq:homotopy_formula_for_currents})
discussed above for smooth maps, is therefore extended to maps $h:A\times\oset\to\reals^{m}$
which are locally Lipschitz maps. We note that a similar definition
and Wolfe's representation theorem are applied in \cite[Sction X.9]{Whitney1957},
to define the pullback of a flat form by a Lipschitz map.
\section{The Lie derivative \label{sec:Lie-derivative}}
In this section we examine the regularity of the Lie derivative of
a differential $r$-form. Cartan's (magic) formula is a key element
in the following analysis and as a first step we examine the contraction
of a differential form by a smooth vector field.
We first introduce a component representation that will be useful
throughout this section. The summation convention will be used unless
otherwise stated. Let $v=v^{i}e_{i}$, $e_{i}=\nicefrac{\partial}{\partial x^{i}}$,
and $\omega=\omega_{\lambda}dx^{\lambda}$ with $\lambda\in\Lambda\left(n,r+1\right)$.
Then,
\begin{equation}
\omega\rest v=\left(\omega_{\lambda}dx^{\lambda}\right)\rest\left(v^{i}e_{i}\right)=v^{i}\omega_{\lambda}dx^{\lambda}\rest e_{i},\label{eq:v_omega_com}
\end{equation}
and the exterior derivative of $\omega\rest v$ is given by
\begin{equation}
\begin{split}\cbnd(\omega\rest v) & =\left(v^{i}\omega_{\lambda}\right)_{,j}dx^{j}\wedge\left(dx^{\lambda}\rest e_{i}\right),\\
& =\left(v_{,j}^{i}\omega_{\lambda}+v^{i}\omega_{\lambda,j}\right)dx^{j}\wedge\left(dx^{\lambda}\rest e_{i}\right).
\end{split}
\label{eq:d(v_omega)_com}
\end{equation}
Consider the $M_{\compact}$-seminorm of $\omega\rest v$ and recall
that
\begin{equation}
\Flmass{\compact}{\omega\rest v}=\ess\sup_{x\in\compact}\left\{ \|\left(\omega\rest v\right)(x)\|\right\} .\label{eq:M_K(v_w)}
\end{equation}
The $r$-form $\omega\rest v$ has $C(n,r)$ components each of which
is a sum of $(n-r)$ terms each of which is a multiplication of a
component of $\omega$ with a component of $v$. Hence,
\begin{equation}
\Flmass{\compact}{\omega\rest v}\leq C(n,r)\sup_{x\in\compact}\|v(x)\|\Flmass{\compact}{\omega}.\label{eq:Mass_v_w}
\end{equation}
\begin{rem}
\label{rem:v_wedge_T_not_flat}For a smooth vector field, $v:\oset\to\reals^{n}$,
and $\omega\in\D^{r}\left(\oset\right)$ such that $\spt(\omega)\subset\compact$,
note that $\spt\left(\cbnd(\omega\rest v)\right)\subset\compact$
and the components of $\cbnd(v\irest\omega)$ are functions in $C_{0}^{\infty}\left(\oset\right)$.
Thus,
\[
\Fmass{\cbnd(\omega\rest v)}=\ess\sup_{x\in\compact}\left\{ \|\cbnd(\omega\rest v)(x)\|\right\} <\infty.
\]
However, one cannot find a $C<\infty$ such that
\[
\sup_{x\in\compact}\left\{ \|\cbnd(\omega\rest v)(x)\|\right\} \leq C\norm v{\Lip,\compact}\sup_{x\in\compact}\left\{ \|\cbnd\omega(x)\|\right\} .
\]
\end{rem}
\begin{defn}
The Lie derivative of the differential form $\omega\in\D^{r}\left(\oset\right)$
with respect to the vector field $v$ on $\oset$ is the differential
$r$-form in $\oset$ denoted by $\Lie_{v}\omega$ and defined by
\begin{equation}
\emb_{t}^{\#}\left(\Lie_{v}\omega\right)=\frac{d}{dt}\left(\emb_{t}^{\#}\omega\right),\label{eq:Lie_form_def}
\end{equation}
where $\emb:\reals\times\oset\to\oset$ is the flow associated with
the vector field $v$ \cite[p.~370]{Marsden1988}. \label{def:Lie-derivative}
\end{defn}
The classical Cauchy-Lipschitz theory asserts the existence of a
flow for $v$, a time dependent vector field which is Lipschitz continuous
in the spatial variable and uniformly continuous with respect to the
time variable \cite[Sec. IV.1]{Lang1999}. The existence of flows
for vector fields with reduced regularity, such as vector fields of
bounded variation, is an active field of research and for further
discussion see \cite{Bouchut2006} and references cited therein. It
is noted that in the rest of this work the existence of a flow for
the Lipschitz vector field follows from the assumptions regarding
the motion described below.
The Lie derivative of a differential form $\omega$ with respect to
the vector field $v$ satisfies the identity
\begin{equation}
\Lie_{v}\omega=\cbnd\left(\omega\rest v\right)+\cbnd\omega\rest v,\label{eq:Cartan_forms}
\end{equation}
which is commonly known as \textit{Cartan's (magic) formula}. Note
that in case $v$ is a smooth vector field on $\oset$, it follows
that $\Lie_{v}\omega\in\D^{r}\left(\oset\right)$ for every $\omega\in\D^{r}\left(\oset\right)$.
\begin{lem}
Let $\omega\in\D^{r}\left(\oset\right)$ and let $v:\oset\to\reals^{n}$
be a smooth vector field. Then, there exists a constant $C\left(n,r\right)$
such that
\begin{equation}
\Flmass{\compact}{\Lie_{v}\omega}\leq C(n,r)S_{\compact}(\omega)\norm v{\Lip,\compact}.\label{eq:M_Lie_deriv}
\end{equation}
Moreover, the Lie derivative of a sharp $r$-form with respect to
a smooth vector field, taken in the weak sense, is an $r$-form of
locally finite mass.\label{lem:Reugularity_of_Lie_derivative}\end{lem}
\begin{proof}
By Equations (\ref{eq:v_omega_com}), (\ref{eq:d(v_omega)_com}) and
(\ref{eq:Cartan_forms}), a local representation of the Lie derivative
$\Lie_{v}\omega$ is given by
\begin{equation}
\Lie_{v}\omega=D(\omega)(v)+v_{,j}^{i}\omega_{\lambda}dx^{j}\wedge\left(dx^{\lambda}\rest e_{i}\right).\label{eq:Lie_derive_components}
\end{equation}
For a given $\lambda\in\bigwedge(n,r)$, the contraction $dx^{\lambda}\rest e_{i}$
does not vanish for $r$ base vectors $e_{k}$.For a selection of
$\lambda$ and $k$ such that $dx^{\lambda}\rest e_{k}\not=0$, the
wedge product $dx^{j}\wedge\left(dx^{\lambda}\rest e_{k}\right)$
will not vanish for a subset of $\{dx^{j}\}$ containing $(n-(r-1))$
elements. Hence,
\begin{equation}
\begin{split}\Flmass{\compact}{\Lie_{v}\omega} & \leq\Flmass{\compact}{D(\omega)(v)}+\Flmass{\compact}{v_{,j}^{i}\omega_{\lambda}dx^{j}\wedge\left(dx^{\lambda}\rest e_{i}\right)},\\
& \leq\Lip_{\omega,\compact}\norm v{\infty,\compact}+\frac{n!}{r!\left(n-r\right)!}r(n-(r-1))\Flmass{\compact}{\omega}\Lip_{v,\compact}\\
& \leq C(n,r)\norm v{\Lip,\compact}S_{\compact}(\omega).
\end{split}
,\label{eq:M(Lie(w))<S(w)}
\end{equation}
The extension to sharp forms follows from Rademacher's theorem which
implies that for a sharp form $\omega$, $D\omega$ exists almost
everywhere in $\oset$.
\end{proof}
For a smooth vector field $v:\oset\to\reals^{n}$ defined on the open
set $\oset\subset\reals^{n}$ and an $r$-current $T\in\D_{r}\left(\oset\right)$,
the $\left(r+1\right)$-current $v\wedge T$ is defined by
\begin{equation}
v\wedge T\left(\omega\right)=T\left(\omega\rest v\right),\quad\text{for all }\omega\in\D^{r+1}\left(\oset\right).\label{eq:v_wedge_smooth}
\end{equation}
As $v$ is a smooth vector field it follows that $\omega\rest v\in\D^{r}\left(\oset\right)$
and $T\left(\omega\rest v\right)$ is well defined.
By the remark preceding Definition \ref{def:Lie-derivative}, it follows
that given $T\in F_{r,\compact}\left(\oset\right)$ and a smooth vector
field $v:\oset\to\reals^{n}$, the $(r+1)$-current $v\wedge T$ is
not necessarily a flat $(r+1)$-chain. Moreover, even if $T\in N_{r,\compact}\left(\oset\right)$
the current $v\wedge T$ may not be a flat $(r+1)$-chain. The analysis
of $v\wedge T$ is a key element in what follows and the foregoing
remark is an example for the restricted applicability of the flat
norm.
The contraction of a vector field and a differential form defined
above may be extended to include locally Lipschitz vector fields.
Let $\omega\in\D^{r+1}\left(\oset\right)$, where $\oset\subset\reals^{n}$
is an open set, and let $v:\oset\to\reals^{n}$ be a locally Lipschitz
vector field. Define $\omega\rest v$ as the pointwise limit of the
contractions with the mollified vector fields $\mol{\rho}v$, \emph{i.e.,}
\[
\omega\rest v(x)=\lim_{\rho\to0}\left\{ \omega(x)\rest\left(\mol{\rho}v\right)(x)\right\} .
\]
We note that the convergence of the above limit is locally uniform
with respect to $x$. In a similar manner to the estimate in Equation
(\ref{eq:Mass_v_w}) we have
\[
\begin{split}\Flmass{\compact}{\omega\rest v} & =\lim_{\rho\to0}\Flmass{\compact}{\omega\rest\left(\mol{\rho}v\right)}\\
& \leq\lim_{\rho\to0}C(n,r)\ess\sup_{x\in\compact}\|\mol{\rho}v(x)\|\Flmass{\compact}{\omega}\\
& =C\left(n,r\right)\|v\|_{\infty,\compact}\Flmass{\compact}{\omega}.
\end{split}
\]
For the Lie derivative of $\omega\in\D^{r}\left(\oset\right)$ with
respect to the Lipschitz vector field we have
\[
\Lie_{v}\omega=\lim_{\rho\to0}\left(\Lie_{\mol{\rho}v}\omega\right)=\lim_{\rho\to0}\left(\cbnd\left(\omega\rest\left(\mol{\rho}v\right)\right)+\cbnd\omega\rest\left(\mol{\rho}v\right)\right),
\]
where the above limit is taken with respect to the $\Flmass{\compact}{\cdot}$-semi-norm.
The existence of the limit follows from the fact that
\[
\lim_{\rho\to0}\norm{\mol{\rho}v-v}{\Lip,\compact}=0,
\]
and the bound given in Equation (\ref{eq:M(Lie(w))<S(w)}). Moreover,
the bound in Equation (\ref{eq:M(Lie(w))<S(w)}) holds for locally
Lipschitz vector field.
\section{Smooth Configurations and Motions \label{sec:Smooth_motion}}
Let $\body\subset\reals^{n}$ and $\ti\subset\reals$ bounded open
subsets. Recalling (\ref{eq:DefineEmb}), a \emph{smooth motion} $\motion$
defined over the time interval $\ti$ is viewed as a curve
\begin{equation}
\motion:\ti\to\Emb{\body,\reals^{n}}.\label{eq:Def:motion}
\end{equation}
We assume that the motion $\motion$ is a $C^{1}$-curve with respect
to strong Lipschitz topology, that is, the derivative of the curve,
denoted by $\dot{\motion}$, is viewed as a curve
\[
\dot{\motion}:\ti\to C^{\infty}\left(\body,\reals^{n}\right)\cap\Lip\left(\body,\reals^{n}\right),
\]
which is continuous with respect to the strong Lipschitz topology.
The motion $\motion$, induces a map
\begin{equation}
\emb:\ti\times\body\to\reals^{n},\label{eq:emb_motion}
\end{equation}
by
\[
\emb(\tau,x)=\motion(\tau)(x),\quad\text{for all}\;\tau\in\ti,\: x\in\body,
\]
and so
\[
\frac{\partial\emb}{\partial t}(\tau,x)=\dot{\motion}(\tau)(x),\quad\text{for all}\;\tau\in\ti,\: x\in\body.
\]
It follows from Lemma \ref{lem:compact_strong_topology}, that there
exists a compact subset $\compm\subset\body$ such that for any $x\not\in\compm$
and every $t,\, t'\in\ti$
\begin{equation}
\emb(t,x)=\emb\left(t',x\right).\label{eq:emb_on_compact}
\end{equation}
Hence, for $x\not\in\compm$ and $t\in\ti$
\begin{equation}
\dot{\emb}\left(t,x\right)=0.\label{eq:velocity_on_compact}
\end{equation}
For some $t\in\ti$, set $\body'=\emb_{t}\left\{ \body\right\} $
and $\compm'=\emb_{t}\left\{ \compm\right\} $. By the preceding argument,
$B'$ and $\compm'$ are independent of the particular choice of $t\in\ti$.
\begin{rem}
Equations (\ref{eq:emb_on_compact}, \ref{eq:velocity_on_compact})
and the existence of $\compact_{\motion}$ are key features of the
motion examined and stem from the use of the strong Lipschitz topology.
A drawback to the use of the strong Lipschitz topology is in the relatively
small supply of converging sequences of maps in the form of Equation
(\ref{eq:emb_motion}) converging to a motion as given in Equation
(\ref{eq:Def:motion}).
\end{rem}
For each $t\in\ti$, $\emb_{t}=\motion(t)\in\Emb{\body,\reals^{n}}$,
so there exists an inverse $\eta_{t}:\imag{\emb_{t}}=\body'\to\body$
such that $\emb_{t}\circ\eta_{t}=I_{\body'}$ with $I_{\body'}$ the
identity map on the set $\body'$. Consider the vector field
\begin{equation}
v_{t}:\body'\to\reals^{n},\qquad v_{t}=\dot{\emb}_{t}\circ\eta_{t},\label{eq:EulerianVelocity-2}
\end{equation}
viewed as a vector field on $\body'$ such that $v_{t}(y)=0$ for
every $x\in\body'\backslash\compm'$. The vector field $v_{t}$ is
naturally extended to a vector field $\ve_{t}:\reals^{n}\to\reals^{n}$,
by setting
\begin{equation}
\ve_{t}(x)=\begin{cases}
v_{t}(x),\: & x\in\body',\\
0, & x\not\in\body'.
\end{cases}\label{eq:V_extension}
\end{equation}
Thus,$\ve_{t}$ is a smooth vector field which vanishes on $\reals^{n}\backslash\compm'$
. It follows that,
\[
\ve:\ti\times\reals^{n}\to\reals^{n},
\]
is a time dependent Lipschitz vector field defined on $\reals^{n}$.
For $s,\, t\in\ti$, define
\begin{equation}
J_{s,t}(x)=\begin{cases}
\emb_{s}\circ\eta_{t}(x), & x\in\body',\\
x, & x\not\in\body'.
\end{cases}\label{eq:smooth_flow}
\end{equation}
As shown in \cite{Falach2014},
\begin{equation}
\frac{\partial J_{s,t}}{\partial s}(x)=\ve_{s}\left(J_{s,t}\left(x\right)\right),\quad J_{t,t}(x)=x.\label{eq:ODE_flow_smooth}
\end{equation}
The map $J_{s,t}$, is the flow associated with the time dependent
vector field $\ve$. For $\omega\in\D^{r}\left(\reals^{n}\right)$,
\[
\frac{\partial\left(J_{\tau,t}^{\#}\omega\right)}{\partial\tau}\mid_{\tau=s}=J_{s,t}^{\#}\left(\Lie_{\ve_{s}}\omega\right),
\]
and as $\emb_{\tau}^{\#}\omega=\emb_{t}^{\#}\left(J_{\tau,t}^{\#}\omega\right)$,
it follows from Definition \ref{eq:Lie_form_def} and a direct computation
that
\begin{equation}
\frac{\partial\left(\emb_{\tau}^{\#}\omega\right)}{\partial\tau}\mid_{\tau=t}=\emb_{t}^{\#}\left(\Lie_{\ve_{t}}\omega\right).\label{eq:d/dt_pullback_smooth}
\end{equation}
We now derive a representation formula for the $\emb$-deformation
chain associated with the motion $\emb:\ti\times\body\to\reals$ and
a general current $T\in\D_{r}\left(\body\right)$. Let $[a,b]\subset\ti$,
then, applying Equation (\ref{eq:product_comput}) to $\omega\in\D^{r+1}\left(\reals^{n}\right)$,
one has
\begin{equation}
\begin{split}\emb_{\#}\left(\left[a,b\right]\times T\right)\left(\omega\right) & =\left(\left[a,b\right]\times T\right)\left(\emb^{\#}\left(\omega\right)\right),\\
& =\int_{a}^{b}T\left(\emb_{\tau}^{\#}\left(\omega\right)_{H}\right)d\lusb_{\tau}^{1},\\
& =\int_{a}^{b}T\left(\emb_{\tau}^{\#}\left(\omega\right)\rest e_{t}\right)d\lusb_{\tau}^{1}.
\end{split}
\label{eq:integral_homotopy}
\end{equation}
In order to examine the $r$-form $\emb_{\tau}^{\#}\left(\omega\right)\rest e_{t}$,
we apply it to an $r$-vector $\xi$ which we can assume to be ``space-like'',
that is, $\xi=v_{1}\wedge\dots\wedge v_{r}$ with $v_{i}\in\reals^{n}$,
for $i=1,\dots,r$. Otherwise, $\left(\emb_{\tau}^{\#}\left(\omega\right)\rest e_{t}\right)(\xi)=0$,
identically. One obtains,
\begin{equation}
\begin{split}\left(\emb_{\tau}^{\#}\left(\omega\right)(x)\rest e_{t}\right)(\xi) & =\emb^{\sharp}(\omega)(\tau,x)(e_{t}\wedge v_{1}\wedge\dots\wedge v_{r}),\\
& =\omega\circ\emb_{\tau}(x)\left(D\emb(\tau,x)\left(e_{t}\right)\wedge\bigwedge_{r}D\emb_{\tau}(x)\left(\xi\right)\right),\\
& =\omega\circ\emb_{\tau}(x)\left(\dot{\emb}_{\tau}(x)\wedge\bigwedge_{r}D\emb_{\tau}(x)\left(\xi\right)\right),\\
& =((\omega\circ\emb_{\tau})\rest\dot{\emb}_{\tau})(x)\left(\bigwedge_{r}D\emb_{\tau}(x)(\xi)\right).
\end{split}
\end{equation}
As $\dot{\emb}_{\tau}(x)=v_{\tau}(\emb_{\tau}(x))=\ve_{\tau}\circ\emb_{\tau}(x),$
\begin{equation}
\begin{split}\left(\emb^{\#}(\omega)(\tau,x)\rest e_{t}\right)(\xi) & =\left((\omega\circ\emb_{\tau})\rest\left(\ve_{\tau}\circ\emb_{\tau}\right)\right)(x)\left(\bigwedge_{r}D\emb_{\tau}(x)(\xi)\right),\\
& =(\omega\rest\ve_{\tau})\circ\emb_{\tau}(x)\left(\bigwedge_{r}D\emb_{\tau}(x)(\xi)\right),\\
& =\emb_{\tau}^{\#}(\omega\rest\ve_{\tau})(x)(\xi).
\end{split}
\end{equation}
It is concluded that,
\begin{equation}
\emb_{\tau}^{\#}(\omega)\rest e_{t}=\emb_{\tau}^{\#}(\omega\rest\ve_{\tau}).\label{eq:e_contraction_pullback_form}
\end{equation}
Returning to Equation (\ref{eq:integral_homotopy}), note that the
integrand may be rewritten as
\begin{equation}
\begin{split}T(\emb_{\tau}^{\#}(\omega)\rest e_{t}) & =T(\emb_{\tau}^{\#}(\omega\rest\ve_{\tau})),\\
& =\emb_{\tau\#}(T)(\omega\rest\ve_{\tau}),\\
& =\ve_{\tau}\we\emb_{\tau\#}(T)\left(\omega\right),
\end{split}
\end{equation}
and Equation (\ref{eq:integral_homotopy}) assumes the form
\begin{equation}
\emb_{\#}\left(\left[a,b\right]\times T\right)\left(\omega\right)=\int_{a}^{b}(\ve_{\tau}\we\emb_{\tau\#}T)(\omega)d\lusb_{\tau}^{1}.\label{eq:emb_pushforward_T}
\end{equation}
Applying (\ref{eq:homotopy_formula_for_currents}) to (\ref{eq:homotopy_formula_for_currents}),
one finally has,
\[
\begin{split}\left(\emb_{b\#}\left(T\right)-\emb_{a\#}\left(T\right)\right)\omega & =\left(\bnd\emb_{\#}\left(\left[a,b\right]\times T\right)+\emb_{\#}\left(\left[a,b\right]\times\bnd T\right)\right)\omega,\\
& =\int_{a}^{b}\left[(\ve_{\tau}\we\emb_{\tau\#}T)(\cbnd\omega)+(\ve_{\tau}\we\emb_{\tau\#}\bnd T)(\omega)\right]d\lusb_{\tau}^{1},\\
& =\int_{a}^{b}\left[(\emb_{\tau\#}T)\left(\cbnd\omega\rest\ve_{\tau}+\cbnd\left(\omega\rest\ve_{\tau}\right)\right)\right]d\lusb_{\tau}^{1},\\
& =\int_{a}^{b}\left[(\emb_{\tau\#}T)\left(\Lie_{\ve_{\tau}}\omega\right)\right]d\lusb_{\tau}^{1}.
\end{split}
\]
\section{The Kinematics of Currents under a Smooth Motion \label{sec:kinematics_cerrents}}
This section is devoted to the examination of kinematic properties
of generalized domains. A generalized $r$-dimensional oriented domain,
is naturally viewed as an $r$-current. In the selection of the appropriate
class of domains, the collection of all $r$-currents is far greater
than what we would consider as suitable. The selection of the appropriate
class is motivated by the following guidelines. Firstly, a current
representing a generalized domain must have a local character, at
least in some measure theoretic sense. Secondly, such a current must
have a definite, quantitative notion of a boundary.Finally, such a
current should be well behaved under the image of a Lipschitz map
\footnote{We feel that these requirement are in the spirit put forth by Noll
\& Virga in \cite{Noll1988} where the class admissible bodies should
include all those that can be imagined by an engineer but exclude
those that can be dreamt up only by an ingenious mathematician
} The introductory discussion in Sections \ref{sec:Notation-and-Preliminaries}
and \ref{sec:fushforward_homotopy} indicates that a convenient choice
for the class of domains is the collection of flat chains of finite
mass. Thus, a prototypical control volume $T$, is viewed as a flat
$r$-chain of finite mass in $\body$. Using the properties of a motion
we outlined above and the corresponding notation of Section \ref{sec:Smooth_motion},
we consider $T\in F_{r,\compm}\left(\body\right)$, where $\compm\subset\body$
is the compact set containing the region where the motion is nontrivial.
\begin{rem}
Let $T\in F_{r}\left(\body\right)$ and $\gamma:\body\to\reals$ a
locally Lipschitz function, the multiplication $\gamma\wedge T$ is
flat $r$-chain. Thus, a flat $r$-chain may represent not only a
geometric domain but may also be represent some intensive property.
See \cite{Falach2013} for further details.
\end{rem}
Consider a map $\emb$ induced by a motion as defined in Equation
(\ref{eq:emb_motion}) and a flat chain $T\in F_{r,\compm}\left(\body\right)$
such that $\Fmass T<\infty$. The curve $t\mapsto\emb_{t\#}\left(T\right)$
will be viewed in this work as the time evolution of the control volume
represented by the current $T$.
\begin{lem}
Let $\emb$ be the map associated with a motion as defined by Equation
(\ref{eq:emb_motion}) and $T\in F_{r,\compm}\left(\body\right)$
a flat chain of finite mass. The curve induced by the pushforward
$t\mapsto\emb_{t\#}T$ is a continuous curve with respect to the $M_{K_{m}^{'}}$-norm
on $\D_{r}\left(\reals^{n}\right)$.\label{lem:emb_t(T)_smooth_con}\end{lem}
\begin{proof}
Let $t\in\ti$ and select $\eps$ such that $t+\eps\in\ti$. Then,
\[
\begin{split}\Flmass{\compm'}{\emb_{t+\eps\#}T-\emb_{t\#}T} & =\sup_{\omega\in\D^{r}\left(\reals^{n}\right)}\frac{\left(\emb_{t+\eps\#}T-\emb_{t\#}T\right)\omega}{\Flmass{\compm'}{\omega}},\\
& =\sup_{\omega\in\D^{r}\left(\reals^{n}\right)}\frac{T\left(\emb_{t+\eps}^{\#}\omega-\emb_{t}^{\#}\omega\right)}{\Flmass{\compm'}{\omega}},\\
& \leq\sup_{\omega\in\D^{r}\left(\reals^{n}\right)}\frac{\Fmass T\Flmass{\compm}{\emb_{t+\eps}^{\#}\omega-\emb_{t}^{\#}\omega}}{\Flmass{\compm'}{\omega}},
\end{split}
\]
where the last line follows from the integral representation of $T$.
For the term $\Flmass{\compm}{\emb_{t+\eps}^{\#}\omega-\emb_{t}^{\#}\omega}$,
a direct computation shows that
\[
\begin{split} & \Flmass{\compm}{\emb_{t+\eps}^{\#}\omega-\emb_{t}^{\#}\omega}\\
& =\sup_{x\in\compm}\left\{ \sup_{\xi}\left\{ \omega\left(\emb_{t+\eps}\left(x\right)\right)\left[\bigwedge_{r}D\emb_{t+\eps}(x)(\xi)\right]-\omega\left(\emb_{t}\left(x\right)\right)\left[\bigwedge_{r}D\emb_{t}(x)(\xi)\right]\right\} \right\} ,\\
& \leq\sup_{x\in\compm}\left\{ \sup_{\xi}\left\{ \omega\left(\emb_{t+\eps}\left(x\right)\right)\left[\bigwedge_{r}D\emb_{t+\eps}(x)(\xi)\right]-\omega\left(\emb_{t+\eps}\left(x\right)\right)\left[\bigwedge_{r}D\emb_{t}(x)(\xi)\right]\right\} \right\} \\
& \quad+\sup_{x\in\compm}\left\{ \sup_{\xi}\left\{ \omega\left(\emb_{t+\eps}\left(x\right)\right)\left[\bigwedge_{r}D\emb_{t}(x)(\xi)\right]-\omega\left(\emb_{t}\left(x\right)\right)\left[\bigwedge_{r}D\emb_{t}(x)(\xi)\right]\right\} \right\} ,\\
& \leq\sup_{x\in\compm}\left\{ \sup_{\xi}\left\{ \omega\left(\emb_{t+\eps}\left(x\right)\right)\left[\bigwedge_{r}D\left(\emb_{t+\eps}-\emb_{t}\right)(x)(\xi)\right]\right\} \right\} \\
& \quad+\sup_{x\in\compm}\left\{ \sup_{\xi}\left\{ \left(\omega\left(\emb_{t+\eps}\left(x\right)\right)-\omega\left(\emb_{t}\left(x\right)\right)\right)\left[\bigwedge_{r}D\emb_{t}(x)(\xi)\right]\right\} \right\} ,\\
& \leq\Flmass{\compm'}{\omega}\left(\Lip_{\emb_{t+\eps}-\emb_{t},\compm}\right)^{r}+\left(\Lip_{\emb_{t},\compm}\right)^{r}\Flmass{\compm'}{\omega\circ\emb_{t+\eps}-\omega\circ\emb_{t}}.
\end{split}
\]
By the continuity of the motion with respect to the strong Lipschitz
topology it follows that as $\eps\to0$ we have $\left(\Lip_{\emb_{t+\eps}-\emb_{t},\compm}\right)^{r}\to0$.
For the second term, as $\norm{\emb_{t+\eps}-\emb_{t}}{\infty,\compm}\to0$,
and since $\omega$ is smooth, it follows that $\Flmass{\compm'}{\omega\circ\emb_{t+\eps}-\omega\circ\emb_{t}}\to0$,
which completes the proof.
\end{proof}
As considered in \cite{Falach2014}, for all $\omega\in\D^{r}\left(\oset\right)$,
\[
\begin{split}\frac{d}{dt}\left(\left(\emb_{t\#}T\right)(\omega)\right)\mid_{t=\tau} & =\frac{d}{dt}\left(T\emb_{t}^{\#}(\omega)\right)\mid_{t=\tau},\\
& =T\left(\frac{d}{dt}\emb_{t}^{\#}(\omega)\mid_{t=\tau}\right),\\
& =T\left(\emb_{\tau}^{\#}\left(\Lie_{\ve_{\tau}}\omega\right)\right).
\end{split}
\]
Thus, using Equation (\ref{eq:Cartan_forms})
\[
\begin{split}T\left(\emb_{\tau}^{\#}\left(\Lie_{\ve_{\tau}}\omega\right)\right) & =\emb_{\tau\#}T\left(\cbnd\left(\omega\rest\ve_{\tau}\right)+\left(\cbnd\omega\right)\rest\ve_{\tau}\right),\\
& =\left(\ve_{\tau}\wedge\bnd\left(\emb_{\tau\#}T\right)+\bnd\left(\ve_{\tau}\wedge\emb_{\tau\#}T\right)\right)\omega.
\end{split}
\]
The foregoing result applies to general currents and is not restricted
to flat chains of finite mass. It may also be written with the introduction
of $\Reyo_{\ve_{t}}=\left(\Lie_{\ve_{t}}\right)^{*}$ as the dual
operator of the Lie derivative, that is
\[
T\left(\emb_{t}^{\#}\left(\Lie_{\ve_{t}}\omega\right)\right)=\Reyo_{\ve_{t}}\left(\emb_{t\#}T\right)\omega.
\]
Thus, for any de Rham current $T\in\D_{r}\left(\reals^{n}\right)$
\begin{equation}
\Reyo_{\ve_{t}}(T)=\ve_{t}\wedge\bnd T+\bnd\left(\ve_{t}\wedge T\right).\label{eq:Reynold_operator}
\end{equation}
It is observed that similar results have been reported in \cite{Chi2012}.
The previous analysis shows that the derivative $\nicefrac{d(\emb_{t\#}T)}{dt}$
converges in the topology of $\D_{r}\left(\reals^{n}\right)$. The
following theorem considers the convergence of the limit above in
the sharp norm topology.
\begin{thm}
\label{thm:driv_smooth_motion} Let $\emb$ be a map associated with
a motion as defined by Equation (\ref{eq:emb_motion}) and let $T\in F_{r,\compm}\left(\body\right)$
be a flat chain of finite mass. The derivative $\frac{d}{dt}\left(\emb_{t\#}T\right)\mid_{t=\tau}$,
exists in the topology of $\Fsharp{\compm'}{\reals^{n}}$ and is given
by
\begin{equation}
\frac{d}{dt}\left(\emb_{t\#}T\right)\mid_{t=\tau}=\bnd\left(\ve_{\tau}\wedge\emb_{\tau\#}T\right)+\ve_{\tau}\wedge\bnd\left(\emb_{\tau\#}T\right).\label{eq:driv_formal}
\end{equation}
\end{thm}
\begin{proof}
By the homotopy formula (\ref{eq:homotopy_formula_for_currents}),
we may write
\begin{equation}
\frac{d}{dt}\left(\emb_{t\#}T\right)=\lim_{\eps\to0}\frac{\bnd\emb_{\#}\left(\left[\tau,\tau+\eps\right]\times T\right)+\emb_{\#}\left(\left[\tau,\tau+\eps\right]\times\bnd T\right)}{\eps}.\label{eq:driv_by_homotopy}
\end{equation}
Thus,
\begin{multline*}
\frac{1}{\eps}\Fsharp{\compm'}{\emb_{t+\eps\#}T-\emb_{t\#}T-\eps\left(\bnd\left(\ve_{t}\wedge\emb_{t\#}T\right)+\ve_{t}\wedge\bnd\left(\emb_{t\#}T\right)\right)}\\
\begin{split} & =\frac{1}{\eps}\sup_{\omega\in\D^{r}\left(\reals^{n}\right)}\frac{\left(\emb_{t+\eps\#}T-\emb_{t\#}T-\eps\left(\bnd\left(\ve_{t}\wedge\emb_{t\#}T\right)+\ve_{t}\wedge\bnd\left(\emb_{t\#}T\right)\right)\right)\omega}{\Fsharp{\compm'}{\omega}},\\
& =\frac{1}{\eps}\sup_{\omega\in\D^{r}\left(\reals^{n}\right)}\left\{ \frac{\int_{t}^{t+\eps}\left[(\ve_{\tau}\we\emb_{\tau\#}T)(\cbnd\omega)+(\ve_{\tau}\we\emb_{\tau\#}\bnd T)(\omega)\right]d\lusb_{\tau}^{1}}{\Fsharp{\compm'}{\omega}}\right.\\
& \:\quad\qquad\qquad\left.-\frac{\eps\left(\ve_{t}\wedge\emb_{t\#}T\right)\left(\cbnd\omega\right)+\ve_{t}\wedge\bnd\left(\emb_{t\#}T\right)\omega}{\Fsharp{\compm'}{\omega}}\right\} ,\\
& \leq\frac{1}{\eps}\sup_{\omega\in\D^{r}\left(\reals^{n}\right)}\left\{ \frac{\int_{t}^{t+\eps}\left[T\left(\emb_{\tau}^{\#}\Lie_{\ve_{\tau}}\omega\right)\right]d\lusb_{\tau}^{1}-\eps T\left(\left(\emb_{t}^{\#}\Lie_{\ve_{t}}\omega\right)\right)}{\Fsharp{\compm'}{\omega}}\right\} ,\\
& \leq\sup_{\omega\in\D^{r}\left(\reals^{n}\right)}\left\{ \Fmass T\sup_{s\in[t,t+\eps]}\frac{\Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{s}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega}}{\Fsharp{\compm'}{\omega}}\right\} .
\end{split}
\end{multline*}
For the term $\Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{s}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega}$,
one has the estimate
\begin{multline*}
\Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{s}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega}\\
\begin{split}\leq & \Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{s}}\omega-\emb_{s}^{\#}\Lie_{\ve_{t}}\omega+\emb_{s}^{\#}\Lie_{\ve_{t}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega}\\
\leq & \Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{s}}\omega-\emb_{s}^{\#}\Lie_{\ve_{t}}\omega}+\Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{t}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega}\\
\leq & \Flmass{\compm}{\emb_{s}^{\#}\left(\Lie_{\ve_{s}-\ve_{t}}\omega\right)}+\Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{t}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega}.
\end{split}
\end{multline*}
By Equation (\ref{eq:M_Lie_deriv}),
\begin{eqnarray*}
\Flmass{\compm}{\emb_{s}^{\#}\left(\Lie_{\ve_{s}-\ve_{t}}\omega\right)} & \leq & \left(\Lip_{\emb_{s},\compm}\right)^{r}\Flmass{\compm'}{\Lie_{\ve_{s}-\ve_{t}}\omega},\\
& \leq & \left(\Lip_{\emb_{s},\compm}\right)^{r}C(n,r)S_{\compm'}(\omega)\norm{\ve_{s}-\ve_{t}}{\Lip,\compm'},
\end{eqnarray*}
and since $\norm{\ve_{s}-\ve_{t}}{\Lip,\compm'}\to0$ as $s\to t$,
it follows that
\[
\lim_{s\to t}\Flmass{\compm}{\emb_{s}^{\#}\left(\Lie_{\ve_{s}-\ve_{t}}\omega\right)}=0.
\]
Applying Lemma \ref{lem:emb_t(T)_smooth_con}, it follows that
\[
\lim_{s\to t}\Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{t}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega}=0.
\]
Setting
\[
T_{k}=\frac{\bnd\emb_{\#}\left(\left[t,t+\frac{1}{k}\right]\times T\right)+\emb_{\#}\left(\left[t,t+\frac{1}{k}\right]\times\bnd T\right)}{\frac{1}{k}},
\]
we obtain a sequence $\left\{ T_{k}\right\} $ of flat $r$-chains
converging to $\ve_{t}\wedge\bnd\left(\emb_{t\#}T\right)+\bnd\left(\ve_{t}\wedge\emb_{t\#}T\right)$
in $S_{r}\left(\reals^{n}\right)$. As normal currents are dense in
the space of flat chains one may obtain a sequence of normal $r$-currents
converging to $\ve_{t}\wedge\bnd\left(\emb_{t\#}T\right)+\bnd\left(\ve_{t}\wedge\emb_{t\#}T\right)$
in the $S_{\compm'}$-norm. \end{proof}
\begin{rem}
Note that the map $t\mapsto\bnd\left(\ve_{t}\wedge\emb_{t\#}T\right)+\hat{v}_{t}\wedge\bnd\left(\emb_{t\#}T\right)$
need not be continuous with respect to the $S_{K_{m}^{'}}$-norm,
as $\Fsharp{\compact'}{\omega}$ may depend of the second derivative
of $\omega\in\D^{r}\left(\reals^{n}\right)$. As an example, consider
the simple case of $\body\subset\reals$, and $T\in N(\oset)$ given
by $T(\omega)=\omega(x_{0})+\cbnd\omega(x_{0})$. For $v=e_{1}$,
\[
T\left(\Lie_{v}\omega\right)=D\omega(x_{0})+D^{2}\omega(x_{0}).
\]
\end{rem}
\section{Lipschitz Type Configurations and Motion \label{sec:Lipschitz_motion}}
This section extends the foregoing discussion to non smooth motions.
In particular, configurations represented by bi-Lipschitz maps, as
well as the corresponding motions, will be examined.
The definition of a motion, as introduced in Section \ref{sec:Smooth_motion},
is generalized by considering curves of the form
\begin{equation}
\motion:\ti\to\Limb\left(\body,\reals^{n}\right),\label{eq:Def:motion_Lipschitz}
\end{equation}
which we assume are continuously differentiable with respect to the
strong Lipschitz topology. Thus, the time derivative of the map is
\[
\dot{\motion}:\ti\to\Lip\left(\body,\reals\right),
\]
a continuous curve with respect to the strong Lipschitz topology.
The motion $m$ induces a map
\[
\conf:\ti\times\body\to\reals^{n},
\]
such that
\begin{equation}
\conf(\tau,x)=\motion(\tau)(x),\quad\text{for all}\;\tau\in\ti,\: x\in\body,\label{eq:Lips_map_motion}
\end{equation}
and so
\[
\dot{\conf}(\tau,x)=\dot{\motion}(\tau)(x),\quad\text{for all}\;\tau\in\ti,\: x\in\body.
\]
Using the results of Fukui \cite{Fukui2005}, we can make the analogous
definitions of $\compm$, $\compm'$, and $\body'$ as in Section
\ref{sec:Smooth_motion}. We consider the flow $J_{s,t}$ and the
vector field $\ve:\ti\times\reals^{n}\to\reals^{n}$ as in the smooth
case replacing $\emb$ with $\conf$. The vector field
\[
\ve:\ti\times\reals^{n}\to\reals^{n},
\]
is a time dependent Lipschitz vector field on $\reals^{n}$. As in
the smooth case, we consider the flow $J_{s,t}$ associated with the
time dependent vector field $\ve$ satisfying
\[
\frac{\partial J_{s,t}}{\partial s}(x)=\ve_{s}\left(J_{s,t}\left(x\right)\right),\quad J_{t,t}(x)=x.
\]
For $\omega\in\D^{r}\left(\reals^{n}\right)$ it follows that
\[
\frac{\partial(J_{\tau,t}^{\#}\omega)}{\partial\tau}\mid_{\tau=s}=J_{s,t}^{\#}\left(\Lie_{\ve_{s}}\omega\right),
\]
where $\Lie_{\ve_{s}}\omega$ is the Lie derivative of $\omega$ with
respect to Lipschitz vector field $\ve_{s}$ as discussed is Section
\ref{sec:Lie-derivative}. In particular,
\begin{equation}
\frac{\partial(\conf_{\tau}^{\#}\omega)}{\partial\tau}\mid_{\tau=t}=\conf_{t}^{\#}\left(\Lie_{\ve_{t}}\omega\right).\label{eq:d/dt_pullback_Lipschitz}
\end{equation}
We note that for a given $\omega\in\D^{r}\left(\reals^{n}\right)$,
the curve $t\mapsto\conf_{t}^{\#}\omega$ is valued in the space of
flat $r$-forms in $\body$. As in section \ref{sec:Smooth_motion},
the curve $t\mapsto\conf_{t\#}\left(T\right)$ is used to model the
time evolution of the generalized domain. The main results described
in Section \ref{sec:kinematics_cerrents} apply in the case of Lipschitz
motions.
\begin{lem}
\label{lem:conf_t(T)_Lipschitz}For the mapping $\conf$, as defined
in Equation (\ref{eq:Lips_map_motion}), and a flat chain $T\in F_{r,\compm}\left(\body\right)$
with $\Fmass T<\infty$, the curve $t\mapsto\conf_{t\#}\left(T\right)$,where
$\conf_{t\#}\left(T\right)$ is defined in Section \ref{sec:fushforward_homotopy}
by $\conf_{t\#}\left(T\right)=\lim_{\rho\to0}\left(\mol{\rho}\conf_{t}\right)\left(T\right)$,
is continuous with respect to the $M_{\compm'}$-norm on $\D_{r}\left(\reals^{n}\right)$. \end{lem}
\begin{proof}
let $t\in\ti$ and select $\eps$ such that $t+\eps\in\ti$. Then,
\begin{multline*}
\Flmass{\compm'}{\conf_{t+\eps\#}T-\conf_{t\#}T}\\
\begin{split} & =\sup_{\omega\in\D^{r}\left(\reals^{n}\right)}\left\{ \lim_{\rho\to0}\frac{T\left(\left(\mol{\rho}\conf_{t+\eps}\right)^{\#}\omega-\left(\mol{\rho}\conf_{t}\right)^{\#}\omega\right)}{\Flmass{\compm'}{\omega}}\right\} ,\\
& \leq\sup_{\omega\in\D^{r}\left(\reals^{n}\right)}\left\{ \lim_{\rho\to0}\frac{\Fmass T\Flmass{\compm}{\left(\mol{\rho}\conf_{t+\eps}\right)^{\#}\omega-\left(\mol{\rho}\conf_{t}\right)^{\#}\omega}}{\Flmass{\compm'}{\omega}}\right\} ,
\end{split}
\end{multline*}
and as in the proof of Lemma \ref{lem:emb_t(T)_smooth_con}
\begin{multline*}
\Flmass{\compm}{\left(\mol{\rho}\conf_{t+\eps}\right)^{\#}\omega-\left(\mol{\rho}\conf_{t}\right)^{\#}\omega}\\
\begin{split} & \leq\Flmass{\compm'}{\omega}\left(\Lip_{\left(\mol{\rho}\conf_{t+\eps}-\mol{\rho}\conf_{t}\right),\compm}\right)^{r}\\
& \quad+\left(\Lip_{\mol{\rho}\conf_{t},\compm}\right)^{r}\Flmass{\compm'}{\omega\circ\left(\mol{\rho}\conf_{t+\eps}\right)-\omega\circ\left(\mol{\rho}\conf_{t}\right)}.
\end{split}
\end{multline*}
In addition,
\[
\lim_{\rho\to0}\Lip_{\left(\mol{\rho}\conf_{t+\eps}-\mol{\rho}\conf_{t}\right),\compm}=\lim_{\rho\to0}\Lip_{\left(\mol{\rho}\left(\conf_{t+\eps}-\conf_{t}\right)\right),\compm}=\Lip_{\conf_{t+\eps}-\conf_{t},\compm}.
\]
We now consider the limit of the foregoing estimate when $\eps\to0$.
By the continuity of the motion it follows that $\Lip_{\conf_{t+\eps}-\conf_{t},\compm}\to0$
as $\eps\to0$. The second term in the estimate above vanishes by
the continuity of $\omega$ and the continuity of the mollified motion
$\mol{\rho}\conf$ with respect to the strong Lipschitz topology.
In conclusion,
\[
\lim_{\eps\to0}\Flmass{\compm'}{\conf_{t+\eps\#}T-\conf_{t\#}T}=0.
\]
\end{proof}
The following theorem is a generalization of Theorem \ref{thm:driv_smooth_motion}.
\begin{thm}
Let $\conf$ be a map associated with a motion as defined by Equation
(\ref{eq:Lips_map_motion}) and let $T\in F_{r,\compm}\left(\body\right)$
be a flat chain of finite mass. The derivative $\frac{d}{dt}\left(\conf_{t\#}T\right)\mid_{t=\tau}$,
exists in the topology of $\Fsharp{\compm'}{\reals^{n}}$ and is given
by
\begin{equation}
\frac{d}{dt}\left(\conf_{t\#}T\right)\mid_{t=\tau}=\bnd\left(\ve_{\tau}\wedge\conf_{\tau\#}T\right)+\ve_{\tau}\wedge\bnd\left(\conf_{\tau\#}T\right).\label{eq:ddt(conf(T))}
\end{equation}
\end{thm}
\begin{proof}
We have to show that
\[
\lim_{\eps\to0}\Fsharp{\compm'}{\frac{\conf_{\tau+\eps\#}T-\conf_{\tau\#}T}{\eps}-\left(\bnd\left(\ve_{\tau}\wedge\conf_{\tau\#}T\right)+\ve_{\tau}\wedge\bnd\left(\conf_{\tau\#}T\right)\right)}=0.
\]
In an analogous manner to the proof of theorem \ref{thm:driv_smooth_motion},
and as the homotopy theorem holds for the Lipschitz case, it follows
that
\[
\begin{split} & \frac{1}{\eps}\Fsharp{\compm'}{\emb_{t+\eps\#}T-\emb_{t\#}T-\eps\left(\bnd\left(\ve_{t}\wedge\emb_{t\#}T\right)+\ve_{t}\wedge\bnd\left(\emb_{t\#}T\right)\right)}\\
& \quad\leq\frac{1}{\eps}\sup_{\omega\in\D^{r}\left(\reals^{n}\right)}\left\{ \frac{\int_{t}^{t+\eps}\left[T\left(\emb_{\tau}^{\#}\Lie_{\ve_{\tau}}\omega\right)\right]d\lusb_{\tau}^{1}-\eps T\left(\left(\emb_{t}^{\#}\Lie_{\ve_{t}}\omega\right)\right)}{\Fsharp{\compm'}{\omega}}\right\} ,\\
& \quad\leq\sup_{\omega\in\D^{r}\left(\reals^{n}\right)}\left\{ \Fmass T\sup_{s\in[t,t+\eps]}\frac{\Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{s}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega}}{\Fsharp{\compm'}{\omega}}\right\} .
\end{split}
\]
The term $\Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{s}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega}$
may be estimated as
\begin{multline*}
\Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{s}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega}\\
\begin{split}\leq & \Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{s}}\omega-\emb_{s}^{\#}\Lie_{\ve_{t}}\omega+\emb_{s}^{\#}\Lie_{\ve_{t}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega},\\
\leq & \Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{s}}\omega-\emb_{s}^{\#}\Lie_{\ve_{t}}\omega}+\Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{t}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega},\\
\leq & \Flmass{\compm}{\emb_{s}^{\#}\left(\Lie_{\ve_{s}-\ve_{t}}\omega\right)}+\Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{t}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega}.
\end{split}
\end{multline*}
By Equation (\ref{eq:M_Lie_deriv}) we have
\begin{eqnarray*}
\Flmass{\compm}{\emb_{s}^{\#}\left(\Lie_{\ve_{s}-\ve_{t}}\omega\right)} & \leq & \left(\Lip_{\emb_{s},\compm}\right)^{r}\Flmass{\compm'}{\Lie_{\ve_{s}-\ve_{t}}\omega},\\
& \leq & \left(\Lip_{\emb_{s},\compm}\right)^{r}C(n,r)S_{\compm'}(\omega)\norm{\ve_{s}-\ve_{t}}{\Lip,\compm'}
\end{eqnarray*}
and since $\norm{\ve_{s}-\ve_{t}}{\Lip,\compm'}\to0$ as $s\to t$,
it follows that
\[
\lim_{s\to t}\Flmass{\compm}{\emb_{s}^{\#}\left(\Lie_{\ve_{s}-\ve_{t}}\omega\right)}=0.
\]
Lemma \ref{lem:conf_t(T)_Lipschitz}, implies that
\[
\lim_{s\to t}\Flmass{\compm}{\emb_{s}^{\#}\Lie_{\ve_{t}}\omega-\emb_{t}^{\#}\Lie_{\ve_{t}}\omega}=0
\]
which completes the proof.
\end{proof}
\section{The Product Rule and the Transport Theorem \label{sec:Transport_thm}}
In this section we apply the foregoing results to obtain a generalized
formulation of the transport theorem for a region of any dimension,
and in particular, the surface transport theorem. In view of the postulates
for a motion described in Section \ref{sec:Lipschitz_motion}, we
recall that for an $r$-current of finite mass, $T\in F_{r,\compm}\left(\body\right)$,
and the map $\conf:\ti\times\body\to\reals^{n}$ associated with a
motion by Equation (\ref{eq:Lips_map_motion}), one has
\[
\conf_{t\#}\left(T\right)\in F_{r,\compm'}\left(\reals^{n}\right),\quad\text{and}\quad\frac{d}{dt}\left(\conf_{t\#}\left(T\right)\right)\in S_{r,\compm'}\left(\reals^{n}\right),\quad\text{for all }t\in\ti.
\]
The standard transport theorem is concerned with the integration of
a density of some extensive property over an evolving region. Here,
the evolving region is generalized and is represented by the evolving
flat chain $\conf_{t\#}\left(T\right)$, and the extensive property
at any instant is represented by a sharp cochain. As any sharp cochain
is a flat cochain, the integration operation in the classical formulation
of the theorem is replaced by the action of a cochain on a chain.
Thus, an evolving extensive property, $\psi$, is viewed as a continuous
curve in the space of sharp $r$-cochains, that is, a continuous mapping
\[
\cochain_{\psi}:\ti\to\left[S_{r}\left(\reals^{n}\right)\right]^{*}.
\]
Moreover, we assume that the curve is differentiable in the topology
of flat $r$-cochains in $\reals^{n}$. That is, the limit
\[
\lim_{\eps\to0}\frac{\cochain_{\psi}(\tau+\varepsilon)-\cochain_{\psi}(\tau)}{\eps},
\]
exists as a flat $r$-cochain, and it will denoted by $\dot{\cochain}_{\psi}(\tau)$.
For each $t\in\ti$ the total of the extensive property in the flat
$r$-chain $T$ is therefore
\[
\cochain_{\psi}\left(\conf_{\#}\left(T\right)\right)(t)=\cochain_{\psi}(t)\left(\conf_{t\#}\left(T\right)\right).
\]
For the time derivative of $\cochain_{\psi}\left(\conf_{\#}\left(T\right)\right)$,
it easily follows that we have
\[
\begin{split}\frac{d}{dt}\cochain_{\psi}\left(\conf_{\#}\left(T\right)\right) & =\lim_{\varepsilon\to0}\frac{\cochain_{\psi}(t+\varepsilon)\left(\conf_{t+\epsilon\#}\left(T\right)\right)-\cochain_{\psi}(t)\left(\conf_{t\#}\left(T\right)\right)}{\varepsilon},\\
& =\cochain_{\psi}(t)\left(\frac{d}{dt}\conf_{t\#}\left(T\right)\right)+\frac{d\cochain_{\psi}}{dt}(t)(\conf_{t\#}\left(T\right)).
\end{split}
\]
By Equation (\ref{eq:ddt(conf(T))}), it follows that
\begin{multline}
\frac{d}{dt}\left(\cochain_{\psi}\left(\conf_{\#}\left(T\right)\right)\right)\mid_{t=\tau}\\
\begin{split} & =\dot{\cochain}_{\psi}(\tau)\left(\conf_{\tau\#}\left(T\right)\right)+\cochain_{\psi}(\tau)\left(\bnd\left(\ve_{\tau}\wedge\conf_{\tau\#}\left(T\right)\right)+\ve_{\tau}\wedge\conf_{\tau\#}\left(\bnd T\right)\right).\end{split}
\label{eq:Transport}
\end{multline}
In the case where $\cochain_{\psi}$ is represented by a smooth differential
$r$-form $\fform{\psi}$ and $\conf$ is a smooth embedding we obtain
\[
\begin{split}\frac{d}{dt}\left(\cochain_{\psi}(t)\left(\conf_{t\#}\left(T\right)\right)\right)\mid_{t=\tau} & =\left(\conf_{\tau\#}\left(T\right)\right)\left(\dot{\fform{\psi}}+\cbnd\fform{\psi}\rest\ve_{\tau}+\cbnd\left(\fform{\psi}\rest\ve_{\tau}\right)\right),\\
& =\conf_{\tau\#}\left(T\right)\left(\dot{\fform{\psi}}+\Lie_{\ve_{\tau}}\fform{\psi}\right),
\end{split}
\]
which is a generalization of \cite[equation (T0)]{Betounes1986}.
For a time dependent $r$-form $\fform{\psi}$, using \cite[pp.~367]{Marsden1988}
and the notation of Section \ref{sec:Lipschitz_motion}, it follows
that
\begin{eqnarray*}
\frac{d}{dt}\conf_{t}^{\#}\left(\fform{\psi}(t)\right)\mid_{t=\tau} & = & \conf_{\tau}^{\#}\left(\dot{\fform{\psi}}(\tau)+\cbnd\fform{\psi}(\tau)\rest\ve_{\tau}+\cbnd\left(\fform{\psi}(\tau)\rest\ve_{\tau}\right)\right).
\end{eqnarray*}
Applying the last expression to Equation (\ref{eq:Transport}), we
obtain the following relation between the Eulerian and Lagrangian
formulations of the transport theorem
\[
\frac{d}{dt}\left(\cochain_{\psi}\left(\conf_{\#}\left(T\right)\right)(t)\right)\mid_{t=\tau}=\left(\frac{d}{dt}\conf_{t}^{\#}\left(\cochain_{\psi}(t)\right)\mid_{t=\tau}\right)T.
\]
For the general case, the sharp $r$-cochain $\cochain_{\psi}(\tau)$
is represented by a sharp $r$-form $\fform{\psi}(\tau)$ and the
sharp $r$-cochain $\dot{\cochain}_{\psi}(\tau)$ is represented by
the flat $r$-form $\dot{\fform{\psi}}(\tau)$. As $T$ is a flat
chain of finite mass it may be represented by $\lusb^{n}$-integrable
vector field $T=\eta\wedge\lusb^{n}$. It is concluded that
\begin{multline*}
\frac{d}{dt}\left(\cochain_{\psi}\left(\conf_{\#}\left(T\right)\right)\right)\mid_{t=\tau}\\
\begin{split} & =\left(\conf_{\tau\#}\left(T\right)\right)\left(\dot{\fform{\psi}}(\tau)+\cbnd\fform{\psi}(\tau)\rest\ve_{\tau}+\cbnd\left(\fform{\psi}(\tau)\rest\ve_{\tau}\right)\right),\\
& =\int_{\body}\conf_{\tau}^{\#}\left(\dot{\fform{\psi}}(\tau)+\cbnd\fform{\psi}(\tau)\rest\ve_{\tau}+\cbnd\left(\fform{\psi}(\tau)\rest\ve_{\tau}\right)\right)(\eta)d\lusb_{x}^{n}.
\end{split}
\end{multline*}
As an example consider the situation where the flat form $\fform{\psi}$
satisfies Cauchy's postulates. Then, associated with the property
are a time dependent flat $r$-form $\fform{\form}$ representing
the source, and a time dependent flat $(r-1)$-form $\fform{\xi}$
representing the flux, such that the differential balance equation
for the property is (see \cite{Segev2012})
\begin{equation}
\dot{\fform{\psi}}+\cbnd\fform{\xi}=\fform{\form}.
\end{equation}
Each of the forms above represents a flat cochain denoted by $\cochain_{\omega},\:\cochain_{\xi},\;\cochain_{\form}$,
respectively, such that
\[
\dot{\cochain}_{\omega}+\cbnd\cochain_{\xi}=\cochain_{\form}.
\]
Thus, the transport formula assumes the form
\begin{multline*}
\frac{d}{dt}\left(\cochain_{\psi}\left(\conf_{\#}\left(T\right)\right)\right)\mid_{t=\tau}\\
\begin{split}= & \dot{\cochain}_{\psi}(\tau)\left(\conf_{\tau\#}\left(T\right)\right)+\cochain_{\psi}(\tau)\left(\frac{\partial}{\partial t}\conf_{t\#}\left(T\right)\mid_{t=\tau}\right),\\
= & \dot{\cochain}_{\psi}(\tau)\left(\conf_{\tau\#}\left(T\right)\right)+\cochain_{\psi}(\tau)\left(\bnd\left(\ve_{\tau}\wedge\conf_{\tau\#}T\right)+\ve_{\tau}\wedge\bnd\left(\conf_{\tau\#}T\right)\right),\\
= & \left(\cochain_{\form}\right)\left(\conf_{\#}\left(T\right)\right)(\tau)+\left(\cochain_{\psi}\rest\ve-\cochain_{\xi}\right)\left(\conf_{\#}\left(\bnd T\right)\right)(\tau)+\cochain_{\psi}\left(\bnd\left(\ve\wedge\conf_{\#}T\right)\right)(\tau).
\end{split}
\end{multline*}
\bigskip{}
\noindent \textbf{\textit{Acknowledgments.}} This work was partially
supported by the Perlstone Center for Aeronautical Engineering Studies,
the Kreitman Post-Doctoral Scholarship and the H.~Greenhill Chair
for Theoretical and Applied Mechanics at Ben-Gurion University of
the Negev. The authors wish to thank M. Silhavy for the discussions
and comments he has made on earlier versions of this work.
\bibliographystyle{plain}
|
\section{Introduction}
The complex action theory (CAT) is a trial to extend
quantum theories so that the action is complex
at a fundamental level, but effectively looks real.
So far, the CAT has been investigated
with the expectation that the imaginary part of the action
would give some falsifiable predictions\cite{Bled2006,Nielsen:2008cm,Nielsen:2007ak,Nielsen:2005ub},
and various interesting suggestions have been made for the Higgs mass\cite{Nielsen:2007mj},
quantum mechanical philosophy\cite{newer1,Vaxjo2009,newer2},
some fine-tuning problems\cite{Nielsen2010qq,degenerate},
black holes\cite{Nielsen2009hq},
de Broglie-Bohm particles, and a cut-off in loop diagrams\cite{Bled2010B}.
In the CAT, the Hamiltonian $\hat{H}$ is generically non-normal,
so it is not contained in
the class of PT-symmetric non-Hermitian Hamiltonians
that have been intensively studied\cite{Bender:1998ke,Bender:1998gh,Mostafazadeh_CPT_ip_2002,
Mostafazadeh_CPT_ip_2003}.
In ref.\cite{Nagao:2010xu},
introducing what we call the proper inner product $I_Q$
so that the eigenstates of the Hamiltonian become orthogonal to
each other with respect to it,
we presented a mechanism to effectively
obtain a Hamiltonian that is $Q$-Hermitian, i.e.,
Hermitian with respect to the proper inner product,
after a long time development.
In ref.\cite{Nagao:2011za}, we proposed
a complex coordinate and momentum formalism by
explicitly constructing non-Hermitian operators
of complex coordinate $q$ and momentum $p$
and their eigenstates, so that we can deal with complex $q$ and
$p$ properly.
In general, the CAT could be classified into two theories:
one is the future-not-included theory, i.e., the theory
including only a past time as an integration interval of time,
and the other one is the future-included theory\cite{Bled2006},
which includes not only a past time but also a future time.
Using the complex coordinate and momentum
formalism\cite{Nagao:2011za} in the Feynman path integral,
we found that the momentum relation
is given by the usual expression $p= m \dot{q}$, where
$m$ is a complex mass, in the future-included theory\cite{Nagao:2011is}, and another expression
$p=\left(m_R + \frac{m_I^2}{m_R}\right) \dot{q}$,
where $m_R$ and $m_I$ are the real and imaginary parts of $m$,
respectively,
in the future-not-included theory\cite{Nagao:2013eda}.
The future-included theory is described by using
the future state $| B (T_B) \rangle$ at the final time $T_B$
and the past state $| A (T_A) \rangle$ at the initial time $T_A$.
In refs.\cite{Nagao:2012mj,Nagao:2012ye}
we studied the normalized matrix element\footnote{In the real action theory (RAT), the normalized matrix element
$\langle \hat{\cal O} \rangle^{BA}$ is called the weak value\cite{AAV}, and has been intensively studied.
For details, see ref.\cite{review_wv} and references therein.}
$\langle \hat{\cal O} \rangle^{BA}
\equiv \frac{ \langle B(t) | \hat{\cal O} | A(t) \rangle }{ \langle B(t) | A(t) \rangle }$,
where $t$ is an arbitrary time ($T_A \leq t \leq T_B$),
in the future-included theory, and found that,
if we regard $\langle \hat{\cal O} \rangle^{BA}$
as an expectation value in the future-included theory, then we
obtain the Heisenberg equation, Ehrenfest's theorem, and a conserved probability current density.
This suggests that $\langle \hat{\cal O} \rangle^{BA}$ is a
strong candidate for an expectation value
in the future-included theory.
In this letter we study in the future-included CAT
a slightly modified quantity
$\langle \hat{\cal O} \rangle_Q^{BA}
\equiv \frac{ \langle B(t) |_Q \hat{\cal O} | A(t) \rangle }{ \langle B(t) |_Q A(t) \rangle }$,
where $\langle B(t)|_Q\equiv \langle B(t)|Q$, and $Q$ is a Hermitian operator\footnote{In the special case of the Hamiltonian $\hat{H}$ being
Hermitian, $Q$ is just a unit operator,
so $\langle \hat{\cal O} \rangle_Q^{BA}$
corresponds to $\langle \hat{\cal O} \rangle^{BA}$.} that is used
to define the proper inner product $I_Q$.
The choice of $\langle \hat{\cal O} \rangle_Q^{BA}$ or
$\langle \hat{\cal O} \rangle^{BA}$ is
only a matter of notation as to
what the state symbol $\langle B(t)|$ shall precisely mean.
On the other hand,
the choice of the inner product used in the normalization of the initial and final states $|A(T_A) \rangle$ and $\langle B(T_B)|$
is not just a matter of notation, once we have chosen
$\langle \hat{\cal O} \rangle_Q^{BA}$
as the expression of the candidate for our expectation value.
That is to say, according to the choice of the inner product
used in the normalization of
the initial and final states, two slightly different versions
could be defined.
The normalization defined with the usual inner product
$I$ has the true meaning of normalization, of course,
but includes unphysical transitions between different
eigenstates with different eigenvalues of the Hamiltonian $\hat{H}$.
The normalization defined with the proper inner product $I_Q$,
which we call $Q$-normalization, excludes such unphysical transitions,
but does not have the original meaning of normalization.
Thus, each choice seems to have both advantages and disadvantages,
so we are interested in the study of both versions.
However, let us admit that, in the version with the usually normalized
initial and final states, it is not easy to evaluate
$\langle \hat{\cal O} \rangle_Q^{BA}$ clearly, because
we cannot exhaustively make use of the orthogonality
of the eigenstates of the Hamiltonian $\hat{H}$.
Therefore, we postpone the study of this version to the future,
and concentrate in this letter on the analysis of the version
with the $Q$-normalized initial and final states,
which is much easier to study than the other version.
Assuming that a given Hamiltonian $\hat{H}$
is non-normal but diagonalizable, and that
the imaginary parts of the eigenvalues
of $\hat{H}$ are bounded from above,
we present a theorem that claims that
$\langle \hat{\cal O} \rangle_Q^{BA}$ becomes real
and time-develops under a $Q$-Hermitian Hamiltonian
for any $Q$-Hermitian operator $\hat{\cal O}$,
provided that $| B (t) \rangle$ and $| A (t) \rangle$
are the time-developed states
maximizing the absolute value of the transition amplitude
$|\langle B(t) |_Q A(t) \rangle|$.
Such states would represent an approximation to
$|\langle B(t) |_Q A(t) \rangle|$
in the situation that $| B (T_B) \rangle$ and $| A (T_A) \rangle$
were randomly given.
In fact, in the large $T\equiv T_B-T_A$ case,
only terms associated with the largest imaginary parts of the eigenvalues of the Hamiltonian would dominate,
and even with random initial and final states
the dominant term would give the biggest value.
We call this thinking the maximization principle.
We shall prove this theorem by finding that
$\langle \hat{\cal O} \rangle_Q^{BA}$
for the states maximizing $|\langle B(t) |_Q A(t) \rangle|$
becomes an expression similar to
an expectation value defined with $I_Q$ in the future-not-included
theory.
Indeed, it is very important to obtain a real expectation value
and a Hermitian Hamiltonian in the CAT so that it can
survive as a possible true fundamental quantum theory.
The maximization principle
is regarded as a method of obtaining not only a real expectation
value but also a $Q$-Hermitian Hamiltonian.
Furthermore, assuming that the non-normal Hamiltonian given
at first is written in terms of the Hermitian
coordinate and momentum operators $\hat{q}$ and $\hat{p}$,
we give a possible procedure to formulate
the $Q$-Hermitian Hamiltonian
in terms of $Q$-Hermitian coordinate and momentum
operators $\hat{q}_Q$ and $\hat{p}_Q$.
We also provide a $Q$-Hermitian
probability density operator
and construct a conserved probability current density.
\section{Proper inner product and future-included complex action theory}
We consider a general non-normal diagonalizable
Hamiltonian $\hat{H}$, i.e., $[\hat{H}, \hat{H}^{\dagger}]\neq 0$,
for a general quantum mechanical system that
could be the whole world, and review a
proper inner product for $\hat{H}$ that makes $\hat{H}$
normal with respect to it by following refs.\cite{Nagao:2010xu,Nagao:2011za}.
We define the eigenstates
$| \lambda_i \rangle (i=1,2,\dots)$ of $\hat{H}$ such that
\begin{equation}
\hat{H} | \lambda_i \rangle = \lambda_i | \lambda_i \rangle,
\end{equation}
where $\lambda_i (i=1,2,\dots)$ are the eigenvalues of
$\hat{H}$, and introduce the diagonalizing operator
$P=(| \lambda_1 \rangle , | \lambda_2 \rangle , \ldots)$,
so that $\hat{H}$ is diagonalized as
$\hat{H} = PD P^{-1}$,
where $D$ is given by $\text{diag}(\lambda_1, \lambda_2, \dots)$.
Let us consider a transition from an eigenstate
$| \lambda_i \rangle$ to another
$| \lambda_j \rangle ~(i \neq j)$ fast in time $\Delta t$.
Since $| \lambda_i \rangle$ are not orthogonal to
each other in the usual inner product $I$,
$I(| \lambda_i \rangle , | \lambda_j \rangle ) \equiv \langle \lambda_i | \lambda_j \rangle \neq \delta_{ij}$,
the transition can be measured, i.e.,
$|I(| \lambda_j \rangle, \exp\left( -\frac{i}{\hbar} \hat{H} \Delta t \right) |\lambda_i \rangle )|^2 \neq 0$,
though $\hat{H}$ cannot bring the system from $| \lambda_i \rangle$
to $| \lambda_j \rangle ~(i \neq j)$.
Such an unphysical transition from one eigenstate to another
with a different eigenvalue should be prohibited in a reasonable theory.
In order to have reasonable probabilistic results,
we introduce
a proper inner product \cite{Nagao:2010xu, Nagao:2011za}\footnote{Similar inner products are also studied
in refs.\cite{Geyer,Mostafazadeh_CPT_ip_2002,Mostafazadeh_CPT_ip_2003}.}
for arbitrary kets $|u \rangle$ and $|v \rangle$ as
\begin{equation}
I_Q(|u \rangle , |v \rangle) \equiv \langle u |_Q v \rangle
\equiv \langle u | Q | v \rangle,
\end{equation}
where $Q$ is a Hermitian operator chosen as
$Q=(P^\dag)^{-1} P^{-1}$,
so that $| \lambda_i \rangle$ become orthogonal to each other
with regard to $I_Q$:
\begin{equation}
\langle \lambda_i |_Q \lambda_j \rangle = \delta_{ij}.
\end{equation}
This implies the orthogonality relation
$\sum_i | \lambda_i \rangle \langle \lambda_i |_{Q} = 1$.
In the special case of the Hamiltonian $\hat{H}$ being
Hermitian, $Q$ would be the unit operator.
We introduce the ``$Q$-Hermitian'' conjugate $\dag^Q$ of
an operator $A$ by
$\langle \psi_2 |_Q A | \psi_1 \rangle^* \equiv \langle \psi_1 |_Q A^{\dag^Q} | \psi_2 \rangle$, so
\begin{equation}
A^{\dag^Q} \equiv Q^{-1} A^\dag Q.
\end{equation}
If $A$ obeys $A^{\dag^Q} = A$, $A$ is $Q$-Hermitian.
We also define $\dag^Q$ for kets and bras as
$| \lambda \rangle^{\dag^Q} \equiv \langle \lambda |_Q $ and
$\left(\langle \lambda |_Q \right)^{\dag^Q} \equiv | \lambda \rangle$.
In addition,
$P^{-1}=
\left(
\begin{array}{c}
\langle \lambda_1 |_Q \\
\langle \lambda_2 |_Q \\
\vdots
\end{array}
\right)$
satisfies $P^{-1} \hat{H} P = D$ and
$P^{-1} \hat{H}^{\dag^Q} P = D^{\dag}$,
so $\hat{H}$ is ``$Q$-normal'',
$[\hat{H}, \hat{H}^{\dag^Q} ] = P [D, D^\dag ] P^{-1} =0$.
Thus the inner product $I_Q$ makes $\hat{H}$
$Q$-normal.
We note that $\hat{H}$ can be decomposed as
$\hat{H}=\hat{H}_{Qh} + \hat{H}_{Qa}$,
where $\hat{H}_{Qh}= \frac{\hat{H} + \hat{H}^{\dag^Q} }{2}$ and
$\hat{H}_{Qa} = \frac{\hat{H} - \hat{H}^{\dag^Q} }{2}$ are
$Q$-Hermitian and anti-$Q$-Hermitian parts of $\hat{H}$, respectively.
In refs.\cite{Bled2006,Nagao:2012mj,Nagao:2012ye}
the future-included theory is
described by using
the future state $| B (T_B) \rangle$ at the final time $T_B$
and the past state $| A (T_A) \rangle$ at the initial time $T_A$,
where $| A (T_A) \rangle$ and $| B (T_B) \rangle$
time-develop as follows:
\begin{eqnarray}
&&i \hbar \frac{d}{dt} | A(t) \rangle = \hat{H} | A(t) \rangle , \label{schro_eq_Astate} \\
&&-i \hbar \frac{d}{dt} \langle B(t) |
= \langle B(t) | \hat{H} , \label{schro_eq_Bstate_old}
\end{eqnarray}
and the ``normalized'' matrix element
$\langle \hat{\cal O} \rangle^{BA}
\equiv \frac{ \langle B(t) | \hat{\cal O} | A(t) \rangle }{ \langle B(t) | A(t) \rangle }$
is studied.
The quantity $\langle \hat{\cal O} \rangle^{BA}$
is called the weak value\cite{AAV,review_wv}
in the real action theory (RAT).
In refs.\cite{Nagao:2012mj,Nagao:2012ye}
we investigated $\langle \hat{\cal O} \rangle^{BA}$, and found that,
if we regard $\langle \hat{\cal O} \rangle^{BA}$
as an expectation value in the future-included theory,
then we obtain the Heisenberg equation, Ehrenfest's theorem,
and a conserved probability current density.
Thus $\langle \hat{\cal O} \rangle^{BA}$ seems to play the
role of an expectation value in the future-included theory.
In this letter, we adopt the proper inner product $I_Q$
for all quantities,
and hence slightly modify the final state
$\langle B(T_B) |$ as
$\langle B(T_B) | \rightarrow \langle B(T_B) |_Q$
so that the Hermitian operator $Q$ pops out
and the usual inner product $I$ is replaced with $I_Q$.
Our new final state $\langle B(T_B) |$ time-develops
according not to
eq.(\ref{schro_eq_Bstate_old}) but to
\begin{eqnarray}
-i \hbar \frac{d}{dt} \langle B(t) |_Q
= \langle B(t) |_Q \hat{H}
\quad \Leftrightarrow \quad
i \hbar \frac{d}{dt} | B(t) \rangle = {\hat{H}}^{\dag^Q} | B(t) \rangle .
\label{schro_eq_Bstate}
\end{eqnarray}
Thus the normalized matrix element
$\langle \hat{\cal O} \rangle^{BA}$ is modified into
the following expression:
\begin{equation}
\langle \hat{\cal O} \rangle_Q^{BA}
\equiv \frac{ \langle B(t) |_Q \hat{\cal O} | A(t) \rangle }{ \langle B(t) |_Q A(t) \rangle },
\end{equation}
where $I_Q$ is used for both the denominator and numerator.
As far as the construction of $\langle \hat{\cal O} \rangle_Q^{BA}$
is concerned,
the shift between $\langle B(t)|$ and $\langle B(t)|_Q$ is
just a change of notation, but when it comes to our
maximization principle, we need to normalize
the initial and final states $|A(T_A) \rangle$ and $\langle B(T_B)|$.
There are two choices:
the normalization
defined with the usual inner product $I$
or
the normalization defined with
the proper inner product $I_Q$,
which we call $Q$-normalization.
The choice of the inner product used in the normalization
is not just a matter of notation,
once we have chosen
$\langle \hat{\cal O} \rangle_Q^{BA}$
as the expression of the candidate for our expectation value.
That is to say, according to the choice of the inner product
used in the normalization of
the initial and final states, two slightly different versions
could be defined.
As we have explained in the introduction,
each choice seems to have both advantages and disadvantages,
and it is not easy to evaluate
$\langle \hat{\cal O} \rangle_Q^{BA}$
clearly in the version with the usually normalized
initial and final states, because
we cannot exhaustively make use of the orthogonality
of the eigenstates of the Hamiltonian $\hat{H}$.
Therefore, we postpone the study of this version to the future, and in the following
we investigate
the quantity $\langle \hat{\cal O} \rangle_Q^{BA}$
with the $Q$-normalized initial and final states $|A(T_A) \rangle$ and $\langle B(T_B)|$,
which is much easier to study than the other version.
\section{Theorem on the normalized matrix element and its proof}
We present the following theorem:
{\bf Theorem}
{\em
As a prerequisite, assume that a given Hamiltonian
$\hat{H}$ is non-normal but diagonalizable
and that the imaginary parts of the eigenvalues
of $\hat{H}$ are bounded from above,
and define a modified inner product $I_Q$ by means
of a Hermitian operator $Q$ arranged so
that $\hat{H}$ becomes normal with respect to $I_Q$.
Let the two states $| A(t) \rangle$ and $ | B(t) \rangle$
time-develop according to the Schr\"{o}dinger
equations\footnote{See eqs.(\ref{schro_eq_Astate}) and (\ref{schro_eq_Bstate}).}
with $\hat{H}$ and $\hat{H}^{\dag^Q}$, respectively:
$|A (t) \rangle =
e^{-\frac{i}{\hbar}\hat{H} (t-T_A) }| A(T_A) \rangle$,
$|B (t) \rangle =
e^{-\frac{i}{\hbar} {\hat{H}}^{\dag^Q} (t-T_B) }
| B(T_B)\rangle$,
and be normalized with $I_Q$
at the initial time $T_A$ and the final time $T_B$, respectively:
$\langle A(T_A) |_{Q} A(T_A) \rangle = 1$,
$\langle B(T_B) |_{Q} B(T_B) \rangle = 1$.
Next determine
$|A(T_A) \rangle$
and $|B(T_B) \rangle$ so as to maximize
the absolute value of the transition
amplitude $|\langle B(t) |_Q A(t) \rangle|=
|\langle B(T_B)|_Q \exp(-i\hat{H}(T_B-T_A))
|A(T_A) \rangle|$.
Then, provided that an operator $\hat{\cal O}$
is $Q$-Hermitian, i.e., Hermitian with respect to
the inner product $I_Q$,
$\hat{\cal O}^{\dag^Q} = \hat{\cal O}$,
the normalized matrix element of
the operator $\hat{\cal O}$ defined by
$\langle \hat{\cal O} \rangle_Q^{BA}
\equiv
\frac{\langle B(t) |_Q \hat{\cal O} | A(t) \rangle}{\langle B(t) |_Q A(t) \rangle}$
becomes {\rm real}
and time-develops under
a {\rm $Q$-Hermitian} Hamiltonian. }
\vspace*{0.5cm}
Before proving the theorem,
we make a couple of remarks on it.
The procedure of maximizing
the absolute
value of the transition amplitude
$|\langle B(t) |_Q A(t) \rangle|$,
which we call the maximization principle,
can be understood as an approximation
to what will be with very large likelihood
the result of just taking the initial state $| A(T_A) \rangle$
and the final state $| B(T_B) \rangle$
at random.
In fact, we would like to show
in a later publication that with the random states
$| A(T_A) \rangle$ and $| B(T_B) \rangle$
we obtain approximately
the same result for $\langle \hat{\cal O} \rangle_Q^{BA}$
as if we used the
maximization principle as just stated in the theorem.
The crucial point of the theorem is that
$\langle \hat{\cal O} \rangle_Q^{BA}$, which is taken as an average
for an operator $\hat{\cal O}$ obeying $\hat{\cal O}^{\dag^Q} =\hat{\cal O}$,
turns out to be real almost unavoidably.
This is under the restriction that $\hat{H}$ be $Q$-normal,
i.e., normal with regard to the proper inner product $I_Q$,
but that $\hat{H}$ is not required to be $Q$-Hermitian,
$\hat{H} \neq \hat{H}^{\dag^Q}$.
Now let us prove the above theorem
by expanding $| A(t) \rangle$ and $| B(t) \rangle$
in terms of the eigenstates $| \lambda_i \rangle$
as follows:
$|A (t) \rangle = \sum_i a_i (t) | \lambda_i \rangle$,
$|B (t) \rangle = \sum_i b_i (t) | \lambda_i \rangle$,
where
$a_i (t) = a_i (T_A) e^{-\frac{i}{\hbar}\lambda_i (t-T_A) }$,
$b_i (t) = b_i (T_B) e^{-\frac{i}{\hbar}\lambda_i^* (t-T_B) }$.
Since $\langle B (t) |_Q A (t) \rangle$ is expressed as
$\langle B (t) |_Q A (t) \rangle
= \sum_i R_i e^{i \Theta_i}$,
where we have introduced
$a_i(T_A)= | a_i(T_A) | e^{i \theta_{a_i}}$,
$b_i(T_B) = | b_i(T_B) | e^{i \theta_{b_i}}$,
$T\equiv T_B - T_A$,
$R_i
\equiv |a_i (T_A)| |b_i (T_B)| e^{\frac{1}{\hbar} T \text{Im} \lambda_i }$,
and $\Theta_i \equiv \theta_{a_i} - \theta_{b_i}
- \frac{1}{\hbar} T \text{Re} \lambda_i$,
$| \langle B (t) |_Q A (t) \rangle |^2$ is calculated as
$| \langle B (t) |_Q A (t) \rangle |^2
= \sum_i R_i^2 + 2 \sum_{i<j} R_i R_j \cos(\Theta_i - \Theta_j)$.
On the other hand, the normalization conditions
are expressed as
$\sum_i | a_i (T_A) |^2 = 1$ and
$\sum_i | b_i (T_B) |^2 = 1$, respectively.
Here we note that
the imaginary parts of the eigenvalues
of $\hat{H}$ have to be bounded from above
to avoid the Feynman path integral
$\int e^{\frac{i}{\hbar}S} {\cal D} \text{path}$
being divergently meaningless.
So we assume that some of the $\text{Im} \lambda_i$
take the maximal value $B$, and denote
the corresponding subset of $\{ i \}$ as $A$.
Then, since $R_i \geq 0$,
$| \langle B (t) |_Q A (t) \rangle |$ can take
a maximal value only under the following conditions:
\begin{eqnarray}
&& | a_i (T_A) | = | b_i (T_B) | =0 \quad \text{for $\forall i \notin A$} , \label{abinotinA0} \\
&& \Theta_i
\equiv \Theta_c
\quad \text{for $\forall i \in A$} \label{max_cond_theta} , \\
&& \sum_{i \in A} | a_i (T_A) |^2 =\sum_{i \in A}|b_i (T_B)|^2 = 1, \label{nc_ATABTB3}
\end{eqnarray}
and $| \langle B (t) |_Q A (t) \rangle |^2$ is estimated as
\begin{eqnarray}
| \langle B (t) |_Q A (t) \rangle |^2
&=& \left( \sum_{i \in A} R_i \right)^2 \nonumber \\
&=& e^{\frac{2 B T}{\hbar} }
\left( \sum_{i \in A} |a_i (T_A)| |b_i (T_B)| \right)^2 \nonumber \\
&\leq& e^{\frac{2 B T}{\hbar} }
\left\{ \sum_{i \in A} \left( \frac{ |a_i (T_A)| + |b_i (T_B)|}{2} \right)^2 \right\}^2
=e^{\frac{2}{\hbar} B T} ,
\end{eqnarray}
where the third equality is realized for
\begin{equation}
|a_i (T_A)| = |b_i (T_B)| \quad \text{for $\forall i \in A$}.
\label{max_cond_ab}
\end{equation}
In the last equality we have used this relation
and eq.(\ref{nc_ATABTB3}).
The maximization condition of $| \langle B (t) |_Q A (t) \rangle |$
is represented by
eqs.(\ref{abinotinA0})-(\ref{nc_ATABTB3}) and (\ref{max_cond_ab}).
That is to say,
the states to maximize $| \langle B (t) |_Q A (t) \rangle |$,
$| A(t) \rangle_{\rm{max}}$ and $| B(t) \rangle_{\rm{max}}$, are
expressed as
\begin{eqnarray}
&&|A (t) \rangle_{\rm{max}} = \sum_{i \in A} a_i (t) | \lambda_i \rangle ,
\label{Atketmax_sum_inA_ai} \\
&&|B (t) \rangle_{\rm{max}} = \sum_{i \in A} b_i (t) | \lambda_i \rangle,
\label{Btketmax_sum_inA_bi}
\end{eqnarray}
where $a_i (t)$ and $b_i (t)$ obey
eqs.(\ref{max_cond_theta}), (\ref{nc_ATABTB3}), and (\ref{max_cond_ab}).
Intuitively, it might be rather obvious that, to
get the biggest transition amplitude
$| \langle B (t) |_Q A (t) \rangle |$ for states
$| A (t) \rangle $ and $| B(t) \rangle$ normalized
at the initial time $T_A$ and the final time
$T_B$, respectively, we should
seek the eigenstates leading to the
biggest increase with time development under the Schr\"{o}dinger equations, i.e., with the biggest
imaginary parts of the eigenvalues of $\hat{H}$.
We evaluate $\langle \hat{\cal O} \rangle_Q^{BA}$
for $| A(t) \rangle_{\rm{max}} $ and $| B(t) \rangle_{\rm{max}}$.
Using eqs.(\ref{abinotinA0})-(\ref{nc_ATABTB3}) and (\ref{max_cond_ab}),
we obtain
${}_{\rm{max}} \langle B (t) |_Q A (t) \rangle_{\rm{max}}
= e^{i \Theta_c} \sum_{i \in A} R_i
= e^{i \Theta_c} e^{\frac{B T}{\hbar} }$,
and
\begin{eqnarray}
{}_{\rm{max}} \langle B (t) |_Q \hat{\cal O} | A (t) \rangle_{\rm{max}}
&=&
e^{i \Theta_c} e^{\frac{B T}{\hbar} }
\sum_{i , j \in A} a_j(T_A)^* a_i(T_A)
e^{\frac{i}{\hbar}(t-T_A) (\text{Re}\lambda_j - \text{Re}\lambda_i)}
\langle \lambda_j |_Q \hat{\cal O} | \lambda_i \rangle \nonumber \\
&=&
e^{i \Theta_c} e^{\frac{B T}{\hbar} }
\langle \tilde{A}(t) |_Q \hat{\cal O} | \tilde{A}(t) \rangle ,
\end{eqnarray}
where we have introduced
$| \tilde{A}(t) \rangle \equiv
e^{-\frac{i}{\hbar}(t-T_A) \hat{H}_{Qh}} | A(T_A) \rangle_{\rm{max}}$,
which is normalized as
$\langle \tilde{A}(t) |_Q \tilde{A}(t) \rangle = 1$
and obeys the Schr\"{o}dinger equation
\begin{eqnarray}
i\hbar \frac{d}{d t}| \tilde{A}(t) \rangle
&=& \hat{H}_{Qh} | \tilde{A}(t) \rangle . \label{ScheqAtildetket}
\end{eqnarray}
Thus the normalized matrix element
$\langle \hat{\cal O} \rangle_Q^{BA}$
for $| A(t) \rangle_{\rm{max}}$ and $| B(t) \rangle_{\rm{max}}$
is evaluated as
\begin{eqnarray}
\langle \hat{\cal O} \rangle_Q^{BA}
&=&
\langle \tilde{A}(t) |_Q \hat{\cal O} | \tilde{A}(t) \rangle
\equiv
\langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}} . \label{OBAmaxtilde}
\end{eqnarray}
Now we see that $\langle \hat{\cal O} \rangle_Q^{BA}$
for $| A(t) \rangle_{\rm{max}} $ and $| B(t) \rangle_{\rm{max}}$
has become the form of an average
defined with the proper inner product $I_Q$.
Since the complex conjugate of
$\langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}}$
is expressed as
$\left\{ \langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}} \right\}^*=\langle \hat{\cal O}^{\dag^Q} \rangle_Q^{\tilde{A} \tilde{A}}$,
$\langle \hat{\cal O} \rangle_Q^{BA}$
for $| A(t) \rangle_{\rm{max}} $ and $| B(t) \rangle_{\rm{max}}$
is shown to be real for $Q$-Hermitian $\hat{\cal O}$.
Next we study the time development of
$\langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}}$.
We express $\langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}}$ as
$\langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}}
=\langle \tilde{A}(T_A) |_Q$ $\hat{\cal O}_{H}(t, T_A)
| \tilde{A}(T_A) \rangle$,
where we have introduced the Heisenberg operator
$\hat{\cal O}_{H}(t, T_A)
\equiv
e^{ \frac{i}{\hbar} \hat{H}_{Qh} (t-T_A) }$
$\hat{\cal O}
e^{ -\frac{i}{\hbar} \hat{H}_{Qh} (t-T_A)}$.
This operator $\hat{\cal O}_{H}(t, T_A)$ obeys
the Heisenberg equation
\begin{equation}
i\hbar \frac{d}{d t} \hat{\cal O}_{H}(t, T_A)
= [ \hat{\cal O}_{H}(t, T_A) , \hat{H}_{Qh} ],
\end{equation}
so we find that
$\langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}}$ time-develops
under the $Q$-Hermitian Hamiltonian $\hat{H}_{Qh}$ as
\begin{eqnarray}
\frac{d}{dt} \langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}}
&=&
\frac{i}{\hbar}
\langle \left[ \hat{H}_{Qh}, \hat{\cal O} \right]
\rangle_Q^{\tilde{A} \tilde{A}} .
\label{ddtOAtildeAtildeQ}
\end{eqnarray}
Now, for pedagogical reasons, let us suppose that
$\langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}}$ time-develops
under some Hamiltonian $\hat{H}_{1}$ as
$\frac{d}{dt} \langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}}
=\frac{i}{\hbar}
\langle \left[ \hat{H}_{1}, \hat{\cal O} \right] \rangle_Q^{\tilde{A} \tilde{A}}$.
The complex conjugate of this relation is given by
$\left\{ \frac{d}{dt} \langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}} \right\}^*
=\frac{i}{\hbar}
\langle \left[ \hat{H}_{1}^{\dag^Q}, \hat{\cal O}^{\dag^Q} \right] \rangle_Q^{\tilde{A} \tilde{A}}$.
Since $\langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}}$
is real for $Q$-Hermitian $\hat{\cal O}$,
these relations
claim that $\hat{H}_{1}$ has to be $Q$-Hermitian.
Therefore,
the reality of $\langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}}$
implies that it has to time-develop
under some $Q$-Hermitian Hamiltonian.
As shown in eq.(\ref{ddtOAtildeAtildeQ}),
$\langle \hat{\cal O} \rangle_Q^{\tilde{A} \tilde{A}}$ time-develops
under $\hat{H}_{Qh}$, which is consistent with the implication.
We emphasize that
the maximization principle
provides not only
the reality of $\langle \hat{\cal O} \rangle_Q^{BA}$
for $Q$-Hermitian $\hat{\cal O}$
but also the $Q$-Hermitian Hamiltonian.
\section{Discussion}
In this letter,
we first reviewed the proper inner product $I_Q$
defined with a Hermitian operator $Q$, which
is constructed from a diagonalizing operator of
a given non-normal diagonalizable Hamiltonian $\hat{H}$,
so that the eigenstates of $\hat{H}$
become orthogonal to each other with regard to
the proper inner product $I_Q$,
and the $Q$-Hermitian conjugate $\dag^Q$, i.e.,
Hermitian conjugate with regard to $I_Q$.
We also explained the property of
the normalized matrix element $\langle \hat{\cal O} \rangle^{BA}=\frac{\langle B(t) | \hat{\cal O} | A(t) \rangle}{\langle B(t) | A(t) \rangle}$
in the future-included complex action theory (CAT).
Next we introduced a slightly modified
normalized matrix element
$\langle \hat{\cal O} \rangle_Q^{BA}=\frac{\langle B(t) |_Q \hat{\cal O} | A(t) \rangle}{\langle B(t) |_Q A(t) \rangle}$,
which is defined with $I_Q$,
and explained that two versions could be defined
according to the choice of the normalization of
the initial and final states $|A(T_A)\rangle$ and $\langle B(T_B)|$.
One is the usual normalization defined with the usual inner
product $I$,
and the other is the $Q$-normalization defined with
the proper inner product $I_Q$.
Assuming that a given Hamiltonian $\hat{H}$ is
non-normal but diagonalizable, and that
the imaginary parts of the eigenvalues of $\hat{H}$
are bounded from above,
we presented a theorem
that states that, provided that $\hat{\cal O}$ is $Q$-Hermitian, i.e.,
$\hat{\cal O}^{\dag^Q}=\hat{\cal O}$,
and that $|A(t) \rangle $ and $|B(t) \rangle$ time-develop
according to the Schr\"{o}dinger equations
with $\hat{H}$ and $\hat{H}^{\dag^Q}$ and are
$Q$-normalized
at the initial time $T_A$ and at the final time $T_B$, respectively,
$\langle \hat{\cal O} \rangle_Q^{BA}$ becomes real
and time-develops under a $Q$-Hermitian Hamiltonian
for $|A(t) \rangle $ and $|B(t) \rangle$ such that the
absolute value of the transition amplitude
$|\langle B(t)|_Q A(t) \rangle|$ is maximized.
We proved the theorem by expanding
$|A(t) \rangle $ and $|B(t) \rangle$ in terms of
the eigenstates of $\hat{H}$.
It is noteworthy that, in the future-included CAT
with a priori non-normal Hamiltonian $\hat{H}$,
we nevertheless have got a real average for $\hat{\cal O}$
at any time $t$
by means of the simple expression
$\langle \hat{\cal O} \rangle_Q^{BA}$.
As for an emerging hermiticity, in ref.\cite{Nagao:2010xu}
we presented a mechanism to obtain
a $Q$-Hermitian Hamiltonian
by considering a long time development.
The maximization principle studied in this letter
is another approach to obtaining
such a $Q$-Hermitian Hamiltonian.
We have seen that the non-hermiticity of
the fundamental Hamiltonian $\hat{H}$ has
disappeared
from the usually expected results of the model.
It is this remarkable result of our works with
non-Hermitian Hamiltonians or complex actions
that allows us to consider such models to be viable.
We would not have been able to see any effects
of the anti-Hermitian part as far as the reality
of the dynamical variables and the equations of motion
are concerned.
However, as earlier discussed in ref.\cite{Bled2006}
and also seen in eqs.(\ref{abinotinA0})-(\ref{nc_ATABTB3}),
the anti-Hermitian part has a strong influence
on the initial state, which should effectively be seen.
Indeed, the maximization principle has resulted in a periodicity
of the history of the universe
that the initial and final states become basically the same.
Such an influence would be more recognizable
in a system defined with a time-dependent non-Hermitian
Hamiltonian\cite{Fukuma:2013mx}.
We expect the future-included CAT to
have the feature that it can provide a unification of an
initial condition prediction and an equation of motion.
In this letter, we studied the version defined with
the $Q$-normalized initial and final states.
It would be interesting to see what kind of result
we could obtain in the other version
defined with the usually normalized initial and final states,
which is more difficult to study than the
the version studied here,
because we cannot fully utilize the
orthogonality of the eigenstates of the Hamiltonian $\hat{H}$.
In the future we hope to investigate this version and
to see if the reality of $\langle \hat{\cal O} \rangle_Q^{BA}$,
emerging Hermitian Hamiltonian, and such a periodicity
are suggested or not.
Finally, assuming that the fundamental non-normal Hamiltonian
$\hat{H}$ is written in terms of Hermitian
coordinate and momentum operators $\hat{q}$ and $\hat{p}$
as $\hat{H}=H(\hat{q}, \hat{p})$,
we give a possible procedure\footnote{For simplicity,
we do not use
the complex coordinate and
momentum formalism\cite{Nagao:2011za}
just by supposing the case where the eigenvalues $q$ and $p$
are essentially real.
If we like, we could generalize the argument here
by following ref.\cite{Nagao:2011za}
so that we could deal with complex $q$ and $p$.}
to formulate the $Q$-Hermitian Hamiltonian $\hat{H}_{Qh}$
in terms of $Q$-Hermitian
coordinate and momentum operators $\hat{q}_{Q}$ and $\hat{p}_{Q}$.
We also introduce a $Q$-Hermitian probability density operator as
an example of $Q$-Hermitian $\hat{\cal O}$, and construct
a conserved probability current density.
Let us begin with defining
$\hat{q}_{Q}$ and $\hat{p}_{Q}$ by
\begin{eqnarray}
&&\hat{q}_{Q} \equiv \frac{\hat{q} + \hat{q}^{\dag^Q} }{2 }, \\
&&\hat{p}_{Q} \equiv \frac{\hat{p} + \hat{p}^{\dag^Q} }{2 }.
\end{eqnarray}
Since $Q$ depends on $\hat{q}$ and $\hat{p}$ via $\hat{H}$,
$\hat{q}_{Q}$ and $\hat{p}_{Q}$ could be written
in terms of $\hat{q}$ and $\hat{p}$,
and vice versa.\footnote{In the harmonic oscillator model\cite{KNHBN_ho} defined by the Hamiltonian
$\hat{H}_{\rm{ho}} \equiv \frac{\hat{p}^2}{2m}+\frac{1}{2}m\omega^2 \hat{q}^2$
with a mass $m$ and an angular frequency $\omega$, we obtain
$\hat{q}_{Q}= e^{i \frac{\theta}{2}} \hat{q}$ and
$\hat{p}_{Q}= e^{-i \frac{\theta}{2}} \hat{p}$, where
$\theta=\arg (m\omega)$. $\hat{H}_{\rm{ho}}$ is rewritten as $\hat{H}_{\rm{ho}}=\frac{\hat{p}_Q^2}{2m_{\rm{eff}}}+\frac{1}{2}m_{\rm{eff}}\omega^2 \hat{q}_Q^2$, where
$m_{\rm{eff}}=me^{-i\theta}$.}
Then $\hat{H}$ would be rewritten as
$\hat{H}=H_{\rm{eff}}(\hat{q}_{Q}, \hat{p}_{Q})$,
where $H_{\rm{eff}}$ is some analytic function of
$\hat{q}_{Q}$ and $\hat{p}_{Q}$,
and $\hat{H}_{Qh}$ is expressed in terms of
$\hat{q}_{Q}$ and $\hat{p}_{Q}$ as
\begin{equation}
\hat{H}_{Qh}
=\frac{1}{2} \left( H_{\rm{eff}}(\hat{q}_{Q}, \hat{p}_{Q})
+ H_{\rm{eff}}(\hat{q}_{Q}, \hat{p}_{Q})^{\dag^Q} \right) .
\end{equation}
Next we define $|q \rangle^Q$
as the eigenstate of $\hat{q}_{Q}$ by
$\hat{q}_{Q} |q \rangle^Q = q|q \rangle^Q$ and
${}^Q\langle q |_Q q' \rangle^Q = \delta(q-q')$,
which suggests
$\int_{-\infty}^\infty dq |q \rangle^Q {}^Q\langle q |_Q =1$.
Similarly, $|p \rangle^Q$ is introduced as the eigenstate of $\hat{p}_{Q}$ by $\hat{p}_{Q} |p \rangle^Q = p|p \rangle^Q$ and ${}^Q\langle p |_Q p' \rangle^Q = \delta(p-p')$.
Now, utilizing $|q \rangle^Q$,
we define the $Q$-Hermitian probability density operator
\begin{equation}
\hat{\rho} \equiv | q \rangle^Q {}^Q\langle q |_Q
\end{equation}
as an example of $Q$-Hermitian $\hat{\cal O}$,
and write a $q$-representation of the maximizing state
$| \tilde{A}(t) \rangle$ as
\begin{equation}
\psi_{\tilde{A}}(q) \equiv {}^Q\langle q |_Q \tilde{A}(t) \rangle.
\end{equation}
Then the probability density
$\rho \equiv \langle \hat{\rho} \rangle_Q^{BA}$
is given via the maximization principle by
$\rho =\langle \hat{\rho} \rangle_Q^{\tilde{A} \tilde{A}}
= | \psi_{\tilde{A}}(q) |^2$,
which obeys $\int_{-\infty}^{\infty} dq \rho =1$,
so we could construct
a conserved probability current density
\begin{equation}
j(q,t) = \frac{i\hbar}{2m}
\left( \frac{\partial \psi_{\tilde{A}}^{*} }{\partial q}
\psi_{\tilde{A}} - \psi_{\tilde{A}}^{*}
\frac{\partial \psi_{\tilde{A}} }{\partial q} \right),
\end{equation}
which satisfies the continuity equation
$\frac{\partial \rho}{\partial t } + \frac{\partial }{\partial q} j(q,t) = 0$.
In realistic cases, not only the maximizing state but also many other states contribute to the transition amplitude,
while the above relations are obtained
by considering only the maximizing state, which is
a kind of approximation in the sense that
we are ignoring the effects of the other states.
But we expect that their contribution becomes very small
in the large $T=T_B-T_A$ case, which we are interested in
from a phenomenological point of view.
The larger $T$ we consider, the more the states with the largest positive imaginary part of energy get to dominate.
Thus we have briefly given
a possible procedure to formulate
$\hat{H}_{Qh}$ in terms of $Q$-Hermitian $\hat{q}_Q$
and $\hat{p}_Q$, and also constructed
a conserved probability current density for the maximizing state.
However, it is not trivial at all
to determine the local expression of $\hat{H}_{Qh}$ in $q$-space, nor
to examine the classical behavior of
$\langle \hat{\cal O} \rangle_Q^{BA}$ explicitly.
We postpone these problems to future studies.
\section*{Acknowledgements}
K.N. would like to thank the members and visitors of NBI
for their kind hospitality.
H.B.N. is grateful to NBI for allowing him to work at the Institute as emeritus.
|
\section{Introduction}
Control and synchronization of large-scale dynamical systems
have attracted increasing interests over the last several decades \cite{Liu12}-\cite{Lu07}. When a
networked system is unstable by itself, many control strategies are
designed to stabilize the networked system. Among them, pinning control is effective, because it is economically realized by controlling a partial of the nodes, instead of all nodes
in the network.
The general idea behind pinning control is that when applying some local
feedback controllers only to a fraction of nodes, the rest
of nodes can be propagated through the network interactions among nodes \cite{Xiang}-\cite{Chen07}.
More related to the present paper, for example in \cite{han}, the authors investigated the pinning control problem of coupled dynamical systems with Markovian switching couplings
and Markovian switching controller-set. In \cite{han} and most existing works in linearly coupled dynamical systems, each node needs to gather information of its own state and neighbors states and update them continuously or in a fixed sampling rate \cite{Lu04}. However, as pointed out in \cite{Astrom},
the event-based sampling technique showed better performance than sampling periodically in time for some simple systems. Hence, a number of researchers suggested that the event-based control algorithms can be utilized for the purpose to reduce communication and computation load in networked systems \cite{Tabuada}-\cite{Wang11} but still maintain control performance \cite{Wang11}-\cite{Yi1}. Therefore, the event-based control is particularly suitable for networked systems with limited resources and so has attracted wide interests in the scope of distributed control of networked systems. The idea of event-triggered control can be regarded as a specific temporal discretization approach, which was studied before \cite{Lu04,Lu07}. As application, the event-based control was applied for consensus of multi-agent systems. For instance,
\cite{Dimarogonas} investigated centralized and
distributed formulation of event-driven strategies for consensus of multi-agent systems and proposed a self-triggered setup; \cite{Johannesson,Rabi} studied the stochastic event-driven strategies;
\cite{Seyboth} introduced event-based control strategies for both networks of single-integrators with time-delay in communication links and networks of double-integrators; By using scattering transformation,
\cite{YuHan} studied the output synchronization problem of multi-agent systems with event-driven communication in the presence of constant communication delays.
In some recent papers \cite{Alderisio}-\cite{ZLiu}, the authors addressed event-triggered algorithms for pinning control of networks. \cite{Alderisio} gave an exponentially decreasing threshold function, while the event-triggering threshold in \cite{Gao,ZLiu} is prescribed by the continuous or discrete states of agents and target. In these works, sufficient conditions were proposed, which are based on the control gain, some quantities of the uncoupled node dynamics and the minimum eigenvalue of the augmented Laplacian.
Motivated by these works including our previous work \cite{han}, in the present paper, we employ the event-triggered strategy in both coupling configuration and pinning control terms to realize stability in coupled dynamical systems with Markovian switching couplings and pinned node set.
At each node, the diffusion coupling and feedback pinning terms are piecewise static based on the information of its local neighborhood and the target trajectory only at the latest time of event, which is triggered by some specified criteria derived from the information of its local neighborhood and target. In other words,
once the triggering criterion of node is satisfied, the diffusion coupling and pinning terms will be updated; otherwise these terms are constant between two successive event time points.
We consider two scenarios: continuous monitoring and discrete monitoring.
In the continuous-monitoring scenario, each node observes its neighborhood's and the target's states in an instantaneous way; on the contrary, in discrete-monitoring scenario, each node can only obtain its neighborhood's and target's information at the last event-triggering time point, which results in a small cost of monitoring (communication load) but the triggering events happening more frequently than continuous monitoring, namely, higher computation load. For both scenarios, it is proved that the proposed event-triggered rules guarantee the stability of the coupled dynamical systems under the local pinning algorithm.
This paper is organized as follows. After formulating the underlying problem in Sec. 2, the event-triggering rules of diffusion and pinning terms are proposed to pin
the coupled systems to a homogenous preassigned trajectory of the uncoupled node system by using continuous and discrete monitoring scenarios in Sec. 3 and 4 respectively. Simulations are given in Sec. 5 to verify the theoretical results. Strength, limitations of the work and possible orients of future study are discussed in Sec. 6.
Finally, this paper is concluded in Sec. 7.
\section{Problem formation}
In this paper, we consider a network of linearly coupled dynamical systems with discontinuous diffusions and feedback pinning terms as follows:
\begin{eqnarray}
\dot{x}_i(t)=f(x_{i}(t))+\theta_i(\sigma_t,t_k^i),~~~t^{i}_{k}\le t<t^{i}_{k+1},~i=1,\cdots,m,
\label{pinning_ds1}
\end{eqnarray}
where
\begin{eqnarray}
\theta_i(\sigma_t,t_k^i)=-c\sum_{j=1}^{m}L_{ij}(\sigma_t)\Gamma\bigg[x_{j}(t^{i}_{k})-x_{i}(t^{i}_{k})\bigg]
-c\epsilon D_i(\sigma_t)\bigg[x_{i}(t^{i}_{k})-s(t^{i}_{k})\bigg].
\label{pinning_ds2}
\end{eqnarray}
Here, the system contains $m$ nodes, $x_{i}(t) = [x_i^1(t),\cdots,x_i^n(t)]^{\top} \in\mathbb R^{n}$ denotes the state vector of node $i$, the continuous map $f(\cdot):\mathbb R^{n}\to\mathbb R^{n}$ denotes the identical node dynamics. $\sigma_t$ is a homogeneous Markov chain, which will be specified later. $L(\sigma_t)=[L_{ij}(\sigma_t)]_{i,j=1}^{m}\in\mathbb R^{m,m}$ is the time-varying Laplacian matrix of the underlying time-varying bi-graph $\mathcal G(\sigma_t)=\{V,E(\sigma_t)\}$, with the node set $V$ and time-varying link set $E(\sigma_t)$: for each pair of nodes $i\ne j$, $L_{ij}(\sigma_t)=-1$ if $i$ is linked to $j$ at time $t$, otherwise $L_{ij}(\sigma_t)=0$, and $L_{ii}(\sigma_t)=-\sum_{j=1}^{m}L_{ij}(\sigma_t)$.
$c$ is uniform coupling strength at each node. $\Gamma=[\gamma_{ij}]_{i,j=1}^{n}\in\mathbb R^{n,n}$ stands for the inner configuration matrix that describes the coupling of components between the state vectors.
$D_i(\sigma_t)=\delta_{\mathcal D(\sigma_t)}(i)$, where ${\delta}_{\cdot}(\cdot)$ is the characteristic function, i.e. $\delta_{\mathcal D(\sigma_t)}(i) = 1$ if $i \in\mathcal D(\sigma_t) $, otherwise $\delta_{\mathcal D(\sigma_t)}(i) = 0$ for the pinned node subset $\mathcal D(\sigma_t)\subset\{1,\cdots,m\}$, where
$\mathcal D(\sigma_t)$ denotes the node subset in $\mathcal V$ that are pinned at time $t$ by a specific node dynamic trajectory $s(t)$ with $\dot{s}=f(s(t))$, $s(0)=s_{0}$. $\epsilon$ is the pinning strength gain over the coupling strength.
Our aim is to provide sufficient conditions to guarantee that $s(t)$ is a global stable trajectory for the coupled system, namely
\begin{eqnarray*}
\lim_{t\to\infty}\|x_{i}(t)-s(t)\|=0,~\forall~i=1,\cdots,m.
\end{eqnarray*}
Here, we consider the $L_{2}$-vector norm and denote it by $\|\cdot\|$ throughout this paper.
Let $\hat{x}_{i}(t)=x_{i}(t)-s(t)$. Then (\ref{pinning_ds1} and \ref{pinning_ds2}) become
\begin{eqnarray}\label{pinning_ds3}
\dot{\hat{x}}_{i}(t)=f(\hat{x}_{i}(t)+s(t))-f(s(t))-c\sum_{j=1}^{m}L_{ij}(\sigma_t)
\Gamma\bigg[\hat{x}_{j}(t^{i}_{k})-\hat{x}_{i}(t^{i}_{k})\bigg]
-c\epsilon D_i(\sigma_t)\Gamma\hat{x}_{i}(t^{i}_{k}),~t^{i}_{k}\le t<t^{i}_{k+1}
\end{eqnarray}
Suppose $\sigma_{t}$ is a homogeneous continuous Markov chain with a finite
state space $\mathbb{S}=\{1,2,\cdots,N\}$ and its infinitesimal generator
$Q=[q_{uv}]_{N\times N}$ is given by
\begin{eqnarray*}
\mathbb{P}\{\sigma_{t+\Delta}=v|\sigma_{t}=u\}
=\left\{\begin{array}{lr}q_{uv}\Delta +o(\Delta), & u\ne v,\\
1+q_{uu}\Delta+o(\Delta),& u=v,
\end{array}
\right.
\end{eqnarray*}
where $\Delta>0$, $o(\Delta)$ is a infinitesimal as $\Delta\to 0$, i.e, $\lim_{\Delta \to 0}(o(\Delta)/\Delta)=0$,
$p_{uv}=-\frac{q_{uv}}{q_{uu}}>0$ is the transition probability from $u$ to
$v$ if $v\ne u$, while $q_{uu}=-\sum_{v=1,v\ne u}^{N}q_{uv}$. Denote $P=[p_{uv}]$ the transition matrix of the Markov chain.
The sojourn time in state $u$ is exponentially distributed with parameter
$q_{u}\triangleq -q_{uu}$.
Let $\mathbb P_{t,u}(s,\Lambda)=\mathbb P(\sigma_{t+s}\in\Lambda|\sigma_t=u)$ and
$\mathbb E_{t,u} (f(\sigma_s))=\int f(y)\mathbb P_{t,u}(s,dy)$.
Denote by $\mathcal A$ the weak infinitesimal operator of $\sigma_t$.
A function $f(\sigma_t,t)$ is said to be
in the domain of $\mathcal A$ if
\begin{align*}
&\lim_{\Delta\to 0}\frac{\mathbb E_{t,u} (f(\sigma_{t+\Delta},t+\Delta))-f(u,t)}{\Delta}\\
=&\lim_{\Delta\to 0}\frac{\mathbb E_{t,u} (f(\sigma_{t+\Delta},t+\Delta))-\mathbb E_{t,u} (f(\sigma_{t+\Delta},t))}{\Delta}
+\lim_{\Delta\to 0}\frac{\mathbb E_{t,u} (f(\sigma_{t+\Delta},t))-f(u,t)}{\Delta}\\
=&f_{t}(u,t)+h(u,t)
\end{align*}
and
\begin{align*}
\lim_{\Delta\to 0} \mathbb E_{t,u}(f_t(\sigma_{t+\Delta},t+\Delta)+h(\sigma_{t+\Delta},t+\Delta))
=f_t(u,t)+h(u,t).
\end{align*}
Then, we write $\mathcal A f(u,t)=f_t(u,t)+h(u,t)$.
Note that for fixed $t$,
\begin{align*}
\mathbb E_{t,u} (f(\sigma_{t+\Delta},t))=\int f(y,t)\mathbb P_{t,u}(\Delta,dy)
=\sum_{v}f(v,t)\mathbb P_{t,u}(\Delta,v)
=\sum_{v}f(v,t)q_{uv}\Delta+f(u,t).
\end{align*}
Hence, by the Dynkin's formula, we have
\begin{align}\label{infinitesimal}
\mathcal A f(u,t)=f_t(u,t)+\sum_{v}f(v,t)q_{uv}.
\end{align}
Throughout the paper, we assume $f(\cdot,t)$ belongs to the following function
class.
\begin{definition}
Function class QUAD $(G,\alpha\Gamma,\beta)$: let $G$ be an $n\times n$ positive definite matrix and $\Gamma$ be an $n\times n$ matrix. QUAD$ (G,\alpha\Gamma,\beta)$ denotes a class of
continuous functions $f(\xi,t):\mathbb{R}^{n}\times[0,+\infty)\mapsto
\mathbb{R}^{n}$ satisfying
\begin{eqnarray*}\label{f}
(\xi-\zeta)^{\top}G[f(\xi,t)-f(\zeta,t)-\alpha \Gamma (\xi-\zeta)]
\le -\beta(\xi-\zeta)^{\top}G(\xi-\zeta)
\end{eqnarray*}
holds for all $\xi, \zeta \in \mathbb{R}^n$.
\end{definition}
\begin{definition}
System (\ref{pinning_ds1},\ref{pinning_ds2}) is said to be exponentially stable at $s(t)$ in mean
square sense, if there exists constants $\delta>0$ and $M>0$, such that
\begin{eqnarray}
\mathbb{E}\bigg[\|x_{i}(t)-s(t)\|^2\bigg]\leq M e^{-\delta t}
\end{eqnarray}
holds for all $t>0$ and any $i=1,\cdots,m$.
\end{definition}
\section{Continuous monitoring}
We briefly provide the basic idea of the setup of the coupling and pinning
terms. Instead of using the simultaneous state
from the neighborhood and the target trajectory to realize stability, an economic
alternative for the node $i$ is to use the neighbors' constant
states at the nearest time point $t_k^i$ until some pre-defined
event is triggered at time $t_{k+1}^i$; if node $i$ is pinned at time $t$, it also obtains the target trajectory's constant state at time point $t_{k_i(t)}^i$; then the incoming neighbors' and the target trajectory's information is updated by the states at $t_{k+1}^i$ until the next event is triggered, and so on. The event is defined based on
the neighbors', the target trajectory's and its own states with some prescribed rule.
This process goes on through all nodes in a parallel fashion.
To depict the event that triggers the next
time point, we introduce following Lyapunov function:
\begin{eqnarray}\label{V}
V(\hat{x},t,\sigma_t)&=&\frac{1}{2}\hat{x}^{\top}\bigg(P(\sigma_t)\otimes G\bigg)\hat{x},
\end{eqnarray}
where $\hat{x}=[\hat{x}_{1}^{\top},\cdots,\hat{x}_{m}^{\top}]^{\top}$, $P(\sigma_t)\in\mathbb R^{m,m}$ are diagonal positive definite matrices, induced by $\sigma_{t}$, and
$G\in\mathbb R^{n,n}$ is a positive definite matrix.
Let $\hat{F}(\hat{x})=[(f(x_{i})-f(s))^{\top},\cdots,(f(x_{m})-f(s))^{\top}]^{\top}$,
$D(\sigma_t)=diag[D_i(\sigma_t)]_{i=1}^{m}$,
$\hat{L}(\sigma_t)=L(\sigma_t)+\epsilon D(\sigma_t)$.
Note that $V(\hat{x},t,\sigma_t)$ is in the domain of the weak infinitesimal
operator of $\sigma_t$. Denoting $\sigma_t=u$, by (\ref{infinitesimal}), we have
\begin{eqnarray}\label{derivative_of_V}
\mathcal A V(\hat{x},t,u) = \sum_{v=1}^{N}q_{uv}V(\hat{x},t,v)+(\frac{\partial
V(\hat{x},t,u)}{\partial \hat{x}})^{\top}\frac{d{\hat{x}}}{dt}.
\end{eqnarray}
Substitute (\ref{pinning_ds3}) into (\ref{derivative_of_V}), we get
\begin{eqnarray}\label{derivative_of_V1}
\mathcal A V(\hat{x},t,\sigma_t) \left|_{(\ref{pinning_ds3})}\right.&=& \hat{x}^{\top}(t)\left[P(\sigma_t)\otimes G\right]\left[\hat{F}(\hat{x}(t))-\alpha I_m\otimes\Gamma \hat{x}(t)+ cz(t)\right]\\
\nonumber&+&\hat{x}^{\top}(t)\left\{P(\sigma_t)\left[(\alpha I_{m}-c \hat{L}(\sigma_t))\otimes G\Gamma\right]+\frac{1}{2}\sum_{v}q_{\sigma_t v} P(v)\otimes G\right\}^{sym}\hat{x}(t),
\end{eqnarray}
where $z(t)=[z_1^{\top}(t),\cdots,z^{\top}_m(t)]^{\top}$ with
\begin{eqnarray}\label{z_i}
z_i(t)&=&\sum_{j}L_{ij}(\sigma_t)\Gamma
\left[x_j(t)-x_i(t)-x_j(t_{k_i(t)}^i)+x_i(t_{k_i(t)}^i)\right]\nonumber\\
&+&\epsilon D_i(\sigma_t)\left[x_i(t)-s(t)-x_i(t_{k_i(t)}^i)-s(t_{k_i(t)}^i)\right],
\end{eqnarray}
and $\{S\}^{sym}$ denotes the symmetry part of a square matrix $S$, i.e., $S^{sym}=(S+S^{\top})/2$.
Let $\lambda_m(\cdot)$ and $\lambda_M(\cdot)$ denote the smallest and largest eigenvalues in module of a symmetry real matrix, and $\underline{\lambda}=\min_v\lambda_m\left(P(v)\otimes G\right)$, $\bar{\lambda}=\max_v\lambda_M\left(P(v)\otimes G\right)$. From the condition $f\in QUAD(P,\alpha\Gamma,\beta)$, we have
\begin{eqnarray}\label{V_term1}
\hat{x}^{\top}(t)\left[P(\sigma_t)\otimes G\right]\left[\hat{F}(\hat{x}(t))-\alpha I_m\otimes\Gamma \hat{x}(t)\right]\le
-\beta\underline{\lambda}\hat{x}^{\top}(t)\hat{x}(t).
\end{eqnarray}
For the term $\hat{x}^{\top}(t)\left[P(\sigma_t)\otimes G\right]z(t)$, we have
\begin{eqnarray}\label{V_term2}
\hat{x}^{\top}(t)\left[P(\sigma_t)\otimes G\right]z(t)&\le &
\frac{\upsilon}{2}\hat{x}^{\top}(t)(P^2(\sigma_t)\otimes G^2)\hat{x}(t)+\frac{1}{2\upsilon} z^{\top}(t)z(t)\\
\nonumber&\leq& \frac{\upsilon\bar{\lambda}^2}{2}
\hat{x}^{\top}(t)\hat{x}(t)+\frac{1}{2\upsilon}z^{\top}(t)z(t)
\end{eqnarray}
holds for any $\upsilon>0$.
Then, we have the following theorem.
\begin{theorem}\label{thm1}
Suppose that $f$ belongs to $QUAD(G,\alpha\Gamma,\beta)$ with the positive matrix $G$ and $\alpha>0, \beta>0$, and there exist diagonal
positive definite matrices $P(u), u=1,\cdots,N$ such that
\begin{eqnarray}\label{slow_condition}
\left\{P(u)[\alpha I_{m}- cL(u)-c\epsilon D(u)]\otimes G\Gamma\right\}^{sym}
+\frac{1}{2}\sum_{v=1}^{N}q_{uv}P(v)\otimes G\leq 0,~\rm{for~ all}~ u\in \mathbb{S}.
\end{eqnarray}
Then, under either of the following two updating rules, system (\ref{pinning_ds1}) is exponentially stable at the homogeneous trajectory $s(t)$
in mean square sense:
\begin{enumerate}
\item[(1)]
set $t^{i}_{k+1}$ by the rule
\begin{eqnarray}
t_{k+1}^{i}=\max\left\{\tau\ge t_{k}^{i}:~\|z_i(\tau)\|\le\frac{(\beta\underline{\lambda}-\frac{1}{2}\delta\bar{\lambda})}
{\sqrt{c}\bar{\lambda}}\|\hat{x}_i(\tau)\|\right\}
\label{event1}
\end{eqnarray}
where $0<\delta\le{2\beta\underline{\lambda}}/{\bar{\lambda}}$ is a constant;
\item[(2)]
set $t^{i}_{k+1}$ by the rule
\begin{eqnarray}
t_{k+1}^{i}=\max\left\{\tau\ge t_{k}^{i}:~\|z_i(\tau)\|\le a\exp{(-b\tau)}\right\}
\label{event2}
\end{eqnarray}
where $a>0$ and $b>0$ are constants.
\end{enumerate}
\end{theorem}
\begin{proof}
[Case (1).] Consider the event-triggering rule (\ref{event1}) and
pick a constant $\delta$ with $0<\delta\le{2\beta\underline{\lambda}}/{\bar{\lambda}}$. By Dynkin Formula \cite{Mao06}, we have
\begin{eqnarray}
\mathbb{E}e^{\delta t}V(\hat{x},t,\sigma_{t})
=\mathbb{E}V(\hat{x}(0),0,\sigma_0)+\delta\mathbb{E}\int_{0}^{t}e^{\delta \tau}V(\hat{x},\tau,\sigma_{\tau})d\tau
+\mathbb{E}\int_{0}^{t}e^{\delta \tau}\mathcal{A}V(\hat{x},\tau,\sigma_{\tau})d\tau.
\end{eqnarray}
From (\ref{derivative_of_V1})-(\ref{slow_condition}), we have
\begin{eqnarray}\label{dv}
\nonumber\mathbb{E}e^{\delta t}V(\hat{x},t,\sigma_{t})
\nonumber&\le& \mathbb{E}V(\hat{x}(0),0,\sigma_0)+\delta\mathbb{E}\int_{0}^{t}e^{\delta \tau}V(\hat{x},\tau,\sigma_{\tau})d\tau
-\beta\underline{\lambda}\mathbb{E}\int_{0}^{t}e^{\delta \tau}\hat{x}^{\top}(\tau)\hat{x}(\tau)d\tau\\
\nonumber&&+ \frac{c\upsilon\bar{\lambda}^2}{2}\mathbb{E}\int_{0}^{t}e^{\delta \tau}\hat{x}^{\top}(\tau)\hat{x}(\tau)d\tau+\frac{c}{2\upsilon}\mathbb{E}\int_{0}^{t}e^{\delta \tau}z^{\top}(\tau)z(\tau)d\tau\\
\nonumber&&+\mathbb{E}\int_{0}^{t}e^{\delta \tau}\hat{x}^{\top}(\tau)\left\{P(\sigma_{\tau})\left[(\alpha I_{m}-c \hat{L}(\sigma_{\tau}))\right]\otimes G\Gamma+\frac{1}{2}\sum_{v}q_{\sigma_{\tau} v} P(v)\otimes G\right\}^{sym}\hat{x}(\tau)d\tau\\
&\le& \mathbb{E}V(\hat{x}(0),0,\sigma_0)+\mathbb{E}\int_{0}^{t}e^{\delta \tau}\left\{\left[
-\beta\underline{\lambda}+\frac{\delta\bar{\lambda}}{2}+\frac{c\upsilon\bar{\lambda}^2}{2}\right]
\hat{x}^{\top}(\tau)\hat{x}(\tau)+\frac{c}{2\upsilon}z^{\top}(\tau)z(\tau)\right\}d\tau
\end{eqnarray}
for any $\upsilon>0$.
Note that
\begin{eqnarray*}
\max_{\upsilon>0}\frac{2\upsilon}{c}\left[\beta\underline{\lambda}-\frac{
\delta\bar{\lambda}}{2}-\frac{c\upsilon\bar{\lambda}^2}{2}\right]=
\frac{(\beta\underline{\lambda}-\frac{1}{2}\delta\bar{\lambda})^2}
{c\bar{\lambda}^2}
\end{eqnarray*}
and the maximum is reached if and only if $\upsilon=\frac{(\beta\underline{\lambda}-\frac{1}{2}\delta\bar{\lambda})}
{\sqrt{c}\bar{\lambda}}$. Hence, letting $\upsilon=\frac{(\beta\underline{\lambda}-\frac{1}{2}\delta\bar{\lambda})}
{\sqrt{c}\bar{\lambda}}$, (\ref{event1}) implies
\begin{eqnarray*}
\|z_i(\tau)\|^2\leq \frac{2\upsilon}{c}\left[\beta\underline{\lambda}-\frac{
\delta\bar{\lambda}}{2}-\frac{c\upsilon\bar{\lambda}^2}{2}\right]\|\hat{x}_i(\tau)\|^2,~~~
i=1,\cdots,m,
\end{eqnarray*}
for all $\tau\le t^{i}_{k+1}$. Therefore, we have
\begin{eqnarray}\label{exp_step1}
\mathbb{E}e^{\delta t}V(\hat{x},t,\sigma_{t})\le \mathbb{E}V(\hat{x}(0),0,\sigma_{0}),
\end{eqnarray}
which implies
\begin{eqnarray}\label{exp_step2}
\mathbb{E}e^{\delta t}\|x_{j}(t)-s(t)\|^{2}
\le \frac{2}{\underline{\lambda}}\mathbb{E}e^{\delta t}V(\hat{x},t,\sigma_{t})
\le \frac{2}{\underline{\lambda}}\mathbb{E}V(\hat{x}(0),0,\sigma_{0}),~~~\forall j=1,\cdots,m.
\end{eqnarray}
Therefore, we have
\begin{eqnarray}\label{exp_step3}
\mathbb{E}\|x_{j}(t)-s(t)\|^{2}\le \frac{2e^{-\delta t}}{\underline{\lambda}}\mathbb{E}V(\hat{x}(0),0,\sigma_{0}),~~~~\forall j=1,\cdots,m.
\end{eqnarray}
[Case (2).] Consider the event-triggering rule (\ref{event2}) and pick $\upsilon=\frac{2\beta\underline{\lambda}-\delta\bar{\lambda}}{c\bar{\lambda}^2}$. Then, we have
\begin{eqnarray*}
-\beta\underline{\lambda}+\frac{
\delta\bar{\lambda}}{2}+\frac{c\upsilon\bar{\lambda}^2}{2}=0.
\end{eqnarray*}
Substituting $\|z_{i}(\tau)\|\le a\exp(-b\tau)$ into (\ref{dv}) gives
\begin{eqnarray*}
\mathbb{E}e^{\delta t}V(\hat{x},t,\sigma_{t})&\le & \mathbb{E}V(\hat{x}(0),0,\sigma_{0})+\frac{a^2c^2\bar{\lambda}^2}
{2(2\beta\underline{\lambda}-\delta\bar{\lambda})}\mathbb{E}\int_{0}^{t}e^{(-2b+\delta) \tau}d\tau\\
&=&\mathbb{E}V(\hat{x}(0),0,\sigma_{0})+\frac{a^2c^2\bar{\lambda}^2}
{2(2\beta\underline{\lambda}-\delta\bar{\lambda})}\frac{1}{2b-\delta}
\left[1-e^{(-2b+\delta)t}\right].
\end{eqnarray*}
Let $C_0=\frac{a^2c^2\bar{\lambda}^2}
{2(2\beta\underline{\lambda}-\delta\bar{\lambda})}\frac{1}{2b-\delta}$. By the similar arguments as (\ref{exp_step1})-(\ref{exp_step3}), we have
\begin{eqnarray*}
\mathbb{E}\|x_{j}(t)-s(t)\|^{2}\le \frac{2e^{-\delta t}}{\underline{\lambda}}\left(\mathbb{E}V(\hat{x}(0),0,\sigma_{0})+
C_0\left[1-e^{(-2b+\delta)t}\right]\right)
\le C_1e^{-\min(2b,\delta)t},~~~\forall j=1,\cdots,m,
\end{eqnarray*}
where $C_1=\frac{2}{\underline{\lambda}}\max(|\mathbb{E}V(\hat{x}(0),0,\sigma_{0})+C_0|,|C_0|)$. This completes the proof.
\end{proof}
\begin{remark}
If $\hat{x}(t_1)=0$ holds, then from (\ref{event1}), at time $t_1$, every node updated its feedback term and pinning term (if pinned). Therefore, from (\ref{pinning_ds1},\ref{pinning_ds2}), for any $t\ge t_1$, $\hat{x}(t)=0$ holds, which means $\hat{x}=0$ is an equilibrium of the system under the event-triggering rules.
\end{remark}
\begin{remark}
In fact, in our previous work \cite{han}, we studied pinning dynamic systems of networks with Markovian switching couplings and controller-node set.
By this theorem, we showed and proved that if the condition for the stability of the coupled system with spontaneous coupling and control in \cite{han} can be satisfied, then the event-trigger strategies can stabilize the system too. Therefore, the issue of selection of pinned node in terms of guaranteeing stability is totally the same with the system with spontaneous diffusion and control.
\end{remark}
Under the updating rule (\ref{event2}), it can be proved that the Zeno behaviors \cite{Joh} is excluded by the arguments as the same fashion as in \cite{Alderisio}. While for the updating rule (\ref{event1}), similar to work \cite{Dimarogonas}, it should be pointed out that there exists at least one node such that its next inter-event interval is strictly positive.
\begin{proposition}\label{Zeno1}
Suppose that all hypotheses of Theorem \ref{thm1} hold. Under the updating rule (\ref{event1}), if the system does not reach stability, then there exists at least one node $i\in\{1,\cdots,m\}$ such that the next inter-event interval is strictly positive; Under the updating rule (\ref{event2}), if the system does not reach stability, the expectation of next inter-event interval of every node is strictly positive, further, it is lower bounded by some positive constant.
\end{proposition}
The proof is the similar to those in \cite{Alderisio,Dimarogonas} with some modifications. In fact, if at any time $t$, there exists one node $i$ such that (\ref{event1}) cannot hold as an equality for this node $i$. Hence, $t_{k+1}^{i}>t$ can be derived, which implies that the inter-event interval for node $i$ is positive. Otherwise, at time $t$, (\ref{event1}) holds as an equality for every node, that is, all nodes update their control law at this moment, which implies $z_i(t)=0$ holds for all $i$. However, since the network has not been stabilized at $s(t)$ yet, there exists at least node $j$ with $\hat{x}_{j}(t)\neq 0$, which implies $t^{j}_{k+1}>t$ holds, which implies that the next inter-event interval for node $j$ should be positive.
In comparison, under the rule (\ref{event2}), suppose $b<\frac{\delta}{2}$, for each node $i$, at $\tau=t^{i}_{k}$, we have $z_{i}(t^i_k)=0$ but the right-hand side of (\ref{event2}) is nonzero. Therefore, $t^{i}_{k+1}>t^{i}_{k}$ always holds and the low bound of the expectation of inter-event intervals can be estimated as
$\frac{1}{b}\log{\{1+\frac{1}{A+B}\}}$ with $L_f$ the Lipschtiz constant of $f(\cdot)$, $A = \frac{2mL_f+2c m(m+\epsilon)+L_f+cm}{ab}$ and $B =\frac{2m+1}{b}$.
\section{Discrete monitoring}
In the discrete monitoring scenario, each node $i$ can obtain its local neighborhood's state only at the time points $t^{i}_{k}$, $k=1,2,\cdots$. Meanwhile, if node $i$ is pinned, it can also obtain the target's state at latest time points $t_{k_i(t)}^i$. By this way, the rule to determine the next time point $t^{i}_{k+1}$ of obtaining state information only depends on the local states at $t^{i}_{k}$. In comparison, the triggering event rules (\ref{event1}) and (\ref{event2}) demand the instantaneous states after $t^{i}_{k}$.
Consider system (\ref{pinning_ds1},\ref{pinning_ds2}) and $V(x)$ as the candidate Lyapunov function with its derivative (\ref{derivative_of_V1}). We are to derive a triggering event rule from (\ref{event1}) in Theorem \ref{thm1}, which only depends on $t_{k}^{i}$. The estimations of the upper bounds of $\|z_i(t)\|$ and the lower bounds of $\|x_i(t)-s(t)\|$ for all $i$ are essential.
First of all, we take the switching time points of the Markov chain $\sigma_{t}$ to trigger the state information updating for all nodes. Then, we are to estimate the upper-bound of $\|(x_{j}(t)-x_{j}(t_{k}^{i}))-(x_{i}(t)-x_{i}(t_{k}^{i}))\|$ with $L_{ij}(\sigma_t)\neq 0$, of which the evolution equation can be written as:
\begin{eqnarray}\label{evolution1}
\begin{cases}\frac{d{x}_{i}(t)}{dt}=f(x_{i}(t))+
\theta_{i}(\sigma_t,t_{k}^i)&\\
\frac{d{x}_{j}(t)}{dt}=f(x_{j}(t))+\theta_{j}(\sigma_t,t_{k_j(t)}^j)
&
\end{cases}
\end{eqnarray}
for $t^{i}_{k}\le t<\min\{t^{i}_{k+1},t^{j}_{k_{j}(t)+1}\} $ and initials $x_{i}(t^{i}_{k}), x_{j}(t^{i}_{k})$. Here, $\theta_{i}(\sigma_t,t_{k}^i)$ and $\theta_{j}(\sigma_t,t_{k_j(t)}^j)$ are constants in this time interval
Let us consider a general form of (\ref{evolution1}) as follows:
\begin{eqnarray}
\begin{cases}\frac{du}{dt}=f(u(t))+\theta&u(0)=u_{0}\\
\frac{dv}{dt}=f(v(t))+\vartheta&v(0)=v_{0},
\end{cases}\label{r2}
\end{eqnarray}
where $u,v,u_{0},v_{0}\in\mathbb R^{n}$.
Suppose that there exists a nonnegative-valued continuous map $\rho:\mathbb R_{\ge 0}\times\mathbb R^{4n}\to\mathbb R_{\ge 0}$ such that the solutions of (\ref{r2}) satisfy the following inequality:
\begin{eqnarray}
\|(u(t)-u_{0})-(v(t)-v_{0})\|\le\rho(t,\theta,\vartheta,u_{0},v_{0}).\label{r1}
\end{eqnarray}
Here the map $\rho$ depends on the node dynamics map $f(\cdot)$, the initial value $u_{0},v_{0}$ and inputs $\theta,\vartheta$, and satisfies $\rho(0,\cdot,\cdot,\cdot,\cdot)\equiv 0$. Geometrically, $\rho$ is an upper-bound estimation of the difference between the two trajectories of (\ref{r2}) starting at $0$ of time-length $t$:
\begin{align*}
\|(u(t)-u_{0})-(v(t)-v_{0})\|
=\left\|\int_{0}^{t}[f(u(s))-f(v(s))]d s+(\theta-\vartheta) t\right\|.
\end{align*}
For example, if $f(\cdot)$ is Lipschitz (on the two trajectories): $\|f(u(s))-f(v(s))\|\le L_{f}\|u(s)-v(s)\|$ for all $s\ge 0$, then we have
\begin{align*}
\|(u(t)-u_{0})-(v(t)-v_{0})\|
\le &L_{f}\int_{0}^{t}\|(u(s)-u_{0})-(v(s)-v_{0})\|ds\\
&+(\|\theta-\vartheta\|+L_f\|u_{0}-v_{0}\|)~t.
\end{align*}
By the Gronwall-Bellman inequality \cite{Gron,Bell}, we have
\begin{align}
\label{Lips}
\|(u(t)-u_{0})-(v(t)-v_{0})\|\
\le\frac{(\|\theta-\vartheta\|+L_f\|u_{0}-v_{0}\|)}{L_{f}}[\exp(L_{f}t)-1].
\end{align}
We can take $\rho(t,\theta,\vartheta,u_{0},v_{0})$ as the right-hand side above, which equals to zero at $t=0$.
Second, we suppose that there exists a nonnegative map $\varrho:\mathbb R_{\ge 0}\times\mathbb R^{4n}\to\mathbb R_{\ge 0}$ such that the solutions of (\ref{r2}) satisfy:
\begin{eqnarray}
\|u(t)-v(t)\|\geq\varrho(t,\theta,\vartheta,u_{0},v_{0}).
\end{eqnarray}
Here, $\varrho$ can be regarded as the lower-bound estimation of the distance between two trajectories:
\begin{align*}
\|u(t)-v(t)\|
=\left\|\int_{0}^{t}[f(u(s))-f(v(s))]ds
+(\theta-\vartheta)t
+(u_0-v_0)\right\|
\end{align*}
and satisfies (i). $\varrho(\cdot,\theta,\theta,u_{0},u_{0})\equiv 0$; (ii). $\varrho(0,\cdot,\cdot,u_{0},u_{0})\equiv 0$.
For example, assuming that there exists some constant $\sigma$ (possibly negative) such that
\begin{eqnarray*}
(u-v)^{\top}(f(u)-f(v))\ge\sigma (u-v)^{\top}(u-v)
\end{eqnarray*}
holds for all $u,v\in\mathbb R^{n}$.
We have
\begin{align*}
&\frac{d}{dt}[(u(t)-v(t))^{\top}(u(t)-v(t))]\left|_{(\ref{r2})}\right.
=2(u-v)^{\top}[f(u)-f(v)+\theta-\vartheta]\\
&\ge{2\sigma}(u-v)^{\top}(u-v)-\mu (u-v)^{\top}(u-v)
-\frac{1}{\mu}(\theta-\vartheta)^{\top}(\theta-\vartheta)
\end{align*}
hold for any $\mu>0$.
By Gronwall-Bellman inequality, we have
\begin{align*}
(u(t)-v(t))^{\top}(u(t)-v(t))
\ge&\exp{[({2\sigma}-\mu)t]}(u_{0}-v_{0})^{\top}(u_{0}-v_{0})\\
&-\frac{(\theta-\vartheta)^{\top}
(\theta-\vartheta)/\mu}{2\sigma-\mu}\bigg\{\exp[({2\sigma}-\mu)t]-1\bigg\}.
\end{align*}
We take $\varrho$ as the right-hand side above, which is positive for a small interval of $t$, starting from $0$, for any $u_{0}\ne v_{0}$.
\iffalse
Suppose the solution of (\ref{r2}) satisfy the following inequality:
\begin{eqnarray}
\|(u(t)-u_{0})-(v(t)-v_{0})\|\le\rho(t,u_{0},v_{0}),\label{r1}
\end{eqnarray}
where $\rho$ is nonnegative-valued continuous map that depends on the node dynamics map $f(\cdot)$, the initial value $u_{0},v_{0}$, and inputs $\theta,\vartheta$,
and satisfies $\rho(0,\cdot,\cdot)\equiv 0$. Geometrically, $\rho$ is an upper-bound estimation of the difference between the of the two trajectories of (\ref{r2}) starting at $0$ of time-length $t$:
\begin{align*}
\|(u(t)-u_{0})-(v(t)-v_{0})\|
=\left\|\int_{0}^{t}[f(u(s))-f(v(s))]d s+(\theta-\vartheta) t\right\|.
\end{align*}
\fi
We highlight that there is no uniform approach to get precise estimation for a general function of $f(\cdot)$ but one can do it case by case. Therefore, an efficient way is to use integrators that simulates the node dynamics of $\dot{u}=f(u)+\theta$ to realize generators that calculate the maps of $\rho$ and $\varrho$. Noting that these generators are independent of the states of the nodes, they can be built parallel to the networked system. Figures \ref{rho} and \ref{varrho} show the generators of $\rho$ and $\varrho$ respectively.
\begin{figure}[!t]
\begin{center}
\includegraphics[height=.4\textwidth,width=.45\textwidth]{rho.eps}
\caption{$\rho$ generator.} \label{rho}
\end{center}
\end{figure}
\begin{figure}[!t]
\begin{center}
\includegraphics[height=.4\textwidth,width=.45\textwidth]{varrho.eps}
\caption{$\varrho$ generator.} \label{varrho}
\end{center}
\end{figure}
Let $\vartheta^{i}_{k} = \theta_i(\sigma_t,t_k^i)$ and $\vartheta_{k_{j}(t)}^{j} = \theta_j(\sigma_t,t_{k_j(t)}(t))$.
Based on the event-triggering rules (\ref{event1}), (\ref{event2}), we have the following theorem.
\begin{theorem}\label{thm2}
Suppose that $f$ belongs to $QUAD(G,\alpha\Gamma,\beta)$ with positive matrix $G$ and $\alpha, \beta>0$. Suppose there exist diagonal
positive definite matrices $P(u), u=1,\cdots,N$ such that
\begin{eqnarray*}
\left\{P(u)[\alpha I_{m}- cL(u)-c\epsilon D(u)]\otimes G\Gamma\right\}^{sym}
+\frac{1}{2}\sum_{v=1}^{N}q_{uv}P(v)\otimes G\leq 0,~\rm{for~ all}~ u\in \mathbb{S}.
\end{eqnarray*}
define a sequence of $\xi^{i}_{k}$ under either of the following two updating rules,
\begin{enumerate}
\item[(1)]
\begin{eqnarray}
\xi^{i}_{k}=\max\bigg\{&\xi:&~\sum_{j\ne i}(-L_{ij}(\sigma_{\xi+t_k^i}))\rho\left(\xi,{\vartheta}^{i}_k,
{\vartheta}^{j}_{k_j(\xi+t_k^i)},x_{i}(t^{i}_{k}),x_{j}(t^{i}_{k})\right)+\nonumber\\
&&+\epsilon D_i(\sigma_{\xi+t_k^i})\rho\left(\xi,{\vartheta}^{i}_k,
0,x_{i}(t^{i}_{k}),s(t^{i}_{k})\right)\nonumber\\
&&\le\frac{(\beta\underline{\lambda}-\frac{1}{2}\delta\bar{\lambda})}
{\sqrt{c}\bar{\lambda}}\varrho\left(\xi,\vartheta^{i}_k,
0,x_{i}(t^{i}_{k}),s(t^{i}_{k})\right)\bigg\}\label{event3}
\end{eqnarray}
where $0<\delta\le{2\beta\underline{\lambda}}/{\bar{\lambda}}$ is a constant;
\item[(2)]
\begin{eqnarray}
\xi^{i}_{k}=\max\bigg\{&\xi:&~\sum_{j\ne i}(-L_{ij}(\sigma_{\xi+t_k^i}))\rho\left(\xi,{\vartheta}^{i}_k,
{\vartheta}^{j}_{k_j(\xi+t_k^i)},x_{i}(t^{i}_{k}),x_{j}(t^{i}_{k})\right)+\nonumber\\
&&+\epsilon D_i(\sigma_{\xi+t_k^i})\rho\left(\xi,{\vartheta}^{i}_k,
0,x_{i}(t^{i}_{k}),s(t^{i}_{k})\right)\nonumber\\
&&\le a\exp{(-b(\xi+t_k^i))}\bigg\}\label{event4}
\end{eqnarray}
where $a>0$ and $b>0$ are constants.
\end{enumerate}
If the triggering event time points $\{t^{i}_{k}\}$ are picked by the following scheme:
\begin{enumerate}
\item Initialization: $t_{0}^{i}=0$ for all $i=1,\cdots,m$;
\item At $t=t^{i}_{k}$, node $i$ obtains $\xi^{i}_{k}$ by the rule (\ref{event3})(or (\ref{event4}));
\item
At $t>t^{i}_{k}$, if one of its neighbor, for example, denoted by $j$, is triggered at $t=t^{j}_{k'+1}$ (let $k'$ be the latest event at node $j$ before $t$), then $j$ broadcasts its current updating law, $\vartheta^{j}_{k'}$, to node $i$, and the rule (\ref{event3})(or (\ref{event4})) is updated by replacing the diffusion term from node $j$, $\vartheta^{j}_{k'}$, by $\vartheta^{j}_{k'+1}$, and $t^{i}_{k}$ by $t$. And, go to Step 2;
\item If $\sigma_{t}$ switches at $t$, then we update the rule (\ref{event3})(or (\ref{event4})) by replacing $\vartheta^{i}_{k}$ by the current state $\theta_i(\sigma_t,t)$, and $t^{i}_{k}$ by $t$. And go to Step 2;
\item Let $t^{i}_{k+1}=t^{i}_{k}+\xi^{i}_{k}$, an event is triggered at node $i$ by updating the state information in (\ref{pinning_ds1},\ref{pinning_ds2}) from $t^{i}_{k}$ by $t^{i}_{k+1}$,
\end{enumerate}
then system (\ref{pinning_ds1},\ref{pinning_ds2}) is stabilized at $s(t)$ in mean square sense.
\end{theorem}
This theorem can be derived from Theorem \ref{thm1} immediately. In fact, event (\ref{event3}) is an estimation of event (\ref{event1}), event (\ref{event4}) is an estimation of event (\ref{event2}).
There is substantial difference between the discrete and continuous monitoring strategies. Generally speaking, the continuous monitoring require that every node collects its neighborhood states at every instant time, while discrete monitoring does not need this step. As shown in Table \ref{table1}, the continuous monitoring scheme costs higher communication load than the discrete monitoring. As a pay-off, we will show in the numerical example section that the frequencies of triggering events in the continuous monitoring are much lower than that the discrete monitoring requires. That is, the continuous monitoring costs lower computation load than the discrete monitoring.
\begin{table}[h]
\centering
\caption{Continuous \emph{vs} discrete time monitoring schemes\label{table1}}
\begin{tabular}{|r|p{.45\textwidth}|p{.45\textwidth}|}
\hline
Step&Continuous monitoring & Discrete-time monitoring\\
\hline
\multirow{2}{*}{1 }
& At time $t_k^i$, agent $i$ updates feedback control law $\theta_i(\sigma_t,t_k^i)$
& At time $t_k^i$, agent $i$ updates feedback control law $\theta_i(\sigma_t,t_k^i)$
\\
\hline
\multirow{1}{*}{2 }
& If $t<t_{k+1}^i$ in (\ref{event1}) or (\ref{event2}) & If $\xi<\xi_k^i$ in (\ref{event3}) or (\ref{event4})
\\
\hline
\multirow{4}{*}{3}
& then
&\\
& monitoring the states of $i$'s neighborhood $x_j(t)$, $j\in \mathcal N_i$ and target $s(t)$ (if $i$ is pinned at time $t$), $t\ge t_k^i$
&
\\
\hline
\multirow{2}{*}{4 }
&else
&else\\
&
go to step 1, replace $t_k^i$ by $t_{k+1}^i$
& go to step 1, replace $t_k^i$ by $t_k^i+\xi_k^i$
\\
\hline
\end{tabular}
\end{table}
\begin{remark}
In the discrete monitoring scenario, each node does not need to observe the information of its neighbors at every instants, but each node has to broadcast its updating law, $\theta^{i}_{k}$, to all its neighborhood once it is triggered.
\end{remark}
Similar to Proposition \ref{Zeno1}, we have:
\begin{proposition}\label{Zeno2}
\begin{itemize}
\item [(1)]
Suppose that hypotheses in Theorem \ref{thm2} hold. Under the rule (\ref{event3}) and the scheme described in Theorem \ref{thm2}, if system (\ref{pinning_ds1},\ref{pinning_ds2}) is not stable at $t$, there exists at least one node $i\in\{1,\cdots,m\}$ such that the next triggering event time strictly greater than $t$, namely, inter-event interval is strictly positive.
\item [(2)] Under the rule (\ref{event4}) and the scheme described in Theorem \ref{thm2}, if system (\ref{pinning_ds1},\ref{pinning_ds2}) does not reach stability,
the expectation of next inter-event interval of every node is strictly positive, further, it is lower bounded by some positive constant.
\end{itemize}
\end{proposition}
\begin{remark}
The discrete monitoring strategy implies the triggering events happen more frequently than continuous monitoring as a reward of a smaller cost of monitoring.
\end{remark}
\begin{remark}
For the discrete monitoring strategy, the computation complexity for every task depends on the number of multiplies in $\rho(\cdot)$ and $\varrho(\cdot)$. We suppose the number of multiplies of $\rho(\cdot)$ and $\varrho(\cdot)$ are respectively $N_1$, $N_2$. From the updating rules (\ref{event3}) and (\ref{event4}), the computation complexity for the next triggering time of every agent is at most $(m+1)N_1+N_2$.
\end{remark}
\section{Examples}
In this section, we present several numerical examples to illustrate these theoretical results. The system is an array of $5$ coupled Chua circuits with the map $f(\cdot)$ of node dynamics as follows:
\begin{eqnarray}
f(z)=\left[\begin{array}{c}p*(-z_{1}+z_{2}-g(z_{1}))\\
z_{1}-z_{2}+z_{3}\\
-q*z_{2}\end{array}\right]
\end{eqnarray}
where $g(z_{1})=m_{1}*z_{1}+1/2*(m_{0}-m_{1})*(|z_{1}+1|-|z_{1}-1|)$, with the parameters taken values as $p=9.78$, $q=14.97$, $m_{0}=-1.31$ and $m_{1}=-0.75$, which implies that the intrinsic node dynamics (without coupling terms) have a double-scrolling chaotic attractor \cite{chua}. Let $P=\Gamma=G=I_{3}$, where $I_{3}$ stands for the identity matrix of three dimensions. To estimate the parameter $\beta$ in the $QUAD$ condition, noting the Jacobin matrices of $f$ is one of the following
\begin{eqnarray*}
A_{1}=\left[\begin{array}{lll}-2.445&9.78&0\\
1&-1&1\\0&-14.97&0\end{array}\right],A_{2}=\left[\begin{array}{lll}3.0318&9.78&0\\
1&-1&1\\0&-14.97&0\end{array}\right],
\end{eqnarray*}
then we estimate $\beta'=\alpha-\lambda_{\max}((A_{2})^{s})=\alpha-9.1207$, where $9.1207$ is the largest eigenvalue of the symmetry parts of all Jacobin matrices of $f$.
The possible coupling graph topologies are shown in Fig.\ref{network_topology}. To select the pinned nodes, we add an extra virtual node (on behalf of $s(t)$) to the original network, which has a few links to the node that are pinned, then have an {\em extended network}. From the results in \cite{han}, one can see that if every extended network topology among the switching is strongly connected and the duration time at each network topology is sufficiently long, then there always exist positive matrices $P(\cdot)$, coupling gain $c$ and pinning gain $\epsilon$ such that condition (\ref{slow_condition}) holds, which implies that the system with persistent coupling and control can be stabilized. Hence, according to this viewpoint, one node from every strongly connected component is picked to be pinned. Noting topologies given in Fig.\ref{network_topology} are all connected, every possible pinned node set is applicable.
In this simulation, the graph topologies and pinned nodes of coupled system (\ref{pinning_ds1},\ref{pinning_ds2}) switch among these four states, as shown in Fig.\ref{network_topology} (a)-(d) respectively, induced by a homogeneous Markov chain, $\sigma_t$. $L(\sigma_t)$ is picked as the Laplacian of the graph $\mathcal G(\sigma_t)$, where each link has uniform weight $1$. Here, we pick the state space of the Markov chain $\sigma_t$ by $\{1,2,3,4\}$, and its transition matrix is given by
\begin{eqnarray*}
T=\left[
\begin{array}{cccc}
-10 & 6.5 & 0 & 3.5 \\
7 & -10 & 3 & 0 \\
0 & 1 & -10 & 9\\
4 & 6 & 0 & -10
\end{array}
\right].
\end{eqnarray*}
It can be seen from the transition matrix of $\sigma_t$, the expected sojourn time in each graph follows an exponential distribution with parameter $0.1$.
In the following, we pick $\alpha=10$ and $\beta=0.8803.$
The ODEs (\ref{pinning_ds1},\ref{pinning_ds2}) are numerically solved by the Euler method with time step $0.001$ (sec) and the time duration of the numerical simulations is $[0,10]$ (sec).
\begin{figure}[!t]
\centering
\subfigure[Pinned set $\{2,6\}$]{
\includegraphics[height=.35\textwidth,width=.4\textwidth]{figtst1.eps}}
\subfigure[Pinned set $\{5,8\}$]{\includegraphics[height=.35\textwidth,width=.4\textwidth]{figtst2.eps}}
\subfigure[Pinned set $\{2,6\}$]{\includegraphics[height=.35\textwidth,width=.4\textwidth]{figtst3.eps}}
\subfigure[Pinned set $\{2,5\}$]{\includegraphics[height=.35\textwidth,width=.4\textwidth]{figtst4.eps}}
\label{network_topology}
\caption{The topologies of the graph of the coupled system and the pinned set.}
\end{figure}
\subsection{Continuous monitoring}
We give two examples to illustrate the updating rules (\ref{event1}) and (\ref{event2}) respectively. Pick $\delta =0.03$, $c=20$, $a = b =0.5$ and $\epsilon=0.5$. Then, it can be verified that $\{\alpha I_{m}-kL(u)-k\epsilon D(u)\}^s$ are negative definite and so the matrix inequality (\ref{slow_condition}) is satisfied. And we have $\frac{(\beta\underline{\lambda}-\frac{1}{2}\delta\bar{\lambda})}
{\sqrt{c}\bar{\lambda}}=0.2736$.
We employ rule (\ref{event1}). Fig. \ref{fig_variation_x_con1} shows the dynamics of each components of the $10$ nodes and Fig. \ref{fig_variation_v} shows the dynamics of $V(t)$. All show that the coupled system (\ref{pinning_ds1}, \ref{pinning_ds2}) is stable. Similarly, we also employ the triggering event rule (\ref{event2}). Fig. \ref{fig_variation_x_con2} illustrates the dynamics of each components of all nodes, and Fig. \ref{fig_variation_v} illustrates the dynamics of $V(t)$. One can see that the coupled systems (\ref{pinning_ds1},\ref{pinning_ds2}) is asymptotically stable at certain chaotic homogeneous trajectory.
\begin{figure}[!t]
\centering
\includegraphics[height=.6\textwidth]{fig_variation_x_con1.eps}
\caption{For continuous monitoring with event triggering (\ref{event1}), the dynamics of components of the coupled system ((\ref{pinning_ds1},\ref{pinning_ds2})).}
\label{fig_variation_x_con1}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[height=.6\textwidth]{fig_variation_x_con2.eps}
\caption{For continuous monitoring with event triggering (\ref{event2}), the dynamics of components of the coupled system ((\ref{pinning_ds1},\ref{pinning_ds2})).}
\label{fig_variation_x_con2}
\end{figure}
\subsection{Discrete monitoring}
In this subsection, we illustrate the discrete-time monitoring strategies as described in Theorem \ref{thm2}. In these examples, we also take $c=20$, $a = b = 0.5$, and $\epsilon=0.5$. Fig.\ref{fig_variation_x_dis1} shows the dynamics of each components of the $10$ nodes and Fig.\ref{fig_variation_v} shows the dynamics of $V(t)$ under rule (\ref{event3}) in Theorem \ref{thm2}. All of them show that the coupled system (\ref{pinning_ds1},\ref{pinning_ds2}) is stable. By employing rule (\ref{event4}) in Theorem \ref{thm2}, Fig. \ref{fig_variation_x_dis2} illustrates the dynamics of each components of all nodes and Fig. \ref{fig_variation_v} illustrates the dynamics of $V(t)$. These plots show that the coupled system (\ref{pinning_ds1},\ref{pinning_ds2}) is asymptotically stable.
\begin{figure}[!t]
\centering
\includegraphics[height=.6\textwidth]{fig_variation_x_dis1.eps}
\caption{For discrete monitoring with event triggering (\ref{event3}), the dynamics of components of the coupled system ((\ref{pinning_ds1},\ref{pinning_ds2})).}
\label{fig_variation_x_dis1}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[height=.6\textwidth]{fig_variation_x_dis2.eps}
\caption{For discrete monitoring with event triggering (\ref{event4}), the dynamics of components of the coupled system ((\ref{pinning_ds1},\ref{pinning_ds2})).}
\label{fig_variation_x_dis2}
\end{figure}
\begin{figure}[!t]
\begin{center}
\includegraphics[height=.4\textwidth]{fig_compare_variation_v.eps}
\caption{The dynamics of Lyapunov function $V(t)$ for systems with continuous control law, continuous monitoring with rules (\ref{event1}),(\ref{event2}), discrete monitoring with rules (\ref{event3}), (\ref{event4}). } \label{fig_variation_v}
\end{center}
\end{figure}
\begin{figure}[!t]
\centering
\subfigure[]{
\includegraphics[width=.8\textwidth]{fig_con_vs_dis1.eps}
\label{fig_con_vs_dis1}
}
\subfigure[u]{
\includegraphics[width=.8\textwidth]{fig_con_vs_dis2.eps}
\label{fig_con_vs_dis2}
}
\caption{Histogram of triggering times of each node in $[9,10]$s. Under rules (a) (\ref{event1}) and (\ref{event3}), (b) (\ref{event2}) and (\ref{event4}). }
\end{figure}
Furthermore, as shown in Figs.\ref{fig_con_vs_dis1}-\ref{fig_con_vs_dis2}, the events of updating the diffusion and pinning terms in the discrete monitoring strategy is much more than the continuous monitoring strategy, as we expected as Table \ref{table1}. As a trade-off, the performance of the discrete monitoring in terms of convergence rate of $V(t)$ is higher than the continuous monitoring, as shown by Fig.\ref{fig_variation_v}.Further, it can be seen from Fig. \ref{fig_variation_v} that the rules (\ref{event1}) and (\ref{event3}) have higher convergence rates than (\ref{event2}) and (\ref{event4}) and close to the original coupled system with simultaneous diffusion and pinning, as a reward of high event frequency of updating diffusion and pinning terms, as shown in Figs.\ref{fig_con_vs_dis1}-\ref{fig_con_vs_dis2}.
\section{Strength, limitations of the work and orients of future research}
Event-triggered algorithm is a new issue in the coordination control. Despite attracting increasing interests recently, there are a small number of papers, for instance \cite{Gao}-\cite{ZLiu}, which were concerned with pinning control of networks with event-triggered algorithm. Moreover, all of them can not handle the scenario considered in this paper. \cite{Gao} studied distributed event-triggered mechanism for pinning control of networks with {\em static} topology, while in our paper, we studied pinning networks with Markovian switching topologies and Markovian switching pinned node set.
\cite{Alderisio,ZLiu} investigated event-triggered pinning control of networks with static and switching topologies. However, from the sufficient conditions for complete synchronization that were given in \cite{Alderisio} and \cite{ZLiu}, the pinned coupled system with each possible topology among the switching topologies should be able to stabilize the coupled system; in comparison, in our work, there may exist some network topology and pinned node set in the state space of the Markovian chain that cannot stabilize the coupled subsystem. Moreover, they have not considered the switching of pinned node set, which were taken into consideration in the present paper.
However, there are a few limitations of the present methods. First, the present study assumes there is no delay in information transitions. But real networks may have limit bandwidth limitation that will cause delays in message delivery. An interesting future research may take the time-delays into consideration. Second, in this work, we assume that the possible graph topologies and corresponding coupling matrices are already given and induced by a Markovian chain. It is sufficient for constructing a Lyapunov function to prove the stability of system. But in reality, the coupling weights of every possible graph topology may be time-varying. it is an important issue, and will be addressed in the near future.
Third, this work mainly shows that if the linearly coupled system with persistent diffusion and control can be stabilized, then the proposed event-triggered rules can stabilize the system, too. It is important to extend the model to nonlinear cases. In \cite{lxwang1}-\cite{lxwang3}, the authors proved that fuzzy systems are universal approximators for nonlinear dynamic systems. Hence, applying event-triggered strategies to fuzzy systems can be seen as a modest step. For this issue, we refer readers to \cite{Qiu1}-\cite{Qiu3} . Recently, \cite{Peng} proposed a centralized event-triggered communication scheme for networked Takagi-Sugeno fuzzy systems, while distributed event-triggered algorithms for fuzzy systems are absent. This also leads to interesting orients of our future work.
\section{Conclusions}
In this paper, event-triggered configurations and pinning control are employed to realize stability in linearly coupled dynamical systems with Markovian switching in both coupling matrix and pinned node set. Two monitoring scenarios are considered.
For continuous monitoring, each node observes its neighborhood's state and the target's state (if it is pinned) in an instantaneous way to determine the next triggering event time for updating state information. Instead, for discrete monitoring, each node can only obtain the state information at the event time or switching time of the underlying Markov chain to predict the next triggering event time for updating state information.
Once an event for a node is triggered, the diffusion coupling term and feedback control term of this node is updated.
Event triggering criteria are derived for each node that can be computed in a parallel way.
For both scenarios, it is proved that the coupled system can realize stability and the rule of piece-wise constant diffusion and pinning (if pinned) terms can efficiently reduce the computation load of the networked system, in comparison to the original coupled system. In addition, the discrete monitoring strategy can also reduce the communication load as well. Zeno behaviors can be proved excluded by proving the positivity of the lengths of the inter-event time intervals for some rules. Simulations are given to verify these theoretical results.
\section*{Acknowledgements}
This work is jointly supported by the National Natural Sciences
Foundation of China under Grant (Nos. 61273211 and 61273309), the Marie
Curie International Incoming Fellowship from the European Commission
(FP7-PEOPLE-2011-IIF-302421), the Program for New Century Excellent Talents
in University (NCET-13-0139), and the Programme of Introducing Talents of Discipline to
Universities (B08018).
|
\section{Introduction}
For decades, options have been widely used as a tool for investment and risk management. As of 2012, the daily market notional for S\&P 500 options is about US\$90 billion and the average daily volume has grown rapidly from 119,808 in 2002 to 839,108 as of Jan 2013\footnote{See \url{http://www.cboe.com/micro/spx/introduction.aspx}}. Empirical studies on options returns often assume that the options are held to maturity (see \cite{Broadie2009} and references therein). For every liquidly traded option, there is an embedded timing flexibility to liquidate the position through the market prior to expiry. Hence, an important question for effective risk management is: when is the best time to sell an option? In this paper, we propose a risk-adjusted optimal stopping framework to address this problem for a variety of options under different underlying price dynamics.
In addition to maximizing the expected discounted market value to be received from option sale, we incorporate a risk penalty that accounts for adverse price movements till the liquidation time. For every candidate strategy, we measure the associated risk by integrating over time the realized shortfall, or more generally its transformation in terms of a loss function, of the option position. As such, our integrated shortfall risk penalty is path dependent and introduces the trade-off between risk and return for every liquidation timing strategy.
Under a general diffusion model for the underlying stock price, we formulate an optimal stopping problem that includes an integral penalization term. To this end, we define and apply the concept of optimal liquidation premium which represents the additional value from optimally waiting to sell, as opposed to immediate liquidation. As it turns out, it is optimal for the option holder to sell as soon as this premium vanishes. This observation leads to a number of useful mathematical characterizations and financial interpretations of the optimal liquidation strategies for various positions.
We first identify the conditions under which it is optimal to immediately liquidate or hold the option position through expiration. The investigation of the non-trivial liquidation strategies involves the analytical and numerical studies of the inhomogeneous variational inequality associated with the optimal stopping problem. In a related work, \cite{budhi12} examines the solution structure for a finite maturity optimal stopping problem under L\'{e}vy processes with a running cost and other features, and an inhomogeneous variational inequality also arises from the associated partial integro-differential free-boundary problem. In the context of asset management, \cite{egamiasset} study a perpetual optimal stopping problem with a running cash flow generated from dividend and
coupon payments, and they solve a time-independent inhomogeneous variational inequality. For the variational inequalities in our liquidation problems, we prove the existence and uniqueness of a strong solution \`a la \cite{Bensoussan78} (see Section \ref{Sect-ExUn} below) under general conditions applicable to both geometric Brownian motion (GBM) (see \cite{Merton1973}) and exponential Ornstein-Uhlenbeck (OU) (see \cite{Ornstein1930}) models for the underlying dynamics. We also provide some mathematical characterizations and numerical examples of the optimal liquidation strategies for stocks, calls, puts, and straddles.
The incorporation of the risk penalty gives rise to optimal liquidation strategies that are distinctly different from the unpenalized case. For instance, if the option's Delta (derivative of the option price with respect to the underlying price) is of the same constant sign as the excess return of the underlying, then it is optimal to hold the option till maturity when there is no risk penalty
(see Prop. \ref{trivial_timing}). This applies to the case of a call option (resp. a put option) if the investor is bullish (resp. bearish) on the stock. However, under risk penalization the investor may find it optimal to liquidate a call (resp. a put) early even in the bullish (resp. bearish) scenario (see e.g. Prop. \ref{Put4}). Furthermore, the shape of the optimal liquidation region depends significantly on the risk penalty. We show that higher risk penalization coefficient always reduces the delay region, which intuitively means that the investor is more likely to sell earlier. Moreover, in some cases the optimal delay and sell regions can exhibit some interesting structures, such as disconnectedness (see Figures \ref{Put_SF} and \ref{Call_OU}). These analytical results are significantly facilitated by the properties of the optimal liquidation premium (see Theorems \ref{prop1_L_G} and \ref{compactsupport}).
Our path-dependent risk penalization model can also be viewed as an alternative way to incorporate the investor's risk sensitivity in option liquidation/exercise timing problems, as compared to the utility maximization/indifference pricing approach \citep{henderson2011optimal, LeungLudkovski2, LSZ12}. On the other hand, \cite{LeungLudkovski2011} investigate the optimal timing to buy equity European and American options without risk penalty under incomplete markets, where the investor is assumed to select risk-neutral pricing measure different from the market's. \cite{LeungLiubookchap} also discuss the timing to sell an option under the GBM model without any risk penalty, which is a special example of our model.
As is well known, the concept of risk measures based on shortfall risk has been applied to many portfolio optimization problems; see \cite{Artzner,Rockafellar, FS02, FScB, sircarRM}, and references therein. Our model applies this idea to options trading as a path-penalty associated with each liquidation strategy. As a variation of the shortfall we also introduce a risk penalty based on the quadratic variation of option price process. In particular, we obtain an explicit closed-form solution for the liquidation of a stock with quadratic penalty under the GBM model (see Theorem \ref{perpL}). Through examining the optimal liquidation premium, we also compare the liquidation strategies for calls and puts under the shortfall-based and quadratic risk penalties. \cite{Forsyth2012} also adopt the mean-quadratic-variation as a criterion for determining the optimal stock trading strategy in the presence of price impact.
The recent paper by \cite{ziemba} considers a discrete-time portfolio optimization problem with a convex loss function that accounts for the shortfall of the wealth trajectory from a benchmark. While we consider the problem of optimal liquidation of stocks and options, their investigation focuses on the optimal capital growth or Kelly strategy. On the other hand, \cite{frei13} study the optimal liquidation of a stock position subject to temporary price impact. Specifically, they minimize the mean and variance of the order slippage with respect to the VWAP
(volume weighted average price) as the benchmark. These papers adopt the stochastic control approach to solve for the optimal position over time, whereas our problems concern only the optimal timing to liquidate.
The rest of the paper is organized as follows. In Section \ref{sec:ProbForm}, we formulate the optimal liquidation problem for a generic European claim in a diffusion market. In subsequent sections, we focus on the liquidation of a stock or an option under the GBM and exponential OU models. In Sections \ref{sect-GBM} and \ref{sect-OU}, we study the optimal liquidation timing with a shortfall risk penalty. In Section \ref{sect:quadr}, we conduct our analysis with a quadratic variation risk penalty. Section \ref{sect:concl} concludes the paper. In Section \ref{Sect-ExUn}, we discuss the existence of a strong solution to the variational inequality as well as the probabilistic representation satisfied by the optimal liquidation premium.
\section{Problem Overview}\label{sec:ProbForm}
In the background, we fix a probability space $(\Omega, {\mathcal F}, \P)$, where $\P$ is the historical probability measure. The market consists of a risky asset $S$ and a money market account with a constant positive interest rate $r$. The risky asset price is modeled by a positive diffusion process following the stochastic differential equation
\begin{equation}\label{SDE_S}
dS_t =\mu(t,S_t)S_tdt + \sigma(t,S_t)S_tdW_t, \qquad S_0=s,
\end{equation}
where $W$ is a standard Brownian motion under measure $\P$ and $s>0$. Here, the deterministic coefficients $\mu(t,s)$ and $\sigma(t,s)$ are assumed to satisfy standard Lipschitz and growth conditions \cite[\S 5.2]{KaratzasShreve91} to ensure a unique strong solution to \eqref{SDE_S}. We let $\mathbb{F}=({\mathcal F}_t)_{t\geq 0}$ be the filtration generated by the Brownian motion $W$.
Let us consider a market-traded European option with payoff $h(S_T)$ on expiration date $T$ written on the underlying asset $S$. If the Sharpe ratio $\lambda(t,s) := \frac{\mu(t,s) -r}{\sigma(t,s)}$ satisfies the Novikov condition: $\E\{ \exp (\int_{0}^T\frac{1}{2}{\lambda^2(u,S_u)}\,du ) \}<\infty$, the density process
\begin{align}\label{dqdp}
\frac{d\Q}{d\P}\bigg|_{{\mathcal F}_t} = \exp\left({-\frac{1}{2}\int_{0}^t{\lambda^2(u,S_u)}\,du + \int_{0}^t{\lambda(u,S_u)}\,dW_u }\right), \qquad 0\le t\le T,
\end{align}
is a $(\P,\mathbb{F})-$martingale (see \cite{KaratzasShreve91}, Prop. 3.5.12). This defines a unique equivalent martingale (risk-neutral) measure $\Q$, and the market price of the option is given by
\begin{align}
V(t,s) = \widetilde{\E}_{t,s}\left\{e^{-r(T-t)} h(S_T)\right\}, \qquad (t,s) \in [0,T]\times \R^+.\label{mkt_pr}
\end{align}
The shorthand notation $\widetilde{\E}_{t,s}\{\cdot\} \equiv \widetilde{\E}\{\cdot|S_t = s\}$ denotes the conditional expectation under $\Q$. Note that the market price function $V(t,s)$ does not depend on the drift function $\mu (t,s)$.
Observing the stock and option price movements over time, the investor has the timing flexibility to sell the option before expiry. While seeking to maximize the expected discounted market value of the option, we incorporate a risk penalty that accounts for the downside risk up to the liquidation time. Specifically, we define the shortfall at time $t$ by
\begin{equation} \label{shortfall}
\ell(t,S_t)=(m-V(t,S_t))^+,
\end{equation}
where $m>0$ is a constant benchmark set by the investor. Then, the risk penalty is modeled as a \emph{loss function} of the shortfall, denoted by $\psi(\ell(t,S_t))$. Here, the loss function $\psi:\R^+ \to \R$ is assumed to be increasing, convex, continuously differentiable, with $\psi(0)=0$ (see e.g. \citep[Chap 4.9]{FScB}). As a result, the investor faces the penalized optimal stopping problem
\begin{equation} \label{V_prob}\\
J^{\alpha}(t,s) = \sup_{\tau \in \mathcal{T}_{t,T}}\E_{t,s}\left\{e^{-r(\tau-t)}V(\tau,S_\tau)-\alpha \int_t^{\tau} e^{-r(u-t)} \psi\left((m-V(t,S_t))^+\right)du\right\},
\end{equation}
where $\alpha\ge 0$ is a penalization coefficient and $\mathcal{T}_{t,T}$ is the set of $\mathbb F$-stopping times taking values in $[t,T]$.
Unless otherwise noted, our analysis applies to a general loss function $\psi$ satisfying the conditions above. Here, let us give an example to visualize the penalization mechanism. For instance, one can set the benchmark to be the initial option price, and take $\psi(\ell) = \ell$. Then, the penalty term amounts to accumulating the (discounted) area when the option is below its initial cost. We illustrate this in Figure \ref{realized_shortfall}. Notice that the realized shortfall stays flat when the option price is above the benchmark, and continues to increase as long as the option is under water. Other viable specifications include the power penalty $\psi(\ell)=\ell^p$, $p\ge 1$, and the exponential penalty $\psi(\ell)=\exp(\gamma \ell)-1$, $\gamma >0$, and more.
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.65]{realized_shortfall.eps}
\caption{\small{The realized shortfall (dashed) based on a simulated price path (solid) of a European call option under the GBM model, with parameters $S_0=100$, $r=0.03$, $\mu=-0.05$ and $\sigma=0.3$, $K=100$, $T=1$, $\alpha =1$. The benchmark $m$ is the initial call option price. }}
\label{realized_shortfall}
\end{figure}
In order to quantify the value of optimal waiting, we define the optimal liquidation premium by the difference between the value function $J^{\alpha}$ and the current market price of the option, namely,
\begin{equation}
L^{\alpha}(t,s):= J^{\alpha}(t,s)-V(t,s). \label{delayed_liquidation}
\end{equation}
Alternatively, the optimal liquidation premium $L^{\alpha}$ can be interpreted as the risk-adjusted expected return from a simple buy-now-sell-later strategy.
Denote the discounted penalized liquidation value process by
\begin{align*}
Y_u=e^{-ru}V(u,S_u)-\alpha \int_0^{u} e^{-rt} \psi((m-V(t,S_t))^+)dt.
\end{align*}
In order to guarantee the existence of an optimal stopping time to problem \eqref{V_prob}, we require that $\E\{\sup_{0\leq u \leq T}Y_u\}<\infty$. For a European call option, the option value $V(t,S_t)$ is dominated by the stock price $S_t$, while the put option price is bounded by the strike price. Consequently, for any linear combination of calls and puts, it suffices to impose $\E\{\sup_{0\leq u \leq T}S_u\}<\infty$. We also require that $\P\{\min_{0\le t\le \hat{t}} S_t>0\}= 1$, which means that the asset price stays strictly positive before any finite time $\hat{t}$ a.s. Then by standard optimal stopping theory \citep[Theorem D.12]{KaratzasShreve01}, the optimal liquidation time, associated with $L(t,s)$, is given by
\begin{equation}
\tau^* = \inf\{ \,u \in [t,T]\,:\, L^{\alpha}(u,S_u) =0\, \}.\label{Ltau}
\end{equation}
In other words, it is optimal for the investor to sell the option as soon as the optimal liquidation premium $L^{\alpha}$ vanishes, meaning that the timing flexibility has no value. Accordingly, the investor's optimal liquidation strategy can be described by the sell region $\mathcal S$ and delay region $\mathcal D$, namely,
\begin{align}
\mathcal S&=\{(t, s)\in [0,T]\times \R^+ : \ L^{\alpha}(t,s)=0\},\\
\mathcal D &=\{(t, s)\in [0,T]\times \R^+ : \ L^{\alpha}(t,s)>0\}.
\end{align}
Our framework can be readily applied to the \emph{reverse} problem of optimally timing to buy an option. This amounts to changing the $\sup$ to $\inf$ in $L^\alpha$. In this paper, we shall focus on the liquidation problem.
\subsection{Analysis of the Optimal Liquidation Premium}\label{sect-ana}
\begin{theorem}\label{prop1_L_G}
Given the underlying price dynamics in \eqref{SDE_S}, the optimal liquidation premium admits the probabilistic representation
\begin{align}
L^{\alpha}(t,s)
&=\sup_{\tau \in \mathcal{T}_{t,T}} \E_{t,s} \left\{ \int_t^\tau e^{-r(u-t)} G^{\alpha}(u,S_u) \,du\right\},\label{L_G}
\end{align}
where we denote
\begin{align}
G^{\alpha}(t,s):=\big( \mu(t, s) - r\big)sV_s(t,s)-\alpha \psi\left((m-V(t,s))^+\right). \label{Gdrift}
\end{align}
\end{theorem}
\begin{proof} Applying Ito's formula to the market price in \eqref{mkt_pr}, we get
\[\E_{t,s} \left\{e^{-r(\tau-t)}V(\tau,S_\tau) \right\} -V(t,s) = \E_{t,s} \left\{ \int_t^\tau e^{-r(u-t)} \big( \mu(u, S_u) - r\big)S_uV_s(u,S_u) du\right\}.\]
Substituting this into the optimal liquidation premium in \eqref{delayed_liquidation} gives
\begin{equation*}
L^{\alpha}(t,s) = \sup_{\tau \in \mathcal{T}_{t,T}}~\E_{t,s} \bigg\{ \int_t^\tau e^{-r(u-t)}
\underbrace{\big[\big(\mu(u, S_u)-r\big)S_uV_s(u,S_u)-\alpha \psi((m-V(u,S_u))^+)\big]}_
{=\,G^\alpha(u,S_u)\text{ in \eqref{Gdrift}}}\,du\bigg\}.
\end{equation*} \end{proof}
We shall call $G^{\alpha}(t,s)$ in \eqref{Gdrift} the \emph{drive function}. We observe that it depends on the Delta $V_s \equiv \frac{\partial V}{\partial s}$ of the option and the penalty coefficient $\alpha$ reduces the drive function for every $(t,s)$. Many properties of the optimal liquidation premium $L^\alpha$ can be deduced by studying the drive function.\\
\begin{proposition} \label{trivial_timing}Let $t\in[0,T]$ be the current time.
If the drive function $G^{\alpha}(u,s)$ is positive, $\forall (u,s) \in [t,T]\times \R^+$, then it is optimal to sell at maturity, namely, $\tau^*=T$. ~
If the drive function $G^{\alpha}(u,s)$ is negative, $\forall (u,s) \in [t,T]\times \R^+$, then it is optimal to sell immediately, namely, $\tau^*=t$.
\end{proposition}
\begin{proof}
We observe from the integral in \eqref{L_G} that if the drive function $G^{\alpha}$ is positive (resp. negative), $\forall (u,s) \in [t,T]\times \R^+$, then we can maximize the expectation by selecting the largest (resp. smallest) stopping time, namely, $\tau^*=T$ (resp. $\tau^*=t$).
\end{proof}
In particular, if $V_s(t,s)$ and $(\mu(t,s)-r)$ are of different signs $\forall\,(t,s)$, then the drive function $G^{\alpha}$ is always negative, so it is optimal to sell immediately. Proposition \ref{trivial_timing} can also be applied to the perpetual case if we set $T= \infty$. In general, the delay region always contains the region where the drive function is positive, namely,
\begin{equation} \label{GsubL}
\{G^\alpha>0\}\subset\{L^\alpha>0\},
\end{equation}see e.g. \citep[Prop. 2.3]{OksendalSulemBook}. Intuitively, this means that if $G(t, s) > 0$, then the investor
should not sell immediately since an incremental positive infinitesimal premium can be obtained by waiting for an infinitesimally small amount of time.
In addition, we can infer from \eqref{L_G} the ordering of optimal liquidation premium based on the drive function.
\begin{corollary}\label{Compare_G} Consider two options $A$ and $B$, along with two penalty coefficients $\alpha_A$ and $\alpha_B$ respectively. If the drive function of $A$ dominates that of $B$, i.e. $G^{\alpha_A}_A(t,s)\geq G^{\alpha_B}_B(t,s),\forall (t,s)\in [0,T]\times \R^+$, then the optimal liquidation premium for $A$, $L^{\alpha_A}_A$, dominates that for $B$, $L^{\alpha_B}_B(t,s)$, i.e. $L^{\alpha_A}_A(t,s) \geq L^{\alpha_B}_B(t,s),\forall (t,s)\in[0,T]\times \R^+$.
\end{corollary}
The corollary allows us to compare the liquidation timing of different penalties. For example, for $0\leq \alpha_1\leq \alpha_2$, we have $G^{\alpha_1}(t,s)\geq G^{\alpha_2}(t,s)$ for the same option. It follows from \eqref{Ltau} and Corollary \ref{Compare_G} that the optimal liquidation time with penalty $\alpha_1$ is later than that with penalty $\alpha_2$.
In general, a variety of delay and sell regions can occur depending on the underlying dynamics and option payoff. Next, we give sufficient conditions so that the delay region is bounded.
\begin{theorem}\label{compactsupport}
Let $T<\infty$ and $S$ be time homogeneous. Then, the delay region is bounded provided that\\
(i) $\exists \,c>0$ s.t. $G^\alpha(t,s)<c$ for every $(t,s)\in[0,T]\times \R^+$; and\\
(ii) there exist constants $b,k>0$ such that $G^\alpha(t,s)<-b$ in $[0,T]\times[k,\infty)$.
\end{theorem}
\begin{proof}
\textbf{Step 1.} \textit{We find a function $\widehat{L}(t,s)$ that dominates $L^\alpha(t,s)$ and is decreasing in both $t$ and $s$.} To this end, we define
\begin{align*}
\widehat{G}^\alpha(s)
& := \max\{G^\alpha(t,\xi):(t,\xi)\in[0,T]\times[s,\infty)\},\\
\widehat{L}(t,s)
& := \sup_{\tau \in \mathcal{T}_{t,T}} \E_{t,s}\left\{ \int_t^\tau e^{-r(u-t)}\widehat{G}^\alpha(S_u)\,du,\right\}.
\end{align*}
By construction $\widehat{G}^\alpha:[0,T]\times\R^+ \to \R$ is constant in $t$ and decreasing in $s$. It also satisfies conditions (i) and (ii). Consequently, using the time homogeneity of $S$, we have, for $t>t'$,
\begin{equation*}
\widehat{L}(t,s)
= \sup_{\tau \in \mathcal{T}_{0,T-t}} \E_{0,s} \left\{ \int_0^\tau e^{-ru} \widehat{G}^\alpha(S_u) \,du\right\}
\leq \sup_{\tau \in \mathcal{T}_{0,T-t'}} \E_{0,s}\left\{ \int_0^\tau e^{-ru} \widehat{G}^\alpha(S_u) \,du\right\}
= \widehat{L}(t',s).
\end{equation*}
Hence, $\widehat{L}(t,s)$ is decreasing in $t$. Moreover, since $\widehat{G}^\alpha$ is decreasing in $s$, we have, for $s'>s$,
\begin{align*}
\widehat{L}(t,s')
& = \sup_{\tau \in \mathcal{T}_{t,T}} \E_{t,s'} \left\{ \int_t^\tau e^{-r(u-t)} \widehat{G}^\alpha(S_u) \,du\right\}\\
& = \sup_{\tau \in \mathcal{T}_{t,T}} \E_{t,s} \left\{ \int_t^\tau e^{-r(u-t)} \widehat{G}^\alpha(S_u+s'-s) \,du\right\}\\
& \leq \sup_{\tau \in \mathcal{T}_{t,T}} \E_{t,s} \left\{ \int_t^\tau e^{-r(u-t)} \widehat{G}^\alpha(S_u) \,du\right\} = \widehat{L}(t,s).
\end{align*}
Therefore, $\widehat{L}(t,s)$ is also decreasing in $s$.
Since by definition $\widehat{G}^\alpha$ dominates $G^\alpha$, Corollary \ref{Compare_G} implies that $L^\alpha(t,s)$ has a bounded support as long as $\widehat{L}(t,s)$ has a bounded support. Henceforth, we can assume without loss of generality that $L^\alpha$ is decreasing in both variables $t$ and $s$ and that $G^\alpha$ is time homogeneous and decreasing in $s$. In particular, we denote $G^\alpha(s) \equiv G^\alpha(t,s)$.
\textbf{Step 2.} \textit{We prove that for every $t>0$, there exists $\hat{s}<\infty$ such that $L^\alpha(t,s)=0$ for every $s>\hat{s}$}. Since $L^\alpha(t,s)$ is decreasing, it is equivalent to show that there exists no $\hat{t}\in (0,T]$ s.t. $L^\alpha(t,s)>0$ for $0\leq t \leq \hat{t}$ and $s\in \R^+$. To this end, let's suppose that such a time $\hat{t}$ exists. In other words, $\tau^* = \inf\{t\le u\le T\,:\, L(u,S_u) =0\} > \hat{t}.$
Now, we show that this leads to a contradiction. Fix $t\in[0,\hat{t})$. Condition (ii) means that there exists $k$ s.t. $G^\alpha(s)<-b<0$ in $[k,\infty)$. For $s>k$, we let $\tau_k:=\inf\{u \ge t :S_{u}\le k\}$. Since $S$ has continuous paths, we have $\tau_k>t$. Define
\begin{align}
\mathcal{K}(t,s)
:= \frac{c}{r}\E_{t,s}\left\{e^{-r(\tau_k-t)}\indic{\tau_k\leq \tau^*}\right\}
-\E_{t,s} \left\{b\int_t^{\tau^*\wedge\tau_k}e^{-r(u-t)} \,du\right\} \notag \\
= \frac{c}{r}\E_{t,s}\left\{e^{-r(\tau_k-t)} \indic{\tau_k\leq \tau^*}\right\}
-b\left(1-\E_{t,s}\left\{e^{-r(\tau^*\wedge\tau_k-t)}\right\}\right), \label{proofK}
\end{align}where $c$ is the upper bound of $G^\alpha$ in condition (i).
Next, taking $s\uparrow\infty$ yields that $\P_{t,s}(\tau_k\le T) \downarrow 0$, while $\E_{t,s}\left\{e^{-r(\tau^*\wedge\tau_k-t)}\right\}< e^{-r(\hat{t} -t)}$ since $\tau^*>\hat{t}>t$ a.s. Therefore, we obtain
\begin{equation}\notag
\beta(t,s) :=\frac{c\,\P_{t,s}(\tau_k\leq \tau^*)}{r(1-\E_{t,s}\left\{e^{-r(\tau^*\wedge\tau_k-t)}\right\})}\to 0.
\end{equation}
As a result, for a sufficiently large $s>k$, we get $b \ge \beta(t,s)$, which implies that $\mathcal{K}(t,s)\leq 0$ (see \eqref{proofK}).
Next we consider the difference
\begin{align}
L^\alpha(t,s)-\mathcal{K}(t,s)
& \leq \E_{t,s} \left\{\int_{\tau^*\wedge\tau_k}^{\tau^*} e^{-r(u-t)}G^\alpha(S_u) \,du
- \frac{c}{r}e^{-r(\tau_k-t)}\indic{\tau_k\leq \tau^*}\right\}\notag \\
& \leq e^{rt}\E_{t,s}\left\{\frac{c}{r}(e^{-r\tau^*\wedge\tau_k}
- e^{-r\tau^*})- \frac{c}{r}e^{-r\tau_k}\indic{\tau^*\leq \tau_k}\right\}\notag \\
& = -\frac{ce^{rt}}{r}\E_{t,s}\left\{e^{-r\tau^*}\indic{\tau_k\leq \tau^*}\right\}\leq0.\notag
\end{align}
This means that $L^\alpha(t,s)\leq \mathcal{K} (t,s) \leq 0$. This contradicts the assumption $L^\alpha(t,s)>0$.
\textbf{Step 3.} It remains to show at time $0$ that $\exists\,\hat{s}>0$ such that $L^\alpha(0,s)=0$ for every $s>\hat{s}$. Let $\hat{t}\in[0,T]$ and consider for every $t\in[0,T+\hat{t}]$ the optimal stopping problem
\begin{equation*}
\overline{L}^\alpha(t,s) := \sup_{\tau \in \mathcal{T}_{t,T+\hat{t}}} \E_{t,s} \left\{ \int_t^\tau e^{-r(u-t)} G^\alpha(S_u) \,du\right\}.
\end{equation*}
The time homogeneity of $S$ yields that $\overline{L}^\alpha(\hat{t},s)=L^\alpha(0,s)$. Now we apply Step 2 and conclude that there exists $\hat{s}>0$ such that $\overline{L}^\alpha(\hat{t},s)=0$ for every $s>\hat{s}$. Hence, the delay region is bounded above.
\end{proof}
We remark that the statement and the proof of Theorem \ref{compactsupport} do not involve the properties of the loss function. In other words, as long as the resulting drive function satisfies conditions (i) and (ii), the delay region is bounded. We notice that if the delay region is bounded, then there exists a constant $\bar{s}$ such that $\{L^\alpha>0\}\subseteq[0,T]\times (0,\bar{s})$. We will utilize Theorem \ref{compactsupport} repeatedly when we discuss the liquidation strategies in subsequent sections.
\subsection{Applications to GBM and Exponential OU Underlyings} \label{sect:GBM_OU}
Henceforth, we shall investigate analytically and numerically the optimal liquidation timing when $S$ follows (i) the geometric Brownian motion (GBM) model with $\mu(t,s)=\mu$ and $\sigma(t,s)=\sigma>0$, as well as (ii) the exponential OU model with $\mu(t,s)=\beta(\theta-\log(s))$ and $\sigma(t,s)=\sigma>0$.
We will study the liquidation timing of a stock, European put and call options. For both the GBM and exponential OU cases, the risk-neutral measure $\Q$ is uniquely defined by \eqref{dqdp}, and the Novikov condition is satisfied (see Appendix \ref{appendix_A}). Furthermore, the $\Q$ dynamics of $S$ is a GBM with drift $r$ and the no-arbitrage prices (see \ref{mkt_pr}) of a call and a put with strike $K$ and maturity $T$ are given by
\begin{equation}
C(t,s)=s\,\Phi(d_1)-Ke^{-r(T-t)}\Phi(d_2), \quad \text{ and } \quad P(t,s)=Ke^{-r(T-t)}\Phi(-d_2)-s \,\Phi(-d_1), \label{CP}
\end{equation}
where $\Phi$ is the standard normal c.d.f. and
\begin{equation}
d_1=\frac{\log(\frac{s}{K})+(r+\frac{\sigma^2}{2})(T-t)}{\sigma\sqrt{T-t}} , \qquad d_2=d_1-\sigma\sqrt{T-t}. \notag
\end{equation}
In order to numerically compute the non-trivial liquidation strategy, we solve the variational inequality (VI) of the form
\begin{align}
\text{min}\bigg\{-L^{\alpha}_t-\mu(t,s)s\,L^{\alpha}_s-\frac{\sigma^2(t,s)s^2}{2}L^{\alpha}_{ss} + r L^{\alpha}-G^{\alpha}, \ L^{\alpha}\bigg\}=0, \label{L_general_VI}
\end{align}
with terminal condition $L^{\alpha}(T,s)=0$ and where $(t,s) \in [0,T) \times \mathbb{R}^+$. In Section \ref{Sect-ExUn}, we show that the above VI admits a unique strong solution in the terminology of \cite{Bensoussan78} under conditions that include the GBM and exponential OU cases (see Theorem \ref{ExistenceStrongSolution}). For implementation, we adopt the Crank-Nicholson scheme for the VI \eqref{L_general_VI} on a finite (discretized) grid $D=[s_{\min},s_{\max}]\times[0,T]$. We refer to the book by Glowinski \citep[Chap.3]{Glowinski} for details on numerical methods for solving inhomogeneous VIs of parabolic type.
\section{Optimal Liquidation with a GBM Underlying}\label{sect-GBM}
We begin our first series of illustrative examples under the GBM model. In view of Proposition \ref{trivial_timing}, we observe that, if $\mu\leq r$, it is never optimal to hold a stock or a call (or in general any positive delta position) regardless whether we introduce a risk penalty or not. On the other hand, Proposition \ref{trivial_timing} also implies that, if $\mu>r$ and $\alpha=0$, it is always optimal to delay.
However, with a non-zero risk penalty ($\alpha>0$), the solution can be non-trivial. To see this, we note that the drive function associated with a call is given by $G^\alpha_{Call}(t,s)=(\mu-r)sC_s-\alpha\psi((m-C(t,s))^+)$ where $C(t,s)$ is the call price in \eqref{CP}. In particular, the penalty term is strictly positive at $s=0$ and decreasing, and it vanishes for large $s$. On the other hand, the first term $(\mu-r)sC_s$ is strictly increasing from zero at $s=0$. This implies that there exists a price level $\hat{s}$ such that $G^\alpha_{Call}(t,s)$ is positive in $[0,T]\times[\hat{s},\infty)$. In turn, it follows from \eqref{GsubL} that the sell region must be bounded (possibly empty) and the delay region is unbounded. The same argument applies to the case with a stock. Figure \ref{Stock_Call_SF_1} illustrates this.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.55]{Stock_Neumann_SF_1.eps}\qquad ~~
\includegraphics[scale=0.55]{Call_Neumann_SF_1.eps}
\caption{\small{The optimal liquidation boundaries (solid) and the zero contours of $G^{\alpha}$ (dashed) of a stock (left panel) and a call option (right panel). We take $T=0.5$, $r=0.03$, $\mu=0.08$, $\sigma=0.3$, $K=50$, $\alpha=0.1$. The loss function is given by $\psi(\ell)=\ell$, with the benchmark $m=50$ for the stock and $m=C(0,K)$ for the call. }}
\label{Stock_Call_SF_1}
\end{figure}
Next, we consider the liquidation of a put option. Recall the put price $P(t,s)$ given in \eqref{CP}. Its negative Delta implies that for $\mu\geq r$ the drive function $G_{Put}^\alpha(t,s) \le 0$,\, $\forall (t,s)$, meaning that it is optimal to sell immediately by Proposition \ref{trivial_timing}. In contrast, when $\mu<r$, the sell region is empty if $\alpha=0$, but under risk penalization the optimal strategy may be non-trivial. \\
\begin{proposition}\label{Put4}
Consider the optimal liquidation of a put under the GBM model with $\mu<r$ and $\alpha>0$. Then, the delay region is {bounded}.
Furthermore, it is non-empty if $m<K$ and $\exists \, \hat{t} \in [0,T]$ such that $\alpha\psi'((m-P(\hat{t},0))^+)<r-\mu$.
\end{proposition}
\begin{proof}
The drive function for the put $G^\alpha \equiv G^\alpha_{Put}(t,s)=(r-\mu)s\Phi(-d_1)-\alpha\psi((m-P(t,s))^+)$ satisfies
\begin{align}
\lim_{s\to 0} G^\alpha(t,s)&=-\alpha\psi((m-K)^+) \leq 0, \label{Put1} \\
\lim_{s\to \infty} G^\alpha(t,s)&=-\alpha\psi(m) < 0, \label{Put2} \\
\frac{\partial G^\alpha}{\partial s}(t,s)&=[r-\mu-\alpha\psi'((m-P(t,s))^+)\indic{m>P(t,s)}]\Phi(-d_1)-(r-\mu)s\Gamma, \label{Put3}
\end{align}
where $\Gamma=\frac{\partial^2 P}{\partial s^2}\geq 0$. In turn, we fix any $\hat{b}\in (0,\alpha \psi (m))$ and define $\overline{\psi}(\ell):=\min\{\psi(\ell),\hat{b}\}$. This implies the inequality\begin{equation*}
\overline{G}^\alpha(t,s):=(r-\mu)s\Phi(-d_1) - \alpha \overline{\psi}((m-P(t,s))^+) \geq G^\alpha(t,s).
\end{equation*}
Then by Corollary \ref{Compare_G}, we only need to show that $\overline{G}^\alpha$ satisfies the assumptions of Theorem \ref{compactsupport}. We observe that $\overline{G}^\alpha$ is bounded above and it follows from \eqref{Put2} that $\lim_{s\to\infty}\overline{G}^\alpha(t,s)\to -\alpha \hat{b} <0$ for every $t\in[0,T]$. Moreover, there exists $\hat{s}>0$ such that for every $s>\hat{s}$, $\overline{\psi}((m-P(t,s))^+)=\hat{b}$. Consequently, we have
\begin{equation*}
\frac{\partial \overline{G}^\alpha}{\partial t}
= (\mu-r)s\phi(d_1)\frac{\log(\frac{s}{K})-(r+\frac{\sigma^2}{2})(T-t)}{2\sigma(T-t)^\frac{3}{2}}\leq 0,
\end{equation*}
for $s>\max\{\hat{s}, K\exp((r+\sigma^2/2)T)\}$ and $t\in[0,T]$. Since $\overline{G}^\alpha (0,s)\to -\alpha \hat{b}$ as $s\to-\infty$, we can choose $b\in(0,\alpha \hat{b})$ such that $\exists k>\max\{\hat{s}, K\exp((r+\sigma^2/2)T)\}$ and $-b>\overline{G}(0,s)>\overline{G}(t,s)$ in $[0,T]\times [k,\infty)$. Therefore, $\overline{G}$ satisfies the assumptions of Theorem \ref{compactsupport}.
Finally, suppose $\exists \, \hat{t} \in [0,T]$ such that $\alpha\psi'((m-P(\hat{t},0))^+)<r-\mu$, where $m<K$. It follows from \eqref{Put1} and \eqref{Put3} that $G^{\alpha}(\hat{t},0)=0$ and $\frac{\partial G^{\alpha}}{\partial s}(\hat{t},0)>0$, so that the set $\{G^\alpha>0\}$ is non-empty. In turn, the inclusion \eqref{GsubL} implies that the delay region is also non-empty.
\end{proof}
\begin{remark} As an example, the delay region is empty if $\alpha\psi'((m-P(t,0))^+)\geq r-\mu>0,\,\forall \,t \in [0,T]$. Indeed, since we have $G^\alpha(t,0)\leq 0$ and $\frac{\partial G^{\alpha}}{\partial s}(t,0)\leq 0,\,\forall\,t\in[0,T]$, $G^\alpha$ cannot be strictly positive. By Proposition \ref{trivial_timing}, it is optimal to sell immediately.
\end{remark}
Proposition \ref{Put4} is illustrated in Figure \ref{Put_SF}. In these examples, the delay region is non-empty and the sell region is unbounded but may be disconnected (Figure \ref{Put_SF} (right)). This can arise when, for example, $G^\alpha(t,0)<0$ for every $t\in[0,T]$, but $\min_t\max G^\alpha(t,s)>0$. The intuition for a disconnected sell region is as follows. If the put is deeply in the money (i.e. when $S_t$ is close to zero), its market price has very limited room to increase since it is bounded above $Ke^{-r(T-t)}$. At the same time, delaying sale further will incur a penalty. Therefore, when the penalization coefficient $\alpha$ is high, it is optimal to sell at a low stock price level. On the other hand, if the put is deep out of the money (i.e. when $S_t$ is very high), the market price and the Delta of the put are close to zero, meaning the drive function becomes more negative and selling immediately is optimal.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.55]{Put_Neumann_SF_1.eps}\qquad ~~
\includegraphics[scale=0.55]{Put_Neumann_SF_2.eps}
\caption{\small{The optimal liquidation boundary (solid) and the zero contour of $G^{\alpha}$ (dashed) of a put option under GBM dynamics with the loss function $\psi(\ell)=\ell$. We take $m=2K$, $\alpha=0.001$ (left panel), and $m=P(0,K)$, $\alpha=0.01$ (right panel). Parameters: $T=0.5$, $r=0.03$, $\mu=0.02$, $\sigma=0.3$, $K=50$.}}
\label{Put_SF}
\end{figure}
For a long position in calls and the underlying stock, or in puts, the Delta $C_s$ takes a constant sign. As an example of a derivative with a Delta of non-constant sign, we consider a long straddle. This is a combination of a call and a put with strike prices $K_1 \leq K_2$ respectively and the same maturity $T$. The payoff of a straddle is given by $h^{STD}(S_T):=(S_T-K_1)^+ + (K_2-S_T)^+$. The market price of a long straddle, denoted by $C^{STD}$, is simply the sum of the respective Black-Scholes call and put prices, i.e. $C^{STD}(t,s)=C(t,s)+P(t,s)$. For simplicity, we set $K_1=K_2=K$.
\begin{proposition}\label{STD}
For the optimal liquidation of a long straddle position under the GBM model, it follows that \\
(i) if $\mu=r$, the delay region must be empty;\\
(ii) if $\mu>r$, the delay region is unbounded;\\
(iii) if $\mu<r$, the delay region is bounded.
\end{proposition}
\begin{proof}
The straddle's drive function is $G^{\alpha}_{STD}(t,s)=(\mu-r)sC_s^{STD}(t,s)-\alpha \psi((m-C^{STD}(t,s))^+)$. For $\mu=r$, the conclusion follows immediately by Proposition \ref{trivial_timing}. If $\mu>r$ we simply notice that $G_{STD}^\alpha(t,s) \to \infty$ as $s\to\infty$ for every $t\in[0,T]$, and the assertion follows from the inclusion \eqref{GsubL}.
Now suppose $\mu<r$. We will show that $G^\alpha$ satisfies the assumptions of Theorem \ref{compactsupport}. Clearly, $G_{STD}^\alpha$ is bounded above. Since $C^{STD}(t,s) \to \infty$ as $s\to\infty$ for every $t\in[0,T]$, then there exists $\hat{s}>0$ such that, for every $s>\hat{s}$ and $t\in[0,T]$, $\psi((m-C^{STD}(t,s))^+) = 0$. Moreover, for $s>\max\{\hat{s},K\exp\left((r+\sigma^2/2)T\right)\}$, we have
\begin{equation*}
\frac{\partial \Phi(d_1)}{\partial t} = \phi(d_1)\frac{\log(\frac{s}{K})-(r+\frac{\sigma^2}{2})(T-t)}{2\sigma(T-t)^{\frac{3}{2}}} > 0,
\end{equation*}
and thus
\begin{equation*}
\frac{\partial G_{STD}^\alpha}{\partial t}(t,s)=2(\mu-r)s\frac{\partial \Phi(d_1)}{\partial t} \leq 0.
\end{equation*}
This implies $G_{STD}^\alpha(0,s)\geq G_{STD}^\alpha(t,s)$ for every $t\in[0,T]$. Since $G_{STD}^\alpha(0,s)\to -\infty$ as $s\to \infty$, for a fixed $b>0$ there exists $k_b>0$ such that $G^\alpha_{STD}(0,s)<-b$ for every $s\geq k_b$. Therefore, setting $k=\max\{\hat{s},K\exp\left((r+\sigma^2/2)T\right),k_b\}$, we have $G^\alpha_{STD}(t,s)\leq G^\alpha_{STD}(0,s)<-b$ in $[0,T]\times[k,\infty)$. Therefore, the assumptions of Theorem \ref{compactsupport} are satisfied and we conclude.
\end{proof}
In particular, Proposition \ref{STD} suggests that when $\mu<r$, the sell region is unbounded, even if $\alpha=0$. In Figure \ref{Straddle_SF2}, we illustrate the optimal liquidation boundaries for cases (ii) and (iii). When the investor is bullish (left panel: $\mu = 0.08 > 0.03 = r$), the liquidation boundary is increasing and the delay region is on top of the sell region. Interestingly, the opposite is observed when the investor is bearish (right panel: $\mu = 0.02 < 0.03 = r$).
\begin{figure}[h!]
\centering
\includegraphics[scale=0.55]{Straddle_Bull_Neumann_SF_1_2.eps}\qquad ~~
\includegraphics[scale=0.55]{Straddle_Bear_Neumann_SF_1_2.eps}
\caption{\small{Optimal liquidation boundary and the zero contour of $G^{\alpha}$ for a straddle under the GBM model with the loss function $\psi(\ell)=\ell$. We set $K=50$, $m=C^{STD}(0,K)$, $\alpha=0.1$, $r=0.03$, $\mu=0.08$ (left panel) and $\mu=0.02$ (right panel).}}
\label{Straddle_SF2}
\end{figure}
We end this section by discussing the liquidation timing of a stock with an infinite horizon ($T=\infty$). This leads to the following stationary optimal stopping problem
\begin{equation}\label{PerpetualStock}
L(s) = \sup_{\tau \in \mathcal{T}} \E_{s} \left\{ \int_0^\tau e^{-ru} G^{\alpha}(S_u) \,du\right\}.
\end{equation}
where $G^\alpha(s) = (\mu-r)s - \alpha \psi((m-s)^+)$ and $\mathcal{T}$ is the set of $\Fil$-stopping times taking values in $[0,\infty]$. When $\mu\leq r$, selling immediately is optimal according to Proposition \ref{trivial_timing}, as for the case with finite maturity. As it turns out, the liquidation problem has the opposite trivial solution when $\mu>r$, that is, it is optimal to hold forever. \\
\begin{proposition}
If $\mu>r$, then the value function $L(s)$ in \eqref{PerpetualStock} is infinite and it is optimal to never sell the stock.
\end{proposition}
\begin{proof}
Consider a candidate stopping time $\tau=\infty$. Then, by applying Tonelli's theorem, we have
\begin{align*}\notag
\E_s\left\{\int_0^{\infty} e^{-ru}G^\alpha(S_u)du\right\}
& = \E_s\left\{\int_0^{\infty}e^{-ru}(\mu-r)S_udu - \alpha\int_0^{\infty} e^{-ru}\psi((m-S_u)^+)du\right\} \notag \\
& = \int_0^{\infty}e^{(\mu-r)u}(\mu-r)sdu - \alpha\int_0^{\infty} e^{-ru}\E\left\{\psi((m-S_u)^+)\right\}du \notag \\
& \geq \int_0^{\infty}e^{(\mu-r)u}(\mu-r)sdu - \alpha\int_0^{\infty} e^{-ru}\psi(m)du = \infty, \notag
\end{align*}
since $\mu>r$ and $\psi$ is increasing. Hence, $L(s)=\infty$ and it is never optimal to sell.
\end{proof}
\section{Optimal Liquidation with an Exponential OU Underlying}\label{sect-OU}
In the exponential OU model, the stock price satisfies the SDE
\begin{equation}\label{MRSDE}
dS_t = \beta(\theta - \log S_t )S_t\,dt + \sigma S_t dW_t,
\end{equation}
with $\theta \in\R$ and $\beta, \sigma >0$. Therefore, the optimal liquidation premium $L(t,s)$ is given by equation \eqref{L_G} with the drive function
\begin{equation}\label{GOUgen}
G^{\alpha}(t,s)=[\beta(\theta-\log(s))-r]sV_s(t,s)-\alpha \psi((m-V(t,s))^+),
\end{equation}where $V(t,s)$ is a generic option price in \eqref{mkt_pr}.
In contrast to the GBM case, the optimal liquidation strategy can now be non-trivial for a stock or a call when there is no penalty. More generally, we can prove that the delay region is in fact bounded. The intuition should be clear: when $S_t$ is very high, it is expected to revert back to its long-term mean, so that selling immediately becomes optimal.
\begin{proposition}\label{prob_repr_OU}
Under the exponential OU model, the delay region for a call is bounded.
\end{proposition}
\begin{proof}
The drive function $G^\alpha_{Call}$ for the call is given by \eqref{GOUgen} with $V(t,s) = C(t,s)$ (see \eqref{CP} for the call price). It is bounded above, so it satisfies condition (i) of Theorem \eqref{compactsupport}. As is well known, the call price satisfies $\frac{\partial C(t,s)}{\partial t}\leq0$. In addition, $\beta(\theta-\log(s))-r \leq 0$ iff $s \geq \exp(\theta-\frac{r}{\beta})$, and $\frac{\partial \Phi(d_1)}{\partial t} \geq 0$ for $s \geq K\exp\left((r+\sigma^2/2)T\right)$. In turn, we have
\begin{equation*}
\frac{\partial G^\alpha_{Call}}{\partial t}(t,s) = [\beta(\theta-\log(s))-r]s\frac{\partial \Phi(d_1)}{\partial t}
+ \alpha\frac{\partial C(t,s)}{\partial t}\psi'((m-C(t,s))^+)\indic{m>C(t,s)} \leq 0,
\end{equation*} for $s>\max\{\exp(\theta-\frac{r}{\beta}), K\exp\left((r+\sigma^2/2)T\right)\}$ and $t\in[0,T]$. This implies $G^\alpha_{Call}(0,s) \geq G^\alpha_{Call}(t,s)$. Fix $b>0$. Since $G^\alpha_{Call}(0,s)\to -\infty$, $\exists k_b>0$ s.t., $\forall s>k_b$, $G^\alpha_{Call}(0,s)<-b$. Hence, if we set $k=\max\{\exp(\theta-\frac{r}{\beta}), K\exp\left((r+\sigma^2/2)T\right), k_b\}$, we are guaranteed that $G^\alpha_{Call}(t,s)\leq G^\alpha_{Call}(0,s)<-b$ in $[0,T]\times[k,\infty)$, thus satisfying condition (ii) of Theorem \ref{compactsupport}. As a result, Theorem \ref{compactsupport} applies and gives the boundedness of the delay region for a call.
\end{proof}
Since a stock can be viewed as a call with strike $K=0$, Proposition \ref{Call_OU} also applies to the optimal liquidation of a stock over a finite time horizon. Also, we notice the delay region can be empty, and we can identify this case by finding the maximum of the drive function. As an example, we consider the case of the stock with penalty function $\psi((m-S_t)^+)=(m-S_t)^+$, and we obtain the maximizer of $G^\alpha$ in different scenarios
\begin{equation*}
\arg\max G^\alpha = \begin{cases}
\exp(\theta-1-\frac{r-\alpha}{\beta}) & \text { if } \exp(\theta-1-\frac{r-\alpha}{\beta}) < m,\\
\exp(\theta-1-\frac{r}{\beta}) & \text { if } \exp(\theta-1-\frac{r}{\beta}) > m,\\
m & \text { otherwise},
\end{cases}
\end{equation*}
and the corresponding maximum values
\begin{equation*}
\max G^\alpha = \begin{cases}
(\beta-\alpha)\hat{s}_1-\alpha(m-s^*_1) & \text { if } \hat{s}_1 < m,\\
\beta \hat{s}_2 & \text { if } \hat{s}_2 > m,\\
(\beta(\theta-\log(m))-r)m & \text { otherwise,}
\end{cases}
\end{equation*}
where
\begin{equation*}
\hat{s}_1 = \exp(\theta-1-\frac{r-\alpha}{\beta}), \qquad \hat{s}_2 = \exp(\theta-1-\frac{r}{\beta}).
\end{equation*}
Thus, the delay region is non-empty if and only if $\max G^\alpha>0$.
The optimal liquidation boundary for stock is shown in Figure \ref{Stock_OU} for $\alpha=0$ (left panel) and $\alpha>0$ (right panel). We notice that, in both cases, the optimal strategy is to sell immediately if $S_t$ is high enough. Intuitively, if $S_t$ is high, it is expected to revert back to its long-term mean, so selling immediately becomes optimal. However, if $S_t$ is low, the optimal behavior depends on the parameter $\alpha$. On one hand, $S_t$ is expected to increase and thus the investor should wait to sell at a better price (Figure \ref{Stock_OU}, right panel). On the other hand, such benefit is countered (if the penalization coefficient is high enough) by the risk incurred from holding the position, and this induces the investor to sell immediately. As a consequence, the sell region is disconnected (Figure \ref{Stock_OU}, right panel).
\begin{figure}[ht]
\centering
\includegraphics[scale=0.55]{Stock_OU_Neumann_SF_1_1.eps}\qquad ~~
\includegraphics[scale=0.55]{Stock_OU_Neumann_SF_1_2.eps}
\caption{\small{The liquidation boundary (solid) and the zero contour of $G^{\alpha}$ (dashed) for a stock under exponential OU dynamics. Parameters: $T=0.5$, $r=0.03$, $\theta=\log(60)$, $\beta=4$, $\sigma=0.3$, $\psi(\ell)=\ell$, $\alpha=0$ (left), $\alpha=1.5$ (right). }}
\label{Stock_OU}
\end{figure}
Figure \ref{Call_OU} illustrates the delay region for a call option with penalty.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.55]{Call_OU_Neumann_SF_1_1.eps}\qquad ~~
\includegraphics[scale=0.55]{Call_OU_Neumann_SF_1_2.eps}
\caption{\small{The liquidation boundary (red solid) and the zero contour of $G^{\alpha}$ (dashed) for a call under exponential OU dynamics. We take $\alpha=0.2$, $\theta=\log(60)$, $\beta=4$ in the left panel, and $\alpha=0.001$, $\theta=\log(50)$ and $\beta=0.2$ in the right panel, with common parameters $T=0.5$, $r=0.03$, $\sigma=0.3$, $\psi(\ell)=\ell$.}}
\label{Call_OU}
\end{figure}
In the right panel, we observe the interesting phenomena where the sell region is connected and contains the nonempty delay region. If the parameter $\beta$ (which measures the speed of mean reversion) is not sufficiently high, there may be no time for the price of the option to revert back to its long-term mean before expiration, so that selling immediately becomes optimal close to maturity.
\begin{proposition} \label{Put_OU_prop}
For the liquidation of a put option under the exponential OU model, the delay region is bounded if and only if $\alpha>0$.
\end{proposition}
\begin{proof}
The drive function is given by
\begin{equation}
G^{\alpha}_{Put}(t,s)=[r-\beta(\theta-\log(s))]s\Phi(-d_1)-\alpha\psi((m-P(t,s))^+),
\end{equation}
If $\alpha=0$, then we have $\{G^{\alpha}_{Put}>0\}=\{s>\exp(\frac{r}{\beta}-\theta)\}$. By \eqref{GsubL}, the delay region contains this set, so it is unbounded.
Now let $\alpha>0$, and we have the limit
\begin{equation}
\lim_{s\to \infty} G^\alpha_{Put}(t,s)=-\alpha\psi(m) < 0. \label{OUPut2}
\end{equation}
Next, we fix any $\hat{b}\in (0,\alpha \psi (m))$ and define $\overline{\psi}(\ell):=\min\{\psi(\ell),\hat{b}\}$. With this, we have
\begin{equation*}
\overline{G}^\alpha(t,s):=[r-\beta(\theta-\log(s))]s\Phi(-d_1) - \alpha \overline{\psi}\left((m-P(t,s))^+\right) \geq G^\alpha_{Put}(t,s).
\end{equation*}
We observe that $\overline{G}^\alpha$ is bounded above and by \eqref{OUPut2} $\lim_{s\to\infty}\overline{G}^\alpha(t,s)\to -\alpha \hat{b} <0$ for every $t\in[0,T]$. Moreover, there exists $\hat{s}>0$ such that for every $s>\hat{s}$, $\overline{\psi}((m-P(t,s))^+)=\hat{b}$. As a result, we have
\begin{equation*}
\frac{\partial \overline{G}^\alpha}{\partial t}
= (\beta(\theta-\log(s))-r)s\phi(d_1)\frac{\log(\frac{s}{K}-(r+\frac{\sigma^2}{2})(T-t)}{2\sigma(T-t)^\frac{3}{2}}\leq 0,
\end{equation*} for $s>\max\{\hat{s}, \exp(\frac{r}{\beta}-\theta), K\exp((r+\sigma^2/2)T)\}$ and $t\in[0,T]$. Also, we notice that $\overline{G}^\alpha (0,s)\to -\alpha \hat{b}$ as $s\to-\infty$. This allows us to choose a $b\in(0,\alpha \hat{b})$, then there exists $k>\max\{\hat{s}, K\exp((r+\sigma^2/2)T)\}$ such that $-b>\overline{G}(0,s)>\overline{G}(t,s)$ in for $(t,s) \in [0,T]\times [k,\infty)$. Therefore, $\overline{G}$ satisfies the assumptions of Theorem \ref{compactsupport}. By Corollary \ref{Compare_G}, we conclude the boundedness of the delay region.
\end{proof}
Proposition \ref{Put_OU_prop} is illustrated in Figure \ref{Put_OU}. When $S_t$ is low, it is expected to revert back to the (higher) long-term mean, and the put price will decrease. This generates an incentive to sell at a low stock price level. If $\alpha=0$, when $S_t$ is high, there is no reason to sell since the put price is very low and expected to increase. Consequently, the delay region is on top of the sell region (Figure \ref{Put_OU}, left panel).
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.55]{Put_OU_Neumann_SF_1_1.eps}\qquad ~~
\includegraphics[scale=0.55]{Put_OU_Neumann_SF_1_2.eps}
\caption{\small{The liquidation boundary (solid) and the zero contour of $G^{\alpha}$ (dashed) for a put option under the exponential OU model. We take $\alpha=0$ and $K=50$ in the left panel, and $\alpha=0.01$ and $K=40$ in the right panel. Common parameters: $T=0.5$, $r=0.03$, $\sigma=0.3$, $\beta=4$ and $\theta=\log(60)$, $\psi(\ell)=\ell$. }}
\label{Put_OU}
\end{figure}
However, this is no longer true when we incorporate a non-zero risk penalty which reduces the value of waiting. As a result, the holder may sell the put at high and low stock prices. In fact, if the penalization coefficient is large and/or when the time-to-maturity is very short, the optimal liquidation premium may be zero at all stock price levels, resulting in an empty delay region (Figure \ref{Put_OU}, right panel).
\section{Quadratic Penalty} \label{sect:quadr}
As a variation to the shortfall-based penalty, we consider a risk penalty based on the realized variance of the option price process from the starting time up to the liquidation time. Precisely, the investor now faces the penalized optimal stopping problem
\begin{align*}
\tilde{J}^{\alpha}(t,s)&:= \sup_{\tau \in \mathcal{T}_{t,T}}\E_{t,s}\left\{e^{-r(\tau-t)}V(\tau,S_\tau)
-\alpha \int_t^{\tau} e^{-r(u-t)}d[V,V]_u\right\} \\
&= \sup_{\tau \in \mathcal{T}_{t,T}}\E_{t,s}\left\{e^{-r(\tau-t)}V(\tau,S_\tau)
-\alpha \int_t^{\tau} e^{-r(u-t)} \sigma^2(u,S_u) S_u^2 V_s^2(u,S_u)du\right\},
\end{align*}
where $[V,V]$ denotes the quadratic variation of option price process $V$ defined in \eqref{mkt_pr}. Figure \ref{quadratic} illustrates the realized quadratic penalty associated with a simulated call option price path. Compared to the shortfall penalty in Figure \ref{shortfall}, the realized quadratic penalty is increasing at all times, even when the option price is above its initial price.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.65]{quadratic.eps}
\caption{\small{Realized quadratic penalty (dashed) based on a simulated price path (solid) of a call under the GBM model with $\alpha =0.05$. The price path and other parameters are the same as Figure \ref{realized_shortfall}. }}
\label{quadratic}
\end{figure}
Following \eqref{delayed_liquidation}, we define the optimal liquidation premium by $\tilde{L}^{\alpha}(t,s) := \tilde{J}^{\alpha}(t,s)-V(t,s)$. Again, we shall discuss the stock or option liquidation problems under the GBM and exponential OU models.
\subsection{Optimal Timing to Sell a Stock }\label{AnalyticalSolution}
We first consider the liquidation of a stock with the GBM dynamics in terms of the perpetual optimal stopping problem:
\begin{equation} \label{prob_repr_inf}
\tilde{L}^\alpha(s):=\sup_{\tau \in \mathcal{T}} \E_s\left\{ \int_0^\tau e^{-ru} \tilde{G}^{\alpha}(S_u) \,du \right\},
\end{equation}
with the drive function $\tilde{G}^\alpha(s):=(\mu-r)s-\alpha\sigma^2s^2$. If $\mu\leq r$, then selling immediately is always optimal since $\tilde{G}^\alpha$ is always negative. In contrast if $\mu>r$, then we obtain a non-trivial closed-form solution.
\begin{theorem}\label{perpL}
Let $\mu>r$. The value function $\tilde{L}^\alpha(s)$ in \eqref{prob_repr_inf} is given by the formula
\begin{equation}\label{PerpGBM_solution}
\tilde{L}^\alpha(s)=\left\{\frac{\left(s^*\right)^{1-\lambda}}{2-\lambda}\,s^{\lambda}-s+B\,s^{2}\right\}\indic{s \,\leq \,s^*},
\end{equation}
where
\begin{align}
B & = \frac{\alpha\sigma^2}{2\mu+\sigma^2-r}, \quad \lambda = \frac{1}{\sigma ^{2}}\left[\, \frac{\sigma ^2}{2} - \mu + \sqrt{\left(\frac{\sigma ^{2}}{2}-\mu \right)^{2}+2r\sigma^{2}}\,\right],\label{parameterB} \\
s^{\ast} & = \frac{1-\lambda}{( 2-\lambda)B}, \label{threshold}
\end{align}
and the stopping time $\tau^*=\inf\{t\geq0:S_t \ge s^*\}$ is optimal for \eqref{prob_repr_inf}.
\end{theorem}
\begin{proof}
We first show that \eqref{PerpGBM_solution} is the solution of
\begin{equation*}
\min \left\{r\Lambda(s)-\mu s\Lambda'(s)-\frac{\sigma ^2 s^2}{2}\Lambda''(s)-\tilde{G}^{\alpha}(s),\,\Lambda(s)\right\}=0, \quad s>0,
\end{equation*}
with $\Lambda(0)=0$. To do this, we split $\R^+$ into two regions: $\mathcal{D}_1=(0,s^*)$ and $\mathcal{D}_2=[s^*,\infty)$ with $s^*>0$ to be determined. We conjecture that $\Lambda(s)=0$ in $\mathcal{D}_2$, and for $s\in \mathcal{D}_1$ $\Lambda(s)$ solves
\begin{equation}\label{PerpGBM_PDE}
r\Lambda(s)-\mu s\Lambda'(s)-\frac{\sigma^2s^2}{2}\Lambda''(s)-\tilde{G}^\alpha(s)=0.
\end{equation}
By direct substitution, the general solution to equation \eqref{PerpGBM_PDE} is of the form
\begin{equation*}
\Lambda(s)=C_{1}s^{\lambda_{1}}+C_{2}s^{\lambda _{2}}-s+Bs^2,
\end{equation*}
where $C_{1}$ and $C_{2}$ are constants to be determined, $B$ is specified in \eqref{parameterB} and
\begin{equation*}
\lambda_{k} = \frac{1}{\sigma ^{2}}\left[\, \frac{\sigma ^2}{2} - \mu +(-1)^k\sqrt{\left(\mu -\frac{\sigma ^{2}}{2}\right)^{2}+2r\sigma^{2}}\,\right], \quad k \in \{1,2\}.
\end{equation*}
We apply the continuity and smooth pasting conditions at $s=0$ and $s=s^*$ to get
\begin{align}
&\lim_{s\downarrow 0}\Lambda(s)=0 ~\Rightarrow ~C_1=0,\notag\\
&\lim_{s\uparrow s^*}\Lambda(s)=0 ~\Rightarrow~ C_2(s^*)^{\lambda_2}-s^*+B(s^*)^2=0,\label{Lcont}\\
&\lim_{s\uparrow s^*}\Lambda'(s)=0~ \Rightarrow~ \lambda_2C_2(s^*)^{\lambda_2-1}-1+2Bs^*=0. \label{smoothL}
\end{align}
Solving the system of equations \eqref{Lcont}--\eqref{smoothL} gives $C_2$ and $s^*$ as in \eqref{parameterB}-\eqref{threshold}. One can verify by substitution that $\Lambda(s)$ is indeed a classical solution of \eqref{PerpGBM_PDE}.
By Ito's formula and \eqref{PerpGBM_PDE}, $(\Lambda(S_t))_{t\ge 0}$ is a $(\P, \Fil)$-supermartingale, so for every $\Fil$-stopping time $\tau$ and $n\in\N$, we have
\begin{equation} \label{F-K}
\Lambda(s)\geq \E_{0,s} \left\{\int_0^{\tau\wedge n} e^{-ru}\tilde{G}^\alpha(S_u)du\right\}.
\end{equation}
Maximizing \eqref{F-K} over $\tau$ and $n$ yields that $\Lambda(s)\geq \tilde{L}^\alpha(s)$ for $s \ge 0$. The reverse inequality is deduced from the probabilistic representation $\Lambda(s) = \E_{0,s} \left\{\int_0^{\tau^*} e^{-ru}\tilde{G}^\alpha(S_u)du\right\}$, with the candidate stopping time $\tau^*:=\inf\{t\geq0\,:\,S_t \ge s^*\}$. Hence, we conclude that $\Lambda(s)=\tilde{L}^\alpha(s)$ and $\tau^*$ is optimal. \end{proof}
The optimal liquidation threshold $s^{\ast}~$ in \eqref{threshold} is non-negative if and only if $\lambda_{2}<1$, which is equivalent to the condition $\mu>r$ in Theorem \ref{perpL}. Otherwise, $\tilde{L}^\alpha(s)=0$ and the optimal strategy is to sell immediately.
In Figure \ref{PerpGBM_Fig1}, we illustrate the optimal liquidation premium $\tilde{L}^\alpha(s)$ for various values of $\mu$ and $\sigma$. As $\mu$ increases, the optimal threshold as well as the optimal liquidation premium (at all stock price levels) increase (left panel). On the other hand, a higher volatility reduces the optimal liquidation premium at every initial stock price. We also observe that $\tilde{L}^\alpha(s)$ smooth-pastes the level 0 at the optimal threshold $s^*$, as is expected from \eqref{Lcont} and \eqref{smoothL}.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.55]{Stock_Quadr_GBM_mu_010.eps}\qquad ~~
\includegraphics[scale=0.55]{Stock_Quadr_GBM_sigma_010.eps}
\end{center}
\caption{\small{The optimal liquidation premium for a stock under the GBM model for different values of $\mu$ and $\sigma$. In the left panel, we take $r=0.03$, $\sigma=0.3$ and $\alpha=0.2$, and the liquidation threshold $s^* = 9.37, 7.97, 6.52$ for $\mu = 0.09, 0.08, 0.07$ respectively. In the right panel, we take $r=0.03$, $\mu=0.08$, and $\alpha=0.1$, and the liquidation threshold $s^* = 10.63, 7.97, 6.26$ for $\sigma=0.25, 0.30, 0.35$. }}
\label{PerpGBM_Fig1}
\end{figure}
If $S$ follows the exponential OU dynamics, the drive function for liquidating a stock is
\begin{equation}\label{G_PerpStock_OU}
\tilde{G}^\alpha(s)=[\beta(\theta-\log(s))-r-\alpha s]s.
\end{equation}
In this case, we do not have a closed-form solution. Nevertheless we observe from \eqref{G_PerpStock_OU} that the delay region is non-empty, namely, $\{\tilde{L}^\alpha>0\}\supseteq\{s<\tilde{s}\}$, where $\tilde{s}$ is determined uniquely from the equation
\begin{equation}\notag
\beta(\theta-\log(\tilde{s}))-r-\alpha \tilde{s} = 0.
\end{equation}
On the other hand, since $\tilde{G}^\alpha\to-\infty$ as $s\to\infty$, we expect intuitively that the investor will sell when the stock price is high.
\subsection{Liquidation of Options}
We now discuss some numerical examples to demonstrate the liquidation strategies for European call and put options. With strike $K$ and maturity $T$, the drive functions are respectively given by
\begin{align}
\tilde{G}^{\alpha}_{Call}(t,s) & = s\Phi(d_1)\big(\mu-r-\alpha \sigma^2 s \Phi(d_1)\big), \label{call_quadr_gbm} \\
\tilde{G}^{\alpha}_{Put}(t,s) & = s\Phi(-d_1)\big(r-\mu-\alpha \sigma^2 s\Phi(-d_1)\big). \label{put_quadr_gbm}
\end{align}
When $\mu\leq r$ and $\alpha>0$, the drive function $\tilde{G}^{\alpha}_{Call}(t,s)$ is negative for all $(t,s)$, so it is optimal to sell the call immediately. However, when $\mu>r$ and $\alpha>0$, we notice from \eqref{call_quadr_gbm} that, when the stock price is sufficiently large (resp. small), the drive function of a call is negative (resp. positive). Hence, as we see in Figure \ref{fig_Quadr_GBM}, it is optimal to sell the call when the stock price is high, and the optimal liquidation boundary is lower as the penalization coefficient increases. In contrast to the shortfall penalty, the investor now is subject to a higher penalty when the stock price is high under the quadratic penalty. Consequently, the sell region is now above the delay region, as opposed to being at the bottom in the shortfall case in Figure \ref{Stock_Call_SF_1} (right panel).
In the put option case, we observe from \eqref{put_quadr_gbm} that
\begin{equation*}
\lim_{s\rightarrow 0}r-\mu-\alpha \sigma^2 s\Phi(-d_1)=\lim_{s\rightarrow \infty}r-\mu-\alpha \sigma^2 s\Phi(-d_1)=r-\mu.
\end{equation*}
Consequently, when $\mu<r$ and the stock price is sufficiently large or small, the drive function is strictly positive and it is optimal to hold the position. In contrast, the shortfall converges to $\psi(m)>0$ as $s$ increases (see \eqref{Put2}), which means that it is optimal to sell when the stock price is high (see Figure \ref{Put_SF}). We illustrate the timing strategies under quadratic penalty in Figure \ref{fig_Quadr_GBM} (right). As expected there is a low and a high delay regions which are separated by a sell region in the middle. Also we notice that as the penalization coefficient $\alpha$ increases, the sell region expands.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.55]{call_ext.eps}
\includegraphics[scale=0.55]{put_ext.eps}
\end{center}
\caption{\small{The liquidation boundaries for a call option (left panel) and a put option (right panel) under the GBM mode with different values of $\alpha$. Parameters: $T=0.5$, $r=0.03$, $\sigma=0.3$, $K=50$, $\mu=0.08$ (call) and $\mu=0.02$ (put). }}
\label{fig_Quadr_GBM}
\end{figure}
Under the exponential OU model, the drive functions for selling a call and a put are, respectively,
\begin{align*}
\tilde{G}_{Call}^\alpha(t,s)
&= s\Phi(d_1) \big( \theta - r - \beta \log s - \alpha \sigma^2 s \Phi(d_1) \big) ,\\
\tilde{G}^{\alpha}_{Put}(t,s)
&= s\Phi(-d_1)\bigg(r-\theta + \beta \log s-\alpha \sigma^2 s\Phi(-d_1)\bigg).
\end{align*}
In Figure \ref{fig_Quadr_OU}, we can visualize the optimal liquidation premium $\tilde{L}^\alpha(t,s)$ for a call (right panel) and a put (left panel). In the call case, the delay region, which corresponds to the area where $\tilde{L}^\alpha>0$, is bounded. When $s$ is sufficiently high, $\tilde{L}^\alpha$ vanishes and it is optimal to sell. This is intuitive since $\lim_{s\to\infty}\tilde{G}_{Call}^\alpha(t,s)=-\infty$ and $\tilde{G}_{Call}^\alpha$ is positive for sufficiently small $s$.
In contrast, the drive function for the put $\tilde{G}^{\alpha}_{Put}(t,s)$ is negative when $s < \exp\left(\frac{\theta - r}{\beta}\right)$, for every $\alpha \ge 0$. Therefore, as Figure \ref{fig_Quadr_OU} indicates, one expects the optimal liquidation premium to vanish for small $s$, so the investor will sell when the put price is high.
Compared to Figure \ref{Put_OU} with a shortfall penalty, the investor does not sell when the underlying stock price is very high. This is because the drive function $\tilde{G}^{\alpha}_{Put}(t,s)$ stays positive for large $s$ (recall \eqref{GsubL}). As time approaches maturity, the delay liquidation premium decreases to its terminal condition of value zero.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.55]{OLP_Call_Quadr_Neumann.eps}\qquad ~~
\includegraphics[scale=0.55]{OLP_Put_Quadr_Neumann.eps}
\end{center}
\caption{\small{The optimal liquidation premium for a call option (left) and a put option (right) with exponential OU dynamics. We take $T=0.5$, $r=0.03$, $\sigma=0.3$, $K=50$, $\alpha=0.1$, $\beta=4$ and $\theta=\log(60)$. }}
\label{fig_Quadr_OU}
\end{figure}
\section{Conluding Remarks} \label{sect:concl}
In summary, we have provided a flexible mathematical model for the optimal liquidation of option positions under a path-dependent penalty. We have identified the situations where the optimal timing is trivial, and solved for non-trivial liquidation strategy via variational inequality. The penalty type as well as the penalization coefficient can give rise to very different liquidation timing. Our findings are useful for both individual and institutional investors who use options for speculative investments or risk management purposes.
For future research, a natural direction is to adapt our model to the problem of sequentially buying and selling an option. Moreover, one can consider applying the methodology to derivatives other than equity options. For example, we refer to \cite{LeungLiu2012} for a recent study on the liquidation of credit derivatives with pricing measure discrepancy but without risk penalty. It would be both mathematically interesting and challenging to study option liquidation under incomplete markets. On the other hand, our model can be extended to markets with liquidity cost and price impact (see e.g. \cite{Almgren2003,Lorenz11,Schied2009}). Finally, the path-dependent risk penalization can also be incorporated to dynamic portfolio optimization problems to account for adverse performance during the investment horizon.
\section{Strong Solution to the Inhomogeneous Variational Inequality}\label{Sect-ExUn}
In this section, we follow the terminology and procedures in \cite{Bensoussan78}, and establish the existence and uniqueness of a strong solution to the variational inequality \eqref{L_general_VI} under conditions that are applicable to the GBM and exponential OU models.
\noindent \textbf{Preliminaries.} We express prices in logarithmic scale by setting $X_t=\log(S_t)$. Equation \eqref{SDE_S} then becomes
\begin{equation} \label{SDE_X}
dX_t = \eta(t,X_t)dt + \kappa(t,X_t)dW_t,
\end{equation}
for some functions $\kappa(t,x)$ and $\eta(t,x)$. Next, we define the operator $\mathcal{A}$ by
\begin{align}
\mathcal{A}[\cdot]
&= -\frac{\kappa^2(t,x)}{2} \frac{\partial^2\, \cdot}{\partial x^2}-\eta(t,x) \frac{\partial\, \cdot}{\partial x} + r \,\cdot \notag \\
&= -\frac{\partial\,\cdot}{\partial x}\left(a_2(t,x)\frac{\partial \,\cdot}{\partial x}\right) + a_1(t,x)\frac{\partial \,\cdot}{\partial x} + r\,\cdot \,, \label{divergenge_form}
\end{align}
where
\begin{equation*}
a_1(t,x) = \frac{1}{2}\frac{\partial }{\partial x}\kappa^2(t,x)-\eta(t,x), \qquad
a_2(t,x) = \frac{\kappa^2(t,x)}{2}.
\end{equation*} In term of log-prices, we express the drive function as $g(t,x)=G^\alpha(t,e^x)$ and the optimal liquidation premium as $u(t,x) = L(t,e^x)$. Throughout, we denote the domain $\mathcal{D}=[0,T]\times \R$. In order to solve the VI \eqref{L_general_VI}, it is equivalent to solve the VI problem:
\begin{equation} \label{u_general_VI}
\begin{cases}
-\frac{\partial u}{\partial t} + \mathcal{A}[u] - g(t,x) \geq 0, \,u(t,x)\geq0, \quad (t,x) \in \mathcal{D}, \\
\left(-\frac{\partial u}{\partial t} + \mathcal{A}[u] - g(t,x)\right) u = 0, \quad (t,x) \in \mathcal{D}, \\
u(T,x)=0, \quad x \in \R.
\end{cases}
\end{equation}
We describe an appropriate class of solutions for \eqref{u_general_VI} in a suitable Sobolev space and prove that such a solution exists and is unique. First, let us define, for $\lambda(x)=\exp(-n|x|)$, $n\in \mathbb{N}$,
\begin{align*}
\mathcal{L}^2_{\lambda}(\R) & = \{v\,|\,\sqrt{\lambda} v \in L^2(\R)\}, \\
\mathcal{H}^1_{\lambda}(\R) & = \{v\in L^2_{\lambda}(\R)\,|\,\frac{\partial v}{\partial x} \in L^2_{\lambda}(\R)\},\\
\mathcal{H}^1_{0,\lambda}(\R) & = \{v\in H^1_{\lambda}(\R)\,|\,\lim_{|x|\to\infty}v(x)=0\}.
\end{align*} These are Hilbert spaces when endowed with the following inner products
\begin{align*}
(f,g)_{L^2} & = \int_{\R}\lambda fgdx, \,\quad f,g \in L^2_{\lambda}(\R),\\
(f,g)_{H^1} & = \int_{\R}\lambda fgdx + \int_{\R}\lambda \frac{\partial f}{\partial x}\frac{\partial g}{\partial x}dx,\quad f,g \in H^1_{\lambda}(\R).
\end{align*}
We denote by $\mathcal{H}^1_{c,\lambda}(\R)$ the set of functions $w \in \mathcal{H}^1_{\lambda}(\R)$ with compact support. For $u \in \mathcal{H}^1_{0,\lambda}(\R)$, $w \in \mathcal{H}^1_{c,\lambda}(\R)$, we define the operator
\begin{equation*}
\mathcal{I}_{\lambda}(t,u,w) = \int_{\R} a_2(t,x)\left(\lambda\frac{\partial u}{\partial x}\frac{\partial w}{\partial x}
+ w\frac{\partial u}{\partial x}\frac{\partial \lambda}{\partial x}\right)dx
+ \int_{\R} a_1(t,x)\lambda\frac{\partial u}{\partial x}w dx +r\int_{\R}\lambda uw dx.
\end{equation*}
We can assume without loss of generality \citep[Sect. 3.2.17]{Bensoussan78} that $\mathcal{I}_{\lambda}$ is coercive on $\mathcal{H}^1_{c,\lambda}(\R)$, i.e.
\begin{equation*}
\mathcal{I}_{\lambda}(t,w,w)\geq\alpha |\!|w|\!|_{H^1} \quad \forall \, w \in \mathcal{H}^1_{c,\lambda}(\R), \, \alpha>0.
\end{equation*}
Integrating by parts allows us to extend $\mathcal{I}_{\lambda}$ to a bilinear form on the whole space $\mathcal{H}^1_{0,\lambda}(\R)$. In particular, we set
\begin{equation*}
\mathcal{I}_{\lambda}(t,u,v) = \int_{\R}\left[a_2(t,x)\lambda\frac{\partial u}{\partial x}\frac{\partial v}{\partial x}
+ a_2(t,x)\frac{\partial \lambda}{\partial x}\frac{\partial u}{\partial x}v\right]dx
+ \int_{\R} \left(r - \frac{1}{2}\frac{\partial a_1}{\partial x}
- \frac{1}{2\lambda}a_1\frac{\partial \lambda}{\partial x} \right) \lambda uvdx.
\end{equation*}
with $u,v \in \mathcal{H}^1_{0,\lambda}(\R)$.
Following Chapter 5.9.2 of \cite{Evans1998} and Chapter 2.6 of \cite{Bensoussan78}, we define the space $\mathcal{L}^p(0,T;X)$ consisting of all strongly measurable functions $\chi:[0,T]\to X$ with
\begin{equation*}
|\!|\chi|\!|_{\mathcal{L}^p(0,T;X)}=\left(\int_0^T|\!|\chi(t)|\!|^p_X\,dt\right)^{1/p}, \quad 1\leq p<\infty,
\end{equation*}
and for $p=\infty$,
\begin{equation*}
|\!|\chi|\!|_{\mathcal{L}^{\infty}(0,T;X)}=\esssup_{0\leq t \leq T}|\!|\chi(t)|\!|_X,
\end{equation*}
For $\chi\in \mathcal{L}^1(0,T;X)$, we say $\nu\in \mathcal{L}^1(0,T;X)$ is the weak derivative of $\chi$, denoted by $\nu= \frac{\partial \chi}{\partial t}$, if
\begin{equation*}
\int_0^T\frac{\partial w}{\partial t}\chi(t) dt = -\int_0^T w(t)\nu(t)dt, \quad \forall \,w \in C_c^\infty([0,T]).
\end{equation*}
The Sobolev space $\mathcal{H}^1(0,T;X)$ consists of all functions $\chi \in \mathcal{L}^2(0,T;X)$ such that the weak derivative exists and belongs to $\mathcal{L}^2(0,T;X)$. Furthermore, we set
\begin{equation} \label{H1norm}
|\!|\chi|\!|_{\mathcal{H}^1(0,T;X)}=\left(\int_0^T|\!|\chi(t)|\!|^2_X+|\!|\frac{\partial }{\partial t}\chi(t)|\!|^2_X\,dt\right)^{1/2},
\end{equation}
which makes $\mathcal{H}^1(0,T;X)$ an Hilbert space (see Chapter 5.9.2 in \cite{Evans1998}).
\vspace{5 mm}
\noindent \textbf{Main Results.}
\begin{definition}
A function $u:\mathcal{D}\to\R$ is a \emph{strong solution} of problem \eqref{u_general_VI} if, $\forall$ $v \in \mathcal{H}^1_{\lambda}(\R)$, $v\geq0$ a.e., the following conditions are satisfied:
\begin{equation}\label{L_general_VI_strong}
\begin{cases}
& u\in \mathcal{L}^2(0,T;\mathcal{H}^1_{0,\lambda}(\R)), \, \frac{\partial u}{\partial t} \in \mathcal{L}^2(0,T;\mathcal{L}^2_{\lambda}(\R)), \\
& -\left(\frac{\partial u}{\partial t},v-u\right)-\mathcal{I}_{\lambda}(t;u,v-u)\leq(g,v-u), \\
& u\geq0 \text{ a.e. in } \mathcal{D}, \\
& u(T,x)=0, \, x\in\R.
\end{cases}
\end{equation}
\end{definition}
We shall impose the following conditions on $a_2$, $a_1$, $g$.\\
\textbf{Assumption A}. $a_2$, $\frac{\partial a_2}{\partial t}$ and $\frac{\partial a_1}{\partial x}\in\mathcal{L}^{\infty}(\mathcal{D})$; $a_1$ and $\frac{\partial a_1}{\partial t} \in C^0(\overline{\mathcal{D}})$; $g \in \mathcal{H}^1(0,T;\mathcal{L}^2_{\lambda}(\R))$.
\begin{theorem}\label{ExistenceStrongSolution}
Under Assumption {A}, the variational inequality in \eqref{L_general_VI_strong} has a unique strong solution.
\end{theorem}
\begin{proof}
Assumption A is equivalent to assumptions (2.223), (2.224), (2.238), (2.239), (2.240) of \citep[Chap. 3]{Bensoussan78}, and we also follow their Remark 2.24 to use $\lambda(x)=e^{-n|x|}$ for some arbitrarily fixed $n>0$ in our definition of Hilbert spaces. In turn, we can apply their Theorem 2.21 and our statement follows.
\end{proof}
Our main objective is to verify that Assumption {A} is satisfied for our applications so that Theorem \ref{ExistenceStrongSolution} applies to ensure the existence of a unique strong solution to the VI \eqref{L_general_VI}.
To see this, we first write down the operators associated with the log-price $X_t=\log(S_t)$ under the GBM and exponential OU models, namely,
\begin{equation} \notag
\mathcal{A}[v]=\frac{\sigma^2}{2}\frac{\partial^2 v}{\partial x^2}+\mu\frac{\partial v}{\partial x}, \qquad
\mathcal{A}[v]=\frac{\sigma^2}{2}\frac{\partial^2 v}{\partial x^2}+(\hat{\theta}-\beta x)\frac{\partial v}{\partial x}.
\end{equation} Therefore, $a_2 = \sigma^2/2$ is constant and $a_1$ is an affine function in $x$ for both cases, so these coefficients meet the requirements in Assumption A.
It remains to verify that the drive function $g(t,x) =G(t,e^x) \in \mathcal{H}^1(0,T;\mathcal{L}^2_{\lambda}(\R))$. In view of \eqref{H1norm}, we want to show that there exists $n>0$ such that
\begin{align*}
\int_0^T|\!|g(t,x)|\!|_{\mathcal{L}^2_{\lambda}(\R)}^2dt
&=\int_0^T\int_{\R}\left(g(t,x)e^{-\frac{n}{2}|x|}\right)^2dx\,dt, \quad \text{and}\\
\int_0^T|\!|\frac{\partial g}{\partial t}(t,x)|\!|_{\mathcal{L}^2_{\lambda}(\R)}^2dt
&=\int_0^T\int_{\R}\left(\frac{\partial g}{\partial t}(t,x)e^{-\frac{n}{2}|x|}\right)^2dx\,dt \\
\end{align*}
are finite, where
\begin{align*}
g(t,x)&= (r-\mu(t,e^x))e^xV_s(t,e^x)-\alpha\psi((m-V(t,e^x))^+),\\
\frac{\partial g}{\partial t}(t,x)&=(r-\mu(t,e^x))e^xV_{ts}(t,e^x)+\alpha\psi'((m-V(t,e^x))^+)V_t(t,e^x)\indic{m>V(t,e^x)}.
\end{align*} Here, the subscripts of $V$ indicate the partial derivatives in $t$ and $s$. Recall the drift functions $\mu(t,e^x)=\mu$ under GBM and $\mu(t,e^x)=\beta(\theta-x)$ under exponential OU models. We notice that, in both cases, the drift does not depend on $t$, so we just write $\mu(e^x)$. Also, we observe that $\psi$ and $\psi'$ are increasing, and $\psi'(\ell)$ is bounded for any finite $\ell$. For both call and put options, there exist positive constants $h_1,\,q_1,\,h_2,\,q_2$ such that $|V_s(t,e^x)|\leq 1, |V_t(t,e^x)|\leq h_1e^x+q_1,\,|V_{st}(t,e^x)|\leq h_2e^x+q_2$. Together, these imply the time-independent bounds for both models:
\begin{align*}
|g(t,x)|&\leq |r-\mu(e^x)|e^x+\alpha\psi(m)=\emph{o}({e^{2|x|}}), \\
|\frac{\partial g}{\partial t}(t,x)|&\leq |r-\mu(e^x)|(h_1e^x+q_1)e^x+\alpha\psi'(m)(h_2e^x+q_2)=\emph{o}(e^{2|x|}).
\end{align*}
This implies that by choosing $n>4$, we have
\begin{align*}
\int_0^T|\!|g(t,x)|\!|_{\mathcal{L}^2_{\lambda}(\R)}^2dt &\leq \int_0^T\big|\!\big| |r-\mu(e^x)|e^x+\alpha\psi(m)\big|\!\big|_{\mathcal{L}^2_{\lambda}(\R)}^2dt<\infty,\\
\int_0^T|\!|\frac{\partial g}{\partial t}(t,x)|\!|_{\mathcal{L}^2_{\lambda}(\R)}^2dt&\leq \int_0^T\big|\!\big| |r-\mu(e^x)|(h_1e^x+q_1)e^x+\alpha\psi'(m)(h_2e^x+q_2)\big|\!\big|_{\mathcal{L}^2_{\lambda}(\R)}^2dt<\infty.
\end{align*}
Hence, we conclude that $g\in \mathcal{H}^1(0,T;\mathcal{L}^2_{\lambda}(\R))$ for both puts and calls under the GBM and exponential OU models, and Assumption A is satisfied.
\vspace{5 mm}
As a final remark, Sect. 3.4 of \cite{Bensoussan78} also provides the probabilistic representation of the strong solution $u(t,x)$ of the VI \eqref{u_general_VI}, given by \begin{equation} \label{log_prices}
u(t,x)=\sup_{\tau \in \mathcal{T}_{t,T}} \E_{t,x} \left\{ \int_t^\tau e^{-r(u-t)} g(u,X_u) \,du\right\},
\end{equation}
where $dX_u=\eta(u,X_u)du+\kappa(u,X_u)dW_u$ and $X_t=x$. By the definition $L(t,e^x) = u(t,x)$, the optimal stopping problem in \eqref{log_prices} resembles that for the optimal liquidation premium in \eqref{L_G}.
|
\section{Introduction}
\label{Introduction}
The complex networks is a model which can used to describe those complex relationship in the real system, such as the biological, social and technological systems \cite{albert2000error,newman2003structure}. Many property of the complex networks have illuminated by these researchers in this filed, such as the network topology and dynamics \cite{watts1998collective,newman2006structure}, the property of the network structure \cite{newman2003structure,barthelemy2004betweenness}, the self-similarity and fractal property of the complex networks\cite{song2005self,wei2014informationdimension}, the controllability and the synchronization of the complex networks \cite{liu2011controllability,arenas2008synchronization} and so on \cite{barrat2004architecture,barabasi2009scale,barabasi1999emergence,barabasi2009scale,song2005self,teixeira2010complex}.
How to identify the influential nodes in the complex networks has attracted many researchers to study it. Recently, a local structure entropy of the complex networks is proposed to identify the influential nodes in the complex networks \cite{zhang2014local}. In the local structure entropy the node's influence on the whole network is replaced by the local network. The degree entropy of the local network is used as the measure of the influence of the node on the whole network. The local structure entropy is based on the shannon entropy. In this paper the Tsalli entropy which is proposed by Tsalli et.al \cite{tsallis1988possible} is used to analysis the property of the local structure entropy.
Depends on the nonextensive statistical mechanics, the relationship between each node can be described by the nonextensive additivity.
In this paper, the property of the local structure entropy is analysed by the nonextensive statistical mechanics.
Depends on the Tsallis entropy, a new form of the local structure entropy is proposed in this paper. In the nonextensive local structure entropy, the influences of the node on the whole network is changed by the entropic index $q$. The nonextensive in the local structure entropy is changed correspond to value of $q$ and the nonextensive additivity is restricted by the value of $q$. We also find the nonextensive threshold value of $q$ in the nonextensive local structure entropy. When the value of $q$ is bigger than the nonextensive threshold value, then the property of the local structure entropy will not be controlled by the $q$. When the value of $q$ is equal to 0, then the local structure entropy is degenerated to another form of the degree centrality. The nonextensive local structure entropy is a generalised of the local structure entropy.
The rest of this paper is organised as follows. Section \ref{Rreparatorywork} introduces some preliminaries of this work, such as the local structure entropy of complex networks and the nonextensive statistical mechanics. In section \ref{newmethod}, the analysis of the local structure entropy based on the nonextensive is proposed. The application of the nonextensive analysis in these real networks is shown in the section \ref{application}. Conclusion is given in Section \ref{conclusion}.
\section{Preliminaries}
\label{Rreparatorywork}
\subsection{Local structure entropy of complex networks}
\label{Existing methods}
There many methods are proposed to identify the influential nodes in the complex networks. The degree centrality and the betweenneess centrality are the wildly used method to identify the influential nodes in the complex networks.
Recently, the "Local structure entropy" of the complex networks which is based on the degree centrality and the shannon entropy is proposed \cite{zhang2014local}. The details of the local structure entropy of the complex networks is shown as follows \cite{zhang2014local}.
The definition of the local structure entropy can be divided into three steps, the details are shown as follows \cite{zhang2014local}.
\textbf{Step 1 Creating a local network}: First, choose one of the node in the network as the central node. Second, find all of the nodes in the network which are connect with the central node in directly. Third, create a local network which contains the central node and his neighbour nodes.
\textbf{Step 2 Calculating the unit of the local structure entropy}: Calculate the degree of each node and the total number of the degree in the local network. The unit of the local structure entropy can be represents as the ${p_{ij}}$, it is defined in the Eq.(\ref{P_ij}).
\textbf{Step 3 Calculating the local structure entropy of each node}: The definition of the local structure entropy for each node is shown in the Eq.(\ref{Local_E}).
The definition of the local structure entropy of the complex networks is shown as follows \cite{zhang2014local}.
\begin{equation}\label{Local_E}
{LE_i} = -\sum\limits_{j = 1}^n {{p_{ij}}} \log {p_{ij}}
\end{equation}
Where the ${LE_i}$ represents the local structure entropy of the $i$th node in the complex networks. The $n$ is the total number of the nodes in the local network. The ${p_{ij}}$ represents the percentage of degree for the $j$th node in the local network. The definition of the ${p_{ij}}$ is shown in the Eq.(\ref{P_ij}).
\begin{equation}\label{P_ij}
{p_{ij}} = \frac{{\ degree(j)}}{{\sum\limits_{j = 1}^n {\ degree(j)} }}
\end{equation}
An example of the process to calculate the local structure entropy is shown in the Fig.\ref{E_E}.
\begin{figure}[!htbp]
\centering
\subfigure[The example network A ]{
\label{karate_E:subfig:a}
\centering
\includegraphics[scale=0.6]{A21.eps}}
\subfigure[The 5$th$ node. ]{
\label{karate_E:subfig:a}
\centering
\includegraphics[scale=0.8]{A21_5_U.eps}}
\hspace{2cm}
\subfigure[The 15$th$ node]{
\label{karate_E:subfig:b}
\includegraphics[scale=0.8]{A21_15_U.eps}}
\centering
\caption{In the example network A, different node has different value of degree. The ${LE_i}$ of each node is different to each others. We use the node 5 and node 15 to show the details of the calculation of the local structure entropy of each node. First, the local structure entropy of the node 5. The node 5 has 5 neighbours, the node 2, 3, 20, 8, 7. The degree of each neighbour node is 3, 3, 3, 3, 5. The degree of the node 5 is 5. The total degree in the local network is 19. Then the set of the degree in the local network is ${D_5}={3/19, 3/19, 3/19, 5/19, 5/19}$. Then the $LE_5$=1.8684. Second, the local structure entropy of the node 15. The node 15 has 6 neighbours, the node 12, 13, 17, 18, 19, 21. The degree of each neighbour node is 3, 2, 3, 4, 4, 3. The degree of node 15 is 6. The total degree in the local network is 25. The set of the degree in the local network is ${D_15}={3/25, 2/25, 3/25, 4/25, 4/25, 3/25, 6/25}$ Then the $LE_15$=1.89426. From the definition of the local structure entropy the node 15 is more influential than the node 5 in the example network.}\label{E_E}
\end{figure}
It is clear that in the local structure entropy the influence of the node on the whole network is replaced by the influence of the local network on the whole network \cite{zhang2014local}.
\subsection{Nonextensive statistical mechanics}
\label{Tsallis entropy}
The entropy is defined by Clausius for thermodynamics \cite{clausius1867mechanical}. For a finite discrete set of probabilities the definition of the Boltzmann-Gibbs entropy \cite{gibbs2010elementary} is given as follows:
\begin{equation}\label{S_BG}
{S_{BG}} = -k \sum\limits_{i = 1}^N {{p_i}} \ln {p_i}
\end{equation}
The conventional constant $k$ is the Boltzmann universal constant for thermodynamic systems. The value of $k$ will be taken to be unity in information theory \cite{shannon2001mathematical}.
In 1988, a generalised entropy have been proposed by Tsallis \cite{tsallis1988possible}. It is shown as follows:
\begin{equation}\label{S_q}
{S_q} = - k\sum\limits_{i = 1}^N {{p_i}} {\ln _q}\frac{1}{{{p_i}}}
\end{equation}
The $q-logarithmic$ function in the Eq. (\ref{S_q}) is presented as follows:
\begin{equation}\label{ln_q}
{\ln _q}{p_i} = \frac{{{p_i}^{1 - q} - 1}}{{1 - q}}({p_i} > 0;q \in \Re ;l{n_1}{p_i} = ln{p_i})
\end{equation}
Based on the Eq. (\ref{ln_q}), the Eq. (\ref{S_q}) can be rewritten as follows:
\begin{equation}\label{S_q1}
{S_q} = - k\sum\limits_{i = 1}^N {{p_i}} \frac{{{p_i}^{q - 1} - 1}}{{1 - q}}
\end{equation}
\begin{equation}\label{S_q2}
{S_q} = - k\sum\limits_{i = 1}^N {\frac{{{p_i}^q - {p_i}}}{{1 - q}}}
\end{equation}
\begin{equation}\label{S_q1}
{S_q} = k\frac{{1 - \sum\limits_{i = 1}^N {{p_i}^q} }}{{q - 1}}
\end{equation}
Where ${N}$ is the number of the subsystems.
Based on the Tsallis entropy, the nonextensive theory is established by Tsallis et.al. The nonextensive statistical mechanics is a generalised statistical mechanics.
\section{Nonextensive analysis of the local structure entropy of complex networks}
\label{newmethod}
The main idea of the local structure entropy is try to use the influence of the local network to replace the influence of the node on the whole network \cite{zhang2014local}. However, in the definition of the local structure entropy of each node, the relationship between each node in the local network is extensive. In order to illuminate the property of the local structure entropy, in this paper the nonextensive statistical mechanic is used in the definition of the local structure entropy.
Depends on the Tsallis entropy, the definition of the local structure entropy is redefined as follows:
\begin{equation}\label{T_local_E}
{S_q}_i = - k\sum\limits_{i = 1}^N {{p_i}} {\ln _q}\frac{1}{{{p_i}}}
\end{equation}
Where in the Eq.(\ref{T_local_E}), the logarithmic function in the local structure entropy is replaced by the $q-logarithmic$ function in the Eq.(\ref{ln_q}).
The ${{S_q}_i}$ is the new local structure entropy of the node $i$. It is defined based on the Tsallis entropy \cite{tsallis2009introduction}. The ${{p_i}_j}$ is defined in the Eq.(\ref{P_ij}). The $q$ is the nonextensive entropic index.
We use the example network A which is shown in the Fig.\ref{E_E} to show the nonextensive property of the local structure entropy.
\begin{figure}
\centering
\subfigure[q=0.1]{
\label{Yeast-infect-local1}
\centering
\includegraphics[scale=0.25]{1.eps}}
\subfigure[q=0.2]{
\label{Yeast-infect-local:a}
\centering
\includegraphics[scale=0.25]{2.eps}}
\subfigure[q=0.3]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{3.eps}}
\subfigure[q=0.4]{
\label{Yeast-infect-local:a}
\centering
\includegraphics[scale=0.25]{4.eps}}
\subfigure[q=0.5]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{5.eps}}
\subfigure[q=0.6]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{6.eps}}
\subfigure[q=0.7]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{7.eps}}
\subfigure[q=0.8]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{8.eps}}
\subfigure[q=0.9]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{9.eps}}
\subfigure[q=1.0]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{10.eps}}
\subfigure[q=1.1]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{11.eps}}
\subfigure[q=1.2]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{12.eps}}
\caption{The figure show the value of the nonextensive local structure entropy of each node in the example network A. The value of $q$ is big than 0.1 and small than 1.2. The caption of the subfigure show the value of $q$. The Abscissa in those subfigure represents the node's number and the ordinate represents the value of nonextensive local structure entropy. }\label{Yeast-infect-local}
\end{figure}
\begin{figure}
\centering
\subfigure[q=1.3]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{13.eps}}
\subfigure[q=1.4]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{14.eps}}
\subfigure[q=1.5]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{15.eps}}
\subfigure[q=1.6]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{16.eps}}
\subfigure[q=1.7]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{17.eps}}
\subfigure[q=1.8]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{18.eps}}
\subfigure[q=1.9]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{19.eps}}
\subfigure[q=2.0]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{20.eps}}
\subfigure[q=2.2]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{21.eps}}
\subfigure[q=2.4]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{22.eps}}
\subfigure[q=2.6]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{23.eps}}
\subfigure[q=2.8]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{24.eps}}
\caption{The figure show the value of the nonextensive local structure entropy of each node in the example network A. The value of $q$ is big than 1.3 and small than 2.8. The caption of the subfigure show the value of $q$. The Abscissa in those subfigure represents the node's number and the ordinate represents the value of nonextensive local structure entropy. }\label{Yeast-infect-local}
\end{figure}
\begin{figure}
\centering
\subfigure[q=3.0]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{25.eps}}
\subfigure[q=3.2]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{26.eps}}
\subfigure[q=3.4]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{27.eps}}
\subfigure[q=3.6]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{28.eps}}
\subfigure[q=3.8]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{29.eps}}
\subfigure[q=4.0]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{30.eps}}
\subfigure[q=4.5]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{31.eps}}
\subfigure[q=5.0]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{32.eps}}
\subfigure[q=5.5]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{33.eps}}
\subfigure[q=6.0]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{34.eps}}
\subfigure[q=6.5]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{35.eps}}
\subfigure[q=7.0]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{36.eps}}
\caption{The figure show the value of the nonextensive local structure entropy of each node in the example network A. The value of $q$ is big than 3.0 and small than 7.0. The caption of the subfigure show the value of $q$. The Abscissa in those subfigure represents the node's number and the ordinate represents the value of nonextensive local structure entropy. }\label{Yeast-infect-local}
\end{figure}
\begin{figure}
\centering
\subfigure[q=7.5]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{37.eps}}
\subfigure[q=8.0]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{38.eps}}
\subfigure[q=8.5]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{39.eps}}
\subfigure[q=9.0]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{40.eps}}
\subfigure[q=9.5]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{41.eps}}
\subfigure[q=10.0]{
\label{Yeast-infect-local:b}
\centering
\includegraphics[scale=0.25]{42.eps}}
\caption{The figure show the value of the nonextensive local structure entropy of each node in the example network A. The value of $q$ is big than 7.5 and small than 10. The caption of the subfigure show the value of $q$. The Abscissa in those subfigure represents the node's number and the ordinate represents the value of nonextensive local structure entropy. }\label{Yeast-infect-local}
\end{figure}
\begin{table}[htbp]
\tiny
\addtolength{\tabcolsep}{-2pt}
\centering
\caption{The order of the influential nodes in the example network A with the change of the value of $q$}
\begin{tabular}{l|ccccccccccccccccccccc}
\hline
Node order&1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 & 16 & 17 & 18 & 19 & 20 & 21 \\
\hline
\rowcolor{gray}
q=0 & 15 & 7 & 5 & 10 & 19 & 18 & 6 & 1 & 21 & 3 & 20 & 17 & 8 & 2 & 12 & 16 & 4 & 14 & 11 & 13 & 9 \\
q=0.1 & 15 & 5 & 7 & 10 & 18 & 19 & 1 & 3 & 6 & 20 & 8 & 2 & 21 & 17 & 12 & 16 & 4 & 14 & 11 & 13 & 9 \\
q=0.2 & 15 & 5 & 7 & 10 & 19 & 18 & 20 & 12 & 8 & 2 & 3 & 17 & 1 & 21 & 6 & 16 & 4 & 14 & 11 & 13 & 9 \\
q=0.3 & 15 & 5 & 7 & 19 & 18 & 10 & 12 & 8 & 2 & 20 & 17 & 3 & 1 & 21 & 6 & 14 & 16 & 4 & 11 & 13 & 9 \\
q=0.4 & 15 & 5 & 7 & 19 & 18 & 10 & 12 & 8 & 2 & 20 & 17 & 3 & 21 & 1 & 6 & 14 & 11 & 13 & 16 & 4 & 9 \\
q=0.5 & 15 & 5 & 7 & 19 & 18 & 10 & 12 & 8 & 2 & 17 & 20 & 3 & 21 & 1 & 6 & 11 & 14 & 13 & 16 & 4 & 9 \\
q=0.6 & 15 & 5 & 7 & 19 & 18 & 10 & 12 & 8 & 2 & 17 & 20 & 3 & 21 & 1 & 6 & 11 & 13 & 14 & 16 & 4 & 9 \\
q=0.7 & 15 & 5 & 7 & 19 & 18 & 10 & 12 & 8 & 2 & 17 & 20 & 3 & 21 & 1 & 6 & 13 & 11 & 14 & 16 & 4 & 9 \\
q=0.8 & 15 & 5 & 7 & 19 & 18 & 12 & 10 & 8 & 2 & 17 & 20 & 3 & 21 & 1 & 6 & 13 & 11 & 14 & 16 & 4 & 9 \\
q=0.9 & 15 & 5 & 7 & 12 & 19 & 18 & 8 & 2 & 17 & 20 & 10 & 3 & 21 & 1 & 6 & 13 & 11 & 14 & 16 & 4 & 9 \\
\rowcolor{gray}
q=1 & 15 & 5 & 7 & 12 & 8 & 2 & 19 & 18 & 17 & 20 & 3 & 10 & 21 & 1 & 13 & 6 & 11 & 14 & 16 & 4 & 9 \\
q=1.1 & 15 & 5 & 7 & 12 & 8 & 2 & 17 & 19 & 18 & 20 & 3 & 21 & 10 & 1 & 13 & 11 & 14 & 6 & 16 & 4 & 9 \\
q=1.2 & 15 & 5 & 12 & 7 & 8 & 2 & 17 & 20 & 19 & 18 & 3 & 21 & 10 & 13 & 11 & 1 & 14 & 6 & 16 & 4 & 9 \\
q=1.3 & 15 & 5 & 12 & 7 & 8 & 2 & 17 & 20 & 19 & 18 & 3 & 21 & 13 & 10 & 11 & 14 & 1 & 16 & 4 & 6 & 9 \\
q=1.4 & 12 & 15 & 5 & 8 & 2 & 7 & 17 & 20 & 3 & 19 & 18 & 21 & 13 & 11 & 14 & 10 & 1 & 16 & 4 & 6 & 9 \\
q=1.5 & 12 & 5 & 15 & 8 & 2 & 17 & 7 & 20 & 3 & 19 & 18 & 13 & 21 & 11 & 14 & 10 & 16 & 4 & 1 & 6 & 9 \\
q=1.6 & 12 & 8 & 2 & 17 & 5 & 15 & 20 & 7 & 3 & 13 & 19 & 18 & 21 & 11 & 14 & 16 & 4 & 10 & 1 & 6 & 9 \\
q=1.7 & 12 & 8 & 2 & 17 & 5 & 20 & 15 & 7 & 13 & 3 & 11 & 21 & 18 & 19 & 14 & 16 & 4 & 10 & 1 & 6 & 9 \\
q=1.8 & 12 & 8 & 2 & 17 & 20 & 5 & 13 & 15 & 7 & 3 & 11 & 21 & 18 & 19 & 14 & 16 & 4 & 1 & 10 & 6 & 9 \\
q=1.9 & 12 & 8 & 2 & 17 & 20 & 13 & 5 & 11 & 15 & 3 & 7 & 21 & 14 & 19 & 18 & 16 & 4 & 1 & 10 & 6 & 9 \\
q=2 & 12 & 8 & 2 & 17 & 13 & 20 & 11 & 5 & 3 & 15 & 7 & 21 & 14 & 19 & 18 & 16 & 4 & 1 & 10 & 6 & 9 \\
q=2.2 & 12 & 8 & 2 & 17 & 13 & 20 & 11 & 5 & 3 & 14 & 21 & 15 & 7 & 19 & 18 & 16 & 4 & 1 & 6 & 10 & 9 \\
q=2.4 & 12 & 13 & 8 & 2 & 17 & 11 & 20 & 14 & 3 & 5 & 21 & 15 & 7 & 18 & 19 & 16 & 4 & 1 & 6 & 10 & 9 \\
q=2.6 & 12 & 13 & 8 & 2 & 17 & 11 & 20 & 14 & 3 & 21 & 5 & 16 & 4 & 7 & 15 & 19 & 18 & 1 & 6 & 10 & 9 \\
q=2.8 & 12 & 13 & 2 & 8 & 17 & 11 & 20 & 14 & 3 & 21 & 5 & 16 & 4 & 7 & 19 & 18 & 15 & 1 & 6 & 10 & 9 \\
q=3 & 12 & 13 & 8 & 2 & 17 & 11 & 20 & 14 & 3 & 21 & 5 & 16 & 4 & 19 & 18 & 7 & 15 & 1 & 6 & 9 & 10 \\
q=3.2 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 5 & 16 & 4 & 18 & 19 & 7 & 15 & 1 & 6 & 9 & 10 \\
q=3.4 & 12 & 13 & 11 & 2 & 8 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\hline
\rowcolor{gray}
q=3.6 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=3.8 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=4 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=4.5 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=5 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 18 & 19 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=5.5 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=6 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=6.5 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=7 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=7.5 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 18 & 19 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=8 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 18 & 19 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=8.5 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=9 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=9.5 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\rowcolor{gray}
q=10 & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\hline
\end{tabular}%
\label{tab:order}%
\end{table}%
The order of the influential nodes in the example network A is shown in the Table \ref{tab:degree}.
\begin{table}[htbp]
\scriptsize
\addtolength{\tabcolsep}{-2pt}
\centering
\caption{The degree of each node in the example network A}
\begin{tabular}{cccccccccccccccccccccc}
\hline
Node number & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 & 16 & 17 & 18 & 19 & 20 & 21 \\
\hline
Degree & 3 & 3 & 3 & 2 & 5 & 3 & 5 & 3 & 1 & 4 & 2 & 3 & 2 & 2 & 6 & 2 & 3 & 4 & 4 & 3 & 3 \\
\hline
\end{tabular}%
\label{tab:degree}%
\end{table}%
The order of the influential nodes in the example network A is shown in the Table \ref{tab:degree_order}. Where in the Table \ref{tab:degree_order}, the $D_{order}$ represents the order of the influential nodes in the example network A, the $D_{order1}$ represents another order of the influential nodes in the example network A. Both of the $D_{order}$ and the $D_{order1}$ is based on the degree of each node. The $LE_{order}q=x$ represents the order of the influential nodes which is identified by the nonextensive local structure entropy with different value of $q$. The $LE_{order}$ represents the order of the influential nodes which is identified by the local structure entropy.
\begin{table}[htbp]
\scriptsize
\addtolength{\tabcolsep}{-2pt}
\centering
\caption{The order of the influential nodes in the example network A}
\begin{tabular}{l|ccccccccccccccccccccc}
\toprule
$D_{order}$ & 15 & 5 & 7 & 10 & 18 & 19 & 1 & 2 & 3 & 6 & 8 & 12 & 17 & 20 & 21 & 4 & 11 & 13 & 14 & 16 & 9 \\
\rowcolor{gray}
$D_{order1}$& 15 & 7 & 5 & 10 & 19 & 18 & 6 & 1 & 21 & 3 & 20 & 17 & 8 & 2 & 12 & 16 & 4 & 14 & 11 & 13 & 9 \\
Degree & 6 & 5 & 5 & 4 & 4 & 4 & 3 & 3 & 3 & 3 & 3 & 3 & 3 & 3 & 3 & 2 & 2 & 2 & 2 & 2 & 1 \\
\midrule
\rowcolor{gray}
$LE_{order}q=0 $ & 15 & 7 & 5 & 10 & 19 & 18 & 6 & 1 & 21 & 3 & 20 & 17 & 8 & 2 & 12 & 16 & 4 & 14 & 11 & 13 & 9 \\
\midrule
$LE_{order}q=1$ & 15 & 5 & 7 & 12 & 8 & 2 & 19 & 18 & 17 & 20 & 3 & 10 & 21 & 1 & 13 & 6 & 11 & 14 & 16 & 4 & 9 \\
\midrule
$LE_{order}$ & 15 & 5 & 7 & 12 & 8 & 2 & 19 & 18 & 17 & 20 & 3 & 10 & 21 & 1 & 13 & 6 & 11 & 14 & 16 & 4 & 9 \\
\bottomrule
\end{tabular}%
\label{tab:degree_order}%
\end{table}%
The results in of the test of the local structure entropy based on the nonextensive statistical have show the influence of the nonextensive additivity between each node on the local structure entropy.
When the value of the entropic index $q$ is equal to 0. Then the value of the local structure entropy on each nodes is corresponded to the number of the degree of each node. The influence of the local network is degenerated to the degree's influence on the whole network. Therefore, the order of the influential nodes which is identified by the local structure entropy in the example network A is the same as the the order based on the degree value. In other word, when the value of $q$ is equal to 0, then influence of the components on the local structure entropy is equal to others. The value of the local structure entropy for each node is decided by the node's degree. When the value of $q$ is equal to 1, then the additivity among there components of local structure entropy is based on the degree of the node in the local network. When the value of $q$ is bigger than 3.6, the order of the influential nodes in the example networks is stable. It means that when the value of $q$ is bigger than 3.6, the nonextensive additivity among those components is stable. Change the value of $q$ has no influence on the order of the local structure entropy. The 3.6 is a threshold value of the nonextensive in the local structure entropy of example network A. The $P_{value}$ is used to represents the threshold value of the nonextensive in the local structure entropy.
The details can be illuminated in six parts:
\begin{itemize}
\item[Case 1] When \textbf{q=0}, the relationship between the components in the local structure entropy is equal to each others. The value of the local structure entropy is decided by the number of the components in it. In the local structure entropy which is based on the degree distribution, the order of the influential nodes in the network is equal to the order which is identified by the degree centrality.
\item[Case 2] When \textbf{0$<$q$<$1}, these components which have small value are the main part of the local structure entropy.
\item[Case 3] When \textbf{q=1}, the nonextensive additivity in the local structure entropy is degenerated to the extensive additivity. The nonextensive local structure entropy degenerate to the local structure entropy.
\item[Case 4] When \textbf{1$<$q$<$$P_{value}$}, the components which have big value are the main part of the local structure entropy.
\item[Case 5] When \textbf{q=$P_{value}$}, the nonextensive additivity in the local structure entropy achieve to a stable state.
\item[Case 6] When \textbf{q$>$$P_{value}$}, the order of the influential nodes in the complex networks is achieve to stable. Change the value of $q$ will have no influence on the order of the influential nodes in the complex networks.
\end{itemize}
The order of the influential nodes in the complex can be described in three state: 1. The initial state $Order_{q=0}$. When the value of $q$ is equal to 0, the order of the influential nodes in the complex networks. This state is the same as the order which is identified by the degree centrality. 2. The local structure entropy state $Order_{q=1}$. When the value of $q$ is equal to 1, the order of the influential nodes in the complex networks. The order is decided by the local structure entropy. 3. The stable state $Order_{stable}$. When the value of $q$ is bigger than the nonextensive threshold value, the order of the influential nodes in the complex networks is stable.
The three state of the example network A is shown as follows:
\begin{table}[htbp]
\scriptsize
\addtolength{\tabcolsep}{-2pt}
\centering
\caption{The three state of the influential nodes in the example network A}
\begin{tabular}{l|ccccccccccccccccccccc}
\toprule
Node order&1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 & 16 & 17 & 18 & 19 & 20 & 21 \\
\midrule
$Order_{q=0}$ & 15 & 5 & 7 & 10 & 18 & 19 & 1 & 2 & 3 & 6 & 8 & 12 & 17 & 20 & 21 & 4 & 11 & 13 & 14 & 16 & 9 \\
$Order_{q=1}$ & 15 & 5 & 7 & 12 & 8 & 2 & 19 & 18 & 17 & 20 & 3 & 10 & 21 & 1 & 13 & 6 & 11 & 14 & 16 & 4 & 9 \\
$Order_{stable}$ & 12 & 13 & 11 & 8 & 2 & 17 & 20 & 14 & 3 & 21 & 16 & 4 & 5 & 19 & 18 & 7 & 15 & 9 & 1 & 6 & 10 \\
\bottomrule
\end{tabular}%
\label{tab:degree_order}%
\end{table}%
The results show that the nonextensive local structure entropy is more useful and more reasonable than the local structure entropy. The nonextensive local structure entropy is a generalised method to identify the influential nodes in the complex networks.
\section{Application}
\label{application}
In this section four real networks is use to analysis the nonextensive in the local structure entropy of it. The four networks are the Zachary's Karate Club network (Karate) \cite{uci}, the US-airport network (Us-airport) \cite{networkdata}, Email networks (Email) \cite{networkdata}and the Germany highway networks (Highway) \cite{nettt}.
The results of the nonextensive threshold value ($P_{value}$) of each network and the three state of the four networks is shown as follows:
\begin{table}[htbp]
\scriptsize
\addtolength{\tabcolsep}{-2pt}
\centering
\caption{The three states of the (Karate)network}
\begin{tabular}{c|cccccccccc|c|cccccccccc}
\toprule
& \multicolumn{10}{c}{Top 10 influential nodes} & \multicolumn{1}{c}{} & \multicolumn{10}{c}{Low 10 influential nodes} \\
\midrule
$Order_{q=0}$ & 34 & 1 & 33 & 3 & 2 & 4 & 32 & 24 & 14 & 9 & & 13 & 18 & 22 & 10 & 15 & 16 & 19 & 21 & 23 & 12 \\
$Order_{q=1}$ & 14 & 9 & 3 & 32 & 31 & 8 & 4 & 1 & 33 & 2 & & 6 & 7 & 5 & 11 & 13 & 26 & 27 & 25 & 17 & 12 \\
$Order_{stable}$ & 12 & 15 & 16 & 19 & 21 & 23 & 20 & 10 & 18 & 22 & & 3 & 6 & 7 & 2 & 33 & 26 & 25 & 1 & 17 & 34 \\
\bottomrule
\end{tabular}%
\label{tab:Karate}%
\end{table}%
\begin{table}[htbp]
\tiny
\addtolength{\tabcolsep}{-3pt}
\centering
\caption{The three states of the (Us-airport) network}
\begin{tabular}{c|cccccccccc|c|cccccccccc}
\toprule
& \multicolumn{10}{c}{Top 10 influential nodes} & \multicolumn{1}{c}{} & \multicolumn{10}{c}{Low 10 influential nodes} \\
\midrule
$Order_{q=0}$ & 118 & 261 & 255 & 152 & 182 & 230 & 166 & 67 & 112 & 201 & & 277 & 278 & 279 & 280 & 282 & 291 & 294 & 304 & 88 & 114 \\
\midrule
$Order_{q=1}$ & 159 & 292 & 172 & 131 & 94 & 109 & 307 & 301 & 310 & 305 & & 188 & 31 & 22 & 171 & 328 & 330 & 32 & 122 & 332 & 327 \\
$Order_{stable}$ & 88 & 114 & 241 & 247 & 257 & 268 & 277 & 278 & 279 & 280 & & 22 & 31 & 13 & 171 & 328 & 330 & 122 & 332 & 32 & 327 \\
\bottomrule
\end{tabular}%
\label{tab:Us-airport}%
\end{table}%
\begin{table}[htbp]
\tiny
\addtolength{\tabcolsep}{-4pt}
\centering
\caption{The three states of the (Email) network}
\begin{tabular}{c|cccccccccc|c|cccccccccc}
\toprule
& \multicolumn{10}{c}{Top 10 influential nodes} & \multicolumn{1}{c}{} & \multicolumn{10}{c}{Low 10 influential nodes} \\
\midrule
$Order_{q=0}$ & 105 & 333 & 23 & 16 & 42 & 41 & 196 & 233 & 21 & 76 & & 791 & 955 & 956 & 959 & 294 & 253 & 261 & 779 & 374 & 644 \\
\midrule
$Order_{q=1}$ & 105 & 3 & 39 & 16 & 42 & 54 & 210 & 390 & 50 & 332 & & 1091 & 1099 & 1118 & 1119 & 1121 & 1130 & 1052 & 1103 & 1125 & 1133 \\
$Order_{stable}$ & 644 & 236 & 374 & 779 & 247 & 394 & 253 & 261 & 294 & 955 & & 1091 & 1099 & 1118 & 1119 & 1121 & 1130 & 1052 & 1103 & 1125 & 1133 \\
\bottomrule
\end{tabular}%
\label{tab:Email}%
\end{table}%
\begin{table}[htbp]
\tiny
\addtolength{\tabcolsep}{-4pt}
\centering
\caption{The three states of the (Highway) network}
\begin{tabular}{c|cccccccccc|c|cccccccccc}
\toprule
& \multicolumn{10}{c}{Top 10 influential nodes} & \multicolumn{1}{c}{} & \multicolumn{10}{c}{Low 10 influential nodes} \\
\midrule
$Order_{q=0}$ & 693 & 403 & 300 & 410 & 758 & 373 & 217 & 556 & 207 & 331 & & 247 & 587 & 678 & 898 & 779 & 326 & 1031 & 265 & 787 & 798 \\
\midrule
$Order_{q=1}$ & 219 & 393 & 698 & 217 & 404 & 450 & 543 & 267 & 331 & 763 & & 1154 & 1155 & 1158 & 1160 & 1162 & 1163 & 1165 & 1166 & 1167 & 1168 \\
$Order_{stable}$ & 265 & 787 & 798 & 326 & 1031 & 779 & 687 & 709 & 161 & 247 & & 1128 & 1129 & 1131 & 1133 & 1138 & 1144 & 1145 & 1156 & 1159 & 1164 \\
\bottomrule
\end{tabular}%
\label{tab:Highway}%
\end{table}%
\begin{table}[htbp]
\centering
\caption{The nonextensive threshold value of the four networks}
\begin{tabular}{cccc}
\toprule
Network &Nodes& edages& $P_{value}$\\
\midrule
Karate & 34 & 78 & 4.5 \\
Us-airport & 332 & 2126 & 7.6 \\
Email & 1133 & 10902 & 6.2 \\
Highway & 1168 & 2486 & 4.1 \\
\bottomrule
\end{tabular}%
\label{tab:threshold value}%
\end{table}%
The threshold value of the nonextensive ($P_{value}$) in the four real networks is shown in the Table \ref{tab:threshold value}. It is clear that the nonextensive in the local structure entropy if not based on the scale of the network.
\section{Conclusion}
\label{conclusion}
The local structure entropy is a new method which is used to identify the influential nodes in the complex networks. In this paper, the local structure entropy of the complex networks is redefined by the nonextensive statistical mechanics. The results of the nonextensive analysis on the local structure entropy shown that when the entropic index $q$ is equal to 0, the order of the influential nodes which is identified by the nonextensvie local structure entropy the same as the degree centrality. When the value of $q$ is equal to 1, then the nonextensive local structure entropy is degenerated to the traditional local structure entropy. When the value of $q$ is bigger than the nonextensive threshold value ($P_{value}$) the order of the influential nodes in the complex networks is stable.
It is clear that the value of $q$ will influence the property of the local structure entropy, but it also have a range. When the value of $q$ is smaller than $P_{value}$ and bigger than 1, the components with big value in the local structure entropy play an important roles in the local structure entropy. When the value of $q$ is smaller than 1, bigger than 0 the components with small value in the local structure entropy play an important roles in the local structure entropy. When the value of $q$ is equal to 0, then the performance of the local structure entropy is decided by the numbers of the components in the local structure entropy. The nonextensive local structure entropy is degenerates to the degree centrality.
The new form of the local structure entropy which is defined based on the Tsallis entropy is more reasonable and more useful than the existing one. It is a generalised method which can be used to identify the influential nodes in the complex networks.
\section*{Acknowledgments}
The work is partially supported by National Natural Science Foundation of China (Grant No. 61174022), Specialized Research Fund for the Doctoral Program of Higher Education (Grant No. 20131102130002), R$\&$D Program of China (2012BAH07B01), National High Technology Research and Development Program of China (863 Program) (Grant No. 2013AA013801), the open funding project of State Key Laboratory of Virtual Reality Technology and Systems, Beihang University (Grant No.BUAA-VR-14KF-02). Fundamental Research Funds for the Central Universities No. XDJK2015D009. Chongqing Graduate Student Research Innovation Project (Grant No. CYS14062).
\bibliographystyle{elsarticle-num}
|
\section{Introduction}\label{INTRODUCTION}
\IEEEPARstart{F}{ree-space} optical (FSO) communication enables wireless connectivity through atmosphere using laser transmitters at infrared bands. These systems provide high data rates comparable to fiber optics while they offer much more flexibility in installation and deployment. Since they operate in unregulated spectrum, no licensing fee is required leading to a cost-effective solution \cite{1,2,kashani2}.
A major performance limiting factor in FSO systems is atmospheric turbulence-induced fading (also called as scintillation) \cite{3}. Inhomogenities in the temperature and the pressure of the atmosphere result in variations of the refractive index and cause atmospheric turbulence. This manifests itself as random fluctuations in the received signal. In the literature, several statistical models have been proposed to model this random phenomenon. Historically, log-normal distribution has been the most widely used statistical model for the random irradiance experienced over atmospheric channels \cite{4,5,6,6.1,6.2,kashani3}. This model is however restricted to weak turbulence conditions and has large deviations from the experimental data when the strength of turbulence increases.
In an effort to come up with a more general model to cover a wide range of turbulence conditions, other statistical models have been proposed in the literature which include the K \cite{7}, I-K \cite{8}, log-normal Rician \cite{9}, Gamma-Gamma \cite{10}, M \cite{11} and Double Weibull \cite{12} distributions. In our recent work, we proposed the so-called Double Generalized Gamma (Double GG) as a unifying distribution for the irradiance fluctuations \cite{kashani}. This model is valid under all range of turbulence conditions and contains most of the existing models in the literature as special cases.
In our previous work \cite{kashani}, as an initial performance study, we derived the BER performance of single-input single-output (SISO) FSO link over Double GG channels. In this work, we extend our performance analysis to multiple-input multiple-output (MIMO) FSO systems. MIMO FSO systems are known to mitigate turbulence-induced fading and significantly improve the performance. Some earlier results on MIMO FSO systems over log-normal, K, negative exponential and Gamma-Gamma channels can be found in \cite{15,20,21,Farid, Antonio}. In this paper, we study the error rate performance of single-input multiple-output (SIMO), multiple-input single-output (MISO) and MIMO FSO systems employing intensity modulation/direct detection (IM/DD) with on-off keying (OOK) over independent and not necessarily identically distributed (i.n.i.d.) Double GG turbulence channels.
The rest of the paper is organized as follows: In Section II, we introduce the MIMO FSO system model. In Section III, we provide the BER expressions for SIMO, MISO and MIMO FSO links. In Section IV, we present numerical results to confirm the accuracy of the derived expressions and demonstrate the advantages of employing spatial diversity over SISO links. Finally, Section V concludes the paper.
\section{System Model}
We consider an FSO system employing IM/DD with OOK where the information signal is transmitted via $M$ apertures and received by $N$ apertures over the Double GG channel. The received signal at the $n^{\text{th}}$ receive aperture is then given by
\begin{equation}\label{eq12}
{{r}_{n}}=\eta x\sum\limits_{m=1}^{M}{{{I}_{mn}}}+{{\upsilon }_{n}},\,\,\,n=1,\ldots ,N
\end{equation}
where $x$ represents the information bits and can be either 0 or 1, ${\upsilon }_{n}$ is the Additive White Gaussian noise (AWGN) term with zero mean and variance $\sigma _{\upsilon}^{2}={{N}_{0}}/2$ , and $\eta$ is the optical-to-electrical conversion coefficient. Here, $I_{mn}$ is the normalized irradiance from the $m^{\text{th}}$ transmitter to the $n^{\text{th}}$ receiver whose pdf follows [\citen{23}, Eq. (1)]
\begin{align}\label{eq4}
&{{f}_{I}}\left( I \right)=\frac{{{\gamma }_{2}}p{{p}^{{{\beta}_{2}}-1/2}}{{q}^{{{\beta}_{1}}-1/2}}{{\left( 2\pi \right)}^{{1-\left( p+q \right)}/{2}\;}}{{I}^{-1}}}{\Gamma \left( {{\beta}_{1}} \right)\Gamma \left( {{\beta}_{2}} \right)}\\\nonumber
&\times G_{p+q,0}^{0,p+q}\left[ {{\left( \frac{{{\Omega }_{2}}}{{{I}^{{{\gamma }_{2}}}}} \right)}^{p}}\frac{{{p}^{p}}{{q}^{q}}\Omega _{1}^{q}}{\beta_{1}^{q}\beta_{2}^{p}}|\begin{matrix}
\Delta \left( q:1-{{\beta}_{1}} \right),\Delta \left( p:1-{{\beta}_{2}} \right) \\
- \\
\end{matrix} \right]
\end{align}
where $G_{p,q}^{m,n}\left[ . \right]$ is the Meijer’s G-function defined in [\citen{24}, Eq.(9.301)], $p$ and $q$ are positive integer numbers that satisfy ${p}/{q}\;={{{\gamma }_{1}}}/{{{\gamma }_{2}}}\;$ and $\Delta (j;x)\triangleq {x}/{j}\;,...,{\left( x+j-1 \right)}/{j}\;$, and ${{\beta }_{i}}\ge 0.5$ is a shaping parameter modeling the severity of fading. The distribution parameters ${{\gamma }_{i}}$ and ${{\Omega }_{i}}$, $i=1,2$ , of the Double GG model can be identified using the following equations
\begin{align}\label{eq9}
&{{\Omega }_{i}}={{\left( \frac{\Gamma \left( {{\beta }_{i}} \right)}{\Gamma \left( {{\beta }_{i}}+1/{{\gamma }_{i}} \right)} \right)}^{{{\gamma }_{i}}}}{{\beta }_{i}},~~i=1,2\\\label{eq8a}
\newsubeqblock
\mysubeq &\sigma _{x}^{2}=\frac{\Gamma \left( {{\beta }_{1}}+{2}/{{{\gamma }_{1}}}\; \right)\Gamma \left( {{\beta }_{1}} \right)}{{{\Gamma }^{2}}\left( {{\beta }_{1}}+{1}/{{{\gamma }_{1}}}\; \right)}-1\\\label{eq8b}
\mysubeq &\sigma _{y}^{2}=\frac{\Gamma \left( {{\beta }_{2}}+{2}/{{{\gamma }_{2}}}\; \right)\Gamma \left( {{\beta }_{2}} \right)}{{{\Gamma }^{2}}\left( {{\beta }_{2}}+{1}/{{{\gamma }_{2}}}\; \right)}-1
\end{align}
where $\sigma _{x}^{2}$ and $\sigma _{y}^{2}$ are respectively normalized variances of small and large scale irradiance fluctuations.
The cumulative distribution function (cdf) of Double GG distribution can be derived from (\ref{eq4}) as \cite{kashani}
\begin{align}\label{eq5}
&{{F}_{I}}\left( I \right)=\frac{{{p}^{{{\beta}_{2}}-1/2}}{{q}^{{{\beta}_{1}}-1/2}}{{\left( 2\pi \right)}^{{1-\left( p+q \right)}/{2}\;}}}{\Gamma \left( {{\beta}_{1}} \right)\Gamma \left( {{\beta}_{2}} \right)}\\\nonumber
&\times G_{1,p+q+1}^{p+q,1}\left[ {{\left( \frac{{{I}^{{{\gamma }_{2}}}}}{{{\Omega }_{2}}} \right)}^{p}}\frac{\beta_{1}^{q}\beta_{2}^{p}}{{{p}^{p}}{{q}^{q}}\Omega _{1}^{q}}|\begin{matrix}
1 \\
\Delta \left( q:{{\beta}_{1}} \right),\Delta \left( p:{{\beta}_{2}} \right),0 \\
\end{matrix} \right]
\end{align}
\section{BER Performance}
The optimum decision metric for OOK is given by \cite{21}
\begin{equation}\label{eq10}
P\left( \mathbf{r}|\text{on,}{{\text{I}}_{mn}} \right)\underset{\text{off}}{\mathop{\overset{\text{on}}{\mathop{\lessgtr }}\,}}\,P\left( \mathbf{r}|\text{off,}{{\text{I}}_{mn}} \right)
\end{equation}
where $\mathbf{r}=\left( {{r}_{1}},{{r}_{2}},...,{{r}_{N}} \right)$ is the received signal vector. Following the same approach as \cite{20,21}, the conditional bit error probabilities are given by (see \cite{20} for details of derivation)
\begin{align}\nonumber
&{{P}_{e}}(\text{off }|{{I}_{mn}})={{P}_{e}}(\text{on }|{{I}_{mn}})\\\label{eq20}
&=Q\left( \frac{1}{MN}\sqrt{\frac{{\bar{\gamma }}}{2}\sum\limits_{n=1}^{N}{{{\left( \sum\limits_{m=1}^{M}{{{I}_{mn}}} \right)}^{2}}}} \right)
\end{align}
Therefore, the average error rate can be expressed as
\begin{equation}\label{eq21}
{{P}_{\text{MIMO}}}=\int\limits_{\mathbf{I}}{{{f}_{\mathbf{I}}}}\left( \mathbf{I} \right)Q\left( \frac{1}{MN}\sqrt{\frac{{\bar{\gamma }}}{2}\sum\limits_{n=1}^{N}{{{\left( \sum\limits_{m=1}^{M}{{{I}_{mn}}} \right)}^{2}}}} \right)d\mathbf{I}
\end{equation}
where ${{f}_{\mathbf{I}}}\left( \mathbf{I} \right)$ is the joint pdf of vector $\mathbf{I}=\left( {{I}_{11}},{{I}_{12}},\ldots ,{{I}_{MN}} \right)$. The factor $M$ in (\ref{eq21}) ensures that the total transmitted powers of diversity system and SISO link are equal for a fair comparison. On the other hand, the factor $N$ is used to ensure that sum of the $N$ receive aperture areas is the same as the area of the receive aperture of the SISO link. The integral expressed in (\ref{eq21}) does not yield a closed-form solution even for simpler turbulence distributions; however, it can be calculated through numerical multi-dimensional integration. Similarly, we can use multidimensional Gaussian quadrature rule (GQR) \cite{handbook} techniques to calculate the BER for MISO case, i.e.,
\begin{equation}\label{eq22}
{{P}_{\text{MISO}}}=\int\limits_{\mathbf{I}}{{{f}_{\mathbf{I}}}}\left( \mathbf{I} \right)Q\left( \frac{\sqrt{{\bar{\gamma }}}}{M\sqrt{2}}\sum\limits_{m=1}^{M}{{{I}_{m}}} \right)d\mathbf{I}
\end{equation}
which does not yield a closed form expression either.
In the following, we focus on the SIMO case and investigate the BER performance under the assumption of optimal combining (OC) with perfect CSI where the variance of the noise in each receiver is given by $\sigma _{n}^{2}={{N}_{0}}/2N$. Therefore, replacing $M=1$ in (\ref{eq21}) we obtain
\begin{equation}\label{eq23}
{{P}_{\text{SIMO,OC}}}=\int\limits_{\mathbf{I}}{{{f}_{\mathbf{I}}}}\left( \mathbf{I} \right)Q\left( \sqrt{\frac{{\bar{\gamma }}}{2N}\sum\limits_{n=1}^{N}{I_{n}^{2}}} \right)d\mathbf{I}
\end{equation}
Eq. (\ref{eq23}) does not yield a closed-form solution and requires N-dimensional integration. Nevertheless, the Q-function can be well-approximated as $Q(x)\approx {{{e}^{-\frac{{{x}^{2}}}{2}}}}/{12}\;+{{{e}^{-\frac{2{{x}^{2}}}{3}}}}/{4}\;$ \cite{30}, and thus the average BER can be obtained as
\begin{align}\nonumber
{{P}_{\text{SIMO,OC}}}&\approx \frac{1}{12}\prod\limits_{n=1}^{N}{\int_{0}^{\infty }{{{f}_{{{I}_{n}}}}\left( {{I}_{n}} \right)}}\exp \left( \frac{-\bar{\gamma }}{4N}I_{n}^{2} \right)d{{I}_{n}}\\\label{eq24}
&+\frac{1}{4}\prod\limits_{n=1}^{N}{\int_{0}^{\infty }{{{f}_{{{I}_{n}}}}\left( {{I}_{n}} \right)}}\exp \left( \frac{-\bar{\gamma }}{3N}I_{n}^{2} \right)d{{I}_{n}}
\end{align}
The above integral can be evaluated by first expressing the exponential function in terms of the Meijer G-function presented in [\citen{29}, eq. (11)] as
\begin{equation}\label{eq25}
\exp \left( -x \right)=\operatorname{G}_{0,1}^{1,0}\left[ x\left| _{0}^{-} \right. \right]
\end{equation}
Then, a closed-form expression for (\ref{eq24}) is obtained using [\citen{29}, Eq. (21)] as
\begin{equation}\label{eq26}
{{P}_{\operatorname{SIMO},OC}}\approx \frac{1}{12}\prod\limits_{n=1}^{N}{\Lambda \left( n,4 \right)}+\frac{1}{4}\prod\limits_{n=1}^{N}{\Lambda \left( n,3 \right)}
\end{equation}
where $\Lambda \left( n,\upsilon \right)$ is defined in (\ref{eq27}) at the top of the next page.
\begin{figure*}[t]
\begin{equation}\label{eq27}
\Lambda \left( n,\upsilon \right)=\frac{{{\alpha }_{n}}l_{n}^{-0.5}k_{n}^{{{\beta}_{1,n}}+{{\beta}_{2,n}}}}{2{{\left( 2\pi \right)}^{0.5\left( {{l}_{n}}-1+\left( {{k}_{n}}-1 \right)\left( {{p}_{n}}+{{q}_{n}} \right) \right)}}}G_{{{l}_{n}},{{k}_{n}}\left( {{p}_{n}}+{{q}_{n}} \right)}^{{{k}_{n}}\left( {{p}_{n}}+{{q}_{n}} \right),{{l}_{n}}}\left[ \frac{{{\left( \upsilon N \right)}^{{{l}_{n}}}}\omega _{n}^{-{{k}_{n}}}l_{n}^{{{l}_{n}}}}{{{{\bar{\gamma }}}^{{{l}_{n}}}}k_{n}^{{{k}_{n}}\left( {{p}_{n}}+{{q}_{n}} \right)}}\left| \begin{matrix}
\Delta \left( {{l}_{n}},1 \right) \\
{{J}_{{{k}_{n}}}}\left( {{q}_{n}},1-{{\beta}_{1,n}} \right),{{J}_{{{k}_{n}}}}\left( {{p}_{n}},1-{{\beta}_{2,n}} \right) \\
\end{matrix} \right. \right]
\end{equation}
\hrulefill
\end{figure*}
In (\ref{eq27}), ${{l}_{n}}$ and ${{k}_{n}}$ are positive integer numbers that satisfy ${{{p}_{n}}{{\gamma }_{2,n}}}/{2}\;={{{l}_{n}}}/{{{k}_{n}}}\;$, and ${{\operatorname{J}}_{\xi }}\left( y,x \right)$, ${{\alpha }_{n}}$ and ${{\omega }_{n}}$, $n\in \left\{ 1,2,\ldots ,N \right\}$, are defined as
\begin{align}\label{eq28c}
&{{\operatorname{J}}_{\xi }}\left( y,x \right)\\\nonumber
&=\Delta \left( \xi ,\frac{y-x}{y} \right),\Delta \left( \xi ,\frac{y-1-x}{y} \right),\ldots ,\Delta \left( \xi ,\frac{1-x}{y} \right)\\\label{eq28}
&{{\alpha }_{n}}=\frac{{{\gamma }_{2,n}}p_{n}^{{{\beta}_{2,n}}+1/2}q_{n}^{{{\beta}_{1,n}}-1/2}{{\left( 2\pi \right)}^{{1-\left( {{p}_{n}}+{{q}_{n}} \right)}/{2}\;}}}{\Gamma \left( {{\beta}_{1,n}} \right)\Gamma \left( {{\beta}_{2,n}} \right)}\\\label{eq28b}
&{{\omega }_{n}}={{\left( {{\Omega }_{2,n}}{{p}_{n}}\beta_{2,n}^{-1} \right)}^{{{p}_{n}}}}{{\left( {{q}_{n}}{{\Omega }_{1,n}}\beta_{1,n}^{-1} \right)}^{{{q}_{n}}}}
\end{align}
The derived BER expression in (\ref{eq26}) for SIMO FSO system with OC can be seen as a generalization of BER results over other atmospheric turbulence models. For example, if we insert ${{\gamma }_{i}}=1$ and ${{\Omega }_{i}}=1$ in (\ref{eq26}), we obtain the BER expression over Gamma-Gamma channel. Setting ${{\beta }_{i}}=1$ in (\ref{eq26}), we obtain the BER for Double Weibull channel. On the other hand, for ${{\gamma }_{i}}=1$, ${{\Omega }_{i}}=1$ and ${{\beta }_{2}}=1$, (\ref{eq26}) reduces to (12) of \cite{20} reported for the K-channel. Appendix provides the details on these.
As an alternative to OC, we also consider equal gain combining (EGC) where the receiver adds the receiver branches. In this case, the average BER can be expressed as
\begin{equation}\label{eq29}
{{P}_{\text{SIMO,ECG}}}=\int\limits_{\mathbf{I}}{{{f}_{\mathbf{I}}}}\left( \mathbf{I} \right)Q\left( \frac{\sqrt{{\bar{\gamma }}}}{N\sqrt{2}}\sum\limits_{n=1}^{N}{{{I}_{n}}} \right)d\mathbf{I}
\end{equation}
It should be noted that (\ref{eq29}) is equivalent to (\ref{eq22}) obtained for the MISO FSO links. Another method is selection combining (SC) which is the least complicated of the combining schemes since it only processes one of the diversity apertures. Specifically, the SC chooses the aperture with the maximum received irradiance (or electrical SNR). Therefore, the pdf of the output of SC receiver can be obtained as
\begin{align}\nonumber
{{f}_{{{I}_{\max }}}}\left( {{I}_{\max }} \right)&=\frac{d{{F}_{{{I}_{\max }}}}\left( {{I}_{\max }} \right)}{d{{I}_{\max }}}\\\label{eq30}
&=\sum\limits_{n=1}^{N}{\prod\limits_{k=1,k\ne n}^{N}{{{f}_{{{I}_{n}}}}\left( {{I}_{\max }} \right)}}{{F}_{{{I}_{k}}}}\left( {{I}_{\max }} \right)
\end{align}
The average BER can be then expressed as
\begin{align}\nonumber
{{P}_{\text{SC}}}&=\sum\limits_{n=1}^{N}{\prod\limits_{k=1,k\ne n}^{N}{\int_{0}^{\infty }{{{f}_{{{I}_{n}}}}\left( {{I}_{\max }} \right)}}}{{F}_{{{I}_{k}}}}\left( {{I}_{\max }} \right)\\\label{eq31}
&\times \operatorname{Q}\left( {{I}_{\max }}\sqrt{\frac{{\bar{\gamma }}}{2N}} \right)d{{I}_{\max }}
\end{align}
which can be efficiently calculated through numerical means.
\begin{figure}[t]
\centering
\includegraphics[width = 8cm, height = 7.5cm]{f3.eps}
\caption{ Comparison of the average BER between SISO and different diversity techniques for plane wave assuming i.i.d. turbulent channel defined as channel $b$.}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width = 8cm, height = 7.5cm]{f2.eps}
\caption{ Comparison of the average BER between SISO and different diversity techniques for plane wave assuming i.i.d. turbulent channel defined as channel $c$.}
\end{figure}
\section{Numerical Results}
In this section, we present BER performance results of SIMO, MISO and MIMO FSO systems over Double GG channels and quantify the performance improvements over SISO systems. We consider the following four scenarios of atmospheric turbulence conditions reported in \cite{kashani}.
\begin{itemize}
\item \textbf{Channel a}: \emph{Plane wave and moderate irradiance fluctuations} with ${{\gamma }_{1}}=2.1690$, ${{\gamma }_{2}}=0.8530$, ${{m}_{1}}=0.55$, ${{m}_{2}}=2.35$, ${{\Omega }_{1}}=1.5793$, ${{\Omega }_{2}}=0.9671$, $p=28$ and $q=11$
\item \textbf{Channel b}: \emph{Plane wave and strong irradiance fluctuations} with ${{\gamma }_{1}}=1.8621$, ${{\gamma }_{2}}=0.7638$, ${{m}_{1}}=0.5$, ${{m}_{2}}=1.8$, ${{\Omega }_{1}}=1.5074$, ${{\Omega }_{2}}=0.9280$, $p=17$ and $q=7$.
\item \textbf{Channel c}: \emph{Spherical wave and moderate irradiance fluctuations} with ${{\gamma }_{1}}=0.9135$, ${{\gamma }_{2}}=1.4385$, ${{m}_{1}}=2.65$, ${{m}_{2}}=0.85$, ${{\Omega }_{1}}=0.9836$ and ${{\Omega }_{2}}=1.1745$, $p=7$ and $q=11$.
\item \textbf{Channel d}: \emph{Spherical wave and strong irradiance fluctuations} with ${{\gamma }_{1}}=0.4205$, ${{\gamma }_{2}}=0.6643$, ${{m}_{1}}=3.2$, ${{m}_{2}}=2.8$, ${\Omega_{1}}=0.8336$ and ${{\Omega }_{2}}=0.9224$, $p=7$ and $q=11$.
\end{itemize}
Figs. 1-2 present the average BER over i.i.d. channels defined as channel $b$ and channel $c$ respectively. For SIMO FSO links employing OC receivers, we present approximate analytical results which have been obtained through (\ref{eq26}) along with the Monte-Carlo simulation of (\ref{eq23}). As clearly seen from Figs. 1-2, our approximate expressions provide an excellent match to simulation results. As a benchmark, the average BER of SISO FSO link is also included in these figures. It is observed that multiple receive aperture deployment significantly improves the performance. For instance, at a target bit error rate of ${{10}^{-5}}$, we observe performance improvements of 47.2 dB and 67.2 dB for SIMO FSO links with $N=2$ and 3 receive apertures employing OC with respect to the SISO transmission over channel $b$. Similarly, for channel $c$, at a BER of ${{10}^{-5}}$, impressive performance improvements of 51.5 dB and 64.3 dB are achieved for SIMO links with $N=2$ and 3 employing OC compared to the SISO deployment. It is also illustrated that EGC receivers yield nearly the same performance as OC receivers. For example, in Fig 2, for $N=2$ the performance difference between OC and EGC receivers is merely 0.4 dB at a BER of ${{10}^{-5}}$. Also as expected, EGC and OC receivers outperform SC counterpart.
Figs. 3-4 demonstrate the BER performance of SIMO FSO links employing OC, EGC and SC receivers over non-identically distributed (i.n.i.d.) Double GG channels. Similar to i.i.d. results, our approximate closed-form expressions again yield nearly identical match to simulation results. We further compare the performance of i.n.i.d. case with respect to i.i.d. case presented in Figs1-2. For example, to achieve a BER of ${{10}^{-5}}$ in SIMO links with $N=2$ over i.n.i.d. channels $a$ and $b$, we need 8.2 dB, 8.5 dB and 4.9 dB less in comparison to i.i.d. case respectively for OC, EGC and SC receivers. Note that in Fig. 1, we assume that both of the two channels between the transmitter and receivers are described by channel $b$. Thus, since the channel $a$ is less severe than the channel $b$, we need less SNR in comparison to i.i.d case to obtain the same BER. On the other hand, to achieve a BER of ${{10}^{-5}}$ for SIMO links with $N=2$ over i.n.i.d. channels $c$ and $d$, we need 6.8 dB more for OC and EGC receivers and 11.6 dB more for SC receiver in comparison to i.i.d channels. Note that in Fig. 2, both of the two channels between the transmitter and receivers are described by channel $c$. Therefore, as the channel $d$ is more severe than the channel $c$, we need more SNR in comparison to i.i.d case to achieve the same performance.
\begin{figure}
\centering
\includegraphics[width = 8cm, height = 7.5cm]{f1f2.eps}
\caption{Comparison of the OC, EGC and SC receivers for SIMO FSO links over two i.n.i.d. atmospheric turbulence channels defined as channel $a$ and channel $b$.}
\end{figure}
\begin{figure}
\centering
\includegraphics[width = 8cm, height = 7.5cm]{f3f4.eps}
\caption{Comparison of the OC, EGC and SC receivers for SIMO FSO links over two i.n.i.d. atmospheric turbulence channels defined as channel $c$ and channel $d$.}
\end{figure}
\section{Conclusions}\label{Con}
In this paper, we have investigated the BER performance of FSO links with spatial diversity over atmospheric turbulence channels described by the Double GG distribution. We have obtained an efficient and unified closed-form expression for the BER of SIMO FSO systems with OC receiver which generalizes existing results as special cases. For MISO and MIMO systems, we have presented BER performance based on numerical calculations of the integral expressions. Our numerical results have demonstrated that spatial diversity schemes can significantly improve the system performance and bring impressive performance gains over SISO systems. Our comparisons among SIMO FSO links employing OC, EGC and SC receivers have further demonstrated that EGC scheme presents a favorable trade-off between complexity and performance.
\section*{Appendix}\label{Appendix D}
\noindent\textbf{Proposition 1}: \emph{Inserting ${{\gamma }_{i}}=1$, ${{\Omega }_{i}}=1$ and ${{\beta }_{1}}=1$ in (\ref{eq26}), the BER of SIMO FSO links using optimal combining over K-channel is obtained.}
\noindent\textbf{Proof}: First by replacing ${{\gamma }_{i}}={{\Omega }_{i}}=p=q={{\beta }_{1}}=1$ in (\ref{eq28}) and (\ref{eq28b}), we obtain ${{\alpha }_{n}}={1}/{\Gamma \left( {{\beta }_{2,n}} \right)}$ and ${{\omega }_{n}}=\beta _{2,n}^{-1}$. Then by plugging all the values in (\ref{eq27}), we obtain
\begin{align}\nonumber
& {{\Lambda }_{KC}}\left( n,\upsilon \right)\\
&=\frac{{{2}^{{{\beta}_{2,n}}-1}}}{\pi \Gamma \left( {{\beta}_{2,n}} \right)}G_{1,4}^{4,1}\left[ \frac{\left( \upsilon N \right)\beta_{2,n}^{2}}{16\bar{\gamma }}\left| \begin{matrix}
1 \\
\frac{1}{2},1,\frac{{{\beta}_{_{2,n}}}}{2},\frac{{{\beta}_{_{2,n}}}+1}{2} \\
\end{matrix} \right. \right] \\\nonumber
&=\frac{{{2}^{{{\beta}_{2,n}}-1}}}{\pi \Gamma \left( {{\beta}_{2,n}} \right)}G_{4,1}^{1,4}\left[ \frac{16\bar{\gamma }}{\upsilon N\beta_{2,n}^{2}}\left| \begin{matrix}
\frac{1}{2},0,\frac{2-{{\beta}_{_{2,n}}}}{2},\frac{1-{{\beta}_{_{2,n}}}}{2} \\
0 \\
\end{matrix} \right. \right]
\end{align}
Therefore,
\begin{equation}
{{P}_{\operatorname{SIMO},OC\_KC}}\approx \frac{1}{12}\prod\limits_{n=1}^{N}{{{\Lambda }_{KC}}\left( n,4 \right)}+\frac{1}{4}\prod\limits_{n=1}^{N}{{{\Lambda }_{KC}}\left( n,3 \right)}
\end{equation}
which coincides with (21) of \cite{20}.
\noindent\textbf{Proposition 2}: \emph{Inserting ${{\gamma }_{i}}=1$, ${{\Omega }_{i}}=1$ in (\ref{eq26}), the BER of SIMO FSO links using optimal combining over Gamma-Gamma channel is obtained.}
\noindent\textbf{Proof}: First by replacing ${{\gamma }_{i}}={{\Omega }_{i}}=p=q=1$ in (\ref{eq28}) and (\ref{eq28b}), we obtain ${{\alpha }_{n}}={1}/{\Gamma \left( {{\beta }_{1,n}} \right)\Gamma \left( {{\beta }_{2,n}} \right)}$ and ${{\omega }_{n}}=\beta _{1,n}^{-1}\beta _{2,n}^{-1}$. Then by inserting all the values as well as $l=1$ and $k=2$ in (\ref{eq27}), we obtain
\begin{align}\nonumber
&{{\Lambda }_{GG}}\left( n,\upsilon \right)=\frac{{{2}^{{{\beta }_{1,n}}+{{\beta }_{2,n}}-2}}}{\pi \Gamma \left( {{\beta }_{1,n}} \right)\Gamma \left( {{\beta }_{2,n}} \right)}\\
&\times G_{1,4}^{4,1}\left[ \frac{\upsilon N{{\left( {{\beta }_{2,n}}{{\beta }_{2,n}} \right)}^{2}}}{16\bar{\gamma }}\left| \begin{matrix}
1 \\
\frac{{{\beta }_{_{1,n}}}}{2},\frac{{{\beta }_{_{1,n}}}+1}{2}\frac{{{\beta }_{_{2,n}}}}{2},\frac{{{\beta }_{_{2,n}}}+1}{2} \\
\end{matrix} \right. \right]
\end{align}
Therefore,
\begin{equation}
{{P}_{\operatorname{SIMO},OC\_GG}}\approx \frac{1}{12}\prod\limits_{n=1}^{N}{{{\Lambda }_{GG}}\left( n,4 \right)}+\frac{1}{4}\prod\limits_{n=1}^{N}{{{\Lambda }_{GG}}\left( n,3 \right)}
\end{equation}
\noindent\textbf{Proposition 3}: \emph{Inserting ${{\beta }_{i}}=1$ in (\ref{eq26}), the BER of SIMO FSO links using optimal combining over Double-Weibull channel is obtained.}
\noindent\textbf{Proof}: Inserting ${{\beta }_{i}}=1$in (\ref{eq28}) and (\ref{eq28b}), we obtain
\begin{align}
&{{\alpha }_{n}}={{\gamma }_{2,n}}p_{n}^{3/2}q_{n}^{1/2}{{\left( 2\pi \right)}^{1-\left( {{p}_{n}}+{{q}_{n}} \right)/2\ }}\\
&{{\omega }_{n}}={{\left( {{\Omega }_{2,n}}{{p}_{n}} \right)}^{{{p}_{n}}}}{{\left( {{q}_{n}}{{\Omega }_{1,n}} \right)}^{{{q}_{n}}}}
\end{align}
Then by plugging all the values in (\ref{eq27}), we obtain
\begin{align}\nonumber
&{{\Lambda }_{DW}}\left( n,\upsilon \right)=\frac{{{\gamma }_{2,n}}p_{n}^{3/2}q_{n}^{1/2}l_{n}^{-0.5}k_{n}^{2}}{2{{\left( 2\pi \right)}^{0.5\left( {{l}_{n}}-3+{{k}_{n}}\left( {{p}_{n}}+{{q}_{n}} \right) \right)}}}\\\nonumber
&\times G_{{{l}_{n}},{{k}_{n}}\left( {{p}_{n}}+{{q}_{n}} \right)}^{{{k}_{n}}\left( {{p}_{n}}+{{q}_{n}} \right),{{l}_{n}}}\left[ \frac{{{\left( \upsilon N \right)}^{{{l}_{n}}}}}{{{\left( {{\Omega }_{2,n}}{{p}_{n}}{{k}_{n}} \right)}^{{{k}_{n}}{{p}_{n}}}}}\right.\\\label{dw}
&\times \left.\frac{{{\left( {{{\bar{\gamma }}}^{-1}}{{l}_{n}} \right)}^{{{l}_{n}}}}}{{{\left( {{q}_{n}}{{\Omega }_{1,n}}{{k}_{n}} \right)}^{{{k}_{n}}{{q}_{n}}}}}\left| \begin{matrix}
\Delta \left( {{l}_{n}},1 \right) \\
{{J}_{{{k}_{n}}}}\left( {{q}_{n}},0 \right),{{J}_{{{k}_{n}}}}\left( {{p}_{n}},0 \right) \\
\end{matrix} \right. \right]
\end{align}
Thus,
\begin{equation}
{{P}_{\operatorname{SIMO},OC\_DW}}\approx \frac{1}{12}\prod\limits_{n=1}^{N}{{{\Lambda }_{DW}}\left( n,4 \right)}+\frac{1}{4}\prod\limits_{n=1}^{N}{{{\Lambda }_{DW}}\left( n,3 \right)}
\end{equation}
\balance
\bibliographystyle{IEEEtran}
|
\section{Appendix: Proofs of Section \ref{s:tr}}\label{app:proofs:s:tr}
\input{_Proof_3Rate.tex}
\section{Appendix: Proofs of Section \ref{s:ad}}\label{app:proofs:s:ad}
\input{_Proof_4Adaptive.tex}
\section{Introduction}\label{s:in}
Accounting for the fact that inverse problems are widely used in many fields of science,
there has been over the last decades a growing interest in statistical
inverse problems (see, \textit{e.g.}, \cite{KorostelevTsybakov1993},
\cite{MairRuymgaart1996}, \cite{EvansStark2002}, \cite{KaipioSomersalo2005}, \cite{BissantzHohageMunkRuymgaart2007} and references therein).
Mathematical statistics has paid special attention to oracle or
minimax optimal nonparametric estimation and adaptation in the framework of
inverse problems (see \cite{EfromovichKoltchinskii2001}, \cite{CavalierGolubevLepskiTsybakov2003},
\cite{Cavalier2008} and \cite{HoffmannReiss2008}, to name but a few). Nonparametric estimation in general requires
to choose a tuning parameter which is challenging in practise. Oracle
and minimax estimation is achieved, respectively, if the tuning
parameter is set to an optimal value which relies either on a
knowledge of the unknown parameter of interest or of certain
characteristics of it (such as smoothness). Since both the parameter and its smoothness are unknown, it
is necessary to design a feasible procedure to select the tuning
parameter that adapts to the unknown underlying function or to its
regularity and achieves the oracle or minimax rate. Among the most
prominent approaches stand without doubts model selection
(cf. \cite{BarronBirgeMassart1999} and its exhaustive discussion in \cite{Massart07}),
Stein's unbiased risk estimation and its extensions (cf. \cite{CavalierGolubevPicardTsybakov2002},
\cite{CavalierGolubevPicardTsybakov2002} or
\cite{CavalierHengartner2005}),
Lepski's method (see, e.g., \cite{Lepskij1990},
\cite{Birge2001}, \cite{EfromovichKoltchinskii2001} or \cite{Mathe2006})
or combinations of the aforementioned strategies (cf. \cite{GoldenshlugerLepski2011} and \cite{ComteJohannes2012}).
On the other hand side, it seems natural to adopt a Bayesian point of view where the tuning parameter can be
endowed with a prior. As the theory for a general inverse problem --
with a possibly unknown or noisy operator -- is technically highly
involved, we consider in this paper as a starting point an indirect Gaussian regression which
is well known to be equivalent to an indirect Gaussian sequence
space model (in a \cite{LeCam1964} sense, see, e.g., \cite{BrownLow1996a}
for the direct case and \cite{Meister2011} for the indirect case).\\
Let $\Hspace$ be the Hilbert space of square summable real valued sequences endowed with the usual inner
product $\Hskalar$ and associated norm $\Hnorm$. In an indirect Gaussian
sequence space model (iGSSM) one aim is to recover a parameter sequence
$\So=\suite{\So}\in\Hspace$ from a transformed version
$(\Ev_j\So_j)_{j\geq1}$ that is blurred by a Gaussian white
noise. Precisely, an observable sequence of random variables
$(\ObSo)_{j\geq1}$, $\ObSo$ for short,
obeys an indirect Gaussian sequence space model, if
\begin{equation}
\label{in:de:mod}
\ObSo_j = \Ev_j\So_j + \sqrt{\ObSoNoL}\ObSoNo_j ,\qquad j\in\Nz,
\end{equation}
\noindent where $\{\ObSoNo_j\}_{j\geq1}$ are unobservable error terms,
which are independent and standard normally distributed, and
$0<\ObSoNoL<1$ is the noise level. The sequence
$\Ev=\suite{\Ev}$ represents the operator that
transforms the signal $\So$. In the particular case of a constant
sequence $\Ev$ the sequence space model is called direct while it
is called an indirect sequence space model if the sequence $\Ev$
tends to zero. We assume throughout the paper that the sequence is bounded.
In this paper we adopt a Bayesian approach, where the parameter
sequence of interest $\So=(\So_j)_{j\geq1}$ itself is a realisation of a
random variable $\RvSo=(\RvSo_j)_{j\geq1}$ and the observable random
variable $\ObSo=(\ObSo_j)_{j\geq1}$ satisfies
\begin{equation}
\label{in:de:ob}
\ObSo_j = \Ev_j\RvSo_j + \sqrt{\ObSoNoL}\ObSoNo_j,\quad j\in\Nz
\end{equation}
with independent and standard normally distributed error terms
$\{\ObSoNo_j\}_{j\geq1}$ and noise level $0<\ObSoNoL<1$. Throughout the paper we assume that random parameters
$\{\RvSo_j\}_{j\geq1}$ and the error terms $\{\ObSoNo_j\}_{j\geq1}$
are independent. Consequently, \eqref{in:de:ob} and a specification of the prior
distribution $P_{\RvSo}$ of $\RvSo$ determine completely the joint distribution of $\ObSo$
and $\RvSo$. For a broader overview on Bayesian procedures we refer the reader to the
monograph by \cite{Robert2007}.\\
Typical prior specifications studied in the
direct sequence space model literature are compound priors, also known as Sieve priors (see,
e.g., \cite{Zhao2000}, \cite{ShenWasserman2001} or
\cite{ArbelGayraudRousseau2013}, Gaussian series priors (cf. \cite{freedman1999}, \cite{Cox1993} or \cite{Castillo2008}), block
priors (cf. \cite{GaoZhou2014}), countable mixture of normal priors
(cf. \cite{BelitserGhosal2003}) and finite mixtures of normal and
Dirac priors (\textit{e.g.} \cite{Abramovichetal1998}). In the context
of an iGSSM,
\cite{KnapikVanderVaartVanZanten2011} and
\cite{KnapikSzaboVanderVaartVanZanten2014} consider Gaussian series
priors and continuous mixture of Gaussian series priors, respectively. \\
By considering an iGSSM we derive in this paper theoretical properties of a Bayes
procedure with a Sieve prior specification from a frequentist point of
view, meaning that there exists a true parameter value $\TrSo =
(\TrSo[j])_{j\geq 1}$ associated with the data generating process of
$\suite{\ObSo}$. A broader overview of frequentist asymptotic
properties of nonparametric Bayes procedures can be found, for example,
in \cite{GhoshRamamoorthi2003}, while direct and indirect models, respectively, are considered by
e.g., \cite{Zhao2000}, \cite{BelitserGhosal2003},
\cite{Castillo2008} and \cite{GaoZhou2014}, and, e.g.,
\cite{KnapikVanderVaartVanZanten2011} and
\cite{KnapikSzaboVanderVaartVanZanten2014}. Bayesian procedures in the context of
slightly different Gaussian inverse problems and their asymptotic
properties are studied in, e.g., \cite{AgapiouLarssonStuart2013} and \cite{FlorensSimoni2010}.
However, our special attention is given to posterior consistency and optimal posterior concentration in
an oracle or minimax sense, which we
elaborate in the following.\\
In this paper we consider a sieve prior family $\{P_{\DiRvSo}\}_{\Di}$
where the prior distribution $P_{\DiRvSo}$ of the random parameter
sequence $\DiRvSo=(\DiRvSo_j)_{j\geq1}$ is Gaussian and degenerated
for all $j>m$. More precisely, the first $m$ coordinates
$\{\DiRvSo_j\}_{j=1}^{\Di}$ are independent and normally distributed
random variables while the remaining coordinates
$\{\DiRvSo_j\}_{j>\Di}$ are degenerated at a point. Note that the dimension
parameter $\Di$ plays the role of a tuning parameter. Assuming an
observation $\ObSo=(\ObSo_j)_{j\geq1}$ satisfying
$\ObSo_j=\DiRvSo_j+\sqrt{\ObSoNoL}\ObSoNo_j$, we denote by
$P_{\DiRvSo|\ObSo}$ the corresponding posterior distribution of
$\DiRvSo$ given $\ObSo$. Given a prior sub-family
$\{P_{\DiRvSo[\eDi]}\}_{\eDi}$ in dependence of the noise level
$\ObSoNoL$, our objective is the study of frequentist properties of the
associated posterior sub-family $\{P_{\DiRvSo[\eDi]|\ObSo}\}_{\eDi}$.
To be more precise, let $\TrSo$ be the realization of the random
parameter $\RvSo$ associated with the data-generating distribution and
denote by $\Ex_{\TrSo}$ the corresponding expectation. A
quantity $\eRa$ which is up to a constant a lower and an upper bound of
the concentration of the posterior sub-family $\{P_{\DiRvSo[\eDi]|\ObSo}\}_{\eDi}$, i.e.,
\begin{equation}\label{in:de:ra}
\lim_{\ObSoNoL\to0}\Ex_{\TrSo}P_{\DiRvSo[\eDi]|\ObSo}((K)^{-1}\,
\eRa \leq \HnormV{\DiRvSo[\eDi]-\TrSo}^2\leq K\, \eRa)=1
\quad\text{with $1\leq K<\infty$},\hfill
\end{equation}
is called exact posterior
concentration (see, e.g., \cite{BarronBirgeMassart1999},
\cite{GhosalGhoshVanDerVaart2000} or \cite{Castillo2008} for a broader
discussion of the concept of posterior concentration). We shall
emphasise that the derivation of the posterior concentration relies strongly
on tail bounds for non-central $\chi^2$ distributions established
in \cite{Birge2001}. Moreover, if $\eRa\to0$ as $\ObSoNoL\to 0$ then the lower and
upper bound given in \eqref{in:de:ra} establish posterior consistency
and $\eRa$ is called exact posterior concentration rate.
Obviously, the exact rate depends on the prior sub-family
$\{P_{\DiRvSo[\eDi]}\}_{\eDi}$ as well as on the
unknown parameter $\TrSo$. \\In the spirit of a frequentist oracle
approach, given a parameter $\TrSo$ we derive in this paper
a prior sub-family $\{P_{\DiRvSo[\treDi]}\}_{\treDi}$ with smallest
possible exact posterior concentration rate $\treRa$ which we
call, respectively, an oracle prior sub-family and an oracle posterior concentration rate. On the other hand side, following a minimax approach, \cite{JohannesSchwarz2013},
for example, derive the minimax rate of convergence $\oeRa$ of the
maximal mean integrated squared error (MISE) over a given class $\cwSo$ of parameters (introduced below).
We construct a sub-family $\{P_{\DiRvSo[\oeDi]}\}_{\oeDi}$ of prior
distributions with exact posterior concentration rate $\oeRa$
uniformly over $\cwSo$ which does not depend on the true parameter $\TrSo$ but only on the set of possible
parameters $\cwSo$. It is interesting to note that in a direct GSSM
\cite{Castillo2008} establishes up to a constant the minimax-rate as
an upper bound of the posterior concentration, while
the derived lower bound features a logarithmic factor compared to the
minimax rate. \cite{ArbelGayraudRousseau2013}, for example,
in a direct GSSM and \cite{KnapikSzaboVanderVaartVanZanten2014} in an
indirect GSSM provide only upper bounds of the posterior
concentration rate which differ up to a logarithmic factor from the
minimax rate.
We shall emphasize, that the prior specifications we propose in this
paper lead to exact posterior concentration rates that are optimal
in an oracle or minimax sense over certain classes of parameters not only in the direct model but also in the
more general indirect model. However, both oracle and minimax sieve prior are unfeasible in
practise since they rely on the knowledge of either $\TrSo$ itself or
its smoothness.
Our main contribution in this paper is the construction of a
hierarchical prior $P_{\RvDiRvSo}$ that is adaptive. Meaning that, given a
parameter $\TrSo\in\Hspace$ or a classes $\cwSo\subset\Hspace$ of
parameters, the posterior distribution $P_{\RvDiRvSo|\ObSo}$ contracts,
respectively, at the oracle rate or the minimax rate over $\cwSo$
while the hierarchical prior $P_{\RvDiRvSo}$ does not rely neither on the
knowledge of $\TrSo$ nor the class $\cwSo$. Let us briefly elaborate on
the hierarchical structure of the prior which induces an additional prior on the
tuning parameter $\Di$, i.e., $\Di$ itself is a realisation of a
random variable $\RvDi$. We construct a prior for $\RvDi$ such
that the marginal posterior for $\RvDiRvSo$ (obtained by integrating out
$\RvDi$ with respect to its posterior) contracts exactly at the
oracle concentration rate. This is possible for every $\TrSo$ whose
components differ from the components of the prior mean infinitely many times. In addition, for
every $\TrSo$ in the class $\cwSo$ we show that the posterior
distribution $P_{\RvDiRvSo|\ObSo}$ contracts at least at the minimax
rate $\oeRa$ and that the corresponding Bayes estimate is
minimax-optimal. Thereby, the proposed Bayesian procedure is \emph{minimax
adaptive} over the class $\cwSo$.
Although adaptation has attracted remarkable interest in the frequentist
literature, only few contributions are available in the Bayesian
literature on Gaussian sequence space models. In a direct model
\cite{BelitserGhosal2003}, \cite{SzaboVanderVaartVanZanten2013},
\cite{ArbelGayraudRousseau2013} and \cite{GaoZhou2014} derive Bayesian
methods that achieve minimax adaptation while in an indirect Gaussian
sequence space model, to the best of our knowledge, only
\cite{KnapikSzaboVanderVaartVanZanten2014} has derived an adaptive
Bayesian procedure. In this paper, we extend previous results on
adaptation obtained through sieve priors to the indirect Gaussian
sequence space model. This requires a specification of the prior on
the tuning parameter $\RvDi$ different from the one used by, e.g.,
\cite{Zhao2000} and \cite{ArbelGayraudRousseau2013}. Interestingly,
our novel prior specification on $\RvDi$ improves the general results
of \cite{ArbelGayraudRousseau2013} since it allows to obtain
adaptation without a rate loss (given by a logarithmic factor) even in
the direct model. Compared to
\cite{KnapikSzaboVanderVaartVanZanten2014} our procedure relies on a
sieve prior while they use a family of Gaussian prior for $\RvSo$
that is not degenerate in any component of $\RvSo$ and where the
hyper-parameter is represented by the smoothness of the prior
variance. Their procedure is minimax-adaptive up to a logarithmic
deterioration of the minimax rate on certain smoothness classes
for $\TrSo$ which is, instead, avoided by our
procedure.
The rest of the paper is organised as follows. The prior scheme is
specified in Section \ref{s:mo}. In Section \ref{s:tr} we derive the
lower and upper bound of the posterior concentration, the oracle
posterior concentration rate and the minimax rate. In Section
\ref{s:ad} we introduce a prior distribution $P_{\RvDi}$ for the
random dimension $\RvDi$ and we prove adaptation of the hierarchical Bayes procedure.
The proofs are given in the appendix.\\
\section{Basic model assumptions}\label{s:mo}
Let us consider a Gaussian {prior} distribution for the parameter
$\RvSo=(\RvSo_j)_{j\geq1}$, that is, $\{\RvSo_j\}_{j\geq1}$ are independent, normally distributed with prior means
$(\RvSoEr_j)_{j\geq1}$ and prior variances $(\RvSoVa_j)_{j\geq1}$.
Standard calculus shows that the posterior distribution of $\RvSo$
given $\ObSo=(\ObSo_j)_{j\geq1}$ is Gaussian, that is, given $\ObSo$,
$\{\RvSo_j\}_{j\geq1}$ are conditionally independent, normally distributed random variables with
posterior variance
$\sigma_j:=\Var(\RvSo_j\vert\ObSo)=(\Ev_j^2\ObSoNoL^{-1}+\RvSoVa^{-1}_j)^{-1}$
and posterior mean
$ \So^{\ObSo}_j:=\Ex[\RvSo_j\vert\ObSo]=\sigma_j({\RvSoVa^{-1}_j\RvSoEr_j+\Ev_j\ObSoNoL^{-1}\ObSo_j})$, for all $j\in\Nz$.
Taking this as a starting point, we construct a sequence of hierarchical Sieve prior distributions. To be more precise,
let us denote by $\dirac_x$ the Dirac measure in the point $x$. Given $\Di\in\Nz$, we
consider the independent random variables $\set{\DiRvSo_j}_{j\geq1}$
with marginal distributions
\begin{equation}\label{mo:de:DiRvSo}
\DiRvSo_j\sim
\cN(\RvSoEr_j,\RvSoVa_j),\; 1\leq j\leq \Di \mbox{ and } \DiRvSo_j\sim
\dirac_{\RvSoEr_j},\; \Di<j,
\end{equation}
resulting in the degenerate prior distribution
$P_{\DiRvSo}$. Here, we use the notation $\DiRvSo=\suite{\DiRvSo}$.
Consequently, $\{\DiRvSo_j\}_{j\geq1}$ are conditionally independent given $\ObSo$ and their { posterior}
distribution $P_{\DiRvSo_j|\ObSo}$ is Gaussian with mean $\So^{\ObSo}_j$ and variance $\sigma_j$ for
$1\leq j\leq \Di$ while being degenerate on $\RvSoEr_j$ for $j>\Di$.\\
Let $\Ind{A}$ denote the
indicator function which takes the value one if the condition $A$ holds true, and the value zero otherwise.
We consider the posterior mean $\hSo^\Di=\suite{\hSo^\Di}:=\Ex[\DiRvSo\vert \ObSo]$ given for $j\geq1$ by
$\hSo_j^{\Di}:=\So^{\ObSo}_j \Ind{\set{j\leq \Di}}+
\RvSoEr_j \Ind{\set{j> \Di}}$ as Bayes estimator of $\So$. We shall
emphasize an improper specification of the prior, that is, $\RvSoEr=\suite{\RvSoEr}\equiv 0$ and $\RvSoVa=\suite{\RvSoVa}\equiv
\infty$. Obviously, in this situation
$\So^{\ObSo}=Y/\Ev=\suite{Y_j/\Ev}$ and $\PoVa=\ObSoNoL/\Ev^2=\suite{\ObSoNoL/\Ev^2}$ are the posterior mean and
variance sequences, respectively. Consequently, under the improper prior
specification, for each $\Di\in\Nz$ the posterior mean
$\hSo^\Di=\Ex[\RvSo^\Di|\ObSo]$ of $\RvSo^\Di$ corresponds to an
orthogonal projection estimator, i.e.,
$\hSo^\Di=(Y/\Ev)^m$ with $(Y/\Ev)^m_j=Y_j/\Ev_j\Ind{\set{1\leq j\leq \Di}}$.
\\
From a Bayesian point of view the thresholding parameter $\Di$ is a hyper-parameter and hence, we may complete
the prior specification by introducing a prior distribution on it. Consider a random
thresholding parameter $\RvDi$ taking its values in $\set{1,\dotsc,\DiMa}$ for some $\DiMa\in\Nz$ with
prior distribution $P_{\RvDi}$. Both
$\DiMa$
and $P_{\RvDi}$ will be specified in Section
\ref{s:ad}.
Moreover, the distribution of the random variables $\set{\ObSo_j}_{j\geq1}$ and $\{\RvDiRvSo_j\}_{j\geq1}$ conditionally on $\RvDi$ are determined by
\begin{equation*}
\ObSo_j=\Ev_j\RvDiRvSo+\sqrt{\ObSoNoL}\ObSoNo_j\quad\text{ and }\quad
\RvDiRvSo_j=\RvSoEr_j+\sqrt{\RvSoVa}_j \RvSoNo_j \Ind{\set{1\leq j\leq \RvDi}}
\end{equation*}
where $\{\ObSoNo_j,\RvSoNo_j\}_{j\geq1}$
are iid. standard normal random variables independent of
$\RvDi$. Furthermore, the posterior mean $\hSo:=\Ex[\RvDiRvSo|\ObSo]$ satisfies
$\hSo_j=\RvSoEr_j$ for $j>\DiMa$ and $\hSo_j=\RvSoEr_j\,P(1\leq \RvDi<
j\vert\ObSo) + \So_j^{\ObSo}\,P(j\leq \RvDi\leq \DiMa\vert\ObSo)$ for
all $1\leq j\leq \DiMa$. It is important to note, that the marginal posterior
distribution $P_{\RvDiRvSo|\ObSo}$ of $\RvDiRvSo=\suite{\RvDiRvSo}$
given the observation $\ObSo$ does depend on the prior specification
and the observation only, and hence it is fully data-driven. Revisiting
the improper prior specification introduced above,
the data-driven Bayes estimator equals a shrunk orthogonal
projection estimator. More precisely, we have $\hSo_j= P(j\leq
\RvDi\leq \DiMa\vert\ObSo)\times Y_j/\Ev_j\Ind{\set{1\leq j\leq
\DiMa}}$. Interestingly, rather than using the data to select the dimension
parameter $\Di$ in the set of possible values $\{1,\dotsc,\DiMa\}$,
the Bayes estimator uses all components, up to $\DiMa$, shrunk by a
weight decreasing with the index.
\section{Optimal concentration rate}\label{s:tr}
\subsection{Consistency}
Note that conditional on $\ObSo$ the random variables $\{\DiRvSo_j-\TrSo[j]\}_{j=1}^m$ are independent and
normally distributed with conditional mean $\So^{\ObSo}_j-\TrSo[j]$ and conditional variance
$\sigma_j$.
The next assertion
presents a version of tail bounds for sums of independent squared
Gaussian random variables. It is shown in the appendix using a result
due to \cite{Birge2001} which can be shown along the lines of the proof of Lemma 1 in \cite{LaurentLM2012}.
\begin{lem}\label{tr:le:te}
Let $\{X_j\}_{j\geq1}$ be independent and normally distributed r.v.\ with mean
$\alpha_j\in\Rz$ and standard deviation $\beta_j\geq0$, $j\in\Nz$. For $m\in\Nz$ set
$S_{m}:=\sum_{j=1}^{m}X_j^2$ and consider $v_m\geq \sum_{j=1}^m\beta_j^2$, $t_m\geq \max_{1\leq
j\leq m}\beta_j^2$ and $r_m\geq\sum_{j=1}^m\alpha_j^2$. Then for all $c\geq 0$ we have
\begin{align}\label{tr:le:te:e1}
&\sup_{m\geq 1}\exp\Bigl({\frac{c(c\wedge1)(v_m+ 2r_m)}{4 t_m}}\Bigr)P\big(S_m-\Ex S_m\leq -c(v_m+2r_m)\big) \leq 1;\\ \label{tr:le:te:e2}
&\sup_{m\geq 1}\exp\Bigl({\frac{c(c\wedge1)(v_m + 2r_m)}{4t_m}}\Bigr) P\big(S_m-\Ex S_m \geq \frac{3c}{2}(v_m +
2r_m)\big)\leq1.
\end{align}
\end{lem}
\noindent A major step towards establishing a concentration rate of the
{posterior} distribution consists in finding a finite sample bound
for a fixed $\Di\in\Nz$. We express these bounds in terms of
\begin{gather*}
\gb_\Di:=\sum_{j>\Di}(\TrSo[j]-\RvSoEr_j)^2,\quad\oPoVa:=\sum_{j=1}^{\Di}\sigma_j \quad\mbox{with }\sigma_j=(\Ev_j^2\ObSoNoL^{-1}+\RvSoVa_j^{-1})^{-1};\\
\mPoVa:=\max_{1\leq j\leq \Di}\sigma_j\quad\mbox{and}\quad \gr_\Di:= \sum_{j=1}^{m}(\Ex_{\TrSo}[\So^{\ObSo}_j]-\TrSo[j])^2=\sum_{j=1}^{m}\sigma_j^2\RvSoVa^{-2}_j(\RvSoEr_j-\TrSo[j])^2.
\end{gather*}
\begin{prop}\label{tr:pr:pbm}For all $\Di\in\Nz$, for all $\ObSoNoL>0$
and for all $ 0<c< 1/5$ we
have
\begin{align}\label{tr:pr:pbm:e1}
& \Ex_{\TrSo}P_{\DiRvSo|\ObSo}(
\HnormV{\DiRvSo-\TrSo}^2> \gb_\Di+ 3\oPoVa + {3}\,\Di\,\mPoVa/2+4\gr_\Di)\leq 2\exp(-{\Di}/{36});\\\label{tr:pr:pbm:e2}
& \Ex_{\TrSo}P_{\DiRvSo|\ObSo}( \HnormV{\DiRvSo-\TrSo}^2< \gb_\Di+\oPoVa- 4\, c\, (\Di\,\mPoVa+\gr_\Di))\leq 2\exp(-c^2\Di/2).\hfill
\end{align}
\end{prop}
\noindent
The desired convergence to zero of all the aforementioned sequences
necessitates to consider an appropriate sub-family
$\{P_{\RvSo^{\eDi}}\}_{\eDi}$ in dependence of the noise level
$\ObSoNoL$, notably introducing consequently sub-sequences $(\oPoVa[\eDi])_{\eDi\geq1},(\mPoVa[\eDi])_{\eDi\geq1}$ and
$(\gr_{\eDi})_{\eDi\geq1}$.
\begin{ass}\label{tr:as:pr}
There exist constants $
0<\ObSoNoL_{\TrSy}:=\ObSoNoL_{\TrSy}{(\TrSo,\Ev,\RvSoEr,\RvSoVa)}<1 $ and $1\leq
K:=K{(\TrSo,\Ev,\RvSoEr,\RvSoVa)}<\infty$ such
that
the Sieve sub-family
$\{P_{\RvSo^{\eDi}}\}_{\eDi}$ of prior
distributions satisfies the condition
$\sup_{0<\ObSoNoL<\ObSoNoL_{\TrSy}}(\gr_{\eDi}\vee \eDi\mPoVa[\eDi])/(\gb_{\eDi}\vee\oPoVa[\eDi])\leq K$.
\end{ass}%
\noindent The following corollary can be immediately deduced from
Proposition \ref{tr:pr:pbm} and we omit its proof.%
\begin{coro}\label{tr:co:pbm} Under Assumption \ref{tr:as:pr}
for all $0<\ObSoNoL<\ObSoNoL_{\TrSy}$ and $0<c<1/(8K)$ hold
\begin{align}\label{tr:co:pbm:e1}
& \Ex_{\TrSo}P_{\DiRvSo[\eDi]|\ObSo}(
\HnormV{\DiRvSo[\eDi]-\TrSo}^2> (4+({11}/{2})K)[\gb_{\eDi}\vee\oPoVa[\eDi]])\leq 2\exp(-\frac{\eDi}{36});\\\label{tr:co:pbm:e2}
& \Ex_{\TrSo}P_{\DiRvSo[\eDi]|\ObSo}( \HnormV{\DiRvSo[\eDi]-\TrSo}^2<(1-8\,c\,K)[\gb_{\eDi}\vee\oPoVa[\eDi]])\leq 2\exp(-c^2\eDi/2).\hfill
\end{align}
\end{coro}
\noindent Note that the sequence $(
\gb_{\eDi}\vee\oPoVa[\eDi])_{\eDi\geq 1}$ generally does not converge
to zero. However, supposing that $\eDi\to\infty$ as $\ObSoNoL\to0$
then it follows from the dominated convergence
theorem that $\gb_{\eDi}=o(1)$. Hence, assuming additionally that $\oPoVa[\eDi]=o(1)$
holds true is sufficient to ensure that $( \gb_{\eDi}\vee\oPoVa[\eDi])_{\eDi\geq 1}$
converges to zero and it is indeed
a posterior concentration rate. The next assertion summarises this
result and we omit its elementary proof.
\begin{prop}[Posterior consistency]\label{tr:th:pc} Let Assumption
\ref{tr:as:pr} be satisfied. If
$\eDi\to\infty$ and $\oPoVa[\eDi]=o(1)$ as $\ObSoNoL\to0$, then
\begin{equation*}
\lim_{\ObSoNoL\to0}\Ex_{\TrSo}P_{\DiRvSo[\eDi]|\ObSo}\big( (10K)^{-1}[\gb_{\eDi}\vee\oPoVa[\eDi]]\leq\HnormV{\DiRvSo[\eDi]-\TrSo}^2\leq 10K[\gb_{\eDi}\vee\oPoVa[\eDi]]\big)=1.
\end{equation*}
\end{prop}
\noindent The last assertion shows that $(\gb_{\eDi}\vee\oPoVa[\eDi])_{\eDi\geq1}$ is up to a constant a lower and upper
bound of the concentration rate associated with the Sieve sub-family
$\{P_{\RvSo^{\eDi}}\}_{\eDi}$ of prior distributions. It is easily
shown that it also provides an upper bound of the frequentist risk of
the associated Bayes estimator.
\begin{prop}[Bayes estimator consistency]\label{tr:pr:bc} Let the assumptions of Proposition \ref{tr:th:pc} be
satisfied. Consider the Bayes estimator $\hSo^{\eDi}:=\Ex[\DiRvSo[\eDi]|\ObSo]$ then
\begin{equation*}
\Ex_{\TrSo} \HnormV{\hSo^{\eDi}-\TrSo}^2 \leq (2+K)[\gb_{\eDi}\vee\oPoVa[\eDi]]
\end{equation*}
and consequently $\Ex_{\TrSo} \HnormV{\hSo^{\eDi}-\TrSo}^2=o(1)$ as $\ObSoNoL\to0$.
\end{prop}
\noindent The previous results are obtained under Assumption \ref{tr:as:pr}. However, it may be difficult to verify whether a
given sub-family of priors $\{P_{\RvSo^{\eDi}}\}_{\eDi}$ satisfies such an assumption.
Therefore, we now introduce an assumption which states a more
precise requirement on the prior variance and that can be more easily
verified. Define for $j,m\in\Nz$
\begin{equation*}
\Evs_j:=\Ev_j^{-2},\quad \mEvs:=\max_{1\leq j\leq
\Di}\Evs_{j},\quad\oEvs:=m^{-1}\sum_{j=1}^\Di\Evs_{j}\quad\text{ and
}\quad \eRa^{\Di}:=[\gb_{\Di}\vee \ObSoNoL\,\Di\,\oEvs[\Di]].
\end{equation*}
\begin{ass}\label{tr:as:pv} Let $\DiMa:=\max\{1\leq m\leq \gauss{\ObSoNoL^{-1}}:\ObSoNoL\mEvs\leq \Evs_1\}$.
There exists a finite constant $d>0$ such that
$\RvSoVa_j\geq d[\ObSoNoL^{1/2}\Evs_j^{1/2} \vee\ObSoNoL\Evs_j]$ for all $1\leq j \leq \DiMa$ and for all $\ObSoNoL\in(0,1)$.
\end{ass}
\noindent Note that in the last Assumption the defining set of $\DiMa$ is not empty, since
$\ObSoNoL \mEvs[1]\leq\Evs_1$ for all $\ObSoNoL\leq 1$.
Moreover, under Assumption \ref{tr:as:pv}, by some elementary algebra, it is readily verified
for all $1\leq j \leq \DiMa$ that
\begin{equation*}
1 \leq \ObSoNoL\Evs_j/\PoVa[j]\leq (1+1/d) \quad\text{
and } \quad\PoVa[j]/\RvSoVa_j \leq (1\wedge d^{-1}\ObSoNoL^{1/2} \Evs_j^{1/2})
\end{equation*}
which in turn implies for all $1\leq \Di \leq \DiMa$ that
\begin{equation*}
\gr_{\Di}\leq
d^{-2}\HnormV{\RvSoEr-\TrSo}^2\ObSoNoL\,\mEvs[\Di],\; 1\leq \ObSoNoL\,
\Di\,\mEvs[\Di](\Di\mPoVa)^{-1}\;\text{and}\; 1\leq\ObSoNoL \,\Di \oEvs[\Di]\,(\oPoVa[\Di])^{-1} \leq(1+1/d).
\end{equation*}
We will use these elementary bounds in the sequel without further reference.
Returning to the Sieve sub-family $\{P_{\RvSo^{\eDi}}\}_{\eDi}$ of prior distributions, if in addition
to Assumption \ref{tr:as:pv} there exists a constant
$1\leq L:=L(\TrSo,\Ev,\RvSoEr)<\infty$ such that
\begin{equation}\label{tr:as:pv:ac}
\sup_{0<\ObSoNoL<1}
\ObSoNoL\,\eDi\,\mEvs[\eDi](\eRa^{\eDi})^{-1}\leq L
\end{equation}
and $\eRa^{\eDi}=o(1)$ as $\ObSoNoL\to0$ hold true, then the sub-family $\{P_{\RvSo^{\eDi}}\}_{\eDi}$ satisfies Assumption \ref{tr:as:pr} with $K:=((1+d^{-1})\vee d^{-2}\HnormV{\TrSo-\RvSoEr}^2)L$. Indeed, if $\eRa^{\eDi}=o(1)$ and, hence $\eRa[{\ObSoNoL}]^{\Di_{\ObSoNoL}}\leq\Evs_1/L$ for all
$\ObSoNoL\in(0,\ObSoNoL_{\TrSy})$, then $\eDi\leq\DiMa$ holds true for all
$\ObSoNoL\in(0,\ObSoNoL_{\TrSy})$ since
$\ObSoNoL\eDi\Evs_1\leq\ObSoNoL\eDi\mEvs[\eDi]\leq L
\eRa[{\ObSoNoL}]^{\Di_{\ObSoNoL}}\leq \Evs_1$ and thus $\eDi\leq
\gauss{\ObSoNoL^{-1}}$ and $\ObSoNoL \mEvs[\eDi]\leq
\Evs_1$.
In other words, for all $\ObSoNoL\in(0,\ObSoNoL_{\TrSy})$ we can
apply Assumption \ref{tr:as:pv} and the claim follows taking into account
the aforementioned elementary bounds. Note further that the constant $K$ does not depend on the prior variances
$\RvSoVa$ but only on the constant $d$ given by Assumption
\ref{tr:as:pv}. The next assertion follows immediately from Corollary
\ref{tr:co:pbm} and we omit its proof.
\begin{coro}\label{tr:co:pb}Under Assumption \ref{tr:as:pv} consider a sub-family $\{P_{\RvSo^{\eDi}}\}_{\eDi}$ such
that \eqref{tr:as:pv:ac} and $\eRa^{\eDi}=o(1)$ as $\ObSoNoL\to0$
are satisfied, then there exists $\ObSoNoL_{\TrSy}\in(0,1)$ such that for
all $0<\ObSoNoL<\ObSoNoL_{\TrSy}$ and $0<c<1/(8K)$ with $K=((1+d^{-1})\vee
d^{-2}\HnormV{\TrSo-\RvSoEr}^2)L$ hold
\begin{align}\label{tr:co:pb:e1}
& \Ex_{\TrSo}P_{\DiRvSo[\eDi]|\ObSo}\big(
\HnormV{\DiRvSo[\eDi]-\TrSo}^2> (4+({11}/{2})K)\eRa^{\eDi}\big)\leq 2\exp(-\frac{\eDi}{36});\\\label{tr:co:pb:e2}
& \Ex_{\TrSo}P_{\DiRvSo[\eDi]|\ObSo}\big( \HnormV{\DiRvSo[\eDi]-\TrSo}^2<(1-8\,c\,K)(1+d^{-1})^{-1}\eRa^{\eDi}\big)\leq 2\exp(-c^2\eDi/2).\hfill
\end{align}
\end{coro}
\noindent The result implies consistency if $\eDi\to\infty$ as $\ObSoNoL\to0$ but it does not answer the question
of an optimal rate in a satisfactory way.
\subsection{Oracle concentration rate}
Considering the Sieve family $\{P_{\RvSo^{\Di}}\}_\Di$ of prior
distributions, the sequence $(\eRa^{\eDi})_{\eDi\geq1}$ provides up to constants a lower and upper
bound for the posterior concentration rate for each sub-family
$\{P_{\RvSo^{\eDi}}\}_{\eDi}$ satisfying the conditions of Corollary \ref{tr:co:pb}. Observe that the term $\gb_{\eDi}$ and hence the rate
depends on the parameter of interest $\TrSo$. Let us minimise the rate for each
$\TrSo$ separately. For a sequence $(a_m)_{m\geq1}$ with minimal
value in $A$ we set $\argmin\nolimits_{m\in A}\set{a_m}:=\min\set{m:a_m\leq
a_{k},\forall k\in A}$ and define for all $\ObSoNoL>0$
\begin{multline}\label{tr:de:tre}
\treDi:=\treDi(\TrSo,\RvSoEr,\Ev):=\argmin_{\Di\geq 1} \set{\eRa^{\Di}}\text{ and }\\
\treRa:=\treRa(\TrSo,\RvSoEr,\Ev):=\eRa^{\treDi}=\min_{\Di\geq 1}\eRa^{\Di}\quad .
\end{multline}
We may emphasise that $\treRa=o(1)$ as $\ObSoNoL\to 0$. Indeed, for
all $\delta>0$ there exists a dimension $\Di_\delta$ and a noise level $\ObSoNoL_\delta$ such that $\treRa\leq[\gb_{\Di_\delta}\vee\ObSoNoL_{\delta}\, \Di_\delta\,
\oEvs[\Di_\delta]]\leq \delta$ for all
$0<\ObSoNoL\leq\ObSoNoL_\delta$.
Obviously, given $\TrSo\in\cSo$ the rate $\treRa$ is a lower bound for all posterior concentration
rates $\eRa^{\eDi}$ associated with a
prior sub-family $\{P_{\RvSo^{\eDi}}\}_{\eDi}$ satisfying the conditions of
Corollary \ref{tr:co:pb}.
Moreover, the next assertion establishes $\treRa$ up to constants as upper and lower bound for
the concentration rate associated with the sub-family
$\{P_{\RvSo^{\treDi}}\}_{\treDi}$. Consequently, $\treRa$ is called oracle posterior concentration rate
and $\{P_{\RvSo^{\treDi}}\}_{\treDi}$ oracle prior sub-family. The assertion follows again from
Corollary \ref{tr:co:pbm} (with $c=1/(9K)$) and we omit its proof.
\begin{theo}[Oracle posterior concentration rate]\label{tr:pr:ora}
Suppose that Assumption \ref{tr:as:pv} holds true and that there
exists a constant $1\leq L^{\TrSy}:=L^{\TrSy}(\TrSo,\Ev,\RvSoEr)<\infty$ such that
\begin{equation}\label{tr:pr:ora:ac}
\sup_{0<\ObSoNoL<1} \ObSoNoL\,\treDi\,\mEvs[\treDi](\treRa)^{-1}\leq L^{\TrSy}.
\end{equation}
If in addition $\treDi\to\infty$ as $\ObSoNoL\to0$ and $K^{\TrSy}:=10((1+d^{-1})\vee d^{-2}\HnormV{\TrSo-\RvSoEr}^2)L^{\TrSy}$, then
\[\lim_{\ObSoNoL\to0}\Ex_{\TrSo}P_{\DiRvSo[\treDi]|\ObSo}(
(K^{\TrSy})^{-1} \treRa\leq \HnormV{\DiRvSo[\treDi]-\TrSo}^2\leq K^{\TrSy} \treRa)=1.\]
\end{theo}
\noindent Note that $\treDi\to\infty$ as $\ObSoNoL\to 0$ if and only if
$\gb_{\Di}>0$ for all $m\geq1$. Roughly speaking, the last assertion
establishes $\treRa$ as oracle posterior concentration rate for all
parameter of interest $\TrSo$ with components differing from the
components of the prior mean
$\RvSoEr$ infinitely many times.
However, we do not need this additional assumption to prove the next
assertion which establishes $\treRa$ as oracle
rate for the family $\{\hSo^{\Di}\}_{\Di}$ of Bayes estimator
and that $\hSo^{\treDi}$ is an oracle Bayes estimator.
\begin{theo}[Oracle Bayes estimator]\label{tr:pr:bo} Consider the family $\{\hSo^{\Di}\}_{\Di}$ of Bayes
estimators. Under Assumption
\ref{tr:as:pv} we have \begin{inparaenum}[\color{dred}\upshape(i\upshape)]
\item\label{tr:pr:bo:i}$\Ex_{\TrSo}
\HnormV{\hSo^{\treDi}-\TrSo}^2 \leq (2+d^{-2}\HnormV{\TrSo-\RvSoEr}^2)\treRa$ and
\item\label{tr:pr:bo:ii}$\inf_{\Di\geq 1} \Ex_{\TrSo}
\HnormV{\hSo^{\Di}-\TrSo}^2\geq (1+1/d)^{-2}\treRa$
\end{inparaenum} for all $\ObSoNoL\in(0,\ObSoNoL_o)$. \end{theo}
\noindent Note that, the oracle choice $\treDi$ depends on the parameter of interest $\TrSo$ and thus the
oracle Bayes estimator $\hSo^{\treDi}$ as well as the associated oracle
sub-family $\{P_{\RvSo^{\treDi}}\}_{\treDi}$ of prior distributions
are generally not feasible.
\subsection{Minimax concentration rate}
In the spirit of a minimax theory
we are interested in the following in a uniform rate over a class
of parameters rather than optimising the rate for each $\TrSo$
separately. Given a strictly positive and non-increasing sequence
$\wSo=\suite{\wSo}$ with $\wSo_1=1$ and $\lim_{j\to\infty}\wSo_j=0$ consider for $\theta\in\Hspace$ its weighted norm
$\normV{\theta}_\wSo^2:=\sum_{j\geq1}\theta_j^2/\wSo_j$. We define $\wHspace$ as the completion of $\Hspace$ with
respect to $\norm_\wSo$.
In order to formulate the optimality of the posterior concentration rate let us
define
\begin{multline}\label{tr:de:oe}
\oeDi:=\oeDi(\wSo,\Ev):=\argmin_{\Di\geq 1} \set{\wSo_{\Di}\vee \ObSoNoL\, \Di\, \oEvs}\text{ and }\\
\oeRa:=\oeRa(\wSo,\Ev):=[\wSo_{\oeDi}\vee \ObSoNoL\, \oeDi\, \oEvs[\oeDi]]\quad \text{for all }\ObSoNoL>0.
\end{multline}
We remark that $\oeRa=o(1)$ and $\oeDi\to\infty$ as $\ObSoNoL\to 0$
since $\wSo$ is strictly positive and tends monotonically to
zero. We assume in the following that the
parameter $\TrSo$ belongs to the ellipsoid $\cwrSo:=\set{\theta\in \wHspace:
\normV{\theta-\RvSoEr}_\wSo^2\leq\rSo}$ and therefore, $\gb_\Di(\TrSo)\leq
\ga_{\Di}\rSo$. Note that $\treRa=\min_{\Di\geq1}[\gb_{\Di}\vee
\ObSoNoL\, \Di\, \oEvs]\leq (1\vee\rSo)\min_{\Di\geq1}[\wSo_{\Di}\vee \ObSoNoL\,
\Di\, \oEvs]=(1\vee\rSo)\oeRa$ and $\HnormV{\TrSo-\RvSoEr}^2\leq\rSo$, and hence from Theorem \ref{tr:pr:bo} it follows $\Ex_{\TrSo}
\HnormV{\hSo^{\treDi}-\TrSo}^2 \leq (2+\rSo/d^2)(1\vee\rSo)\oeRa$. On the
other hand side, given an estimator $\hSo$ of $\So$ let $\sup_{\So\in\cwrSo}\Ex_\So\HnormV{\hSo-\So}^2$
denote the maximal mean integrated squared error over the class
$\cwrSo$. It has been shown in \cite{JohannesSchwarz2013} that $\oeRa$
provides up to a constant a lower bound for the maximal MISE over
the class
$\cwrSo$ (assuming a prior mean $\RvSoEr=0$) if the next assumption is satisfied.
\begin{ass}\label{tr:as:mi}
Let $\wSo$ and $\Ev$ be sequences such that
\begin{equation}
0<\kappa^{\OpSy}:=\kappa^{\OpSy}(\wSo,\Ev):=\inf_{0<\ObSoNoL<\ObSoNoL_o}\set{(\oeRa)^{-1}[\wSo_{\oeDi}\wedge \ObSoNoL\,
\oeDi\, \oEvs[\oeDi]]}\leq 1.\label{tr:as:mi:e1}
\end{equation}
\end{ass}
\noindent We may emphasise that under Assumption \ref{tr:as:mi} the rate
$\oeRa=\oeRa(\wSo,\Ev)$ is optimal in a minimax sense and the Bayes
estimate $\hSo^{\treDi}$ attains the minimax rate up to a constant.
However, the dimension parameter $\treDi$ depends still on the
parameter of interest $\TrSo$. Therefore, let us consider the Bayes
estimate $\hSo^{\oeDi}$ and the sub-family
$\{P_{\RvSo^{\oeDi}}\}_{\oeDi}$ of prior distributions
which do not depend anymore on the parameter
of interest $\TrSo$ but only on the set of possible parameters $\cwrSo$
characterised by the weight sequence $\wSo$. The next assertion can be shown
along the lines of the proof of Theorem \ref{tr:pr:bo},
and, hence we omit its proof.
\begin{theo}[Minimax optimal Bayes estimator]\label{tr:th:obe} Let Assumption
\ref{tr:as:pv} be satisfied. Considering the Bayes estimator
$\hSo^{\oeDi}:=\Ex[\DiRvSo[\oeDi]|\ObSo]$ we have
\begin{equation*}
\sup_{\TrSo\in\cwrSo}\Ex_{\TrSo}
\HnormV{\hSo^{\oeDi}-\TrSo}^2 \leq (2+\rSo/d^2)(1\vee\rSo)\oeRa\quad\text{for all }\ObSoNoL\in(0,\ObSoNoL_o).
\end{equation*}
\end{theo}
\noindent The last assertion establishes the minimax optimality of the Bayes
estimate $\hSo^{\oeDi}$ over the class $\cwrSo$. Moreover,
the minimax rate $\oeRa$ provides up to a constant a lower and an
upper bound for the posterior concentration rate associated with the
prior sub-family $\{P_{\RvSo^{\oeDi}}\}_{\oeDi}$, which
is summarised in the next assertion.
\begin{theo}[Minimax optimal posterior concentration rate]\label{tr:th:opc}
Let Assumption \ref{tr:as:pv} and \ref{tr:as:mi} hold
true. If there
exists a constant $1\leq L^{\OpSy}:=L^{\OpSy}(\wSo,\Ev)<\infty$ such that
\begin{equation}\label{tr:as:mi:ac}
\sup_{0<\ObSoNoL<\ObSoNoL_o}\ObSoNoL\,\oeDi\,\mEvs[\oeDi](\oeRa)^{-1}\leq L^{\OpSy}
\end{equation}
and $K^{\OpSy}:=K^{\OpSy}(\rSo,\wSo,\Ev, d, \kappa):=10 ((1+1/d)\vee \rSo/d^{2})(1\vee\rSo)(L^{\OpSy}/\kappa^{\OpSy})$, then
\[\lim_{\ObSoNoL\to0}\inf_{\TrSo\in\cwrSo}\Ex_{\TrSo}P_{\DiRvSo[\oeDi]|\ObSo}(
(K^{\OpSy})^{-1} \oeRa\leq \HnormV{\DiRvSo[\oeDi]-\TrSo}^2\leq K^{\OpSy} \oeRa)=1.\]
\end{theo}
\noindent Comparing the last result with the result of Theorem \ref{tr:pr:ora} and keeping
in mind that $(1\vee \rSo)\oeRa\geq\treRa$, the posterior
concentration rate associated with the prior sub-family
$\{P_{\RvSo^{\oeDi}}\}_{\oeDi}$ is of order of the minimax rate $\oeRa$ uniformly for all parameter of
interest $\TrSo\in\cwrSo$. However, for certain parameter $\TrSo$ the minimax rate $\oeRa$ may be far slower than
the oracle rate $\treRa$. For example, as shown in case
{\bf\small[P-P]} in the following illustration the
minimax rate $\oeRa$ is of order $O(\ObSoNoL^{2p/(2a+2p+1)}$
) while it is not hard to see, that for all parameter $\TrSo$ with $\gb_\Di\asymp \exp (-\Di^{2p})$ the
oracle rate is of order $O(\ObSoNoL|\log\ObSoNoL|^{(2a+1)/(2p)})$ (see
case {\bf\small[E-P]}). Moreover, the optimal choice $\oeDi$ of the
dimension parameter still depends on the class $\cwrSo$, which might
be unknown in practise, therefore we will consider in the next section
a fully data-driven choice using a hierarchical specification of the prior distribution.
\begin{illu}\label{tr:il:mi} We illustrate the last assumptions and the minimax rate for typical choices of the sequences $\wSo$ and $\Ev$. For two strictly
positive sequences $(a_j)_{j\geq1}$ and $(b_j)_{j\geq1}$ we write $a_j\asymp b_j$, if $(a_j/b_j)_{j\geq 1}$ is bounded away from $0$ and infinity.
\begin{itemize}
\item[{\bf\small[P-P]}] Consider $\wSo_j\asymp j^{-2p}$ and $\Ev_j^2\asymp
j^{-2a}$ with $p> 0$ and $a>0$ then $\oeDi\asymp
\ObSoNoL^{-1/(2p+2a+1)}$ and $\oeRa\asymp\ObSoNoL^{2p/(2a+2p+1)}$.
\item[{\bf\small[E-P]}] Consider $\wSo_j\asymp\exp (-j^{2p}+1)$ and
$\Ev_j^2\asymp j^{-2a}$ with $p> 0$ and $a>0$ then $\oeDi\asymp |\log
\ObSoNoL -\frac{2a+1}{2p}(\log|\log \ObSoNoL|)|^{1/(2p)}$ and
$\oeRa\asymp \ObSoNoL |\log \ObSoNoL|^{(2a+1)/(2p)}$.
\item[{\bf\small[P-E]}] Consider $\wSo_j\asymp j^{-2p}$ and $\Ev_j^2\asymp
\exp (-j^{2a}+1)$, with $p> 0$ and $a>0$ then $\oeDi\asymp |\log
\ObSoNoL -\frac{2p+(2a-1)_{+}}{2a}(\log|\log \ObSoNoL|)|^{1/(2a)}$
and $\oeRa\asymp|\log \ObSoNoL|^{-p/a}$.
\end{itemize}
In all three cases Assumption \ref{tr:as:mi} and \eqref{tr:as:mi:ac} hold true.\qed
\end{illu}
\section{Data-driven Bayesian estimation}\label{s:ad}
We will derive in this section a concentration rate given the
aforementioned hierarchical prior distribution. For this purpose
we impose additional conditions on the behaviour of the sequence
$\Ev=\suite{\Ev}$.
\begin{ass}\label{ad:as:ev}
There exist finite constants $C_\Ev\geq 1$ and $L_\Ev\geq1$ such that
for all $k,l\in\Nz$ hold
\begin{inparaenum}[\color{dred}\upshape(i\upshape)]
\item\label{ad:as:ev:a} $\max_{j>k}\Ev_j^2\leq C_\Ev \min_{1\leq j\leq k}\Ev_j^2=C_\Ev
\mEvs[k]^{-1}$;
\item\label{ad:as:ev:b}$\mEvs[kl]\leq\mEvs[k]\mEvs[l]$;
\item\label{ad:as:ev:c}$1\leq \mEvs[k]/\oEvs[k]\leq L_\Ev$.
\end{inparaenum}
\end{ass}%
\noindent We may emphasise that Assumption \ref{ad:as:ev} \eqref{ad:as:ev:a} holds
trivially with $C_\Ev=1$ if the sequence $\Ev$ is monotonically
decreasing. Moreover, considering the typical choices of the sequence
$\Ev$ presented in Illustration \ref{tr:il:mi}, Assumption
\ref{ad:as:ev} \eqref{ad:as:ev:b} and \eqref{ad:as:ev:c} hold only true
in case of a polynomial decay, i.e., {\bf\small[P-P]} and
{\bf\small[E-P]}. In other words, Assumption
\ref{ad:as:ev} excludes an exponential decay of $\Ev$, i.e., {\bf\small[P-E]}.
\begin{ass}\label{ad:as:tr}
Let $\RvSoEr$, $\TrSo$ and $\Ev$ be sequences such that
\begin{equation}
0<\kappa^{\TrSy}:=\kappa^{\TrSy}(\RvSoEr,\TrSo,\Ev):=\inf_{0<\ObSoNoL<\ObSoNoL_o}\set{(\treRa)^{-1}[\gb_{\treDi}\wedge \ObSoNoL\,
\treDi\, \oEvs[\treDi]]}\leq 1.\label{ad:as:tr:e1}
\end{equation}
\end{ass}%
\noindent Observe that $\gb_{\treDi}\geq
\kappa^{\TrSy}\treRa>0$ due to Assumption \ref{ad:as:tr} which in turn implies $\gb_k>0$ for all
$k\in\Nz$ and, hence $\treDi\to\infty$ as
$\ObSoNoL\to0$. Indeed, if there exists $K\in\Nz$ such
that $\gb_K=0$ and $\gb_{K-1}>0$ then there
exists $\ObSoNoL_{\TrSy}\in(0,1)$ with
$\ObSoNoL_{\TrSy}K\oEvs[K]<\gb_{K-1}$ and for all
$\ObSoNoL\in(0,\ObSoNoL_{\TrSy})$ it is easily seen that
$\treDi=K$ and hence $\gb_{\treDi}=0$.
Moreover, due to Assumption
\ref{ad:as:ev} \eqref{ad:as:ev:c} there exists a constant $L_\Ev$
depending only on $\Ev$ such that $\ObSoNoL\,\treDi\mEvs[\treDi] (\treRa)^{-1}\leq
\mEvs[\treDi](\oEvs[\treDi])^{-1}\leq L_{\Ev}$, i.e., condition
\eqref{tr:as:pv:ac} holds true uniformly for all parameters
$\So\in\Hspace$. If we suppose in addition to Assumption \ref{ad:as:ev}
and \ref{ad:as:tr} that the sequence
of prior variances meets Assumption \ref{tr:as:pv} and that
$\treDi\to\infty$ as $\ObSoNoL\to0$, then the
assumptions of Theorem \ref{tr:pr:ora} are satisfied and $\treRa$
provides up to a constant an upper and lower bound of the posterior
concentration rate associated with the oracle prior sub-family
$\{P_{\treDi}\}_{\treDi}$.
\noindent Let us specify the prior distribution $P_{\RvDi}$
of the thresholding parameter $\RvDi$ taking its values in
$\{1,\dotsc,\DiMa\}$ with $\DiMa$ as in Assumption \ref{tr:as:pv},
and for $1\leq
m\leq \DiMa$
\begin{equation} \label{ad:de:pr:di}
p_{\RvDi}(\Di):=P_{\RvDi}(\RvDi=\Di)=\frac{\exp(-3 C_\Ev
\Di/2)\prod_{j=1}^\Di(\RvSoVa_j/\sigma_j)^{1/2}}{\sum_{k=1}^{\DiMa}\exp(-3
C_\Ev k/2)\prod_{j=1}^{k}(\RvSoVa_j/\sigma_j)^{1/2}}.
\end{equation}
\noindent Keeping in mind the sequences $\So^{\ObSo}=\suite{\So^{\ObSo}}$ and $\sigma=\suite{\sigma}$ of conditional means and
variances, respectively, given by $\So^{\ObSo}_j=\sigma_j(\Ev_j\ObSoNoL^{-1}Y_j+
\RvSoVa^{-1}_j\RvSoEr_j)$ and
$\sigma_j=(\RvSoVa_j^{-1}+\Ev_j^2\ObSoNoL^{-1})^{-1}$, for each
$\Di\in\Nz$ the sequence $\hSo^\Di=\suite{\hSo^\Di}=\Ex[\RvSo^\Di|\ObSo]$ of posterior
means of $\RvSo^\Di$ satisfies $\hSo^\Di_j=\So^{\ObSo}_j \Ind{\set{1\leq j\leq \Di}}+
\RvSoEr_j \Ind{\set{j> \Di}}$. Introducing further the weighted norm
$\normV{\theta}_{\sigma}^2:=\sum_{j\geq1}\theta_j^2/\sigma_j$ for $\theta\in\Hspace$ the
posterior distribution $P_{\RvDi|\ObSo}$ of the thresholding parameter
$\RvDi$ is given by
\begin{equation} \label{ad:po:di}
p_{\RvDi|\ObSo}(m)=P_{\RvDi|\ObSo}(\RvDi=m)
=\frac{\exp(-\frac{1}{2}\{-\normV{\hSo^\Di-\RvSoEr}_{\sigma}^2+3
C_\Ev m\})}{\sum_{k=1}^{\DiMa}\exp(-\frac{1}{2}\{-\normV{\hSo^k-\RvSoEr}_{\sigma}^2+3
C_\Ev k\})}
\end{equation}
\noindent Interestingly, the posterior distribution $P_{\RvDi|\ObSo}$ of the thresholding
parameter $\RvDi$ is concentrating around the oracle dimension
parameter $\treDi$ as $\ObSoNoL$ tends to zero. To be more precise,
there exists $\ObSoNoL_{\TrSy}\in (0,1)$ such that $\treDi\leq \DiMa$
for all $\ObSoNoL\in (0,\ObSoNoL_{\TrSy})$ since $\treRa=o(1)$ for
$\ObSoNoL\to0$. Let us further define for all $\ObSoNoL\in (0,\ObSoNoL_{\TrSy})$
\begin{multline}\label{ad:de:mp}
\meDi:=\min\set{m\in\set{1,\dotsc,\treDi}: \gb_\Di\leq 8 L_{\Ev}C_{\Ev}(1+1/d)\treRa}\quad\text{and}\\\peDi:=\max\set{m\in\set{\treDi,\dotsc,\DiMa}: \Di\leq 5L_{\Ev}(\ObSoNoL\mEvs[\treDi])^{-1} \treRa }
\end{multline}
where the defining sets are not empty under Assumption
\ref{ad:as:ev} since $8 L_{\Ev}C_{\Ev}(1+1/d)\treRa\geq 8
L_{\Ev}C_{\Ev}(1+1/d)\gb_{\treDi}\geq \gb_{\treDi}$ and $5L_{\Ev}(\ObSoNoL\mEvs[\treDi])^{-1} \treRa \geq
5\treDi\geq \treDi$. Moreover, under Assumption \ref{ad:as:tr} it is
easily verified that $\meDi\to \infty$ as
$\ObSoNoL\to 0$.
\begin{lem}\label{ad:le:pm}
If Assumptions \ref{tr:as:pv} and \ref{ad:as:ev} hold true then for all $\ObSoNoL\in(0,\ObSoNoL_{\TrSy})$
\begin{enumerate}[label=\emph{\textbf{(\roman*)}},ref=\emph{\textbf{(\roman*)}}]
\item\label{ad:le:pm:i} $\Ex_{\TrSo}P_{\RvDi|\ObSo}(1\leq
\RvDi<\meDi)\leq
2\exp\big(-\frac{7C_{\Ev}}{32}\treDi+\log\DiMa\big)\leq
2\exp\big(-\frac{C_{\Ev}}{5}\treDi+\log\DiMa\big)$;
\item\label{ad:le:pm:ii}
$\Ex_{\TrSo}P_{\RvDi|\ObSo}(\peDi<
\RvDi\leq\DiMa)\leq 2\exp\big(-\frac{4C_\Ev}{9}\treDi+\log\DiMa\big)\leq
2\exp\big(-\frac{C_{\Ev}}{5}\treDi+\log\DiMa\big)$.
\end{enumerate}
\end{lem}
\noindent Recall that $\treDi\to\infty$ as $\ObSoNoL\to0$ under Assumption
\ref{ad:as:tr}. If in addition $\treDi/(\log
\DiMa)\to\infty$ as $\ObSoNoL\to0$ then Lemma \ref{ad:le:pm} states that the posterior distribution of
the thresholding parameter $\RvDi$ is vanishing outside the set
$\{\meDi,\dotsc,\peDi\}$ as $\ObSoNoL\to0$. On the other hand side, the posterior distribution $P_{\RvDiRvSo|\ObSo}$
of $\RvDiRvSo=\suite{\RvDiRvSo}$ associated with the hierarchical
prior is a weighted mixture of the posterior distributions
$\{P_{\DiRvSo|\ObSo}\}_{\Di=1}^{\DiMa}$ studied in section \ref{s:tr}, that is,
$P_{\RvDiRvSo|\ObSo}=\sum_{\Di=1}^{\DiMa}p_{\RvDi|\ObSo}(\Di)P_{\DiRvSo|\ObSo}$.
The next assertion shows that considering posterior distributions
$\{P_{\DiRvSo|\ObSo}\}_{\Di=\meDi}^{\peDi}$ associated with
thresholding parameters belonging to $\{\meDi,\dotsc,\peDi\}$ only, then their
concentration rate equals $\treRa$ up to a constant.
\begin{lem}\label{ad:le:pv}
If Assumptions \ref{tr:as:pv}, \ref{ad:as:ev} and \ref{ad:as:tr} hold
true then for all $\ObSoNoL\in(0,\ObSoNoL_{\TrSy})$
\begin{enumerate}[label=\emph{\textbf{(\roman*)}},ref=\emph{\textbf{(\roman*)}}]
\item\label{ad:le:pv:i}$\sum_{\meDi\leq\Di\leq\peDi}\Ex_{\TrSo}
P_{\DiRvSo[\Di]|\ObSo}\big(\HnormV{\DiRvSo[\Di]-\TrSo}^2>K^o\treRa\big)\leq 74 \exp(-\meDi/36)$;
\item\label{ad:le:pv:ii}$\sum_{\meDi\leq\Di\leq\peDi}\Ex_{\TrSo}
P_{\DiRvSo[\Di]|\ObSo}\big(\HnormV{\DiRvSo[\Di]-\TrSo}^2<(K^o)^{-1}\treRa
\big)\leq 4 (K^{\TrSy})^2 \exp(-\meDi/(K^{\TrSy})^2)$,
\end{enumerate}
where $K^{\TrSy}:=10 ((1+1/d)\vee \HnormV{\TrSo-\RvSoEr}^2/d^{2})L^2_\Ev(8C_\Ev(1+1/d)\vee D^{\TrSy}\mEvs[D^{\TrSy}])$ with $D^{\TrSy}:=D^{\TrSy}(\RvSoEr,\TrSo,\Ev):=\ceil{5 L_\Ev/\kappa^{\TrSy}}$.
\end{lem}
\noindent From Lemma \ref{ad:le:pm} and \ref{ad:le:pv} we derive next
upper and lower bounds for the concentration rate of the posterior distribution $P_{\RvDiRvSo|\ObSo}$
by decomposing the weighted mixture into three parts with respect to
$\meDi$ and $\peDi$ which we bound separately.
\begin{theo}[Oracle posterior concentration rate]\label{ad:th:ora}
Let Assumptions \ref{tr:as:pv}, \ref{ad:as:ev} and \ref{ad:as:tr} hold
true. If in addition $(\log \DiMa)/\treDi\to0$ as
$\ObSoNoL\to0$, then
\[\lim_{\ObSoNoL\to0}\Ex_{\TrSo}P_{\RvDiRvSo|\ObSo}(
(K^{\TrSy})^{-1} \treRa\leq \HnormV{\RvDiRvSo-\TrSo}^2\leq K^{\TrSy} \treRa)=1\]
where $K^{\TrSy}$ is given in Lemma \ref{ad:le:pv}.
\end{theo}
\noindent We shall emphasise that the Bayes
estimator $\hSo:=\suite{\hSo}:=\Ex[\RvDiRvSo|\ObSo]$ associated with the
hierarchical prior and given by
$\hSo_j=\RvSoEr_j$ for $j>\DiMa$ and $\hSo_j=\RvSoEr_j\,P(1\leq \RvDi<
j\vert\ObSo) + \So_j^{\ObSo}\,P(j\leq \RvDi\leq \DiMa\vert\ObSo)$ for
all $1\leq j\leq \DiMa$, does not take into account any
prior information related to the parameter of interest, and hence it
is fully data-driven. The next assertion provides an upper bound of
its MISE.
\begin{theo}[Oracle optimal Bayes estimator]\label{ad:th:bo}
Under Assumptions \ref{tr:as:pv}, \ref{ad:as:ev} and \ref{ad:as:tr} consider the Bayes estimator
$\hSo:=\Ex[\RvDiRvSo|\ObSo]$. If in addition $\log
(\DiMa/\treRa)/\treDi\to0$ as $\ObSoNoL\to0$, then there exists a
constant $K^{\TrSy}:=K^{\TrSy}(\TrSo,\RvSoEr,\Ev,
d, L)<\infty$ such that $\Ex_{\TrSo} \HnormV{\hSo-\TrSo}^2\leq K^{\TrSy}
\treRa$ for all $\ObSoNoL\in(0,\ObSoNoL_{\TrSy})$.
\end{theo}
Both Theorems, \ref{ad:th:ora} and \ref{ad:th:bo} hold true only under
Assumption \ref{ad:as:tr}, which we have seen before imposes an
additional restriction on the parameter of
interest $\TrSo$, i.e., its components differ from the components of the prior mean $\RvSoEr$
infinitely many times. However, for all parameters of interest
satisfying Assumption \ref{ad:as:tr}, the hierarchical prior sequence
allows to recover the oracle posterior concentration rate and the
fully data driven Bayes estimator attains the oracle rate. {In the last part
of this section we show that for all $\TrSo\in\cwrSo$ the
posterior concentration rate and the MISE of the Bayes estimator
associated with the hierarchical prior are bounded from above by the
minimax rate $\oeRa$ up to a constant. In other words, the fully data-driven
hierarchical prior and the associated Bayes estimator are
minimax-rate optimal.}
Recall the definition \eqref{tr:de:oe} of $\oeDi$ and
$\oeRa$. Consider the prior distribution
$P_{\RvDi}$ of the thresholding parameter $\RvDi$, and observe that
there exists $\ObSoNoL_\star$ such that $\oeDi\leq \DiMa$ for all
$\ObSoNoL\in(0,\ObSoNoL_\star)$ since $\oeRa=o(1)$ as $\ObSoNoL\to0$.
Remark that $\ObSoNoL\,\oeDi\mEvs[\oeDi] (\oeRa)^{-1}\leq
\mEvs[\oeDi](\oEvs[\oeDi])^{-1}\leq L_{\Ev}$ with $L_\Ev$
depending only on $\Ev$ due to Assumption
\ref{ad:as:ev} \eqref{ad:as:ev:c}, i.e., condition \eqref{tr:as:mi:ac} holds true uniformly for all parameters
$\So\in\Hspace$. If we assume in addition that the sequence
of prior variances satisfies Assumption \ref{tr:as:pv} and that Assumption
\ref{tr:as:mi} holds true, then the
conditions of Theorem \ref{tr:th:opc} are satisfied and $\oeRa$
provides up to a constant an upper and lower bound of the posterior
concentration rate associated with the minimax prior sub-family $\{P_{\oeDi}\}_{\oeDi}$. On the other hand side, the
posterior distribution $P_{\RvDi|\ObSo}$ of the thresholding
parameter $\RvDi$ is concentrating around the minimax-optimal dimension
parameter $\oeDi$ as $\ObSoNoL$ tends to zero. To be more precise, for
$\ObSoNoL\in(0,\ObSoNoL_\star)$ let us define
\begin{multline}\label{ad:de:omp}
\moeDi:=\min\set{m\in\set{1,\dotsc,\oeDi}: \gb_\Di\leq 8 L_{\Ev}C_{\Ev}(1+1/d)(1\vee\rSo)\oeRa}\quad\text{and}\\\poeDi:=\max\set{m\in\set{\oeDi,\dotsc,\DiMa}: \Di\leq 5L_{\Ev}(\ObSoNoL\mEvs[\oeDi])^{-1}(1\vee\rSo) \oeRa }
\end{multline}
where the defining sets are not empty under Assumption
\ref{ad:as:ev} since $8 L_{\Ev}C_{\Ev}(1+1/d)(1\vee\rSo)\oeRa\geq 8
L_{\Ev}C_{\Ev}(1+1/d)\rSo\wSo_{\oeDi}\geq 8
L_{\Ev}C_{\Ev}(1+1/d)\gb_{\oeDi}\geq \gb_{\oeDi}$ and $5L_{\Ev}(\ObSoNoL\mEvs[\oeDi])^{-1} (1\vee\rSo)\oeRa \geq
5\oeDi\geq \oeDi$. Moreover, it is again straightforward to see that $\moeDi\to \infty$ as
$\ObSoNoL\to 0$.
\begin{lem}\label{ad:le:opm}
If Assumption \ref{tr:as:pv} and \ref{ad:as:ev} hold true then for all
$\TrSo\in\cwrSo$ and $\ObSoNoL\in(0,\ObSoNoL_\star)$
\begin{enumerate}[label=\emph{\textbf{(\roman*)}},ref=\emph{\textbf{(\roman*)}}]
\item\label{ad:le:opm:i} $\Ex_{\TrSo}P_{\RvDi|\ObSo}(1\leq \RvDi<\moeDi
\leq
2\exp\big(-\frac{C_{\Ev}(1\vee\rSo)}{5}\oeDi+\log\DiMa\big)$;
\item\label{ad:le:opm:ii}
$\Ex_{\TrSo}P_{\RvDi|\ObSo}(\poeDi<
\RvDi\leq\DiMa
\leq
2\exp\big(-\frac{C_{\Ev}(1\vee\rSo)}{5}\oeDi+\log\DiMa\big)$.
\end{enumerate}
\end{lem}
\noindent By employing Lemma \ref{ad:le:opm} we show next for
each $\TrSo\in\cwrSo$ that
the minimax rate $\oeRa$ provides up to a constant an upper bound
for the posterior concentration rate associated with the fully
data-driven hierarchical prior distribution $P_{\RvDiRvSo}$.
\begin{theo}[Minimax optimal posterior concentration rate]\label{ad:th:ra:mm}
Let Assumption \ref{tr:as:pv}, \ref{tr:as:mi} and \ref{ad:as:ev} hold
true. If in addition $(\log \DiMa)/\oeDi\to0$ as
$\ObSoNoL\to0$, then
\begin{enumerate}[label=\emph{\textbf{(\roman*)}},ref=\emph{\textbf{(\roman*)}}]
\item\label{ad:th:ra:mm:i} for all $\TrSo\in\cwrSo$ we have
\[\lim_{\ObSoNoL\to0}\Ex_{\TrSo}P_{\RvDiRvSo|\ObSo}(\HnormV{\RvDiRvSo-\TrSo}^2\leq K^{\OpSy} \oeRa)=1\]
where $K^{\OpSy}:={16} ((1+1/d)\vee \rSo/d^{2})L^2_\Ev(8C_\Ev(1+1/d)\vee
D^{\OpSy}\mEvs[D^{\OpSy}]){(1\vee\rSo)}$ with
$D^{\OpSy}:=D^{\OpSy}(\wSo,\Ev):=\ceil{5 L_\Ev/\kappa^{\OpSy}}$;
\item\label{ad:th:ra:mm:ii} for any monotonically increasing and
unbounded sequence $(K_\ObSoNoL)_{\ObSoNoL}$ holds
\[\lim_{\ObSoNoL\to0}\inf_{\TrSo\in\cwrSo}\Ex_{\TrSo}P_{\RvDiRvSo|\ObSo}(\HnormV{\RvDiRvSo-\TrSo}^2\leq
K_\ObSoNoL \oeRa)=1.\]
\end{enumerate}
\end{theo}
\noindent We shall emphasise that due to Theorem \ref{ad:th:ora} for all $\TrSo\in\cwrSo$ satisfying
Assumption \ref{ad:as:tr} the posterior concentration rate associated
with the hierarchical prior attains the oracle rate $\treRa$ which
might be far smaller than the minimax-rate $\oeRa$. Consequently, the
minimax rate cannot provide an uniform lower bound over $\cwrSo$ for the posterior concentration rate associated with the
hierarchical prior. However, due to Theorem \ref{ad:th:ra:mm} the
posterior concentration rate is for all $\TrSo\in\cwrSo$, independently
that Assumption \ref{ad:as:tr} holds, at least of
the order of the minimax rate $\oeRa$. The next assertion establishes the minimax-rate optimality of the
fully data-driven Bayes estimator.
\begin{theo}[Minimax optimal Bayes estimate]\label{ad:th:be:mm}
Under Assumption \ref{tr:as:pv}, \ref{tr:as:mi} and \ref{ad:as:ev} consider the Bayes estimator
$\hSo:=\Ex[\RvDiRvSo|\ObSo]$. If in addition $\log(
\DiMa/\oeRa)/\oeDi\to0$ as $\ObSoNoL\to0$, then there exists $K^\star:=K^\star(\cwrSo,\Ev,d)<\infty$ such that $\sup_{\TrSo\in\cwrSo}\Ex_{\TrSo} \HnormV{\hSo-\TrSo}^2\leq K^\star
\oeRa$ for all $\ObSoNoL\in(0,\ObSoNoL_\star)$.
\end{theo}
\noindent Let us briefly comment on the last assertion by considering
again the improper
specification of the prior family
$\{P_{\RvSo^{\Di}}\}_{\Di}$ introduced in Section \ref{s:mo}. Recall
that in this situation for each
$\Di\in\Nz$ the Bayes estimator $\hSo^\Di=\Ex[\RvSo^\Di|\ObSo]$ of
$\RvSo^\Di$ equals an orthogonal projection estimator, i.e.,
$\hSo^\Di=(Y/\Ev)^m$. Moreover, the
posterior probability of the thresholding parameter
$\RvDi$ taking a value $m\in\{1,\dotsc,\DiMa\}$ is proportional to $\exp(-\frac{1}{2}\{-\normV{(Y/\Ev)^m}_{\ObSoNoL\Evs}^2+3
C_\Ev m\})$, and hence the data-driven Bayes estimator
$\hSo=\suite{\hSo}=\Ex[\RvDiRvSo|\ObSo]$ equals the shrinked
orthogonal projection estimator given by
\begin{equation*}
\hSo_j
=\frac{\sum_{\Di=j}^{\DiMa}\exp(-\frac{1}{2}\{-\normV{(Y/\Ev)^\Di}_{\ObSoNoL\Evs}^2+3
C_\Ev m\})}{\sum_{\Di=1}^{\DiMa}\exp(-\frac{1}{2}\{-\normV{(Y/\Ev)^\Di}_{\ObSoNoL\Evs}^2+3
C_\Ev \Di\})}\times \frac{Y_j}{\Ev_j} \Ind{\set{1\leq j\leq
\DiMa}}.
\end{equation*}
From Theorem \ref{ad:th:be:mm} it follows now, that the fully data-driven shrinkage estimator $\hSo$ is
minimax-optimal up to a constant for a wide variety of parameter
spaces $\cwrSo$ provided Assumptions \ref{tr:as:mi} and \ref{ad:as:ev}
hold true. Interestingly, identifying
$\Upsilon(\hSo^\Di):=-(1/2)\normV{(Y/\Ev)^\Di}_{\ObSoNoL\Evs}^2$ as a
contrast and
$\pen_\Di:=3/2 C_\Ev \Di$ as a penalty term the $j$-th shrinkage
weight is proportional to $\sum_{\Di=j}^{\DiMa}\exp(-\{\Upsilon(\hSo^\Di)+\pen_m\})$. Roughly
speaking, in comparison to a classical model selection approach where
a data-driven estimator $\hSo^{\whm}=(Y/\Ev)^{\whm}$ is obtained by
selecting the dimension parameter $\whm$ as minimum of a
penalised contrast criterion over a
class of admissible models $\{1,\dotsc,\DiMa\}$, i.e.,
$\whm=\argmin\nolimits_{1\leq m\leq \DiMa}\{\Upsilon(\hSo^\Di)+\pen_m\}$, following the
Bayesian approach each of the
$\DiMa$ components of the data-driven Bayes estimator is shrunk
proportional to the associated values of the penalised contrast criterion.
\subsection{Proof of Theorem \ref{ad:th:ora}}
\begin{proof}[\noindent{\color{darkred}\sc Proof of Lemma
\ref{ad:le:pm}.}]\label{ad:le:pm:pr}
Consider \ref{ad:le:pm:i}. The claim holds trivially true
in case $\meDi=1$, thus suppose $\meDi>1$ and let
$1\leq \Di<\meDi\leq \treDi$. Define
$S_{\Di}:=\normV{\hSo^{\treDi}-\RvSoEr}_{\sigma}^2-\normV{\hSo^\Di-\RvSoEr}_{\sigma}^2$. Given
an event $\cA_{\Di}$ and its complement
$\cA_{\Di}^c$ (to be specified below) it follows
\begin{multline}\label{ad:le:pm:pr:e1}
p_{\RvDi|\ObSo}(\Di)=\frac{\exp(\frac{1}{2}\{\normV{\hSo^\Di-\RvSoEr}_{\sigma}^2-3
C_\Ev m\})}{\sum_{k=1}^{\DiMa}\exp(\frac{1}{2}\{\normV{\hSo^k-\RvSoEr}_{\sigma}^2-3
C_\Ev k\})}\\
=
\exp\bigg(\frac{1}{2}\big\{-S_{\Di}+3 C_\Ev [\treDi-\Di]\big\}\bigg)\Ind{\cA_{\Di}}
+ \Ind{\cA_{\Di}^c}
\end{multline}
Moreover, elementary algebra shows
\begin{equation*}
S_{\Di}
=\sum_{j=\Di+1}^{\treDi}\frac{\Ev_j^2\sigma_j}{\ObSoNoL^2}(Y_j-\Ev_j\RvSoEr_j)^2
\end{equation*}
where the random variables $\{\Ev_j\sigma_j^{1/2}\ObSoNoL^{-1}(Y_j-\Ev_j\RvSoEr_j)\}_{j\geq1}$ are independent
and normally distributed with standard deviation $\beta_j=
\Ev_j\sigma_j^{1/2}\ObSoNoL^{-1/2}$ and mean $\alpha_j=
\beta_j\ObSoNoL^{-1/2}\Ev_j(\TrSo[j]-\RvSoEr_j)$. Keeping in mind the
notations used in Lemma \ref{tr:le:te} define
$v_{\Di}:=\sum_{j=\Di+1}^{\treDi}\beta_j^2$ and $r_{\Di}:=\sum_{j=\Di+1}^{\treDi}\alpha_j^2$. We observe that Assumption
\ref{tr:as:pv} implies that $1\geq \beta_j^2\geq (1+1/d)^{-1}$ and
hence it follows by
employing $\min_{\Di<j\leq\treDi}\Ev_j^2\geq \min_{1\leq
j\leq\treDi}\Ev_j^2=\mEvs[\treDi]^{-1}$ and Assumption
\ref{ad:as:ev} \eqref{ad:as:ev:c} that
\begin{multline}\label{ad:le:pm:pr:e2}
L_\Ev (\ObSoNoL\mEvs[\treDi])^{-1}\treRa\geq L_\Ev (\ObSoNoL\mEvs[\treDi])^{-1} \ObSoNoL \treDi \oEvs[\treDi] \geq \treDi
\quad\text{ and }\\
(1+1/d)^{-1}(\ObSoNoL\mEvs[\treDi])^{-1} [\gb_\Di-\treRa] \leq (1+1/d)^{-1}(\ObSoNoL\mEvs[\treDi])^{-1} [\gb_\Di-\gb_{\treDi}] \leq
r_{\Di}.
\end{multline}
Moreover, we set $t_{\Di}:=1\geq
\max_{\Di<j\leq\treDi}\beta_j^2$ and $\mu_{\Di} := \Ex
S_{\Di}=v_{\Di}+r_{\Di}$. Introduce the event $\cA_{\Di}:=\{S_{\Di}-\mu_{\Di}\geq - (1/4)(v_{\Di}+2r_{\Di})\}$
and its complement $\cA_{\Di}^c:=\{S_{\Di}-\mu_{\Di}<
-(1/4)(v_{\Di}+2r_{\Di})\}$. By employing successively Lemma
\ref{tr:le:te}, \eqref{ad:le:pm:pr:e2} and $\gb_{\treDi}\leq \treRa$ it follows now from \eqref{ad:le:pm:pr:e1} that
\begin{multline*}
\Ex_{\TrSo} p_{\RvDi|\ObSo}(\Di) \leq
\Ex_{\TrSo}\exp\big(\{-(S_{\Di}-\mu_{\Di})-\mu_{\Di}+3C_\Ev[\treDi-\Di]\}/2\big)\Ind{\cA_{\Di}}
+ \Ex_{\TrSo}\Ind{\cA_{\Di}^c}\\
\leq\exp
\big(\{-3v_{\Di}/4-r_{\Di}/2+3C_\Ev[\treDi-\Di]\}/2\big)
+\exp
\big(-(1/64)(v_{\Di} + 2r_{\Di})\big)\\
\leq \exp
\big(-r_{\Di}/4+3C_\Ev\treDi/2\big)
+\exp
\big(-r_{\Di}/32\big)\\
\leq \exp
\big( -\frac{[\gb_\Di-\treRa]}{4(1+1/d)\ObSoNoL\mEvs[\treDi]}+\frac{3C_\Ev L_\Ev\treRa}{2\ObSoNoL\mEvs[\treDi]}\}\big)+\exp
\big(-\frac{[\gb_\Di-\treRa]}{32(1+1/d)\ObSoNoL\mEvs[\treDi]})\big)\\
\leq \exp
\big(-\frac{\gb_\Di}{4(1+1/d)\ObSoNoL\mEvs[\treDi]} + \frac{2C_\Ev
L_\Ev\treRa}{\ObSoNoL\mEvs[\treDi]}\big)
\times\exp\big(-\frac{L_\Ev C_{\Ev}\treRa}{4\ObSoNoL\mEvs[\treDi]}\big)
\\
+\exp
\big(-\frac{[\gb_\Di-\treRa]}{32(1+1/d)\ObSoNoL\mEvs[\treDi]}\big)
\end{multline*}
Taking into account the definition \eqref{ad:de:mp} of $\meDi$, i.e.,
$\gb_\Di> 8 L_{\Ev}C_{\Ev}(1+1/d)\treRa$ for all $1\leq\Di<\meDi$, and
$L_\Ev\treRa(\ObSoNoL\mEvs[\treDi])^{-1}\geq \treDi$ due to Assumption \ref{ad:as:ev} \eqref{ad:as:ev:c}, we obtain
\begin{equation*}
\Ex_{\TrSo} p_{\RvDi|\ObSo}(\Di) \leq \exp\big(-\frac{L_\Ev C_{\Ev}\treRa}{4\ObSoNoL\mEvs[\treDi]}\big)
+\exp
\big(-\frac{7
L_{\Ev}C_{\Ev}\treRa}{32\ObSoNoL\mEvs[\treDi]}\big)\leq 2 \exp
\big(-\frac{7C_{\Ev}}{32}\treDi\big).
\end{equation*}
Thereby, $\Ex_{\TrSo}P_{\RvDi|\ObSo}(1\leq
\RvDi<\meDi)=\sum_{\Di=1}^{\meDi-1}\Ex_{\TrSo} p_{\RvDi|\ObSo}(\Di)\leq 2 \exp
\big(-\frac{7C_{\Ev}}{32}\treDi + \log \DiMa \big)$ using that
$\DiMa\geq \meDi$ which proves the assertion \ref{ad:le:pm:i}.
Consider now \ref{ad:le:pm:ii}. The claim holds trivially true
in case $\peDi=\DiMa$, thus suppose $\peDi<\DiMa$ and let
$\DiMa\geq\Di>\peDi\geq \treDi$. Consider again the upper bound given
in \eqref{ad:le:pm:pr:e1} where now
\begin{equation*}
-S_{\Di}
=\sum_{j=\treDi+1}^{\Di}\frac{\Ev_j^2\sigma_j}{\ObSoNoL^2}(Y_j-\Ev_j\RvSoEr_j)^2.
\end{equation*}
Employing the notations $\alpha_j$ and $\beta_j$ introduced in the
proof of \ref{ad:le:pm:i} and keeping in mind Lemma \ref{tr:le:te} we
define $v_{\Di}:=\sum_{j=\treDi+1}^{\Di}\beta_j^2$ and
$r_{\Di}:=\sum_{j=\treDi+1}^{\Di}\alpha_j^2$ where $1\geq
\beta_j^2\geq (1+1/d)^{-1}$ due to Assumption
\ref{tr:as:pv}. Moreover, from Assumption
\ref{ad:as:ev} \eqref{ad:as:ev:a} follows
$\max_{\treDi<j\leq\Di}\Ev_j^2\geq \max_{\treDi<j}\Ev_j^2\leq C_\Ev
\min_{1\leq j\leq \treDi}\Ev_j^2 =C_\Ev\mEvs[\treDi]^{-1}$ and
taking into account in addition Assumption
\ref{ad:as:ev} \eqref{ad:as:ev:c} that
\begin{multline}\label{ad:le:pm:pr:e3}
L_\Ev (\ObSoNoL\mEvs[\treDi])^{-1}\treRa\geq \treDi,\quad v_\Di\leq m-\treDi
\quad\text{ and }\\
C_\Ev(\ObSoNoL\mEvs[\treDi])^{-1}\treRa \geq C_\Ev(\ObSoNoL\mEvs[\treDi])^{-1} [\gb_{\treDi}-\gb_\Di] \geq
r_{\Di}.
\end{multline}
Moreover, we set $t_{\Di}:=1\geq
\max_{\treDi<j\leq\Di}\beta_j^2$ and $\mu_{\Di}:=C_{\Ev}[\Di-\treDi]+C_{\Ev}(\mEvs[\treDi]\ObSoNoL)^{-1}[\gb_{\treDi}-\gb_\Di] \geq \Ex
S_{\Di}=v_{\Di}+r_{\Di}$. Consider now the event
$\cA_{\Di}:=\{-S_{\Di}-\mu_{\Di}\leq (C_{\Ev}[\Di -
\treDi]+2C_{\Ev}(\mEvs[\treDi]\ObSoNoL)^{-1}[\gb_{\treDi}-\gb_\Di])\}$
and its complement $\cA_{\Di}^c:=\{-S_{\Di}-\mu_{\Di}> (C_{\Ev}[\Di -
\treDi]+2C_{\Ev}(\mEvs[\treDi]\ObSoNoL)^{-1}[\gb_{\treDi}-\gb_\Di])\}$. By employing successively Lemma
\ref{tr:le:te}, \eqref{ad:le:pm:pr:e3} and $\gb_{\treDi}\leq \treRa$ it follows now from \eqref{ad:le:pm:pr:e1} that
\begin{multline*}
\Ex_{\TrSo} p_{\RvDi|\ObSo}(\Di)
\leq \Ex_{\TrSo}\exp\big(\{(-S_{\Di}-\mu_{\Di})+\mu_{\Di}+3C_\Ev[\treDi-\Di]\}/2\big)\Ind{\cA_{\Di}}
+ \Ex_{\TrSo}\Ind{\cA_{\Di}^c}\\
\leq\exp
\big(\{2C_{\Ev}[\Di - \treDi]+3C_{\Ev}(\mEvs[\treDi]\ObSoNoL)^{-1}[\gb_{\treDi}-\gb_\Di]+3C_\Ev[\treDi-\Di]\}/2\big)\\+\exp
\big(-\{C_{\Ev}[\Di - \treDi]+2C_{\Ev}(\mEvs[\treDi]\ObSoNoL)^{-1}[\gb_{\treDi}-\gb_\Di]\}/9\big)\\
\leq\exp
\big(\{C_{\Ev}[\treDi-\Di]+3C_{\Ev}(\mEvs[\treDi]\ObSoNoL)^{-1}\treRa\}/2\big)+\exp
\big(-C_\Ev [\Di-\treDi]/9 \big)\\
\leq\exp
\big(C_{\Ev}\{- \Di +3(\mEvs[\treDi]\ObSoNoL)^{-1}\treRa +L_\Ev(\ObSoNoL\mEvs[\treDi])^{-1}\treRa\}/2\big)\\+\exp
\big(-C_{\Ev}(\Di -L_\Ev(\ObSoNoL\mEvs[\treDi])^{-1}\treRa)/9\big)\\
\leq\exp
\big(C_{\Ev}\{- \Di
+5L_\Ev(\mEvs[\treDi]\ObSoNoL)^{-1}\treRa\}/2\big)\times \exp
\big(-\frac{C_{\Ev}L_\Ev\treRa}{2\mEvs[\treDi]\ObSoNoL} \big)\\+\exp
\big(-C_{\Ev}(\Di -L_\Ev(\ObSoNoL\mEvs[\treDi])^{-1}\treRa)/9\big)
\end{multline*}
Taking into account the definition \eqref{ad:de:mp} of $\peDi$, i.e.,
$\Di> 5 L_{\Ev}(\ObSoNoL\mEvs[\treDi])^{-1}\treRa$ for all $\DiMa\geq\Di>\peDi$, and
$L_\Ev\treRa(\ObSoNoL\mEvs[\treDi])^{-1}\geq \treDi$ due to Assumption \ref{ad:as:ev} \eqref{ad:as:ev:c}, we obtain
\begin{equation*}
\Ex_{\TrSo} p_{\RvDi|\ObSo}(\Di) \leq \exp\big(-\frac{L_\Ev C_{\Ev}\treRa}{2\ObSoNoL\mEvs[\treDi]}\big)
+\exp
\big(-\frac{4
L_{\Ev}C_{\Ev}\treRa}{9\ObSoNoL\mEvs[\treDi]}\big)\leq 2 \exp
\big(-\frac{4C_{\Ev}}{9}\treDi\big).
\end{equation*}
Thereby, $\Ex_{\TrSo}P_{\RvDi|\ObSo}(\peDi<\RvDi\leq \DiMa)=\sum_{\Di=\peDi+1}^{\DiMa}\Ex_{\TrSo} p_{\RvDi|\ObSo}(\Di)\leq 2 \exp
\big(-\frac{4C_{\Ev}}{9}\treDi + \log \DiMa \big)$ which shows the
assertion \ref{ad:le:pm:ii} and completes the proof.
\end{proof}
\begin{proof}[\noindent{\color{darkred}\sc Proof of Lemma
\ref{ad:le:pv}.}]\label{ad:le:pv:pr}
Consider \ref{ad:le:pv:i}. We start the proof with the observation that
due to Assumption
\ref{ad:as:ev} \eqref{ad:as:ev:c} the condition
\eqref{tr:as:pv:ac} holds true with $L=L_\Ev$ uniformly for all $m\in\Nz$ and
$\ObSoNoL\in(0,1)$, and hence imposing Assumption \ref{tr:as:pv} the conditions of Corollary
\ref{tr:co:pb} are satisfied, which in turn setting $c:=1/(9K)$ with $K:=((1+d^{-1})\vee
d^{-2}\HnormV{\TrSo-\RvSoEr}^2)L_\Ev$ implies
for all $1\leq \Di\leq\DiMa$ and
$\ObSoNoL\in(0,\ObSoNoL_{\TrSy})$ that
\begin{align}\label{ad:le:pv:pr:e1}
& \Ex_{\TrSo}P_{\DiRvSo[\Di]|\ObSo}(
\HnormV{\DiRvSo[\Di]-\TrSo}^2> (4+({11}/{2})K)[\gb_{\Di}\vee\ObSoNoL {\Di}
\oEvs[\Di]])\leq 2\exp(-{\Di}/{36});\\\label{ad:le:pv:pr:e2}
& \Ex_{\TrSo}P_{\DiRvSo[\Di]|\ObSo}( \HnormV{\DiRvSo[\Di]-\TrSo}^2<\{9(1+1/d)\}^{-1}[\gb_{\Di}\vee\ObSoNoL {\Di}
\oEvs])\leq 2\exp(-\Di/(162K^2)).\hfill
\end{align}
On the other hand side, taking into account the definition \eqref{ad:de:mp}
of $\peDi$ and $\meDi$, and the monotonicity of $(\gb_\Di)_{\Di\geq1}$ and $(\ObSoNoL \Di \oEvs[\Di])_{\Di\geq1}$
we have for all $\meDi\leq \Di\leq \treDi$ that
\[\ObSoNoL \Di \oEvs\leq \ObSoNoL \treDi
\oEvs[\treDi]\leq \treRa\quad\text{and}\quad\gb_{\Di}\leq 8 L_\Ev C_{\Ev}(1+1/d) \treRa \]
while for all $\peDi\geq\Di\geq\treDi$ (keeping in mind Assumption
\ref{ad:as:tr}) hold
\begin{multline*}
\Di \leq 5 L_\Ev (\ObSoNoL\mEvs[\treDi])^{-1} \treRa\leq 5 L_\Ev
(\ObSoNoL\mEvs[\treDi])^{-1} (\kappa^o)^{-1}\ObSoNoL \treDi
\oEvs[\treDi] \leq
(5
L_\Ev/\kappa^{\TrSy}) \treDi\leq D^{\TrSy}\treDi
\quad\text{and}\\\quad\gb_{\Di}\leq\gb_{\treDi}\leq \treRa
\end{multline*}
where $D^{\TrSy}:=D^{\TrSy}(\RvSoEr,\TrSo,\Ev):=\ceil{5 L_\Ev/\kappa^{\TrSy}}$. Due to
Assumption \ref{ad:as:ev} \eqref{ad:as:ev:b} and \eqref{ad:as:ev:c}
it follows from $\Di\leq D^{\TrSy}\treDi$ that $\mEvs\leq
\mEvs[D^{\TrSy}\treDi]\leq \mEvs[D^{\TrSy}]\mEvs[\treDi]$ and $\oEvs\leq \mEvs\leq
\mEvs[D^{\TrSy}]\mEvs[\treDi]\leq \mEvs[D^{\TrSy}] L_\Ev \oEvs[\treDi]$ which in
turn implies $\ObSoNoL \Di \oEvs\leq L_\Ev D^o\mEvs[D^{\TrSy}]\ObSoNoL
\treDi \oEvs[\treDi]\leq L_\Ev D^{\TrSy}\mEvs[D^{\TrSy}]\treRa$ for all $m\leq\peDi$. Combining the
upper bounds we have $(4+11K/2)[\gb_\Di\vee\ObSoNoL \Di \oEvs]\leq K^{\TrSy}\treRa$ for all $\meDi\leq\Di\leq\peDi$ since
$K^{\TrSy}\geq(4+11K/2)(8L_\Ev C_\Ev(1+1/d)\vee L_\Ev D^{\TrSy}\mEvs[D^{\TrSy}])$, and
together with \eqref{ad:le:pv:pr:e1} follows
\begin{multline*}
\sum_{\Di=\meDi}^{\peDi}\Ex_{\TrSo} P_{\DiRvSo|\ObSo}\big(
\HnormV{\DiRvSo[\Di]-\TrSo}^2> K^o\treRa\big)\\\leq \sum_{\Di=\meDi}^{\peDi}\Ex_{\TrSo}P_{\DiRvSo|\ObSo}\big(
\HnormV{\DiRvSo[\Di]-\TrSo}^2> (4+({11}/{2})K)[\gb_{\Di}\vee\ObSoNoL
{\Di}\oEvs[\Di]]\big)\\\leq 2\sum_{\Di=\meDi}^{\peDi}\exp(-m/36)\leq 74 \exp(-\meDi/36)
\end{multline*}
which proves the assertion \ref{ad:le:pv:i}. Consider now
\ref{ad:le:pv:ii}. We observe that by definition \eqref{tr:de:tre} of
$\treRa$ for all $\Di\in\Nz$ holds $\treRa\leq [\ObSoNoL \Di
\oEvs\vee\gb_\Di]$, and hence
$\{9(1+1/d)\}^{-1}[\gb_{\Di}\vee\ObSoNoL \Di \oEvs]\geq
(K^{\TrSy})^{-1}\treRa$ since $K^{\TrSy}\geq 9(1+1/d)$. Combining the last
estimate, \eqref{ad:le:pv:pr:e2} and $K^{\TrSy}\geq 10 K$ it follows that
\begin{multline*}
\sum_{\Di=\meDi}^{\peDi}\Ex_{\TrSo}
P_{\DiRvSo|\ObSo}\big(
\HnormV{\DiRvSo[\Di]-\TrSo}^2<(K^{\TrSy})^{-1}\treRa\big)\\\leq
\sum_{\Di=\meDi}^{\peDi}\Ex_{\TrSo}
P_{\DiRvSo|\ObSo}\big(
\HnormV{\DiRvSo[\Di]-\TrSo}^2<\{9(1+1/d)\}^{-1}[\gb_{\Di}\vee\ObSoNoL \Di \oEvs]
\big)\\\leq
2\sum_{\Di=\meDi}^{\peDi}\exp(-m/(K^{\TrSy})^2)\leq 4(K^{\TrSy})^2\exp(-\meDi/(K^{\TrSy})^2)
\end{multline*}
which shows the
assertion \ref{ad:le:pv:ii} and completes the proof.
\end{proof}
\begin{proof}[\noindent{\color{darkred}\sc Proof of Theorem
\ref{ad:th:ora}.}]\label{ad:th:ora:pr}
We start the proof with the observation that Lemma \ref{ad:le:pm}
together with Lemma \ref{ad:le:pv} \ref{ad:le:pv:i} imply
\begin{multline}\label{ad:th:ora:pr:e1}
\Ex_{\TrSo}P_{\RvDiRvSo|\ObSo}(
\HnormV{\RvDiRvSo-\TrSo}^2> K^o\treRa)=\Ex_{\TrSo}\sum_{\Di=1}^{\DiMa}p_{\RvDi|\ObSo}(\Di)P_{\DiRvSo|\ObSo}(
\HnormV{\DiRvSo[\Di]-\TrSo}^2> K^o\treRa)\\
\leq \Ex_{\TrSo}P_{\RvDi|\ObSo}(1\leq
\RvDi<\meDi)+\Ex_{\TrSo}P_{\RvDi|\ObSo}(\peDi<
\RvDi\leq\DiMa)\\\hfill + \sum_{\Di=\meDi}^{\peDi}\Ex_{\TrSo}P_{\DiRvSo|\ObSo}(
\HnormV{\DiRvSo[\Di]-\TrSo}^2> K^o\treRa)\\
\leq 4\exp\big(-\treDi\{C_{\Ev}/5-\log\DiMa/\treDi\}\big) + 74
\exp(-\meDi/36)
\end{multline}
On the other hand side, from Lemma \ref{ad:le:pm}
together with Lemma \ref{ad:le:pv} \ref{ad:le:pv:ii} also follows that
\begin{multline}\label{ad:th:ora:pr:e2}
\Ex_{\TrSo}P_{\RvDiRvSo|\ObSo}(
\HnormV{\RvDiRvSo-\TrSo}^2< (K^o)^{-1}\treRa)
\leq \Ex_{\TrSo}P_{\RvDi|\ObSo}(1\leq
\RvDi<\meDi)\\\hfill+\Ex_{\TrSo}P_{\RvDi|\ObSo}(\peDi<
\RvDi\leq\DiMa) + \Ex_{\TrSo}\sum_{\Di=\meDi}^{\peDi}p_{\RvDi|\ObSo}(\Di)P_{\DiRvSo|\ObSo}(
\HnormV{\DiRvSo[\Di]-\TrSo}^2< (K^o)^{-1}\treRa)\\
\leq 4\exp\big(-\treDi\{C_{\Ev}/5-\log\DiMa/\treDi\}\big) + 4(K^{\TrSy})^2
\exp(-\meDi/(K^{\TrSy})^2)
\end{multline}
By combining \eqref{ad:th:ora:pr:e1} and \eqref{ad:th:ora:pr:e2} we
obtain the assertion of the theorem since $\meDi,\treDi \to \infty $ and
$\log\DiMa/\treDi=o(1)$ as $\ObSoNoL\to0$ which completes the proof.
\end{proof}
\subsection{Proof of Theorem \ref{ad:th:bo}}
The next assertion presents a concentration inequality for
Gaussian random variables.
\begin{lem}\label{ad:le:te}Let the assumptions of Lemma \ref{tr:le:te}
be satisfied. For all $c\geq 0$ we have
\begin{align}\label{ad:le:te:e1}
&\sup_{m\geq 1}(6t_m)^{-1}\exp
\bigg(\frac{c(v_m+2r_m)}{4t_m}\bigg)\Ex\bigg(S_m-\Ex S_m-\frac{3}{2}c(v_m+2r_m)\bigg)_+\leq1
\end{align}
where $(a)_+:=(a\vee0)$.\end{lem}
\begin{proof}[\noindent{\color{darkred}\sc Proof of Lemma \ref{ad:le:te}.}]\label{ad:le:te:pr}
The assertion follows from Lemma \ref{tr:le:te} (keeping in mind that
$c\geq 1$), indeed
\begin{multline*}
\Ex\bigg(S_m-\Ex S_m-\frac{3}{2}c(v_m+2r_m)\bigg)_+=\int_{0}^\infty P(S_m-\Ex S_m \geq x+ \frac{3}{2}c(v_m+2r_m))dx\\ =
\int_{0}^\infty P(S_m-\Ex S_m \geq \frac{3}{2}(2x/(3(v_m+2r_m))+ c)(v_m+2r_m))dx\\
\leq\int_{0}^\infty \exp \bigg(-\frac{(2x/(3(v_m+2r_m))+
c)(v_m+2r_m)}{4t_m}\bigg)dx\\=\int_{0}^\infty \exp \bigg(-\frac{2x/3+
c(v_m+2r_m)}{4t_m}\bigg)dx\\
=\exp \bigg(-\frac{c(v_m+2r_m)}{4t_m}\bigg)\int_{0}^\infty \exp \bigg(-\frac{x}{6t_m}\bigg)dx=\exp \bigg(-\frac{c(v_m+2r_m)}{4t_m}\bigg)(6t_m)
\end{multline*}
\end{proof}
\begin{lem}\label{ad:le:bo:rem}
If Assumption \ref{tr:as:pv} and \ref{ad:as:ev} hold true then for all $\ObSoNoL\in(0,\ObSoNoL_{\TrSy})$
\begin{enumerate}[label=\emph{\textbf{(\roman*)}},ref=\emph{\textbf{(\roman*)}}]
\item\label{ad:le:bo:rem:i} $\sum_{j=1}^{\DiMa}\sigma_j^2\Ev_j^2\ObSoNoL^{-2}\Ex_{\TrSo}\{(\ObSo_j-\Ev_j\TrSo[j])P_{\RvDi|\ObSo}(j\leq
\RvDi\leq\DiMa)\}^2$\\
\null\hfill
$\leq\ObSoNoL \peDi \oEvs[\peDi]+
10\Evs_1\exp\big(-\treDi/5+2\log\DiMa\big)$;
\item\label{ad:le:bo:rem:ii}
$\sum_{j=1}^{\DiMa}(\RvSoEr_j-\TrSo[j])^2\Ex_{\TrSo}\Ex_{\RvDi|\ObSo}\{\Ind{\{1\leq\RvDi<j\}}
+ (\sigma_j/\RvSoVa_j)^2\Ind{\{j\leq\RvDi\leq \DiMa\}}\}+
\sum_{j>\DiMa}(\RvSoEr_j-\TrSo[j])^2$\\\null\hfill$\leq
\gb_{\meDi} +\HnormV{\RvSoEr-\TrSo}^2\{d^{-2}\ObSoNoL\mEvs[\meDi]+ 2\exp\big(-\treDi/5+\log\DiMa\big)\}$.
\end{enumerate}
\end{lem}
\begin{proof}[\noindent{\color{darkred}\sc Proof of Lemma \ref{ad:le:bo:rem}.}]\label{ad:le:bo:rem:pr}
Consider \ref{ad:le:bo:rem:i}. We start with the observation that the random variables
$\{\ObSoNo_j:=\ObSoNoL^{-1/2}(\ObSo_j-\Ev_j\TrSo[j])\}_{j\geq1}$ are
independent and standard normally distributed. Moreover, applying
Jensen's inequality we have
\begin{multline*}
\{\ObSoNo_jP_{\RvDi|\ObSo}(j\leq
\RvDi\leq\DiMa)\}^2=\{\Ex_{\RvDi|\ObSo}\ObSoNo_j\Ind{\{j\leq\RvDi\leq\DiMa\}}\}^2\leq \Ex_{\RvDi|\ObSo}\ObSoNo_j^2\Ind{\{j\leq\RvDi\leq\DiMa\}}
\\
=\ObSoNo_j^2P_{\RvDi|\ObSo}(j\leq
\RvDi\leq\DiMa).
\end{multline*}
We split the sum into two parts which we bound separately. Precisely,
\begin{multline}\label{ad:le:bo:rem:pr:e1}
\sum_{j=1}^{\DiMa}\sigma_j^2\Ev_j^2\ObSoNoL^{-2}\{(\ObSo_j-\Ev_j\TrSo[j])P_{\RvDi|\ObSo}(j\leq
\RvDi\leq\DiMa)\}^2\\
\leq \sum_{j=1}^{\peDi}\ObSoNoL\Evs_j\ObSoNo^2_j +
\sum_{j=1}^{\DiMa}\ObSoNoL\Evs_j\ObSoNo_j^2P_{\RvDi|\ObSo}(\peDi<
\RvDi\leq\DiMa)
\end{multline}
where we used that $\sigma_j\leq \ObSoNoL\Evs_j$.
Keeping in mind the notations used in Lemma \ref{ad:le:te} let
$S_{\DiMa}:=\sum_{j=1}^{\DiMa}\ObSoNoL\Evs_j\ObSoNo_j^2$ and
observe that $\alpha_j=0$ and $\beta_j^2=\ObSoNoL\Evs_j$, and hence
$r_{\DiMa}=0$. Keeping in mind that
$\DiMa:=\max\{1\leq m\leq \gauss{\ObSoNoL^{-1}}:
\ObSoNoL\mEvs\leq \Evs_1\}$ we set $t_{\DiMa}:=\Evs_1\geq \ObSoNoL \mEvs[\DiMa]
=\max_{1\leq j\leq\DiMa}\beta_j^2$ and $v_{\DiMa} := \Evs_1\DiMa =
\DiMa t_{\DiMa}\geq\sum_{j=1}^{\DiMa}\beta_j^2$, where $\Ex_{\TrSo} S_{\DiMa}\leq
v_{\DiMa}$. From Lemma \ref{ad:le:te} with $c=2/3$ follows that $
\Ex_{\TrSo}(S_{\DiMa}-2\Evs_1\DiMa)_+\leq
(6t_{\DiMa})\exp(-v_{\DiMa}/(6t_{\DiMa}))=(6\Evs_1)\exp(-\DiMa/6)$,
and hence
\begin{multline}\label{ad:le:bo:rem:pr:e2}
\sum_{j=1}^{\DiMa}\ObSoNoL\Evs_j\Ex_{\TrSo}\{\ObSoNo_j^2 P_{\RvDi|\ObSo}(\peDi<\RvDi\leq\DiMa)\}
\\\leq
\Ex\big(S_{\DiMa}-2\Evs_1\DiMa\big)_++2\Evs_1\DiMa\Ex_{\TrSo}P_{\RvDi|\ObSo}(\peDi<
\RvDi\leq\DiMa)\\
\leq 6\Evs_1\exp(-\DiMa/6)+2\Evs_1\DiMa\Ex_{\TrSo}P_{\RvDi|\ObSo}(\peDi<\RvDi\leq\DiMa).
\end{multline}
We distinguish two cases. First, if $\peDi=\DiMa$, then assertion \ref{ad:le:bo:rem:i} follows by combining
\eqref{ad:le:bo:rem:pr:e1} and $\Ex_{\TrSo}\ObSoNo_j^2=1$. Second, if
$\peDi<\DiMa$, then the definition \eqref{ad:de:mp} of $\peDi$ implies
$\DiMa>5\treDi$ which in turn implies the assertion \ref{ad:le:bo:rem:i} by combining
\eqref{ad:le:bo:rem:pr:e1}, $\Ex_{\TrSo}\ObSoNo_j^2=1$, \eqref{ad:le:bo:rem:pr:e2} and Lemma \ref{ad:le:pm} \ref{ad:le:pm:ii}. Consider
\ref{ad:le:bo:rem:ii}. Due to Assumption \ref{tr:as:pv} we have
$(\sigma_j/\RvSoVa_j)^{2}\leq (1\wedge
d^{-2}\ObSoNoL\Evs_j)$ which we will use without further reference. Splitting the first sum into two parts we
obtain
\begin{multline*}
\sum_{j=1}^{\DiMa}(\RvSoEr_j-\TrSo[j])^2\Ex_{\TrSo}\{\Ind{\{1\leq\RvDi<j\}}
+ (\sigma_j/\RvSoVa_j)^2\Ind{\{j\leq\RvDi\leq \DiMa\}}\}+ \sum_{j>\DiMa}(\RvSoEr_j-\TrSo[j])^2\\
\hspace*{5ex}\leq \sum_{j=1}^{\meDi}(\RvSoEr_j-\TrSo[j])^2\Ex_{\TrSo}\{\Ind{\{1\leq\RvDi<j\}}
+ d^{-2}\ObSoNoL\Evs_j\}\hfill\\
+ \sum_{j=\meDi+1}^{\DiMa}(\RvSoEr_j-\TrSo[j])^2\Ex_{\TrSo}\{\Ind{\{1\leq\RvDi<j\}}
+ \Ind{\{j\leq\RvDi\leq \DiMa\}}\}+
\sum_{j>\DiMa}(\RvSoEr_j-\TrSo[j])^2\\\hfill
\leq \HnormV{\RvSoEr-\TrSo}^2\{\Ex_{\TrSo}P_{\RvDi|\ObSo}(1\leq\RvDi<\meDi)+
d^{-2}\ObSoNoL\mEvs[\meDi]\}+\sum_{j>\meDi}(\RvSoEr_j-\TrSo[j])^2
\end{multline*}
The assertion \ref{ad:le:bo:rem:ii} follows now by combining
the last estimate and Lemma \ref{ad:le:pm} \ref{ad:le:pm:i}, which
completes the proof.
\end{proof}
\begin{proof}[\noindent{\color{darkred}\sc Proof of Theorem
\ref{ad:th:bo}.}]\label{ad:th:bo:pr}
We start the proof with the observation that
$\hSo_j-\TrSo[j]=(\DiPoEr{}{j}-\TrSo[j])P_{\RvDi|\ObSo}(j\leq \RvDi\leq
\DiMa)+(\RvSoEr_j-\TrSo[j])P_{\RvDi|\ObSo}(1\leq\RvDi<j)\}$ for all $1\leq
j\leq\DiMa$ and $\hSo_j-\TrSo[j]=\RvSoEr_j-\TrSo[j]$ for all $j>\DiMa$. From the identity
$\DiPoEr{}{j}-\TrSo[j]=(\sigma_j/\RvSoVa_j)(\RvSoEr_j-\TrSo[j])+(\sigma_j\Ev_j\ObSoNoL^{-1})(\ObSo_j-\Ev_j\TrSo[j])$
and Lemma \ref{ad:le:bo:rem} follows that
\begin{multline*}
\Ex_{\TrSo} \HnormV{\hSo-\TrSo}^2\leq \sum_{j=1}^{\DiMa}2\sigma^2\Ev^2_j\ObSoNoL^{-2}\Ex_{\TrSo}\{(\ObSo_j-\Ev_j\TrSo[j])P_{\RvDi|\ObSo}(j\leq \RvDi\leq
\DiMa)\\ + \sum_{j=1}^{\DiMa}2(\RvSoEr_j-\TrSo[j])^2\Ex_{\TrSo}\{(\sigma_j/\RvSoVa_j)P_{\RvDi|\ObSo}(j\leq \RvDi\leq
\DiMa)+
P_{\RvDi|\ObSo}(1\leq\RvDi<j)\}^2+\sum_{j>\DiMa}(\RvSoEr_j-\TrSo[j])^2\\
\leq 2\{\ObSoNoL \peDi \oEvs[\peDi]
+10\Evs_1\exp\big(-\treDi/5+2\log\DiMa\big)\}\\
+ 2\{\gb_{\meDi} +\HnormV{\RvSoEr-\TrSo}^2\{d^{-2}\ObSoNoL\mEvs[\meDi]+ 2\exp\big(-\treDi/5+\log\DiMa\big)\}\}.
\end{multline*}
On the other hand side, taking into account the definition \eqref{ad:de:mp} of $\meDi$ and
$\peDi$, we have show in the proof of Lemma \ref{ad:le:pv} that $\gb_{\meDi}\leq 8L_\Ev C_{\Ev}(1+1/d)\treRa$ and
$\ObSoNoL \peDi \oEvs[\peDi]\leq L_\Ev D^o\mEvs[D^o]\treRa$, while
trivially
$\ObSoNoL\mEvs[\meDi]\leq\ObSoNoL\mEvs[\treDi]\leq\treRa$. By
combination of these estimates we obtain
\begin{multline*}
\Ex_{\TrSo} \HnormV{\hSo-\TrSo}^2\leq \{2L_\Ev
D^o\mEvs[D^o]+16 L_\Ev C_{\Ev}(1+1/d)+2
d^{-2}\HnormV{\RvSoEr-\TrSo}^2\}\treRa\\
+(20\Evs_1+ 4\HnormV{\RvSoEr-\TrSo}^2)\exp\big(-\treDi/5+2\log\DiMa-\log\treRa\big)\}\treRa
\end{multline*}
From the last bound follows the assertion of the theorem since
$(2\log\DiMa-\log\treRa)/\treDi \to 0$ as $\ObSoNoL\to0$ which completes the proof.
\end{proof}
\subsection{Proof of Theorem \ref{ad:th:ra:mm}}
\begin{proof}[\noindent{\color{darkred}\sc Proof of Lemma
\ref{ad:le:opm}.}]\label{ad:le:opm:pr} The proof follows along the
lines of the proof of Lemma \ref{ad:le:pm}, where we replace
$\meDi$, $\peDi$, $\treDi$ and $\treRa$ by its counterpart
$\moeDi$, $\poeDi$, $\oeDi$ and $\oeRa$, respectively.
Moreover, we will use without further reference, that for all
$\TrSo\in\cwrSo$ the bias is bound by $\gb_\Di\leq
\rSo \wSo_m$, for all $m\in\Nz$, and hence $\gb_{\oeDi}\leq (1\vee\rSo)\oeRa$.
Consider \ref{ad:le:opm:i}. The claim holds trivially true
in case $\moeDi=1$, thus suppose $\moeDi>1$ and let
$1\leq \Di<\moeDi\leq \oeDi$. Define
$S_{\Di}:=\normV{\hSo^{\oeDi}-\RvSoEr}_{\sigma}^2-\normV{\hSo^\Di-\RvSoEr}_{\sigma}^2$. Let
$\cA_{\Di}$ and $\cA_{\Di}^c$,
respectively, be an event and its complement defined as in the Proof of Lemma \ref{ad:le:pm}, then it follows
\begin{equation}\label{ad:le:opm:pr:e1}
p_{\RvDi|\ObSo}(\Di)\leq \exp\bigg(\frac{1}{2}\big\{-S_{\Di}+3 C_\Ev [\oeDi-\Di]\big\}\bigg)\Ind{\cA_{\Di}}
+ \Ind{\cA_{\Di}^c}
\end{equation}
where $S_{\Di}=\sum_{j=\Di+1}^{\oeDi}\frac{\Ev_j^2\sigma_j}{\ObSoNoL^2}(Y_j-\Ev_j\RvSoEr_j)^2$.
We use the
notation introduced in Lemma \ref{ad:le:pm}, where again $1\geq
\beta_j^2\geq (1+1/d)^{-1}$ due to Assumption
\ref{tr:as:pv} and by
employing $\min_{\Di<j\leq\oeDi}\Ev_j^2\geq \mEvs[\oeDi]^{-1}$
together with Assumption
\ref{ad:as:ev} \eqref{ad:as:ev:c}
\begin{multline}\label{ad:le:opm:pr:e2}
L_\Ev (\ObSoNoL\mEvs[\oeDi])^{-1}(1\vee\rSo)\oeRa\geq L_\Ev (\ObSoNoL\mEvs[\oeDi])^{-1} \ObSoNoL \oeDi \oEvs[\oeDi] \geq \oeDi
\quad\text{ and }\\
(1+1/d)^{-1}(\ObSoNoL\mEvs[\oeDi])^{-1} [\gb_\Di-(1\vee\rSo)\oeRa] \leq (1+1/d)^{-1}(\ObSoNoL\mEvs[\oeDi])^{-1} [\gb_\Di-\gb_{\oeDi}] \leq
r_{\Di}.
\end{multline}
By employing successively Lemma
\ref{tr:le:te}, \eqref{ad:le:opm:pr:e2} and $\gb_{\oeDi}\leq
(1\vee\rSo)\oeRa$ for all $\TrSo\in\cwrSo$ it follows now from \eqref{ad:le:opm:pr:e1} that
\begin{multline*}
\Ex_{\TrSo} p_{\RvDi|\ObSo}(\Di)
\leq \exp
\big(-\frac{\gb_\Di}{4(1+1/d)\ObSoNoL\mEvs[\oeDi]} + \frac{2C_\Ev
L_\Ev(1\vee\rSo)\oeRa}{\ObSoNoL\mEvs[\oeDi]}\big)
\times\exp\big(-\frac{L_\Ev C_{\Ev}(1\vee\rSo)\oeRa}{4\ObSoNoL\mEvs[\oeDi]}\big)
\\
+\exp
\big(-\frac{[\gb_\Di-(1\vee\rSo)\oeRa]}{32(1+1/d)\ObSoNoL\mEvs[\oeDi]}\big).
\end{multline*}
Taking into account the definition \eqref{ad:de:omp} of $\moeDi$, i.e.,
$\gb_\Di> 8 L_{\Ev}C_{\Ev}(1+1/d)(1\vee\rSo)\oeRa$ for all $1\leq\Di<\moeDi$, and
$L_\Ev\oeRa(\ObSoNoL\mEvs[\oeDi])^{-1}\geq \oeDi$ due to Assumption \ref{ad:as:ev} \eqref{ad:as:ev:c}, we obtain
\begin{equation*}
\Ex_{\TrSo} p_{\RvDi|\ObSo}(\Di) \leq \exp\big(-\frac{L_\Ev C_{\Ev}(1\vee\rSo)\oeRa}{4\ObSoNoL\mEvs[\oeDi]}\big)
+\exp
\big(-\frac{7
L_{\Ev}C_{\Ev}(1\vee\rSo)\oeRa}{32\ObSoNoL\mEvs[\oeDi]}\big)\leq 2 \exp
\big(-\frac{7C_{\Ev}(1\vee\rSo)}{32}\oeDi\big).
\end{equation*}
Thereby, $\Ex_{\TrSo}P_{\RvDi|\ObSo}(1\leq
\RvDi<\moeDi)=\sum_{\Di=1}^{\moeDi-1}\Ex_{\TrSo} p_{\RvDi|\ObSo}(\Di)\leq 2 \exp
\big(-\frac{7C_{\Ev}(1\vee\rSo)}{32}\oeDi + \log \DiMa \big)$ using that
$\DiMa\geq \moeDi$ which proves the assertion \ref{ad:le:opm:i}.
Consider now \ref{ad:le:opm:ii}. The claim holds trivially true
in case $\poeDi=\DiMa$, thus suppose $\poeDi<\DiMa$ and let
$\DiMa\geq\Di>\poeDi\geq \oeDi$. Consider the upper bound \eqref{ad:le:opm:pr:e1} where $-S_{\Di}=\sum_{j=\oeDi+1}^{\Di}\frac{\Ev_j^2\sigma_j}{\ObSoNoL^2}(Y_j-\Ev_j\RvSoEr_j)^2$.
Employing the notations introduced in the
Proof of Lemma \ref{ad:le:pm} where we had $1\geq
\beta_j^2\geq (1+1/d)^{-1}$ due to Assumption
\ref{tr:as:pv}, we obtain from Assumption
\ref{ad:as:ev} \eqref{ad:as:ev:a} that
$\max_{\oeDi<j\leq\Di}\Ev_j^2\leq C_\Ev\mEvs[\oeDi]^{-1}$
and
taking into account in addition Assumption
\ref{ad:as:ev} \eqref{ad:as:ev:c} that
\begin{multline}\label{ad:le:opm:pr:e3}
L_\Ev (\ObSoNoL\mEvs[\oeDi])^{-1}(1\vee\rSo)\oeRa\geq \oeDi,\quad v_\Di\leq m-\oeDi
\quad\text{ and }\\
C_\Ev(\ObSoNoL\mEvs[\oeDi])^{-1}(1\vee\rSo)\oeRa \geq C_\Ev(\ObSoNoL\mEvs[\oeDi])^{-1} [\gb_{\oeDi}-\gb_\Di] \geq
r_{\Di}.
\end{multline}
By employing successively Lemma
\ref{tr:le:te}, \eqref{ad:le:opm:pr:e3} and $\gb_{\oeDi}\leq (1\vee\rSo)\treRa$ it follows now from \eqref{ad:le:opm:pr:e1} that
\begin{multline*}
\Ex_{\TrSo} p_{\RvDi|\ObSo}(\Di)
\leq\exp
\big(C_{\Ev}\{- \Di
+5L_\Ev(\mEvs[\oeDi]\ObSoNoL)^{-1}(1\vee\rSo)\oeRa\}/2\big)\times \exp
\big(-\frac{C_{\Ev}L_\Ev(1\vee\rSo)\oeRa}{2\mEvs[\oeDi]\ObSoNoL} \big)\\+\exp
\big(-C_{\Ev}(\Di -L_\Ev(\ObSoNoL\mEvs[\oeDi])^{-1}(1\vee\rSo)\oeRa)/9\big)
\end{multline*}
Taking into account the definition \eqref{ad:de:omp} of $\poeDi$, i.e.,
$\Di> 5 L_{\Ev}(\ObSoNoL\mEvs[\oeDi])^{-1}(1\vee\rSo)\oeRa$ for all $\DiMa\geq\Di>\poeDi$, and
$L_\Ev(1\vee\rSo)\oeRa(\ObSoNoL\mEvs[\oeDi])^{-1}\geq (1\vee\rSo)\oeDi$ due to Assumption \ref{ad:as:ev} \eqref{ad:as:ev:c}, we obtain
\begin{equation*}
\Ex_{\TrSo} p_{\RvDi|\ObSo}(\Di) \leq \exp\big(-\frac{L_\Ev C_{\Ev}(1\vee\rSo)\oeRa}{2\ObSoNoL\mEvs[\oeDi]}\big)
+\exp
\big(-\frac{4
L_{\Ev}C_{\Ev}(1\vee\rSo)\oeRa}{9\ObSoNoL\mEvs[\oeDi]}\big)\leq 2 \exp
\big(-\frac{4C_{\Ev}(1\vee\rSo)}{9}\oeDi\big).
\end{equation*}
Thereby, $\Ex_{\TrSo}P_{\RvDi|\ObSo}(\poeDi<\RvDi\leq \DiMa)=\sum_{\Di=\poeDi+1}^{\DiMa}\Ex_{\TrSo} p_{\RvDi|\ObSo}(\Di)\leq 2 \exp
\big(-\frac{4C_{\Ev}(1\vee\rSo)}{9}\oeDi + \log \DiMa \big)$ which shows the
assertion \ref{ad:le:opm:ii} and completes the proof.
\end{proof}
\begin{proof}[\noindent{\color{darkred}\sc Proof of Theorem
\ref{ad:th:ra:mm}.}]\label{ad:th:ra:mm:pr}
We start the proof with the observation that
due to Assumption
\ref{ad:as:ev} \eqref{ad:as:ev:c} the condition
\eqref{tr:as:pv:ac} holds true with $L=L_\Ev$ uniformly for all $m\in\Nz$ and
$\ObSoNoL\in(0,1)$, and hence imposing Assumption \ref{tr:as:pv} the conditions of Corollary
\ref{tr:co:pb} \eqref{tr:co:pb:e1} are satisfied, which in turn implies, by
setting $K:=((1+1/d)\vee
\rSo/d^{2})L_\Ev\geq ((1+d^{-1})\vee
d^{-2}\HnormV{\TrSo-\RvSoEr}^2)L_\Ev$,
that for all $1\leq \Di\leq\DiMa$ and
$\ObSoNoL\in(0,\ObSoNoL_{\OpSy})$
\begin{align}\label{ad:th:ra:mm:pr:e1}
& \Ex_{\TrSo}P_{\DiRvSo[\Di]|\ObSo}(
\HnormV{\DiRvSo[\Di]-\TrSo}^2> (4+(11/{2})K)[\gb_{\Di}\vee\ObSoNoL {\Di}
\oEvs[\Di]])\leq 2\exp(-{\Di}/{36}).\hfill
\end{align}
Moreover, exploiting the inequality below \eqref{tr:pr:pbm:pr:e2} with
$c_1=1/3$ and $c_2\geq 1$, it is possible to prove a slightly modified version of Corollary \ref{tr:co:pb} \eqref{tr:co:pb:e1} which implies for all $c_2\geq 1$
\begin{equation}\label{ad:th:ra:mm:pr:e2}
\Ex_{\TrSo}P_{\DiRvSo[\Di]|\ObSo}(
\HnormV{\DiRvSo[\Di]-\TrSo}^2> 16c_2K[\gb_{\Di}\vee\ObSoNoL {\Di}
\oEvs[\Di]])
\leq 2\exp(-c_2{\Di}/{12}).
\end{equation}%
Consider \ref{ad:th:ra:mm:i}. Following line by line the proof of
Lemma \ref{ad:le:pv} \ref{ad:le:pv:i}, using
\eqref{ad:th:ra:mm:pr:e1} rather than \eqref{ad:le:pv:pr:e1} and
exploiting $[\gb_\Di\vee\ObSoNoL\Di\oEvs]\leq 8 L_\Ev
C_\Ev(1+1/d){(1\vee\rSo)}\oeRa$ for all $\moeDi\leq \Di\leq\oeDi$ and
$[\gb_\Di\vee\ObSoNoL\Di\oEvs]\leq L_\Ev D^{\OpSy}\mEvs[D^{\OpSy}]{(1\vee\rSo)}\oeRa$ with $D^{\OpSy}:=\ceil{5 L_\Ev/\kappa^{\OpSy}}$
for all $\oeDi\leq \Di\leq\poeDi$ (keep in mind that $m\leq D^{\OpSy}\oeDi$) , we obtain
\begin{multline*}
\sum_{\Di=\moeDi}^{\poeDi}\Ex_{\TrSo} P_{\DiRvSo|\ObSo}\big(
\HnormV{\DiRvSo[\Di]-\TrSo}^2> K^{\OpSy}\oeRa\big)\\\leq \sum_{\Di=\moeDi}^{\poeDi}\Ex_{\TrSo}P_{\DiRvSo|\ObSo}\big(
\HnormV{\DiRvSo[\Di]-\TrSo}^2> (4+({11}/{2})K)[\gb_{\Di}\vee\ObSoNoL
{\Di}\oEvs[\Di]]\big)\\\leq 2\sum_{\Di=\moeDi}^{\poeDi}\exp(-m/36)\leq 74 \exp(-\moeDi/36).
\end{multline*}
Combining the last estimate, Lemma \ref{ad:le:opm} and the decomposition
\eqref{ad:th:ora:pr:e1} used in the proof of Theorem
\ref{ad:th:ora} (with $\meDi$ and $\peDi$ replaced by
$\moeDi$, $\poeDi$) it follows that
\begin{multline}\label{ad:th:ra:mm:pr:e3}
\Ex_{\TrSo}P_{\RvDiRvSo|\ObSo}(
\HnormV{\RvDiRvSo-\TrSo}^2> K^{\OpSy}\oeRa)
\\
\leq 4\exp\big(-\oeDi\{C_{\Ev}/5-\log\DiMa/\oeDi\}\big) + 74
\exp(-\moeDi/36)
\end{multline}
Taking into account that $\oeDi\to\infty$ and $\log \DiMa/\oeDi=o(1)$
as $\ObSoNoL\to0$, we obtain the assertion \ref{ad:th:ra:mm:i} of the
Theorem for any $\TrSo\in\cwrSo$ such that $\moeDi\to\infty$ as
$\ObSoNoL\to0$. On the other hand side, if $\TrSo\in\cwrSo$ such
that $\moeDi\not\to\infty$, i.e., $\sup_{\ObSoNoL}\moeDi<\infty$, then there exists
$\ObSoNoL_\TrSy\in(0,1)$ such that $\moeDi[G_{\ObSoNoL_{\TrSy}}]=\moeDi$ for all
$\ObSoNoL\in(0,\ObSoNoL_\TrSy)$ (keep in mind that
$(\moeDi)_{\ObSoNoL}$ is an integer-valued monotonically increasing
sequence). Moreover, by construction
$\gb_{\moeDi[G_{\ObSoNoL_{\TrSy}}]}\leq 8 L_\Ev C_\Ev(1+1/d){(1\vee\rSo)}\oeRa$ for
all $\ObSoNoL\in(0,\ObSoNoL_\TrSy)$ which in turn implies
$\gb_\Di\leq \gb_{\moeDi[G_{\ObSoNoL_{\TrSy}}]}=0$ for all $\Di\geq \moeDi[G_{\ObSoNoL_{\TrSy}}]$, since $\oeRa=o(1)$ as
$\ObSoNoL\to0$. Thereby, for all $\Di\geq \moeDi[G_{\ObSoNoL_{\TrSy}}]$
follows $\oeRa/[\gb_{\Di}\vee\ObSoNoL {\Di}
\oEvs[\Di]]=\oeRa/[\ObSoNoL {\Di}\oEvs[\Di]]\geq[\ObSoNoL
{\oeDi}\oEvs[\oeDi]]/[\ObSoNoL {\Di}\oEvs[\Di]]\geq \oeDi/[L_\Ev
\Di]$ using that $L_\Ev\oEvs[\oeDi]\geq\mEvs[\oeDi]\geq\mEvs\geq\oEvs$ due
to Assumption \ref{ad:as:ev} \eqref{ad:as:ev:c}, which in turn together
with $K^{\OpSy}\oeRa/[\gb_{\Di}\vee\ObSoNoL {\Di}
\oEvs[\Di]]\geq K^{\OpSy}\oeDi/[L_\Ev {\Di}] = 16 c_2 K$, $c_2:=(8C_\Ev(1+1/d)\vee
D^{\OpSy}\mEvs[D^{\OpSy}]){(1\vee\rSo)} \oeDi/{\Di}\geq1$ and \eqref{ad:th:ra:mm:pr:e2} implies
\begin{multline*}
\sum_{\Di=\moeDi}^{\poeDi}\Ex_{\TrSo} P_{\DiRvSo|\ObSo}\big(
\HnormV{\DiRvSo[\Di]-\TrSo}^2> K^{\OpSy}\oeRa\big)
\\\leq 2\exp(-(8C_\Ev(1+1/d)\vee
D^{\OpSy}\mEvs[D^{\OpSy}]){(1\vee\rSo)} \oeDi/12 + \log \DiMa)
\\\leq 2 \exp(-C_\Ev\oeDi/5 + \log \DiMa).
\end{multline*}
Consequently, we have
\begin{equation*}\Ex_{\TrSo}P_{\RvDiRvSo|\ObSo}(
\HnormV{\RvDiRvSo-\TrSo}^2> K^{\OpSy}\oeRa)
\leq 6\exp\big(-\oeDi\{C_{\Ev}/5-\log\DiMa/\oeDi\}\big)
\end{equation*}
which shows that assertion \ref{ad:th:ra:mm:i} holds for any
$\TrSo\in\cwrSo$ since $\oeDi\to\infty$ and $\log \DiMa/\oeDi=o(1)$
as $\ObSoNoL\to0$. Consider \ref{ad:th:ra:mm:ii}. Employing that
for all $\TrSo\in\cwrSo$ it holds $K^{\OpSy}\oeRa\geq 16
K[\gb_{\Di}\vee\ObSoNoL {\Di}\oEvs[\Di]]$ for all $\moeDi\leq
\Di\leq\poeDi$ it follows that $K_\ObSoNoL\oeRa/[\gb_{\Di}\vee\ObSoNoL {\Di}
\oEvs[\Di]]\geq 16 c_2 K$ where $c_2:= K_{\ObSoNoL}/K^{\OpSy}\geq 12$ for
all $\ObSoNoL\in(0,\widetilde{\ObSoNoL}_{\OpSy})$ since $K_{\ObSoNoL}\to\infty$ as
$\ObSoNoL\to0$. Therefore, by applying
\eqref{ad:th:ra:mm:pr:e2} we have
\begin{equation*}
\sum_{\Di=\moeDi}^{\poeDi}\Ex_{\TrSo} P_{\DiRvSo|\ObSo}\big(
\HnormV{\DiRvSo[\Di]-\TrSo}^2> {K_{\ObSoNoL}}\oeRa\big)
\leq 4 \exp(-K_{\epsilon}/[12K^{\OpSy}]).
\end{equation*}
and hence from Lemma \ref{ad:le:opm} follows for all
$\ObSoNoL\leq (\widetilde{\ObSoNoL}_{\OpSy}\wedge\ObSoNoL_{\OpSy})$
\begin{multline*}
\Ex_{\TrSo}P_{\RvDiRvSo|\ObSo}(
\HnormV{\RvDiRvSo-\TrSo}^2> K_{\ObSoNoL}\oeRa)
\\\leq 4\exp\big(-\oeDi\{C_{\Ev}/5-\log\DiMa/\oeDi\}\big) + 4\exp(-K_{\ObSoNoL}/[12K^{\OpSy}]).
\end{multline*}
Observe, that $(\widetilde{\ObSoNoL}_{\OpSy}\wedge\ObSoNoL_{\OpSy})$
depends only on the class $\cwrSo$ and thus the upper bound given in the last display holds true
uniformly for all $\TrSo\in\cwrSo$,
which implies the assertion
\ref{ad:th:ra:mm:ii} by using that $K_{\ObSoNoL}\to\infty$, $\oeDi\to\infty$ and $\log \DiMa/\oeDi=o(1)$
as $\ObSoNoL\to0$, and completes the proof.
\end{proof}
\subsection{Proof of Theorem \ref{ad:th:be:mm}}
\begin{lem}\label{ad:le:be:mm:rem}
If Assumption \ref{tr:as:pv} and \ref{ad:as:ev} hold true then for all
$\TrSo\in\cwrSo$ and $\ObSoNoL\in(0,\ObSoNoL_o)$
\begin{enumerate}[label=\emph{\textbf{(\roman*)}},ref=\emph{\textbf{(\roman*)}}]
\item\label{ad:le:be:mm:rem:i} $\sum_{j=1}^{\DiMa}\sigma_j^2\Ev_j^2\ObSoNoL^{-2}\Ex_{\TrSo}\{(\ObSo_j-\Ev_j\TrSo[j])P_{\RvDi|\ObSo}(j\leq
\RvDi\leq\DiMa)\}^2$
\\\null\hfill
$\leq\ObSoNoL \poeDi \oEvs[\poeDi]+ 10\Evs_1\exp\big(-\oeDi/5+2\log\DiMa\big)$;
\item\label{ad:le:be:mm:rem:ii}
$\sum_{j=1}^{\DiMa}(\RvSoEr_j-\TrSo[j])^2\Ex_{\TrSo}\Ex_{\RvDi|\ObSo}\{\Ind{\{1\leq\RvDi<j\}}
+ (\sigma_j^2\RvSoVa_j^{-1})^2\Ind{\{j\leq\RvDi\leq \DiMa\}}\}+
\sum_{j>\DiMa}(\RvSoEr_j-\TrSo[j])^2$\\\null\hfill$\leq
\gb_{\moeDi} +\HnormV{\RvSoEr-\TrSo}^2\{d^{-2}\ObSoNoL\mEvs[\moeDi]+ 2\exp\big(-\frac{C_{\Ev}(1\vee\rSo)}{5}\oeDi+\log\DiMa\big)\}$.
\end{enumerate}
\end{lem}
\begin{proof}[\noindent{\color{darkred}\sc Proof of Lemma
\ref{ad:le:be:mm:rem}.}]\label{ad:le:be:mm:rem:pr}
The proof follows along the lines of the proof of Lemma
\ref{ad:le:bo:rem}, where we replace
$\meDi$, $\peDi$, $\treDi$ and $\treRa$ by its counterpart
$\moeDi$, $\poeDi$, $\oeDi$ and $\oeRa$, respectively.
Consider \ref{ad:le:be:mm:rem:i}.
Following the proof of \eqref{ad:le:bo:rem:pr:e1} it is
straightforward to see that
\begin{multline}\label{ad:le:be:mm:rem:pr:e1}
\sum_{j=1}^{\DiMa}\sigma_j^2\Ev_j^2\ObSoNoL^{-2}\{(\ObSo_j-\Ev_j\TrSo[j])P_{\RvDi|\ObSo}(j\leq
\RvDi\leq\DiMa)\}^2
\\
\leq \sum_{j=1}^{\poeDi}\ObSoNoL\Evs_j\ObSoNo^2_j +
\sum_{j=1}^{\DiMa}\ObSoNoL\Evs_j\ObSoNo_j^2P_{\RvDi|\ObSo}(\poeDi<
\RvDi\leq\DiMa)
\end{multline}
and following line by line the proof of \eqref{ad:le:bo:rem:pr:e2} we
conclude
\begin{multline}\label{ad:le:be:mm:rem:pr:e2}
\sum_{j=1}^{\DiMa}\ObSoNoL\Evs_j\Ex_{\TrSo}\{\ObSoNo_j^2 P_{\RvDi|\ObSo}(\poeDi<\RvDi\leq\DiMa)\}\\
\leq 6\Evs_1\exp(-\DiMa/6)+2\Evs_1\DiMa\Ex_{\TrSo}P_{\RvDi|\ObSo}(\poeDi<\RvDi\leq\DiMa).
\end{multline}
We distinguish two cases. First, if $\poeDi=\DiMa$, then assertion \ref{ad:le:be:mm:rem:i} follows by combining
\eqref{ad:le:be:mm:rem:pr:e1} and $\Ex_{\TrSo}\ObSoNo_j^2=1$. Second, if
$\poeDi<\DiMa$, then the definition \eqref{ad:de:omp} of $\poeDi$ implies
$\DiMa>5\oeDi$ which in turn implies the assertion \ref{ad:le:be:mm:rem:i} by combining
\eqref{ad:le:be:mm:rem:pr:e1}, $\Ex_{\TrSo}\ObSoNo_j^2=1$, \eqref{ad:le:be:mm:rem:pr:e2} and Lemma \ref{ad:le:opm} \ref{ad:le:opm:i}. Consider
\ref{ad:le:bo:rem:ii}.
Following the proof of Lemma \ref{ad:le:bo:rem} \ref{ad:le:bo:rem:ii}
we obtain
\begin{multline*}
\sum_{j=1}^{\DiMa}(\RvSoEr_j-\TrSo[j])^2\Ex_{\TrSo}\{\Ind{\{1\leq\RvDi<j\}}
+ (\sigma_j/\RvSoVa_j)^2\Ind{\{j\leq\RvDi\leq \DiMa\}}\}+ \sum_{j>\DiMa}(\RvSoEr_j-\TrSo[j])^2\\
\leq \HnormV{\RvSoEr-\TrSo}^2\{\Ex_{\TrSo}P_{\RvDi|\ObSo}(1\leq\RvDi<\moeDi)+
d^{-2}\ObSoNoL\mEvs[\moeDi]\}+\sum_{j>\moeDi}(\RvSoEr_j-\TrSo[j])^2
\end{multline*}
The assertion \ref{ad:le:be:mm:rem:ii} follows now by combining
the last estimate and Lemma \ref{ad:le:opm} \ref{ad:le:opm:ii}, which
completes the proof.
\end{proof}
\begin{proof}[\noindent{\color{darkred}\sc Proof of Theorem
\ref{ad:th:be:mm}.}]\label{ad:th:be:mm:pr}
The proof follows line by line the proof of Theorem \ref{ad:th:bo}
using Lemma \ref{ad:le:be:mm:rem} rather than Lemma
\ref{ad:le:bo:rem}, more precisely from Lemma \ref{ad:le:be:mm:rem} follows
\begin{multline*}
\Ex_{\TrSo} \HnormV{\hSo-\TrSo}^2\leq \sum_{j=1}^{\DiMa}2\sigma^2\Ev^2_j\ObSoNoL^{-2}\Ex_{\TrSo}\{(\ObSo_j-\Ev_j\TrSo[j])P_{\RvDi|\ObSo}(j\leq \RvDi\leq
\DiMa)\\ + \sum_{j=1}^{\DiMa}2(\RvSoEr_j-\TrSo[j])^2\Ex_{\TrSo}\{(\sigma_j/\RvSoVa_j)P_{\RvDi|\ObSo}(j\leq \RvDi\leq
\DiMa)+
P_{\RvDi|\ObSo}(1\leq\RvDi<j)\}^2+\sum_{j>\DiMa}(\RvSoEr_j-\TrSo[j])^2\\
\leq 2\{\ObSoNoL \poeDi \oEvs[\poeDi]
+10\Evs_1\exp\big(-\frac{C_{\Ev}(1\vee\rSo)}{5}\oeDi+2\log\DiMa\big)\}\\
+ 2\{\gb_{\moeDi} +\HnormV{\RvSoEr-\TrSo}^2\{d^{-2}\ObSoNoL\mEvs[\moeDi]+ 2\exp\big(-\frac{C_{\Ev}(1\vee\rSo)}{5}\oeDi+\log\DiMa\big)\}\}.
\end{multline*}
Taking further into account the definition \eqref{ad:de:omp} of $\moeDi$ and
$\poeDi$, we have $\gb_{\moeDi}\leq 8L_\Ev
C_{\Ev}(1+1/d)(1\vee\rSo)\oeRa$ and (keeping in mind Assumption \ref{tr:as:mi})
$\poeDi \leq
D^\star\oeDi$
with $D^\star:=D^\star(\cwrSo,\Ev):=\ceil{5 L_\Ev(1\vee\rSo)/\kappa}$,
which in turn implies $\ObSoNoL \poeDi \oEvs[\poeDi]\leq L_\Ev D^\star\mEvs[D^\star]\oeRa$, while
trivially
$\ObSoNoL\mEvs[\moeDi]\leq\ObSoNoL\mEvs[\oeDi]\leq\oeRa$ and $\HnormV{\RvSoEr-\TrSo}^2\leq\rSo$. By
combination of these estimates we obtain uniformly for all $\TrSo\in\cwrSo$ that
\begin{multline*}
\Ex_{\TrSo} \HnormV{\hSo-\TrSo}^2\leq \{2L_\Ev
D^\star\mEvs[D^\star]+16 L_\Ev C_{\Ev}(1+1/d)(1\vee\rSo)+2
d^{-2}\rSo\}\oeRa\\
+(20\Evs_1+ 4\rSo)\exp\big(-\frac{C_{\Ev}(1\vee\rSo)}{5}\oeDi+2\log\DiMa-\log\oeRa\big)\}\oeRa.
\end{multline*}
Note that in the last display the multiplicative factors of $\oeRa$ depend only on the class
$\cwrSo$, the constant $d$ and the sequence $\Ev$. Thereby, the
assertion of the theorem follows from $\log(\DiMa/\oeRa)/\oeDi \to 0$ as $\ObSoNoL\to0$ which completes the proof.
\end{proof}
|
\section{Introduction}
Mathematical conservation laws are integro-differential evolution equations, such as Navier-Stokes and Burgers equations, expressing the physical principles of conservation of mass, energy, momentum, enstrophy, etc., in different dynamical situations. With this paper we initiate a program of investigation of explicit representations for asymptotic behavior of solutions of conservation laws driven by multiscale, $(\alpha_1,\dots,\alpha_k)$-stable, L\'{e}vy processes (multifractal anomalous diffusions). Since the asymptotic behavior of such conservation laws is determined, in some cases, by their linearized versions, the starting point here is to obtain exact representation, via known special functions such as Meijer G functions, of the solutions of linear multiscale evolution equations, that is, for the PDFs of the multiscale L\'evy processes themselves.
The idea is to produce a framework that permits a straightforward calculation of probabilities related to the mutiscale diffusions using a symbolic manipulation platform such as {\it Mathematica}, and a fairly standard set of special functions that have been in use in this area for a long time.
The plan of the paper is as follows. We begin, in Section 2, with a review of the known general results from \cite{BFW1999, BKW1999, BKW2001, BKW2001a} on the asymptotics of solutions of multifractal conservation laws, and apply them to the case of general asymmetric two sided multiscale diffusion. In the case of supercritical nonlinearity asymptotics is dictated by the linear part of the equation so, in Section 3, we produce an exact representation of solutions of linearized equations in the general asymmetric case; simpler representation are then deduced in the symmetric case. In Section 4, anticipating our future needs to obtain explicit solutions for equations describing subdiffusive anomalous diffusions, where the time is also "fractal" \cite{PSW2005}, we obtain an explicit representation for totally asymmetric $\alpha$-stable diffusions, for $0 < \alpha < 1$. Conclusions, as well as a discussion of the relevant moment problem, can be found in Section 5.
In the remainder of this section we establish the notation and provide basic definitions of the integral transforms and special functions we are going to work with. We start with the Fourier transform of an integrable function $f(x)$ defined for $x \in \mathbb{R}$, and real $\omega$ \cite{INSneddon72},
\begin{align}\label{26/08/14-1a}
\tilde{f} (\omega) &= \mathcal{F}[f (x); \omega] = \int_{-\infty}^{\infty} e^{i \omega x} f (x) dx,\\ \label{26/08/14-1b}
f (x) &= \mathcal{F}^{-1}[\tilde{f} (\omega); x] = \frac{1}{2\pi} \int_{-\infty}^{\infty} e^{-i \omega x} \tilde{f} (\omega) d\omega.
\end{align}
For a function $f (x) \equiv 0$, for $x < 0$, such that $ e^{- c x} f (x)$ is integrable on the positive half-line for some fixed number $c > 0$, the Laplace transform
\begin{align}\label{26/08/14-2a}
f^{\star} (p) &= \mathcal{L}[f (x); p] = \int_{0}^{\infty} e^{-p x} f (x) dx, \\ \label{26/08/14-2b}
f (x) &= \mathcal{L}^{-1}[f^{\star} (p); x] = \frac{1}{2\pi i} \int_{c-i \infty}^{c+i \infty}e^{p x} f^{\star} (p) dp,
\end{align}
where $p = c + i \omega$. There is an obvious relationship between the Fourier transform of $f_{1}(x) = e^{- c x} f_{2}(x)$ and the Laplace transform of $f_2(x)$, see e.g., \cite{INSneddon72} for more information. Finally, the Mellin transform of $f (x)$ is here defined as follows:
\begin{align}
\label{26/08/14-3a}
\hat{f} (s) &= \mathcal{M}[f (x); s] = \int_{0}^{\infty} x^{s-1} f (x) dx,\\
\label{26/08/14-3b}
f (x) &= \mathcal{M}^{-1}[\hat{f} (s); x] = \frac{1}{2\pi i} \int_{L} x^{-s} \hat{f} (s) ds,
\end{align}
where $s$ is a complex variable \cite{INSneddon72}. The contour of integration $L$ is determined by the domain of analyticity of $\hat{f}_{2}(s)$ and, usually, it is an infinite strip parallel to the imaginary axis. The conditions under which the integrals in Eqs. \eqref{26/08/14-1a}-\eqref{26/08/14-3b} converge can be found in \cite{INSneddon72}.
The Meijer $G$ function will play a pivotal role in what follows. It is defined as the inverse Mellin transform of products and ratios of the classical Euler's gamma functions. More precisely, see \cite{APPrudnikov-v3, NIST},
\begin{equation}\label{26/08/14-4}
G^{m, n}_{p, q}\left(z\Big\vert {A_{1} \ldots A_{p} \atop B_{1} \ldots B_{q}}\right) = \mathcal{M}^{-1}\left[\frac{\prod_{j=1}^{m} \Gamma(B_{j} + s)\, \prod_{j=1}^{n} \Gamma(1 - A_{j} - s)}{\prod_{j=m+1}^{q} \Gamma(1 - B_{j} - s) \prod_{j=n+1}^{p} \Gamma(A_{j} + s)}; z\right],
\end{equation}
where empty products in Eq. \eqref{26/08/14-4} are taken to be equal to 1. Eq. \eqref{26/08/14-4} holds under the following assumptions:
\begin{align}
\label{26/08/14-5}
&z\neq 0, \quad 0 \leq m \leq q, \quad 0 \leq n \leq p, \nonumber\\
&A_{j}\in\mathbb{C}, \quad j = 1, \ldots, p; \quad B_{j}\in\mathbb{C}, \quad j = 1, \ldots, q.
\end{align}
A description of the integration contours in Eq. \eqref{26/08/14-4}, and the general properties and special cases of the Meijer $G$ functions can be found in \cite{APPrudnikov-v3}. If the integral in Eq. \eqref{26/08/14-4} converges and if no confluent poles appear among $\Gamma(1 - A_{j} - s)$ or $\Gamma(1- B_{j} - s)$, then the Meijer $G$ function can be expressed as a finite sum of the generalized hypergeometric function, see formulas (8.2.2.3) and (8.2.2.4) on p. 520 of \cite{APPrudnikov-v3}. Recall that a generalized hypergeometric function can be represented in terms of the following series, see Eq. (7.2.3.1) on p. 368 of \cite{APPrudnikov-v3}:
\begin{equation}\label{26/08/14-6}
{_{p}F_{q}}\left({ a_{1}, \ldots, a_{p} \atop b_{1}, \ldots, b_{q}}; x\right) = \sum_{n=0}^{\infty} \frac{x^{n}}{n!} \frac{\prod_{j=1}^{p} (a_{j})_{n}}{\prod_{j=1}^{q} (b_{j})_{n}},
\end{equation}
where the upper and lower lists of parameters are denoted by $(a_{p})$ and $(b_{q})$, respectively, and $(a)_{n} = \Gamma(a + n)/\Gamma(a)$ is the Pochhammer symbol.
To conclude the introduction we find it convenient to introduce the special notation for a specific uniform partition of the unit interval,
\begin{equation}\label{speciallist}
\Delta(n, a)=\{ \ulamek{a}{n}, \ulamek{a+1}{n}, \ldots, \ulamek{a+n-1}{n}\}.
\end{equation}
For later reference, we also quote the Euler's reflection formula,
\begin{equation}\label{26/08/14-7}
\Gamma(z) \Gamma(1-z) = \frac{\pi}{\sin(\pi z)},
\end{equation}
see Eq. (8.334.3) on p. 896 in \cite{Gradshteyn},
and the Gauss-Legendre multiplication formula,
\begin{equation}\label{26/08/14-8}
\Gamma(n a) = (2\pi)^{\frac{1-n}{2}} n^{n a - \frac{1}{2}} \prod_{j=0}^{n-1} \Gamma\left(a + \frac{j}{n}\right),
\end{equation}
see Eq. (8.335) on p. 896 of \cite{Gradshteyn}.
\section {Multiscale L\'evy processes}
In this Section we are turning to a review of infinitesimal generators $\cal A$ of semigroups associated with 1-D multiscale $(\alpha_1,\dots,\alpha_k)$-stable L\'evy processes driving the evolution equations of the form,
\begin{equation}\label{29/12/2014-1}
\frac{\partial u}{\partial t}+{\cal A} u + \frac{\partial}{\partial x}g(u) =0, \qquad u(0, x) = u_{0}(x),
\end{equation}
where $u=u(t,x), t\ge 0, x \in \mathbb R$, $u_{0}$ is an initial condition which will be specified later, and $g:\mathbb R\mapsto \mathbb R$ is a (nonlinear) function. Such equations are often called fractal, or anomalous conservation laws \cite{BKW1999,BKW2001}. Their asymptotic behavior will be discussed in Section 3.
The operators $\cal A$ are easiest to describe in terms of their actions in the Fourier domain; they are so-called Fourier multiplier operators. Let us begin by recalling the basic terminology and establishing the notation.
Like any Markov processes\footnote{See, e.g., \cite{JS2001}, for basic information in this area.}, the L\'evy process, $X_t, t>0$, has associated with it a semigroup $P_t$ of convolution operators\footnote{ That is, $P_{t+s} = P_t P_s$, $t, s > 0$.} acting on a bounded function $\phi(x)$ via the formula,
\begin{equation}\label{29/12/2014-2}
P_t \phi(x)=\bE ^x(\phi(X(t))=\int_{\mathbb R} \phi(x+y)\,P(X(t)\in dy).
\end{equation}
The {\it infinitesimal generator} $\cal A$ of such a semigroup is defined by the formula,
\begin{equation}\label{29/12/2014-3}
-{\cal A}=\lim_{h\to 0} {P_h-P_0\over h} ,
\end{equation}
and the family of functions,$v(t, x)~=~P_t g(x)$, interpreted here as probability density functions (PDFs), clearly satisfies the (generalized) Fokker-Planck evolution equation,
\begin{equation}\label{29/12/2014-4}
\frac{\partial v}{\partial t} = -{\cal A} v,
\end{equation}
because $ \lim\limits_{h\to 0} (P_{t+ h}-{P_t})/h = \lim\limits_{h\to 0}[({P_{ h}-P_0)/ h}]P_t =- {\cal A}P_t$.
In the case of a general L\'evy processes $X_t$, we have the identity,
\begin{equation}\label{29/12/2014-5}
{\cal F}({\cal A} \phi)(\omega)= \psi (\omega){\cal F}\phi(\omega),
\end{equation}
where $\cal F$ stands for the Fourier transform, and
\begin{equation}\label{29/12/2014-6}
\psi (\omega)=\log \mathbb{E}[e^{i\omega X_1}]
\end{equation}
is the characteristic exponent of $X_1$, which is necessarily (see, e.g., \cite{B1996}), of the form
\begin{equation}\label{29/12/2014-7}
\psi(\omega )= i\mu \omega - \frac{(\sigma \omega )^2 }{2} + \int_{\mathbb{R}}(e^{i \omega x}-1-i\omega x {\bf I}_{|x|<1})\Lambda(dx),
\end{equation}
where $\mu \in \mathbb{R}$, $ \sigma \in \mathbb{R_{+}}$, and $\Lambda$ is a nonegative measure on $\mathbb R$, satisfying the conditions $\Lambda(\{0\})=0$, and $\int_{\mathbb{R}}(1 \wedge |x|^2)\Lambda(dx) < \infty$. The triplet $(\mu,\sigma,\Lambda)$ is called the \textit{characteristic triplet} of $X$, $\mu\in\mathbb{R}$ -- the drift coefficient, $ \sigma> 0 $ -- the Gaussian, or diffusion coefficient, and $ \Lambda $ -- the L\'evy measure of $X_1$. The L\'evy measure describes the ``intensity'' of jumps of a certain height of a L\'evy process in a time interval of length 1.
Observe that
\begin{align*}
{\cal F}(P_t \phi)(\omega) & = \left(\int_R e^{-i\omega x}\bE \phi(X_t+x)\,dx\right) =\bE \left(\int_Re^{-i\omega (y-X_t)} \phi(y)\,dy\right) \\
& = \bE e^{i\omega X_t}\int_Re^{-i\omega y} \phi(y)\,dy=\exp( t\psi( \omega )){\cal F} \phi(\omega ),
\end{align*}
which, in view of Eq. \eqref{29/12/2014-3}, indeed implies Eq. \eqref{29/12/2014-5}.
In the case of the usual Brownian motion the infinitesimal operator ${\cal A} $ is just the 1-D classical Laplacian $\Delta$ (the second derivative operator). For the self-similar (single-scale) symmetric $\alpha$-stable process $X_t$, the infinitesimal generator is the 1-D fractional Laplacian {$-(-\Delta)^{\alpha/2}, \; 0 < \alpha \leq 2$,} corresponding to the characteristic exponent (Fourier multiplier) $\psi(\omega)=-|\omega|^\alpha$.
In what follows we focus our attention on the multiscale (and not necessarily symmetric) L\'evy processes with the characteristic functions of the form,
\begin{equation}\label{30Jun14-2}
\bE e^{i\omega X_t}=\prod_{j=1}^n \tilde{v}_{\alpha_{j}, \beta_{j},\gamma_j}(\omega, t),
\end{equation}
where, for each $j=1,2,\dots, n$,
\begin{equation}\label{29/12/2014-8}
\tilde{v}_{\alpha_{j}, \beta_{j}, \gamma_{j}}(t, \omega)= \mathcal{F}[v_{\alpha_{j}, \beta_{j}, \gamma_{j}}(t, x);\omega] = \exp\left[- t \gamma_j|\omega|^{\alpha_{j}} e^{\frac{i\pi}{2} \beta_{j}\, {\rm sgn}(\omega)}\right]
\end{equation}
The symbol ${\rm sgn}(\omega)$ denotes the sign of the parameter $\omega$. The multiparameter $(\vec \alpha;\vec \beta;\vec \gamma) = (\alpha_1,\dots,\alpha_n;\beta_1,\dots \beta_n;\gamma_1,\dots,\gamma_n)$ has to satisfy the following conditions: If $0 < \alpha_{j} < 1$ then $|\beta_{j}| \leq \alpha_{j}$, and if $1 < \alpha_{j} \leq 2$ then $|\beta_{j}| \leq 2 - \alpha_{j}$; for all $j$, we assume that $\gamma_j > 0$ . Thus the Fourier multiplier describing the infinitesimal generator of $X_t$ is of the form,
\begin{equation}\label{multisymbol}
\psi_{{(\vec \alpha;\vec \beta;\vec \gamma)} }(\omega)=\sum_{j=1}^n - \gamma_j|\omega|^{\alpha_{j}} e^{\frac{i\pi}{2} \beta_{j}\, {\rm sgn}(\omega)}.
\end{equation}
The generator itself will be denoted ${\cal A}_ {(\vec \alpha;\vec \beta;\vec \gamma)}$. For the sake of convenience, and without loss of generality, in the remainder of the paper we will assume that
$$
\alpha_1<\alpha_2<\dots<\alpha_n.
$$
The densities $v_{\alpha_{j}, \beta_{j}}(x, t)$ appearing in \eqref{29/12/2014-8} are unimodal \cite{ELukacs70, WFeller}. The skewness parameter, $\beta_{j}$, measures the degree of asymmetry of $v_{\alpha_{j}, \beta_{j}}(x, t)$: for $\beta_{j} = 0$ they are just the previously mentioned symmetric $\alpha_j$-stable densities with fractional Laplacians as the corresponding infinitesimal generators. Moreover, all of those densities are self-similar, since, for any $x\in\mathbb{R}$, and $ t > 0$,
\begin{equation}\label{31/08/14-1}
v_{\alpha_{j}, \beta_{j},\gamma_j}(t, x) = \frac{1}{t^{1/\alpha_{j}}} v_{\alpha_{j}, \beta_{j},\gamma_j}\left(1, \frac{x}{t^{1/\alpha_{j}}}\right),
\end{equation}
and
\begin{equation}\label{29/12/2014-8a}
v_{\alpha_{j}, \beta_{j}, \gamma_{j}}(-x, t) = v_{\alpha_{j}, -\beta_{j}, \gamma_{j}}(t, x).
\end{equation}
Eq. \eqref{29/12/2014-8a} is a consequence of the identity $\tilde{v}_{\alpha_{j}, \beta_{j}, \gamma_{j}}(t, -\omega) = \tilde{v}_{\alpha_{j}, -\beta_{j}, \gamma_{j}}(t, \omega)$ satisfied by $\tilde{v}_{\alpha_{j}, \beta_{j}, \gamma_{j}}(t, \omega)$ given in Eq. \eqref{29/12/2014-8}.
\section {Asymptotics of solutions of multifractal conservation laws with supercritical nonlinearity}
Now, we are ready to state the results about existence, uniqueness, and the asymptotic behavior of the solutions of the Cauchy problem for the multifractal conservation laws \eqref{29/12/2014-1} driven by multiscale L\'evy processes (anomalous diffusions) introduced in Section 2. The main point here is the observation that the solutions of Eq. \eqref{29/12/2014-1}, under certain conditions on the generator $\cal A$ and the nonlinearity $g$, have the large time behavior similar to solutions of Eq. \eqref{29/12/2014-4}\footnote{This is in contrast to the phenomena observed for data of Riemann type (nonintegrable, and nonsmooth), when shocks are created, see .e.g., \cite{DI2006, D2010,GW2014}.}. These results provide the motivation for the work presented in the following sections. {The physical justification for considering conservation laws driven by
L\'evy processes are more numerous than can be cited here, but see, e.g., \cite{WAW2001} for a review of the subject.}
The solutions of Eq. \eqref{29/12/2014-1} have to be understood in some weak sense which opens several possibilities presented for example in \cite {BFW1999, BKW1999, BKW2001, BKW2001a, KW2008}. Motivated by the classical Duhamel formula we choose to interpret them as the so-called {\it mild} solutions satisfying the identity,
\begin{equation}\label{D}
u(t, x) = \left[e^{ -t {\cal A}}u_0\right]\!(x) - \int_{0}^{t} \frac{\partial}{\partial x} e^{- (t-\tau){\cal A}}[g(u)](\tau, x) \,d\tau.
\end{equation}
The basic results are summarized in the following Theorem, where
the regularity of the solutions of Eq. \eqref{D} is expressed in terms of the Sobolev space $W^{2,2}$.
\begin{thm}\label{t3.1}
(see {\rm\cite{BKW2001}}) (i) Assume that $g\in C^1({\mathbb R}, {\mathbb R}^{d})$ and ${\cal A}$ is the infinitesimal generator of a L\'evy process with the symbol satisfying the condition
\begin{equation}\label{infH}
\limsup_{|\omega|\to\infty}{{\psi(\omega)-\psi_0|\omega|^2} \over{|\omega|^{\widetilde\alpha}}}<\infty\ \ \mbox{for some } 0< \widetilde\alpha<2, {\;and}\; \psi_0>0.
\end{equation}
Given $u_0\in L^1({\mathbb R} )\cap L^\infty({\mathbb R} )$, there exists a unique solution $u\in {\cal C}([0,\infty); \;L^1({\mathbb R} )\cap L^\infty({\mathbb R}))$ of the problem
\begin{equation}\label{L1}
\frac{\partial u}{\partial t} + {\cal A}u + \nabla\cdot g(u) = 0, \qquad u(x,0)=u_0(x).
\end{equation}
This solution is regular, $u\in C((0,\infty)$; $W^{2,2}({\mathbb R}))\cap C^1((0,\infty); L^2({\mathbb R} ))$, satisfies the conservation of integral property, $\int u(x,t)\, dx=\int u_0(x)\, dx$, and the contraction property in the $L_p(\mathbb R)$ space,
\begin{equation}\label{Lp:es}
\|u(t)\|_p\leq \|u_0\|_p,
\end{equation}
for each $p\in [1,\infty]$, and all $t>0$. Moreover, the maximum and minimum principles hold, that is,
\begin{equation}\label{29/12/2014-3.6}
\mbox{\rm ess}\,\inf u_0\le u(x,t) \le \mbox{\rm ess}\, \sup u_0,\ \ \mbox{a.e.}\ x,t,
\end{equation}
and the comparison principle is valid, which means that if $u_0\le v_0\in L^1({\mathbb R} )$, then
\begin{equation}\label{comp}
u(x,t)\le v(x,t)\ \ \mbox{a.e.}\ \ x, \,t, \mbox{ and}\ \ \|u(t)-v(t)\|_1\le\|u_0-v_0\|_1.
\end{equation}
\noindent
(ii) Under the following additional conditions on the symbol of $\cal A$,
\begin{equation}\label{sym}
0<\liminf_{\omega\to 0}{{\psi(\omega)}\over{|\omega|^\alpha}}\le \limsup_{\omega\to 0}{{\psi(\omega)}\over{|\omega|^\alpha}}<\infty, \qquad
0<\inf_\omega{{\psi(\omega)}\over{|\omega|^2}},
\end{equation}
for some $0<\alpha<2$, the more precise bound,
\begin{equation*}
\|u(t)\|_p\le C_p\min(t^{- (1-1/p)/2},t^{- (1-1/p)/\alpha})\|u_0\|_1
\end{equation*}
holds for all $1\le p\le\infty$. Moreover, if $u_0\in L^1({\mathbb R} )\cap L^\infty({\mathbb R} )$, then
\begin{equation}\label{Lp:decay}
\|u(t)\|_p\le C(1+t)^{- (1-1/p)/\alpha}
\end{equation}
with a constant $C$ which depends only on $\|u_0\|_1$ and $\|u_0\|_p$.\\
\noindent
(iii) Assume that $u$ is a solution of the Cauchy problem \eqref{L1} with $u_0\in L^1({\mathbb R} )\cap L^\infty({\mathbb R} )$, and that the symbol $\psi$ of the generator $\cal A$ satisfies Eqs. \eqref{infH} and \eqref{sym} with some $0 < \alpha < 2$. Furthermore, suppose that the nonlinearity $g$ is supercritical, that is, $g\in C^1$, and $\limsup_{s\to 0} |g(s)|/|s|^r$ $<\infty$, for some $r > {\rm{max}}(\alpha ,1)$. Then the relation
\begin{equation}\label{first}
\lim_{t\to \infty}t^{ (1-1/p)/\alpha}\|u(t)-e^{-t{\cal A}}u_0\|_p = 0
\end{equation}
holds for every $1\le p\le\infty$. As usual, $e^{-t{\cal A}}u_{0}$ denotes the action of the L\'evy semigroup on the function $u_0$, i.e. is a solution of the linear {Eq. \eqref{29/12/2014-4}} with the initial data $u_{0}$.
\end{thm}
On the other hand, the asymptotics of the solution of the linear Cauchy problem {Eq. \eqref{29/12/2014-4}} is well known: there exists a nonnegative function $\eta\in L^\infty(0,\infty)$ satisfying
$\lim\limits_{t\to\infty} \eta(t)=0$ such that
\begin{equation}
\Bigl \| e^{ {-}t{\cal A}}*u_0- \int_{{\mathbb R}} u_0(x)\,dx \cdot p_{\cal A}(t)\Bigr\|_p\le t^{-(1-1/p)/\alpha}\eta(t),
\end{equation}
where $p_{\cal A}(t)$ is the kernel of the operator $\cal A$ in Eq. \eqref{29/12/2014-3} . Higher order asymptotics is also available \cite{BKW1999}.
The above general results have direct consequences for {\it multifractal conservation laws} driven by multiscale anomalous diffusions introduced in Section 2\footnote{The particle approximations and the propagation of chaos results for such systems have been studied in \cite{JMW2005}.}. Note the parabolic regularization included in the operator $\cal A$ because of the conditions \eqref{infH}, and \eqref{sym}.
\begin{cor}\label{c2.1}
All the statements of Theorem \ref{t3.1} are valid for the conservation laws
\begin{equation}\label{LL1}
\frac{\partial u}{\partial t} + {\cal A}_ {(\vec \alpha;\vec \beta, \vec \gamma)}u + \frac{\partial}{\partial x} g(u) = 0, \qquad u(x,0) = u_0(x),
\end{equation}
with
\begin{equation*}
\alpha= \alpha_1<\alpha_2<\dots <\alpha_n =2.
\end{equation*}
In particular, if $u$ is a solution of the Cauchy problem \eqref{LL1} with $u_0\in L^1({\mathbb R} )\cap L^\infty({\mathbb R} )$, and the nonlinearity $g \in C^1$ is supercritical, i.e., $\limsup_{s\to 0} |g(s)|/|s|^r$ $<\infty$, for $r > {\rm{max}}(\alpha ,1)$, then the relation
\begin{equation}\label{second}
\lim_{t\to \infty}t^{ (1-1/p)/\alpha}\|u(t)-e^{-t{\cal A}}u_0\|_p = 0
\end{equation}
holds for every $1\le p\le\infty$. Moreover,
\begin{equation*}
\Bigl \| e^{t{\cal A}_ {(\vec \alpha;\vec \beta, \vec \gamma)}}*u_0- \int_{{\mathbb R}} u_0(x)\,dx \cdot p_{{\cal A}_ {(\vec \alpha;\vec \beta, \vec \gamma)}}(t)\Bigr\|_p\le t^{-(1-1/p)/\alpha}\eta(t),
\end{equation*}
where $p_{{\cal A}_ {(\vec \alpha;\vec \beta, \vec \gamma)}}(t)$ is the kernel of the operator ${\cal A}_ {(\vec \alpha;\vec \beta, \vec \gamma)}$ in Eq. \eqref{LL1}.
\end{cor}
To prove Corollary \ref{c2.1} it suffices to show that conditions Eq. \eqref{infH} and Eq. \eqref{sym} are satisfied. Indeed, for the mutiscale L\'evy process with symbol \eqref{multisymbol}, we have
\begin{equation*}
\limsup_{|\omega|\to\infty} {\psi_ {(\vec \alpha;\vec \beta, \vec \gamma)}(\omega)-\gamma_n|\omega|^2 \over {|\omega|^{ \alpha^*} }} = a_{j^*} < \infty
\end{equation*}
with $\alpha_{j^*}=\alpha^*$ where $ \alpha^*=\max (\alpha_1,\dots,\alpha_{(n-1)})$. Also,
\begin{equation*}
0 < \lim_{\omega\to 0}{{ \psi_ {(\vec \alpha;\vec \beta, \vec \gamma)}(\omega)}\over{|\omega|^ {\alpha_*}}} = a_{j_*} <\infty,
\end{equation*}
with $\alpha_{j_{*}} = \alpha_{*}$ where $\alpha_* = \min (\alpha_1,\dots,\alpha_n) = \alpha_1$; and
\begin{equation*}
\inf_\omega{{ \psi_ {(\vec \alpha;\vec \beta, \vec \gamma)}(\omega)}\over{|\omega|^2}}\ge \gamma_n>0.
\end{equation*}
\medskip
The above results depended on the subcritical behavior
$$
\limsup_{s\to 0}\frac{ |g(s)|}{|s|^r} <\infty , \qquad {\rm for} \qquad r > {\rm{max}}(\alpha ,1),
$$
of the nonlinearity in the conservation laws discussed above. Note that in the classical case of the Burgers equation the situation is dramatically different.
\bigskip
{\bf Remark 1. \it Asymptotics of solutions of the Burgers equation.
{\rm The first order asymptotics of solutions of the Cauchy problem for the Burgers equation
\begin{equation}\label{burgers}
\frac{\partial}{\partial t} u(t, x) - \frac{\partial^{2}}{ \partial x^{2}}u(t, x) + \frac{\partial}{\partial x} [u(t, x)]^2 = 0
\end{equation}
is described by the relation
\begin{equation*}
t^{(1-1/p)/2}\|u( t)-U_M( t)\|_p\to 0,\quad {\rm as} \quad t\to\infty,
\end{equation*}
where
\begin{equation*}
U_M(x,t) = \frac{e^{-x^2/(4t)}}{t^{1/2}} \left(K(M) + \frac{1}{2} \int_0^{\ulamek{x}{2} \sqrt t} e^{-\omega^2/4}\, d\omega\right)^{-1}
\end{equation*}
is the so-called source solution with the initial condition $u(x, 0) = M\delta_0$. It is easy to verify that this solution is self-similar, i.e., $U_M(x,t)=t^{-1/2}U(xt^{-1/2},1)$. Thus, the long time behavior of solutions of Eq. \eqref{burgers} is genuinely nonlinear, i.e., it is not determined by the asymptotics of the linear heat equation. This strongly nonlinear behavior is due to the precisely matched balancing influence of the regularizing Laplacian diffusion operator and the gradient-steepening quadratic inertial nonlinearity, see \cite{Z1993, WAW1998, WAW2001}.
\bigskip
Although not needed explicitly in the remainder of the paper, for the sake of completeness we are providing below a general result showing how such a matching critical nonlinearity exponent for the nonlocal multifractal conservation law yields the solutions of (\ref{L1}) which behave asymptotically like the self-similar source solutions $U$ of (\ref{L1}) with singular initial data $M\delta_0$.
\begin{thm}\label{t2}
(see {\rm \cite{BKW2001a}}) Let $1 < \alpha < 2$, and $u$ be a solution of the Cauchy problem (\ref{L1}) with the operator ${\cal A} = (-\Delta)^{\alpha/2} + {\cal K}$, with the perturbation ${\cal K}$ being another L\'evy infinitesimal generator whose symbol $k$ fulfills the condition,
\begin{equation}\label{critical}
\lim_{\omega\to 0}{k(\omega)\over |\omega|^\alpha} = 0,
\end{equation}
and $u_0\in L^1({\mathbb R} )$, $\int_{{\mathbb R}^d} u_0(x)\, dx = M > 0$. Assume that $g$ satisfies the condition
\begin{equation}
\lim_{s \to 0}{g(s)\over s|s|^{(\alpha-1) }}\in {\mathbb R}.
\end{equation}
Then, for each $1\le p\le\infty$,
\begin{equation}
\lim_{t\to\infty}t^{ (1-1/p)/\alpha}\|u(t)-U(t)\|_p= 0,
\end{equation}
where $U=U_M$ is the unique solution of the problem (\ref{L1}) with $r=\alpha$ and the initial data $M\delta_0$. Moreover, $U$ is of self-similar form $U(x,t)=t^{-1/\alpha}U(xt^{-1/\alpha},1)$, $\int_{{\mathbb R}^d}U(x,1)\, dx= M$, and $U\ge 0$.
\end{thm}
Thus, analogous to Corollary \ref{c2.1}, we also have the following result in the case of multifractal conservation laws with critical nonlinearities. Note that, in contrast to Corollary \ref{c2.1}, the parabolic regularization is not necessary here.
\begin{cor}\label{2.2}
All the statements of Theorem \ref{t2} are valid for the multifractal conservation laws
\begin{equation}
\frac{\partial u}{\partial t} + {\cal A}_ {(\vec \alpha;\vec \beta, \vec \gamma)}u + \frac{\partial}{\partial x}g(u) = 0, \qquad u(x,0) = u_0(x), \label{LL}
\end{equation}
with $\alpha=\alpha_*\equiv \min(\alpha_1,\dots,\alpha_k)=\alpha_1$.
\end{cor}
{
The verification of the condition (\ref{critical}) is immediate. With the symbol of the perturbation $\cal K$,
\begin{equation*}
{k}(\omega) =\sum_{j=2}^n - \gamma_j|\omega|^{\alpha_{j}} e^{\frac{i\pi}{2} \beta_{j}\, {\rm sgn}(\omega)}
\end{equation*}
we do have $\lim\limits_{\omega\to 0}{ k(\omega) }/ |\omega|^\alpha=0$. Recall that, in view of the convention adopted at the beginning of the paper, $\alpha_*=\alpha_1<\alpha_2<\dots\alpha_n$.}
\medskip
{The issue of explicit representations of source solutions of fractal conservation laws with critical nonlinearities is obviously more difficult than the problems we are addressing in the subsequent sections, but we plan to investigate it in the future. }
\section{Explicit representation of the kernels of the two-scale, two-sided L\'evy generators, $0 < \alpha \leq 2$}
In this section our goal is to find explicit representations for kernels $ {v}_ {(\vec \alpha;\vec \beta, \vec \gamma)}$ of the infinitesimal generators ${\cal A}_ {(\vec \alpha;\vec \beta, \vec \gamma)}$ which dictate the long-time behavior of the nonlinear conservation laws discussed in Section 3. For the sake of simplicity, we present the case when the scaling parameter $\vec \gamma =(1,\dots,1)$; the notation is then streamlined to $ {v}_ {(\vec \alpha;\vec \beta, \vec \gamma)}\equiv {v}_ {(\vec \alpha;\vec \beta )}$. Simply stated, we need to find an explicit expression for the Fourier convolution of $v_{\alpha_{j}, \beta_{j}}(t, x)$, $j=1, 2$, $x\in\mathbb{R}$, and $t > 0$:
\begin{equation}\label{30Jun14-3}
H(t, x) = \int_{-\infty}^{\infty} v_{\alpha_{1}, \beta_{1}}(t, y) v_{\alpha_{2}, \beta_{2}}(t, x-y) dy = \int_{-\infty}^{\infty} v_{\alpha_{1}, \beta_{1}}(t, x-y) v_{\alpha_{2}, \beta_{2}}(t, y) dy,
\end{equation}
where $H(t, x) = H(\vec \alpha;\vec \beta; t, x)$, $\vec \alpha=(\alpha_1,\alpha_2)$, and $\vec \beta=(\beta_1,\beta_2)$. The basic properties of $v_{\alpha_{j}, \beta_{j}}(t, x)$, with necessary conditions on $\alpha_{j}$ and $\beta_{j}$, are given in Section 2.
The functions $v_{\alpha_{j}, \beta_{j}}(t, x)$, $j=1, 2$, represent the unimodal probability density functions of two-sided L\'{e}vy stable distributions \cite{ELukacs70, WFeller}, which correspond to one-sided L\'evy stable distributions for $0 < \alpha_{j} < 1$ and $\beta_{j} = -\alpha_{j}$. This case will be discussed in section 5. The series representation of two-sided L\'{e}vy stable distributions for $0 < \alpha_{j} < 1$, and $|\beta_{j}| \leq \alpha_{j}$, can be found in, e.g., Eq. (5.8.8a) on p. 142 \cite{ELukacs70}, Eq. (6.8) on p. 583 of \cite{WFeller}, and Eq. (4) in \cite{HBerstrom52}, whereas, for $ 1 < \alpha_{j} \leq 2$, and $|\beta_{j}| \leq 2-\alpha_{j}$, they are described in, e.g., Eq. (5.8.8b) on p. 142 of \cite{ELukacs70}, and Eq. (6.9) on p. 583 of \cite{WFeller}. Those two different types of series expansions were calculated for rational values of parameter $\alpha_{j}$ and $\beta_{j}$, see Eqs. (4) and (5) in \cite{KGorska11}. We quote some solution which will be used later in the paper: the Gaussian distribution
\begin{equation}\label{29/12/2014-a}
v_{2, 0}(t, x) = \frac{\exp(-\ulamek{x^2}{4t})}{2\sqrt{\pi t}}
\end{equation}
for $\alpha = 2$ and $\beta=0$, the L\'{e}vy-Smirnov distribution
\begin{align}\label{26/09/2014-1a}
v_{\frac{1}{2}, -\frac{1}{2}}(t, x) & = \frac{t \exp(-\ulamek{t^2}{4x})}{2\sqrt{\pi} x^{3/2}} , \qquad x>0 \\ & = 0, \qquad\qquad\qquad\,\, x \leq 0.
\end{align}
for $\alpha = 1/2$ and $\beta = -1/2$, and
\begin{align}\label{29/12/2014-b}
v_{\frac{3}{2}, -\frac{1}{2}}(t, x) & = \frac{(2/t)^{2/3}}{3\sqrt{\pi}} \frac{\Gamma(\ulamek{5}{6})}{\Gamma(\ulamek{2}{3})}\, {_{1}F_{1}}\left({5/6 \atop 2/3}; -\frac{4 x^3}{27 t^2}\right) + \frac{(2/t)^{4/3}}{9\sqrt{\pi}} \frac{\Gamma(\ulamek{7}{6})}{\Gamma(\ulamek{4}{3})}\, x\, {_{1}F_{1}}\left({7/6 \atop 4/3}; -\frac{4 x^3}{27 t^2}\right) \nonumber\\
& = \frac{Re}{2\pi} \int_{-\infty}^{\infty} e^{-i \omega x} e^{-t|\omega|^{3/2}\exp(-\frac{i\pi}{4} {\rm sgn}(\omega)} d\omega
\end{align}
for $\alpha = 3/2$ and $\beta = -1/2$ \cite{KGorska11}. The symbol ${_{1}F_{1}}$ stands for the hypergeometric function introduced in Section 1.
Let us now find the explicit form of $H(t, x)$ given in Eq. \eqref{30Jun14-3}. Applying the property \eqref{29/12/2014-8a} to Eq. \eqref{30Jun14-3}, we can rewrite $H$ in the form,
\begin{align}\label{30Jun14-4}
H(t, x) = \mathcal{F}^{-1}[\tilde{v}_{\alpha_{1}, \beta_{1}}(t, \omega)\, \tilde{v}_{\alpha_{2}, \beta_{2}}(t, \omega); x] = H_{-}(t, -x) \Theta(-x) + H_{+}(t, x)\Theta(x),
\end{align}
where
\begin{equation}\label{30/06/14-5}
H_{+}(t, x) = H_{+}({\alpha_{1}, \beta_{1}, \alpha_{2}, \beta_{2}}; t, x) = \frac{\textit{Re}}{\pi} \int_{0}^{\infty} \hskip -2mm e^{-i x \omega} \exp\left(\!- t \omega^{\alpha_{1}} e^{\frac{i\pi}{2}\beta_{1}} - t \omega^{\alpha_{2}} e^{\frac{i\pi}{2}\beta_{2}}\right) d\omega,
\end{equation}
and
\begin{equation}\label{30/06/14-5a}
H_{-}(t, x) = H_{+}(\alpha_{1}, -\beta_{1}, \alpha_{2}, -\beta_{2}; t, x).
\end{equation}
The function $\Theta(x)$ is here the usual Heaviside step function. The matching, at $x=0$, of these two components is assured by the continuity at the origin of $H(t, x)$, and of all of its higher derivatives. Indeed, the continuity of Eq. \eqref{30Jun14-4} at $x=0$ can be shown by employing Eq. \eqref{30/06/14-5} as follows: for $n=0, 1, 2, \ldots$
\begin{align}\label{30/07/14-3}
\partial^{n}_{x} H_{+}(t, x)\big\vert_{x=0} & = \frac{\textit{Re}}{\pi} \int_{0}^{\infty} (-i \omega)^{n} e^{- t \omega^{\alpha_{1}}\exp\big(\ulamek{i\pi}{2} \beta_{1}\big) - t \omega^{\alpha_{2}}\exp\big(\ulamek{i\pi}{2} \beta_{2}\big)} d\omega \nonumber \\
& = \frac{1}{\pi} \int_{0}^{\infty} e^{- t \omega^{\alpha_{1}} \cos\!\big[\ulamek{\pi}{2}(\beta_{1} - n)\big] - t \omega^{\alpha_{2}} \cos\!\big[\ulamek{\pi}{2}(\beta_{2} - n)\big]} \nonumber \\
& \times \cos\left(t \omega^{\alpha_{1}} \sin\big[\ulamek{\pi}{2}(\beta_{1} - n)] + t \omega^{\alpha_{2}} \sin\big[\ulamek{\pi}{2}(\beta_{2} - n)]\right) d\omega \nonumber \\
& = \frac{\textit{Re}}{\pi} \int_{0}^{\infty} (i \omega)^{n} e^{- t \omega^{\alpha_{1}}\exp\big(-\ulamek{i\pi}{2} \beta_{1}\big) - t \omega^{\alpha_{2}}\exp\big(-\ulamek{i\pi}{2} \beta_{2}\big)} d\omega \nonumber \\
& = \partial^{n}_{x} H_{-}(t, -x)\big\vert_{x=0}.
\end{align}
We would also like to point out that Eqs. \eqref{30Jun14-4} and \eqref{30/06/14-5}, in the case $\alpha_{1} = \alpha_{2}$, and $\beta_{1} = \beta_{2}$, imply the identity,
\begin{equation}
\label{30/07/14-5}
H(x, t) = v_{\alpha_{1}, \beta_{1}}(x, 2t).
\end{equation}
In what follows, without loss of generality, we will consider only the case of $H_{+}(x, t)$. The function $H_{-}(x, t)$ will be used only when necessary. We assume that, for certain values of complex $s$, the Mellin transform of $H_{+}(x, t)$ exists and, according to the notation introduced in Eq. \eqref{26/08/14-3a}, it is denoted by $\hat{H}_{+}(s, t)$. Thereafter, we substitute Eq. \eqref{30/06/14-5} into $\hat{H}_{+}(s, t)$, change the order of integration and use the one of Eqs. (2.3.2.13) of \cite{APPrudnikov-v1}. Those steps imply that for rational $\alpha_{j}$ and $\beta_{j}$, $j=1, 2$, such that $\alpha_{1} = \ulamek{l}{k}$, $\beta_{1} = \ulamek{l-2a}{k}$, $\alpha_{2} = \ulamek{p}{q}$, and $\beta_{2} = \ulamek{p-2b}{q}$, where $l$, $k$, $p$, $q$, $a$, and $b$, are integers, we have
\begin{align}
\label{30/06/14-6a}
\hat{H}_{+}(s, t) & = \frac{1}{M \pi}\sum_{j=0}^{M_{1}-1} \frac{(-1)^{j}}{j!}\, \frac{\Gamma(s)\, \Gamma\big(\frac{1-s}{M} + \frac{m}{M}j\big)}{t^{\frac{1-s}{M} + (\frac{m}{M}-1)j}}\, {\textit Re}\left\{e^{-i\pi\big[\frac{1}{2} - u\frac{1-s}{M} + \big(v-u\frac{m}{M}\big)j\big]} \right.
\nonumber\\
&\qquad \left.\times {_{1+m_{1}}F_{M_{1}}}\left({1, \Delta(m_{1}, \frac{1-s}{M} + \frac{m}{M}j) \atop \Delta(M_{1}, 1+j)}; \big(\!-\ulamek{t e^{-i\pi v}}{M_{1}}\big)^{M_{1}} \big(\ulamek{m_{1}}{t e^{-i\pi u}}\big)^{m_{1}}\right)\right\}
\\
& = \frac{1}{M \pi} \sum_{j=0}^{M_{1}-1}\sum_{r=0}^{\infty} \frac{(-1)^{j+r M_{1}}}{(j + r M_{1})!}\, \frac{\Gamma(s)\, \Gamma[\frac{1-s}{M} + \frac{m}{M}(j+r M_{1})]}{t^{\frac{1-s}{M} + (\frac{m}{M}-1)(j+r M_{1})}} \nonumber
\\
&\qquad \times \sin\left[\pi u\ulamek{1-s}{M} - \pi (j+rM_{1})\big(v - u\ulamek{m}{M}\big) \right]
\label{30/06/14-6b}
\\
& = \frac{1}{M \pi} \sum_{r=0}^{\infty} \frac{(-1)^{r}}{r!} \, \frac{t^{-\frac{1-s}{M} - r(\frac{m}{M}-1)}\, \Gamma(s)\, \Gamma(\ulamek{1-s}{M} + \ulamek{m}{M}r)}{\Gamma\big[1- u \ulamek{1-s}{M} + (v-u\ulamek{m}{M})r\big]\, \Gamma\big[u \ulamek{1-s}{M} - (v-u\ulamek{m}{M})r\big]},
\label{30/06/14-6c}
\end{align}
where $m$, $M$, $m_{1}$, and $M_{1}$ are as follows:
\begin{equation}
\label{26/08/14-10}
m = \min(\ulamek{l}{k}, \ulamek{p}{q}), \quad M = \max(\ulamek{l}{k}, \ulamek{p}{q}), \quad m_{1} = \min(kp, lq), \quad \text{and} \quad M_{1} = \max(kp, lq).
\end{equation}
The parameters $u$, and $v$, are determined by the equalities,
\begin{equation}
\label{31/08/14-1}
u = \frac{a}{k}, \,\, v = \frac{b}{q}, \,\,\, {\rm for} \,\,\, \alpha_{1} > \alpha_{2}, \quad {\rm and} \quad u = \frac{b}{q}, \,\, v = \frac{a}{k} \,\,\, {\rm for} \,\,\, \alpha_{1} < \alpha_{2}.
\end{equation}
In Eq. \eqref{30/06/14-6a} we utilized a series representation of the generalized hypergeometric function given in Eq. (\ref{26/08/14-6}), and the Gauss-Legendre multiplication formula defined in Eq. (\ref{26/08/14-8}). In Eq. \eqref{30/06/14-6b} we applied Eq. \eqref{26/08/14-7}, and also changed the summation index as follows: $j+r M_{1} \to r$.
The next step requires inverting the Mellin transform in Eq. \eqref{30/06/14-6c}. To accomplish this task we will introduce the new variable of integration, $\tilde{s} = (1-s)/(lp)$, Eqs. \eqref{26/08/14-7}, and \eqref{26/08/14-8}. Putting the all of these terms together, we get, for $x > 0$,
\begin{align}
\label{30/06/14-7}
H_{+}(x, t) & = \frac{m_{1}\sqrt{M}}{x (2\pi)^{\frac{lp+m_{1}}{2} - u m_{1}}} \sum_{r=0}^{\infty} \frac{(-t)^{r}}{r!} \left(\frac{m_{1}}{t}\right)^{\!\frac{m}{M}r}\nonumber
\\&\qquad \times G^{m_{1}, lp}_{lp + u m_{1}, m_{1} + u m_{1}}\left(\frac{(lp)^{lp}\, t^{m_{1}}}{x^{lp}\, m_{1}^{m_{1}}} \Big\vert {\Delta(lp, 0), \Delta(u m_{1}, -(v-u\frac{m}{M} r)) \atop \Delta\big(m_{1}, \frac{m}{M}r\big), \Delta(u m_{1}, -(v-u\frac{m}{M} r))}\!\right).
\end{align}
The Meijer $G$ functions in Eq. \eqref{30/06/14-7} can be expressed, via formulas (8.2.2.3), and (8.2.2.4), on p. 520 of \cite{APPrudnikov-v3}, in terms of a generalized hypergeometric function. With respect to the values of $\alpha_{1}$ and $\alpha_{2}$, we can consider two different cases:
\begin{itemize}
\item[\textbf{(A)}]
After applying Eq. (8.2.2.3) on p. 520 of \cite{APPrudnikov-v3} to Eq.~\eqref{30/06/14-7}, for $0 < \alpha_{i} < 1$, $i = 1, 2$, we have
\begin{align}
\label{30/07/14-1}
H_{+}(x, t) &= -\frac{1}{\pi} \sum_{r=0}^{\infty} \sum_{j=0}^{m_{1}-1} \frac{(-t)^{r+j}}{r!\, j!} \frac{\Gamma(1 + Mj + mr)}{x^{1 + Mj + rm}} \sin(r\pi v + j\pi u) \nonumber
\\
& \quad \times {_{1+lp} F_{m_{1}}}\left({1, \Delta(lp, 1+ Mj + mr) \atop \Delta(m_{1}, 1 + j)}; (-1)^{m_{1}u - m_{1}} \frac{t^{m_{1}} (lp)^{lp}}{m_{1}^{m_{1}} x^{lp}}\right),
\end{align}
where $u$, and $v$, are given in Eq. \eqref{31/08/14-1}\footnote{Let us observe that, for $u = M$, and $v = m$, and thus, for $a = l$, and $b = p$, Eq. \eqref{30/07/14-1} gives Eq. \eqref{14/07/14-2} of Section 5, this is the Laplace convolution of two one-sided L\'{e}vy stable distributions.}. Moreover, using the series expansion of the function ${_{1+lp} F_{m_{1}}}$, Eq. \eqref{30/07/14-1}, can be expressed as follows:
\begin{equation}
\label{30/07/14-2}
H_{+}(x, t) = -\frac{1}{\pi} \sum_{r, n = 0}^{\infty} \frac{(-t)^{n + r}}{n!\, r!}\, \frac{\Gamma(1 + \alpha_{1} r + \alpha_{2} n)}{x^{1 + \alpha_{1} r + \alpha_{2} n}}\, \sin\big(\pi r\, \ulamek{\alpha_{1} - \beta_{1}}{2} + \pi n\, \ulamek{\alpha_{2} - \beta_{2}}{2}\big),
\end{equation}
which for $t=1$, $x > 0$, and $r=0$ (or $n=0$), is identical with the series expression for two-sided L\'{e}vy stable distribution given in, e.g. Eq. (5.8.8a) on p. 142, in \cite{ELukacs70}.
\item[\textbf{(B)}]
For $ 1 < \alpha_{i} \leq 2$, $i = 1, 2$, Eq. (8.2.2.4) on p. 520 of \cite{APPrudnikov-v3}, applied to Eq. \eqref{30/06/14-7} gives
\begin{align}
\label{30/07/14-4}
H_{+}(x, t) &= \frac{1}{\pi M} \sum_{r=0}^{\infty} \sum_{j=0}^{lp -1} \frac{(-1)^{r+j}}{r!\, j!} \frac{x^{j}}{t^{\frac{1+j}{M} + (\frac{m}{M} - 1)r}}\, \Gamma\big(\ulamek{1+j}{M} + \ulamek{m}{M}r\big) \sin\big[\pi u\ulamek{1+j}{M} - \pi r \big(v - u \ulamek{m}{M}\big)\big] \nonumber \\[0.2\baselineskip]
& \quad \times {_{1+m_{1}}F_{lp}}\left({1, \Delta(m_{1}, \ulamek{1+j}{M} + \ulamek{m}{M} r) \atop \Delta(lp, 1+j)}; (-1)^{m_{1}u + lp} \frac{m_{1}^{m_{1}} x^{lp}}{t^{m_{1}} (lp)^{lp}}\right),
\end{align}
which can be rewritten as
\begin{equation}
\label{30/07/14-5}
H_{+}(x, t) = \frac{1}{\pi M} \sum_{r, n = 0}^{\infty} \frac{(-1)^{r+n}}{r!\, n!} \frac{x^{n}}{t^{\frac{1+n}{M} + (\frac{m}{M} - 1)r}}\, \Gamma\big(\ulamek{1+n}{M} + \ulamek{m}{M}r\big)\,\sin\big[\pi u\ulamek{1+n}{M} - \pi r \big(v - u \ulamek{m}{M}\big)\big].
\end{equation}
\end{itemize}
\noindent
{\it Example 4.1. The bi-Gaussian case.}
The elementary case $\alpha_{1} = \alpha_{2} = 2$, $\beta_{1} = \beta_{2} = 0$, is straightforward and we include it here only for verification's sake. Substituting Eq. \eqref{29/12/2014-a} into Eq. \eqref{30Jun14-3} and employing Eq. (3.323.2) on p. 337 of \cite{Gradshteyn}, we get
\begin{align} \label{30/12/2014-aa}
H(2, 0, 2, 0; t, x) = \frac{\exp(-\frac{x^{2}}{4t})}{4\pi t} \int_{-\infty}^{\infty} e^{-\ulamek{y^{2}}{2 t} + \ulamek{x y}{2t}} dy = \frac{\exp(-\ulamek{x^{2}}{8 t})}{2 \sqrt{2\pi t}} = v_{2, 0}(x, 2t),
\end{align}
which is in agreement with Eq. \eqref{30/07/14-5} and is presented in Fig. \ref{fig2}, see the curve I (red).\\
\noindent
{\it Example 4.2. The Gaussian-L\'{e}vy case.} Here $\alpha_{1} = 2$, $\alpha_{2} = 1/2$, $\beta_{1} = 0$, and $\beta_{2} = -1/2$, so that Eq. \eqref{30Jun14-3} reads
\begin{align}
H(2, 0, \ulamek{1}{2}, -\ulamek{1}{2}; t, x) & = \frac{1}{4 \pi \sqrt{t}} e^{-\frac{x^2}{4 t}} \int_{0}^{\infty} y^{-\frac{3}{2}} e^{-\frac{t^2}{4 y}} e^{-\frac{y^2}{4 t} + \frac{x y}{2 t}} dy \nonumber \\
& = \frac{\sqrt{t}}{4\pi} e^{-\frac{x^2}{4 t}} \sum_{r=0}^{\infty} \frac{(-t^2/4)^r}{r!} \int_{0}^{\infty} y^{-\frac{3}{2}-r} e^{-\frac{y^2}{4 t} + \frac{x y}{2 t}} dy \label{10/12/2014a} \\
& = \frac{1}{2} e^{-\frac{x^2}{8 t}} \sum_{r=0}^{\infty} \frac{(-1)^{2r+1}}{r! \Gamma(\frac{3}{2} + r)}\, \frac{t^{\frac{1}{4} + \frac{3}{4}r}}{2^{\frac{5}{4} + \frac{5}{2}r}}\, D_{\frac{1}{2} + r}(-\ulamek{x}{\sqrt{2 t}}) \label{10/12/2014b} \\
& = \frac{1}{2\sqrt{\pi t}} e^{-\frac{x^2}{8 t}} \sum_{r=0}^{\infty} \frac{(-1)^{2r+1}}{(2r+1)!} \left(\!\frac{t^3}{2}\!\right)^{\!\frac{1}{4}(2r + 1)}\!\! D_{\frac{2r+1}{2}}(-\ulamek{x}{\sqrt{2 t}}) \label{10/12/2014c} \\
& = \frac{1}{2\sqrt{\pi t}} e^{-\frac{x^2}{8 t}} \sum_{n=0}^{\infty} \frac{(-1)^{n}}{n!} \left(\!\frac{t^3}{2}\!\right)^{\!\frac{n}{4}}\! D_{\frac{n}{2}}(-\ulamek{x}{\sqrt{2 t}}), \label{10/12/2014d}
\end{align}
where $D_{\nu}(z)$ is the parabolic cylinder function \cite{Gradshteyn}. In Eq. \eqref{10/12/2014a} we applied Eq. (9.241.2) on p. 1028 of \cite{Gradshteyn}, and in Eq. \eqref{10/12/2014c} we changed the summation index as follows: $n = 2r + 1$. Eq. \eqref{10/12/2014d} can be obtained from Eq. \eqref{30/07/14-4} after using Eqs. (7.11.3.3) and (7.11.3.4) on p. 491 of \cite{APPrudnikov-v3}. The plot of $H(2, 0, \ulamek{1}{2}, -\ulamek{1}{2}; t, x)$ for $t=1$ is illustrated in Fig. \ref{fig2}, see the curve II (blue).
\\
\noindent
{\it Example 4.3.} In Eq. \eqref{30Jun14-3} we take $v_{\alpha_{j}, \beta_{j}}(t, x)$, $j=1, 2$, given in Eqs. \eqref{29/12/2014-a} and \eqref{29/12/2014-b}. Thus, we get
\begin{align}\label{30/12/2014-1a}
& H(2, 0, \ulamek{3}{2}, -\ulamek{1}{2}; t, x) = \int_{-\infty}^{\infty} \frac{\exp(-\frac{y^2}{4 t})}{2\sqrt{\pi t}} \frac{Re}{2\pi} \left\{ \int_{-\infty}^{\infty} e^{-i \omega (x-y)} e^{-t|\omega|^{3/2}\exp[-\frac{i\pi}{4} {\rm sgn}(\omega)]} d\omega \right\} dy \\ \label{30/12/2014-1b}
& \qquad = \frac{Re}{2\pi} \int_{-\infty}^{\infty} e^{-i \omega x} e^{-t|\omega|^{3/2}\exp[-\frac{i\pi}{4} {\rm sgn}(\omega)]} \left[\int_{-\infty}^{\infty} e^{i\omega y}\frac{\exp(-\frac{y^2}{4 t})}{2\sqrt{\pi t}} dy\right] d\omega \\ \label{30/12/2014-1c}
& \qquad = \frac{Re}{2\pi} \int_{-\infty}^{\infty} e^{-i \omega x} e^{-t|\omega|^{3/2}\exp[-\frac{i\pi}{4} {\rm sgn}(\omega)] - t \omega^2} d\omega,
\end{align}
which is in agreement with Eq. \eqref{30Jun14-4} for $\alpha_{1} = 2$, $\alpha_{2} = 3/2$, $\beta_{1} = 0$, and $\beta_{2} = -1/2$. Thus we can repeat all the steps from Eq. \eqref{30Jun14-4} to Eq. \eqref{30/07/14-4} which gives
\begin{align}\label{30/12/2014-2a}
H(2, 0, \ulamek{3}{2}, -\ulamek{1}{2}; t, x) & = \frac{1}{2\pi} \sum_{r=0}^{\infty} \sum_{j=0}^{5} \frac{(-1)^{r+j}}{r! j!} \frac{x^j}{t^{\frac{1+j}{2} - \frac{r}{4}}} \Gamma(\ulamek{1+j}{2} + \ulamek{3}{4} r) \cos(\ulamek{\pi}{2}j - \ulamek{\pi}{4}r) \nonumber \\
& \times {_{4}F_{6}}\left({1, \Delta(3, \ulamek{1+j}{2} + \ulamek{3}{4}r) \atop \Delta(6, 1+j)}; -\frac{x^6}{1728 t^3}\right).
\end{align}
For $t=1$, the function is plotted as the curve III (green) in Fig. \ref{fig2}.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.6]{fig2.eps}
\caption{\label{fig2} Multiscale densities $H(\alpha_{1}, \beta_{1}, \alpha_{2}, \beta_{2}; t, x)$, for $t = 1$, and given values of $\alpha_{1}$, $\beta_{1}$, $\alpha_{2}$, and $\beta_{2}$. In plot I (red), $\alpha_{1} = \alpha_{2} = 2$, and $\beta_{1} = \beta_{2} = 0$, see Eq. \eqref{30/12/2014-aa}; in plot II (blue), $\alpha_{1} = 2$, $\beta_{1} = 0$, $\alpha_{2} = \ulamek{1}{2}$, and $\beta_{2} = -\ulamek{1}{2}$, see Eq. \eqref{10/12/2014d}; in plot III (green), $\alpha_{1} = 2$, $\beta_{1} = 0$, $\alpha_{2} = \ulamek{3}{2}$, and $\beta_{2} = -\ulamek{1}{2}$, see Eq. \eqref{30/12/2014-2a}.
}
\end{center}
\end{figure}
\section{Explicit representation of the kernels of the two-scale, one-sided L\'evy generators, $0<\alpha < 1$}
Although this case has no direct applicability to our work on the asymptotics of the multifractal conservation laws in Section 3, in anticipation of our future work, in this section we provide explicit representations of the kernels of the two-scale, one-sided L\'evy generators with $0<\alpha_1<\alpha_2\le 1$. Here the tool is, of course, the Laplace transform.
To further simplify our exposition we will only consider two one-parameter families, $v_{\alpha_1}(t, x) \equiv v_{(\alpha_1, -\alpha_{1}, 1)}(t, x)$, and $v_{\alpha_2}(t, x) \equiv v_{(\alpha_2, -\alpha_{2}, 1)}(t, x)$, of one-sided stable L\'evy densities whose the Laplace convolution has the form
\begin{equation}\label{25.06.14-1}
h_{\alpha_{1}, \alpha_{2}}(t, x) = \int_{0}^{x} v_{\alpha_{1}}(t, y) v_{\alpha_{2}}(t, x-y) dy = \int_{0}^{x} v_{\alpha_{1}}(t, x-y) v_{\alpha_{2}}(t, y) dy.
\end{equation}
Functions $v_{\alpha_j}(t, x)$, $j =1, 2$ are given by the ``stretched exponential'' Laplace transform $\exp(- t p^{\alpha_{j}}) = \mathcal{L}[v_{\alpha_{j}}(t, x); p]$, see \cite{RSAnderssen04, GDattoli14, HPollard46, KAPenson10}. Thus, the characteristic function of Eq. \eqref{25.06.14-1} is of the form
\begin{equation}
\label{25.06.14-2}
\exp\left(-t p^{\alpha_{1}} - t p^{\alpha_{2}}\right) = \mathcal{L}[h_{\alpha_{1}, \alpha_{2}}(x, t); p].
\end{equation}
Eq. \eqref{25.06.14-2} implies that, for $0 < \alpha_{1} = \alpha_{2} < 1$,
\begin{equation}
\label{26/08/14-9}
h_{\alpha_{1}, \alpha_{1}}(t, x) = v_{\alpha_{1}}(2t, x),
\end{equation}
whereas $h_{\alpha_{1}, \alpha_{2}}(t, x)$ for arbitrary $\alpha_{j}$ can be found by using the series representation of a one sided-L\'{e}vy stable distribution,
\begin{equation}
\label{31/07/14-1}
v_{\alpha_{j}}(t, x) = -\frac{1}{\pi} \sum_{r=0}^{\infty} \frac{(-t)^{r}}{r!} \frac{\Gamma(1 + \alpha_{j} r)}{x^{1 + \alpha_{j} r}} \sin(\pi r \alpha_{j}).
\end{equation}
Eq. \eqref{31/07/14-1} follows from Eq. \eqref{31/08/14-1} applied for the one-sided L\'{e}vy stable distribution, and the formula (4) of \cite{HPollard46}.
\bigskip
{\bf Remark 2. \it
Absolute convergence of $v_{\alpha_{j}}(t, x)$.} {\rm For $x \geq 0$, the series in Eq. \eqref{31/07/14-1} converges absolutely and its radius of convergence is infinite. Indeed, the absolute value of $v_{\alpha_{j}}(x, t)$, for $x > 0$, can be estimated as follows:
\begin{equation*}
|v_{\alpha_{j}}(t, x,)| < \frac{1}{x} \sum_{n=0}^\infty \frac{\Gamma(1+\alpha_{j} n)}{n!}\! \left(\!\frac{t}{x^{\alpha_{j}}}\!\right)^{\!n} < \infty.
\end{equation*}
Next, from the Cauchy ratio test of convergence, \cite{GBArfken05}, and the Stirling formula, see Eq. (8.327.1) on p. 895 of \cite{Gradshteyn}, the absolute convergence of the series \eqref{31/07/14-1} follows. The convergence of the series \eqref{31/07/14-1} at $x=0$ can be verified by employing the asymptotic form of $v_{\alpha_{j}}(x, t)$ at $x=0$:
\begin{equation}\label{28/08/14-2}
v_{\alpha_{j}}(t, x) \approx K t^{\ulamek{1}{2(1-\alpha_{j})}} x^{-\ulamek{2 - \alpha_{j}}{2 - 2\alpha_{j}}} \exp\!\Big(\!-A t^{\ulamek{1}{1 - \alpha_{j}}} x^{- \ulamek{\alpha_{j}}{1-\alpha_{j}}}\Big),
\end{equation}
where $K$ and $A$ are positive constants. Eq. \eqref{28/08/14-2} was obtained with the help of Eq. (4) in \cite{JMikusinski59}. The absolute value of the right-hand side of of Eq. \eqref{28/08/14-2} obviously converges to 0, as $x\to 0+$ (for any fixed $t > 0$, and $0 < \alpha_{j} < 1$).
\bigskip
For rational $\alpha_{j} = l/k$, where $k$ and $l$ are positive integers, Eq. \eqref{31/07/14-1} can be expressed via a finite sum of the generalized hypergeometric function, see Eqs. (3), and (4), in \cite{KAPenson10}, for $t=1$. Moreover, it turns out that, for $k \leq 3$, it can be written down in terms of standard special functions, e.g., for $\alpha = 1/2$,
\begin{equation}
\label{26/09/2014-1a}
v_{\frac{1}{2}}(t, x) = v_{\frac{1}{2}, -\frac{1}{2}}(t, x) \quad {\rm for}\,\, x > 0,
\end{equation}
for $\alpha = 1/3$ \cite{HSher75}
\begin{equation}\label{26/09/2014-1b}
v_{\frac{1}{3}}(t, x) = \frac{t^{3/2}}{3\pi x^{3/2}} K_{\frac{1}{3}}\!\left(\!\frac{2}{\sqrt{x}}\frac{t^{3/2}}{3^{3/2}}\right),
\end{equation}
and for $\alpha = 2/3$ \cite{EWMontroll84}
\begin{equation}\label{26/09/2014-1c}
v_{\frac{2}{3}}(x, t) = \sqrt{\frac{3}{\pi}}\, \frac{\exp\big(\!-\ulamek{2 t^3}{27 x^2}\big)}{x} W_{\frac{1}{2}, \frac{1}{6}}\left(\frac{4 t^3}{27 x^2}\right),
\end{equation}
where $K_{\nu}(u)$ is a modified Bessel function of the second kind \cite{Gradshteyn}, and $W_{\nu, \mu}(u)$ is the Whittaker W function \cite{Gradshteyn}.
The substitution of Eq. \eqref{31/07/14-1} into Eq. \eqref{25.06.14-1} allows us to write, with help from Eq. \eqref{26/08/14-7}, that
\begin{align}
\label{8/07/14-1}
h_{\alpha_{1}, \alpha_{1}}(t, x) &= \int_{0}^{x} \sum_{r, j = 0}^{\infty} \frac{(-t)^{r+j}}{r! j!} \frac{\Gamma(1+\alpha_{1}j) \Gamma(1+\alpha_{2}r)}{y^{1 + j\alpha_{1}} (x-y)^{1 + r \alpha_{2}}} \sin(\pi\alpha_{1} j) \sin(\pi\alpha_{2} r) \frac{dy}{\pi^{2}} \nonumber \\
& = \sum_{r, j = 0}^{\infty} \frac{(-t)^{r+j}}{r! j!} \frac{1}{\Gamma(-\alpha_{1} j) \Gamma(-\alpha_{2} r)} \int_{0}^{x} y^{-1-j\alpha_{1}} (x-y)^{-1-r \alpha_{2}} dy \nonumber \\
& = -\frac{1}{\pi} \sum_{r, j = 0}^{\infty} \frac{(-t)^{r+j}}{r! j!}\, \frac{\Gamma(1+ \alpha_{1} j + \alpha_{2} r)}{x^{1+ \alpha_{1} j + \alpha_{2} r}}\, \sin(\pi\alpha_{1} j + \pi\alpha_{2} r).
\end{align}
Observe that the expressions in Eq. \eqref{8/07/14-1} are invariant with respect to the change of order of the parameters $\alpha_{1}$, and $\alpha_{2}$, so that $h_{\alpha_{1}, \alpha_{2}}(x, t) = h_{\alpha_{2}, \alpha_{1}}(x, t)$. Moreover, the first term in the sum in Eq. \eqref{8/07/14-1}, corresponding to $j=r=0$, vanishes. The terms with indices $j=0$, $r=1$, and $j=1$, $r=0$, provide the 'heavy-tailed' asymptotic behavior of the one-sided L\'{e}vy stable distributions. Consequently, we can get see immediately that the asymptotic behavior of $h_{\alpha_{1}, \alpha_{2}}(t, x)$ is proportional to $x^{-1-\min(\alpha_{1}, \alpha_{2})}$.
The first series in Eq. \eqref{8/07/14-1} can be expressed as a finite sum of the generalized hypergeometric functions given in Eq. \eqref{26/08/14-6}. Indeed, let us consider, without loss of generality, the case of rational $\alpha_{1} > \alpha_{2}$, with $\alpha_{1} = \ulamek{l}{k}$ and $\alpha_{2} = \ulamek{p}{q}$, where $l$, $k$, $p$, and $q$ are integers. In this case, Eq. \eqref{8/07/14-1} takes the form,
\begin{align}
\label{14/07/14-1}
h_{\frac{l}{k}, \frac{p}{q}}(t, x) & = \frac{1}{\pi} \sum_{r, j = 0}^{\infty} \frac{(-t)^{r+j+1}}{r! j!} \frac{\Gamma(1 + \ulamek{l}{k} j + \ulamek{p}{q} r)}{x^{1 + \ulamek{l}{k} j + \ulamek{p}{q} r}} \sin\big[\pi\big(\ulamek{l}{k} j + \ulamek{p}{q} r\big)\big] \nonumber \\
& = \frac{1}{\pi} \sum_{r=0}^{\infty} \sum_{n=0}^{kp-1}\sum_{j=0}^{\infty} \frac{(-t)^{r+1+(n+kpj)}}{r! (n+kpj)!} \frac{\Gamma[1 + \ulamek{l}{k} (n + kpj) + \ulamek{p}{q} r]}{x^{1 + \ulamek{l}{k} (n + kpj) + \ulamek{p}{q} r}} \sin\big[\pi\big(\ulamek{l}{k} n + \ulamek{p}{q} r\big) + \pi l p j\big] \nonumber \\
& = \sum_{r=0}^{\infty} \sum_{n=0}^{kp-1} \frac{(-t)^{n+r}}{n!\, r!} \frac{x^{-1 -\ulamek{l}{k} n - \ulamek{p}{q} r}}{\Gamma\big(\!-\ulamek{l}{k} n - \ulamek{p}{q} r\big)} \sum_{j=0}^{\infty} \frac{(-1)^{kpj-lnj}}{j!} \left(\!\frac{t}{kp}\!\right)^{jkp} \left(\!\frac{lp}{x}\!\right)^{lp j} \nonumber \\
& \qquad \qquad \qquad \times \frac{(1)_{j} \prod_{s=0}^{lp-1} \big(\ulamek{1+s}{lp} + \ulamek{n}{kp} + \ulamek{r}{lq}\big)_{j}}{\prod_{s=0}^{kp-1} \big(\ulamek{1+j+s}{kp}\big)_{j}}.
\end{align}
To obtain Eq. \eqref{14/07/14-1} we applied the Gauss-Legendre multiplication formula, and the Euler's reflection formula. Now, using Eq. \eqref{26/08/14-6} we can represent the sum over $j$ in terms of the generalized hypergeometric functions. Namely, for $\ulamek{l}{k} \neq \ulamek{p}{q}$, the function $h_{l/k, p/q}(t, x)$ can be written as follows:
\begin{align}
\label{14/07/14-2}
h_{\frac{l}{k}, \frac{p}{q}}(t, x) &= \sum_{r=0}^{\infty} \sum_{n=0}^{m_{1}-1} \frac{(-t)^{n+r}}{n!\, r!}\, \frac{\Gamma(1 + Mn + mr)}{x^{1+ M n + mr}} \sin(\pi M n + \pi mr) \nonumber \\
&\qquad \qquad \times {_{1+lp}F_{m_{1}}}\left({1, \Delta(lp, 1 + M n + mr) \atop \Delta(m_{1}, 1+n)}; (-1)^{m_{1} - lp} \kappa\right),
\end{align}
with $\kappa = (lp)^{lp}/m_{1}^{m_{1}} (t^{1/M}/x)^{lp}$, with $m,M,m_1,$ and $M_1$, defined in Eq.\eqref{26/08/14-10}.
Moreover, it turns out that, after applying Eq. (8.2.2.3) of \cite{APPrudnikov-v3} to the finite sum in Eq. \eqref{14/07/14-2}, we can write
\begin{equation}
\label{15/07/14-1}
h_{\frac{l}{k}, \frac{p}{q}}(t, x) = \frac{x^{-1} m_{1}\sqrt{M}}{(2\pi)^{\ulamek{m_{1}-lp}{2}}} \sum_{r=0}^{\infty} \frac{(-t)^{r}}{r!} \left(\frac{m_{1}}{t}\right)^{\ulamek{m}{M} r} G^{m_{1}, 0}_{lp, m_{1}}\left(\kappa\Big\vert {\Delta(lp, 0) \atop \Delta(m_{1}, \ulamek{m}{M} r)}\right),
\end{equation}
where $G^{m, n}_{p, q}$ is the Meijer $G$ function defined in Eq. \eqref{26/08/14-4}.
The series representing $h_{l/k, p/q}(t, x)$ given in Eq. \eqref{14/07/14-2} converges absolutely for $x \geq 0$. The proof of this fact can be split into two cases: $x > 0$, and for $x = 0$. For $x > 0$, and fixed values of $l$, $k$, $p$, $q$, and $n=0, \ldots, m_{1}-1$, the convergence of the series $|h_{l/k, p/q}(t, x)|$ follows from the convergence of the ${_{1+lp}F_{m_{1}}}$ functions which, for a given $n$, converges to a constant $C_{n}$ \cite{APPrudnikov-v3, NIST}. Thus, we get
\begin{align}
\label{29/08/14-1}
|h_{\frac{l}{k}, \frac{p}{q}}(t, x)| &\leq \sum_{r=0}^{\infty}\sum_{n=0}^{m_{1}-1} \frac{t^{n+r}}{n! r!} \frac{\Gamma(1 + Mn + mr)}{x^{1+Mn + mr}} |\sin(\pi M n + \pi m r) C_{n}| \nonumber \\
& \leq \sum_{n=0}^{m_{1}-1} \frac{|C_{n}| t^{n}}{n!\, x^{M n}} \sum_{r=0}^{\infty} \frac{t^{r}}{r!} \frac{\Gamma(1 + M n + m r)}{x^{1 + mr}}.
\end{align}
An application of the Cauchy ratio test of convergence \cite{GBArfken05} to Eq. \eqref{29/08/14-1} completes the proof of convergence for $|h_{l/k, p/q}(t, x)|$, for $x > 0$.
Let us now show that the series representing $h_{l/k, p/q}(t, x)$ converges absolutely at $x=0$. For this purpose we will find the asymptotic behavior of the ${_{1+lp}F_{m_{1}}}$ function in Eq. \eqref{14/07/14-2} for $\kappa \gg 1$, where the relation between $\kappa$ and $x$ is shown below Eq. \eqref{14/07/14-2}. It follows from \cite{EWBarnes07}, that
\begin{equation}
\label{14/07/14-3}
{_{1+ lp}F_{m_{1}}}(\ldots) \approx (-1)^{\frac{1}{2} + (M-1)n + mr} \frac{(lp)^{Mn + mr }}{m_{1}^{n}} \kappa^{\frac{1/2 - n + Mn + mr}{m_{1} - lp}} e^{- (m_{1}-lp) \kappa^{\ulamek{1}{m_{1} - lp}}},
\end{equation}
where, at the point $x = 0$, there exists an essential singularity. That gives
\begin{equation}
\label{14/07/14-4}
h_{\frac{l}{k}, \frac{p}{q}}(t, x) \approx t^{\ulamek{1}{M}} \kappa^{1 + \ulamek{1}{2(m_{1} - lp)}} e^{-\big[m_{1} - lp + \big(\ulamek{t}{m_{1}}\big)^{m/M}\big] \kappa^{\ulamek{1}{m_{1}-lp}}} \sum_{n=0}^{m_{1}-1} (-1)^{M n} \sin(\pi M n) \kappa^{\ulamek{n}{m_{1}} + \ulamek{M-1}{m_{1} - lp}n}
\end{equation}
and, consequently,
\begin{align}
\label{14/07/14-5}
|h_{\frac{l}{k}, \frac{p}{q}}(t, x)| &\leq t^{\ulamek{1}{M}} \kappa^{1 + \ulamek{1}{2(m_{1}-lp)}} e^{-\big[m_{1} - lp + \big(\ulamek{t}{m_{1}}\big)^{m/M}\big] \kappa^{\ulamek{1}{m_{1}-lp}}} \sum_{n=0}^{m_{1}-1} \kappa^{\big(\ulamek{1}{m_{1}} + \ulamek{M-1}{m_{1} - lp}\big)n} \nonumber \\
& \leq t^{\ulamek{1}{M}} \kappa^{1 + \ulamek{1}{2(m_{1}-lp)}} e^{-\big[m_{1} - lp + \big(\ulamek{t}{m_{1}}\big)^{m/M}\big] \kappa^{\ulamek{1}{m_{1}-lp}}} m_{1} \kappa^{\big(\ulamek{1}{m_{1}} + \ulamek{M-1}{m_{1} - lp}\big)(m_{1}-1)} \nonumber \\
& = m_{1} t^{\ulamek{1}{M}} \kappa^{1 - \ulamek{1}{m_{1}} + \ulamek{1/2-M}{m_{1}-lp}} e^{-\big[m_{1} - lp + \big(\ulamek{t}{m_{1}}\big)^{m/M}\big] \kappa^{\ulamek{1}{m_{1}-lp}}}.
\end{align}
Taking into account the fact that, for $t > 0$, and fixed $l$, $k$, $p$, and $q$, such that $0 < l/k, p/q < 1$, and $l/k \neq p/q$, we can estimate Eq. \eqref{14/07/14-5} as follows:
\begin{equation}\label{14/07/14-6}
|h_{\frac{l}{k}, \frac{p}{q}}(t, x)| < B \kappa^{\ulamek{5}{2}} e^{- A \kappa},
\end{equation}
where $A$ and $B$ are positive constants. Now , from Eq. \eqref{14/07/14-6} it is easy to see that $|h_{l/k, p/q}(t, x)|\to 0$, for $x\to 0$.
\bigskip
We will conclude this section with several concrete examples of explicit expressions for bi-scale totally asymmetric L\'evy densities, where we can show that some of our expressions can be reduced to expressions in terms of more classical special functions. \\
\noindent
{\it Example 5.1. The case $\alpha_{1} = \alpha_{2} = 1/2$. }
Substituting the corresponding $v_{\alpha}(t, x)$ given in Eq. \eqref{26/09/2014-1a} into Eq. \eqref{25.06.14-1}, we have
\begin{align}
\label{19/09/2014-2a}
h_{\frac{1}{2}, \frac{1}{2}}(t, x) &= \frac{t^2}{4\pi} \int_{0}^{x} e^{-\ulamek{t^2}{4}\frac{x}{y(x-y)}} \frac{dy}{[y(x-y)]^{3/2}} = \frac{t^2}{2\pi} \int_{0}^{x/2} e^{-\ulamek{t^2}{4}\frac{x}{y(x-y)}} \frac{dy}{[y(x-y)]^{3/2}} \\ \label{19/09/2014-2b}
& = \frac{t^2}{\pi x^2} \int_{1}^{\infty} e^{-\ulamek{t^2}{x} z} \frac{dz}{\sqrt{z-1}} = \frac{t}{\sqrt{\pi} x^{3/2}} e^{-\ulamek{t^2}{x}} = g_{\frac{1}{2}}(x, 2t).
\end{align}
In Eq. \eqref{19/09/2014-2a} we changed the variable of integration as follows: $z = 1/ [1-(1 - 2y/x)^2]$. In Eq. \eqref{19/09/2014-2b} we also employed formula (3.362.1) on p. 344 of \cite{Gradshteyn}. The same result is obtained from Eq. \eqref{14/07/14-2} where, for $\alpha_{1} = \alpha_{2} = 1/2$, we employed the formulas (7.11.3.3), (7.11.3.4) of \cite{APPrudnikov-v3} and (5.6.1.1) of \cite{APPrudnikov-v2}. \\
\noindent
\textit{Example 5.2. The case $\alpha_{1} = 1/2$ and $\alpha_{2} = 1/3$.}
Using Eqs. \eqref{26/09/2014-1a} and \eqref{26/09/2014-1b}, the Laplace convolution of $v_{\frac{1}{3}}(t, x)$, and $v_{\frac{2}{3}}(t, x)$, takes the form,
\begin{align}
\label{26/09/2014-2}
h_{\frac{1}{2}, \frac{1}{3}}(t, x) & = \frac{t^{5/2}}{6 \pi^{3/2}} \int_{0}^{x} K_{\frac{1}{3}}\big(\ulamek{2}{\sqrt{x-y}} \ulamek{t^{3/2}}{3^{3/2}}\big) e^{-\frac{t^2}{4y}} \frac{dy}{[y(x-y)]^{3/2}} \nonumber \\
& = \frac{t^{5/2}}{6 \pi^{3/2} x^2} e^{-\frac{t^2}{4x}} \int_{0}^{\infty} u^{-3/2} (1+u) e^{-\frac{t^2}{4x} u} K_{\frac{1}{3}}\big(\sqrt{\ulamek{1+u}{x u}} \ulamek{2\, t^{3/2}}{3^{3/2}}\big) du
\end{align}
where $u = (x-y)/y$. Employing the integer representation of $K_{\nu}(z)$ given in Eq. (8.432) on p. 917 in \cite{Gradshteyn}
we get
\begin{align}
\label{26/09/2014-3a}
h_{\frac{1}{2}, \frac{1}{3}}(t, x) & = \frac{t^{3} x^{-\frac{13}{6}}}{12\sqrt{3}\pi^{\frac{3}{2}}} e^{-\frac{t^2}{4x}} \int_{0}^{\infty} u^{-\frac{5}{3}} (1+u)^{\frac{7}{6}} e^{-\frac{t^2}{4x} u}\left[ \int_{0}^{\infty} \omega^{-\frac{4}{3}} e^{-\omega-\frac{t^3}{27 x \omega} \frac{1+u}{u}} d\omega\right] du \\ \label{26/09/2014-3b}
& = \frac{t^{3} x^{-\frac{13}{6}}}{12\sqrt{3}\pi^{\frac{3}{2}}} e^{-\frac{t^2}{4x}} \int_{0}^{\infty} u^{-\frac{5}{3}} (1+u)^{\frac{7}{6}} e^{-\frac{t^2}{4x} u}\left\{\int_{0}^{\infty} \omega^{-\frac{4}{3}} e^{-\omega} \sum_{r=0}^{\infty} \frac{[-\frac{t^3}{27 x \omega} \frac{1+u}{u}]^r}{r!} d\omega\right\} du \\ \label{26/09/2014-3c}
& = \frac{t^{3} x^{-\frac{13}{6}}}{12\sqrt{3}\pi^{\frac{3}{2}}} e^{-\frac{t^2}{4x}} \sum_{r=0}^{\infty} \frac{(-\frac{t^3}{27 x})^r}{r!} \int_{0}^{\infty} u^{-\frac{5}{3}-r} (1+u)^{\frac{7}{6} + r} e^{-\frac{t^2}{4 x} u} du \int_{0}^{\infty} \omega^{-\frac{4}{3}-r} e^{-\omega} d\omega\\ \label{26/09/2014-3d}
& = \frac{t^{3} x^{-\frac{13}{6}}}{12\sqrt{3}\pi^{\frac{3}{2}}} e^{-\frac{t^2}{4x}} \sum_{r=0}^{\infty} \frac{(-\frac{t^3}{27 x})^r}{r!} \Gamma(-\ulamek{1}{3}-r) \Gamma(-\ulamek{2}{3}-r) \psi(-\ulamek{2}{3} - r; \ulamek{3}{2}; \ulamek{t^2}{4 x}) \\ \label{26/09/2014-3e}
& = \frac{2^{-\frac{1}{6}}\sqrt{\pi} t^2}{9\sqrt{3} x^{\frac{5}{3}}} e^{-\frac{t^2}{8 x}} \sum_{r=0}^{\infty} \frac{(-\frac{t^3}{3^3 2x})^r\, D_{\frac{7}{3} + 2 r}(\ulamek{t}{\sqrt{2 x}})}{r! \Gamma(\frac{4}{3}+r) \Gamma(\frac{5}{3}+r)} \\ \label{26/09/2014-3f}
&= \frac{\exp(-\frac{t^2}{8x})}{\sqrt{2\pi} x} \sum_{r=0}^{\infty} \frac{(-t)^{2+3r}}{(2+3r)!} (2 x)^{-\frac{2+3r}{3}} D_{1 + \frac{2}{3}(2+3r)}(\ulamek{t}{\sqrt{2 x}}) \\ \label{26/09/2014-3g}
& = \frac{\exp(-\frac{t^2}{8x})}{\sqrt{2\pi} x} \sum_{n=0}^{\infty} \frac{(-t)^{n}}{n!} (2 x)^{-\frac{n}{3}} D_{1 + \frac{2}{3}n}(\ulamek{t}{\sqrt{2 x}})
\end{align}
To obtain Eq. \eqref{26/09/2014-3a} we used the series expansion of the exponential function, and thereafter we changed the order of integrals and summation. We also applied the definition of Tricomi's function (the confluent hypergeometric function of the second kind) $\psi(a; b; z)$ given in Eq. (9.211.4) on p. 1023 of \cite{Gradshteyn}. In Eq. \eqref{26/09/2014-3d} we employed Eq. \eqref{26/08/14-7} and formula (7.11.4.8) on p. 491 of \cite{APPrudnikov-v3}, whereas in Eq. \eqref{26/09/2014-3e} we used Eq. \eqref{26/08/14-8}. In Eq. \eqref{26/09/2014-3f} we changed the summation index from $2+ 3r$ to $n$.
Note that Eq. \eqref{26/09/2014-3a} can be also obtained from Eq. \eqref{14/07/14-2}. Indeed, using the formulas (7.11.3.3) and (7.11.3.4) of \cite{APPrudnikov-v3} we get
\begin{align}
\label{29.06.14-3a}
h_{\frac{1}{2}, \frac{1}{3}}(t, x) & = \sum_{r=0}^{\infty} \frac{(-t)^{r}}{r!} \frac{x^{-\ulamek{r}{3}-1}}{\Gamma\big(\!-\!\ulamek{r}{3}\big)} {_{1}F_{1}}\left({1 + \frac{r}{3} \atop \frac{1}{2}} ; -\frac{t^{2}}{4x}\right) + \sum_{r=0}^{\infty} \frac{(-t)^{r+1}}{r!} \frac{x^{-\ulamek{r}{3} - \ulamek{3}{2}}}{\Gamma\big(\!-\!\ulamek{r}{3}-\ulamek{1}{2}\big)} {_{1}F_{1}}\left({\frac{r}{3} + \frac{3}{2} \atop \frac{3}{2}} ; -\frac{t^{2}}{4x}\right) \nonumber \\
& = \frac{e^{-t^2/(8x)}}{\sqrt{2\pi} x} \sum_{r=0}^{\infty} \frac{(-t)^r}{r!} (2x)^{-r/3} D_{1+ \frac{2}{3} r}(\ulamek{t}{\sqrt{2x}}).
\end{align}
\newpage
\noindent
\textit{Example 5.3. The case $\alpha_{1} = 2/3$, $\alpha_{2} = 1/3$.}
Here, the Laplace convolution defined in Eq. \eqref{25.06.14-1} is equal to
\begin{align}
\label{8/10/2014-1a}
h_{\frac{1}{3}, \frac{2}{3}}(t, x) & = \frac{t^{\frac{3}{2}}}{\sqrt{3} \pi^{\frac{3}{2}}} \int_{0}^{x} e^{-\frac{2t^3}{27 y^2}} W_{\frac{1}{2}, \frac{1}{6}}\big(\ulamek{4 t^3}{27 y^2}\big) K_{\frac{1}{3}}\big(\ulamek{2}{\sqrt{x-y}}\ulamek{t^{3/2}}{3^{3/2}}\big) \frac{d y}{y (x-y)^{3/2}} \nonumber \\
& = \frac{[t/(3\pi)]^{3/2}}{2\sqrt{3}} \int_{1}^{\infty} \frac{\exp(-\frac{2t^3}{27 x^2} u)}{u^{\frac{1}{6}} (\sqrt{u}-1)^{\frac{5}{3}}}\, W_{\frac{1}{2}, \frac{1}{6}}\big(\ulamek{4 t^3}{27 x^2} u\big) K_{\frac{1}{3}}\big(\ulamek{2 t^{3/2}}{3^{3/2}} \ulamek{u^{1/4}}{(\sqrt{u} - 1)^{1/2}}\big) du \\
& = \frac{t^2 x^{-\frac{5}{3}}}{12\pi^{\frac{3}{2}}} \int_{1}^{\infty} \frac{\exp(-\frac{2t^3}{27 x^2} u)}{u^{\frac{1}{6}} (\sqrt{u}-1)^{\frac{5}{3}}}\, W_{\frac{1}{2}, \frac{1}{6}}\big(\ulamek{4 t^3}{27 x^2} u\big) \left[\int_{0}^{\infty} \omega^{-\frac{4}{3}} e^{-\omega - \frac{t^3}{27 x\omega} \frac{\sqrt{u}}{\sqrt{u}-1}} d\omega\right] du \nonumber \\
& = \frac{t^2 x^{-\frac{5}{3}}}{12\pi^{\frac{3}{2}}} \int_{1}^{\infty} \frac{\exp(-\frac{2t^3}{27 x^2} u)}{u^{\frac{1}{6}} (\sqrt{u}-1)^{\frac{5}{3}}}\, W_{\frac{1}{2}, \frac{1}{6}}\big(\ulamek{4 t^3}{27 x^2} u\big) \left\{\int_{0}^{\infty} e^{-\omega} \sum_{r=0}^{\infty} \frac{[-\ulamek{t^3}{27 x \omega}\ulamek{\sqrt{u}}{\sqrt{u}-1}]^r}{r!} \frac{d\omega}{\omega^{\frac{4}{3}} }\right\} du \nonumber \\
& = \frac{t^2 x^{-\frac{5}{3}}}{12\pi^{\frac{3}{2}}} \sum_{r=0}^{\infty} \frac{(-\frac{t^3}{27 x})^r}{r!} {\rm Int}(x, t) \int_{0}^{\infty} \omega^{-\frac{4}{3} - r} e^{-\omega}\, d\omega
\nonumber \\ \label{8/10/2014-1b}
&= \frac{t^2 x^{-\frac{5}{3}}}{12\pi^{\frac{3}{2}}} \sum_{r=0}^{\infty} \frac{(-\frac{t^3}{27 x})^r}{r!} \Gamma(-\ulamek{1}{3} - r) {\rm Int}(x, t).
\end{align}
Above, the integral representation of $K_{\mu}(z)$, see Eq. (8.432) on p. 917 in \cite{Gradshteyn}, was used. The auxiliary function ${\rm Int}(x, t)$ is defined as follows:
\begin{align}
\label{8/10/2014-2}
{\rm Int}(t, x) &= \int_{1}^{\infty} u^{-\frac{1}{6}+\frac{r}{2}} (\sqrt{u}-1)^{-\frac{5}{3}+\frac{r}{2}} e^{-\frac{2 t^3}{27 x^2} u}\, W_{\frac{1}{2}, \frac{1}{6}}\big(\ulamek{4 t^3}{27 x^2} u\big) du \nonumber \\
& = 2\sqrt{\frac{\pi}{3}}\, {_{2}F_{2}}\left({\ulamek{5}{6} + \ulamek{r}{2}, \ulamek{4}{3} + \ulamek{r}{2} \atop \ulamek{1}{3}, \ulamek{2}{3}}; -\frac{4 t^3}{27 x^2}\right) \nonumber \\
&- 2\sqrt{\frac{\pi}{3}} \frac{t}{x^{2/3}} \frac{\Gamma(-\ulamek{2}{3}-r)}{\Gamma(-\ulamek{4}{3}-r)}\, {_{2}F_{2}}\left({\ulamek{7}{6} + \ulamek{r}{2}, \ulamek{5}{3} + \ulamek{r}{2} \atop \ulamek{2}{3}, \ulamek{4}{3}}; -\frac{4 t^3}{27 x^2}\right) \nonumber \\
& + \sqrt{\frac{\pi}{3}} \frac{t^2}{x^{4/3}} \frac{\Gamma(-\ulamek{2}{3}-r)}{\Gamma(-2-r)}\, {_{2}F_{2}}\left({\ulamek{3}{2} + \ulamek{r}{2}, 2 + \ulamek{r}{2} \atop \ulamek{4}{3}, \ulamek{5}{3}}; -\frac{4 t^3}{27 x^2}\right),
\end{align}
where Eq. (2.19.1.4) on p. 201 of \cite{APPrudnikov-v3} was employed. Substituting Eq. \eqref{8/10/2014-2} into Eq. \eqref{8/10/2014-1b}, using the Euler's reflection formula, the Gauss-Legendre multiplication formula, and changing the summation index as follows $n = 2 + 3r$, we get
\begin{align}
\label{15/07/14-4}
h_{\frac{1}{3}, \frac{2}{3}}(t, x) & = \sum_{n=0}^{\infty} \frac{(-t)^{n}}{n!} \frac{x^{-1-\ulamek{n}{3}}}{\Gamma(-\ulamek{n}{3})} {_{2}F_{2}}\left({\ulamek{1}{2} + \ulamek{n}{6}, 1 + \ulamek{n}{6} \atop \ulamek{1}{3}, \ulamek{2}{3}}; -\frac{4 t^{3}}{27 x^{2}}\right) \nonumber \\
& + \sum_{n=0}^{\infty} \frac{(-t)^{1+n}}{n!} \frac{x^{-\ulamek{5}{3}-\ulamek{n}{3}}}{\Gamma(-\ulamek{2}{3}-\ulamek{n}{3})} {_{2}F_{2}}\left({\ulamek{5}{6} + \ulamek{n}{6}, \ulamek{4}{3} + \ulamek{r}{6} \atop \ulamek{2}{3}, \ulamek{4}{3}}; -\frac{4 t^{3}}{27 x^{2}}\right) \nonumber \\
& + \sum_{n=0}^{\infty} \frac{(-t)^{2+n}}{2 n!} \frac{x^{-\ulamek{7}{3}-\ulamek{n}{3}}}{\Gamma(-\ulamek{4}{3}-\ulamek{n}{3})} {_{2}F_{2}}\left({\ulamek{7}{6} + \ulamek{n}{6}, \ulamek{5}{3} + \ulamek{n}{6} \atop \ulamek{4}{3}, \ulamek{5}{3}}; -\frac{4 t^{3}}{27 x^{2}}\right).
\end{align}
Finally, we would like to point out that ${_{2}F_{2}}$'s cannot be expressed in terms of any other special functions. And, of course, the explicit form of $h_{\frac{1}{3}, \frac{2}{3}}(x, t)$ can be obtained from Eq. \eqref{14/07/14-2} and Eq. \eqref{15/07/14-4}. In Fig. 5.2 we show plots of the densities discussed in Examples 5.1-3. They were efficiently obtained via the symbolic manipulation platform {\it Mathematica}
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.8]{fig1.eps
\caption{\label{fig1} The plots of $h_{\alpha_{1}, \alpha_{2}}(x, 1)$. Plot I (red): $\alpha_{1} = \alpha_{2} = \ulamek{1}{2}$, see Eq. \eqref{19/09/2014-2a}); plot II (blue): $\alpha_{1} = \ulamek{1}{2}$, $\alpha_{2} = \ulamek{1}{3}$, see Eq. \eqref{29.06.14-3a}; and plot III (green): $\alpha_{1} = \ulamek{2}{3}$, $\alpha_{2} = \ulamek{1}{3}$, see Eq. \eqref{15/07/14-4}.}
\end{center}
\end{figure}
\section{Conclusions and comments; the Stieltjes moment problem}
The goal of this work was to show how the asymptotic behavior of solutions of certain nonlinear conservation laws can be explicitly described in terms of some special functions which make the efficient computation of the related probability densities possible. In our future work we plan to expand this work to other types of asymptotic problems for other nonlinear evolution equations discussed in, e.g., \cite{KW2008}, and other papers referenced therein.
Here, we want to conclude with an example of how the results obtained in Section 5
can be applied to the solution of the classical Stieltjes moment problem for special cases considered in Examples 5.1-3.
Recall, that the Stieltjes moment problem \cite{Akhiezer1965, KAPenson2009} can be formulated as follows: Find a positive function $W(x)$ which satisfies the infinite set of equations,
\begin{equation}
\label{10/10/2014-1}
\int_{0}^{\infty} x^{n} W(x) dx = \rho(n), \quad n=0, 1, 2, \ldots,
\end{equation}
for a given moment sequence $\{\rho (n)\}$.
From the practical point of view, it plays a special role in probability theory where the moment estimators can be usually conveniently obtained, and the issue is whether they determine the relevant probability distribution. In general, the solution to the Stieltjes moment problem can be unique or non-unique. The examples of non-unique solutions can be found, e.g., in \cite{KAPenson2009}. Below, we quote one uniqueness criterion, and one non-uniqueness criterion to be used below.
\begin{itemize}
\item[\textbf{C1}] {\it Carleman uniqueness criterion.} (see, .e.g., \cite{Akhiezer1965}).
{If $\,S = \sum_{n=1}^{\infty} [\rho(n)]^{-\frac{1}{2n}}=\infty$, then $W(x)$ is uniquely determined by Eq. \eqref {10/10/2014-1}.}
\item[\textbf{C2}] {\it Carleman non-uniqueness criterion, see, .e.g., {\rm\cite{AGPeaks01, AGut02}}.} If $\sum_{n=1}^{\infty} [\rho(n)]^{-\frac{1}{2n}} < \infty$, and if there exists ${x' > 0}$ such that, for all $x > x'$, $W(x) > 0$, and $\psi(y) = -\ln[W(e^{y})]$ is convex in $y' > 0$, where $y' = \ln(x')$, then $W(x)$ is non-unique.
\end{itemize}
Let us begin with finding the $\mu$-th moments of $h_{\alpha_{1}, \alpha_{2}}(x, t)$:
\begin{equation}\label{31/07/14-4}
\rho_{\alpha_{1}, \alpha_{2}}(\mu) = \int_{0}^{\infty} x^{\mu} h_{\alpha_{1}, \alpha_{2}}(t, x) dx, \quad \mu\in\mathbb{R}.
\end{equation}
Using Eq. \eqref{15/07/14-1}, changing the variable $x$ to $y = x^{-lp}$, applying Eq. (2.24.2.1) in \cite{APPrudnikov-v3}, and the Gauss-Legendre multiplication formula (see Eq. \eqref{26/08/14-8}), we arrive at
\begin{align}
\label{15/07/14-2a}
\rho_{\alpha_{1}, \alpha_{2}}(\mu)
& = \frac{1}{lp} \int_{0}^{\infty} y^{-\frac{\mu+1}{lp}-1} h_{\alpha_{1}, \alpha_{2}}(y^{-\frac{1}{lp}}, t) dy \\
& = \frac{1}{M} \sum_{r=0}^{\infty} \frac{(-1)^r}{r!}\, \frac{\Gamma(\frac{m r - \mu}{M})}{\Gamma(-\mu)}\, t^{(1-\frac{m}{M})r + \frac{\mu}{M}}, \quad t >0. \label{15/07/14-2b}
\end{align}
The absolute convergence of the series in Eq. \eqref{15/07/14-2b} is easy to check through the Cauchy ratio test. Eq. \eqref{15/07/14-2b} implies that the $\mu$th moment is finite for a given $r$, and $\mu < m r$, whereas it is infinite for $\mu\geq m r$. The moment $\rho_{\alpha_{1}, \alpha_{2}}(\mu)$ is equal to one for $\mu =0$. The $\mu$-th moment of $h_{\alpha_{1}, \alpha_{2}}(x, t)$, for $\mu = -lp n$, $n=0, 1, 2, \ldots$, gives rise to the Stieltjes moment problem for which, from the comparison of Eq. \eqref{15/07/14-2a} with Eq. \eqref{10/10/2014-1}, $W(x)$ and $\rho(n)$ are of the form,
\begin{equation}
\label{16/10/2014-1}
W_{\alpha_{1}, \alpha_{2}}(t, x) = \frac{1}{lp} x^{-1-\frac{1}{lp}} h_{\alpha_{1}, \alpha_{2}}(t, x^{-\frac{1}{lp}}), \quad \text{\rm and} \quad \rho(n) = \rho_{\alpha_{1}, \alpha_{2}}(-lp n).
\end{equation}
Now, we can employ the above criteria \textbf{C1} and \textbf{C2} to take a closer look at the Stieltjes moment problem related to the Examples 5.1-3.
In Examples 5.1-2 functions $W_{\alpha_{1}, \alpha_{2}}(x)$ are unique. Indeed, for $\alpha_{1} = \alpha_{2} = 1/2$, we have $\rho(n) = 2 (2t)^{-2n} \Gamma(2n)/\Gamma(n)$, which gives $S = 2 t \sum_{n=1}^{\infty} [2\Gamma(2n)/\Gamma(n)]^{-\frac{1}{2n}}$. Using the Stirling formula for the gamma function, Eq. (8.327.1) on page 895 of \cite{Gradshteyn}, we get that the series $S \approx t \sqrt{e}/(2\sqrt{2}) \sum_{n=1}^{\infty} n^{-1/2}$; the latter series is divergent. Thus, \textbf{C1} leads to the unique function $W_{\frac{1}{2}, \frac{1}{2}}(t, x) = t/(\sqrt{\pi x}) \exp(- x t^2)$. Similar considerations are valid of the case of $\alpha_{1} = 1/2$ and $\alpha_{2} = 1/3$ which leads to the uniqueness of the appropriate function $W(t, x)$ in that case as well.
However, Example 5.3 leads to a non-unique Stieltjes moment problem. Indeed, for $\alpha_{1} = 2/3$, and $\alpha_{2} = 1/3$, we have $\rho(n) = 3 t^{-6n} \Gamma(6n)/\Gamma(2n)$, and the series $S = t^3 \sum_{n=1}^{\infty} 3^{-\frac{1}{2n}} [\Gamma(6n)/\Gamma(2n)]^{-\frac{1}{2n}} \approx t^3 e^2/108 \sum_{n=1}^{\infty} 3^{-\frac{1}{4n}} n^{-2}$ is convergent (we used the Stirling formula Eq. (8.327.1) on page 895 of \cite{Gradshteyn} here). The second condition in \textbf{C2} requires verification of the sign of the second derivative of $\psi_{\frac{1}{3}, \frac{2}{3}}(t, x) = -\ln[W_{\frac{1}{3}, \frac{2}{3}}(t, e^{x})]$,
\begin{align}\label{15/01/2015-1}
\frac{d^2 \psi_{\frac{1}{3}, \frac{2}{3}}(t, x)}{d x^2} =& -\frac{d^2 \ln[W_{\frac{1}{3}, \frac{2}{3}}(t, e^{x})]}{d x^2}
= - u h_{\frac{1}{3}, \frac{2}{3}}(t, u) \frac{d}{d u} h_{\frac{1}{3}, \frac{2}{3}}(t, u)h_{\frac{1}{3}, \frac{2}{3}}(t, u) \nonumber \\
& - u^2 h_{\frac{1}{3}, \frac{2}{3}}(t, u) \frac{d^2}{d u^2} + u^2 \left[\frac{d}{d u} h_{\frac{1}{3}, \frac{2}{3}}(t, u)\right]^2,
\end{align}
for $u = \exp(-x/2)$. If Eq. \eqref{15/01/2015-1} is positive then $\psi(y)$ is convex and $W_{\frac{1}{3}, \frac{2}{3}}(t, x)$ is non-unique.
The positivity of Eq. \eqref{15/01/2015-1} with $h_{\frac{1}{3}, \frac{2}{3}}(t, u)$ given in Eq. \eqref{15/07/14-4} can be established analytically but for the purpose of this commentary section we are just showing the plot of Eq. \eqref{15/01/2015-1} in Fig. \ref{fig3}.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.6]{fig3.eps}
\caption{\label{fig3} The plots of $\frac{d^2}{d y^2} \psi_{\frac{1}{3}, \frac{2}{3}}(t, x)$ for different values of $t$. Plot I (red): $t=0.8$, plot II (blue): $t=1$, and plot III (green): $t=1.2$. }
\end{center}
\end{figure}
\section*{Acknowledgment}
K.~G\'{o}rska acknowledges support from the PHC Polonium, Campus France, project no. 28837QA and MNiSW under "Iuventus Plus 2015-2016" program no IP2014 013073.
|
\section{Introduction}\label{S:Intro}
Since its introduction in 1995 by Li-Tian \cite{LiTi} and Behrend-Fantechi \cite{BeFa}, the theory of virtual fundamental classes has played a key role in algebraic geometry, defining important invariants such as the Gromov-Witten invariant and the Donaldson-Thomas invariant. Quite a few methods for handling the virtual fundamental classes were discovered such as torus localization (\cite{GrPa}), degeneration (\cite{LiD}), virtual pullback (\cite{Mano1}) and cosection localization (\cite{KiemLi}). Often combining these methods turns out to be quite effective.
The purpose of this paper is to prove \begin{enumerate}
\item virtual pullback formula,
\item torus localization formula and
\item wall crossing formula
\end{enumerate}
for cosection localized virtual cycles. Our results can be thought of as generalizations of the corresponding results for the ordinary virtual cycles because when the cosection is trivial, these formulas coincide with those for the ordinary virtual cycles. For (2), we remove a technical assumption in \cite{GrPa} on the existence of an equivariant global embedding into a smooth Deligne-Mumford stack.
\medskip
A Deligne-Mumford stack $X$ is equipped with the intrinsic normal cone $\mathfrak{C}_X$ which is \'etale locally $[C_{U/V}/T_V|_U]$ if $U\to X$ is \'etale and $U\hookrightarrow V$ is an embedding into a smooth variety where $C_{U/V}$ is the normal cone of $U$ in $V$.
A perfect obstruction theory (\cite{BeFa}) gives us a vector bundle stack $\mathfrak{E}$ together with an embedding $\mathfrak{C}_X\subset \mathfrak{E}$.
The virtual fundamental class (\cite{BeFa, LiTi}) is then defined by applying the Gysin map to the intrinsic normal cone
$$[X]^{\mathrm{vir}} =0^!_\mathfrak{E}[\mathfrak{C}_X].$$
When there is a torus action on $X$ with respect to which the perfect obstruction theory is equivariant, the virtual fundamental class is
localized to the fixed locus $F=X^{\mathbb{C}^*}$ under suitable assumptions ({\cite{GrPa}):
\begin{equation}\label{1e1}
[X]^{\mathrm{vir}} =\imath_*\frac{[F]^{\mathrm{vir}} }{e(N^{\mathrm{vir}} )}\in A_\ast^{\mathbb{C}^*} X\otimes_{\mathbb{Q}[t]}\mathbb{Q}[t,t^{-1}].
\end{equation}
Here $\iota: F\to X$ is the inclusion and $t$ is the generator of the equivariant ring of $\mathbb{C}^*$.
\medskip
The construction of virtual fundamental class can be relativized for morphisms $f:X\to Y$ to give the virtual pullback
$$f^!:A_*(Y)\to A_*(X)$$
when the intrinsic normal cone $\mathfrak{C}_{X/Y}$ is embedded into a vector bundle stack $\mathfrak{E}$ on $X$.
When the perfect obstruction theories of $X$ and $Y$ are compatible with $\mathfrak{E}$, the virtual pullback gives us the formula (\cite{Mano1})
\begin{equation}\label{1e2}
f^![Y]^{\mathrm{vir}} =[X]^{\mathrm{vir}} .
\end{equation}
A wall crossing formula (\cite{KL3}) compares $[M_+]^{\mathrm{vir}} $ with $[M_-]^{\mathrm{vir}} $ when $M_+$ and $M_-$ are open Deligne-Mumford substacks of the quotient $[X/\mathbb{C}^*]$ of a Deligne-Mumford stack $X$ which are simple $\mathbb{C}^*$-wall crossings.
The cosection localization says that when there is an open $U\subset X$ and a surjective
$\sigma:\mathfrak{E}|_U\to \mathbb{C}_U$, we can define a
cosection localized virtual fundamental class $$[X]\virt_\loc\in A_*(X(\sigma))\quad \text{where }X(\sigma)=X-U$$
which satisfies usual expected properties such as deformation invariance and
$$\imath_*[X]\virt_\loc=[X]^{\mathrm{vir}} \in A_*(X)\quad \text{where } \imath:X(\sigma)\hookrightarrow X.$$
The construction of $[X]\virt_\loc$ in (\cite{KiemLi}) is obtained in two steps:
\begin{itemize}
\item (cone reduction) the intrinsic normal cone $\mathfrak{C}_{X}$ has support contained in $\mathfrak{E}(\sigma)$ where $\mathfrak{E}(\sigma)=\mathfrak{E}|_{X(\sigma)}\cup \mathrm{ker} (\mathfrak{E}|_U\to \mathbb{C}_U)$;
\item (localized Gysin map) there is a cosection localized Gysin map $$0_{\mathfrak{E},_{\mathrm{loc}} }^!:A_*(\mathfrak{E}(\sigma))\longrightarrow A_*(X(\sigma))$$ compatible with the usual Gysin map.
\end{itemize}
Then the cosection localized virtual fundamental class is defined as $$[X]\virt_\loc=0^!_{\mathfrak{E},_{\mathrm{loc}} }[\mathfrak{C}_X].$$
The cosection localized virtual fundamental class turned out to be quite useful (\cite{Buss1, CK, CLL, Clad, GS1, GS2, HLQ, JT, KL1, KL2, KoTh, LQ, MPT, PP, PT}).
For further applications, it seems desirable to have
cosection localized analogues for torus localization formula, virtual pullback and wall crossing formulas.
For instance, recently there arose a tremendous interest in the Landau-Ginzburg theory whose key invariants such as the Fan-Jarvis-Ruan-Witten invariants are defined algebro-geometrically by cosection localized virtual cycles (\cite{CLL, CKL}). The formulas proved in this paper will be useful in the theory of MSP fields developed in \cite{CLLL} tod
study the Gromov-Witten invariants and the Fan-Jarvis-Ruan-Witten invariants of quintic Calabi-Yau threefolds.
\medskip
In \S\ref{S:Man}, we prove the cosection localized virtual pullback formulas (cf. Theorems \ref{1thmain1} and \ref{1thmain2}). The proofs in \cite{Mano1} work with necessary modifications as long as the rational equivalences used in the proofs lie in the suitable substacks for localized Gysin maps. In \S\ref{S:Torus}, we prove the torus localization formula for the cosection localized virtual fundamental classes (cf. Theorem \ref{2thmain}). In this new proof, we do not require (1) the existence of an equivariant global embedding of $X$ into a smooth Deligne-Mumford stack and (2) the existence of a global resolution of the perfect obstruction theory as in \cite{GrPa}. The technical condition (1) is completely gone while (2) is significantly weakened to ($2'$) the existence of a global resolution of the virtual normal bundle $N^{\mathrm{vir}} $ only on the fixed locus $F$. Finally, in \S\ref{S:Wall}, we prove a wall crossing formula for cosection localized virtual fundamental classes. We remark that in \cite{KL2}, a degeneration formula for cosection localized virtual fundamental class was proved and it was effectively used to prove Maulik-Pandharipande's formulas for Gromov-Witten invariants of spin surfaces.
All schemes or Deligne-Mumford stacks in this paper are defined over the complex number field $\mathbb{C}$.
\section{Virtual pullback for cosection localized virtual cycles}\label{S:Man}
In this section, we show that Manolache's virtual pullback formula (\cite{Mano1}) holds for cosection localized virtual cycles (cf. Theorems \ref{1thmain1} and \ref{1thmain2}).
\def\mathcal{S} {\mathcal{S} }
\subsection{Virtual pullback of cosection localized virtual cycle}\label{ss2.1}
Let $f:X\to Y$ be a morphism of Deligne-Mumford stacks.
Let $\phi_X:E_{X}\to \mathbb{L} _{X}$ and $\phi_Y:E_{Y }\to \mathbb{L} _{Y }$ be (relative) perfect obstruction theories that fit into a commutative diagram of distinguished triangles
\begin{equation}\label{16}\xymatrix{
f^*E_{Y }\ar[r]^\varphi\ar[d]_{f^*\phi_Y} & E_{X } \ar[r]\ar[d]_{\phi_X} & E_{X/Y}\ar[r]\ar[d]^{\phi_{X/Y}} & \\
f^*\mathbb{L} _{Y }\ar[r] & \mathbb{L} _{X }\ar[r] & \mathbb{L} _{X/Y}\ar[r] &
}\end{equation}
\begin{defi}\label{11}
We say $f:X\to Y$ is \emph{virtually smooth} if $E_{X/Y}$ is perfect of amplitude $[-1,0]$.
\end{defi}
By \cite[\S3.2]{Mano1},
if $f$ is virtually smooth, then $\phi_{X/Y}$ is a perfect obstruction theory.
In the remainder of this section, we assume $f:X\to Y$ is virtually smooth.
By \cite{BeFa}, the perfect obstruction theory $\phi_{X/Y}:E_{X/Y}\to \mathbb{L} _{X/Y}$ gives us an embedding of the intrinsic normal sheaf into the vector bundle stack
$$h^1/h^0(\mathbb{L} _{X/Y}^\vee)\hookrightarrow h^1/h^0(E_{X/Y}^\vee)=:\mathfrak{E}_{X/Y}.$$
Moreover the intrinsic normal cone of the morphism $f$ is naturally embedded
into the intrinsic normal sheaf
$$\mathfrak{C}_{X/Y}\hookrightarrow \mathfrak{E}_{X/Y}.
$$
Let ${Ob}_{X }=h^1(E_{X }^\vee)$ and $Ob_{Y }=h^1(E_{Y }^\vee)$ be the obstruction sheaves and let $\sigma_Y:Ob_Y\to {\mathscr O}_Y$ be a cosection.
The morphism $f^*E_{Y }\to E_{X }$ induces a morphism $Ob_{X }=h^1(E_{X }^\vee)\to h^1(f^*E_Y^\vee)= f^*h^1(E_{Y }^\vee)=f^*Ob_{Y }$. Hence $\sigma_Y$ induces a cosection
$$\sigma_X:Ob_{X }\longrightarrow f^*Ob_{Y }\longrightarrow f^*{\mathscr O}_Y={\mathscr O}_X$$
of $Ob_{X }$. We call $\sigma_X$ the cosection induced from $\sigma_Y$.
\begin{defi} We denote by $X(\sigma)=\sigma_X^{-1}(0)$ and $Y(\sigma)=\sigma_Y^{-1}(0)$ the vanishing loci of the cosections $\sigma_X$ and $\sigma_Y$ respectively.
\end{defi}
Here $\sigma_X^{-1}(0)$ is the subvariety defined by the image (ideal) of $\sigma_X$.
\begin{lemm}\label{13} If $f$ is virtually smooth, $X(\sigma)=f^{-1}(Y(\sigma))=Y(\sigma)\times_Y X$.\end{lemm}
\begin{proof}
From the distinguished triangle $E_{X/Y}^\vee\to E_{X }^\vee\to f^*E_{Y }^\vee\mapright{+1}$, we obtain an exact sequence
\[ \cdots\longrightarrow Ob_{X }\longrightarrow f^*Ob_{Y }\longrightarrow h^2(E_{X/Y}^\vee)=0\]
where the identity holds because $E_{X/Y}$ is perfect of amplitude $[-1,0]$. Since $Ob_{X }\to f^*Ob_{Y }$ is surjective,
\[ \sigma_X:Ob_{X }\twoheadrightarrow f^*Ob_{Y }\mapright{f^*\sigma_Y} {\mathscr O}_X\]
is zero if and only if $f^*\sigma_Y$ is zero.
This proves the lemma.
\end{proof}
By Lemma \ref{13}, we have a Cartesian square
\begin{equation}\label{17}\xymatrix{
X(\sigma)\ar[d]_{\imath'}\ar[r]^{f_\sigma} & Y(\sigma)\ar[d]^{\imath}\\
X\ar[r]_f & Y
}\end{equation}
where the vertical arrows are inclusion maps.
\medskip
We recall Manolache's virtual pullback.
\begin{defi}\label{14} \cite{Mano1}
Suppose we have an embedding of the intrinsic normal cone $\mathfrak{C}_{X/Y}$ of $f:X\to Y$ into a vector bundle stack $\mathfrak{E}_{X/Y}$. Consider a fiber product diagram
\[\xymatrix{
X'\ar[r]^{f'} \ar[d]_{p} & Y'\ar[d]^q\\
X\ar[r]^f& Y.
}\]
Then the \emph{virtual pullback} is defined as the composite
$$f^!:A_*(Y')\to A_*(\mathfrak{C}_{X'/Y'})\to A_*(\mathfrak{C}_{X/Y}\times_XX')\to A_*(p^*\mathfrak{E}_{X/Y})\to A_{*+d}(X'),
$$
where the first arrow is $[B]\mapsto [\mathfrak{C}_{B\times_{Y'}X'/B}]$; the second arrow is via the inclusion $\mathfrak{C}_{X'/Y'}\to
\mathfrak{C}_{X/Y}\times_XX'$; the last arrow is the Gysin map $0^!_{\mathfrak{E}_{X/Y}}: A_*(p^*\mathfrak{E}_{X/Y})\to A_{*+d}(X')$
for the vector bundle stack $\mathfrak{E}_{X/Y}$. Here $d$ is the rank of $E_{X/Y}$.
\end{defi}
If $X$ is not connected, we consider each connected component separately.
Letting $X'=X$ and $Y'=Y$, we get $f^!:A_*(Y)\to A_{*+d}(X)$. Letting $X'=X(\sigma)$ and $Y'=Y(\sigma)$, we obtain $f_\sigma^!:A_*(Y(\sigma))\to A_*(X(\sigma))$. By \cite[Theorem 2 (i)]{Mano1}, these fit into a commutative diagram
\begin{equation}\label{18}\xymatrix{
A_*(Y(\sigma))\ar[r]^{f_\sigma^!}\ar[d]_{\imath_*} & A_{*+d}(X(\sigma))\ar[d]^{\imath'_*}\\
A_*(Y)\ar[r]_{f^!} & A_{*+d}(X)
}\end{equation}
For the virtual pullback formula, we need the following analogue of \cite[Lemma 4.7]{Mano1}.
\def\mathrm{ker} {\mathrm{ker} }
\begin{lemm}\label{110}
Let $f:X\to Y$ be a morphism of Deligne-Mumford stacks and ${{\cal N}}$ be a vector bundle stack on $X$ such that $\mathfrak{C}_{X/Y}\subset {{\cal N}}$. Let $\mathfrak{E}$ be a vector bundle stack on $Y$ with the zero section $0_\mathfrak{E}:Y\to \mathfrak{E}$.
Let $U\subset Y$ be open and $\sigma:\mathfrak{E}|_U\to \mathbb{C}_U$ be a surjective map of vector bundle stacks. Let $Y(\sigma)=Y-U$;
$X(\sigma)=X\times_YY(\sigma)$ and $f_\sigma: X(\sigma)\to Y(\sigma)$ the induced morphism.
Let $\mathfrak{E}(\sigma)=\mathfrak{E}|_{Y(\sigma)}\cup \mathrm{ker} (\sigma)$.
Then $(\mathfrak{C}_{X/\mathfrak{E}})_{\mathrm{red}}\subset f^*\mathfrak{E}\oplus {{\cal N}}$ where $\mathfrak{C}_{X/\mathfrak{E}}$ denotes the normal cone of the morphism
$0_{\cal E}\circ f: X\to Y\to \mathfrak{E}$.
Moreover for each irreducible $B\subset \mathfrak{E}(\sigma)$,
\[ f_\sigma^!0^!_{\mathfrak{E},_{\mathrm{loc}} }[B]=0_{f^*\mathfrak{E}\oplus {{\cal N}},_{\mathrm{loc}} }^![\mathfrak{C}_{X\times_\mathfrak{E} B/B}] \quad\text{in } A_*(X(\sigma))
\]
where $0^!_{\mathfrak{E},_{\mathrm{loc}} }$ and $0_{f^*\mathfrak{E}\oplus {{\cal N}},_{\mathrm{loc}} }^!$ denote the localized Gysin maps with respect to the cosections
$\sigma:\mathfrak{E}|_U\to \mathbb{C}_U$ and $(f^*\sigma,0):f^*\mathfrak{E}\oplus{{\cal N}}|_{f^{-1}(U)}\to \mathbb{C}_{f^{-1}(U)}$ respectively.
\end{lemm}
\begin{proof}
The inclusion $\mathfrak{C}_{X/\mathfrak{E}}\subset f^*\mathfrak{E}\oplus {{\cal N}}=f^*\mathfrak{E}\times_X{{\cal N}}$ follows from the identity $\mathfrak{C}_{X/\mathfrak{E}}=f^*\mathfrak{E}\times_X\mathfrak{C}_{X/Y}$ proved in \cite[Example 2.37]{Mano1}. We prove the last statement. If $B\subset \mathfrak{E}|_{Y(\sigma)}$, the localized Gysin maps are the ordinary Gysin maps and hence the lemma follows from \cite[Lemma 4.7]{Mano1}. So we may suppose $B\nsubseteq \mathfrak{E}|_{Y(\sigma)}$.
Further, by shrinking $Y$ if necessary, we can assume that $Y$ is integral and $B\to Y$ is dominant.
By \cite[\S2]{KiemLi}, we can choose a projective variety $Z$, a generic finite and proper morphism $\rho:Z\to Y$, a Cartier divisor $D$ on $Z$ such that $D$ is a linear combination of integral Cartier divisors; $D$ fits into the commutative diagram
\[\xymatrix{
Z\ar[r]^\rho & Y\\
D\ar@{^(->}[u]\ar[r]^{\rho_\sigma} & Y(\sigma)\ar@{^(->}[u]
}\]
and $\rho^*\sigma$ extends to a surjective map $\tilde{\sigma}:\rho^*\mathfrak{E}\to {\mathscr O}_Z(D)$, where by abuse of notation we think of
${\mathscr O}_Z(D)$ as the total space of the line bundle ${\mathscr O}_Z(D)$.
Let $\tilde\mathfrak{E}=\mathrm{ker} (\tilde\sigma)$, which is a bundle stack, and choose an integral $\tilde B\subset \tilde\mathfrak{E}$ such that $\tilde\rho_*[\tilde B]=k[B]$ for some $k>0$, where $\tilde\rho:\tilde\mathfrak{E}\to\mathfrak{E}(\sigma)$ is the composition $\tilde\mathfrak{E}\subset \rho^*\mathfrak{E}\to \mathfrak{E}$ of the inclusion with the first projection $\rho^*\mathfrak{E}=\mathfrak{E}\times_YZ$. Then by the definition in \cite[\S2]{KiemLi},
$$0^!_{\mathfrak{E},_{\mathrm{loc}} }[B]=\frac1{k}{\rho_\sigma}_*(0^!_{\tilde\mathfrak{E}}[\tilde B]\cdot D),
$$
where $\cdot D$ denotes the refined intersection for the inclusion $D\subset Z$ defined in \cite[Chapter 6]{Fulton}.
We further simplify the situation as follows. Because $\ti{\cal E}$ is a bundle stack, there are integral $Z_i\sub Z$
and rational $c_i$ so that $[\ti B]=\sum c_i[\ti{\cal E}|_{Z_i}]\in A_\ast\ti{\cal E}$. Because rational equivalence in $\ti{\cal E}$ induces
a rational equivalence in ${\cal E}(\sigma)$, to prove the theorem, we only need to consider the case where $\ti B=\ti{\cal E}$.
Thus the above identity becomes $0^!_{\mathfrak{E},_{\mathrm{loc}} }[B]=\frac1{k}{\rho_\sigma}_*[D]$.
Consider the Cartesian diagrams
\[\xymatrix{
D\ar[r]^{\rho_\sigma} & Y(\sigma) \ar[r]^{\sub} & Y\\
D'\ar[u]^{f'_\sigma}\ar[r]_{\rho'_\sigma} & X(\sigma)\ar[u]_{f_\sigma}\ar[r]_{\sub} & X\ar[u]_f.
}\]
Since virtual pullbacks commute with pushforwards
(cf. \cite[Theorem 4.1 (i)]{Mano1}), we have
$$ f_\sigma^!0^!_{\mathfrak{E},_{\mathrm{loc}} }[B]=\frac1{k} f_\sigma^!{\rho_\sigma}_*[D]=\frac1{k}{\rho'_\sigma}_*{f'_\sigma}^![D] =
\frac1{k} {\rho'_\sigma}_*0^!_{{{\cal N}}|_{D'}}[\mathfrak{C}_{D'/D}].$$
Consider the commutative diagram
\[\xymatrix{
Z'\ar[r]^{f'}\ar[d]^{\rho'} & Z\ar[r]^0\ar[d]^{\rho} & \tilde\mathfrak{E}\ar[d]^{\tilde\rho} & \tilde B\ar@{=}[l]\\
X\ar[r]^f & Y\ar[r]^{0} & \mathfrak{E}(\sigma) & B\ar[l]
}\]
where $Z':=X\times_Y Z$.
Let
$$\tilde\rho^f:{f'}^*\tilde\mathfrak{E}\to {f}^*\mathfrak{E}(\sigma)
$$
denote the pullback of $\tilde\rho$.
Since $\tilde\rho_*[\tilde {\cal E}]=k[B]$, we have
$$\tilde\rho^f_*[\mathfrak{C}_{X\times_\mathfrak{E} \tilde {\cal E}/\tilde {\cal E}}]=k[\mathfrak{C}_{X\times_\mathfrak{E} B/B}].$$
By the definition of the localized Gysin map \cite[\S2]{KiemLi}, we have
$$0_{f^*\mathfrak{E}\oplus {{\cal N}},_{\mathrm{loc}} }^![\mathfrak{C}_{X\times_\mathfrak{E} B/B}]=
\frac1{k}{\rho'_\sigma}_*\left( 0^!_{{f'}^*\tilde\mathfrak{E}\oplus {\rho'}^*{{\cal N}}}[ \mathfrak{C}_{X\times_\mathfrak{E} \tilde {\cal E}/\tilde {\cal E}}] \cdot D'\right).
$$
Here by $\cdot D'$ we mean the intersection with $D\sub Z$ via the Cartesian square
$$\begin{CD}
D'@>>> Z'\\
@VVV @VVf'V\\
D @>>> Z
\end{CD}
$$
Since $X\times_\mathfrak{E}\tilde\mathfrak{E}=Z'$, by \cite[Example 2.37]{Mano1}
$$\mathfrak{C}_{X\times_\mathfrak{E} \tilde \mathfrak{E}/\tilde \mathfrak{E}}=\mathfrak{C}_{Z'/\tilde\mathfrak{E}}=\mathfrak{C}_{Z'/Z}\times_{Z'}{f'}^*\tilde\mathfrak{E}.
$$
We thus have
$$0_{f^*\mathfrak{E}\oplus {{\cal N}},_{\mathrm{loc}} }^![\mathfrak{C}_{X\times_\mathfrak{E} B/B}]=\frac1{k}{\rho'_\sigma}_*\bigl(0^!_{{{\cal N}}}[\mathfrak{C}_{Z'/Z}]\cdot D'\bigr)
=\frac1{k}{\rho'_\sigma}_*0^!_{{{\cal N}|_{D'}}}\left([\mathfrak{C}_{Z'/Z}]\cdot D'\right)
$$
by the commutativity of Gysin maps.
Let $L={\mathscr O}_Z(D)$ and let $L'={f'}^*L$. We now prove that
$$[\mathfrak{C}_{D'/D}]
[\mathfrak{C}_{Z'/Z}]\cdot D'.
$$
Indeed, by Vistoli's rational equivalence \cite{Vistoli} and \cite[Example 2.37]{Mano1},
and using $\mathfrak{C}_{D/Z}\cong L|_{D}$, we have
$$[\mathfrak{C}_{Z'/Z}]\cdot D'=0^!_{L'}[\mathfrak{C}_{D'/\mathfrak{C}_{Z'/Z}}]=0^!_{L'}[\mathfrak{C}_{D'/Z}]
=0^!_{L'}[\mathfrak{C}_{D'/\mathfrak{C}_{D/Z}}
$$
$$=0^!_{L'}[\mathfrak{C}_{D'/D}\times_{D'}L'|_{D'}]=[\mathfrak{C}_{D'/D}].
$$
Therefore, we have
$$0_{f^*\mathfrak{E}\oplus {{\cal N}},_{\mathrm{loc}} }^![\mathfrak{C}_{X\times_\mathfrak{E} B/B}]=\frac1{k}{\rho'_\sigma}_*0^!_{{{\cal N}|_{D'}}}[\mathfrak{C}_{D'/D}].$$
This proves the desired equality $ f_\sigma^!0^!_{\mathfrak{E},_{\mathrm{loc}} }[B]=0_{f^*\mathfrak{E}\oplus {{\cal N}},_{\mathrm{loc}} }^![\mathfrak{C}_{X\times_\mathfrak{E} B/B}].$
\end{proof}
The following is a cosection localized analogue of \cite[Corollary 4.9]{Mano1}.
\begin{theo}\label{1thmain1}
Let $f:X\to Y$ be a virtually smooth morphism of Deligne-Mumford stacks, and let $\sigma:Ob_{Y }\to {\mathscr O}_Y$ be a cosection. Then
$$f_\sigma^![Y]\virt_\loc =[X]\virt_\loc.$$
\end{theo}
\begin{proof}
The proofs of Theorem 4 (functoriality) and Corollary 4 in \cite{Mano1} work with necessary modifications.
The reader is invited to go through the proofs in \cite{Mano1} with the proof of deformation invariance \cite[Theorem 5.2]{KiemLi} for cosection localized virtual cycles in mind.
With Lemma \ref{110} at hand, one will find that the only thing to be checked is the inclusion of the support
\begin{equation}\label{19}\mathrm{Supp}\, \mathfrak{C}_{X\times \mathbb{P}^1/M^0_{Y}}\subset \mathrm{ker}\left[ h^1/h^0(c(u)^\vee)\longrightarrow q^*{\mathscr O}_{\mathbb{P}^1}(-1) \right]\end{equation}
where $M^0_Y$ is the deformation space from the reduced point $\{pt\}$ to the intrinsic normal cone $\mathfrak{C}_Y$ and $c(u)$ is the cone of the morphism
$$
u=(x_0\cdot \mathrm{id}, x_1\cdot \varphi):p^*f^*E_Y\otimes q^*{\mathscr O}_{\mathbb{P}^1}(-1)\longrightarrow p^*f^*E_Y\oplus p^*E_X\quad\text{in}\quad D(X\times\mathbb{P}^1).
$$
Here $x_0,x_1$ are the homogeneous coordinates of $\mathbb{P}^1$; $p,q$ are projections from $X\times \mathbb{P}^1$ to $X$ and $\mathbb{P}^1$ respectively; $\varphi:f^*E_Y\to E_X$ is the morphism in \eqref{16}. The morphism $ h^1/h^0(c(u)^\vee)\longrightarrow q^*{\mathscr O}_{\mathbb{P}^1}(-1) $ comes from the commutative diagram
$$\xymatrix{
h^1/h^0(c(u)^\vee)\ar[r]\ar[d] & p^*f^*h^1/h^0(E_Y^\vee)\oplus p^*h^1/h^0(E_X^\vee) \ar[r]\ar[d] & p^*f^*h^1/h^0(E_Y^\vee)\otimes q^*{\mathscr O}_{\mathbb{P}^1}(1)\ar[d]\\
q^*{\mathscr O}_{\mathbb{P}^1}(-1)\ar[r] & q^*{\mathscr O}_{\mathbb{P}^1}\oplus q^*{\mathscr O}_{\mathbb{P}^1} \ar[r] & q^*{\mathscr O}_{\mathbb{P}^1}(1)
}$$
where the middle and right vertical arrows come from the cosection $\sigma$.
By the double deformation space construction,
$$\mathfrak{C}_{X\times \mathbb{P}^1/M^0_{Y}}\times_{\mathbb{P}^1}(\mathbb{P}^1-\{(0:1)\})=\mathfrak{C}_X\times (\mathbb{P}^1-\{(0:1)\}).$$
By the cone reduction (\cite[\S4]{KiemLi}), we have the inclusion of the support
$$\mathrm{Supp}\, \mathfrak{C}_X \subset \mathrm{ker}\left[ h^1/h^0(E^\vee_X)\to {\mathscr O}_X \right].$$
Hence \eqref{19} holds over $\mathbb{P}^1-\{(0:1)\}$.
Let $D=\mathfrak{C}_{X\times \mathbb{P}^1/M^0_{Y}}\times_{\mathbb{P}^1}(\mathbb{P}^1-\{(1:0)\})$ be open in $\mathfrak{C}_{X\times \mathbb{P}^1/M^0_{Y}}$ containing the fiber over $(0:1)$. By diagram chase,
$$h^1/h^0(c(u)^\vee)|_{\{(0:1)\}}\cong f^*h^1/h^0(E_Y^\vee)\times_X h^1/h^0(E_{X/Y}^\vee)$$
and the homomorphism in \eqref{19} over the point $(0:1)$ is
$$(f^*\sigma,0):f^*h^1/h^0(E_Y^\vee)\times_X h^1/h^0(E_{X/Y}^\vee)\longrightarrow {\mathscr O}_X.
$$
Therefore the theorem follows if we show that irreducible components $A$ of
$D$ lying over $X\times \{(0:1)\}$ have support contained in
\begin{equation}\label{1501251}
\mathrm{ker}\left(
f^*h^1/h^0(E_Y^\vee)\mapright{f^*\sigma} {\mathscr O}_X \right)\times_X h^1/h^0(E_{X/Y}^\vee).
\end{equation}
Let $A$ be an irreducible component of $D$ lying over $X\times \{(0:1)\}$ and let $a$ be a general closed point in $A$. We claim that $a$ is contained in \eqref{1501251}.
Since the problem is local, we may assume $X$, $Y$ are affine, equipped with embeddings $X\subset V$, $Y\subset W$ into smooth affine varieties that fit into a commutative diagram
\[\xymatrix{
X\ar@{^(->}[r]\ar[d]_f & V\ar[d]^g\ar[r]^(.4){\gamma\circ g,\eta} & \mathbb{A} ^m\times \mathbb{A} ^n\ar[d]^{pr_1}\\
Y\ar@{^(->}[r] & W \ar[r]^\gamma & \mathbb{A} ^m
}\]
such that the morphism $g:V\to W$ is a smooth, $X=\mathrm{zero} (\gamma\circ g,\eta)$, $Y=\mathrm{zero} (\gamma)$ and
$$E_Y=[{\mathscr O}_Y^{\oplus m}\mapright{d\gamma} \Omega_W|_Y], \quad E_{X/Y}=[{\mathscr O}_X^{\oplus n}\mapright{d\eta} \Omega_{V/W}|_X].$$
Since we have nothing to prove when $\sigma=0$ at general points of the irreducible component $A$, we may assume that $\sigma$ is surjective.
To prove the claim, we recall the double deformation space construction for $D$ (cf. \cite{KKP}): Let $\Gamma$ be the graph of the morphism
\[
V\times (\mathbb{A} ^1_t-\{0\})\times (\mathbb{A} ^1_s-\{0\})\longrightarrow \mathbb{A} ^m\times\mathbb{A} ^n, \quad (v,t,s)\mapsto ((ts)^{-1}\gamma\circ g(v), t^{-1}\eta(v))
\]
and let $\bar\Gamma$ be the closure of $\Gamma$ in $V\times \mathbb{A} ^1_t\times\mathbb{A} ^1_s\times\mathbb{A} ^m\times \mathbb{A} ^n$.
Here $\mathbb{A} ^1_t$ and $\mathbb{A} ^1_s$ denote the affine line with local coordinates $t$ and $s$ respectively. Then we have
$$\bar\Gamma\times_{\mathbb{A} ^1_t}\{0\}/(pr_V^*T_V|_{\bar\Gamma\times_{\mathbb{A} ^1_t}\{0\}}) = D.$$
Now we can prove the claim. Since $D\subset (\bar\Gamma-\Gamma)/(pr_V^*T_V|_{\bar\Gamma-\Gamma})$, we may choose a smooth pointed curve $(\Delta,0)$ with local coordinate $\delta$ and a morphism $\rho:\Delta\to \bar\Gamma$ such that $\rho(\Delta-\{0\})\subset \Gamma$ and $\rho(0)$ represents $a\in A$. Let $t_\Delta:\Delta\mapright{\rho} \bar\Gamma\mapright{pr_t} \mathbb{A} ^1_t$ and $s_\Delta:\Delta\mapright{\rho} \bar\Gamma\mapright{pr_s} \mathbb{A} ^1_s$ denote the compositions of $\rho$ and the projections to $\mathbb{A} ^1_t$ and $\mathbb{A} ^1_s$ respectively. Let $\rho_V:\Delta\to V$ denote the composition of $\rho$ with the projection $pr_V:\bar\Gamma\to V$ and let $v_0=\rho_V(0)\in X$. Then
$\rho(0)=(v_0,0,0,v_1,v_2),$
\begin{equation}\label{00}
v_1=\lim_{\delta\to 0} (t_\Delta s_\Delta)^{-1}\cdot \gamma\circ g\circ \rho_V\in \mathbb{A} ^m, \quad
v_2= \lim_{\delta\to 0} t_\Delta^{-1}\cdot \eta\circ \rho_V \in \mathbb{A} ^n.
\end{equation}
Since ${\mathscr O}_Y^{\oplus m}\twoheadrightarrow Ob_Y\mapright{\sigma} {\mathscr O}_Y$ is surjective, by copying the proofs of Lemma 4.4 and Corollary 4.5 in \cite{KiemLi}, we find that
$v_1$ represents a point in $$\mathrm{ker}\left(
f^*h^1/h^0(E_Y^\vee)\mapright{f^*\sigma} {\mathscr O}_X \right)$$
and $v_2$ a point in $h^1/h^0(E_{X/Y}^\vee)$.
Therefore $a$ lies in \eqref{1501251}. This proves the theorem.
\end{proof}
\medskip
\def{\mathscr D}{{\mathscr D}}
\def\tilde{\sigma}{\tilde{\sigma}}
\subsection{Cosection localized virtual pullback}\label{ss2.2}
In \S\ref{ss2.1}, we considered the virtual pullback of a cosection localized virtual fundamental class
when there is a cosection $\sigma:Ob_Y\to {\mathscr O}_Y$ on $Y$ that induces a
cosection $Ob_X\to f^*Ob_Y\mapright{\sigma} {\mathscr O}_X$ on $X$. Actually there is
another way to combine virtual pullback with cosection localization.
Consider the case when there is a cosection
$\sigma:Ob_X\to {\mathscr O}_X$ that induces a cosection $\tilde{\sigma}:Ob_{X/Y}\to Ob_X\to {\mathscr O}_X$ of the relative obstruction sheaf.
In this subsection, we
define the cosection localized virtual pullback
$$f_\sigma^!:A_*(Y)\longrightarrow A_*(X(\tilde{\sigma}))$$
for a virtually smooth morphism $f:X\to Y$ where $X(\tilde{\sigma})=\tilde{\sigma}^{-1}(0)$ (cf. Definition \ref{1clvp})
and prove the cosection localized virtual pullback formula (cf. Theorem \ref{1thmain2}).
\medskip
We let
$f: X\to Y$ be a virtually smooth morphism between Deligne-Mumford stacks as before; we let $\sigma=\sigma_X: {\mathcal Ob}_X\to{\mathscr O}_X$
be a cosection, and form the (composite)
$$\tilde{\sigma}=\sigma_{X/Y}: {\mathcal Ob}_{X/Y}=h^1(E_{X/Y}^{\vee})\longrightarrow h^1(E_X^{\vee})\mapright{\sigma}{\mathscr O}_X.
$$
Let $X(\sigma)=\sigma^{-1}(0)$ and $X(\tilde{\sigma})=\tilde{\sigma}^{-1}(0)$. Then by definition, we have an inclusion
$$\jmath:X(\sigma)\hookrightarrow X(\tilde{\sigma}).$$
We let
${\cal K}=h^1/h^0(E_{X/Y}^{\vee})$. Then $\tilde{\sigma}$ induces a morphism ${\cal K}\to {\mathscr O}_X$ which we also denote by $\tilde{\sigma}$ by abuse of notation. As before, we denote
$${\cal K}(\tilde{\sigma})=\mathrm{ker} [\tilde{\sigma}:{\cal K} \to{\mathscr O}_X]{\,:=\,} {\cal K}|_{X(\tilde{\sigma})}\cup \mathrm{ker} [
\tilde{\sigma}|_U:{\cal K}|_{U}\to {\mathscr O}_{U}],
$$
where $U=X-X(\tilde{\sigma})$ is the open where $\tilde{\sigma}$ is surjective.
\begin{lemm}\label{deg-vanishing}
We have
$$\mathrm{Supp} \,\mathfrak{C}_{{X}/ {Y}} \sub {\cal K}(\tilde{\sigma}).
$$
\end{lemm}
\begin{proof}
We apply the functoriality of the $h^1/h^0$ construction to \eqref{16}
to obtain the commutative diagram
\begin{equation}\label{h1h0}
\begin{CD}
\mathfrak{C}_{{X}/ {Y}}@>\subset>>h^1/h^0(\LL\dual_{{X}/ {Y}})@>\subset>>h^1/h^0(E^{\vee}_{{X}/ {Y}})={\cal K} @>\tilde{\sigma}>>{\mathscr O}_X\\
@VVV @VVV@VVV @V=VV\\
\mathfrak{C}_{{X}{}}@>\subset>>h^1/h^0(\LL\dual_{{X}}) @>\subset>> h^1/h^0(E^{\vee}_{{X}})=:{\cal E} @>\sigma>> {\mathscr O}_X\\
\end{CD}
\end{equation}
Like before, we have
$\mathrm{Supp} \,\mathfrak{C}_{{X}} \sub \mathrm{ker} [h^1/h^0(E_{X}^{\vee})\longrightarrow {\mathscr O}_X]=:{\cal E}(\sigma)$.
Since $\tilde{\sigma}$ is induced from $\sigma=\sigma_X$, the lemma follows.
\end{proof}
We now define the cosection localized virtual pullback. We let
$Y'\to Y$ be a morphism of stacks where $Y'$ has stratification by global quotients. Form the Cartesian product
\[\xymatrix{
X'\ar[r]^p \ar[d]_{f'} & X\ar[d]^f\\
Y'\ar[r]^q& Y.
}\]
and let $\hat\sigma: p^\ast{\cal K}\to {\mathscr O}_{X'}$
be the pullback of $\tilde{\sigma}:{\cal K}\to {\mathscr O}_X$. Then the vanishing locus of $\hat\sigma$ is $$X'(\ti\sigma):=X(\tilde{\sigma})\times_X X'$$
and $$p^*{\cal K}(\hat\sigma)=\mathrm{ker} [\hat\sigma:p^*{\cal K}\to {\mathscr O}_{X'}]={\cal K}(\tilde{\sigma})\times_XX'.$$
Consider the composite
$$\iota: (\mathfrak{C}_{X'/Y'})_{\mathrm{red}} \sub (\mathfrak{C}_{{X}/ {Y}}\times_{ {X}} {X'})_{\mathrm{red}} \sub {\cal K}(\tilde{\sigma})\times_{ {X}} {X'}
= p^*{\cal K}(\hat\sigma),
$$
where the first inclusion follows from the definition of $X'$, and the second inclusion follows from Lemma
\ref{deg-vanishing}.
\begin{defi}\label{1clvp} The \emph{cosection localized virtual pullback} is defined by
$$f_\sigma^! : A_\ast Y'\mapright{\epsilon}A_\ast \mathfrak{C}_{X'/Y'}\mapright{\iota_\ast}
A_\ast(p^\ast{\cal K}(\hat\sigma))\mapright{0^!_{p^*{\cal K},_{\mathrm{loc}} }} A_{\ast}(X'(\tilde{\sigma})),
$
where $\epsilon$ is defined on the level of cycles by $\epsilon(\sum n_i[V_i]) =\sum n_i[\mathfrak{C}_{V_i\times_{Y'}X'/V_i}]$.
\end{defi}
Note that the way that \cite[Theorem 2.31]{Mano1} was applied to \cite[Construction 3.6]{Mano1} can also be
applied here to conclude that $\epsilon$ descends to maps between Chow groups.
We have the following virtual pullback formula.
\begin{theo}\label{1thmain2} Let $f:X\to Y$ be a virtually smooth morphism, $\sigma:{\mathcal Ob}_X\to{\mathscr O}_X$ be a cosection and
$\tilde{\sigma}:{\mathcal Ob}_{X/Y}\to Ob_X\mapright{\sigma} {\mathscr O}_X$ be the induced cosection.
Let $\jmath: X(\sigma)\to X(\tilde{\sigma})$ be the inclusion of zero loci of $\sigma$ and $\tilde{\sigma}$. Then we have
$$f_\sigma^![Y]^{\mathrm{vir}} =\jmath_\ast [X]\virt_\loc\in A_\ast( X(\tilde{\sigma})).
$$
\end{theo}
The proof is completely parallel to that of Theorem \ref{1thmain1}, so we only provide a sketch.
We need the following analogue of Lemma \ref{110}.
\begin{lemm}\label{1100}
Let $f:X\to Y$ be a morphism of Deligne-Mumford stacks and ${{\cal K}}$ be a vector bundle stack on $X$ such that $\mathfrak{C}_{X/Y}\subset {{\cal K}}$. Let ${\cal F}$ be a vector bundle stack on $Y$ with the zero section $0_{\cal F}:Y\to {\cal F}$.
Let $U\subset X$ be open and $\tilde{\sigma}:{\cal K}|_U\to \mathbb{C}_U$ be a surjective map of vector bundle stacks. Let $X(\tilde{\sigma})=X-U$.
Then
for each irreducible $B\subset {\cal F}$,
\[ f_\sigma^!0^!_{{\cal F}}[B]=0_{f^*{\cal F}\oplus {{\cal K}},_{\mathrm{loc}} }^![\mathfrak{C}_{X\times_{\cal F} B/B}] \quad\text{in } A_*(X(\tilde{\sigma}))
\]
where $0_{f^*{\cal F}\oplus {{\cal K}},_{\mathrm{loc}} }^!$ denotes the localized Gysin map with respect to the cosection
$(0,\tilde{\sigma}):f^*{\cal F}\oplus {{\cal K}}|_{f^{-1}(U)}\to \mathbb{C}_{f^{-1}(U)}$.
\end{lemm}
\begin{proof}
We may assume that there is an irreducible $\tilde{B}\subset Y$ such that $B={\cal F}|_{\tilde{B}}={\cal F}\times_Y\tilde{B}$ and $0^!_{{\cal F}}[B]=\tilde{B}$. The left side is $$f_\sigma^!0^!_{{\cal F}}[B]=f_\sigma^![\ti B]=0^!_{{\cal K},_{\mathrm{loc}} } [\mathfrak{C}_{\ti B\times_YX/\ti B}]=0^!_{f^*{\cal F}\oplus {\cal K},_{\mathrm{loc}} } [f^*{\cal F}\times_X \mathfrak{C}_{\ti B\times_YX/\ti B}].$$
Since $B={\cal F}\times_Y\tilde{B}$, $f^*{\cal F}\times_X \mathfrak{C}_{\ti B\times_YX/\ti B}=\mathfrak{C}_{B\times_{\cal F} X/B}$. Hence
the lemma follows.
\end{proof}
Theorem \ref{1thmain2} follows from the following.
\begin{prop}\label{vanishing2}
Suppose $Y'=h^1/h^0(E_Y^{\vee})={\cal F}$ so that $p|_{X'(\tilde{\sigma})}: X'(\tilde{\sigma})\to X(\tilde{\sigma})$ is a bundle stack
and that we have the Gysin map $0_{{\cal F}}^! :A_\ast(X'(\tilde{\sigma}))\longrightarrow A_{\ast} (X(\tilde{\sigma})).$
Then we have
$$0_{\cal F}^!f_\sigma^![\mathfrak{C}_{Y}]=\jmath_\ast [X]\virt_\loc\in A_\ast( X(\tilde{\sigma})).
$$
\end{prop}
\begin{proof}[Proof of Theorem \ref{1thmain2}] By Lemma \ref{1100} and Proposition \ref{vanishing2},
$$f_\sigma^![Y]^{\mathrm{vir}} =f_\sigma^!0_{\cal F}^![\mathfrak{C}_Y]=
0_{f^*{\cal F}\oplus {{\cal K}},_{\mathrm{loc}} }^![\mathfrak{C}_{X\times_{\cal F} \mathfrak{C}_Y/\mathfrak{C}_Y}]
=0_{\cal F}^!f_\sigma^![\mathfrak{C}_Y]=\jmath_\ast [X]\virt_\loc.$$
\end{proof}
\begin{proof}[Proof of Proposition \ref{vanishing2}]
The proof is almost identical to that of Theorem \ref{1thmain1}, so we only point out the difference. The construction of the double deformation space and the cone $c(u)$ is identical and we have a commutative diagram
\small $$\xymatrix{
h^1/h^0(c(u)^\vee)\ar[r]\ar[d] & p^*f^*h^1/h^0(E_Y^\vee)\oplus p^*h^1/h^0(E_X^\vee) \ar[r]\ar[d]^{(0,\sigma)} & p^*f^*h^1/h^0(E_Y^\vee)\otimes q^*{\mathscr O}_{\mathbb{P}^1}(1)\\
q^*{\mathscr O}_{\mathbb{P}^1}\ar[r]^= & q^*{\mathscr O}_{\mathbb{P}^1} &
}$$\normalsize
where the vertical arrows are defined by $\sigma$.
Again it suffices to show
\begin{equation}\label{claim}
\mathrm{Supp}\, \mathfrak{C}_{X\times\PP^1/M_Y^\circ}\sub \mathrm{ker} [h^1/h^0(c(u)^{\vee})\to q^\ast{\mathscr O}_{\PP^1}].
\end{equation}
Now the proof continues exactly the same as the proof of Theorem \ref{1thmain1}.
By the cone reduction in \cite{KiemLi}, \eqref{claim} holds over the open $\mathbb{P}^1-\{(0:1)\}$.
To prove \eqref{claim} over the point $(0:1)$, we consider a general point $a$ in any irreducible component $A$ of $D$
lying over $(0:1)$ and use the local construction of the double deformation space.
After choosing a morphism $\rho$ from a smooth pointed curve $(\Delta,0)$ with $\rho(0)$ representing $a$,
one finds that we only have to check that $v_2$ represents a point in
$$\mathrm{ker} \left(\tilde{\sigma}:h^1/h^0(E^\vee_{X/Y})\to h^1/h^0(E^\vee_X)\mapright{\sigma} {\mathscr O}_X\right).
$$
Again this follows from the arguments in the proofs of Lemma 4.4 and Corollary 4.5 in \cite{KiemLi}:
When $v_2\ne 0$, because $\lim_{\delta\to 0}
s_\Delta=0$, the image $v'_2$ of $v_2$ in $h^1/h^0( E_{X}^{\vee})$ under the tautological
$$h^1/h^0( E_{X/Y}^{\vee})\longrightarrow h^1/h^0( E_{X}^{\vee})
$$
lies in $\mathfrak{C}_X$. Using the cosection $\sigma$, we see that $v'_2\in \mathrm{ker} [h^1/h^0( E_{X}^{\vee})\to {\mathscr O}_X]$.
Because $\tilde\sigma$ is induced by $\sigma$, we obtain \eqref{claim}.
This proves the proposition.
\end{proof}
\begin{rema}
When $f:X\to Y$ is a morphism over a smooth Artin stack
$\mathcal{S} $, sometimes it is more convenient to work with relative obstruction theories, say with $\mathbb{L} _X$ replaced by
$\mathbb{L} _{X/\mathcal{S} }$, $\mathbb{L} _Y$ by $\mathbb{L} _{Y/\mathcal{S} }$ etc.
It is straightforward to see that all the statements and proofs in this section hold in this case.
Another useful situation is when $Y$ is only assumed to be an Artin stack with $Y\to\mathcal{S} $ assumed to be Deligne-Mumford.
Then Proposition 2.11 holds in this case with obstruction theories replaced by relative (to $\mathcal{S} $) obstruction theories.
\end{rema}
\section{Torus localization for cosection localized virtual cycles}\label{S:Torus}
In this section, we prove the torus localization formula for cosection localized virtual cycles (Theorem \ref{2thmain}). We do not assume the existence of an equivariant global embedding or a global resolution of the perfect obstruction theory. When the cosection is trivial, our argument gives a new proof of the torus localization theorem in \cite{GrPa} without these assumptions.
Let $X$ be a Deligne-Mumford stack acted on by a torus $T=\mathbb{C}^*$. Let $F$ denote the $T$-fixed locus, i.e. locally if $X=\mathrm{Spec}(A)$, then $F=\mathrm{Spec}(A/\langle A^{\text{mv}}\rangle)$ where $A^{\mathrm{mv}}$ denotes the ideal generated by $T$-eigenfunctions with nontrivial characters.
Let $$\imath:F\longrightarrow X$$ denote the inclusion map.
Let $D([X/T])$ be the derived category of sheaves of $T$-equivariant ${\mathscr O}_X$-modules on $X$. It is the same as the ordinary derived category of sheaves of ${\mathscr O}_X$-modules except that all sheaves are $T$-equivariant and all morphisms or arrows are $T$-equivariant. The action of $T$ on $X$ gives the equivariant cotangent complex $\mathbb{L} _X\in D([X/T])$.
\begin{defi} A \emph{$T$-equivariant perfect obstruction theory} consists of an object $E\in D([X/T])$ and a morphism $$\phi:E \to \mathbb{L} _X$$
in $D([X/T])$ which is a perfect obstruction theory on $X$.
\end{defi}
If $A$ is a $T$-equivariant sheaf of ${\mathscr O}_F$-modules on $F$, we let $A^{\mathrm{fix}}$ denote the sheaf of $T$-fixed submodules and $A^{\mathrm{mv}}$ denote the subsheaf generated by $T$-eigensections with nontrivial characters.
Given $E\in D([X/T])$, $\bar E:=E|_F$ is a complex of $T$-equivariant sheaves on $F$,
thus we can decompose $\bar E=\bar E^{\mathrm{fix}}\oplus \bar E^{\mathrm{mv}}$ into the fixed and moving parts.
A $T$-equivariant chain map $\psi:\bar E\to \tilde E$ to an $\tilde E\in D([F/T])$ preserves such decompositions to give us
$\psi^{\mathrm{fix}}:\bar E^{\mathrm{fix}}\to \tilde E^{\mathrm{fix}}$ and
$\psi^{\mathrm{mv}}:\bar E^{\mathrm{mv}}\to \tilde E^{\mathrm{mv}}$. If $\psi$ is a quasi-isomorphism, so are $\psi^{\mathrm{fix}}$ and $\psi^{\mathrm{mv}}$. Therefore a $T$-equivariant perfect obstruction theory $\phi:E\to \mathbb{L} _X$ induces morphisms in $D([F/T])$
$$\phi^{\mathrm{fix}}:E|_F^{\mathrm{fix}}\longrightarrow \mathbb{L} _X|_F^{\mathrm{fix}}\quad{\rm and}\quad \phi^{\mathrm{mv}}:E|_F^{\mathrm{mv}}\longrightarrow \mathbb{L} _X|_F^{\mathrm{mv}}.$$
\begin{lemm} Let the notation be as above.
The composition $\phi_F:E|_F^{\mathrm{fix}}\to \mathbb{L} _X|_F^{\mathrm{fix}}\to \mathbb{L} _F$ of $\phi^{\mathrm{fix}}$ and the natural morphism $\mathbb{L} _X|_F^{\mathrm{fix}}\to \mathbb{L} _F$ is a perfect obstruction theory of $F$.
\end{lemm}
\begin{proof}
For any square zero extension $\Delta\to \bar\Delta$ of $k$-schemes with ideal sheaf $J$ and a morphism
$g:\Delta\to F$, let $\omega(g)\in Ext^1(g^*\mathbb{L} _F, J)$ denote the composition
$g^*\mathbb{L} _F\to \mathbb{L} _\Delta\to J[1]$ of the natural morphisms $g^*\mathbb{L} _F\to \mathbb{L} _\Delta$ from $g$ and $\mathbb{L} _\Delta\to \mathbb{L} _{\Delta/\bar\Delta}\to \mathbb{L} ^{\ge -1}_{\Delta/\bar\Delta}=J[1]$ from
$\Delta\to \bar\Delta$. Let
$$\phi_F^*\omega(g)\in \Ext^1(g^*E|_F^{\mathrm{fix}},J)
$$
be the image of $\omega(g)$ by the map $\Ext^1(g^*\mathbb{L} _F,J)\to \Ext^1(g^*E|_F^{\mathrm{fix}},J)$ induced from $\phi_F: E|_F^{\mathrm{fix}}\to \mathbb{L} _F$. Note that $\Ext^1(g^*E|_F^{\mathrm{fix}},J)$ is a $T$-module and
$\phi_F^*\omega(g)$ is $T$-invariant, where $T$ acts on $\Delta$, $\bar\Delta$ and $J$ trivially.
By \cite[Theorem 4.5]{BeFa}, it suffices to show the following claim: the obstruction assignment
$\phi_F^*(\omega(g))$ vanishes if and only if an extension $\bar g:\bar\Delta\to F$ of $g$ exists;
and if $\phi_F^*(\omega(g))=0$, then the extensions form a torsor under $\Ext^0(g^*E|_F^{\mathrm{fix}},J)$.
Let $h:\Delta\to X$ be the composite of $g$ with the inclusion $F\sub X$.
Since $\phi:E\to \mathbb{L} _X$ is a perfect obstruction theory, $h$ extends to $\bar h:\bar\Delta\to X$ if
and only if $0=\phi_X^*\omega(h)\in \Ext^1(h^*E,J)$. Because $h$ factors through $F\sub X$ and
$J$ is an ${\mathscr O}_{\Delta}$-module,
$$\Ext^1(h^*E,J)=\Ext^1(g^*E|_F^{\mathrm{fix}},J)\oplus \Ext^1(g^*E|_F^{\mathrm{mv}},J),
$$
as $T$-module, and further $\phi_X^*\omega(h)$ is $T$-invariant. Since $\phi_X^*\omega(h)^T=
\phi_F^*\omega(g)$, we see that $\phi_F^*\omega(g)=0$ if and only if $h$ extends to $\bar h: \bar\Delta\to X$.
Because $T$ is reductive, a standard argument shows that we can find a $T$-invariant extension $\bar h$,
which necessarily factors through $F\sub X$. This proves that $\phi_F^*\omega(g)$ is an obstruction class to
extending $g$ to $\bar g:\bar \Delta\to F$.
The part on the space of extensions $\bar g$ follows by the same argument.
\end{proof}
We let $E_F:=E|_F^{\mathrm{fix}}$ and $N^{\mathrm{vir}} :=(E|_F^{\mathrm{mv}})^\vee$.
Since $E$ is perfect, both the fixed part $E_F$ and the moving part $E|_F^{\mathrm{mv}}=(N^{\mathrm{vir}} )^\vee$ of $E|_F$ are perfect. They fit into the following diagram of distinguished triangles:
\[\xymatrix{
E|_F\ar[r] \ar[d] & E_F\ar[r]\ar[d] & (N^{\mathrm{vir}} )^\vee[1]\ar[r]\ar[d] &\\
\mathbb{L} _X|_F\ar[r] & \mathbb{L} _F\ar[r] & \mathbb{L} _{F/X}\ar[r] &
}\]
The morphism $E|_F\to E_F$ induces a homomorphism $$Ob_F=H^1(E_F^\vee)\longrightarrow H^1(E|_F^\vee)\cong H^1(E^\vee)|_F=Ob_X|_F.$$
Let $\sigma:Ob_X=H^1(E^\vee)\to {\mathscr O}_X$ be a $T$-equivariant cosection. Then $\sigma$ induces a $T$-invariant cosection
$$\sigma_F:Ob_F\longrightarrow Ob_X|_F\longrightarrow {\mathscr O}_X|_F={\mathscr O}_F$$
and we have a cosection localized virtual cycle $[F]\virt_\loc$.
\def\mathfrak{N} {\mathfrak{N} }
\begin{defi} Suppose the virtual normal bundle $N^{\mathrm{vir}} $ admits a global resolution $[N_0\to N_1]$
by locally free sheaves $N_0$ and $N_1$ over $F$.
We define the Euler class $e(N^{\mathrm{vir}} )$ of $N^{\mathrm{vir}} $ to be
$$e(N^{\mathrm{vir}} )=e(N_0)/e(N_1) \in A^*(F)\otimes_\mathbb{Q}\QQ[t,t^{-1}].
$$
\end{defi}
The goal of this section is to prove the following.
\begin{theo}\label{2thmain}
Let $X$ be a Deligne-Mumford stack acted on by $T$ and let $E\to \mathbb{L} _X$ be an equivariant
perfect obstruction theory on $X$. Let $F$ be the $T$-fixed locus in $X$. Let $\sigma:Ob_X\to {\mathscr O}_X$ be a
$T$-equivariant cosection on $X$. Suppose there is a global resolution $N^{\mathrm{vir}} \cong[N_0\to N_1]$, where
$N_0$ and $N_1$ are locally free sheaves on $F$ (whose ranks may vary from component to component).
Then we have
$$ [X]^{\mathrm{vir}} __{\mathrm{loc}} = \imath_* \frac{[F]\virt_\loc}{e(N^{\mathrm{vir}} )}\in A_\ast^T X\otimes_{\mathbb{Q}[t]}\mathbb{Q}[t,t^{-1}].
$$
Here the class $[F]\virt_\loc$ is defined with respect to the induced perfect obstruction theory $E_F$ and cosection $\sigma_F$.
\end{theo}
\begin{rema}
In \cite{GrPa}, the localization formula in Theorem \ref{2thmain} was proved for the ordinary virtual fundamental class under the following assumptions:
\begin{enumerate}
\item $X$ admits a global equivariant embedding into a smooth $Y$;
\item the perfect obstruction theory $E$ admits an equivariant global locally free resolution.
\end{enumerate}
Both conditions are nontrivial unless $X$ is a projective scheme. Recent development in moduli theory and enumerative geometry utilizes a plethora of moduli stacks for which (1) is often tedious to verify and hence it is
desirable to give a proof without the assumption (1). Here we remove the first assumption entirely and weaken
the second assumption to
\begin{enumerate}
\item[($2'$)] the virtual normal bundle $N^{\mathrm{vir}} $ admits a global locally free resolution $[N_0\to N_1]$ on the fixed locus $F$,
\end{enumerate}
which is often easier to check.
When $\sigma=0$, Theorem \ref{2thmain} says that the torus localization in \cite{GrPa} works without the assumption $(1)$ and with a much weaker $(2')$.\end{rema}
By our assumption that there is a resolution $[N_0\to N_1]$ of $N^{\mathrm{vir}} $, we find that the normal sheaf $N_{F/X}$ is contained in $h^1/h^0(N^{\mathrm{vir}} [-1])=\mathrm{ker} \{ N_0\to N_1\}$, thus contained in $N_0$.
Hence the normal cone $\mathfrak{C}_{F/X}$ is contained in $N_0$ as well. As in Definition \ref{14}, we define the virtual pullback $$\imath^!:A_*(X(\sigma))\to A_*(F(\sigma))$$ for the inclusion $\imath:F\to X$, by
$$[B]\longmapsto [\mathfrak{C}_{B\times_XF/B}]\longmapsto 0^!_{N_0} [\mathfrak{C}_{B\times_XF/B}].$$
The proof of Theorem \ref{2thmain} is attained through the following two lemmas.
\begin{lemm}\label{le2}
Let $X(\sigma)$ and $F(\sigma)$ denote the vanishing loci of $\sigma$ and $\sigma_F$ respectively. Then $F(\sigma)=X(\sigma)\cap F$ and $\imath^![X]\virt_\loc=[F]\virt_\loc\cap e(N_1)$.
\end{lemm}
\begin{proof} The first identity follows from Lemma \ref{13}. We prove the second identity.
By definitions, $[X]\virt_\loc=0^!_{\mathfrak{E},_{\mathrm{loc}} }[\mathfrak{C}_X]$ and $[F]\virt_\loc=0^!_{\mathfrak{E}_F,_{\mathrm{loc}} }[\mathfrak{C}_F]$, where
$\mathfrak{E}=h^1/h^0(E^\vee)$ and $\mathfrak{E}_F=h^1/h^0(E_F^\vee)$. By Lemma \ref{110}, we have
$$\imath^!0^!_{\mathfrak{E},_{\mathrm{loc}} }[\mathfrak{C}_X]=0^!_{\mathfrak{E}|_F\oplus N_0,_{\mathrm{loc}} }[\mathfrak{C}_F]$$
because $[\mathfrak{C}_{F\times_\mathfrak{E}\mathfrak{C}_X/\mathfrak{C}_X}]=[\mathfrak{C}_{F/\mathfrak{C}_X}]=[\mathfrak{C}_F]$ by Vistoli's rational equivalence \cite{Vistoli}.
The proof of Theorem \ref{1thmain1} guarantees that the rational equivalence lives in the desired locus for the localized Gysin maps.
Therefore
$$\imath^![X]\virt_\loc=\imath^!0^!_{\mathfrak{E},_{\mathrm{loc}} }[\mathfrak{C}_X]=0^!_{\mathfrak{E}|_F\oplus N_0,_{\mathrm{loc}} }[\mathfrak{C}_F]
=0^!_{\mathfrak{E}_F\oplus N_1,_{\mathrm{loc}} }[\mathfrak{C}_F]=[F]\virt_\loc\cap e(N_1),
$$
because $\mathfrak{E}|_F=\mathfrak{E}_F\oplus [N_1/N_0]$.
\end{proof}
\begin{lemm}\label{le3}
$\imath^!\imath_*\alpha =\alpha \cap e(N_0)$ for $\alpha\in A_*(F(\sigma))\otimes_\mathbb{Q}\QQ[t,t^{-1}]$.
\end{lemm}
\begin{proof}
If $B$ is a cycle in $F(\sigma)$, the normal cone of $B\cap F(\sigma)$ in $B$ is $B$ and $\imath^!\imath_*B=0^!_{N_0}B=B\cap e(N_0)$ by the definition of virtual pullback $\imath^!$.
\end{proof}
Now we can prove Theorem \ref{2thmain}.
\begin{proof}[Proof of Theorem \ref{2thmain}]
By \cite[Theorem 6.3.5]{Kres},
$$\imath_*:A^T_*(F(\sigma))\otimes_{\mathbb{Q}[t]}\mathbb{Q}[t,t^{-1}] \longrightarrow A^T_*(X(\sigma))\otimes_{\mathbb{Q}[t]}\mathbb{Q}[t,t^{-1}]
$$
is an isomorphism. Thus
$\imath_*\alpha=[X]\virt_\loc\quad$ for some
$$\alpha\in A^T_*(F(\sigma))\otimes_{\mathbb{Q}[t]}\mathbb{Q}[t,t^{-1}]= A_*(F(\sigma))\otimes_\mathbb{Q} \mathbb{Q}[t,t^{-1}].
$$ By Lemmas \ref{le2} and \ref{le3},
$$[F]\virt_\loc \cap e(N_1)= \imath^![X]\virt_\loc =\imath^!\imath_*\alpha=\alpha\cap e(N_0).$$
Hence $$\alpha=\frac{[F]\virt_\loc}{e(N^{\mathrm{vir}} )}\quad{\rm and}\quad [X]\virt_\loc=\imath_*\alpha=\imath_*\frac{[F]\virt_\loc}{e(N^{\mathrm{vir}} )}$$ as desired.
\end{proof}
\begin{exam} Let $V=\mathbb{C}^d$ be a vector space with $(z_1,\cdots, z_d)$ be its standard
coordinates. We let $T=\mathbb{C}^*$ acts on $V$ via $(z_i)^\alpha=(\alpha z_i)$.
The global differentials $dz_i$ give a trivialization of $\Omega_V$, in the form $\Omega_V\cong
V\times V^\ast$. Let $E=[T_V\mapright{0} \Omega_V]$ and
$\sigma:\Omega_V=V\times V^*\to {\mathscr O}_V$ be the tautological pairing.
Then $\sigma^{-1}(0)=\{O\}\subset V$ is the (reduced) origin $O\in V$, and under the induced $T$-action
on $\Omega_V$, $\sigma$ is $T$-invariant. We observe
$$F=V^{\mathbb{C}^*}=\{O\}, \quad E|_F=[V\mapright{0} V^*]=N^{\mathrm{vir}} ,\quad{\rm and}\quad E_F=[0\to 0].$$
Hence $e(N^{\mathrm{vir}} )=(-1)^d$ and $[F]\virt_\loc$ is the zero cycle $[O]$ consisting of one simple point $O$. Then by
Theorem \ref{2thmain}, we have
$$[V]\virt_\loc=\frac{[F]\virt_\loc}{e(N^{\mathrm{vir}} )}=(-1)^d [O]$$
as expected from \cite[Example 2.4]{KiemLi}.
\end{exam}
In this example, if instead we consider a cosection $\sigma': \Omega_V\to{\mathscr O}_V$
via $dz_1\mapsto 1$ and $dz_{i>1}\mapsto 0$. Since $\sigma'$ is surjective, we obtain
$[V]\virt_\loc=0$. However, $\frac{[F]\virt_\loc}{e(N^{\mathrm{vir}} )}=(-1)^d [O]$ as before.
Hence Theorem \ref{2thmain} does not apply.
\medskip
\section{Wall crossing formulas for cosection localized virtual cycles}\label{S:Wall}
In this section we provide a wall crossing formula for simple $\mathbb{C}^*$-wall crossings. The construction and proof are rather standard (cf. \cite{KL3}).
Let $X$ be a Deligne-Mumford stack acted on by $T=\mathbb{C}^*$. Let $\phi:E\to \mathbb{L} _X$ be a $T$-equivariant perfect obstruction theory, together with an equivariant cosection $\sigma:Ob_X=h^1(E^\vee)\to {\mathscr O}_X$.
Let \begin{enumerate}
\item $F$ be the $T$-fixed locus in $X$;
\item $X^s$ be the open substack of $x\in X$ so that the orbit $T\cdot x$ is 1-dimensional and closed in $X$;
\item $\Sigma^0_\pm=\{x\in X-(X^s\cup F)\,|\, \mathrm{lim}_{t\to 0}t^{\pm 1}\cdot x\in F\}$;
\item $\Sigma_\pm = \Sigma^0_\pm \cup F$;
\item $X_\pm = X-\Sigma_\mp \subset X$;
\item $M_\pm=[X_\pm/T]\subset M=[X/T]$ are separated Deligne-Mumford stacks.
\end{enumerate}
Recall from \S\ref{S:Torus} that we have the induced cosections $\sigma_F:Ob_F\to {\mathscr O}_F$.
We then define the master space of the wall crossing $M_\pm$ to be
$$\mathfrak{M}=\left[ X\times \mathbb{P}^1 - \Sigma_-\times\{0\} -\Sigma_+\times\{\infty\}/\mathbb{C}^*\right]$$
where $\mathbb{C}^*$ acts trivially on $X$ and by $\lambda\cdot(a:b)=(a:\lambda b)$ on $\mathbb{P}^1$. The action of $T$ on $X$ induces an action of $T$ on $\mathfrak{M}$ whose fixed locus is
$$M_+\sqcup F\sqcup M_-$$
as is easy to check.
Since $\mathbb{C}^*$ acts only on the component $\mathbb{P}^1$, the pullback of any sheaf on $X$ by the projection $X\times \mathbb{P}^1\to X$ is $\mathbb{C}^*$-equivariant and hence descends to the free quotient $\mathfrak{M}$. By pulling back the perfect obstruction theory $\phi:E\to\mathbb{L} _X$ and descending, we obtain a morphism $\bar\phi:\bar E\to \mathbb{L} _\mathfrak{M}$.
\begin{lemm}
The morphism $\bar\phi:\bar E\to \mathbb{L} _\mathfrak{M}$ is a $T$-equivariant perfect obstruction theory of $\mathfrak{M}$. Moreover the pullback of $\sigma$ descends to a $T$-equivariant cosection $\bar\sigma:Ob_\mathfrak{M}\to {\mathscr O}_\mathfrak{M}$.
\end{lemm}
\begin{proof} This is straightforward and we omit the proof.\end{proof}
The cosection $\bar\sigma$ induces cosections on the fixed locus $M_\pm$ and $F$ in $\mathfrak{M}$.
We are ready to state the main result of this section.
\begin{theo}\label{3thmain}
Let the notation be as above.
Suppose the virtual normal bundle $N^{\mathrm{vir}} $ admits a resolution $[N_0\to N_1]$ by vector bundles on $F$.
Then we have
$$[M_+]\virt_\loc- [M_-]\virt_\loc= \mathrm{res}_{t=0} \frac{[F]\virt_\loc}{e(N^{\mathrm{vir}} )}\quad\text{in } A^T_*(\mathfrak{M})\otimes_{\mathbb{Q}[t]}\mathbb{Q}[t,t^{-1}].$$
\end{theo}
\begin{proof}
Applying Theorem \ref{2thmain} to the master space $\mathfrak{M}$, we find that
$$[\mathfrak{M}]\virt_\loc = \frac{[M_+]\virt_\loc}{-t} +\frac{[M_-]\virt_\loc}{t} + \frac{[F]\virt_\loc}{e(N^{\mathrm{vir}} )}$$
since the normal bundle of $M_+$ is trivial with weight $1$ while that of $M_-$ is trivial with weight $-1$ by construction.
If we take $\mathrm{res}_{t=0}$, the left side vanishes because $[\mathfrak{M}]\virt_\loc\in A_*^T(\mathfrak{M})$ has trivial principal part. Therefore the residue of the right side
$$-[M_+]\virt_\loc +[M_-]\virt_\loc + \mathrm{res}_{t=0} \frac{[F]\virt_\loc}{e(N^{\mathrm{vir}} )}$$
vanishes. This proves the theorem.
\end{proof}
\bibliographystyle{amsplain}
|
\section{Introduction}
The $Y(4260)$ state was first observed by BaBar collaboration in the
$e^+e^-$ annihilation through initial state radiation \cite{babar1}, and it
was confirmed by CLEO and Belle collaborations \cite{yexp}. The $Y(4260)$ was
also observed in the $B^-\to Y(4260)K^-\to J/\Psi\pi^+\pi^-K^-$ decay
\cite{babary2}, and CLEO reported two additional decay channels:
$J/\Psi\pi^0\pi^0$ and $J/\Psi K^+K^-$ \cite{yexp}.
The $Y(4260)$ is one of the many charmonium-like state, called $X,~Y$ and
$Z$ states, recently observed in $e^+e^-$ collisions
by BaBar and Belle collaborations that do not fit the quarkonia interpretation.
The production mechanism, masses, decay widths, spin-parity
assignments and decay modes of these states have been discussed in some
reviews \cite{Zhu:2007wz,Nielsen:2009uh,Brambilla:2010cs,Nielsen:2014mva}.
The $Y(4260)$ is particularly interesting because some new states have
been identified in the decay channels of the $Y(4260)$, like the
$Z_c^+(3900)$. The $Z_c^+(3900)$ was first observed by the BESIII
collaboration in the $(\pi^\pm J/\psi)$ mass spectrum
of the $Y(4260)\to J/\psi\pi^+\pi^-$ decay channel \cite{Ablikim:2013mio}.
This structure, was also observed at the same time by the
Belle collaboration \cite{Liu:2013dau} and was confirmed by the authors of
Ref. ~\cite{Xiao:2013iha} using CLEO-c data.
The decay modes of the $Y(4260)$ into $J/\psi$ and other charmonium
states indicate the existence of a $\bar{c}c$ in its content. However,
the attempts to classify this state in the charmonium spectrum
have failed since the $\Psi(3S),~\Psi(2D)$ and $\Psi(4S)$ $c\bar{c}$ states
have been assigned to the well established $\Psi(4040),~\Psi(4160),~$ and
$\Psi(4415)$ mesons respectively, and the prediction from quark models
for the $\Psi(3D)$ state is 4.52 GeV. Therefore, the mass of the $Y(4260)$
is not consistent with any of the $1^{--}$ $c\bar{c}$ states
\cite{Zhu:2007wz,Nielsen:2009uh}.
Some theoretical interpretations for the $Y(4260)$ are:
tetraquark state \cite{tetraquark}, hadronic $D_{1} D$, $D_{0} D^*$ molecule
\cite{Ding}, $\chi_{c1} \omega$ molecule \cite{Yuan}, $\chi_{c1} \rho$ molecule
\cite{liu}, $J/\psi f_0(980)$ molecule \cite{oset},
a hybrid charmonium \cite{zhu}, a charm-baryonium \cite{Qiao}, a cusp
\cite{eef1,eef2,eef3}, etc. Within the available experimental information,
none of these suggestions can be completely ruled out. However, there are some
calculations, within the QCD sum rules (QCDSR) approach
\cite{Nielsen:2009uh,svz,rry,SNB}, that can not explain
the mass of the $Y(4260)$ supposing it to be a tetraquark state \cite{rapha},
or a $D_{1} D$, $D_{0} D^*$ hadronic molecule \cite{rapha}, or a
$J/\psi f_0(980)$ molecular state \cite{Albuquerque:2011ix}.
In the framework of the QCDSR the mass and the decay width,
in the channel $J/\psi\pi\pi$, of the $Y(4260)$ were
computed with good agreement with data, considering it as a
mixing between two and four-quark states \cite{Dias:2012ek}.
The mixing is done at the level of the hadronic currents and, physically, this
corresponds to a fluctuation of the $c \overline{c}$ state where a gluon is
emitted and subsequently splits into a light quark-antiquark pair, which
lives for some time and behaves like a tetraquark-like state. The same
approach was applied to the $X(3872)$ state and good agreement with the data
were obtained for its mass and the decay width into $J/\psi\pi\pi$
\cite{x3872mix}, its radiative decay \cite{x3872rad}, and also in the
$X(3872)$ production rate in $B$ decay \cite{x3872prod}.
In this work we will focus on the production of the $Y(4260)$, using the
mixed two-quark and four-quark prescription of Ref.~\cite{Dias:2012ek}
to perform a QCDSR analysis of the process $B^-\to Y(4260)K^-$.
The experimental upper limit on the branching fraction
for such a production in $B$ meson decay has been reported by BaBar
Collaboration \cite{babary2}, with $95\%$ C.L.,
\begin{equation}
\label{branching}
{\mathcal B}_{_Y} <\! 2.9\times10^{-5}
\end{equation}
where ${\mathcal B}_{_Y} \equiv {\mathcal B}
(B^- \!\!\to\! K^- Y(4260),Y(4260) \!\to\! J/\psi\pi^+\pi^-)$.
\begin{figure}[h,t,b]
\centerline{
\includegraphics[width=0.35\textwidth]{byk_eff.eps}
}
\caption{The process for production of the $Y(4260)$ state in B meson decay,
mediated by an effective vertex operator ${\mathcal O}_2$.}
\label{weakdecay}
\end{figure}
The process $B\to Y(4260)K$ occurs via weak decay of the $b$ quark, while
the $u$ quark is a spectator. The $Y$ meson
as a mixed state of tetraquark and charmonium interacts via $\bar{c}c$
component of the weak current. In effective theory, at the scale
$\mu\sim m_b\ll m_W$, the weak decay is treated as a four-quark local
interaction described by the effective Hamiltonian (see Fig.~\ref{weakdecay}):
\begin{equation}\lb{ham}
{\mathcal{H}}_W=\frac{G_F}{\sqrt{2}}V_{cb}V_{cs}^*\left[\left(C_2(\mu)+
\frac{C_1(\mu)}{3}\right)
{\mathcal{O}}_2+\cdots\right]\,,
\end{equation}
where $V_{ik}$ are CKM matrix elements, $C_1(\mu)$ and $C_2(\mu)$ are
short distance Wilson coefficients computed at the renormalization
scale $\mu\sim{\mathcal O}(m_b)$. The four-quark effective operator
is ${\mathcal{O}}_2=J_{\mu}^{(\bar{c}c)}J_{\mu}^W$, with
\begin{equation}\lb{wcurrents} J_{\mu}^W=\bar{s}\Gamma_\mu b\,,\quad
J_{\mu}^{(\bar{c}c)}=\bar{c}\Gamma_\mu c\,,\end{equation}
and
$\Gamma_\mu=\gamma_\mu(1-\gamma_5)$.
Using factorization,
the decay amplitude of the process is calculated from the Hamiltonian
(\ref{ham}), by splitting the matrix element in two pieces:
\begin{eqnarray}\lb{amp}
{\mathcal M}
&=&i\frac{G_F}{\sqrt{2}}V_{cb}V_{cs}^*\left(C_2+\frac{C_1}{3}\right)\nn\\&
\times&\langle B(p)\vert J_{\mu}^W\vert K(p^\prime)\rangle\langle Y(q)
\vert J^{\mu(\bar{c}c)}\vert0\rangle,
\end{eqnarray}
where $p=p^\prime+q$.
Following Ref.~\cite{x3872prod}, the matrix
elements in Eq.~({\ref{amp}}) are parametrized as:
\begin{equation}\lb{2pmatrix}
\langle Y(q)\vert J_{\mu}^{(\bar{c}c)}\vert0\rangle=\lambda_W
\epsilon^\ast_\mu(q)\,,
\end{equation}
and
\begin{equation}\lb{3pmatrix}
\langle B(p)\vert J_{\mu}^W\vert K(p^\prime)\rangle=f_+(q^2)(p_\mu+
p_\mu^\prime)+f_-(q^2)(p_\mu-p_\mu^\prime)\,.
\end{equation}
The parameter $\lambda_W$ in (\ref{2pmatrix}) gives the coupling
between the current $J_\mu^{(\bar{c}c)}$ and the $Y$ state. The form
factors $f_\pm(q^2)$ describe the weak transition $B\to K$. Hence we
can see that the factorization of the matrix element describes the
decay as two separated sub-processes.
The decay width for the process $B^-\to Y(4260)K^-$ is given by
\begin{equation}\lb{eqwidth}
\Gamma(B\to YK)=\frac{\vert{\mathcal{M}}\vert^2}{16\pi m_B^3}\sqrt{\lambda(m_B^2,m_K^2,m_Y^2)},
\end{equation}
with $\lambda(x,y,z)=x^2+y^2+z^2-2xy-2xz-2yz$. The invariant amplitude
squared can be obtained from (\ref{amp}), using (\ref{2pmatrix})
and (\ref{3pmatrix}):
\begin{eqnarray}
\vert\mathcal{M}\vert^2&=&\frac{G_F^2}{2 m_Y^2}\vert V_{cb}V_{cs}\vert^2\left(C_2
+\frac{C_1}{3}\right)^2\nn\\\nn\\
&\times& \lambda(m_B^2,m_K^2,m_Y^2)\lambda_W^2f_+^2
\,.
\end{eqnarray}
The coupling constant $f_ +$ was determined in Ref.\cite{x3872prod} through
extrapolation of the form factor $f_+(Q^2)$ to the meson pole $Q^2 = - m_Y^2$,
using the QCDSR approach for the three-point correlator \cite{bcnn}:
\beqa
\Pi_{\mu}(p,p^\prime)&=&\!\!\int d^4x \,d^4y \,e^{i(p^\prime\cdot x-\,p
\cdot y)}\langle0\vert T\{J_\mu^W(0)\times\nn\\
&\times&J_K(x)J^\dagger_B(y)\}\vert0\rangle,
\enqa
where the weak current, $J^W_\mu$, is defined in
(\ref{wcurrents}) and the interpolating currents of the $B$ and $K$
pseudoscalar mesons are:
\begin{equation}
J_K = i\,\bar{u}_a \gamma_5 s_a \,,\quad J_B=i\,\bar{u}_a
\gamma_u b_a\,.
\end{equation}
The obtained result for the form factor was \cite{x3872prod}:
\begin{equation}\lb{fplus}
f_+(Q^2)=\frac{(17.55\pm0.04) \GeV^2}{(105.0\pm1.8)\GeV^2+Q^2}\,.
\end{equation}
For the decay width calculation, we need the value of the form factor
at $Q^2=-m_Y^2$, where $m_Y$ is the mass of the $Y(4260)$ meson.
Using $m_Y=(4251\pm9)\MeV$ \cite{pdg} we get:
\begin{equation}\lb{fpluspolo}
f_+(Q^2)\vert_{Q^2=-m_Y^2}=0.206\pm0.004\,.
\end{equation}
The parameter $\lambda_W$ can also be determined using the QCDSR approach
for the two-point correlator:
\begin{equation}
\Pi_{\mu\nu}(q)=i\int d^4y~e^{iq\cdot y}\langle0\vert T\{J_\mu^Y(y)
J_\nu^{(\bar{c}c)}(0)\}\vert0\rangle\,,
\label{2point}
\end{equation}
where the current $J_\nu^{(\bar{c}c)}$ is defined in
(\ref{wcurrents}). For the $Y$ meson we will follow \cite{Dias:2012ek} and
consider a mixed charmonium-tetraquark current:
\beq
J_\mu^Y = \sin \theta \:J_\mu^{(4)} + \cos\theta \:J_\mu^{(2)},
\label{jmix}
\enq
where
\beqa
J_\mu^{(4)} &=& \frac{\epsilon_{abc} \epsilon_{dec}}{\sqrt{2}}
\Big[(q_a^T C\ga_5 c_b)(\bar{q}_d \ga_\mu\ga_5 C\bar{c}_e^T)+\nn\\
&&~+(q_a^T C\ga_5\ga_\mu c_b)(\bar{q}_d\ga_5 C\bar{c}_e^T) \Big],
\label{j4q}
\enqa
\beq
J_\mu^{(2)}=\frac{1}{\sqrt{2}}\qq \:\Big(
\bar{c}_a\ga_\mu c_a \Big) ~\equiv~ \frac{1}{\sqrt{2}}\qq \:J_\mu^{'(2)} ~.
\enq
In Eq.~(\ref{jmix}), $\theta$ is the mixing angle that was determined in
\cite{Dias:2012ek} to be: $\theta=(53.0\pm0.5)^0$.
Inserting the currents (\ref{wcurrents})
and (\ref{jmix}) in the correlator we have in the OPE side of the sum
rule
\begin{eqnarray}
\Pi^{\mathrm{OPE}}_{\mu\nu}(q)&=&\sin\theta\,
\Pi^{4,2}_{\mu\nu}(q)
+\frac{\qq}{\sqrt{2}}\cos\theta\,
\Pi^{2,2}_{\mu\nu}(q)\,,
\end{eqnarray}
where
\begin{eqnarray}
\Pi^{4,2}_{\mu\nu}(q)&=&i\int d^4y ~e^{iq\cdot y}\langle0\vert T\{
J_\mu^{(4)}(y)J_{\nu(\bar{c}c)}(0)\}\vert0\rangle\nn\\
\Pi^{2,2}_{\mu\nu}(q)&=&i\int d^4y ~e^{iq\cdot y}\langle0\vert T\{
J_\mu^{'(2)}(y)J_{\nu(\bar{c}c)}(0)\}\vert0\rangle\,.
\label{pi24}
\end{eqnarray}
Only the vector part of the current $J_\nu^{(\bar{c}c)}$ contributes
to the correlators in Eq.~(\ref{pi24}). Therefore, these correlators are the
same as the ones calculated in Ref.~\cite{Dias:2012ek} for the mass of the
$Y(4260)$.
\begin{table}[t]
\setlength{\tabcolsep}{1.25pc}
\caption{\small QCD input parameters.}
\begin{tabular}{ll}
&\\
\hline
Parameters&Values\\
\hline
$\overline{m}_c$ & $(1.23 - 1.47) \GeV$ \\
$\qq$ & $ \hspace{-0.25cm}-(0.23 \pm 0.03)^3\GeV^3$\\
$\lag g_s^2 G^2 \rag$ & $(0.88 \pm 0.25)~\GeV^4$\\
$m_0^2 \equiv \lag\bar{q}Gq\rag/\qq$ & $(0.8 \pm 0.1) ~\GeV^2$\\
\hline
\end{tabular}
\label{Param}
\end{table}
To evaluate the phenomenological side we insert intermediate states of the
$Y$:
\begin{eqnarray}
\Pi_{\mu\nu}^{phen}(q)&=&\frac{i}{q^2-m_Y^2}\langle0\vert J^Y_\mu\vert
Y(q)\rangle\langle Y(q)\vert J^{(\bar{c}c)}_\nu\vert0\rangle\,,\nn\\
&=&\frac{i\lambda_Y\lambda_W}{Q^2+m_Y^2}\left(g_{\mu\nu}-\frac{q_\mu q_\nu}
{m_Y^2}\right)
\end{eqnarray}
where $q^2=-Q^2$, and we have used the definition (\ref{2pmatrix})
and
\begin{equation}
\langle0\vert J^Y_\mu\vert Y(q)\rangle=\lambda_Y\epsilon_\mu(q)\,.
\end{equation}
The parameter $\lambda_Y$, that defines the coupling between the current
$J^Y_\mu$ and the $Y$ meson, was determined in Ref.~\cite{Dias:2012ek}
to be: $\lambda_Y = (2.00 \pm 0.23) \times 10^{-2} ~ \GeV^5$.
As usual in the QCDSR approach, we perform a Borel transform to $Q^2\to M_B^2$
to improve the matching between both sides of the sum rules.
After performing the Borel transform in both sides of the sum rule we get in
the $g_{\mu\nu}$ structure:
\begin{eqnarray}\lb{2psumrule}
\lambda_W\lambda_Ye^{-\frac{m_Y^2}{M_B^2}}=
\frac{\sin \theta}{\sqrt{2}}\,\Pi^{4,2}(M_B^2)
+\frac{\qq}{\sqrt{2}} \cos \theta\,
\Pi^{2,2}(M_B^2) ~~~~~~
\end{eqnarray}
where the invariant functions $\Pi^{2,2}(M_B^2)$ and $\Pi^{4,2}(M_B^2)$ are
written in terms of a dispersion relation,
\begin{eqnarray}
\Pi(M_B^2) = \int\limits_{4m_c^2}^{\infty} \!ds ~e^{-s/M_B^2} \:\rho(s) ~~,
\end{eqnarray}
with their respective spectral densities $\rho^{2,2}(s)$ and $\rho^{4,2}(s)$
given in Appendix.
We perform the calculation of the coupling parameter $\lambda_W$
using the same values for the masses and QCD condensates as in
Ref.~\cite{Dias:2012ek} which are listed in Table \ref{Param}. To be
consistent with the calculation of $\lambda_Y$ we also use the same
region in the threshold parameter $s_0$ as in Ref.~\cite{Dias:2012ek}:
$\sqrt{s_{0}} = (4.70 \pm 0.10)$ GeV. As one can see in Fig.~\ref{figLW},
the region where we get $M_B^2$-stability is given by:
$(8.0 \leq M_B^2 \leq 25.0) \GeV^2$.
Taking into account the variation in the Borel mass parameter, in the
continuum threshold, in the quark condensate, in the coupling constant $\lambda_Y$
and in the mixing angle $\theta$, the result for the $\lambda_W$ parameter is:
\begin{equation}\lb{lambdaW}
\lambda_W=(0.90\pm0.32)\GeV^2\,.
\end{equation}
Thus we can calculate the decay width in Eq.~(\ref{eqwidth}) by using the values of
$f_+(-M_Y^2)$ and $\lambda_W$, determined in Eqs.~(\ref{fpluspolo}) and
(\ref{lambdaW}). The branching ratio is evaluated
dividing the result by the total width of the $B$ meson
$\Gamma_{\mathrm{tot}}=4.280 \times 10^{-4} \:\mbox{eV}$:
\begin{equation}\lb{result}
\mathcal{B}(B\to Y(4260)K)=(1.34\pm0.47)\times10^{-6}\,,
\end{equation}
where we have used the CKM parameters $V_{cs}=1.023$,
$V_{cb}=40.6\times10^{-3}$ \cite{pdg}, and the Wilson coefficients
$C_1(\mu)=1.082$, $C_2(\mu)=-0.185$, computed at $\mu=m_b$ and
$\bar{\Lambda}_{\mathrm{MS}}=225\MeV$ \cite{buras}.
\begin{figure}[bp]
\begin{center}
\includegraphics[width=0.45\textwidth]{LW.eps}
\caption{The coupling parameter $\lambda_W$ as a function of $M_B^2$, for
different values of the continuum threshold.}
\label{figLW}
\end{center}
\end{figure}
In order to compare the branching ratio in Eq.~(\ref{result}) with the branching fraction
obtained experimentally in Eq.~(\ref{branching}), we might use the results found in
Ref.~\cite{Dias:2012ek}:
\begin{equation}
\mathcal{B}(Y(4260) \to J/\psi \:\pi^+\pi^-) = (4.3 \pm 0.9)\times10^{-2}\,,
\label{brY}
\end{equation}
and then, considering the uncertainties, we can estimate
${\mathcal B}_{_Y} >\! 3.0 \times 10^{-8}$.
However, it is important to notice that the authors in Ref.~\cite{Dias:2012ek} have considered
two pions in the final state coming only from intermediate states, e.g. $\sigma$ and $f_0(980)$
mesons, which could indicate that the result in Eq.~(\ref{brY}) can be underestimated.
In this sense, considering that the main decay channel observed for the $Y(4260)$ state is
into $J/\psi \:\pi^+\pi^-$, we would naively expect that the branching ratio into this channel could
also be $\mathcal{B}(Y(4260) \to J/\psi \:\pi^+\pi^-) \sim 1.0$, which would lead to the following
result, ${\mathcal B}_{_Y} <\! 1.8 \times 10^{-6}$.
Therefore, we obtain an interval on the branching fraction
\begin{equation}
3.0 \times 10^{-8} < {\mathcal B}_{_Y} < 1.8 \times 10^{-6}
\end{equation}
which is in agreement with the experimental upper limit reported by Babar Collaboration
given in Eq.~(\ref{branching}). In general the experimental evaluation of the branching fraction
takes into account additional factors related to the numbers of reconstructed events for the
final state ($J/\psi \:\pi^+\pi^- \:K$), for the reference process ($B \to Y(4260) \:K$), and for the
respective reconstruction efficiencies. However, since such information has
not been provided in Ref.~\cite{babary2}, we have neglected these factors in the calculation of
the branching fraction ${\mathcal B}_Y$. Therefore, the comparison of our result with the
experimental result could be affected by these differences.
In conclusion, we have used the QCDSR approach to evaluate the production of the $Y(4260)$
state, considered as a mixed charmonium-tetraquark state, in the decay
$B\to YK$. Using the factorization hypothesis, we find that the sum rules
result in Eq.~(\ref{result}), is compatible with the experimental upper limit.
Our result can be interpreted as a lower limit for the branching ratio,
since we did not considered the non-factorizable contributions.
Our result was obtained by considering the mixing angle in
Eq.~(\ref{jmix}) in the range $\theta=(53.0\pm0.5)^0$. This
angle was determined in Ref.~\cite{Dias:2012ek} where the mass and
the decay width of the $Y(4260)$ in the channel $J/\psi\pi^+\pi^-$
were determined in agreement with experimental values. Therefore,
since there is no new free parameter in the present analysis,
the result presented here strengthens the conclusion reached in
\cite{Dias:2012ek} that the $Y(4260)$ is probably a mixture
between a $c\bar{c}$ state and a tetraquark state.
As discussed in \cite{x3872prod}, it is not simple to determine the
charmonium and the tetraquark contribution to the state described by the
current in Eq.~(\ref{jmix}). From Eq.~(\ref{jmix}) one can see that, besides
the $\sin\theta$, the $c\bar{c}$ component of the current is multiplied by a
dimensional parameter, the quark condensate, in order to have the same
dimension of the tetraquark part of the current.
Therefore, it is not clear that only the
angle in Eq.~(\ref{jmix}) determines the percentage of each component.
One possible way to evaluate the importance of each part of the current it is
to analyze what one would get for the production rate with each component,
{\it i.e.}, using $\theta=0$ and $90^\circ$ in Eq.~(\ref{jmix}). Doing this
we get respectively for the pure tetraquark and pure charmonium:
\begin{eqnarray}
{\mathcal{B}}(B\to Y_{\mathrm{tetra}}K) &=& (1.25\pm0.23)\times10^{-6}\,, \\
{\mathcal{B}}(B\to Y_{\bar{c}c}K) &=&(1.14\pm0.20)\times10^{-5}\,.
\end{eqnarray}
Comparing the results for the pure states with the one for the mixed
state (\ref{result}), we can see that the branching ratio for the pure
tetraquark is one order smaller, while the pure charmonium is larger.
From these
results we see that the $c\bar{c}$ part of the state plays a very important
role in the determination of the branching ratio. On the other hand, in the
decay $Y\to J/\psi\pi^+\pi^-$, the width
obtained in our approach for a pure $c\bar{c}$ state is \cite{Dias:2012ek}:
\begin{equation}\lb{xppcc}
\Gamma(Y_{\bar{c}c}\to J/\psi\pi\pi)=0\,,
\end{equation}
and, therefore, the tetraquark part of the state is the only one
that contributes to this decay, playing an essential role
in the determination of this decay width.
Therefore, although we can not determine the percentages of the $c\bar{c}$
and the tetraquark components in the $Y(4260)$, we may say that both
components are extremely important, and that, in our approach, it is not
possible to explain all the experimental data about the $Y(4260)$ with
only one component.
\vspace{0.8cm}
\section*{Acknowledgment}
\noindent
This work has been partially supported by S\~{a}o Paulo Research Foundation
(FAPESP), grant n.\,2012/22815-3, and National Counsel of Technological and
Scientific Development (CNPq-Brazil).
\vspace{0.8cm}
|
\section{Introduction}
\label{sec:intro}
The dynamics of colloid suspensions is crucially influenced by
flow-mediated correlations~\cite{Happel&Brenner,Russel}.
While these hydrodynamic interactions (HI)
have an important role
in the dynamics of ambient suspensions at thermal
equilibrium~\cite{Russel}, their effect becomes even more pronounced for
objects driven out of equilibrium,
where the total force acting on each object generates a
long-ranged flow, decaying as $1/R$ with the distance $R$ between the objects.
A well-known example is colloid
sedimentation, where HI lead to strongly correlated motions and
large-scale dynamic structures~\cite{Ramaswamy2001}. Various types of
driving, such as electrophoresis,
are widely used to control the transport of colloids and other polyatomic objects~\cite{Russel}.
Theoretical studies of driven colloids traditionally focus on regular particle shapes such as uniform spheres and ellipsoids. The driving
of more asymmetric objects is richer~\cite{Makino&Doi2003,Gonzalez_etal2004,Doi&Makino2005,Makino&Doi2005}
as it generally includes coupling
between translation and rotation\,---\,when the object is subjected to
a force it also rotates, and when it is under torque it also
translates~\cite{Happel&Brenner}.
The choice of a rotation sense under a unidirectional
force implies a chiral response of the driven object. Such
richer responses can be exploited to obtain {\it``steerable
colloids''}\,---\,objects whose orientation and transport can be
controlled in much more detail. For example, applying a torque by a
rotating uniform magnetic field was used to achieve efficient
transport of chiral magnetic objects~\cite{Morozov&Leshansky2014}. Another
example, which is the main issue of the present work, is the ability
to achieve orientational alignment of asymmetric objects by applying an
external force~\cite{Krapf_etal2009,Moths&Witten2013,Moths&Witten2013b}.
The earlier theoretical works of
Refs.\ \cite{Morozov&Leshansky2014,Krapf_etal2009,Moths&Witten2013,Moths&Witten2013b} dealt with
isolated asymmetric objects in Stokes flow,
which exhibit a chiral response. The object's chiral response
is encoded in the off-diagonal block of its self-mobility
matrix, referred to as the {\it twist matrix}.
Some objects have a twist matrix that leads them to align one axis in the body with the applied force.
If the twist matrix has only a single real eigenvalue, the object becomes ``axially aligned''
in this way~\cite{Gonzalez_etal2004,Krapf_etal2009}, and the aligning direction is along
the corresponding eigenvector.
Hence, in the absence of HI and thermal
fluctuations, a set of identical, axially aligning objects reach a
partially aligned state, where all the objects rotate about the same
axis with the same angular velocity, but with an arbitrary phase.
Furthermore, it was shown that, by applying an
appropriate time-dependent forcing, the system can be driven to a
fully aligned state, where all the objects are phase-locked with the force
and rotate in synchrony \cite{Moths&Witten2013,Moths&Witten2013b}.
In view of the above we use throughout this article
the following terminology concerning the response of various objects:
(i) {\it symmetric} objects (such as a uniform sphere); (ii) {\it regular} objects,
which are asymmetric objects with a vanishing twist matrix (such as a uniform ellipsoid);
(iii) {\it irregular} objects, having a non-vanishing twist matrix; (iv)
{\it axially alignable} objects, which are irregular objects,
whose twist matrix has a single real eigenvalue.
We note that the twist matrix depends on the position of the forcing point as well.
For example, an ellipsoid whose forcing point is displaced from its centroid,
i.e., an ellipsoid with a non-uniform mass distribution under gravity, has a non-vanishing
twist matrix, and generally might be alignable.
The theoretical groundwork for treating the HI between arbitrary
objects in Stokes flow was laid by Brenner and O'Neill
\cite{BrennerII,Brenner&Oneill1972}. The theory was subsequently
applied to a pair of particles of various regular shapes
\cite{Goldman_etal1966,Wakiya1965,Felderhof1977,Jeffrey&Onishi1984,Liao&Krueger1980,Kim1985,Kim1986}.
To this one should add many earlier studies of the collective dynamics of suspensions made of ellipsoids
\cite{Hinch&Leal1972,Brenner1974,Claeys&Brady1993II,Jeffery1922,Davis1991}.
We note that there are key differences between asymmetric objects, such as ellipsoids,
and the irregular objects studied here. The symmetries of a uniform ellipsoid lead to:
(a) the absence of a translation-rotation coupling for a single object,
and therefore lack of alignability;
(b) the absence of a $1/R^2$ contribution to the relative velocity
developed between two such objects at mutual distance $R$.
Finally, several numerical techniques have been introduced to treat
suspensions of arbitrarily shaped objects
\cite{Karrila_etal1989,Cong&Thien1989,Carrasco&Torre1999,Kutteh2010,Cichocki_etal1994}.
In this work we focus on simple, general properties of the pair HI
between two arbitrarily shaped objects at zero Reynolds number, and
the resulting effect on their orientational alignment. The study of
translational effects will be presented in a separate publication.
The work is made of two distinct parts. The first part treats rigorously
the instantaneous hydrodynamic interaction, i.e., the pair-mobility matrix.
We use Brenner's analytical framework \cite{BrennerI,BrennerIV}, specializing to
the leading order of the HI in the distance between the objects
(multipole expansion,
also known as the method of reflections~\cite{Happel&Brenner}).
The second part addresses the time-dependent trajectories of forced objects.
This is a multi-variable, highly non-linear dynamical system exhibiting complex and diverse dynamics.
In this part we are limited to numerical integration of the objects' trajectories.
We provide typical examples for the time evolution of pairs of stokeslet objects.
We begin by discussing in Sec.\ \ref{sec:general} the general properties
and symmetries of the pair-mobility matrix for two arbitrarily shaped
objects.
In Sec.\ \ref{sec:multipole} we apply a multipole expansion
to the pair-mobility matrix and obtain results for the instantaneous HI
at large distances.
In Sec.\ \ref{sec:methods} we derive the resulting
properties of stokeslet objects, and in Sec.\ \ref{sec:alignment} we
use them to perform numerical time integration for the evolution of
object pairs and their alignment. Finally, in Sec.\ \ref{sec:discuss}, we
discuss several consequences of our results.
\section{Pair-Mobility Matrix: General Considerations}
\label{sec:general}
\subsection{Structure of the Pair-Mobility Matrix}
\label{sec:structure}
The kinematics of a rigid object is represented by a translational
velocity $\vec{V}$, which refers to an arbitrary reference point
rigidly affixed to the object, and an angular velocity
$\vec{\omega}$. We designate the reference point as \emph{the origin}
of the object. Note that the angular velocity of the object is
independent of the choice of its origin, and that the origin does not
necessarily lie on the instantaneous axis of rotation of the object.
Consider two arbitrarily shaped rigid objects, $a$ and $b$, with typical
size $l$, subject to external forces and torques $\vec{F}^a$,
$\vec{F}^b$ and $\vec{\tau}^a$, $\vec{\tau}^b$ in an unbounded,
otherwise quiescent fluid of viscosity $\eta$. In the creeping flow
regime, the objects respond with linear and angular velocities to the
external forces and torques through a $12 \times 12$
\emph{pair-mobility matrix},
\begin{equation}
\begin{pmatrix}
\Cvvec{V}^a \\
\Cvvec{V}^b
\end{pmatrix} =
\frac{1}{\eta l}
\begin{pmatrix}
\Mmat{M}^{aa} & \Mmat{M}^{ab}\\
\Mmat{M}^{ba} & \Mmat{M}^{bb}
\end{pmatrix}
\begin{pmatrix}
\Cvvec{F}^a \\
\Cvvec{F}^b
\end{pmatrix},
\end{equation}
where we define \emph{generalized velocity} and \emph{generalized
force} 6-vectors, $\Cvvec{V}^x=( \vec{V}^x, l \vec{\omega}^x )^T$
and $\Cvvec{F}^x= (\vec{F}^x, \vec{\tau^x}/l )^T$ for $x=a,b$. The
diagonal blocks, $\Mmat{M}^{aa}$ and $\Mmat{M}^{bb}$, correspond to
the self-mobilities of the objects (which nevertheless depend
on the configuration of both objects). The off-diagonal blocks,
$\Mmat{M}^{ab}$ and $\Mmat{M}^{ba}$, describe the pair hydrodynamic
interaction. We hereafter omit the factor $(\eta l)^{-1}$ (i.e., set $\eta l=1$).
This, together with the representation of the generalized forces and velocities, make
$\Mmat{M}$ dimensionless and dependent on the geometry alone.
Throughout the text we designate 6-vectors and matrices
with calligraphic font and blackboard-bold letters, respectively.
A detailed description of the notation used in the article is given in Appendix~\ref{sec:notation}.
Since $\vec{V}$ and $\vec{\tau}$ depend on the choice of object origins,
so does the pair-mobility matrix. The transformation between pair-mobility
matrices corresponding to different origins is given in Appendix~\ref{sec:origin}.
The pair-mobility matrix is a function of the objects' geometries,
their orientations, and the vector connecting their origins, indicated
hereafter by $\vec{R}$. (We define the direction of $\vec{R}$ from the
origin of object $b$ to the origin of object $a$.) The geometry of
object $x$ is denoted by $\bvec{r}^{x}$. For example, if the object
consists of a discrete set of $N_x$ stokeslets (see Sec.~\ref{sec:stokeslet_properties}),
then $\bvec{r}^{x}$ is a $3N_x$-vector specifying the positions of the
stokeslets; otherwise, it represents the surface of the object.
The pair-mobility matrix is positive-definite and
symmetric~\cite{Condiff&Dahler1966,Happel&Brenner,Landau&Lifshitz}. Hence,
$\Mmat{M}^{ab}=(\Mmat{M}^{ba})^T$, and the self-blocks can be written
as
$$
\Mmat{M}^{xx} = \begin{pmatrix}
\Mmat{A}^{xx} & (\Mmat{T}^{xx})^T \\
\Mmat{T}^{xx} & \Mmat{S}^{xx}
\end{pmatrix} .
$$
As in the analysis for isolated objects \cite{Krapf_etal2009}, the
self-mobility matrix contains the following $3\times 3$ blocks: the
alacrity matrix $\Mmat{A}$ (translational response to force); the
screw matrix $\Mmat{S}$ (rotational response to torque); and the twist
matrix $\Mmat{T}$ (translation--rotation coupling). The twist matrix
characterizes the chiral response of the object (the sense of rotation
under a force). In the present article we deal with alignable objects,
whose individual $\Mmat{T}$ is necessarily non-vanishing. Furthermore,
in the case of a pair of objects, the presence of the other object
makes the self-twist matrix, $\Mmat{T}^{xx}$, differ from the
single-object one. As to the off-diagonal blocks of the pair-mobility
matrix, the symmetry of $\Mmat{M}$ implies the following structure:
$$
\Mmat{M}^{ab} = \begin{pmatrix}
\Mmat{A}^{ab} & (\Mmat{T}^{ba})^T \\
\Mmat{T}^{ab} & \Mmat{S}^{ab}
\end{pmatrix},\ \ \
\Mmat{M}^{ba} = \begin{pmatrix}
(\Mmat{A}^{ab})^T & (\Mmat{T}^{ab})^T \\
\Mmat{T}^{ba} & (\Mmat{S}^{ab})^T
\end{pmatrix} .
$$
\subsection{Further Symmetries of the Pair-Mobility Matrix}
\label{sec:symmetry}
The discussion in the preceding subsection has been for a general pair of objects,
which are not necessarily identical.
In the present subsection, we focus on the case in which the two objects are
\emph{identical in shape and orientation}, i.e.,
$\bvec{r}^a=\bvec{r}^b\equiv\bvec{r}$. Our goal is to understand what
the instantaneous relative velocities (linear and angular) between the
two objects are, when the objects are subjected to the same external
forcing. The restriction to identical objects makes $\Mmat{M}$ invariant under
exchange of objects. This additional symmetry is made of two
operations: interchanging the blocks
$\Mmat{M}^{aa}\leftrightarrow\Mmat{M}^{bb}$ and
$\Mmat{M}^{ab}\leftrightarrow\Mmat{M}^{ba}$; and inversion of $\vec{R}$.
That is,
\begin{equation}
\Mmat{M} (\bvec{r},\vec{R})=
\Mmat{E} \Mmat{M} (\bvec{r},-\vec{R}) \Mmat{E}^{-1} ,
\label{eq:switch}
\end{equation}
where $\Mmat{E}$ is a $12\times12$
matrix which interchanges the objects,
$$
\Mmat{E}=
\begin{pmatrix}
0 & \Mmat{I}_{6\times6}\\
\Mmat{I}_{6\times6} & 0
\end{pmatrix},
$$
with $\Mmat{I}_{6\times6}$ denoting the $6\times 6$ identity matrix.
The symmetry to object exchange, when combined with the
parity of $\Mmat{M}$
(i.e., whether it remains the same or changes sign)
under $\vec{R}$-inversion,
\footnote{\setcounter{footnote}{1} Parity does not mean here symmetry under full spatial inversion,
as such an operation would turn the chiral objects
into their enantiomers; rather, we mean here symmetry under the inversion of $\vec{R}$}
has important consequences for the effect of hydrodynamic interactions on
alignment. If $\Mmat{M}$ has a definite parity one can determine what
the relative response of the objects to forcing is\,---\,i.e., whether
they attain the same or the opposite linear and angular velocities. If
the term is symmetric to inversion, the velocities would be identical,
and if it is antisymmetric, they would be opposite. This is because
\begin{equation}
\begin{pmatrix}
\Mmat{M}^{aa}(\vec{R}) & \Mmat{M}^{ab}(\vec{R})\\
\Mmat{M}^{ba}(\vec{R}) & \Mmat{M}^{bb}(\vec{R})
\end{pmatrix}=
\pm
\begin{pmatrix}
\Mmat{M}^{aa}(-\vec{R}) & \Mmat{M}^{ab}(-\vec{R})\\
\Mmat{M}^{ba}(-\vec{R}) & \Mmat{M}^{bb}(-\vec{R})
\end{pmatrix}
=
\pm
\begin{pmatrix}
\Mmat{M}^{bb}(\vec{R}) & \Mmat{M}^{ba}(\vec{R})\\
\Mmat{M}^{ab}(\vec{R}) & \Mmat{M}^{aa}(\vec{R})
\end{pmatrix},
\label{eq:even_odd}
\end{equation}
where the second equality comes from the response to exchange of
objects, Eq.~\eqref{eq:switch}. Consequently, under identical forcing
of the two objects one finds,
\begin{equation}
\Cvvec{V}^a=\left(\Mmat{M}^{aa}+\Mmat{M}^{ab} \right)\Cvvec{F}
= \pm \left(\Mmat{M}^{bb}+\Mmat{M}^{ba} \right)\Cvvec{F}=\pm \Cvvec{V}^b.
\end{equation}
Thus, since any $\Mmat{M}$ can be decomposed into even and odd terms,
we find that only the odd ones cause relative motions of the two
objects.
The pair-mobility as a whole, however, never has a definite parity
under $\vec{R}$-inversion, i.e., it is made of both even and odd
terms. This becomes clear when $\Mmat{M}(\bvec{r},\vec{R})$ is
expanded in small $l/R$, i.e., in multipoles. A general discussion of
the parity of each multipole term is given in the next section. For now, let
us consider those two leading multipoles which are independent of the
objects' shape, and therefore always exist. The monopole--monopole
interaction (Oseen tensor), which is the leading term in
$\Mmat{A}^{ab}$ making particle $a$ translate due to the force on
particle $b$, is symmetric under $\vec{R}$-inversion. The part of the
monopole--dipole interaction causing the second object to rotate due
to the force on the first, i.e., the leading term in $\Mmat{T}^{ab}$,
is antisymmetric. For example, even the most symmetric pair of
objects\,---\,two spheres\,---\,has an $\vec{R}$-symmetric
$\Mmat{A}^{ab}$, leading to zero relative velocity, and an
$\vec{R}$-antisymmetric $\Mmat{T}^{ab}$, causing them to rotate with
opposite senses \cite{Happel&Brenner}. Thus, for a general object, the highest
order which maintains $\Mmat{M}$ of definite parity is the monopole
$1/R$ Oseen one, which is even. (The self-blocks are constant up to
order $1/R^4$; see below.)
From this discussion we can immediately conclude that, to leading
order in the separation of two identical, fully aligned objects, {\em their
instantaneous hydrodynamic interaction
must linearly degrade the alignment}. The leading degrading term comes from
$\Mmat{T}^{ab}$, their rotational response to force, and is of order $1/R^2$.
It is worthwhile to note again that such a rotational response is present as well for a pair
of uniform spheres or ellipsoids; yet, such regular objects are not alignable to begin with.
The relation between object-exchange symmetry and the symmetry of the
linear-velocity response is intimately related to the issue of
hydrodynamic pseudo-potentials \cite{Squires2001}, which will be discussed
in detail in a forthcoming publication.
\section{Far-Field Interaction: Multipole Expansion}
\label{sec:multipole}
There are two characteristic length scales in our problem: the typical
size of the objects, $l$, and the distance between them,
$R=|\vec{R}|$. If $l \ll R$, we can write the pair-mobility matrix as
a power series in $(l/R)$,
$$
\Mmat{M}=\Mmat{M}_{(0)}+\Mmat{M}_{(1)}+\Mmat{M}_{(2)} + \dots,
$$
where $\Mmat{M}_{(n)} \sim (l/R)^n$. The analysis of this expansion as given below
holds for any pair of objects, whether identical or not.
The zeroth order,
$\Mmat{M}_{(0)}$, is a block diagonal matrix which is made of the self-mobilities of the two
non-interacting objects. (These should be distinguished from $\Mmat{M}^{aa}$ and
$\Mmat{M}^{bb}$, the self-mobilities of the interacting objects.)
The hydrodynamic multipole expansion (also known as the method of reflections)
is based on the Green's function
of Stokes flow, the Oseen tensor \cite{Happel&Brenner},
given in our units ($\eta l=1$) by
\begin{equation}
\Mmat{G}_{ij} (\vec{r})=
\frac{1}{8\pi} \frac{l}{r}
\left( \delta_{ij} + \frac{r_i r_j}{r^2} \right),
\label{eq:Oseen}
\end{equation}
which is a symmetric $3\times 3$ tensor, invariant under
$\vec{r}$-inversion. A point force at $\vec{r}_0$,
$\delta(\vec{r}-\vec{r}_0) \vec{f}$, generates a velocity field
$\vec{u}(\vec{r})=\Mmat{G}(\vec{r}-\vec{r}_0)\cdot \vec{f}$.
We obtain two general results concerning the multipoles of the
hydrodynamic interaction between two arbitrary objects. The two
objects need not be identical. The proofs are given in
Appendix~\ref{sec:proofs}. \footnote{In fact, these results are not
special to the hydrodynamic interaction but can be similarly proven
for any multipole expansion. As such, they were most probably
derived before.}
\begin{enumerate}
{\it
\item The leading interaction multipole in the self-blocks of the
pair-mobility matrix is $n=4$.
That is, any response of one object to forces on itself, owing to the other object,
must fall off with distance R between the objects at least as fast as $R^{-4}$.
\item The $n$th multipole has self-blocks of $(-1)^n$ parity, and coupling blocks
of the opposite, $(-1)^{n+1}$ parity.
Thus, e.g., the leading term in $\Mmat{M}^{aa}$, proportional to $R^{-4}$, is invariant under $\vec{R}$-inversion,
and the $R^{-4}$ part of $\Mmat{M}^{ab}$ changes sign under $\vec{R}$-inversion.
Likewise for the multipole varying as $\vec{R}^{-5}$,
the $\Mmat{M}^{aa}$ changes sign under $\vec{R}$-inversion while $\Mmat{M}^{ab}$ remains invariant.\cite{Note1}
}
\end{enumerate}
These statements pertain to the mobility matrix. As to the propulsion
matrix (the inverse of the mobility matrix), the leading correction to the
self-block becomes $\sim 1/R^2$, and the second statement concerning parity remains intact.
We now consider for a moment two identical objects and specialize to the
first and second multipoles, i.e., the hydrodynamic interaction up to
order $1/R^2$. The discussion in the preceding and current sections
implies the following form of the two leading terms in the
pair-mobility matrix:
\begin{equation}
\Mmat{M}_{(1)} =
\begin{pmatrix}
0 & \Mmat{M}_{(1)}^{ab}\\
\Mmat{M}_{(1)}^{ab} & 0
\end{pmatrix},\qquad
\Mmat{M}_{(2)}=
\begin{pmatrix}
0 & \Mmat{M}_{(2)}^{ab}\\
-\Mmat{M}_{(2)}^{ab} & 0\\
\end{pmatrix}.
\label{eq:M1_M2}
\end{equation}
In more detail: there are no first- and second-order corrections to
the objects' self-mobility. Hence, these two multipoles have definite
parities\,---\,the first is even, and the second is odd.
Consequently, the first multipole does not cause any relative motion
of the two objects, whereas the second mutipole makes them translate
and rotate in opposite linear and angular velocities.
The essential characteristics of the first two multipoles
are schematically illustrated in Fig.~\ref{fig:multipole}.
The first multipole arises directly from the Green's function,
\begin{equation}
\Mmat{M}^{ab}_{(1)}=
\begin{pmatrix}
\Mmat{G}(\vec{R}) & 0\\
0 & 0
\end{pmatrix},
\label{eq:M1}
\end{equation}
where $\Mmat{G}(\vec{R})$ is the Oseen tensor, given in
Eq.~(\ref{eq:Oseen}).
In the interaction described by the second multipole one object sees
the other as a point, see Fig.~\ref{fig:multipole}. Accordingly, this term contains two types of
interaction: (1) the response of object $a$ to the non-uniformity of
the flow due to the force monopole at object $b$ (regarded as a
point); (2) the advection of object $a$ (regarded as a point) by the
flow due to the force dipole acting at object $b$. These two effects
are both proportional to $\vec{\nabla}\Mmat{G}(\vec{R})\sim
1/R^2$. Each can be written as a product of a tensor which arises from
the medium alone, through derivatives of the Oseen tensor
$\vec{\nabla}\Mmat{G}(\vec{R})$, and another tensor which depends on
the objects' geometry. The second-order correction to the velocity of
object $a$ is given by the sum of these two effects,
each expressed in terms of a coupling tensor $\Theta$ and an object tensor $\Phi$
\begin{eqnarray}
\Cvvec{V}^a_{(2)} &=& \Mmat{M}^{ab}_{(2)} \cdot \Cvvec{F}^b \nonumber\\
\Mmat{M}^{ab}_{(2)} &=& \Phi^a : \Theta(\vec{R})
- \Theta^T(\vec{R}) : \tilde{\Phi}^b,
\label{eq:V2a}
\end{eqnarray}
where the double dot notation denotes a contraction over two indices.
Equation~(\ref{eq:V2a}) contains three tensors of rank 3, denoted by capital Greek letters.
The first,
$\Phi$, with dimensions $6\times 3\times 3$, gives the generalized velocity of the object in linear
response to the velocity gradient of the flow in which it is
embedded. The second, $\tilde{\Phi}$, having dimensions $3\times 3\times 6$, gives the force dipole acting on
the fluid around the object's origin in linear response to the
generalized force acting on it. Both $\Phi$ and $\tilde{\Phi}$ depend
on the objects' geometry alone~\footnote{These tensors are related to
the two introduced by Brenner~\cite{BrennerIV}. Brenner's tensors
give the force and torque exerted on an object in linear response to
a flow gradient in which it is embedded. Our $\Phi$ is related to
these two via the individual self-mobility matrix.}. The third tensor, $\Theta$,
with dimensions $3\times 3\times 6$,
describes the coupling of these object responses through the fluid.
It is given by
\begin{equation}
\Theta_{skj}(\vec{R}) \equiv
\left\{
\begin{array}{ll}
\partial_s \Mmat{G}_{kj}(\vec{r})|_{\vec{R}}\ \ & j=1,2,3 \\
0\ \ & j=4,5,6.
\end{array} \right.
\end{equation}
Repeating the same procedure for
$\Cvvec{V}^b$ in response to $\Cvvec{F}^a$ while using the odd parity
of $\Theta$, we get
\begin{equation}
\Mmat{M}^{ba}_{(2)} = \Theta^T(\vec{R}) : \tilde{\Phi}^a
- \Phi^b : \Theta(\vec{R}).
\label{eq:V2b}
\end{equation}
The tensors $\Phi$ and $\tilde{\Phi}$ are not independent~\cite{Kim&Karrila}. We now show
that $\Phi=\tilde{\Phi}^T$. The symmetry of $\Mmat{M}$ implies that
each multipole is also a symmetric matrix. Using Eqs.~(\ref{eq:V2a})
and (\ref{eq:V2b}) and equating $(\Mmat{M}^{ba}_{(2)})^T =
\Mmat{M}^{ab}_{(2)}$, we get $\tilde{\Phi}^a = (\Phi^a)^T$ and
$\tilde{\Phi}^b = (\Phi^b)^T$.
To summarize, the matrix $\Mmat{M}_{(2)}$ is given by
\begin{equation}
\Mmat{M}_{(2)}=
\begin{pmatrix}
0 & \Phi^a : \Theta(\vec{R}) - [ \Phi^b : \Theta(\vec{R}) ]^T\\
-\Phi^b : \Theta(\vec{R}) + [ \Phi^a : \Theta(\vec{R}) ]^T & 0\\
\end{pmatrix}.
\label{eq:Mtensor}
\end{equation}
This results is valid for a general pair of objects. If the two objects
are identical, the off-diagonal blocks have the same form with opposite signs.
The additional condition that the entire $\Mmat{M}$ must be symmetrical implies
then that each block by itself is antisymmetric.
By separating the tensors $\Phi$ and $\Theta$ into their symmetric and
antisymmetric parts, the second-order term of the pair-mobility matrix
can be simplified further.
It should be mentioned, in addition, that
the $\Phi$ tensor depends on the origin selected for the object. These
two technical issues are addressed in Appendices
\ref{sec:general_form} and \ref{sec:Phi}, respectively.
Finally, we note that the terms in these tensors corresponding to the
translational response vanish for spheres and ellipsoids.
Consequently, two such regular objects develop relative velocity only to orders $1/R^3$
and above.
\begin{figure}
\centerline{\resizebox{0.8\textwidth}{!}{\includegraphics[viewport= 0 -1 551 360]{fig1}}}
\caption[]{
Illustration of the two leading orders of the hydrodynamic interaction between two forced objects.
The leading term in the pair-mobility matrix (light blue/dash-dotted arrow between
the objects' origins), decaying as $1/R$,
comes from the point-like response of object $a$ to the local flow caused by the force monopole on object $b$
(blue/thick arrow).
The next-order term, decaying as $l/R^2$, has two contributions:
(i) The point-like response of object $a$ to the local flow caused by the force dipole on object $b$
(red/dashed arrow from the red/thin arrows at object $b$ to the origin of $a$).
(ii) The response of object $a$ to the local flow gradient caused by the force monopole on object $b$
(magenta/dotted arrow from the origin of $b$ to the magenta/thin arrows at object $a$).}
\label{fig:multipole}
\end{figure}
\section{Numerical analysis for stokeslet objects}
\label{sec:methods}
In the preceding sections we have derived the general properties of
the instantaneous hydrodynamic interaction between two arbitrarily shaped objects.
We now move on to the second part of the work, addressing the time evolution
of the two objects. This complicated problem is not tractable analytically, and we
resort to numerical integration of specific examples.
Because of the complexity of the problem, and since we are interested in generic properties,
we allow ourselves to restrict the analysis to the simplest, even if unrealistic, objects.
Arguably the
simplest form of an arbitrarily shaped object is the so-called {\it
stokeslet object}\,---\,a discrete set of small spheres, separated
by much larger, rigid distances, where each sphere is approximated as
a point force.
The sparseness of these objects makes them free-draining, which may be valid for
macromolecules but not for compact objects.
We treat pairs of identical objects, each made of four
stokeslets.
To obtain representative sampling of numerical examples
we do not design these objects but create them randomly.
Four points are placed at random distances ranging between $0$ and
$1$ from an arbitrary origin. The origin is then shifted to the points'
center of mass.
The radius $\rho$ of the stokeslets is taken as $0.01$. The
resulting configuration is checked to be ``sufficiently chiral'', in
the sense that the $\Mmat{T}$-matrix of the individual object is strongly
asymmetric, having a single real eigenvalue of absolute value
$|\lambda_3|>0.005$, which makes the object axially alignable. (See
Sec.\ \ref{sec:intro}.). Examples of the stokeslet objects
we use are provided in Fig.~\ref{fig:objects}.
The way to calculate the mobility of a single stokeslet object was
presented in Ref.\ \citenum{Krapf_etal2009}. First, we briefly present in
Sec.\ \ref{sec:stokeslet_properties} the simple extension of this
method to pair-mobilities. We calculate both the pair mobility and the
tensor $\Phi$ introduced in Secs.\ \ref{sec:general} and \ref{sec:multipole}.
The latter allows us to
calculate pair mobilities up to second order in the multipole
expansion. Section \ref{sec:numerical} describes how we use the pair
mobility to numerically calculate the time evolution of the pair
configuration.
\subsection{Pair-Mobility and $\Phi$ Tensor}
\label{sec:stokeslet_properties}
The properties of a stokeslet object can be derived self-consistently
from the linear relations which describe the stokeslets' configuration.
This is done without finding the stokeslets' strengths explicitly. Below
we find the pair-mobility matrix, and the $\Phi$ tensor associated
with a single object, given the stokeslet configuration and the size
of the spheres that they represent.
Each of the two objects, $x=a,b$, consists of $N_x$ stokeslets,
$\bvec{F}^x=(\vec{F}^{x}_{1},\dots,\vec{F}^{x}_{N_x})$, in a known
configuration,
$\bvec{r}^x=\left(\vec{r}^{x}_{1},\dots,\vec{r}^{x}_{N_x} \right)$.
Here, we use the notation of a bold letter to denote a set of $N$ 3-vectors,
and $\vec{r}^{x}_{n}$ indicates the position 3-vector of the $n$th
stokeslet in object $x$ with respect to the object's origin.
Each
stokeslet is a sphere of radius $\rho$, where $\rho<\min(r^{x}_{1},\dots
,r^{x}_{N_x})$. The boundary conditions at the sphere surface enter only through its
self-mobility coefficient. The velocities of the spheres, $\vec{v}^{x}_{n}$, are
known from the object's linear and angular velocities,
\begin{equation}
\begin{pmatrix}
\bvec{v}^a \\
\bvec{v}^b
\end{pmatrix} =
\begin{pmatrix}
\Mmat{U}^{a} & 0\\
0 & \Mmat{U}^{b}
\end{pmatrix}
\begin{pmatrix}
\Cvvec{V}^a \\
\Cvvec{V}^b
\end{pmatrix}, \qquad
\text{with }
\Mmat{U}^x= \begin{pmatrix}
\Mmat{I}_{3\times3} , -\vec{r}^{\,x\,\times}_{1}/l \\
\vdots \\
\Mmat{I}_{3\times3} , -\vec{r}^{\,x\,\times}_{N_x}/l
\end{pmatrix},
\quad \text{for } x=a,b,
\label{eq:U_velocity}
\end{equation}
where the matrix $\vec{y}^\times$ obtained from the vector $\vec{y}$ is
defined as $(\vec{y}^\times)_{ij}=\epsilon_{ikj}y_k$.
Each stokeslet force is proportional to the relative velocity of the sphere
that it represents, with respect to the flow around it as created by
the other stokeslets. This gives a linear relation between the
stokeslets and the velocities of the spheres~\footnote{More
explicitly, consider the stokeslet at position $\vec{r}^a_n$. The
flow at that point which is created by the other stokeslets,
belonging to the two objects, is $\vec{u}(\vec{r}^a_n)=\Sigma_{m\neq
n} \Mmat{G}(\vec{r}^{\,a}_{n}-\vec{r}^{\,a}_{m}) \cdot
\vec{F}^{\,a}_{m} + \Sigma_{m} \Mmat{G}
(\vec{R}+\vec{r}^{a}_{n}-\vec{r}^{b}_{m}) \cdot \vec{F}^{\,b}_{m}$.
The stokeslet at that point is proportional to the velocity of the
sphere relative to the local flow, $\vec{F}^a_{n}=\gamma
\left(\vec{v}^a_{n}-\vec{u}(\vec{r}^a_n) \right)$. This gives
Eq.~(\ref{eq:linear_L}).},
\begin{equation}
\begin{pmatrix}
\bvec{v}^a \\
\bvec{v}^b
\end{pmatrix} =
\begin{pmatrix}
\Mmat{L}^{aa} & \Mmat{L}^{ab}\\
{\Mmat{L}^{ab}}^{T} & \Mmat{L}^{bb}
\end{pmatrix}
\begin{pmatrix}
\bvec{F}^a \\
\bvec{F}^b
\end{pmatrix}, \quad \text{where }
\label{eq:linear_L}
\end{equation}
\begin{eqnarray}
(\Mmat{L}^{xx}_{nm})_{ij} &=&
\left\{
\begin{array}{l l}
\Mmat{G}_{ij}(\vec{r}^{\,x}_{n}-\vec{r}^{\,x}_{m}) & \quad \text{if } n \neq m \\
\gamma^{-1} \delta_{ij} & \quad \text{else }
\end{array}
\right. \nonumber\\
(\Mmat{L}^{ab}_{nm})_{ij} &=&
\Mmat{G}_{ij} (\vec{R}+\vec{r}^{a}_{n}-\vec{r}^{b}_{m}),
\label{eq:Lblocks}
\end{eqnarray}
and $\gamma=6\pi\rho/ l$.
First we find the pair-mobility matrix as a generalization
of the analysis in Ref.~\citenum{Krapf_etal2009}. The sum of the stokeslets and the
corresponding total torque must be equal to the external generalized forces
applied on the objects. In a matrix form we can write
\begin{equation}
\begin{pmatrix}
\Cvvec{F}^a \\
\Cvvec{F}^b
\end{pmatrix}=
\begin{pmatrix}
(\Mmat{U}^{a})^T & 0\\
0 & (\Mmat{U}^{b})^T
\end{pmatrix}
\begin{pmatrix}
\bvec{F}^a \\
\bvec{F}^b
\end{pmatrix} .
\label{eq:U_force}
\end{equation}
Using Eqs.~(\ref{eq:U_velocity}) and (\ref{eq:linear_L}), we have
$$
\begin{pmatrix}
\Mmat{U}^{a} & 0\\
0 & \Mmat{U}^{b}
\end{pmatrix}^T
\cdot
\begin{pmatrix}
\Mmat{L}^{aa} & \Mmat{L}^{ab}\\
{\Mmat{L}^{ab}}^{T} & \Mmat{L}^{bb}
\end{pmatrix}^{-1}
\cdot
\begin{pmatrix}
\Mmat{U}^{a} & 0\\
0 & \Mmat{U}^{b}
\end{pmatrix}
\cdot
\begin{pmatrix}
\Cvvec{V}^a \\
\Cvvec{V}^b
\end{pmatrix}=
\begin{pmatrix}
\Cvvec{F}^a \\
\Cvvec{F}^b.
\end{pmatrix}
$$
From this expression we identify the pair-mobility matrix as
\begin{equation}
\Mmat{M}=\left[\begin{pmatrix}
\Mmat{U}^{a} & 0\\
0 & \Mmat{U}^{b}
\end{pmatrix}^T
\cdot
\begin{pmatrix}
\Mmat{L}^{aa} & \Mmat{L}^{ab}\\
{\Mmat{L}^{ab}}^{T} & \Mmat{L}^{bb}
\end{pmatrix}^{-1}
\cdot
\begin{pmatrix}
\Mmat{U}^{a} & 0\\
0 & \Mmat{U}^{b}
\end{pmatrix} \right]^{-1}.
\label{eq:fullM}
\end{equation}
This expression allows to calculate the pair-mobility matrix,
with the help of Eqs. (\ref{eq:U_velocity}) and (\ref{eq:Lblocks}),
based on the stokeslets' configuration and the Oseen tensor alone.
Next, we derive the $\Phi^x$ tensor of a stokeslet object $x$.
From this tensor we may readily obtain the second multipole of the pair
interaction (cf.\ Sec.\ \ref{sec:multipole}).
The force dipole around the origin
of a forced object is given by [Eq.\ (\ref{eq:V2a})], $(\bvec{r}\bvec{F})^x
\equiv (\Phi^x)^T \cdot \Cvvec{F}^x$. Similar to the $\Mmat{U}^x$
matrix relating the stokeslets to the total generalized force,
$\Cvvec{F}^x=(\Mmat{U}^x)^T \cdot \bvec{F}^x$, we define a tensor of
rank 3, $\Upsilon^x$, which relates the stokeslet forces to the total force
dipole on the object by $(\bvec{r}\bvec{F})^x=(\Upsilon^x)^T \cdot
\bvec{F}^x$. (Note that no force dipole
is applied on the individual stokeslets; being arbitrarily small they possess only
a force monopole.)
Specifically, it is made of $N$ blocks of $3\times
3\times 3$, given by $(\Upsilon_n)_{ijs} = r_{n,s}\delta_{ij}$,
$n=1\dots N$, $i,j,s=1,2,3$ (i.e., $r_{n,s}$ is the $s$ Cartesian
coordinate of the stokeslet $n$). Using Eqs.~(\ref{eq:U_velocity})
and (\ref{eq:linear_L}), we have $(\bvec{r}\bvec{F})^a=(\Upsilon^a)^T
\cdot (\Mmat{L}^{aa})^{-1} \cdot \Mmat{U}^a \cdot \Cvvec{V}^a$. This
implies $(\Phi^x)^T = (\Upsilon^x)^T \cdot (\Mmat{L}^{xx})^{-1} \cdot
\Mmat{U}^x \cdot \Mmat{M}_{\text{self}}^x$. Recalling that the
matrices $\Mmat{M}_{\text{self}}$ and $\Mmat{L}$ are symmetric, we
finally get
\begin{equation}
\Phi^x = \Mmat{M}_{\text{self}}^x \cdot
(\Mmat{U}^x)^T \cdot (\Mmat{L}^{xx})^{-1} \cdot \Upsilon^x.
\label{eq:stokeslets_Phi}
\end{equation}
It is important to note that in the above derivation we compute
$\Mmat{M}$ and $\Phi$ under the assumption that, for each object, the
stokeslet sizes are arbitrarily small compared to the distances
between them, $\rho \ll l$
(where $l$ is the object's radius of gyration).
However, in a more general case one can
use the Rotne-Prager-Yamakawa
tensor~\cite{Rotne&Prager1969,Yamakawa1970}, which corrects the Oseen
tensor for force distributions with finite
size~\cite{Carrasco&Torre1999}.
\subsection{Numerical Time Integration}
\label{sec:numerical}
We present a numerical integration scheme for the dynamics of two
stokeslet objects. We should first define the reference frames used in
the scheme. Each rigid object is characterized by axes which are
affixed and rotate with it. We define the object reference frame (ORF)
such that its $z$ axis coincides with the object's alignment axis (the
corresponding eigenvector of the $\Mmat{T}$-matrix). The other two axes
are selected arbitrarily. The $z$ axis of the laboratory frame is
defined along the forcing direction. During the evolution we
follow the translation and rotation of the ORF in the laboratory
frame.
We represent the orientation of an object by the Euler{-}Rodrigues
4-parameters \cite{Favro1960}, defined by
$(\Gamma,\vec{\Omega}) \equiv
(\cos\frac{\theta}{2},\hat{n}\sin\frac{\theta}{2})$, where $\hat{n}$
and $\theta$ are the axis and angle of rotation \footnote{This is the
same as the unit-quaternion representation \cite{Favro1960}.}. The
following properties hold for this 4-parameter representation: (a) The
norm of $(\Gamma,\vec{\Omega})$ in 4D-space is unity,
$\Gamma^2+\Omega^2=1$. (b) A rotation matrix is given by Rodrigues'
rotation formula,
\begin{equation}
R(\Gamma,\vec{\Omega})= \Mmat{I}_{3 \times 3} +2\Gamma \vec{\Omega}^{\times}+
2 ( \vec{\Omega}^{\times} )^2.
\end{equation}
(c) The parameters are invariant under inversion, i.e.,
$(\Gamma,\vec{\Omega})$ and $(-\Gamma,-\vec{\Omega})$ correspond to
the same orientation. (d) Given the angular velocity of the object,
the dynamics of its orientation simply reads
\begin{equation}
\begin{pmatrix}
\dot{\Gamma} \\
\dot{\vec{\Omega}}
\end{pmatrix}=
\frac{1}{2}
\begin{pmatrix}
0 & -\vec{\omega}^T\\
\vec{\omega} &\vec{\omega}^{\times}
\end{pmatrix}
\begin{pmatrix}
\Gamma \\
\vec{\Omega}
\end{pmatrix}.
\label{eq:angular}
\end{equation}
Since we choose the ORF such that the $z$-axis is the axis of
alignment, the terminal orientation of an axially alignable object under
a constant force along the $z$-axis is
$(\Gamma,\vec{\Omega})=(\cos(\frac{\omega t + \alpha}{2} ),\hat{z}
\sin(\frac{\omega t+ \alpha}{2}) )$, where $\alpha$ is a constant phase
which depends on the object's initial orientation at time $t=0$.
The state of a pair of objects at time $t$ is described by the
position of the origins of the objects, $\vec{R}^a(t)$ and
$\vec{R}^b(t)$, and their orientation parameters, $(
\Gamma^a(t),\vec{\Omega}^a(t) )$ and $( \Gamma^b(t),\vec{\Omega}^b(t)
)$. We time-integrate from the initial state,
$\vec{R}^a_{0}=(0,0,0)$, $\vec{R}^a_0-\vec{R}^b_{0}=\vec{R}_0$, $(
\Gamma^a_{0},\vec{\Omega}^a_{0} )$ and $(
\Gamma^b_{0},\vec{\Omega}^b_{0} )$, as follows. Given the positions
of the stokeslets at time $t$, the pair-mobility matrix,
$\Mmat{M}(t)$, is calculated as explained in
Sec.\ \ref{sec:stokeslet_properties}, either exactly or using the
multipole approximation. Then, the linear and angular velocities of
the objects are given by $ (\Cvvec{V}^a (t),\Cvvec{V}^b (t))^T =
\Mmat{M}(t) \cdot (\Cvvec{F}^a(t),\Cvvec{F}^b (t) )^T$. From them
the origins and orientations of the objects at time $t+dt$ are derived
according to
\begin{eqnarray}
\vec{R}^x(t+dt)&=&\vec{R}^x(t)+\vec{V}^x(t)dt
\\
\begin{pmatrix}
\Gamma^x(t+dt) \\
\vec{\Omega}^x(t+dt)
\end{pmatrix} &=&
\exp \left[ \frac{dt}{2}
\begin{pmatrix}
0 & -\vec{\omega}^{xT}\\
\vec{\omega}^x &\vec{\omega}^{x\times}
\end{pmatrix}
\right]
\begin{pmatrix}
\Gamma^x(t) \\
\vec{\Omega}^x(t)
\end{pmatrix}
\end{eqnarray}
for $x=a,b$. During the evolution we make sure that the small
stokeslet spheres do not overlap, and that the pair mobility matrix
remains positive-definite. In practice we never encountered such
problems when using the exact pair mobility matrices; when it did
happen in the case of the multipole approximation we stopped the
integration.
We define a scalar order parameter which characterizes the degree of
mutual alignment of the two objects,
\begin{equation}
m \equiv \left[ ( \Gamma^a,\vec{\Omega}^a) \cdot (
\Gamma^b,\vec{\Omega}^b) \right]^2= \left( \Gamma^a\Gamma^b +
\vec{\Omega}^a \cdot \vec{\Omega}^b \right)^2.
\label{eq:phi}
\end{equation}
As required, the order parameter is invariant under 3-rotation. This
can be verified by explicitly applying a 3-rotation to the laboratory
frame, or alternatively, by the following argument. Since 3-rotation
leaves the norm of the 4-parameter orientation unchanged (property (a)
above), it is a unitary transformation in 4-space. Hence, the dot
product is invariant. When the objects are aligned, $ (
\Gamma^a,\vec{\Omega}^a)=\pm ( \Gamma^b,\vec{\Omega}^b)$, and $m=1$;
otherwise $0 \leq m <1$. In the case of partial alignment,
$m=\cos^2(\frac{\Delta\alpha}{2})$, where $\Delta\alpha$ is the mutual
phase difference~\footnote{If the symmetry of the objects is such that
their phase difference is unobservable (e.g., two ellipsoids
rotating around their major axis), then we set it to zero.}.
Another scalar property of the two-object system is the energy dissipation rate.
At time $t$, the latter is given by
$\Cvvec{V}^a(t)\cdot\Cvvec{F}^a(t)+\Cvvec{V}^b(t)\cdot\Cvvec{F}^b(t)$.
Since the pair-mobility matrix is positive definite the energy dissipation of
the driven pair is positive at all times.
\section{Numerical Results: Effect on Alignment}
\label{sec:alignment}
We present in Figs.\ \ref{fig:geometry:t}--\ref{fig:dipole} several examples
for the numerically integrated evolution of object pairs under various conditions.
One can be immediately appreciate the diversity of possible trajectories.
To make your way through this richness it is important to make two distinctions between
types of trajectories. The first distinction is between
constant forcing (as in sedimentation), which can make the objects only partially aligned
without synchronizing their phases of rotation~\cite{Gonzalez_etal2004,Krapf_etal2009},
and a time-dependent forcing protocol, which is known to lock the phase of an individual object onto
that of the force~\cite{Moths&Witten2013,Moths&Witten2013b}.
The main issue examined below is how the presence of hydrodynamic interaction
affects these two behaviors.
The second distinction, therefore, is whether hydrodynamic interactions are included
(dashed, dotted and dash-dotted/colored curves) or turned off (solid gray curves).
In the absence of hydrodynamic interactions
(or when they get weak as the objects move far apart), the time-dependent aligning force
will make the objects fully synchronized, whereas under constant forcing
the objects will generally become unaligned.
The results are presented in a dimensionless form, using units such that
$\eta=|\omega_0|=1$ and $\rho=0.01$. The distances between the stokeslets
of each object are taken randomly between $0$ and $1$; hence, $\rho \ll
l\sim 1$. The time-dependent forcing protocol is $\vec{F} = F_0
\left(-\sin(\omega_0t)\sin(\theta),\cos(\omega_0t)\sin(\theta),-\cos(\theta)
\right)$, where $\theta=0.1\pi$, $F_0=-|\lambda_3|^{-1}$,
$\omega_0=\text{sign}(\lambda_3)$ and $\lambda_3$ is the real eigenvalue of
the single-particle twist matrix. We examine both the trajectories of
the separation vector connecting the origins of the two objects, and
the corresponding evolution of the orientation order parameter.
We begin with the case of a time-dependent forcing,
Figs.\ \ref{fig:geometry:t} and \ref{fig:m:t}. The first observation,
most clearly demonstrated in Fig.\ \ref{fig:m:t}(b), is
that hydrodynamic interaction degrades the alignment of the two
objects, as has been rigorously inferred based on symmetry
considerations in Sec.\ \ref{sec:symmetry}. Another conclusion,
supported by additional examples not shown here, is that most objects,
which start sufficiently far apart, especially if they start fully
aligned, tend to repel each other (Fig.\ \ref{fig:geometry:t}). Even
if they are not fully aligned, the growing distance and weakening
interaction make them individually more aligned with the
forcing, and therefore also mutually synchronized.
Thus, the repulsion helps restore the alignment at long
times. The increasing separation occurs in the $xy$ plane, while along
the $z$ axis the separation decreases and saturates to a finite
distance, dependent on initial conditions, see Fig.\ \ref{fig:geometry:t}.
The repulsion is accompanied by a decrease in dissipation rate (up to small
oscillations), as demonstrated in Fig. \ref{fig:dissip}. When the HI
is turned off, the dissipation rate reaches a constant value as the
two independent objects set into their ultimate aligned state (solid
curves in Fig.\ \ref{fig:dissip}).
The repulsive effect is observed for most examples of our randomly generated
pairs of objects but is not a general law.
For instance, when the objects start at a sufficiently small separation,
some pairs remain ``bound'' in a limit cycle, oscillating about a
certain mean separation and mean orientational alignment, as
demonstrated by the green/dashed curves in Figs.\ \ref{fig:geometry:t} and \ref{fig:m:t}.
In Figs.\ \ref{fig:geometry:const} and \ref{fig:m:const} we examine
the same properties under constant forcing. The two effects--- degradation
in the alignment of a pair which is initially fully aligned, and
mutual repulsion--- are observed here as well. Yet,
in the absence of a time-dependent aligning force, as the
two objects move apart, alignment is not restored.
At long times, and for the common case of repulsion,
we distinguish between two observed behaviors:
a) The order parameter continues to change without saturating to a constant value
(e.g., red/dash-dotted curve and cyan/dotted curve in Figs.~\ref{fig:m:const}(a) and \ref{fig:m:const}(b),
respectively).
This non-intuitive result can be explained as follows. The fact that the interaction becomes weak does not
necessarily imply that the accumulation of phase difference stops. If the two distant objects are partially aligned
we have $m\simeq \cos^2(\delta\omega_z t/2)$, where $\delta\omega_z$ is the difference between the objects' angular velocity
along the alignment axis. Hence, if the decay of $\delta\omega_z$ with time is slower then $t^{-1}$ then
phase difference will continue to accumulate. This depends on the detailed dynamics of repulsion which will be
addressed in publication II.
b) The other option is that $m$ converges to some value dependent on the initial state, with no particular chosen $m$
(green/dash-dotted curve in Fig.~\ref{fig:m:const}(b) and
cyan/dash-dotted curve in Fig.~\ref{fig:m:const}(c)),
i.e., the two objects continue to rotate with a fixed relative orientation.
In the examples that we checked there seems to be a tendency toward ultimate anti-alignment ($m=0$).
Therefore, we also checked the stability of anti-alignment
in pairs which start from such a state.
Fig.~\ref{fig:m:const}(d) examines the stability
of this configuration for objects initially confined to the $xy$ plane (perpendicular to the force).
Whereas the aligned pair (blue/dotted curve) is unstable, the anti-aligned one (red/dashed curve) remains stable
for the duration of integration.
It may well be that this stability survives for a long but finite time, see e.g., dark red/dotted curve
in Fig.~\ref{fig:m:const}(c).
In addition, a separation of the pair along the $z$-axis destabilizes an anti-aligned pair as well (examples not shown). Finally, we note that even if the final phase difference were arbitrary and uniformly distributed, the value of $m$ would be evenly distributed around $1/2$ but non-uniformly, with larger weights on $m=0,1$.
(This follows from the definition of $m$, see Eq.\eqref{eq:phi}.)
Figure \ref{fig:dipole} compares results obtained using the full
pair-mobility matrix of the stokeslet objects with those obtained from
the multipole (dipole) approximation. As expected, the two
calculations agree for objects whose mutual distance increases with
time, and disagree for objects whose trajectories reach close
proximity.
Further investigation (not shown here) of the orientational
dynamics of the objects suggests a possible explanation for the
characteristic repulsion between two chiral objects. In the
absence of HI, each object rotates along $\hat{F}$ and translates
on average along $\hat{F}$. One contribution to the dipolar term of the HI
comes from the effect on each object by the vorticity of the Oseen flow
caused by the other object. This perturbative angular velocity is along an axis
which is perpendicular to the separation vector and the force,
$\hat{\omega}^a_{\rm flow} \propto -\hat{R} \times \hat{F}$ and
$\hat{\omega}^b_{\rm flow} \propto \hat{R} \times \hat{F}$. The
competition between this rotation and the aligning self-response of
each individual object results in an inclination of the
two objects relative to their non-interacting state. This
inclination alters the average unperturbed linear velocity of the object
by a small rotation about the $\hat{R} \times \hat{F}$ direction---
counter-clockwise for object $a$ and clockwise for object $b$.
Hence, the two objects glide away from each other,
$\dot{R}^2=2(\vec{V}^a-\vec{V}^b)\cdot\vec{R} \propto
((-\hat{R} \times \hat{F})\times\hat{F})\cdot\vec{R}
= R(1-(\hat{R}\cdot\hat{F})^2)\geq0$,
where the proportionality constant is positive, i.e., the separation increases with time
(unless $\vec{R}\parallel\vec{F}$, for which the whole argument does not hold).
\begin{figure}
\centerline{\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=55 14 345 292]{fig2a}}
\hspace{2cm}
\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=55 14 345 292]{fig2b}}}
\caption[]{Two examples of axially alignable stokeslet objects, which were used in the simulations.
The objects comprise four stokeslets connected by dragless rods. The origin of the objects is
at point (0,0,0) and the aligning direction is $-\hat{z}$.
The object on the left corresponds to the dark red/dotted trajectories in the left panels of
Figs.~\ref{fig:geometry:const} and~\ref{fig:m:const}, and
the one on the right corresponds to the purple/dashed trajectories in the right panels of
Figs.~\ref{fig:geometry:t} and~\ref{fig:m:t}.}
\label{fig:objects}
\end{figure}
\begin{figure}
\centerline{\resizebox{0.4\textwidth}{!}{\includegraphics[viewport=14 2 389 302]{fig3a}}
\hspace{0.55cm}
\resizebox{0.4\textwidth}{!}{\includegraphics[viewport=14 2 380 304]{fig3b}}}
\vspace{0.3cm}
\centerline{\resizebox{0.4\textwidth}{!}{\includegraphics[viewport=51 0 346 301]{fig3c}}
\hspace{0.55cm}
\resizebox{0.4\textwidth}{!}{\includegraphics[viewport=45 0 346 301]{fig3d}}}
\vspace{0.3cm}
\centerline{\resizebox{0.4\textwidth}{!}{\includegraphics[viewport=11 2 389 302]{fig3e}}
\hspace{0.55cm}
\resizebox{0.4\textwidth}{!}{\includegraphics[viewport=13 4 388 302]{fig3f}}}
\caption[]{Trajectories of object separation under time-dependent
forcing. The three rows, from top to bottom, correspond,
respectively, to the separation along the $z$ direction, its
projection onto the $xy$ plane, and its total magnitude. The squares
in the middle row indicate the state at the end of the
simulation. The panels show results for three different
objects, starting from either a random mutual orientation (left
column) or their fully aligned state (right column).
The green/dashed trajectory on the right panels was integrated
longer than 150 time units to verify that it continues in a limit cycle.}
\label{fig:geometry:t}
\end{figure}
\begin{figure}
\centerline{\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=7 2 390 300]{fig4a}}
\hspace{0.35cm}
\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=7 2 380 302]{fig4b}}}
\caption[]{Orientation order parameter as a function of time, under
time-dependent forcing, for the examples of
Fig.\ \ref{fig:geometry:t}. (a) results for
random initial orientations (examples on the left column of
Fig.\ \ref{fig:geometry:t}); the additional solid gray curves
correspond to non-interacting objects. (b)
results for initially fully aligned object pairs (right column in
Fig.\ \ref{fig:geometry:t}).}
\label{fig:m:t}
\end{figure}
\begin{figure}
\centerline{\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=14 2 389 302]{fig5a}}
\hspace{0.55cm}
\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=14 2 389 304]{fig5b}}}
\vspace{0.3cm}
\centerline{\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=45 0 343 301]{fig5c}}
\hspace{0.55cm}
\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=45 0 346 301]{fig5d}}}
\vspace{0.3cm}
\centerline{\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=11 2 390 303]{fig5e}}
\hspace{0.55cm}
\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=13 4 389 302]{fig5f}}}
\caption[]{Trajectories of particle separation under constant
forcing. The meaning of the various panels is the same as in
Fig.\ \ref{fig:geometry:t}.
In all the examples shown here, the two objects repel each other except of the example
which corresponds to the blue/dashed curve in the left panels.
(The repulsive trajectories were actually integrated to times longer than presented here.).
}
\label{fig:geometry:const}
\end{figure}
\begin{figure}
\centerline{\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=7 2 380 302]{fig6a}}
\hspace{0.35cm}
\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=7 2 380 305]{fig6b}}}
\vspace{0.3cm}
\centerline{\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=7 2 380 302]{fig6c}}
\hspace{0.35cm}
\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=7 2 390 301]{fig6d}}}
\caption[]{Orientation order parameter as a function of time, under
constant forcing, for the examples of
Fig.\ \ref{fig:geometry:const}.
(a) results for
random initial orientations (examples on the left column of
Fig.\ \ref{fig:geometry:const}); the solid gray curves correspond
to non-interacting objects. (b) results for
initially fully aligned object pairs (right column in
Fig.\ \ref{fig:geometry:t}). (c)
results for objects with initial partial alignment (rotating around
the same axis with random initial phases).
(d) the stability of anti-alignment;
shows trajectories of two identical pairs, which start on the
$xy$ plane from the same separation and axes of rotation but with different relative phases.
Blue/dotted and red/dashed curves represent, respectively, a pair which starts aligned (zero relative phase)
and one which starts anti-aligned (relative phase of $\pi$).}
\label{fig:m:const}
\end{figure}
\begin{figure}
\centerline{\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=2 2 380 301]{fig7a}}
\hspace{0.35cm}
\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=2 2 380 301]{fig7b}}}
\caption[]{Dissipation rate as a function of time for object pairs
starting from arbitrary orientations, under time-dependent forcing
(a) and constant forcing (b). Dash-dotted and dotted colored curves correspond to the
examples of the same styles/colors in the preceding figures. Solid curves
show the results in the absence of HI.}
\label{fig:dissip}
\end{figure}
\begin{figure}
\centerline{\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=11 2 389 302]{fig8a}}
\hspace{0.35cm}
\resizebox{0.45\textwidth}{!}{\includegraphics[viewport=11 0 389 302]{fig8b}}}
\caption[]{Comparison between the evolution of pair separations
obtained using the full pair-mobility matrix
(dashed, dotted and dash-dotted colored curves) and its
multipole approximation (solid curves). Each panel
presents three examples of pairs under time-dependent (a)
and constant forcing (b). All pairs start from a fully aligned
state. The multipole approximation includes the monopolar and
dipolar terms.}
\label{fig:dipole}
\end{figure}
\section{Discussion}
\label{sec:discuss}
Irregular objects display rich dynamics
already at the level of a pair of objects, as has been demonstrated
above. In the present work we have focused on the effect of the
hydrodynamic interaction on the orientational alignment of such object pairs.
The hydrodynamic interaction, in general, degrades the alignment.
We have rigorously proven the instantaneous linear degradation for fully aligned objects at large mutual distances.
In other circumstances, such as nearby or unaligned objects, the hydrodynamic interaction
may have an opposite effect.
The leading degradation effect in distance is dipolar rather than monopolar; yet, it
is significant\,---\,a large mutual distance (compared to the object
size) is required to make the degradation negligible. More
quantitatively, the degradation will be significant when the
perturbation to the angular velocity due to HI, $\delta\omega$,
becomes comparable to the inverse of the time required to align a
single object. The unperturbed angular velocity is given by $\omega_0
= \Mmat{T}_{\rm self}F$. The dimensionless eigenvalue of the
self-twist matrix is generally found to be about an order of magnitude
smaller than the dimensionless self-mobility coefficient
\cite{Moths&Witten2013,Morozov&Leshansky2014,Moths&Witten2013b}, i.e.,
$\omega_0\sim 0.1 F/(8\pi
l^2)$. As presented in Sec.\ \ref{sec:symmetry}, $\delta\omega\sim
\Mmat{T}^{ab}F \sim F/(8\pi l^2) (l/R)^2$.
The alignment time is typically
$t_{\rm al} \sim 10/\omega_0$ (see Fig\ \ref{fig:m:t}). Hence, the
degree of degradation is $t_{\rm al}\delta\omega \sim 10^2(l/R)^2$.
The conclusion is that the separation between the objects should be larger
than ten times their size to maintain alignment. In the case of many
objects this implies a maximum volume fraction $(l/R)^3 \sim 10^{-3}$.
At the same time, as shown in Sec.\ \ref{sec:alignment}, for most of our randomly generated
pairs of objects,
the hydrodynamic interaction makes the rotating objects repel each other.
As a result, at long times the hydrodynamic interaction usually becomes negligible
and each of the objects gets aligned again with the time-dependent force.
In that section we also provided a possible explanation for this
repulsion, related to the mutual rotation of the two objects
which causes them to glide away from each other.
In fact, the objects need not be irregular to exhibit this gliding effect;
two forced ellipsoids which start parallel to one another
will experience the same repulsion~\cite{Claeys&Brady1993,Kutteh2010,Kim1986}.
The resulting hydrodynamic ``pseudo-potential'' \cite{Squires&Brenner2000,Squires2001}
will be addressed in a future publication.
For the case of a finite number of objects
the repulsion will help restore the alignment as the objects drift
apart.
It should be kept in mind, however, that the repulsion is not a general law.
We observed it for a few dozens pairs of stokeslet objects. As mentioned above,
it also holds for a pair of well separated ellipsoids. Yet, a few counter-examples
have been also provided in Sec~\ref{sec:alignment}.
An interesting counterpart of the effects discussed here is found in the interaction between a forced
object and a nearby wall \cite{Happel&Brenner,Russel_etal1977}.
The wall can be represented by an image (though not identical) object forced in
the {\em opposite} direction~\cite{Pozrikidis}.
As a result, the object will rotate and, if it is non-spherical, also glide {\em toward} the wall, as was indeed shown
for a rod falling near a wall~\cite{Russel_etal1977}.
Obviously, the interaction of an alignable object with a wall will also degrade the alignment.
An important distinction between regular and irregular objects,
which we have not dealt with here, concerns many-body interactions in forced systems.
A pair of forced spheres does not develop any
relative translational velocity \cite{Happel&Brenner}.
The same holds
for a pair of forced uniform ellipsoids to order $1/R^3$ (for an ellipsoid,
the components of $\Phi$ which correspond to the translational velocity vanish
\cite{BrennerIV}).
For a suspension of many objects this implies that two-body effects on
relative motion are either absent (spheres) or negligible at low
volume fraction (ellipsoids). By contrast, as we have shown here, a
pair of irregular objects develops a relative velocity
already at order $1/R^2$, which should lead to significant two-body interactions in
a suspension. This may bring about qualitative differences between driven
suspensions of regular and irregular objects in relation to such
phenomena as sedimentation.
This work shows that asymmetry in sedimenting objects leads to a wealth of hydrodynamic interaction effects not seen for spheres. This study was undertaken to assess how interactions disrupt the rotational synchronization of such objects. However it proves to have striking effects independent of this alignment. The prevalent repulsion, the occasional entrapment and the intricate quasiperiodic motions shown above are examples. These effects could have significant impacts on real colloidal dispersions, e.g., in fluidized beds of catalyst particles. Though we have studied only pairwise interactions between identical objects, many of these effects are expected to apply more generally. The general treatment of hydrodynamic interaction and its dependence on the
shape of the interacting objects, which we have developed here, should prove useful in exploring these phenomena.
Our work in progress aims to achieve a more general understanding of the rich behavior reported in Sec. \ref{sec:alignment}.
\begin{acknowledgments}
We thank Robert Deegan and Alex Leshansky for helpful discussions, and the James Franck Institute and
Tel Aviv University for their hospitality during part of this work.
This research has been supported by the US--Israel Binational Science Foundation
(Grant no.\ 2012090).
\end{acknowledgments}
|
\section{Introduction}
The leading theory of moon formation is a giant impact event occurring approximately $4.5 \times 10^{9}$ years ago \cite{Hartmann1975}. Recent large scale simulations show that giant impacts are capable of producing the correct Earth - Moon masses and angular momentum~\cite{Canup2012,Stewart2012}. Complicating the giant impact theory, however, is that the Earth and moon have a nearly identical chemical and isotopic composition~\cite{Wiechert2001}. This implies that either the impactor was compositionally similar to the proto-Earth, which is thought to be unlikely, or that extensive mixing of the post impact materials occurred. The simulations needed to test the post-impact mixing are dependent on an accurate understanding of the properties of mantle materials at extreme pressures and temperatures. For example, post impact mixing for chemical equilibration in the proto-Lunar disk has been shown in simulations~\cite{Pahlevan2007}, but requires melting and vaporization of the mantle in order for material to diffuse. Testing such theories is difficult because the melt line of the most common mantle materials is not well constrained at these conditions \cite{Boehler2000}.
Advanced facilities for performing dynamic materials experiments have greatly increased the pressure and temperature regimes that can be probed for important planetary materials \cite{MDKQuartz2009, Kraus2012, Dylan2012, RootCO2}. The Rankine-Hugoniot (RH) equations \cite{Zeldovich} relate the experimentally measured shock velocity to the thermodynamic state, pressure and density, of the shocked material. The ability to perform experiments with steady planar shocks and with well-characterized impactors and targets is critical for determining the equation of state (EOS) and the phase. To fully address the physics relevant to planetary science, this thermodynamic information must be augmented with an understanding of the phase transformations.
In this work we focus on MgO, the end member of magnesiow\"ustite, a major constituent of the earth's mantle \cite{McDonough1995} and likely other terrestrial exoplanets~\cite{kepler10b}. At ambient conditions, MgO exists in a NaCl (B1) lattice structure, which is stable over a wide pressure-temperature range~\cite{Zerr1994, DuffyDAC, Coppari2013}. Dynamic compression experiments starting from ambient~\cite{marsh1980lasl, VassiliouMgO, Duffy1994, Zhang2008, Fratanduono2013} and from T$_0$=1900~K~\cite{OlegSCCM09} show no indications of phase transitions up to 230~GPa. \textit{Ab initio} studies on the MgO phase diagram agree that three phases exist: the B1 solid, the B2 (CsCl) solid, and the liquid~\cite{BeloMgO, boates-bonev, cebulla-MgO}, but disagree on the location of the phase boundaries. Along the Hugoniot, which is relevant for planetary impact scenarios, the location of the B1-B2 transition and the melt transition has not been precisely determined. Recently McWilliams \textit{et al}. have shown that MgO can be dynamically compressed to pressures exceeding 1 TPa using a decaying shock technique~\cite{McWilliams2012}. The authors also proposed locations for the B1-B2 and B2-liquid transitions along the Hugoniot, but the measurements relied heavily on prior MgO Hugoniot data, which was not well known above 200~GPa.
We present a comprehensive study of the MgO Hugoniot using experiments, density functional theory (DFT), and quantum Monte Carlo (QMC) methods over a wide pressure range covering the B1, the B2, and the liquid phases from 0.27 to 1.2~TPa. The high-precision data constrain the Hugoniot at multi-Mbar pressures, and the DFT and QMC results further elucidate information on the phase boundaries. This work provides accurate EOS data at extreme conditions and furthermore reveals lower limits of the relative velocity required in the giant impact scenario for moon formation.
To attain planetary impact conditions, we performed a series of shock compression experiments using the Sandia Z-Machine \cite{ZMachRef}. The Z-machine is a pulsed power system capable of producing shaped current pulses and induced magnetic fields in excess of 20 MA and 10 MG respectively. The combined current and magnetic field densities generate magnetic pressures up to 650 GPa that can accelerate aluminum flyers up to 40~km/s \cite{LemkeZFlyer05}.
Figure~\ref{ExptApproach} shows a schematic view of the target geometry; a more detailed Z target geometry is found elsewhere~\cite{RootFoam2013}. An Al or Cu flyer plate is shocklessly accelerated toward the target stack consisting of a single-crystal MgO sample ([100], 300-500 $\mu$m, Asphera Corp., $\rho_0$ = 3.584 g/cm$^3$) backed by a window. Although the back side of the flyer is melted by the high current, the surface of the flyer impacting the sample typically has a few hundred microns of solid density material at impact~\cite{LemkeZFlyer05}. The velocity interferometer system for any reflector (VISAR) measures the flyer plate velocity ($V_F$) up to impact at the target (Fig.~\ref{ExptApproach}). Impact produces a steady shock in the MgO sample. At low impact velocities (and consequently low shock pressures), the MgO sample scatters light from the VISAR preventing direct measurement of the shock velocity. Instead, sharp fiducials are observed in the VISAR signal (Fig.~\ref{ExptApproach} inset) that correspond to impact and to shock transit into the quartz window. In this case, we calculated the MgO shock velocity ($U_S$) using the transit time determined from the fiducials and the measured thickness. At high impact velocities (and high shock pressures), the shock front in the MgO is reflective and the VISAR directly measures the shock velocity. As the shock transits into the quartz, the VISAR measures the quartz shock velocity directly. Multiple VISAR signals were recorded for each sample eliminating 2$\pi$ ambiguities and providing redundant measurements for improved precision. For directly measured velocities, the uncertainty is better than 1\% and for transit time measurements the uncertainty is on the order of 1-2\%.
\begin{figure}
\includegraphics[width=3.0in]{Figure01.eps}
\caption{The experimental target and representative VISAR data. The VISAR measures the flyer velocity as it approaches the MgO (grey line). For this low velocity impact, as the shock transits the MgO, the VISAR loses signal (red line). As the shock transits into the quartz, the VISAR signal returns and the quartz shock velocity is measured (blue line). The inset shows a VISAR raw signal. The fiducials correspond to impact and to shock transit into the rear window. }
\label{ExptApproach}
\end{figure}
Knowing the initial densities of the MgO and the flyer plate and measuring the flyer velocity and the MgO shock velocity, we calculate the MgO Hugoniot state density ($\rho$), pressure ($P$), and particle velocity ($U_P$). The Hugoniot state is determined using a Monte Carlo impedance matching analysis~\cite{RootCO2} to solve the RH equations~\cite{Zeldovich}. The Monte Carlo method accounts for the uncertainties in the experimental measurement and the Al and Cu Hugoniot standards. Further discussion on the impedance matching method along with the list of Hugoniot data are found in the supplement~\cite{supplemental-material}.
\begin{figure}
\includegraphics[width=3.3in]{Figure02.eps}
\caption{The MgO Hugoniot data in $\rho-P$ space from the Z experiments, previous experimental data\cite{marsh1980lasl, VassiliouMgO, Duffy1994, Zhang2008, McWilliams2012, Fratanduono2013}, and the DFT simulation results from this work. The current data set shows a deviation from the extrapolation of the fit to the B1 MgO data from $<$~230~GPa.}
\label{rhoPdata}
\end{figure}
\begin{figure}
\includegraphics[width=3.3in]{Figure03.eps}
\caption{The experimental $U_S - U_P$ data including results from previous experiments from Refs~{\cite{marsh1980lasl, VassiliouMgO, Duffy1994, Zhang2008, McWilliams2012, Fratanduono2013}}. The optimized linear fits are also plotted. The dashed vertical lines indicate the optimized phase boundaries along the principal Hugoniot and the shaded cyan regions indicate the uncertainty. }
\label{UsUpdata}
\end{figure}
Figure~\ref{rhoPdata} plots the experimental and DFT principal Hugoniot in $\rho$-P space. The Z experimental data span the range from 0.27~TPa up to 1.2~TPa - the highest, directly measured Hugoniot states attained in MgO. Also included are the DFT simulation results for the B1, B2, and liquid phases of MgO (discussed later). Although the VISAR diagnostic does not give direct information about the MgO phase upon shock compression, we can infer phase transitions given our data. Figure~\ref{rhoPdata} shows an extrapolation of the linear fit to the $U_S - U_P$ data for B1-phase Hugoniot states $<$~230~GPa (converted to $\rho$-P using the RH equations) determined from the previous experiments~\cite{marsh1980lasl, VassiliouMgO, Duffy1994, Zhang2008, Fratanduono2013}. Below $\approx$ 360~GPa, the Z experimental data are consistent with the gas-gun data but above 360~GPa they deviate from the extrapolation. This suggests that the B1 phase is stable up to 360~GPa and likely undergoes a phase transition from the B1 state to another phase, presumably the B2 state, at that shock pressure. At higher pressures ($\approx > $700~GPa) the shock front in the MgO was reflective. From the reflective shock front, we infer that the MgO has melted into a conductive fluid, similar to what is observed for quartz~\cite{MDKQuartz2009}.
The interpretation for the MgO phase along the Hugoniot is not as obvious between the low pressure B1 phase and the high pressure liquid phase. Prior \textit{ab initio} calculations suggest that MgO transitions from B1 to B2 and then from B2 to liquid with a small B1-B2 coexistence region and a potentially large coexistence region between the B2 and liquid phases \cite{cebulla-MgO}. To investigate the phase regions, we analyze the $U_S - U_P$ data using a Monte Carlo optimization (MCO) method similar to the method used in previous work on carbon\cite{knudson-carbon}.
Slope changes in the $U_S - U_P$ data often indicate phase transitions and phase boundaries. Using the MCO method, we fit four lines to the experimental $U_S - U_P$ data. The $U_S-U_P$ data were converted to a ``cloud'' of points (as described in Ref.~\citenum{knudson-carbon}), allowing region boundaries to move smoothly during optimization. For a particular set of data clouds, the eleven parameters (four slopes, four intercepts, and three region boundaries) of a four-segment curve were obtained by minimizing the square minimum distances to each cloud point. Revised clouds were generated by randomly drawing a new center for each cloud. Optimization was repeated ($\approx10000$ times) using the revised clouds to characterize the distributions of the parameters. The parameters are listed in the supplement~\cite{supplemental-material}. It is important to note that this analysis is only possible because of the high precision data produced from the steady shocks.
Figure~\ref{UsUpdata} shows the compiled experimental $U_S - U_P$ Hugoniot data, the four linear fits, and the phase regions determined from the MCO method. Following the literature~\cite{cebulla-MgO} and our DFT results, we propose the four regions be classified as follows: 1. The B1 solid from ambient to 363~GPa; 2. The B2 solid from 363 to 462~GPa; 3. The B2-liquid coexistence region between 462 and 620~GPa; and 4. The liquid state above 620~GPa. However, as our continuum level experiments do not provide microstructure information, we performed \textit{ab initio} calculations of the Hugoniot and the phase diagram to better understand the high pressure states of MgO.
The high precision requirements of this work necessitated refinements of previous \textit{ab initio} methods \cite{BeloMgO,boates-bonev,cebulla-MgO}. We performed calculations utilizing DFT and QMC focusing on the solid-solid phase transformation from B1 to B2 and the melting of MgO along the Hugoniot, presumably from the B2 phase. Using DFT to calculate the Hugoniot requires prior knowledge of the phase state, so we first calculated the phase diagram. We used a three-part approach to determine the phase boundaries. In order to determine the melt boundary from both the B1 and B2 phases, we performed two-phase calculations of melting using VASP 5.2.11\cite{VASPshort,PAWshort}; further details are presented in the supplemental material~\cite{supplemental-material}. To determine the solid-solid phase boundaries we decomposed the solid's Helmholtz free energy into two pieces.
\begin{equation}
F_{sol}(V,T) = E(V) + F_{vib}(V,T)
\label{free-energy-decomp}
\end{equation}
The first piece is the density dependent energy of either the B1 or B2 phase. This is calculated via diffusion QMC using \textsc{qmcpack}~\cite{qmcpack} following methodology detailed in Ref.~\citenum{shulenburger-prb} with particular concern paid to the construction of pseudopotentials. The second piece of the free energy is due to the finite temperature motion of the ions and electrons and is calculated in two parts. First the harmonic part of the free energy is calculated using the finite displacement method as implemented in the \textsc{phon} code~\cite{Alfe-phon}. The quasiharmonic approximation (QHA) is known to break down as temperatures increase and this is particularly true for MgO~\cite{wu-anharmonic-prb}. For this reason and because the Hugoniot is expected to cross the phase boundary relatively close to the melt line, we have augmented our QHA calculations of free energy with thermodynamic integration. This is performed by using
\begin{equation}
\Delta S = \int_{T_i}^{T_f} \frac{1}{T} \left(\frac{\partial E}{\partial T}\right)_V dT
\end{equation}
that allows the change in entropy along an isochore to be calculated directly in terms of the internal energy. The energy is calculated using DFT based quantum molecular dynamics (QMD) at points spaced by 250 K along several isochores in the region of the phase transition. Using entropy from the QHA calculation at low temperatures as a reference, we calculate the Gibbs free energy of both phases and determine the phase transition pressure directly. This method also determines the range of validity for the QHA. We find the range to be smaller than previously estimated\cite{BeloMgO} with significant deviations in the free energy occurring by 5000 K and 400 GPa. The positive effect of the anharmonic entropy was significantly larger in the B1 phase than in the B2 phase, moving the phase boundary to higher pressures at high temperature. Specific computational details are in the supplement~\cite{supplemental-material}.
Figure~\ref{PhaseDiagram} shows the DFT determined phase diagram and the calculated P-T states on the MgO Hugoniot. Using the DFT phase diagram, we can infer the microstructure during our plate impact experiments. However, the experimental Hugoniot data is an incomplete thermodynamic description due to the lack of temperature measurements. To assign the data to a given phase within the single phase regions, we use QMD calculations to provide the missing temperature by calculating the Hugoniot. For each of the three phases at given densities, we perform several QMD calculations at various temperatures. We then find the temperature for which the RH energy equation is satisfied to determine the shock state. Finally, the pressure and temperature of these shock states are compared to the phase boundaries to determine whether that state is thermodynamically stable. This approach also provides a validation of the QMD by comparing the calculated Hugoniot to the experimental results. The calculations and the experiments are in good agreement in $\rho-P$ space (Fig.~\ref{rhoPdata}) and in $P-T$ space (Fig.~\ref{PhaseDiagram}). The DFT calculated phase boundaries along the Hugoniot corroborate the MCO fitting method results for the experimental data suggesting the Hugoniot has four major regions: B1, B2, coexistence, and liquid. Table~\ref{PhaseBounds} lists the phase boundaries along the principal Hugoniot from the MCO method and the DFT simulations.
\begin{table}
\caption{Phase boundaries along the principal Hugoniot.}
\label{PhaseBounds}
\begin{tabular}{cccc}
\hline
Method & B1-B2 & B2-Coexist. & Coexist.-Liquid \\
& (GPa) & (GPa) & (GPa) \\
\hline
Expt. (MCO) & 363$\pm$6 & 462$\pm$20 & 620$\pm$17 \\
DFT & 330 & 475 & 600 \\
\hline \hline
\end{tabular}
\end{table}
Both the experimental and DFT results show that a minimum shock pressure of 620~GPa is required to achieve complete melting in MgO that is initially at ambient temperature. In the giant impact scenario, the proto-Earth has also been assumed to have an elevated surface temperature prior to the moon-forming event~\cite{Canup2012}. We have performed additional DFT simulations to calculate the Hugoniot of MgO starting from an initial temperature of 1900~K. From $T_0=1900$K, a minimum shock pressure of 445~GPa is required to achieve complete melt in the MgO. Assuming planar normal impact, we can determine a minimum impact velocity required to melt MgO. Table~\ref{ImpactCompare} lists the required impact velocities for impactors of common planetary materials. In a real impact event, oblique impact \cite{Kraus2014} and shock attenuation through the mantle~\cite{Croft1982} affect the amount of energy transferred and imply even higher velocities are likely required for significant melting of MgO in the mantle.
\begin{table}
\caption{Impactor velocities for common planetary materials required to completely melt MgO with T$_0$ at ambient and at 1900 K in a shock event (assuming planar normal impact).}
\begin{tabular}{ccc}
\hline
Initial MgO Temp. & Impactor & Impact Velocity \\
$[K]$ & $[300 K]$ & $[km/s]$ \\
\hline
300 & MgO & 18.6 \\
300 & Dunite & 19.4 \\
300 & Iron & 15.3 \\
300 & Quartz & 20.1 \\
\hline
1900 & MgO & 16.0 \\
1900 & Dunite & 16.3 \\
1900 & Iron & 12.9 \\
1900 & Quartz & 17.7 \\
\hline \hline
\end{tabular}
\label{ImpactCompare}
\end{table}
We have performed an extensive experimental and computational study of the high pressure - high temperature behavior of MgO. The Hugoniot is experimentally determined up to 1.2 TPa and the data suggests that the Hugoniot crosses the B2 region with a large solid-liquid coexistence region lasting over 100 GPa before complete melt. The DFT simulation results corroborate the findings inferred from the experimental data and also show that at elevated T$_0$ the Hugoniot persists through a large coexistence region before melt. The results help place a lower bound on impact velocities for the moon-forming giant impact scenario. The data and phase diagram provide a solid basis for the development of an MgO equation of state that can be used in planetary collision studies.
\begin{figure}
\noindent\includegraphics[width=3.1in]{Figure04.eps}
\caption{P-T phase diagram of MgO with calculated Hugoniots starting at ambient and elevated initial temperature conditions. Experimental P-T data~\cite{SvendsenMgO, OlegSCCM09} and the low pressure B1-B2 melt line from Ref.~\citenum{Alfe-MgO-2phase} are included.}
\label{PhaseDiagram}
\end{figure}
The authors thank the Z-Operations and Fabrication team for assembling targets and fielding the Z experiments. The authors also thank K. Cochrane for insightful discussions. QMC calculations by LS were supported through the Predictive Theory and Modeling for Materials and Chemical Science program by the Office of Basic Energy Science (BES), Department of Energy (DOE). Sandia National Laboratories is a multiprogram laboratory managed and operated by Sandia Corporation, a wholly owned subsidiary of Lockheed Martin Corporation, for the U.S. Department of Energy's National Nuclear Security Administration under Contract No. DE-AC04-94AL85000.
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
With the growing concern on the global climate change, it becomes critical to develop energy-efficient systems in all industries. In particular, the information and communication technology (ICT) sector was estimated to contribute to more than 830 million tons of carbon dioxide emission in 2013, which is approximately 2\% of the global carbon dioxide emission, and this number is expected to double by the year 2020 \cite{2013CEETAnnualReport}. With the fast growing mobile data communication demand, the wireless cellular networks are playing a more important role in ICT sector than ever before. Thus a promising approach to reduce the global green house gas emission is to reduce the energy consumption of the cellular networks. It is also very important from the economical perspective of network operators as a significant portion of the operational expenditure is due to electricity consumption.
Designing green cellular networks, especially green base stations, is a recent hot research topic. There are at least two mainstream approaches. With the development of smart grid technology, one approach is to utilize renewable energy for base station operation \cite{TaoAnsari2013OptimizingGreenEnergy}. Another approach is to manage the operation profile of the base stations. As the network operators need to deploy their base stations to support the peak mobile data traffic, it is inevitable that during a major portion of the day a large number of the base stations are under-utilized. This phenomenon can be observed from real mobile data usage profile \cite{OhSonKrishnamachari2013DynamicBaseStation}. Unfortunately, the energy consumption of an idle base station is nearly the same as one under full load \cite{CorreiaZellerBlumeFerlingJadingGodorAugerPerre2010Challengesandenabling}. So an under-utilized base station must be switched off in order to save energy, and there naturally rises a problem: Which base stations should be switched off to save the most amount of energy while maintaining adequate service for end users?
Such energy-efficient design of wireless cellular networks has attracted significant attention recently. However, most of the existing efforts focus on the formulation of the problem \cite{OhSonKrishnamachari2013DynamicBaseStation}\cite{MarsanChiaraviglioCiulloMeo2009OptimalEnergySavings}. This paper focuses on designing an swarm intelligence algorithm based on the Social Spider Algorithm (SSA) \cite{YuLi2013SocialSpiderAlgorithm}. In this paper, we concentrate on the modifications on the original problem to make it solvable with metaheuristics, and the implementation of a metaheuristic.
The rest of the paper is organized as follows. In Section \ref{sec:background} we briefly introduce the current progress on green cellular network design. In Section \ref{sec:formulation} we elaborate on the original optimization problem and our modifications. Section \ref{sec:algorithm} presents the detailed implementation of our proposed algorithm. Section \ref{sec:simr} introduces the random network generation protocol, and the simulation results are presented with analysis and discussion. Finally we conclude in Section \ref{sec:conclusion} with potential future research.
\section{Background}\label{sec:background}
Green cellular network design is a recent hot research topic, and much work has been carried out. Chiaraviglio \textit{et al.} initially proposed the idea of dynamic base station operation based on the data traffic profile in \cite{ChiaraviglioCiulloMeoMarsan2009Energyefficientmanagement}. Later they extended their work and set up an analytical model to control the base station switching profile \cite{MarsanChiaraviglioCiulloMeo2009OptimalEnergySavings}. Oh and Krishnamachari also proposed a similar model in \cite{OhKrishnamachari2010EnergySavingsthrough}, and devised a simple base station switching policy. They further developed a complete system model for this problem and proposed a distributed algorithm to solve it \cite{OhSonKrishnamachari2013DynamicBaseStation}. In this work, both the algorithm design and the practical implementation, including a distributed base station switching protocols were elaborated, and the preliminary simulation results indicate a satisfactory energy-saving performance. All the above work focus on base station operations in the same type of networks, i.e. macro cellular networks. There are also some researchers concentrating on the cooperative operation considering network sharing. Fehske \textit{et al.} investigated the possibility of deploying small and low power base stations alongside conventional ones from a deployment perspective in \cite{FehskeRichterFettweis2009EnergyEfficiencyImprovements}. Marsan and Meo evaluated the energy savings achieved with the energy-aware cooperative management of the cellular access networks of different operators over the same area in \cite{MarsanMeo2010Energyefficientmanagement}. In \cite{RostFettweis2010TransmissionComputationEnergy}, Rost and Fettweis discussed the energy-efficient operation based on the cooperative transmission in multi-hop systems. Niu \textit{et al.} introduced the concept of cell zooming for energy savings. In their implementation, the cell size is adaptively adjusted according to different control variables.
\section{Problem Formulation}\label{sec:formulation}
In this paper, we adopt the system model formulated in \cite{OhSonKrishnamachari2013DynamicBaseStation} as it is a simple yet generalized wireless cellular network model. In this model, the system load is formulated as follows:
\begin{equation}
\rho_b(t)=\int_{A_b}\frac{\gamma(x,t)}{r_b(x,t)}dx,
\end{equation}
where $\rho_b(t)$ represents the system load of base station $b$ at time $t$, $A_b$ represents $b$'s coverage, $\gamma(x,t)$ represents the traffic load of user equipment $x$ at $t$, and $r_b(x,t)$ represents the service rate of $x$ from $b$ at $t$. This system load stands for the fraction of time needed to serve all the data transmission load in the base station's coverage. For simplicity, we omit the time variable in the following notations.
With this system model, we further formulate our Base Station Switching Problem (BSSP) as follows. We consider that the system load for a wireless cellular network remains constant in a short time interval. An optimization problem may be formulated as follows:
\begin{equation}\label{eqn:fitness}
\begin{aligned}
\min&\hspace{1em}\sum_{b\in B}a_b\\
s.t.&\hspace{1em}0\leq\rho_b\leq\overline{\rho_b},\hspace{1em}\forall b\in B
\end{aligned}
,
\end{equation}
where $B$ is the collection of all available base stations, $a_b\in\{0,1\}$ is the active indicator of $b$, and $\overline{\rho_b}$ is the maximum allowed system load for $b$. When a base station $b$ is switched off, its traffic load will be handled by all its active neighboring base stations $N_b = \{N_{b,1},N_{b,2},\cdots,N_{b,n}\}$, and we use $\rho_{b\rightarrow i}$ to denote the amount of transferred traffic load from $b$ to the $i$-th neighbor base station in $N_b$.
From (\ref{eqn:fitness}) we can see the fitness evaluation function is very simple, but the rigid constraints may potentially obstruct the generation of feasible solutions to the problem. In order to alleviate the effort in designing an algorithm that can easily generate solutions satisfying all the constraints, we transform the constraint in (\ref{eqn:fitness}) into a penalty function. The new fitness function is described as follows:
\begin{equation}\label{eqn:newfitness}
\min\hspace{1em}\sum_{b\in B}(a_b + |B|\times P_b),
\end{equation}
where $P_b$ is the penalty value for $b$. In order to define $P_b$, we first analyze the possible scenarios a base station may come across during the switching operation:
\begin{enumerate}
\item $b$ is switched off, and $|N_b|=0$. In this case, both $b$ and all its neighboring base stations are switched off and the traffic load originally handled by $b$ cannot be served anymore. $P_b$ is defined as $1+\rho_b$ for all such base stations.
\item $b$ is switched off, and $|N_b|>0$. In this case, $b$ is switched off, but it has some active neighboring base stations. The traffic load of $b$ is handed over to $N_b$ and the transferred load to $N_{b,i}$ is $\rho_{b\rightarrow i}$. $P_b$ is defined as $0$ for all such base stations.
\item $b$ remains active, and a number of its neighboring base stations are switched off. We use $S_b$ to denote the collection of the neighboring inactive base stations. Then the system load for $b$ is $\rho_b+\sum_{s\in S_b}\rho_{s\rightarrow b}$. If this new system load is larger than $\overline{\rho_b}$, $P_b$ is defined as $1+\rho_b+\sum_{s\in S}\rho_{s\rightarrow b}-\overline{\rho_b}$. Otherwise, $P_b=0$ for all such base stations.
\end{enumerate}
This penalty function is designed according to the design principles that are critical for a successful penalty function stated in \cite{KhuriBackHeitkotter1994evolutionaryapproachto}:
\begin{enumerate}
\item The fitness values shall improve as the solutions approach feasible regions of the search space.
\item The unfeasible solutions are guaranteed to be assigned with fitness values inferior to the fitness value of the worst feasible solution.
\end{enumerate}
Thus, any solutions with fitness values larger than $|B|$ are unfeasible ones to the original optimization problem (\ref{eqn:fitness}).
It is noted that BSSP is NP-complete as it can be reduced from a vertex-covering problem, which has been shown to be NP-complete \cite{Karp1972Reducibilityamongcombinatorial}.
\section{Algorithm Design}\label{sec:algorithm}
In this section, we will first briefly introduce SSA. Then the detailed implementation of our proposed methodology will be presented.
\subsection{Social Spider Algorithm}
SSA is a recently proposed general-purpose swarm intelligence algorithm. It mimics the foraging behavior of the social spiders to perform optimization task. In SSA, the search space of the optimization problem is formulated as a hyper-dimensional spider web, and each position on the web represents a feasible solution. Besides the solution space, the spider web also serves as the transmission media of the vibrations generated by the spiders.
The spiders are the basic operating agents of SSA. Each spider holds a memory consisting of its current position on the web, the fitness value of its current position, and the vibration (position and intensity) it was following in the previous iteration. The first two pieces of information describe the characteristic of this spider, and the last one helps the algorithm guide the movement of this spider.
Based on biological observations, spiders are found to be extremely sensitive to vibrations. They can accurately sense the strength and the direction of vibrations, and can even separate different vibrations propagated on the same web\cite{Uetz1992Foragingstrategiesspiders}. SSA utilizes this characteristic of the spiders and established a vibration-based information-loss communication system among spiders. In SSA, a spider generates a new vibration whenever it moves to a different position from the previous one. The vibration will then propagate over the spider web and be sensed by others. By this means, the spiders share their personal information to form a collective social knowledge.
The vibrations are defined by two properties in SSA, namely, the source position and the source intensity. When a spider moves to a new position, it generates a vibration at its current position. We define the intensity of a vibration in the range $[0,+\infty)$, and the value of the intensity is calculated as follows:
\begin{equation}\label{eqn:intensity}
I=
\begin{cases}
1/(C_{max} - f(s)) & \mathrm{for\:maximization} \\
1/(f(s) - C_{min}) & \mathrm{for\:minimization} \\
\end{cases}
,
\end{equation}
where $I$ is the intensity of the vibration at its source position, $f(s)$ is the fitness value of the source position, $C_{max}$ is a confidently large constant selected such that all possible fitness values of the maximization problem are smaller than $C_{max}$, and $C_{min}$ is a confidently small constant such that all possible fitness values of the minimization problem are larger than $C_{min}$.
(\ref{eqn:intensity}) provides the method of calculating the source intensity of the vibration. However, as a kind of energy, the vibration attenuates over time and distance during the process of propagation. This physical phenomenon is also taken into consideration in SSA from two aspects, i.e., time and distance. As a swarm intelligence algorithm, SSA performs the searching task in an iterative manner. The vibrations are attenuated with each iteration as follows:
\begin{equation}\label{eqn:vibratteniter}
I(t+1)=I(t)\times r_a,
\end{equation}
where $I(t)$ is the source vibration intensity at iteration $t$, and $r_a$ is a user-defined vibration attenuation parameter. This design can prevent the algorithm from pre-mature convergence as a non-decaying vibration on the web can potentially attracts all spiders to move continuously towards it, thus facilitating the exploitation searching behavior but obstructing the exploration of the whole search space.
Another attenuation factor besides the time is the propagation distance. In SSA, the vibration attenuation over distance is defined as follows:
\begin{equation}\label{eqn:vibrattendist}
I(p)=I(s)\times\exp(-\frac{D(s,p)}{D_{max}\times r_a}),
\end{equation}
where $s$ and $p$ are the source position and receiving position of the vibration, respectively, $I(s)$ is the vibration intensity at $s$, $D(s,p)$ is the distance between $s$ and $p$, and $D_{max}$ is the maximum possible distance between any two positions on the web. We usually employ the Manhattan distance to reduce the computational time.
There are three phases in SSA, namely, initialization, iteration, and final. In each run of SSA, we start with the initialization phase where the objective function, search space, and the optimization parameters of SSA are initialized. Then a random population of spiders are generated and placed on the spider web, i.e., search space. The positions of these spiders are randomly generated in the search space, and the initial vibration each spider is following has a position of the spider's current position, and an intensity of zero. This finishes the initialization phase and SSA proceeds to the iteration phase.
In the iteration phase, the algorithm performs the optimization task in an iterative manner. A number of iterations are executed. In each iteration, the algorithm first evaluates the fitness values of all spiders on the web and attenuates the vibrations in the previous iteration. Then the spiders generate vibrations using (\ref{eqn:intensity}) at their current positions. The vibrations then propagates using (\ref{eqn:vibrattendist}) over the spider web. After the propagation, each spider will receive a number of different vibrations from all directions. Upon the receipt of these attenuated vibrations, each spider chooses one vibration with the largest intensity, i.e., the strongest vibration $v_{best}$. $v_{best}$ is then compared with the stored vibration this spider followed in the previous iteration $v_{prev}$. The one with a larger intensity is kept and saved as $v_{best}$. The algorithm then manipulate the position of the spiders as follows:
\begin{equation}\label{eqn:walk}
\boldsymbol{P}_s(t+1)=\boldsymbol{P}_s(t)+(\boldsymbol{P}_{best}(t)-\boldsymbol{P}_s(t))\odot(\boldsymbol{1}-\boldsymbol{R}\odot\boldsymbol{R}),
\end{equation}
where $\boldsymbol{P}_s(t)$ is the position of spider $s$ at iteration $t$, and $\odot$ denotes element-wise multiplication. $\boldsymbol{P}_{best}$ is the vibration source position of the best vibration $v_{best}$. $\boldsymbol{R}$ is a vector of random numbers generated from zero to one uniformly, whose length is the number of dimensions of the problem, and $\boldsymbol{1}$ is a vector of ones of the same length as $\boldsymbol{R}$.
After the manipulation of spider positions, an artificial spider jump away process is introduced in order to prevent the algorithm from getting stuck in the local optimums. Each spider has a small possibility of being re-assigned with a new random position in the search space. And this ends one iteration.
The iteration phase loops until the stopping criteria is matched. The algorithm then outputs the best solution with the fitness value. The above three phases constitute SSA and interested readers can refer to \cite{YuLi2013SocialSpiderAlgorithm} for more details.
\subsection{The Proposed Implementation}
SSA was initially proposed to solve global optimization problems, i.e., the solution space is continuous. So we make several modifications to adapt it to solve BSSP, which is a discrete problem.
\subsubsection{Encoding Scheme}
In this work, we adopt the classical encoding scheme for this kind of problem. We use a vector of 0's and 1's to represent the off/on state of each available base station.
\subsubsection{Spider Following Scheme}
In SSA, we use (\ref{eqn:walk}) to manipulate the movement of spiders. However this equation is designed for continuous optimization. So we devise a new spider following scheme to replace (\ref{eqn:walk}) as follows:
\begin{equation}\label{eqn:newwalk}
\boldsymbol{P}_s(t+1)=\boldsymbol{P}_s(t)+round((\boldsymbol{P}_{best}(t)-\boldsymbol{P}_s(t))\odot\boldsymbol{R}),
\end{equation}
where $round()$ is the rounding function. (\ref{eqn:newwalk}) will first determine the dimensions at which $\boldsymbol{P}_{best}(t)$ and $\boldsymbol{P}_s(t)$ are different. Then all different dimensions have a probability of $0.5$ to change.
\subsubsection{Spider Jump Away Scheme}
Besides the spider following scheme, another scheme that can change the position of the spiders is the spider jump away scheme. In SSA, the jump away scheme operates at the spider level, where each selected spider is assigned with a new random position. But in our implementation for solving BSSP, the jump away scheme operates on the dimension level. Right after the spider following scheme has manipulated the position of each spider, each dimension of a spider position has a probability of $1/|B|$ to change from zero to one or one to zero, i.e.,
\begin{equation}\label{eqn:newjump}
\boldsymbol{P}_s(t+1)=|step(\boldsymbol{R})-\boldsymbol{P}_s(t)|,
\end{equation}
where $step(r)=1$ if $r<1/|B|$, otherwise $step(r)=0$.
These are the modifications we made to adapt SSA to solve BSSP. Algorithm \ref{alg:ssa} is the pseudo-code of SSA for BSSP
\begin{algorithm}
\caption{\sc{Social Spider Algorithm for Base Station Switching Problem}}
\begin{algorithmic}[1]
\State Assign values to the parameters of SSA.
\State Create the population of spiders $pop$ and assign memory for them.
\State Initialize $v_{prev}$ for each spider.
\While {stopping criteria not met}
\For {\textbf{each} spider $s$ in $pop$}
\State Evaluate the fitness value of $s$.
\State Attenuate the intensity of $v_{prev}$.
\State Generate a vibration at the position of $s$.
\EndFor
\For {\textbf{each} spider $s$ in $pop$}
\State Calculate the intensity of the vibrations $V$ \par\hspace{0.7em} generated by other spiders.
\State Select the strongest vibration $v_{best}$ from $V$.
\If {The intensity of $v_{best}$ is smaller than $v_{prev}$}
\State Store $v_{prev}$ as $v_{best}$.
\EndIf
\State Perform the spider following move towards $v_{best}$.
\State Perform the spider jump away modification on \par\hspace{0.7em} each spider.
\State Store $v_{best}$ as $v_{prev}$.
\EndFor
\EndWhile
\State Output the best solution found.
\end{algorithmic}
\label{alg:ssa}
\end{algorithm}
\section{Experimental Setting and Simulation Results}\label{sec:simr}
In order to evaluate the performance of saving energy using our proposed algorithm, we performed a series of simulations over some randomly generated instances.
\subsection{Testing Instance}
In our simulation, we consider the 3G network topologies consisting of 20, 40, and 60 base stations in the area of $10 \times 10$ km$^2$. The instances are generated as follows. We first randomly deploy the base stations to different positions in the area. Then we start from the first base station and determine the number of its neighboring base stations using
\begin{equation}\label{eqn:neighbornumber}
|Neigh_b|=|N_b+S_b|=Pois(\lambda-2)+2,
\end{equation}
where $Pois(\lambda)$ is the Poisson distribution and $\lambda$ is the expected number of neighboring base stations. Each base station may have a different number of neighborhoods, and this design guarantees that all base stations will have at least two neighborhoods.
After determining the number of neighborhoods for each base stations, we establish the neighboring links starting from the first base station to the last one. In this step we first check the existing links connected to the current base station. If the number is already larger than $|Neigh_b|$, we go on to the next base station. Otherwise we establish a neighboring link between the current base station and its nearest base station that is not linked yet. This substep will iterate until enough links are established. The pseudo-code for this random instance generation algorithm is presented in Algorithm \ref{alg:random}.
\begin{algorithm}
\caption{\sc{Random Instance Generation Algorithm}}
\begin{algorithmic}[1]
\For {\textbf{each} base station $b$ in $B$}
\State Assign a random position in the area for $b$.
\EndFor
\For {\textbf{each} base station $b$ in $B$}
\State Determine the number of $b$'s neighboring base stations \par $|Neigh_b|$.
\EndFor
\For {\textbf{each} base station $b$ in $B$}
\State Find the number of $b$'s established neighboring links \par $n$.
\If {$n\geq|Neigh_b|$}
\State \textbf{continue}
\Else
\State Find the $|Neigh_b|-n$ nearest base stations to $b$ \par\hspace{0.7em} that not yet linked with $b$.
\State Establish neighboring links between these base \par\hspace{1.1em}stations with $b$.
\EndIf
\EndFor
\State Output the best solution found.
\end{algorithmic}
\label{alg:random}
\end{algorithm}
Using this method we generated 9 instances using the instance parameter $(|B|,\lambda)\in[(20,3),(40,4),(60,6)]$, and 3 different instances are generated for each instance parameter pair. These instances are shown in Fig. \ref{fig:instance}, where the dots are the base stations and the base stations connected with lines are neighboring base stations. The first two digits in the names of the instances are $|B|$ and $\lambda$, respectively, and the last letter stands for different instances with the same parameters.
\begin{figure}
\center
\subfigure[20.3.A]{
\includegraphics[width=0.27\linewidth]{data/data_20_3_1.pdf}
}
\subfigure[20.3.B]{
\includegraphics[width=0.27\linewidth]{data/data_20_3_2.pdf}
}
\subfigure[20.3.C]{
\includegraphics[width=0.27\linewidth]{data/data_20_3_3.pdf}
}
\subfigure[40.4.A]{
\includegraphics[width=0.27\linewidth]{data/data_40_4_1.pdf}
}
\subfigure[40.4.B]{
\includegraphics[width=0.27\linewidth]{data/data_40_4_2.pdf}
}
\subfigure[40.4.C]{
\includegraphics[width=0.27\linewidth]{data/data_40_4_3.pdf}
}
\subfigure[60.5.A]{
\includegraphics[width=0.27\linewidth]{data/data_60_5_1.pdf}
}
\subfigure[60.5.B]{
\includegraphics[width=0.27\linewidth]{data/data_60_5_2.pdf}
}
\subfigure[60.5.C]{
\includegraphics[width=0.27\linewidth]{data/data_60_5_3.pdf}
}
\caption{Plot of Randomly Generated Topology Instances with Different Parameters}
\label{fig:instance}
\end{figure}
\subsection{Experimental Setting}
To determine which base stations to switch off, we need some feedback information from all the base stations. However these feedback information may reduce the overall performance and sometimes is invalid. So in our simulation we use an approximation method to reduce the reliability of these external and internal information. We make two assumptions on the random instances generated:
\begin{enumerate}
\item The system load for each base station is equal to the overall system load: $\rho_b\equiv\rho_s, \forall b\in B$. Here $\rho_s$ is the overall system load.
\item The transferred traffic load from one base station to all its neighbor base stations are equal: $\rho_{b\rightarrow i}\equiv\rho_b/|B|$
\end{enumerate}
With these two assumptions we minimize the impact of all additional information and focus on the optimization performance of our proposed algorithm.
We compare the performance of our proposed SSA for BSSP with the SWitching-on/off based Energy Saving algorithm (SWES) \cite{OhSonKrishnamachari2013DynamicBaseStation} whose system model was adopted in our problem formulation. In particular, the SWES$_{(0,0)}$ algorithm is employed for performance comparison as no additional information (external or internal) is required for this algorithm, thus satisfying our assumptions. For SSA, we use the number of base stations as the size of the population, and the vibration attenuation rate is set to 0.9. As the spider jump away scheme is modified, we do not need the jump away rate as elaborated in \cite{YuLi2013SocialSpiderAlgorithm}. The maximum iteration number is set to 500, so for 20-base-station instances the fitness function is evaluated for 10 000 times. $\overline{\rho_b}$ is set to 0.6 for all base stations in all the instances and we test the performance of the compared algorithms with $\rho_s\in[0.05,0.1,0.15,0.2,0.25,0.3,0.35,0.4,0.45,0.5,0.55]$ for each instance.
\subsection{Simulation Results}
\begin{figure}
\center
\subfigure[Mean results of 20 base stations with $\lambda=3$]{
\includegraphics[width=0.95\linewidth]{result/result_20_3.pdf}
}
\subfigure[Mean results of 40 base stations with $\lambda=4$]{
\includegraphics[width=0.95\linewidth]{result/result_40_4.pdf}
}
\subfigure[Mean results of 60 base stations with $\lambda=5$]{
\includegraphics[width=0.95\linewidth]{result/result_60_5.pdf}
}
\caption{Comparison of Active Base Stations}
\label{fig:result}
\end{figure}
Fig. \ref{fig:result} illustrates the simulation results of our generated random topology instances with SSA for BSSP and SWES$_{(0,0)}$. Each dot in the figure is the mean value of the number of active base stations in all three instances with the corresponding same set of parameters. From the simulation results we have the following observations:
\begin{enumerate}
\item SSA for BSSP always outperforms SWES$_{(0,0)}$ in all three set of simulations. Out of 33 data points, SSA performs better than SWES$_{(0,0)}$ in 29, and the remaining 4 comparisons end up with draws.
\item Both SSA for BSSP and SWES$_{(0,0)}$ are efficient in terms of saving energy. Assume the traffic profile is increasing and decreasing evenly during the day, SSA can save 42.5\% of the total energy and SWES$_{(0,0)}$ can save 31.3\%. These numbers are generated by averaging all data points indicated in Fig. \ref{fig:result}.
\item The advantage of SSA for BSSP over SWES$_{(0,0)}$ is more significant when the system load is relatively low. A potential reason is that when the system load is close to the maximum allowed system load, the number of base stations that can potentially be switched off is very small. This means that the number of feasible solutions to (\ref{eqn:fitness}) is very small and a heuristic like SWES$_{(0,0)}$ can also find a good solution to the problem. In the extreme case of $\rho=0.55$, no base station can be switched off and the performance of the two compared algorithm naturally becomes the same.
\end{enumerate}
\subsection{Impact of $\lambda$ on Energy Saving Performance}
Besides the comparison of SSA for BSSP and SWES$_{(0,0)}$, we also analyze the impact of $\lambda$ on the performance of saving energy. In order to have a complete analysis on this issue, we generated 5 topologies with $\lambda\in[3,5,6,7,8]$, $|B|=40$ and conduct simulations on these instances as well as the ``40.4.A'' instance illustrated in Fig. \ref{fig:instance}. The topology instances to be considered are presented in Fig. \ref{fig:lambda}.
\begin{figure}
\center
\subfigure[40.3]{
\includegraphics[width=0.27\linewidth]{data/data_40_3_1.pdf}
}
\subfigure[40.4.A]{
\includegraphics[width=0.27\linewidth]{data/data_40_4_1.pdf}
}
\subfigure[40.5]{
\includegraphics[width=0.27\linewidth]{data/data_40_5_1.pdf}
}
\subfigure[40.6]{
\includegraphics[width=0.27\linewidth]{data/data_40_6_1.pdf}
}
\subfigure[40.7]{
\includegraphics[width=0.27\linewidth]{data/data_40_7_1.pdf}
}
\subfigure[40.8]{
\includegraphics[width=0.27\linewidth]{data/data_40_8_1.pdf}
}
\caption{Plot of Randomly Generated Topology Instances with Different $\lambda$}
\label{fig:lambda}
\end{figure}
The simulation results on these new instances with SSA for BSSP is shown in Fig. \ref{fig:lambdaresult}. From the results we can observe that although $\lambda=3$ performs slightly worse than others when the system load is high and low, the overall performance of these instances are barely distinguishable. This phenomenon suggests that increasing the neighboring links among the base stations would not significantly improve the overall energy saving performance when the system already has sufficient links. It can also be observed that the performance curves are very close to the ideal best performing curve $\rho_s/\overline{\rho_b}$. This again shows the outstanding performance our proposed algorithm.
\begin{figure}
\center
\includegraphics[width=0.95\linewidth]{result/result_lambda.pdf}
\caption{Simulation Results of Instances with Different $\lambda$}
\label{fig:lambdaresult}
\end{figure}
\section{Conclusion and Future Work}\label{sec:conclusion}
In this paper we consider the possibility of saving energy in wireless cellular network by switching off the under-utilized base stations during non-peak hours. We transform the original optimization problem with constraints into an unconstrained problem with a penalty function to alleviate the design difficulty considering the constraints. Besides the new formulation, we also propose a new binary SSA to solve this optimization problem. SSA is a newly proposed general-purpose metaheuristic. It was originally devised to solve continuous optimization problem. We make several modifications to the canonical SSA to adapt it to solve BSSP. We devise two optimization schemes in SSA and make modifications on the control flow of the algorithm. We employ a series of randomly generated network topologies to test the performance of our proposed SSA for BSSP. The simulation results indicate that SSA can almost always outperform the compared algorithm. We also analyze the impact of the network density on the energy saving performance. The results imply that the performance improvement caused by the increasing network density is limited as the performance is already very close to the ideal best case scenario.
There are some potential future work. In this paper, we consider the base stations consume the same amount of energy despite different system loads and communication ranges. In reality, both factors will influence the energy consumed. This can be taken into consideration but will significantly increase the complexity of the optimization problem. Another extension is to consider the diversity of the base stations. In this work we only consider homogeneous base stations. It will be interesting to study the impact of adopting different types of base stations on the energy-saving performance.
\section*{Acknowledgement}
This research is supported in part by the University of Hong Kong Strategic Research Theme on Computation and Information.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\lettrine[lines=2]{T}{he set} covering problem (SCP) is one of the representative combinatorial optimization problems. It has many real-world applications, e.g. bus, railway and airline crew scheduling, vehicle routing, facility location, and political districting \cite{BalasPadberg1976SetPartitioning:Survey}. More recent applications of SCP are on sensor lifetime maximization \cite{HuZhangYuChungLiShiLuo2010HybridGeneticAlgorithm} and phasor measurement unit placement \cite{XuWenLeungLi2013OptimalPMUPlacement}.
SCP is formally defined as follows. We have a set of $m$ elements $\mathbb{M}=\{1,\cdots,m\}$ and a collection of $n$ subsets $\mathbb{N} = \{S_j\subseteq\mathbb{M},1\leq j\leq n\}$, each of which is associated with a cost $S_j$, denoted as $c_j$. We say a collection of subsets $\mathbb{X}\subseteq\mathbb{N}$ is a cover of $\mathbb{M}$ if $\bigcup_{S_j\in\mathbb{X}}S_j=\mathbb{M}$ holds. $\mathbb{X}$ is a prime cover of $\mathbb{M}$ if there is no redundant subset in $\mathbb{X}$, i.e., $\mathbb{X}$ will not cover $\mathbb{M}$ if any subset is removed from $\mathbb{X}$. The goal of SCP is to find an $\mathbb{X}$ with the minimum cost.
SCP is usually formulated as a binary integer programming problem as follows:
\begin{equation}
\begin{aligned}
& \min && \sum^n_{j=1}c_jx_j \\
& \text{s.t.} && \sum^n_{j=1}a_{ij}x_j\geq1, \; i=1,2,\cdots,m, \\
& && x_j\in\{0,1\}, \; j=1,2,\cdots,n,
\end{aligned}
\end{equation}
where $a_{ij}=1$ if $i\in S_j$ and $a_{ij}=0$ otherwise. The decision variable $x_{j}$ is set to one if subset $S_j$ is selected in the cover $\mathbb{X}$.
It is also common to formulate SCP into matrix form. In this formulation, SCP is the problem of covering the rows of an $m\times n$ matrix by a subset of the columns at a minimal cost. We use $\mathbf{A}=\{a_{ij}, 1\leq i\leq m, 1\leq j\leq n\}$ to represent the matrix, and we say the $k$-th element is covered by the $l$-th subset if $a_{kl}=1$. We use $\mathsf{C}=\{c_j,1\leq j\leq n\}$ as the cost coefficient vector. Then SCP is defined as follows:
\begin{equation}\label{eqn:form2}
\begin{aligned}
& \min && \mathsf{C^\top X} \\
& \text{s.t.} && \mathbf{A}\mathsf{X}\geq\mathsf{b}, \\
\end{aligned}
\end{equation}
where $\mathsf{X}=\{x_j,1\leq j\leq n\}$ is the solution vector and $\mathsf{b}$ is the unit vector of length $m$. The constraint ensures that each row is covered by at least one column. If the costs for all subsets are identical, then the SCP is named \textit{unicost} SCP. Otherwise, it is called \textit{weighted} or \textit{non-unicost} SCP.
SCP is known to be an NP-hard optimization problem \cite{GareyJohnson1979ComputersandIntractability:}, and metaheuristics have been shown to be effective in solving complex problems. Chemical reaction optimization (CRO) is one of the general-purpose metaheuristics which has shown its capability in solving similar NP-hard combinatorial problems, e.g. the quadratic assignment problem \cite{LamLi2010ChemicalReactionInspired}. CRO is inspired by chemical reactions, where reactant molecules collide with container walls or with each other. During the collision, the structure of these molecules may change and the energy hold by these molecules may be transferred to other molecules or transformed into other energy forms. The changes made by the collision follow a natural tendency that the potential energy of product molecules is smaller than that of reactant molecules macroscopically \cite{LamLi2010ChemicalReactionInspired}. CRO utilizes this tendency to perform optimization.
In this paper, we propose a heuristic-based CRO algorithm to solve SCP, named hCRO. We perform a series of simulations to test its performance. This paper is organized as follows. We perform a brief survey on previously proposed approaches to solve SCP, and some prior efforts on CRO in Section II. Section III presents the design and implementation of our proposed hCRO algorithm. The performance of our proposed algorithm is illustrated with the help of a series of benchmark problems in Section IV. Finally we conclude our work and propose some potential future work in Section V.
\section{Background}
SCP is a classical NP-hard combinatorial optimization problem, and has attracted the interests of researchers for several decades. Many exact algorithms, heuristic, and metaheuristic approaches have been proposed and reported in the literature.
Existing exact algorithms to solve SCP is mainly based on the branch-and-bound and branch-and-cut search algorithms \cite{BalachandarKannan2010MetaheuristicAlgorithm}. Fisher and Kedia proposed an exact branch-and-bound algorithm based on a dual heuristic \cite{FisherKedia1990OptimalSolutionSet}. This algorithm is capable of solving SCP instances with up to 200 elements and 2000 subsets (200$\times$2000). Beasley combined a Lagrangian-based heuristic, sub-gradient optimization, and linear programming to improve the branching strategy \cite{Beasley1987AlgorithmSetCovering}. He then enhanced this algorithm using feasible solution exclusion constraints and Gomory's f-cut in \cite{BeasleyJornsten1992EnhancingAlgorithmSet}. The algorithm is tested on instances with matrices up to the order of 400$\times$4000. Harche and Thompson developed a column subtraction exact algorithm to solve sparse instances of SCP up to the order of 900$\times$8000 \cite{HarcheThompson1994ColumnSubtractionAlgorithm:}. All the above exact algorithms are based on the tree-search algorithm, which has limitations such as extremely high computational complexity, very large searching space, and relatively poor performance \cite{BalachandarKannan2010MetaheuristicAlgorithm}. Due to these drawbacks, researchers resorted to approximate algorithms to meet the requirement of less stringent computation with satisfactory solution quality.
Greedy algorithms are one of the heuristic approaches to quickly solve large combinatorial problems. Chvatal proposed the very first greedy algorithm to solve SCP in 1979 \cite{Chvatal1979GreedyHeuristicSet}. However, due to its deterministic and myopic feature, this algorithm can rarely generate good solutions despite it is fast and simple. In order to improve the performance of the canonical greedy algorithm, researchers have tried to introduce some stochastic features into it \cite{FeoResende1989ProbabilisticHeuristicComputationally} \cite{HaouariChaouachi2002probabilisticgreedysearch}. Generally, greedy algorithms with stochastic features can generate better solutions than the canonical greedy ones. There are also some non-greedy-based heuristics. Caprara's work \cite{CapraraFischettiToth1999HeuristicMethodSet} is an example which gives good solutions. This work introduced variable fixing and pricing techniques into a Lagrangian heuristics.
The research on employing metaheuristic algorithms, especially evolutionary algorithms, to solve SCP has been intensely investigated in the last decade. A wide range of metaheuristics have been utilized, including genetic algorithm \cite{BeasleyChu1996GeneticAlgorithmSet} \cite{Aickelin2002IndirectGeneticAlgorithm}, ant colony optimization \cite{LessingDumitrescuStutzle2004ComparisonBetweenACO}, simulated annealing \cite{JacobsBrusco1995Note:Localsearch}, and artificial neural networks \cite{OhlssonPetersonSoderberg2001EfficientMeanField}. The high adaptability and superior performance of metaheuristic make it a competitive approach to solve SCP. Moreover, due to the characteristics of unicost SCP, researchers have proposed some algorithms to solve this special kind of SCP. Grossman and Wool designed an artificial neural network framework to solve unicost SCP \cite{GrossmanWool1997ComputationalExperienceWith}. However, the existing algorithms to solve SCP have two drawbacks in general. Firstly, most algorithms are designed to solve non-unicost SCP. Beasley and Chu pointed out that their algorithms based on Lagrangian relaxation and genetic algorithm were not recommended for unicost problems \cite{BeasleyChu1996GeneticAlgorithmSet}. From the literature, very few algorithms are found to work effectively for both unicost and non-unicost problem. Secondly, most algorithms which can generate satisfactory results are difficult to implement, while the simple algorithms, e.g., greedy algorithms, are less competitive in performance.
In order to overcome these two drawbacks, the goal of this work is to design a robust and simple metaheuristic based on CRO that can generate good results for both unicost and non-unicost SCP. CRO is a recently proposed general-purpose metaheuristic, which has been developed intensely in the past few years. CRO was proposed by Lam \textit{et al.} in 2010 \cite{LamLi2010ChemicalReactionInspired}, and was originally designed for solving combinatorial optimization problems. They solved some classical problems, e.g., quadratic assignment problem and channel assignment problem. CRO is also applied to solve some real world applications like real-time monitoring \cite{YuLiLam2012SensorDeploymentAir} and smart grid \cite{YuLiLam2013OptimalV2GScheduling}. For continuous problems, Lam \textit{et al.} also proposed a variant of CRO, i.e., real-coded CRO, which has demonstrated superior performance in many real-world applications, e.g., training artificial neural networks \cite{YuLamLi2011EvolutionaryArtificialNeural} and optimal power flow problem in power grid \cite{SunLamLiXuYu2012ChemicalReactionOptimization}.
\section{Algorithm Design}
In this section we will first introduce CRO. Then the implementation of operators used in CRO will be presented.
\subsection{Chemical Reaction Optimization}
We consider a number of molecules in a closed container with an attached energy buffer. Molecules are the basic operating agents of CRO. CRO manipulates and controls a collection of molecules to explore the solution space of an optimization problem. CRO considers the molecular structure as the feasible solution of the optimization problem. Besides the molecular structure, a molecule also has some other attributes that help the algorithm to perform optimization. Typically each molecule possesses two different kinds of energy, namely, potential energy (PE) and kinetic energy (KE). We use PE to represent the solution quality, or objective function value, of the corresponding molecular structure. The solution space is decribed by the potential energy surface (PES). The molecules can move freely on the PES. Every position on the PES is associated with a PE value. The lower the PE (for objective function minimization), the better is the solution. KE quantifies the ability of the molecule to move towards an area on the PES with higher values. The larger a molecule's KE, the higher it can position itself on the PES, which means that this molecule can accept worse solutions. This feature is very important for the cases when CRO tries to manipulate the molecules to jump out of a local optimum in the solution space.
CRO controls and manipulates the molecules with four different elementary reactions, namely, on-wall ineffective collision (\textit{on-wall}), decomposition (\textit{dec}), inter-molecular ineffective collision (\textit{inter}), and synthesis (\textit{syn}), each of which is described by an operator. Each operator modifies the molecular structures of some molecules, performing stochastic exploration or exploitation in the solution space. The four elementary reactions have different energy handling schemes and molecular structure operations. Meanwhile, they share a feature, which distinguishes CRO from other metaheuristics. All the operations in CRO must comply with the energy conservation law, which states that although energy is allowed to transform between types, the total energy in an isolated system (i.e. the container in CRO) shall remain constant. In CRO, the total energy of the system before and after an elementary reaction is constant. Interested reader can refer to \cite{LamLi2012ChemicalReactionOptimization:} for the detailed implementation of CRO.
\subsection{Encoding Scheme}
There are two major encoding schemes to solve SCP in the literature. The first one is the natural encoding scheme. This scheme uses a binary vector of length $n$ for a solution. Each element in the vector represents the status of one particular subset. In other words, this encoding scheme uses the $\mathsf{X}$ in (\ref{eqn:form2}) as the solution for optimization. This encoding scheme is easy to implement, but a random solution generated from this scheme is not guaranteed to be a cover of $\mathbb{M}$. In order to overcome this drawback, a second type of encoding scheme is developed for SCP.
The second encoding scheme, which we adopt to solve SCP with CRO, has each value in the solution vector representing a subset index. The values are selected from the indices of those subsets that cover the corresponding element. For example, consider a solution vector $[2,6,7,2,\cdots]$. This solution uses the second subset to cover the first and fourth elements in $\mathbb{M}$, and the sixth subset to cover the second element and so forth. Solutions in the second scheme have shorter length than those in the first in general (for most non-unicost SCP, $m$ is smaller than $n$). Moreover, all the solutions generated by this encoding scheme satisfy the constraints of SCP naturally \cite{BalachandarKannan2010MetaheuristicAlgorithm}.
\subsection{Algorithm Design}
CRO defines four different types of elementary reactions, which possess different functionalities. So we design a corresponding operator for each of them. We also design an initial solution generator to generate the solution structures of new molecules. We generate all random numbers uniformly in the solution space, unless stated otherwise.
\subsubsection{Initial Solution Generator}
This operator is applied whenever a new molecule is generated. Instead of randomly assigning subset indices (this is common when we solve other optimization problems), we use a reverse cumulative scheme to select a random subset to cover this element. The scheme is stated in Algorithm \ref{alg:init}.
\begin{algorithm}
\caption{\sc{Reverse Cumulative Scheme}}
\label{alg:init}
\begin{algorithmic}[1]
\ForAll{Elements $i$ in a solution}
\State Find all subsets $\mathbb{X}_i$ that cover the i-th element in $\mathbb{M}$.
\State Find the maximum cost $c_{max}$ and the minimum cost \par $c_{min}$ of subsets in $\mathbb{X}_i$.
\ForAll{$S_j\subseteq\mathbb{X}_i$}
\State Assign a reverse cumulative value $v_j$.
\State Assign its probability of being selected $p_j$.
\EndFor
\State Select the value of Element $i$ according to each \par subsets' $p_j$.
\EndFor
\end{algorithmic}
\end{algorithm}
In Algorithm \ref{alg:init}, the variables are calculated by
\begin{equation}\label{eqn:rev1}
v_j=c_{max}+c_{min}-c_j
\end{equation}
and
\begin{equation}\label{eqn:rev2}
p_j=\frac{v_j}{\sum^{|\mathbb{X}_i|}_{k=1}v_k}
\end{equation}
This scheme renders the initial solution more likely to include those subsets with lower cost, while ensuring the solution complies with all constraints. This scheme assigns the subsets with lower cost a larger possibility of being selected by (\ref{eqn:rev1}) and (\ref{eqn:rev2}). For example, the first element in $\mathbb{M}$ can be covered by Subsets 1, 3, and 6, whose costs are 2, 4, and 5, respectively. Then the reverse cumulative value of Subset 1 is calculated by $5+2-2=5$, and that for Subsets 3 and 6 are 3 and 2. Thus the probability of being selected are 0.5, 0.3, and 0.2, respectively.
\subsubsection{Neighborhood Search Operator}
This operator is applied to the two ineffective elementary reactions, namely, on-wall and inter. This operator modifies the input solution slightly to perform a local search. We employ a perturbation heuristic to act as the neighborhood search operator in CRO. The algorithm is stated in Algorithm \ref{alg:neigh}.
\begin{algorithm}
\caption{\sc{Perturbation Heuristic}}
\label{alg:neigh}
\begin{algorithmic}[1]
\ForAll{Subsets $S_j$ in the solution}
\State Calculate its cost efficiency value $e^c_j$.
\EndFor
\State Find the Subset $S_i$ with the lowest cost efficiency.
\State Remove all $i$ in the solution and leave blanks.
\While{There are blanks in the solution}
\ForAll{Subsets $S_k\subseteq\mathbb{N}$}
\State Calculate its repair efficiency value $e^r_k$
\EndFor
\ForAll{Subsets $S_k\subseteq\mathbb{N}$}
\State Calculate its probability of being selected $p_k$
\EndFor
\State Select a subset $S_l$to repair the solution according to \par each subsets' $p_k$.
\State Use $l$ to fill all the blanks which can be covered by \par $S_l$.
\EndWhile
\end{algorithmic}
\end{algorithm}
In Algorithm \ref{alg:neigh}, the variables are calculated by
\begin{equation}\label{eqn:heu1}
e^c_j=\frac{n_j}{c_j},
\end{equation}
where $n_j$ is the occurrence of $j$ in the solution, and
\begin{equation}\label{eqn:heu2}
e^r_k=\frac{s_k}{c_k},
\end{equation}
where $s_k$ is the number of blanks that Subset $S_k$ can cover in the solution, and
\begin{equation}\label{eqn:heu3}
p_k=\frac{e^r_k}{\sum^{|\mathbb{X}|}_{j=1}e^r_j}
\end{equation}
This perturbation heuristic can be divided into two major parts: remove and repair. In the remove phase (Lines 1 to 5 of Algorithm \ref{alg:neigh}), the least efficient subset is removed from the solution. Its cost efficiency value is calculated by (\ref{eqn:heu1}). Assume a Subset 2 with cost 10. If it covers two elements in the solution, its cost efficiency value is 5. At the end of the remove phase, we have an incomplete solution with one or several blank positions. Then in the repair phase, we select a most efficient subset to fill in the blank(s). This repair efficiency value is calculated by (\ref{eqn:heu2}). With the repair efficiency values, we can calculate the probability of being selected for repairing using (\ref{eqn:heu3}). For example, assume we have an incomplete solution $[1,\_,4,\_,\_,\cdots]$ where the underline positions are blank positions. Subset 3 can cover two blanks with a cost 20, and Subset 5 can cover three blanks with a cost 40. So the repair efficiency values of Subsets 3 and 5 are 0.1 and 0.075, respectively. If there are no other subsets, the probabilities of selecting Subsets 3 and 5 are 57\% and 43\%, respectively. If the algorithm chooses Subset 3 for repairing, the two blanks which can be covered by Subset 3 is filled with 3. This completes one iteration of the repairing phase and this process iterates until all blanks are filled.
For on-wall, we directly employ this perturbation scheme to the molecule, while for inter, we apply the scheme to the two input molecules simultaneously.
\begin{table}[t]
\centering
\caption{Beasley's OR Library Non-unicost Instances}
\begin{tabular}{rrrrrr}
\hline
Set & Instances & Size ($m\times n$) & Costs & Density & Opt. Solution \\
\hline
4 & 10 & 200$\times$1000 & 1-100 & 2\% & Known \\
5 & 10 & 200$\times$2000 & 1-100 & 2\% & Known \\
6 & 5 & 200$\times$1000 & 1-100 & 5\% & Known \\
A & 5 & 300$\times$3000 & 1-100 & 2\% & Known \\
B & 5 & 300$\times$3000 & 1-100 & 5\% & Known \\
C & 5 & 400$\times$4000 & 1-100 & 2\% & Known \\
D & 5 & 400$\times$4000 & 1-100 & 5\% & Known \\
NRE & 5 & 500$\times$5000 & 1-100 & 10\% & Unknown \\
NRF & 5 & 500$\times$5000 & 1-100 & 20\% & Unknown \\
NRG & 5 & 1000$\times$10000 & 1-100 & 2\% & Unknown \\
NRH & 5 & 1000$\times$10000 & 1-100 & 5\% & Unknown \\
\hline
\end{tabular}
\label{tab:orlib1}
\end{table}
\subsubsection{Dec and Syn Operators}
The main purpose of dec and syn is to help the molecules jump out of local optimum. So we usually impose severe changes to the input molecule(s). For dec operator, we first copy the input molecule to the two output molecules, then perform the neighborhood search operator on each molecule for 10 times separately. The resultant molecules are regarded as the final output molecules. For the syn operator, we introduce a probabilistic combination scheme to combine the two input molecules and create a new one. Assume the two input solutions are $\mathsf{X}_1$ and $\mathsf{X}_2$, and their costs are $c_1$ and $c_2$, respectively. Each element in the output new solution is drawn from the same position of either $\mathsf{X}_1$ or $\mathsf{X}_2$ with the probability of $\frac{c_2}{c_1+c_2}$ and $\frac{c_1}{c_1+c_2}$, respectively.
\section{Simulation Results}
In this section we will first introduce the benchmark instances used to evaluate the performance of our proposed hCRO algorithm. Then the detailed simulation parameter settings, results, and comparisons are presented.
\subsection{Benchmark Instances}
\begin{table}[t]
\centering
\caption{Beasley's OR Library Unicost Instances}
\begin{tabular}{rrrrrr}
\hline
Set & Instances & Size ($m\times n$) & Density & Opt. Solution \\
\hline
E & 5 & 50$\times$500 & 2\% & Known \\
CLR.10 & 1 & 511$\times$210 & 2\% & Unknown \\
CLR.11 & 1 & 1023$\times$330 & 5\% & Unknown \\
CLR.12 & 1 & 2047$\times$495 & 2\% & Unknown \\
CLR.13 & 1 & 4095$\times$715 & 5\% & Unknown \\
CYC.6 & 1 & 240$\times$192 & 2\% & Known \\
CYC.7 & 1 & 672$\times$448 & 5\% & Unknown \\
CYC.8 & 1 & 1792$\times$1024 & 10\% & Unknown \\
CYC.9 & 1 & 4608$\times$2304 & 20\% & Unknown \\
CYC.10 & 1 & 11520$\times$5120 & 2\% & Unknown \\
CYC.11 & 1 & 28160$\times$11264 & 5\% & Unknown \\
\hline
\end{tabular}
\label{tab:orlib2}
\end{table}
\begin{table}[t]
\centering
\caption{hCRO Parameter Values}
\begin{tabular}{rr}
\hline
Parameter & Value \\
\hline
Initial population size & 10 \\
Initial molecular kinetic energy & 1000 \\
Initial central energy buffer size & 10000 \\
Collision rate & 0.1 \\
Energy loss rate & 0.1 \\
Decomposition threshold & 10000 \\
Synthesis threshold & 1000 \\
\hline
\end{tabular}
\label{tab:param}
\end{table}
We will test the performance of our proposed hCRO using 65 non-unicost SCP test instances from Beasley's OR Library \cite{Beasley1990LagrangianHeuristicSet}. We note that almost all SCP algorithms developed in the past two decades were tested using these problems. The instances are divided into 11 different sets, as listed in Table \ref{tab:orlib1}, where the density is the percentage of non-zero entries in the SCP matrix $\mathbf{A}$.
We also test the performance of hCRO with the unicost instances in Beasley's OR Library and the information about these instances are listed in Table \ref{tab:orlib2}.
\subsection{Parameter Tuning and Simulation Environment}
When applying hCRO to perform simulation, several parameters must be set. Coy \textit{et al.} note that it is often very difficult to find appropriate parameter settings for metaheuristics, and common procedures of generating proper parameter values have ranged from simple trial-and-error to complicated sensitivity analysis \cite{CoyGoldenRungerWasil2001UsingExperimentalDesign}. In this work, we use the trial-and-error method to tune the hCRO parameters, as in some previous CRO efforts \cite{YuLiLam2012SensorDeploymentAir} \cite{YuLamLi2012RealcodedChemical}. The first instances in the SCP problem set 4, 5, 6, A, B, C, and D are selected as representative instance for parameter tuning and we perform a series of test runs on these instances, using the methodology described in \cite{YuLamLi2012RealcodedChemical}. The final parameters used for all test instances are listed in Table \ref{tab:param}.
Our proposed approach was implemented in C++ on a computer with an Intel Core i5 3.1-GHz processor and MinGW compiler. In our experimental study, 100 trials of hCRO were conducted for each of these test problems. The maximum function evaluation limit is set to $n\times1000$, which is smaller or equal to that in all the metaheuristics we compare with.
\subsection{Simulation Results and Comparison with Other Algorithms}
\begin{table}
\centering
\caption{hCRO Results on Non-unicost Instances}
\begin{tabular}{l|r|r|rr|rr|rr}
\hline
Inst. & BKS & Opt. & \multicolumn{2}{c|}{Best} & \multicolumn{2}{c|}{Mean} & \multicolumn{2}{c}{Worst} \\ \cline{4-9}
& & & Value & Pct. & Value & Pct. & Value & Pct. \\\hline
4.1 & 429 & 100 & 429 & 0 & 429 & 0 & 429 & 0 \\
4.2 & 512 & 100 & 512 & 0 & 512 & 0 & 512 & 0 \\
4.3 & 516 & 100 & 516 & 0 & 516 & 0 & 516 & 0 \\
4.4 & 494 & 100 & 494 & 0 & 494 & 0 & 494 & 0 \\
4.5 & 512 & 100 & 512 & 0 & 512 & 0 & 512 & 0 \\
4.6 & 560 & 100 & 560 & 0 & 560 & 0 & 560 & 0 \\
4.7 & 430 & 100 & 430 & 0 & 430 & 0 & 430 & 0 \\
4.8 & 492 & 100 & 492 & 0 & 492 & 0 & 492 & 0 \\
4.9 & 641 & 100 & 641 & 0 & 641 & 0 & 641 & 0 \\
4.10 & 514 & 100 & 514 & 0 & 514 & 0 & 514 & 0 \\ \hline
5.1 & 253 & 100 & 253 & 0 & 253 & 0 & 253 & 0 \\
5.2 & 302 & 100 & 302 & 0 & 302 & 0 & 302 & 0 \\
5.3 & 226 & 100 & 226 & 0 & 226 & 0 & 226 & 0 \\
5.4 & 242 & 100 & 242 & 0 & 242 & 0 & 242 & 0 \\
5.5 & 211 & 100 & 211 & 0 & 211 & 0 & 211 & 0 \\
5.6 & 213 & 100 & 213 & 0 & 213 & 0 & 213 & 0 \\
5.7 & 293 & 100 & 293 & 0 & 293 & 0 & 293 & 0 \\
5.8 & 288 & 100 & 288 & 0 & 288 & 0 & 288 & 0 \\
5.9 & 279 & 100 & 279 & 0 & 279 & 0 & 279 & 0 \\
5.10 & 265 & 100 & 265 & 0 & 265 & 0 & 265 & 0 \\ \hline
6.1 & 138 & 100 & 138 & 0 & 138 & 0 & 138 & 0 \\
6.2 & 146 & 100 & 146 & 0 & 146 & 0 & 146 & 0 \\
6.3 & 145 & 100 & 145 & 0 & 145 & 0 & 145 & 0 \\
6.4 & 131 & 100 & 131 & 0 & 131 & 0 & 131 & 0 \\
6.5 & 161 & 100 & 161 & 0 & 161 & 0 & 161 & 0 \\ \hline
A.1 & 253 & 100 & 253 & 0 & 253 & 0 & 253 & 0 \\
A.2 & 252 & 100 & 252 & 0 & 252 & 0 & 252 & 0 \\
A.3 & 232 & 100 & 232 & 0 & 232 & 0 & 232 & 0 \\
A.4 & 234 & 100 & 234 & 0 & 234 & 0 & 234 & 0 \\
A.5 & 236 & 100 & 236 & 0 & 236 & 0 & 236 & 0 \\ \hline
B.1 & 69 & 100 & 69 & 0 & 69 & 0 & 69 & 0 \\
B.2 & 76 & 100 & 76 & 0 & 76 & 0 & 76 & 0 \\
B.3 & 80 & 100 & 80 & 0 & 80 & 0 & 80 & 0 \\
B.4 & 79 & 100 & 79 & 0 & 79 & 0 & 79 & 0 \\
B.5 & 72 & 100 & 72 & 0 & 72 & 0 & 72 & 0 \\ \hline
C.1 & 227 & 100 & 227 & 0 & 227 & 0 & 227 & 0 \\
C.2 & 219 & 100 & 219 & 0 & 219 & 0 & 219 & 0 \\
C.3 & 243 & 100 & 243 & 0 & 243 & 0 & 243 & 0 \\
C.4 & 219 & 100 & 219 & 0 & 219 & 0 & 219 & 0 \\
C.5 & 215 & 100 & 215 & 0 & 215 & 0 & 215 & 0 \\ \hline
D.1 & 60 & 100 & 60 & 0 & 60 & 0 & 60 & 0 \\
D.2 & 66 & 100 & 66 & 0 & 66 & 0 & 66 & 0 \\
D.3 & 72 & 100 & 72 & 0 & 72 & 0 & 72 & 0 \\
D.4 & 62 & 100 & 62 & 0 & 62 & 0 & 62 & 0 \\
D.5 & 61 & 100 & 61 & 0 & 61 & 0 & 61 & 0 \\ \hline
NRE.1 & 29 & 100 & 29 & 0 & 29 & 0 & 29 & 0 \\
NRE.2 & 30 & 100 & 30 & 0 & 30 & 0 & 30 & 0 \\
NRE.3 & 27 & 100 & 27 & 0 & 27 & 0 & 27 & 0 \\
NRE.4 & 28 & 100 & 28 & 0 & 28 & 0 & 28 & 0 \\
NRE.5 & 28 & 100 & 28 & 0 & 28 & 0 & 28 & 0 \\ \hline
NRF.1 & 14 & 100 & 14 & 0 & 14 & 0 & 14 & 0 \\
NRF.2 & 15 & 100 & 15 & 0 & 15 & 0 & 15 & 0 \\
NRF.3 & 14 & 100 & 14 & 0 & 14 & 0 & 14 & 0 \\
NRF.4 & 14 & 100 & 14 & 0 & 14 & 0 & 14 & 0 \\
NRF.5 & 13 & 100 & 13 & 0 & 13 & 0 & 13 & 0 \\ \hline
NRG.1 & 176 & 100 & 176 & 0 & 176 & 0 & 176 & 0 \\
NRG.2 & 154 & 100 & 154 & 0 & 154 & 0 & 154 & 0 \\
NRG.3 & 166 & 100 & 166 & 0 & 166 & 0 & 166 & 0 \\
NRG.4 & 168 & 100 & 168 & 0 & 168 & 0 & 168 & 0 \\
NRG.5 & 168 & 100 & 168 & 0 & 168 & 0 & 168 & 0 \\ \hline
NRH.1 & 63 & 100 & 63 & 0 & 63 & 0 & 63 & 0 \\
NRH.2 & 63 & 100 & 63 & 0 & 63 & 0 & 63 & 0 \\
NRH.3 & 59 & 100 & 59 & 0 & 59 & 0 & 59 & 0 \\
NRH.4 & 58 & 100 & 58 & 0 & 58 & 0 & 58 & 0 \\
NRH.5 & 55 & 100 & 55 & 0 & 55 & 0 & 55 & 0 \\
\hline
\end{tabular}
\label{tab:cro1}
\end{table}
\begin{table*}
\centering
\caption{hCRO Results on Non-unicost Instances}
\begin{tabular}{l|r|r|rr|rr|rr}
\hline
Inst. & BKS & Opt. & \multicolumn{2}{c|}{Best} & \multicolumn{2}{c|}{Mean} & \multicolumn{2}{c}{Worst} \\ \cline{4-9}
& & & Value & Pct. & Value & Pct. & Value & Pct. \\\hline
E.1 & 5 & 100 & 5 & 0 & 5 & 0 & 5 & 0 \\
E.2 & 5 & 100 & 5 & 0 & 5 & 0 & 5 & 0 \\
E.3 & 5 & 100 & 5 & 0 & 5 & 0 & 5 & 0 \\
E.4 & 5 & 100 & 5 & 0 & 5 & 0 & 5 & 0 \\
E.5 & 5 & 100 & 5 & 0 & 5 & 0 & 5 & 0 \\ \hline
CLR.10 & 25 & 100 & 25 & 0 & 25 & 0 & 25 & 0 \\
CLR.11 & 23 & 100 & 23 & 0 & 23 & 0 & 23 & 0 \\
CLR.12 & 23 & 100 & 23 & 0 & 23 & 0 & 23 & 0 \\
CLR.13 & 23 & 100 & 23 & 0 & 23 & 0 & 23 & 0 \\ \hline
CYC.6 & 60 & 100 & 60 & 0 & 60 & 0 & 60 & 0 \\
CYC.7 & 144 & 11 & 144 & 0 & 146.7 & 0.0188 & 152 & 0.0556 \\
CYC.8 & 344 & 14 & 344 & 0 & 349.2 & 0.0151 & 361 & 0.0494 \\
CYC.9 & 780 & 0 & 789 & 0.0115 & 797.3 & 0.0221 & 819 & 0.05 \\
CYC.10 & 1792 & 0 & 1802 & 0.0056 & 1832.6 & 0.0226 & 1872 & 0.0446 \\
CYC.11 & 4103 & 0 & 4113 & 0.0024 & 4159.2 & 0.0137 & 4201 & 0.0239 \\
\hline
\end{tabular}
\label{tab:cro2}
\end{table*}
The simulation results are presented in Table \ref{tab:cro1}. In this table, ``Inst.'' is the instance index, ``BKS" is the optimum solution or the best known value, ``Opt." is the number of trials that hCRO is able to find the optimum solution or the best known value. We also present the mean, best, and worst objective function values of the 100 trials, and their respective percentages above the optimal value. The results of hCRO on unicost SCP instances are also presented in Table \ref{tab:cro2}.
\begin{table*}
\centering
\caption{Performance Gap Comparison on Non-unicost Instances (\%)}
\begin{tabular}{l|rrrrrrrrr}
\hline
Problem Set & hCRO & RaPS & GRA & CFT & BeCh & IGA & PROG & Be & Greedy \\ \hline
4 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.57 & 0.06 & 3.78 \\
5 & 0.00 & 0.00 & 0.00 & 0.00 & 0.09 & 0.00 & 0.88 & 0.18 & 5.51 \\
6 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.69 & 0.56 & 7.22 \\
A & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.75 & 0.82 & 5.61 \\
B & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.81 & 5.57 \\
C & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.87 & 1.93 & 6.88 \\
D & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.32 & 0.00 & 2.75 & 9.79 \\
NRE & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 3.50 & 12.75 \\
NRF & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 1.43 & 7.16 & 12.98 \\
NRG & 0.00 & 0.00 & 0.00 & 0.00 & 0.13 & 0.13 & 1.18 & 4.83 & 8.49 \\
NRH & 0.00 & 0.00 & 0.00 & 0.00 & 0.63 & 1.30 & 1.68 & 8.12 & 11.78 \\ \hline
Overall & 0.00 & 0.00 & 0.00 & 0.00 & 0.08 & 0.16 & 0.72 & 2.36 & 8.21 \\ \hline
Opt. Found & 65/65 & 65/65 & 65/65 & 65/65 & 61/65 & 61/65 & 22/65 & 20/65 & 0/65 \\
\hline
\end{tabular}
\label{tab:comp1}
\end{table*}
From Tables \ref{tab:cro1} and \ref{tab:cro2}, we can see that hCRO has excellent performance in non-unicost instances, and all 65 optimum solutions are generated in every run for all instances. For unicost-SCP instances, hCRO obtains 12 optimums out of 15. For the remaining instances, hCRO can also generate a satisfactory result (error percentage around 1\% of best results).
In order to further demonstrate the performance of hCRO, we also compare hCRO with other algorithms proposed in the recent literature. We compare the performance on non-unicost instances of hCRO with the Lagrangian heuristic by Beasley (Be) \cite{Beasley1990LagrangianHeuristicSet}, the genetic algorithm by Beasley and Chu (BeCh) \cite{BeasleyChu1996GeneticAlgorithmSet}, the Lagrangian heuristic by Caprara \textit{et al.} (CFT) \cite{CapraraFischettiToth1999HeuristicMethodSet}, a probabilistic greedy search heuristic by Haouari and Chaouachi (PROG) \cite{HaouariChaouachi2002probabilisticgreedysearch}, an indirect genetic algorithm by Aickelin (IGA) \cite{Aickelin2002IndirectGeneticAlgorithm}, a metaheuristic for randomized priority search by Lan \textit{et al.} (RaPS) \cite{LanDePuyWhitehouse2007EffectiveAndSimple}, and a metaheuristic algorithm based on gravity by Balachandar and Kannan (GRA) \cite{BalachandarKannan2010MetaheuristicAlgorithm}. The comparison is presented in Table \ref{tab:comp1}, where the row of ``Opt. Found" is the number of instances in which the corresponding algorithm finds the optimal value or the best known solution out of the all 65 instances. The table shows the average gap of the global optimal found by the algorithm and BKS. For example, an algorithm generates an optimal value of 52 on a problem instance with a best known value of 50, the gap is defined as $(52-50)/50=4\%$. From the table we can see, hCRO is one of the best four algorithms that can find the optimal value or best known solution 100\% of the time.
We further demonstrate the performance of hCRO by comparing the simulation results on unicost instances with other algorithms. These algorithms include the heuristic random approximation by Peleg \textit{et al.} (RR) \cite{PelegSchechtmanWool1997RandomizedApproximationBounded}, a greedy heuristic by Chvatal (Gr) \cite{Chvatal1979GreedyHeuristicSet}, three algorithms proposed by Grossman and Wool in \cite{GrossmanWool1997ComputationalExperienceWith} (Alt-Gr, NN, and R-Gr), and a mean-field approach by Ohlsson \textit{et al.} (MF) \cite{OhlssonPetersonSoderberg2001EfficientMeanField}. The results are presented in Table \ref{tab:comp2}. From the comparison we can see hCRO again outperforms all other algorithms dramatically. Therefore, hCRO is an effective algorithm in solving both non-unicost and unicost set covering problems.
\begin{table*}
\centering
\caption{Performance Gap Comparison on Unicost Instances (\%)}
\begin{tabular}{l|rrrrrrrrr}
\hline
Problem Set & hCRO & RR & Gr & Alt-Gr & NN & R-Gr & MF \\ \hline
Overall & 0.62 & 20.34 & 10.50 & 7.29 & 10.07 & 9.85 & 6.51 \\ \hline
\end{tabular}
\label{tab:comp2}
\end{table*}
\subsection{Analysis on Performance Contribution of the Proposed Heuristic Schemes and CRO Framework}
In order to analyze the contribution of different proposed heuristic schemes and the CRO framework to the outstanding performance of hCRO, we also construct different hCRO variations and a heuristic-based Genetic Algorithm (hGA).
To determine the impact of our initialization operator (Algorithm \ref{alg:init}) and our neighborhood search operator (Algorithm \ref{alg:neigh}), we create alternative operators for these two schemes and construct two new hCRO variations:
\begin{itemize}
\item hCRO/IR: This variation of hCRO utilizes our previously proposed perturbation heuristic as the neighborhood search operator, with a ``random pick scheme'' as the initial population generator. In the ``random pick scheme'', we first construct a permutation vector of all elements in $\mathbb{M}$. We then start from the first element $i$ in this permutation vector and find all subsets $\mathbb{X}_i$ that cover $i$. Then we randomly select one such subset and go on to the next element $j$. If $j$ is not covered yet, we repeat the previous step, i.e. find all subsets that cover $j$ and randomly select one. If $j$ is already covered, we go to the next element. This process repeats until all elements are covered by at least one subset.
\item hCRO/NR: This variation of hCRO utilizes our previously proposed reverse cumulative scheme as the initial population generator, with a ``remove-repair scheme'' as the neighborhood search operator. In the ``remove-repair scheme'', we first randomly remove one subset in the solution. Then we determine all elements not yet covered by any subsets. The remaining part is similar to the ``random pick scheme'', where we build a permutation vector, find all subsets that cover the current element and randomly select one such subset.
\end{itemize}
Besides hCRO/IR and hCRO/NR, we also develop hGA to solve SCP in order to reveal the contribution of the CRO framework in hCRO:
\begin{itemize}
\item hGA: We utilize our previously proposed reverse cumulative scheme as the initial population generator. We use the inter-molecular ineffective collision operator in hCRO as the crossover operator in hGA, i.e., two perturbation heuristics are executed simultaneously. As for the mutation operator, we adopt the decomposition operator in hCRO and select the better-performing solution as the mutated chromosome. We set the population size the same with hCRO. The crossover rate and mutation rate are set at 0.8 and 0.2, respectively, which is a commonly used combination of parameters for GA.
\end{itemize}
We perform simulations of hCRO, hCRO/IR, hCRO/NR, and hGA on both unicost and non-unicost instances. The simulation results are presented in Table \ref{tab:comp3}.
\begin{table}
\centering
\caption{Performance Gap Comparison on hCRO, hCRO/IR, hCRO/NR, and hGA (\%)}
\begin{tabular}{l|rrrr}
\hline
Problem Set & hCRO & hCRO/IR & hCRO/NR & hGA \\ \hline
Non-unicost & 0.00 & 0.04 & 1.78 & 0.11 \\
Unicost & 0.62 & 2.19 & 14.64 & 3.82 \\ \hline
\end{tabular}
\label{tab:comp3}
\end{table}
From the results we can see that hCRO always performs the best. While hCRO/IR can generate similar results as hCRO, there is still a small gap between the performance of the two algorithms. This demonstrates the superiority of the reverse cumulative scheme. hCRO/NR performs much worse than all other algorithms, and this observation underlines the importance of the perturbation heuristic in generating good results. The results also show that the CRO framework is superior to conventional problem solver frameworks like GA. This is probably due to the unique energy conservation design in CRO as well as its tolerant, energy-related individual selection pattern \cite{LamLi2010ChemicalReactionInspired}.
\section{Conclusion}
In this paper we develop a heuristic-based CRO algorithm to solve non-unicost and unicost SCP. This algorithm introduces two heuristics into the operators of CRO. We study the performance of hCRO with a series of benchmark test instances from the Beasley's OR Library and show that hCRO enjoys superior performance in terms of the solution quality when compared with other algorithms. hCRO is able to find all 65 optimal solutions in non-unicost instances and it demonstrates outstanding performance when applied to unicost SCP. We also perform a series of test over different variations of hCRO as well as a heuristic-based GA to demonstrate the contribution of these two heuristics and the CRO framework on the final performance.
Further research includes the improvement of the performance for huge SCP problems. We can also introduce some implementation techniques that are commonly used by other SCP algorithms, e.g., group memory between iterations, to CRO. Some of the heuristics developed in our proposed algorithm can also be applied to solve other applications of CRO, such as the bin packing problem and the multi-dimensional knapsack problem. Last but not least, we can use the proposed algorithm to solve real-world applications of SCP.
\section*{Acknowledgement}
This research is supported in part by the University of Hong Kong Strategic Research Theme on Computation and Information. A.Y.S. Lam is supported in part by the Faculty Research Grant of Hong Kong Baptist University, under Grant No. FRG2/13-14/045.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Chain conditions provide a measure of how \emph{small} is a space, from a topological point of view. For example, a space has the \emph{countable chain condition} (ccc) if it does not contain an uncountable family of pairwise disjoint non-empty open sets. This is one of the weakest chain conditions one can put on a space, while separability may be considered one of the strongest. Of course, every separable space satisfies the ccc.
Todorcevic \cite{T} surveys a wealth of chain conditions that are between the ccc and separability, elaborating on their \emph{classifying power}, in the sense that the better the space, the greater number of chain conditions it identifies. A few examples of this phenomenon are Knaster's result that separability and \emph{Knaster's property} (that is, every uncountable family of open sets contains an uncountable family where each pair of elements meets) are equivalent on ordered continua and Shapirovskii's result that separability and the \emph{Shanin condition} (that is, point-countable families of open sets are countable) are equivalent for compact spaces of countable tightness. These are all ZFC results, but the picture is even clearer in certain models of set theory: for example, the ccc and separability are equivalent for linearly ordered spaces under $MA_{\omega_1}$. Some of these characterizations even offer topological equivalents of certain combinatorial principles. Let us just mention Todorcevic and Velickovic's result that $MA_{\omega_1}$ is equivalent to the statement that every compact first-countable ccc space is separable.
In more recent times the framework of selection principles in mathematics and topological games has been applied to chain conditions offering more examples of the phenomenon underlined by Todorcevic. A typical selective chain condition is the one considered by the first-named author in \cite{A} which states that one can diagonalize a family with a dense union from a countable sequence of maximal pairwise disjoint families of non-empty open sets. Daniels, Kunen and Zhou introduced a game-theoretic strengthening of this property in \cite{DKZ} by using a two-player game where each player plays an inning per natural number and at a given inning, the first player picks a maximal pairwise disjoint open family, while the second player picks an open set from it. The second player wins if the set of all open sets he picked has dense union. If the second player has a winning strategy in this game on a given space $X$, then $X$ is selectively ccc, which in turn implies that $X$ is ccc. Unfortunately, there are countable spaces failing the selective versions of the ccc (see \cite{A}), so separability alone does not appear to play any role in this context. What should take the role of separability when dealing with selective chain conditions is countable $\pi$-weight. Recall that a \emph{$\pi$-base} for a space $X$ is a family $\mathcal{P}$ of non-empty open sets such that for every non-empty open $U \subset X$ there is $P \in \mathcal{P}$ such that $P \subset U$. The $\pi$-weight of a space is then the minimum cardinality of a $\pi$-base. It is easy to see that in every space with a countable $\pi$-base the second player has a winning strategy in the ccc game. The second-named author proved in \cite{BS} that this is actually equivalent to having a countable $\pi$-base for spaces with a countable local $\pi$-base at every point. A selective version and a game-theoretic version of separability were defined by Scheepers in \cite{SchSep} in a similar way. The former turns out to be equivalent to a countable $\pi$-base for compact spaces and the latter is equivalent to a countable $\pi$-base for all regular spaces. These results seem to suggest that the same classifying ability of traditional chain condition is shared by their selective versions, as long as one takes countable $\pi$-weight as the \emph{ultimate selective chain condition}.
Even though they measure the \emph{topological smallness} of a space, chain conditions seldom put any bound on its cardinality. Indeed, while a separable regular space cannot have cardinality larger than $2^{\mathfrak{c}}$, there are ccc spaces of arbitrarily large cardinality: it suffices to consider the Cantor Cube $2^\kappa$, where $\kappa$ is any cardinal. Nonetheless, chain conditions feature prominently in a few classical cardinal inequalities. A result of Hajnal and Juh\'asz states that the cardinality of a ccc first countable space does not exceed the continuum. An interesting partial generalization of this theorem, that also generalizes Arhangel'skii's theorem on the cardinality of first-countable compacta, is due to Bell, Ginsburg and Woods. They proved that the cardinality of a first-countable weakly Lindel\"of normal space does not exceed the continuum. Weakly Lindel\"of means that every open cover has a countable subcollection with dense union, a condition which is easily seen to be satisfied by all ccc spaces. The question of whether normality can be relaxed to regularity in this result has remained open, but Angelo Bella gave a partial answer to it in \cite{BS} by considering the natural game-theoretic strengthening of the weak Lindel\"of property. In the third section of our paper we prove this game is equivalent to a sort of dual of Berner and Juh\'asz's classical point-picking game and exploit this equivalence to give a short proof to Bella's Theorem.
Another reason for our interest in game-theoretic strengthenings of chain conditions is that they provide an unexpected ZFC partial positive answer to the old problem of the productivity of the ccc. It was already known by Kurepa that the square of a Suslin Line is not ccc. Thus, consistently, the countable chain condition is not productive. Both Knaster's property and the Shanin condition are productive, and one can use these results along with the above mentioned equivalences to prove that under $MA_{\omega_1}$ the ccc is productive. In \cite{DKZ} Daniels, Kunen and Zhou proved in ZFC that if player II has a winning strategy in the ccc game on every factor of an arbitrary product then it also has it on the full product.
In section 2 we deal with selective properties. In particular, we characterize the selective ccc on Pixley-Roy hyperspaces and we give a consistent topological characterization of $cov(\mathcal{M})$ by means of the weak Rothberger property.
In section 3 we concentrate on game versions. We prove preservation results in finite unions and products and give a characterization of the weak Rothberger game that we then exploit to prove cardinal inequalities in topological spaces. We then construct counterexamples proving the sharpness of our inequalities.
Given a space $X$ with topology $\tau$, we fix some notation about families of open covers and families of subsets of $X$.
\begin{itemize}
\item $\mathcal{O}=\{\mathcal{U}: \mathcal{U} \subset \tau \wedge \bigcup \mathcal{U}=X \}$.
\item $\mathcal{O}_D=\{\mathcal{U}: \mathcal{U} \subset \tau \wedge \overline{\bigcup \mathcal{U}}=X \}$.
\item $\Omega=\{\mathcal{U}: \mathcal{U} \subset \tau \wedge (\forall F \in [X]^{<\omega} \exists O \in \mathcal{U}: F \subset O \}$.
\item $\mathcal{D}= \{D \subset X: \overline{D}=X\}$.
\item $\mathcal{D}_O=\{O \in \tau: \overline{O}=X \}$.
\end{itemize}
An element of $\Omega$ is usually known as an \emph{$\omega$-cover} of $X$.
Let's recall some basic selection principles and two-person infinite games we will deal with in our paper. Let $\mathcal{A}, \mathcal{B} \subset \mathcal{P}(X)$.
\begin{itemize}
\item We say that $X$ satisfies $S^\kappa_1(\mathcal{A}, \mathcal{B})$ if for every sequence $\{A_\alpha: \alpha< \kappa \} \subset \mathcal{A}$, we can choose $B_\alpha \in A_\alpha$ such that $\{B_\alpha: \alpha < \kappa \} \in \mathcal{B}$.
\item We say that $X$ satisfies $S^\kappa_{fin}(\mathcal{A}, \mathcal{B})$ if for every sequence $\{A_\alpha: \alpha < \kappa \} \subset \mathcal{A}$, we can choose $B_\alpha \in [A_\alpha]^{<\omega}$ such that $\bigcup \{B_\alpha: \alpha < \kappa \} \in \mathcal{B}$.
\item We denote by $G^\kappa_1(\mathcal{A}, \mathcal{B})$ the two-person game in $\kappa$ many innings such that, at inning $\alpha<\kappa$, player one picks $A_\alpha \in \mathcal{A}$ and player two picks $B_\alpha \in A_\alpha$. Player two wins if $\{B_\alpha: \alpha < \kappa \} \in \mathcal{B}$.
\item We denote by $G^\kappa_{fin}(\mathcal{A}, \mathcal{B})$ the two-person game in $\kappa$ many innings such that, at inning $\alpha<\kappa$, player one picks $A_\alpha \in \mathcal{A}$ and player two picks $B_\alpha \in [A_\alpha]^{<\omega}$. Player two wins if $\bigcup \{B_\alpha: \alpha < \kappa \} \in \mathcal{B}$.
\end{itemize}
The properties $S^\omega_1(\mathcal{O}, \mathcal{O})$ and $S^\omega_{fin}(\mathcal{O}, \mathcal{O})$ are now known as \emph{Rothberger} and \emph{Menger} respectively. Moreover, we recall that $X$ satisfies $S^\omega_1(\Omega, \Omega)$ if and only if every finite power of $X$ is Rothberger.
\section{Selective versions}
In \cite{SchSep}, Marion Scheepers defined a natural selective version of separability which is now known as $R$-separability (see \cite{BBM}). The $R$ in the name comes, of course, from Rothberger.
\begin{definition}
A space is \emph{$R$-separable} if it satisfies $S^\omega_1(\mathcal{D}, \mathcal{D})$. In other words, for every sequence $\{D_n: n < \omega \}$ of dense sets there is a point $x_n \in D_n$, for every $n<\omega$ such that $\{x_n: n < \omega \}$ is dense in $X$.
\end{definition}
$R$-separability is much stronger than separability. For example, it implies that every dense set is separable, and by a result of Juh\'asz and Shelah \cite{JS}, this is equivalent to countable $\pi$-weight in the realm of compact spaces. Moreover, there are even examples of countable spaces that are not $R$-separable (see \cite{BBMadd} and \cite{SSS}).
In \cite{A} the first named author introduced the following weakening of $R$-separability, following a suggestion of Sakai.
\begin{definition}
We say that the space $X$ has property $S$ if it satisfies $S^\omega_1(\mathcal{D}_O, \mathcal{D})$, that is, for every sequence $\{O_n: n < \omega \}$ of open dense sets we can pick points $x_n \in O_n$ such that $\{x_n: n < \omega \}$ is dense.
\end{definition}
One of the most interesting features about property $S$ is that it lies strictly between $R$-separability and a natural selective version of the countable chain condition that was introduced by Scheepers in \cite{Schccc}.
\begin{definition}
We say that $X$ is \emph{selectively ccc} if it satisfies $S^\omega_1(\mathcal{O}_D, \mathcal{O}_D)$. In other words, for every sequence $\{\mathcal{U}_n: n < \omega \}$ of open families, such that $\bigcup \mathcal{U}_n$ is dense for every $n<\omega$ we can pick an open set $U_n \in \mathcal{U}_n$ such that $\bigcup \{U_n: n <\omega \}$ is dense.
\end{definition}
Obviously, every space having property S is selectively ccc, but the converse does not hold, since $2^\kappa$ is selectively ccc for every cardinal $\kappa$ (see \cite{A}), but fails to have property $S$ for $\kappa > \mathfrak{c}$ because it's not even separable then.
Let $X$ be a space and set $PR(X)=[X]^{<\omega}$. There is a natural topology on $PR(X)$ called the Pixley-Roy topology. A basic open neighbourhood of $F \in PR(X)$ is a set of the form $[F,U]=\{G \in PR(X): F \subset G \subset U \}$, where $F \subset U$ and $U \subset X$ is open in the topology on $X$.
While other selective properties have been characterized on $PR(X)$ (see, for example, \cite{Sak} and \cite{BCPT}), the problem of characterizing the selective ccc of $PR(X)$ has remained open. We introduce a new selection principle to provide such a characterization and then reduce the selective ccc of the Pixley-Roy hyperspace of a separable metrizable space $X$ to a well-known selective covering property of $X$.
\begin{definition}
Let $(X,\tau)$ be a topological space, and $\mathcal{F} \subset \{(F,U) \in [X]^{<\omega} \times \tau: F \subset U \}$. We call $\mathcal{F}$ an \emph{$\omega$-double cover} if for every pair $(G,V) \in [X]^{<\omega} \times \tau$ such that $G \subset V$ there is $(F,U) \in \mathcal{F}$ such that $F \subset V$ and $G \subset U$. The family of all $\omega$-double covers will be indicated with $\Omega^2$.
\end{definition}
\begin{proposition} \label{simpleprop}
$S^\omega_1(\Omega^2, \Omega^2) \Rightarrow S^\omega_1(\Omega, \Omega)$.
\end{proposition}
\begin{proof}
Just note that if $\mathcal{U}_n$ is an $\omega$-cover then $\{(\emptyset,U): U \in \mathcal{U}_n \}$ is an $\omega$-double cover.
\end{proof}
\begin{theorem}
$PR(X)$ has $S^\omega_1(\mathcal{O}_D, \mathcal{O}_D)$ if and only if $X$ has $S^\omega_1(\Omega^2, \Omega^2)$.
\end{theorem}
\begin{proof}
For the direct implication, let $\{\mathcal{F}_n: n <\omega \} \subset \Omega^2$. Now let $\mathcal{O}_n=\{[F,U]: (F,U) \in \mathcal{F}_n \}$. Then $\{\mathcal{O}_n: n < \omega \} \subset \mathcal{O}_D(PR(X))$. By the $S^\omega_1(\mathcal{O}_D, \mathcal{O}_D)$ property of $PR(X)$ we can find $[F_n, U_n] \in \mathcal{O}_n$ such that $\{[F_n, U_n]: n < \omega \} \subset \mathcal{O}_D$. We then have that $\{(F_n,U_n): n < \omega \} \in \Omega^2$. Indeed, let $(F,U) \in \{(F,U) \in [X]^{<\omega} \times \tau: F \subset U \}$. Then $[F,U] \cap [F_n,U_n] \neq \emptyset$ for some $n$. So there is $H$ extending both $F$ and $F_n$ such that $H \subset U_n$ and $H \subset U$, and that implies $F \subset U_n$ and $F_n \subset U$.
Vice versa, suppose $X$ has $S^\omega_1(\Omega^2, \Omega^2)$ and let $\{\mathcal{U}_n: n <\omega \}$ be a sequence of open families with dense union in $PR(X)$. Without loss of generality we can assume that $\mathcal{U}_n$ is made up of basic open sets. Let now $\mathcal{F}_n=\{(F,U): [F,U] \in \mathcal{U}_n\}$. Then $\mathcal{F}_n \in \Omega^2$, for every $n<\omega$. Hence we can find $(F_n,U_n) \in \mathcal{F}_n$ such that $\{(F_n, U_n): n < \omega \} \in \Omega^2$. Now by the same argument $\{[F_n,U_n]: n < \omega \}$ has dense union in $PR(X)$.
\end{proof}
\begin{corollary}
If $PR(X)$ satisfies $S^\omega_1(\mathcal{O}_D, \mathcal{O}_D)$ (that is, $PR(X)$ is selectively ccc) then $X$ satisfies $S^\omega_1(\Omega, \Omega)$.
\end{corollary}
\begin{corollary} \label{opensub}
If $X$ satisfies $S^\omega_1(\Omega^2, \Omega^2)$ and $U \subset X$ is open then $U$ satisfies $S^\omega_1(\Omega^2, \Omega^2)$.
\end{corollary}
\begin{proof}
If $X$ satisfies $S^\omega_1(\Omega^2, \Omega^2)$ then $PR(X)$ is selectively ccc. Now $PR(U)=[\emptyset, U]$ is an open subspace of a selectively ccc space and hence it's selectively ccc. But then $U$ satisfies $S^\omega_1(\Omega^2, \Omega^2)$.
\end{proof}
\begin{theorem}
Let $X$ be a second-countable space. Then $X$ is $S^\omega_1(\Omega^2, \Omega^2)$ if and only if $X$ is $S^\omega_1(\Omega, \Omega)$.
\end{theorem}
\begin{proof}
The direct implication is clear by Proposition $\ref{simpleprop}$. For the converse implication, let $\{B_n: n < \omega \}$ be a countable base for $X$ which is closed under finite unions. Let $\{I_k: k < \omega \}$ be a partition of $\omega$ into infinite sets. Let $\{\mathcal{U}_n: n < \omega \} \subset \Omega^2$. We can assume without loss of generality that for every $n<\omega$ and for every $(F,U) \in \mathcal{U}_n$ there is $k<\omega$ such that $U \subset B_k$. Moreover, we can assume that for every $n< \omega$, and for every $(F,U) \in \mathcal{U}_n$, if $V$ is an open subset of $X$ such that $F \subset V \subset U$ then $(F,V) \in \mathcal{U}_n$. For every $n \in I_k$, define families $\mathcal{V}_n$ as follows:
$$\mathcal{V}_n=\{U: (F,U) \in \mathcal{U}_n \wedge U \subset B_k \}$$
Then $\mathcal{V}_n$ is an $\omega$-cover of $B_k$ and hence we can pick an element $U_n \in \mathcal{V}_n$ for every $n \in I_k$ such that $\{U_n: n \in I_k \}$ is an $\omega$-cover of $B_k$. Let $F_n \in [X]^{<\omega}$ be such that $F_n \subset U_n$ and $(F_n, U_n) \in \mathcal{U}_n$. We claim that $\mathcal{V}=\{(F_n,U_n): n < \omega \}$ is an $\omega$-double cover for $X$. Indeed, let $(G,V)$ be any pair, where $G \in [X]^{<\omega}$, $V \subset X$ is open and $G \subset V$. Let $k<\omega$ be such that $G \subset B_k \subset V$. Then we can find a $n \in I_k$ such that $G \subset U_n$. Now $F_n \subset U_n \subset B_k \subset V$ and hence $\mathcal{V}$ is actually an $\omega$-double cover for $X$.
\end{proof}
\begin{corollary}
Let $X$ be a separable metrizable space. Then $PR(X)$ is selectively ccc if and only if every finite power of $X$ is Rothberger.
\end{corollary}
The weak Lindel\"of property is a covering property that may be considered a chain condition, since it is a consequence of the ccc. We finish this section by considering the natural question of when a weakly Lindel\"of space satisfies the selective version of weak Lindel\"ofness, that is, $S^\omega_1(\mathcal{O}, \mathcal{O}_D)$. As in \cite{BPS}, we will call this property the \emph{weak Rothberger property}.
\begin{theorem}
Let $X$ be a weakly Lindel\"of space such that $\pi w(X) <cov(\mathcal{M})$. Then $X$ is weakly Rothberger.
\end{theorem}
\begin{proof}
Recall that the space $\omega^\omega$ with its usual topology is homeomorphic to the irrationals and $cov(\mathcal{M})$ can be characterized as the least cardinal of a cover of the irrationals by nowhere dense sets.
Let $\{B_\alpha: \alpha < \kappa \}$ enumerate a $\pi$-base of $X$, for some $\kappa < cov(\mathcal{M})$. Let $\{\mathcal{U}_n: n < \omega \}$ be a sequence of open covers. Since $X$ is weakly Lindel\"of we can find a countable subcollection $\{U^n_k: k < \omega \}$ of $\mathcal{U}_n$ such that $\bigcup_{k<\omega} U^n_k$ is dense. Define $N_\alpha$ to be the following set: $$N_\alpha=\{f \in \omega^\omega: (\forall n)(B_\alpha \cap U^n_{f(n)}=\emptyset) \}$$
\noindent {\bf Claim:} The set $N_\alpha$ is nowhere dense in $\omega^\omega$ for every $\alpha < \kappa$.
\begin{proof}[Proof of Claim] It will suffice to prove that for every $\alpha < \kappa$, the set $\omega^\omega \setminus \overline{N_\alpha}$ is dense in $\omega^\omega$. Let $[\sigma]$ be a basic open set in $\omega^\omega$, where $\sigma \in \bigcup_{n<\omega} \omega^n$ and $[\sigma]=\{f \in \omega^\omega: f \supset \sigma \}$. Let $k=dom(\sigma)$ and pick $j<\omega$ such that $B_\alpha \cap U^k_j \neq \emptyset$. Let $\tau=\sigma \cup \{(k,j)\}$ and note that $[\tau] \subset [\sigma]$ and any function in the open set $[\tau]$ misses $N_\alpha$.
\renewcommand{\qedsymbol}{$\triangle$}
\end{proof}
Since $\kappa < cov(\mathcal{M})$ we can pick $f \notin \bigcup_{\alpha < \kappa} N_\alpha$. Then $\{U^n_{f(n)}: n < \omega \}$ is the selection showing that $X$ is weakly Rothberger.
\end{proof}
\begin{theorem}
Under $cov(\mathcal{M})<\mathfrak{b}$ there is a compact space which is not weakly Rothberger and has $\pi$-weight $cov(\mathcal{M})$.
\end{theorem}
\begin{proof}
Assume that $cov(\mathcal{M})<\mathfrak b$. We claim that $\beta (cov(\mathcal{M}))$ is the required example.
We say that $F \subset \omega^\omega$ satisfies property (P) if for every $g \in \omega^\omega$ there is $f \in F$ such that $f(n) \neq g(n)$ for every $n<\omega$.
\noindent {\bf Claim.} There is a family $F \subset \omega^\omega$ of cardinality $cov(\mathcal{M})$ satisfying property (P) such that $f(n) < b(n)$ for some $b \in \omega^\omega$ and every $f \in F$.
\begin{proof}[Proof of Claim]
By Lemma 2.4.2 of \cite{BaJu} there is a family $F' \subset \omega^\omega$ of cardinality $cov(\mathcal{M})$ satisfying (P). Let $\{h_\alpha: \alpha < cov(\mathcal{M}) \}$ be an enumeration of $F'$.
Since $cov(\mathcal{M}) < \mathfrak{b}$, we can fix a function $b \in \omega^\omega$ such that $h_\alpha <^* b$, for every $\alpha < cov(\mathcal{M})$. Now, for every $\alpha < cov(\mathcal{M})$ there is $n_\alpha < \omega$ such that $h_\alpha(n) < b(n)$ for every $n \geq n_\alpha$. For every $\alpha < cov(\mathcal{M})$, $n<\omega$ and $i<b(n)$ define:
$$h^i_\alpha(n)=
\begin{cases}
i & \mbox{if } n<n_\alpha \\
h_\alpha(n) & \mbox{if } n \geq n_\alpha
\end{cases}
$$
It's easy to see that $F=\{h^i_\alpha: i < b(n), \alpha < cov(\mathcal{M}) \}$ is a subfamily of $\omega^\omega$ satisfying (P) and such that $h(n) < b(n)$, for every $h \in F$.
\renewcommand{\qedsymbol}{$\triangle$}
\end{proof}
Let $\{f_\alpha: \alpha < cov(\mathcal{M})\}$ be an enumeration of $F$ and for every $n$ let us consider the following finite clopen cover of
$\beta(cov(M))$: $\mathcal U_n=\{U^n_k:k<b(n)\}$, where
$U^n_k = \beta(\{\xi<cov(\mathcal{M}) : f_\xi(n)=k \})$ and $\beta(A)$ is the set of all
ultrafilters on $cov(\mathcal{M})$ containing $A$.
We claim that the sequence $\{ \mathcal U_n:n\in\omega \}$
witnesses that $\beta(cov(\mathcal{M}))$ fails to have $S^\omega_1(\mathcal O, \mathcal O_D)$.
Indeed, suppose that $\bigcup_{n\in\omega}U^n_{g(n)}$ is dense in $\beta(cov(\mathcal{M}))$
for some $g\in \omega^\omega$. Then, in particular, $cov(\mathcal{M})\subset \bigcup_{n\in\omega}U^n_{g(n)}$.
But since $F$ satisfies property (P), there exists $\xi<cov(\mathcal{M})$ such that
$f_\xi(n)\neq g(n)$ for all $n$, which means that $\xi\not\in U^n_{g(n)}$ for all $n$, and that is a contradiction.
\end{proof}
\begin{corollary}
Assume $cov(\mathcal{M})<\mathfrak{b}$. Then $cov(\mathcal{M})$ can be characterized as the minimum $\pi$-weight of a non-weakly Rothberger weakly Lindel\"of space.
\end{corollary}
\begin{question}
Is it true in ZFC that $cov(\mathcal{M})$ is the minimum $\pi$-weight of a non-weakly Rothberger weakly Lindel\"of space?
\end{question}
\section{Game versions}
The property $S$ defined in the previous section has a natural game version.
\begin{definition}
We say that $X$ has property $S^+$ if the second player has a winning strategy in the game $G^\omega_1(\mathcal{D}_O, \mathcal{D})$. This is the two person game of countable length where at inning $n$ player one picks a dense open set $O_n$ and player two picks a point $x_n \in O_n$. Player two wins if the set of all points he picked is dense in $X$.
\end{definition}
\begin{lemma} \label{openlemma}
Property $S^+$ is hereditary for open sets.
\end{lemma}
\begin{proof}
Let $X$ be a space with property $S^+$ and $U \subset X$ be a non-empty open subset. Let $\tau$ be a winning strategy for player II in $G^\omega_1(\mathcal{D}_O, \mathcal{D})$ on $X$. Let $\sigma$ be the strategy assigning to the open dense subset $O$ of $U$ the point $\tau(O \cup Int(X \setminus W))$ if this last point is in $O$, and any point of $O$ otherwise. Then $\sigma$ is a winning strategy for player II in $G^\omega_1(\mathcal{D}_O, \mathcal{D})$ on $O$.
\end{proof}
The following fact is also clear.
\begin{lemma}
If $D$ is dense in $X$ and $D$ has property $S^+$ then $X$ also has property $S^+$.
\end{lemma}
\begin{proposition}
Property $S^+$ is preserved by finite unions.
\end{proposition}
\begin{proof}
Once this is proved for the union of two spaces, the result will follow easily by induction, so let $X$ be a topological space and $A$ and $B$ be subspaces with property $S^+$ such that $X=A \cup B$. If $Int(\overline{A}) \cap A=\emptyset$ then $B$ is dense in $X$ and we are done. Similarly, if $Int(\overline{B}) \cap B = \emptyset$ we are done. So we can assume that $Int(\overline{A}) \cap A$ and $Int(\overline{B}) \cap B$ are both non-empty subsets of $X$. By Lemma $\ref{openlemma}$ we can fix a winning strategy $\sigma_A$ for player two on $Int(\overline{A}) \cap A$ and a winning strategy $\sigma_B$ for player two on $Int(\overline{B}) \cap B$ in the game $G^\omega_1(\mathcal{D}_O, \mathcal{D})$.
The set $Int(\overline{A}) \cup Int(\overline{B})$ is a dense open subset of $X$. Note that if $U$ is dense open in $X$ then $U \cap Int(\overline{A}) \cap A$ is dense open in $Int(\overline{A}) \cap A$ and $U \cap Int(\overline{B}) \cap B$ is dense open in $Int(\overline{B}) \cap B$. We are now going to define a strategy $\sigma$ for player two on $A \cup B$.
Suppose that in the first $n$ innings, player two played the following sequence of dense open sets $(U_i: i \leq n)$.
If $n=2k$ for some $k<\omega$, we let
$$\sigma((U_i: i \leq n))=\sigma_A((U_{2i} \cap Int(\overline{A}) \cap A: i \leq k))$$
If $n=2k+1$ for some $k<\omega$, we let
$$\sigma((U_i: i \leq n))=\sigma_B((U_{2i+1} \cap Int(\overline{B}) \cap B: i \leq k))$$
Note now that, by the definition of $\sigma_A$ and $\sigma_B$ we have that:
$$\overline{\{\sigma((U_i: i \leq n)): n < \omega\}}=$$
$$=\overline{\{\sigma_A((U_{2i}: i \leq k)): k < \omega \}} \cup \overline{\{\sigma_B((U_{2i+1}: i \leq k)): k <\omega \}}=$$
$$=\overline{Int(\overline{A}) \cap A} \cup \overline{Int(\overline{B}) \cap B}=X$$
So $\sigma$ is a winning strategy for player two in $G^\omega_1(\mathcal{D}_O, \mathcal{D}))$ on $X$ and we are done.
\end{proof}
In a similar way, one can prove the following propositions:
\begin{proposition}
If player two has a winning strategy in $G^\omega_{fin}(\mathcal{D}_O, \mathcal{D})$ on $A_i$, for every $i \leq n$ then he also has a winning strategy in that game on $A_1 \cup A_2 \cup \dots \cup A_n$.
\end{proposition}
\begin{proposition}
The properties $S^\omega_1(\mathcal{D}_O, \mathcal{D})$ and $S^\omega_{fin}(\mathcal{D}_O, \mathcal{D}))$ are preserved by finite unions.
\end{proposition}
The natural game version of the weak Rothberger property, introduced in the end of the previous section is the game $G^\omega_1(\mathcal{O}, \mathcal{O}_D)$. This is tightly connected to the dual version of the point-picking game studied by Berner and Juh\'asz in \cite{BJ}.
\begin{definition}
Let $G^p_o(\kappa)$ be the following game. At inning $\alpha < \kappa$, player one picks a point $x_\alpha \in X$ and player two picks an open set $U_\alpha$ such that $x_\alpha \in U_\alpha$. Player one wins if and only if $\bigcup_{\alpha < \kappa} U_\alpha$ is dense in $X$.
\end{definition}
We may call this the \emph{open-picking game}. We prove that the game $G^\kappa_1(\mathcal{O}, \mathcal{O}_D)$ and the game $G^p_o(\kappa)$ are \emph{dual}, in the following sense.
\begin{theorem} \label{thmequiv} {\ \\}
\begin{enumerate}
\item \label{playerone} Player one has a winning strategy in $G^p_o(\kappa)$ if and only if player two has a winning strategy in $G^\kappa_1(\mathcal{O}, \mathcal{O}_D)$.
\item \label{playertwo} Player two has a winning strategy in $G^p_o(\kappa)$ if and only if player one has a winning strategy in $G^\kappa_1(\mathcal{O}, \mathcal{O}_D)$.
\end{enumerate}
\end{theorem}
\begin{proof}
The direct implication of ($\ref{playerone}$) is easy to see. Indeed, let $\tau$ be a winning strategy for player one in $G^p_o(\kappa)$. Given an open cover $\mathcal{U}$ let $\sigma((\mathcal{U}))$ be any open set $O \in \mathcal{U}$ such that $\tau(\emptyset) \in O$. Assuming we have defined $\sigma$ for the first $\alpha$ many innings, and $\{\mathcal{O}_\beta: \beta \leq \alpha \}$ be a sequence of open covers, let $\sigma((\mathcal{O}_\beta: \beta \leq \alpha))$ be any open set $O \in \mathcal{O}_\alpha$ such that $\tau((\sigma(\mathcal{O}_\gamma: \gamma \leq \beta): \beta < \alpha)) \in O$. Then $\sigma$ is a winning strategy for player two in $G^\kappa_1(\mathcal{O}, \mathcal{O}_D)$. Indeed, let $(\mathcal{O}_0, O_0, \mathcal{O}_1, O_1, \dots \mathcal{O}_\alpha, O_\alpha, \dots)$ be a play, where player two plays according to $\sigma$. Then $\tau((\sigma(\mathcal{O}_\gamma: \gamma \leq \beta): \beta <\alpha)) \in O_\alpha$, and hence $\bigcup \{O_\alpha: \alpha <\kappa \}$ is dense, as $\tau$ is a winning strategy for player one in $G^p_o(\kappa)$.
To prove the converse implication of ($\ref{playerone}$), $\sigma$ be a winning strategy for player two in $G^\kappa_1(\mathcal{O}, \mathcal{O}_D)$ on some space $X$.
\noindent {\bf Claim}. Let $(\mathcal{O}_\alpha: \alpha < \beta)$ be a sequence of open covers. Then there is a point $x \in X$ such that, for every neighbourhood $U$ of $x$ there is an open cover $\mathcal{U}$ with $U=\sigma((\mathcal{O}_\alpha: \alpha < \beta)^\frown (\mathcal{U}))$.
\begin{proof}[Proof of Claim]
Recalling that $\mathcal{O}$ denotes the set of all open covers of $X$, let $\mathcal{V}=\{V$ open: $(\forall \mathcal{U} \in \mathcal{O})(V \neq \sigma((\mathcal{O}_\alpha: \alpha < \beta)^\frown(\mathcal{U})) \}$. Its definition easily implies that $\mathcal{V}$ cannot be an open cover, and hence there is a point $x \in X \setminus \bigcup \mathcal{V}$. By definition of $\mathcal{V}$ we must have that for every neighbourhood $U$ of $x$ there is an open cover $\mathcal{U}$ such that $U=\sigma((\mathcal{O}_\alpha: \alpha < \beta)^\frown (\mathcal{U}))$ and hence we are done.
\renewcommand{\qedsymbol}{$\triangle$}
\end{proof}
Use the Claim to choose a point $x_0$ such that, for every neighbourhood $U$ of $x_0$ there is an open cover $\mathcal{U}$ with $\sigma(\mathcal{U})=U$ and let $\tau(\emptyset)=x_0$.
Suppose we have defined $\tau$ for the first $\alpha$ many innings. Let now $\{V_\beta: \beta \leq \alpha \}$ be a sequence of open sets and $\{\mathcal{O}_\beta: \beta < \alpha\}$ be a sequence of open covers such that $V_\beta=\sigma((\mathcal{O}_\gamma: \gamma \leq \beta ))$, for every $\beta < \alpha$. Use the claim to choose a point $x_\alpha$ such that, for every open neighbourhood $U$ of $x_\alpha$ there is an open cover $\mathcal{O}$ with $U=\sigma((\mathcal{O}_\beta: \beta < \alpha)^\frown (\mathcal{O}))$ and let $\tau((V_\beta: \beta \leq\alpha))=x_\alpha$.
We now claim that $\tau$ is a winning strategy for player one in $G^p_o(\kappa)$. Indeed, let $(x_0, V_0, x_1, V_1, \dots x_\alpha, V_\alpha, \dots )$ be a play where player one uses strategy $\tau$. Then there must be a sequence of open covers $\{\mathcal{O}_\alpha: \alpha < \kappa \}$ such that $V_\beta=\sigma((\mathcal{O}_\alpha: \alpha < \beta ))$, for every $\beta < \kappa$. Since $\sigma$ is a winning strategy for two in $G^\kappa_1(\mathcal{O}, \mathcal{O}_D)$ then $\bigcup \{V_\alpha: \alpha < \kappa \}$ is dense in $X$ and this proves that $\tau$ is a winning strategy for player one in $G^p_o(\kappa)$.
To prove the direct implication of $(\ref{playertwo})$, let $\tau$ be a winning strategy for player two in $G^p_0(\kappa)$. We define a winning strategy $\sigma$ for player one in $G^\kappa_1(\mathcal{O}, \mathcal{O}_D)$ as follows: in his first move player one plays the open cover $\sigma(\emptyset)= \{\tau(x): x \in X \}$. Assuming we have defined $\sigma$ for every inning $\beta$, with $\beta <\alpha$, let $(U_\beta: \beta < \alpha)$ be a sequence of open sets such that there is a sequence $\{x_\beta: \beta < \alpha \}$ of points with $U_\beta=\tau(x_\gamma: \gamma \leq \beta)$. Then we simply define $\sigma((U_\beta: \beta < \alpha))$ to be $\{\tau((x_\beta: \beta < \alpha)^\frown (x)): x \in X \}$. We claim that $\sigma$ is a winning strategy for player one in $G^\kappa_1(\mathcal{O}, \mathcal{O}_D)$.
Indeed, let $(\mathcal{O}_0, U_0, \dots \mathcal{O}_\alpha, U_\alpha \dots)$ be a play of $G^\kappa_1(\mathcal{O}, \mathcal{O}_D)$ where player one plays according to $\sigma$. Therefore, we can find a sequence $\{x_\alpha: \alpha < \kappa \}$ of points, such that $U_\alpha=\tau((x_\gamma: \gamma < \alpha))$, and hence $\bigcup_{\alpha < \kappa} U_\alpha$ is not dense, since $\tau$ is a winning strategy for player two in $G^p_o(\kappa)$. So $\sigma$ must be a winning strategy for player one in $G^\kappa_1(\mathcal{O}, \mathcal{O}_D)$.
To prove the converse implication of $(\ref{playertwo})$, let $\sigma$ be a winning strategy for player one in $G^\kappa_1(\mathcal{O}, \mathcal{O}_D)$. We will use $\sigma$ to define a winning strategy $\tau$ for player two in $G^p_o(\kappa)$. Given a point $x \in X$, let $\tau((x))$ be any open set $U \in \sigma(\emptyset)$ such that $x \in U$.
Now suppose $\tau$ has been defined for all sequences of points of ordinal length less than $\alpha$. Given a sequence $\{x_\beta: \beta \leq \alpha \} \subset X$, let $\tau((x_\beta: \beta \leq \alpha))$ be any open set $U \in \sigma((\tau((x_\gamma: \gamma \leq \beta): \beta < \alpha))$ such that $x_\alpha \in U$. We claim that $\tau$ thus defined is a winning strategy for player two in $G^p_o(\kappa)$.
Indeed, let $x_0, U_0, x_1, U_1, \dots x_\alpha, U_\alpha \dots$ be a play of $G^p_o(\kappa)$, where player two plays according to $\tau$. Then $U_\alpha \in \sigma((\tau((x_\gamma: \gamma \leq \beta): \beta < \alpha))$, for every $\alpha < \kappa$ and hence, since $\sigma$ is a winning strategy for player I in $G^\kappa_1(\mathcal{O}, \mathcal{O}_D)$ we must have that $\bigcup_{\alpha < \kappa} U_\alpha$ is not dense, and we are done.
\end{proof}
We are now going to exploit this result to give a short proof to a result of Angelo Bella from \cite{BS}.
\begin{theorem} \label{thmang}
Let $(X, \tau)$ be a regular space. Suppose that player two has a winning strategy in $G^{\kappa}_1(\mathcal{O}, \mathcal{O}_D)$. Then $d(X) \leq \chi(X)^{<\kappa}$
\end{theorem}
\begin{proof}
Using Theorem $\ref{thmequiv}$ fix a winning strategy $\sigma$ for player one in $G^p_o(\kappa)$. Let $M$ be a $<\kappa$-closed elementary submodel of $H(\theta)$, for large enough regular $\theta$ such that $|M|=\chi(X)^{<\kappa}$, $(X,\tau), \sigma \in M$ and $\chi(X) +1 \subset \kappa$.
We claim that $X \cap M$ is dense in $X$. Suppose not, and let $V \subset X$ be an open set such that $\overline{V} \cap X \cap M=\emptyset$. For every $x \in X \cap M$, let $\mathcal{U}_x \in M$ be a local base of size $\leq \chi(X)$. Since $\chi(X)+1 \subset M$ we have $\mathcal{U}_x \subset M$, and hence we can find in $M$ a neighbourhood $U_x$ of $x$ such that $U_x \cap V=\emptyset$.
Since we have both $X \in M$ and $\sigma \in M$, the first move of player one $\sigma((\emptyset)):=x_0$ is a point of $M$. Let player two pick the open set $U_{x_0}$
Suppose that for some countable ordinal $\alpha$, player two picked open set $U_{x_\beta}$ at inning $\beta$ for every $\beta < \alpha$. Note that $\{U_{x_\beta}: \beta < \alpha \} \subset M$ and since $M$ is $<\kappa$ closed we have $\{U_\beta: \beta < \alpha \} \in M$. Therefore $x_\alpha:=\sigma((X)^\frown (U_\beta: \beta < \alpha)) \in M$. Let player two play $U_{x_\alpha}$ at the $\alpha$th inning. Since $\sigma$ is a winning strategy for player one, we must have that $\bigcup \{U_{x_\alpha}: \alpha < \omega_1 \}$ is a dense set. But this is impossible, because $V \cap U_{x_\alpha}=\emptyset$ for every $\alpha < \kappa$.
\end{proof}
\begin{corollary}
(A. Bella) Let $X$ be a first-countable regular space. If player two has a winning strategy in $G^{\aleph_1}_1(\mathcal{O}, \mathcal{O}_D)$ then $|X| \leq 2^{\aleph_0}$.
\end{corollary}
\begin{proof}
Every first countable space with a dense set of cardinality continuum has cardinality at most $(2^{\aleph_0})^{\aleph_0}=2^{\aleph_0}$.
\end{proof}
\begin{corollary}
Every first countable regular space where player II has a winning strategy in $G^\omega_1(\mathcal{O}, \mathcal{O}_D)$ is separable.
\end{corollary}
\begin{theorem}
Let $X$ be a first countable regular space. If player two has a winning strategy in $G^{\aleph_0}_1(\mathcal{O}, \mathcal{O}_D)$ then $X$ is separable.
\end{theorem}
The above theorem would lead one to conjecture that if player two has a winning strategy in $G^{\aleph_1}_1(\mathcal{O}, \mathcal{O}_D)$ on a first countable regular space $X$, then $X$ should have density $\aleph_1$, but in Example $\ref{exdensity}$ we are going to show that this is not the case, not even if $G^{\omega_1}_1(\mathcal{O}, \mathcal{O}_D)$ is replaced with the stronger (for player II) game $G^{\omega_1}_1(\mathcal{O}, \mathcal{O})$.
\begin{lemma} \label{lemmaex}
Assume $cov(\mathcal{M})>\aleph_1+ \mathfrak{c}=\aleph_2$. Then there is a set $Y \subset \mathbb{I}=[0,1]$ of cardinality $\aleph_2$ such that the intersection of $Y$ with every meager set of $\mathbb{I}$ has cardinality at most $\aleph_1$.
\end{lemma}
\begin{proof}
$MA_{\omega_1}$ implies that $\mathbb{I}$ is not the union of $\aleph_1$ many nowhere dense sets.
Note that $\mathbb{I}$ has continuum many closed sets, so we can use $\mathfrak{c}=\omega_2$ to fix an enumeration $\{F_\alpha: \alpha < \omega_2 \}$ of the closed nowhere dense subsets of $\mathbb{I}$. We are going to construct $Y$ by transfinite induction. Suppose we have chosen points $\{y_\alpha: \alpha < \beta \} \subset \mathbb{I}$, where $\beta < \omega_2$. Choose any point $y_\beta \in \mathbb{I} \setminus (\bigcup_{\alpha < \beta} F_\alpha \cup \{y_\alpha: \alpha < \beta \})$. Then $Y=\{y_\alpha: \alpha < \omega_2\}$ is the desired set.
\end{proof}
\begin{example} \label{exdensity}
($cov(\mathcal{M})>\aleph_1+\mathfrak{c}=\omega_2$) A first countable regular space $X$ such that player two has a winning strategy in $G^{\omega_1}_1(\mathcal{O}, \mathcal{O})$ and $d(X)>\aleph_1$.
\end{example}
\begin{proof}
Recall the construction of the Alexandroff Double $\mathbb{D}$ of the unit interval. We define a topology on $\mathbb{D}=\mathbb{I} \times \{0,1\}$ by declaring every point of $\mathbb{I} \times \{1\}$ to be isolated and declaring a neighbourhood of a point $(x,0) \in \mathbb{I} \times \{0\}$ to be of the form $U \times \{0\} \cup U \times \{1\} \setminus F$, where $U$ is an Euclidean open set and $F$ is a finite set. It is easy to see that $\mathbb{D}$ is a compact Hausdorff space with points $G_\delta$, and hence it's first countable.
Let now $Y$ be the set constructed in Lemma $\ref{lemmaex}$. Without loss of generality we can assume that $Y$ is dense in $\mathbb{I}$. Now, consider the space $X= Y \times \{0,1\}$ with the topology induced by $\mathbb{D}$. Then $X$ is a regular first countable space of density $\aleph_2$. Fix a countable dense set $D \subset Y$. We let $C=D \times \{0\}$.
\noindent {\bf Claim.} The complement of every open set containing $C$ has cardinality at most $\aleph_1$.
\begin{proof}[Proof of Claim] Let $U$ be an open subset of $X$ such that $C \subset U$. Since $Y$ is hereditarily Lindelof, we can assume that $U=\bigcup_{n<\omega} (U_n \times \{0\} \cup U_n \times \{1\} \setminus F_n)$, where $U_n$ is the trace on $Y$ of an Euclidean open set and $F_n$ is finite. Let $O=\bigcup_{n<\omega} U_n$ and let $V$ be an open subset of $\mathbb{I}$ such that $V \cap Y=O$. Since $V$ is dense in $\mathbb{I}$, $\mathbb{I} \setminus V$ is nowhere dense, and hence $A=(\mathbb{I} \setminus V) \cap Y$ has cardinality at most $\aleph_1$. Now $X \setminus U \subset (A \times \{0,1\}) \cup \bigcup_{n<\omega} F_n$ and since the latter set has cardinality at most $\aleph_1$, also $X \setminus U$ has cardinality at most $\aleph_1$, as we wanted.
\renewcommand{\qedsymbol}{$\triangle$}
\end{proof}
A winning strategy for player two in $G^{\omega_1}_1(\mathcal{O}, \mathcal{O})$ is now easy to describe. Let $\{x_n: n < \omega \}$ be an enumeration of $C$, and suppose that $\mathcal{U}_\alpha$ is the open cover played by player one at inning $\alpha$. At inning $n<\omega$ player two picks an open set $U_n \in \mathcal{U}_n$ such that $x_n \in U_n$. Let $\{z_\alpha: \alpha < \omega_1\}$ be an enumeration of $X \setminus \bigcup_{n<\omega} U_n$. Then at inning $\omega+\alpha$ player two simply picks an open set $U_{\omega+\alpha} \in \mathcal{U}_{\omega+\alpha}$ such that $z_\alpha \in U_{\omega+\alpha}$. We then have that $\{U_\alpha: \alpha < \omega_1 \}$ is an open cover, regardless of player one's choices, and hence the strategy we have defined is a winning one.
\end{proof}
\begin{question}
Can we get a space with the features of Example $\ref{exdensity}$ simply from the negation of CH?
\end{question}
Here is another application of Theorem $\ref{thmequiv}$, regarding the behavior of the game version of weak Lindel\"ofness in products.
\begin{theorem}
Suppose player two has a winning strategy in $G^\omega_1(\mathcal{O}, \mathcal{D})$ on $X$ and $Y$ is separable. Then player two has a winning strategy in $G^\omega_1(\mathcal{O}, \mathcal{D})$ on $X \times Y$.
\end{theorem}
\begin{proof}
Exploiting Theorem $\ref{thmequiv}$ fix a winning strategy $\sigma_X$ for player one on $X$ in $G^p_o(\omega)$ and fix a countable dense set $\{d_n: n < \omega \}$ for $Y$. Partition $\omega$ into infinitely many sets $\{I_k: k < \omega \}$ in such a way that if $\{i^k_j: j < \omega \}$ is an increasing enumeration of $I_k$ then $\{i^k_j: k < \omega \}$ is an increasing sequence, for every $j<\omega$. We now define a winning strategy for player one in $G^p_o(\omega)$ .
Assume player two played open set $U_i \times V_i$ at inning $i$, for $i \leq n$, and let $k, j< \omega$ be such that $n=i^k_j$. Let $\sigma_{X \times Y}((U_1 \times V_1, U_2 \times V_2, \dots U_n \times V_n))$ be $(x_n,y_n)$, where $x_n=\sigma_X((U_{i^k_m}: m \leq j ))$ and $y_n=d_k$.
\noindent Let now $$(\sigma_{X \times Y}(\emptyset), U_1 \times V_1, \pi_{X \times Y}(U_1\times V_1), \dots)$$ be a play. We claim that $\bigcup \{U_n \times V_n: n < \omega \}$ is dense. Indeed, let $U \times V$ be a non-empty basic open set in $X \times Y$. Then there is $k<\omega$ such that $d_k \in V$ and since $\pi_X$ is a winning strategy for $G^p_o(\omega)$ on $X$ then there is $n \in I_k$ such that $U \cap U_n \neq \emptyset$. Since $d_k \in V_n$ we have $(U \times V) \cap (U_n \times V_n) \neq \emptyset$ and hence we are done.
\end{proof}
Let's recall the definition of the point open game $G^o_p(\kappa)$, due to Berner and Juh\'asz \cite{BJ}.
\begin{definition}
Two players play $\kappa$ many innings. At inning $\alpha$, player one chooses an open set $O_\alpha$ and player two plays a point $x_\alpha \in O_\alpha$. Player one wins in $G^o_p(\kappa)$ if $\{x_\alpha: \alpha < \kappa \}$ is dense in $X$.
\end{definition}
Scheepers proved in \cite{SchSep} that the game $G^o_p(\kappa)$ is equivalent to the generalized $R$-separability game $G^\kappa_1(\mathcal{D}, \mathcal{D})$, in the following sense.
\begin{theorem} \label{thmequiv2} {\ \\}
\begin{enumerate}
\item \label{Splayerone} Player one has a winning strategy in $G^o_p(\kappa)$ if and only if player two has a winning strategy in $G^\kappa_1(\mathcal{D}, \mathcal{D})$.
\item \label{Splayertwo} Player two has a winning strategy in $G^o_p(\kappa)$ if and only if player one has a winning strategy in $G^\kappa_1(\mathcal{D}, \mathcal{D})$.
\end{enumerate}
\end{theorem}
He actually stated the result only for the case $\kappa=\omega$
We now exploit Scheeper's result to prove that every regular space where player two has a winning strategy in $G^{\omega_1}_1(\mathcal{D}, \mathcal{D})$ has $\pi$-weight at most continuum. As a byproduct we obtain an alternative proof of Scheeper's result countable $\pi$-weight is equivalent to the property that player two has a winning strategy in the $R$-separability game (see \cite{SchSep}).
\begin{theorem}
Let $X$ be a regular space and suppose that player two has a winning strategy in $G^{\kappa}_1(\mathcal{D}, \mathcal{D})$. Then $\pi w(X) \leq 2^{<\kappa}$.
\end{theorem}
\begin{proof}
Let $\tau$ be the set of all open sets of $X$ and fix a strategy $\sigma$ for player one in $G^o_p(\kappa)$. Let $M$ be a $<\kappa$-closed elementary submodel of $H(\theta)$, for some large enough regular $\theta$, such that $X, \tau \in M$, $|M| \leq 2^{<\kappa}$.
\noindent {\bf Claim.} $X \cap M$ is dense in $X$.
\begin{proof}[Proof of Claim]
Play a game of $G^o_p(\omega_1)$ where the first player plays according to $\sigma$ and the second player picks all its points in $M$. In other words, let $\alpha<\kappa$, and suppose that at inning $\beta < \alpha$, player one picked non-empty open set $U_\beta \in M$ and player two picked a point $x_\beta \in U_\beta \cap M$. At inning $\alpha$, player one plays non-empty open set $U_\alpha=\sigma((x_\beta: \beta < \alpha))$ which is an element of $M$ by $\kappa$-closedness and player two plays any point $x_\alpha \in U_\alpha \cap M$. Since $\sigma$ is a winning strategy for player one we must have that $\{x_\alpha: \alpha <\kappa\}$ is a dense set. But $\{x_\alpha: \alpha < \kappa \} \subset X \cap M$, and hence the claim.
\renewcommand{\qedsymbol}{$\triangle$}
\end{proof}
We now claim that $\tau \cap M$ is a $\pi$-base. Suppose this is not the case, and let $V$ be an open set such that $U \nsubseteq V$ for every $U \in \tau \cap M$. By regularity of $X$ we can actually assume that $U \nsubseteq \overline{V}$, for every $U \in \tau \cap M$. Now play a game of $G^o_p(\kappa)$ in a similar way as in the above claim, with the only difference that at inning $\alpha$ player two picks a point $x_\alpha \in U_\alpha \setminus \overline{V} \cap M$. Since player one is playing according to $\sigma$, we again have that $\{x_\alpha: \alpha < \kappa \}$ is dense, but this is impossible, since $V \cap \{x_\alpha: \alpha < \kappa \} = \emptyset$.
\end{proof}
\begin{corollary}
(Scheepers) Let $X$ be a regular space. Then $X$ has a countable $\pi$-base if and only if player two has a winning strategy in $G^\omega_1(\mathcal{D}, \mathcal{D})$.
\end{corollary}
\section{Acknowledgements}
The first listed author was partially supported by FAPESP (2013/05469-7) and by
Austrian Science Funds (FWF) through the grant M1244-N13. He wishes to thank KGRC fellows
for the great hospitality. The second named author was partially supported by FAPESP
postdoctoral grant 2013/14640-1,
\emph{Discrete sets and cardinal invariants in set-theoretic topology}.
Part of the work for the paper was carried out when he visited the third author
at the KGRC in Vienna and when he visited the first author at the ICMC in Sao Carlos.
He wishes to thank his colleagues in Vienna and Sao Carlos for their warm hospitality and
the University of Vienna and FAPESP for financial support.
The third-listed author thanks the Austrian Academy of Sciences and the
Austrian Science Funds (FWF) for generous support through the APART Program
and grants I 1209-N25, M 1244-N13, respectively.
|
\section{Introduction}
In \cite{vk1} Valiente Kroon identified a new class of obstructions to smoothness at infinity for asymptotically-flat solutions of the Einstein vacuum equations. He used the formalism of Friedrich \cite{f1} which begins
by identifying universal structure at null and spatial infinity and then writes out a system of conformal Einstein equations, equivalent to the vacuum equations, with adapted spin frames and coordinates. Friedrich \cite{f1}
found obstructions to smoothness at infinity in these coordinates in the time-symmetric case: necessary conditions for smoothness at infinity in the time-symmetric case are the vanishing at infinity of the symmetrised spinor derivatives of all orders of the spinor corresponding to the Bach tensor;
then finite degrees of smoothness
require finite numbers of these derivatives to vanish.
Valiente Kroon \cite{vk1} considered time-symmetric initial data with the extra condition of conformal-flatness, that is the second fundamental form is zero and the 3-metric of the data surface is conformally-flat, which entails the vanishing of the
Bach tensor. Thus Friedrich's conditions on the Bach tensor are vacuously satisfied but, by careful analysis of Friedrich's system, Valiente Kroon is able to show that there are still obstructions to smoothness. He finds a
heirarchy of obstructions
labelled by quantities $G^{(n)}$ for $n\geq 5$ and where each $G^{(n)}$ is a set of $2n-5$ constants obtained from a spherical harmonic expansion of a harmonic function obtained from the conformal factor to flat-space. He observes
that the first obstruction $G^{(5)}$ is essentially the Newman--Penrose conserved quantity for the gravitational field \cite{np1} (see also \cite{tb1}, \cite{dvk}). In this note, we give reasons for believing that in general
Valient Kroon's obstructions are proportional to a family of tensors which generalise the NP quantities as they relate to multipoles. That there is a connection between his obstructions and multipoles is implicit in the work of
Valiente Kroon, and explicit in the proof of rigidity \cite{vk2}, but the explicit relation conjectured here is thought to be new.
Time-symmetric, conformally-flat initial data then are determined by a 3-metric
\be\label{mm1}h_{ij}dx^idx^j=V^4\delta_{ij}dx^idx^j,\ee
where the constraint equations for the Einstein vacuum equations reduce to the Laplace equation on $V$. Asymptotic flatness can be achieved by $V\rightarrow 1$ at large distances.
Commonly considered solutions (e.g. \cite{bl}) are sums of point masses
\be\label{h1}
V=1+\frac{1}{2}\sum_\alpha\frac{m_\alpha}{|{\bf{r}}-{\bf{a}}_\alpha|}.
\ee
Here the factor $1/2$ before the sum is to ensure that the total (ADM) mass at infinity is $M=\sum m_\alpha$.
As is familiar, the singularities of $V$ at ${\bf{r}}={\bf{a}}_\alpha$ are not singularities of $h_{ij}$ but rather are other `ends' in the sense of asymptotically-flat regions, \cite{bl}. Also one may construct solutions like
the Misner wormhole \cite{mis} by superposing a suitable infinite set of collinear point masses.
The outline of this letter is as follows: in the next section we review the definition of multipole moments in ${\mathbb{R}}^3$ and introduce a set of trace-free, symmetric tensors $\hE^{(n)}_{i_1\ldots i_{n+1}}$ derived from
the multipole moments. In Section 3 we recall Valiente Kroon's obstructions to smoothness at infinity \cite{vk1} and conjecture a relation between them and the tensors $\hE^{(n)}$. In Section 4 we say a few words about the
issue of rigidity in the sense that vanishing of the obstructions constrains the data to be data for Schwarzschild.
\section{Multipole moments in flat space}
Given a harmonic function $V$ in ${\mathbb{R}}^3$ tending to one at large distances, as in the previous section, one has an expansion in spherical harmonics
\[V = 1+\sum_{\ell=1}^\infty\sum_{m=-\ell}^{\ell}r^{-(\ell+1)}c_{\ell m}Y_{\ell m}(\theta,\phi)\]
in the standard spherical polar coordinates. This can alternatively be written as
\[V=1+\sum_{\ell=1}^\infty r^{-(\ell+1)}m_{i_1\ldots i_\ell}e_{i_1}\ldots e_{i_\ell},\]
where the summation convention applies to tensor indices, we introduce the vector ${\bf{e}}=(e_i)=(\sin\theta\cos\phi,\sin\theta\sin\phi,\cos\theta)$ and for each $\ell$ the tensor $m_{i_1\ldots i_\ell}$ is trace-free and symmetric.
For each $\ell$, there are $2\ell+1$ coefficients $c_{\ell m}$ and the tensor $m_{i_1\ldots i_\ell}$ has $2\ell+1$ independent components, so that each can be written uniquely in terms of the other. Following varying conventions these components either are
or are proportional to the $(\ell+1)$-th or $2^\ell$-th multipole moment.
If $V$ solves a Poisson equation
\[\nabla^2V=\kappa\rho\]
with density $\rho$ then the multipole moments can be related to integrals over the source in a familiar manner. The case of interest for us is when the source consists of a set of
point masses of mass $m^\alpha$ at positions ${\bf{a}}^\alpha=(a^{\alpha}_i)$ in ${\mathbb{R}}^3$ with $\alpha$ ranging over some indexing set. It is convenient to define
\be\label{m1}
O^{(n)}_{i_1\ldots i_n}=\sum_{\alpha}m^\alpha a^\alpha_{i_1}\ldots a^\alpha_{i_n},
\ee
with $O^{(0)}=M$, the total mass. Expansion of $V$ in (\ref{h1}) now shows that the tensor expression of the $n$-th multipole moment is proportional to the tensor
\be\label{m2}
\hO^{(n)}_{i_1\ldots i_n}=O^{(n)}_{i_1\ldots i_n}-\mbox{ trace
\ee
where the subtracted trace is an expression that we don't need explicitly, made from a combination of Kronecker deltas and all possible traces of $O^{(n)}$. It is computationally simpler to work with
$O^{(n)}$ but keep in mind that it is the symmetric and trace-free $\hO^{(n)}$ which corresponds to the multipole.
We obtain a compression of notation by use of polynomials: introduce $(X^i)=(X,Y,Z)$ and then symmetric tensors $t_{i_1\ldots i_k}$ are in one-to-one correspondence with polynomials of degree $k$ via
\[t(X,Y,Z):=t_{i_1\ldots i_k}X^{i_1}\ldots X^{i_k}.\]
In particular this gives $n$-th order polynomials $O^{(n)}$ and $\hO^{(n)}$ representing the tensors in (\ref{m1}) and (\ref{m2}).
Note that the tensor $\delta_{ij}$ corresponds to the polynomial $r^2:=X^2+Y^2+Z^2$. Call a tensor $t_{i_1\ldots i_n}$ \emph{pure trace} if it is of the form $\delta_{(i_1i_2}s_{i_3\ldots i_n)}$ for some $s_{i_3\ldots i_n)}$,
since it will then have a zero trace-free part. As polynomials this condition is
\be\label{tf}t^{(n)}=r^2s^{(n-2)},\ee
which we use below.
Under shift of origin by $b_i$ in ${\mathbb{R}}^3$ the point masses move according to:
\[a^{\alpha}_{i}\rightarrow a^\alpha_i+b_i,\]
which can be written in terms of polynomials as:
\[a^\alpha\rightarrow a^\alpha+b.\]
Now introduce $\delta$ for the change under translation so that
\[\delta a^\alpha=b\]
and deduce from (\ref{m1}) that
\be\label{s1}
\delta O^{(n)}=\sum_{k=1}^n\,^{n}C_kb^kO^{(n-k)}.
\ee
For $n\geq 1$, introduce an infinite set of tensors
\be\label{e1}
E^{(n)}_{i_1\ldots i_{n+1}}:=M^{n}O^{(n+1)}_{i_1\ldots i_{n+1}}-O^{(1)}_{(i_1}\ldots O^{(1)}_{i_{n+1})}
\ee
so that, as polynomials
\[E^{(n)}=M^{n}O^{(n+1)}-(O^{(1)})^{n+1}\]
and then write $\hE^{(n)}$
for the trace-free part of $E^{(n)}$.
It is straightforward to see that, under translation of origin, $E^{(n)}$ transforms according to
\be\label{e2}
\delta E^{(n)}=\sum_{k=1}^{n-1}\,^{n+1}C_kM^kb^kE^{(n-k)}
\ee
for $n>1$ while $E^{(1)}$ is independent of origin, and the higher $E^{(n)}$ transform in terms of the lower ones.
Since taking the trace commutes with $\delta$, $\hE^{(1)}$ is also origin-independent and it can in fact be identified as the tensor
defining the NP conserved quantities (see e.g. \cite{dvk}, \cite{vk1}).
From (\ref{e2}) we also deduce that
\begin{itemize}
\item if $E^{(k)}=0$ for $1\leq k<n$ then $E^{(n)}$ is origin-independent.
\end{itemize}
This result also holds for $\hE^{(n)}$ by the following argument: suppose $\hE^{(k)}=0$ for $1\leq k<n$ then for this range $E^{(k)}$ is pure trace i.e. by (\ref{tf})
\[E^{(k)}=r^2S^{(k-2)}\]
for some $S^{(k-2)}$. From (\ref{e2}) it follows that $\delta E^{(n)}$ is also pure trace and therefore, since taking the trace commutes with $\delta$, that $\delta\hE^{(n)}=0$.
We may choose the origin to set $O^{(1)}=0$, and this choice defines the centre-of-mass frame. In the centre-of-mass frame by (\ref{e1}) each $E^{(n)}$ is proportional to the corresponding $O^{(n)}$, and therefore each
$\hE^{(n)}$ is proportional to the corresponding $\hO^{(n)}$. Thus
\begin{itemize}
\item in the centre-of-mass frame if $\hE^{(k)}=0$ for $1\leq k\leq n$ then $\hO^{(k)}=0$ for $1\leq k\leq n$, and vice versa.
\end{itemize}
whence also
\begin{itemize}
\item if $\hE^{(k)}=0$ for all $k$ then in the centre-of-mass frame $\hO^{(k)}=0$ for all $k>0$ and $V$ is spherically symmetric.
\end{itemize}
It is worth noting that there is another set of tensors with properties similar to the $E^{(n)}$ and $\hE^{(n)}$, namely $F^{(n)}$ and $\hF^{(n)}$ where, defined as polynomials:
\be\label{fn}
F^{(n)}:=MO^{(n+1)}-O^{(1)}O^{(n)},\ee
for $n\geq 1$, and $\hF^{(n)}$ is the trace-free part of $F^{(n)}$. In place of (\ref{e2}) one finds
\be\label{f2}
\delta F^{(n)}=\sum_{k=1}^{n-1}\,^{n}C_kb^kF^{(n-k)},
\ee
and the discussion proceeds as before. Evidently one could express the $E^{(n)}$ in terms of the $F^{(n)}$ and vice versa, but it is the $E^{(n)}$ which arise most immediately in Valiente Kroon's obstructions.
It remains to be seen whether the $\hE^{(n)}$ or $\hF^{(n)}$ have significant properties at null infinity, $\scri^+$.
\section{Valiente Kroon's obstructions}
In \cite{vk1} Valiente Kroon studied Friedrich's system \cite{f1} for time-symmetric, conformally-flat initial data. The data are the metric in the form (\ref{mm1}). He was able to integrate the equations at infinity for the first
four orders, giving smooth solutions, but logarithmic terms arise at the intersection of the cylinder at space-like infinity with null infinity at all orders from the fifth order and will obstruct smoothness at infinity. The coefficients of the logarithms at order $n$ for $n\geq 5$ are
quantities $G^{(n)}$, where each $G^{(n)}$ is a set of $2n-5$ constants expressed in terms of the coefficients of the expansion of $V$ in spherical harmonics.
Valiente Kroon expanded $V$ in terms of functions $T^{\;\;k}_{n\;\;\;j}$ which are a complete orthonormal set for $L^2(SU(2,{\mathbb{C}}))$ with standard Haar measure, and are standardly used by Friedrich and his collaborators (see e.g.
\cite{f1}, \cite{fk}). With a suitable choice of conventions, they can be related to spin-weighted spherical harmonics up to numerical constants by
\[T^{\;\;k}_{n\;\;\;j}\sim e^{is\psi}\,_sY_{\ell m}(\theta,\phi),\]
with $s=\frac12n-j,\ell=\frac12n,m=\frac12n-k$ and Euler angles $(\theta,\phi,\psi)$. In particular
\[T^{\;\;k}_{2\ell\;\;\;\ell}\sim Y_{\ell m}\]
with $m=\ell-k$. If one is more familiar with spherical harmonics, this observation simplifies the interpretation of Valiente Kroon's obstructions. His $G^{(5)}$ contains coefficients $w_{2,4,k}$ and $w_{1,2,k}$: $w_{2,4,k}$ is the
coefficient of $T^{\;\;k}_{4\;\;\;2}$ in $V$, and therefore is related to the coefficient of $Y_{2,2-k}$ and therefore is a component of the tensor $\hO^{(2)}$; $w_{1,2,k}$ is similarly related to $\hO^{(1)}$;
so $G^{(5)}$ is a sum of two terms proportional respectively to $M\hO^{(2)}$ and the trace-free part of $(O^{(1)})^2$. It is a trace-free tensor under rotation and, as the leading obstruction, must be origin-independent. Therefore
it must be $\hE^{(1)}$, a fact shown explicitly in \cite{vk1}, but this model of argument can be applied at higher orders.
Moving on to $G^{(6)}$, each component is a sum of two kinds of term, one obtained from $M^2\hO^{(3)}$ and the other from the trace-free part of $(O^{(1)})^3$. Again it's a trace-free tensor under rotation and
if $G^{(5)}=0$ it becomes the leading obstruction and therefore origin-independent. Thus it must be proportional to $\hE^{(2)}$. Inductively we are lead to conjecture that, for all $k\geq 5$, $G^{(k)}$ and $\hE^{(k-4)}$ are proportional.
\section{Rigidity}
Valiente Kroon \cite{vk2} shows that smoothness at infinity implies that the data is that for the Schwarzschild solution. This would follow from the conjecture in the preceding
section since the vanishing of all $G^{(k)}$ entails the vanishing of all $\hE^{(k)}$ and hence, in the centre-of-mass frame, of all multipole moments $\hO^{(n)}$ for $n\geq 1$ which implies Schwarzschild.
One can obtain various intermediate results:
\begin{enumerate}
\item[$\bullet$] in the case that all ${\bf{a}}_\alpha$ are coplanar then the vanishing of $E^{(1)}$ implies Schwarzschild (i.e. all the mass points are coincident); in particular this also holds for coplanar Misner wormholes;
\item[$\bullet$] with 4 mass-points there is a one-parameter family of tetrahedral configurations with $E^{(1)}=0$ which are equilateral if the masses are all equal; however
the vanishing of $E^{(2)}$ forces Schwarzschild;
\item[$\bullet$] so one is lead to conjecture: for each $n$ there will be an $N_n$ such that the vanishing of $\hE^{(k)}$ for $1\leq k\leq n$ forces any configuration of $j\leq N_n$ mass points to be Schwarzschild. A naive count
of equations suggests $3N_n$ is the largest multiple of 3 less than $n^2+4n+7$.
\end{enumerate}
\section*{Acknowledgement}
We are grateful to Prof Valiente Kroon for discussions. JG is grateful to St. John's College, Oxford for a Visiting Scholarship during which this work was carried out. He is also grateful to the Mathematical Institute in Oxford for their hospitality.
|
\section*{\hfil #1\hfil}}
\bibliographystyle{plain}
\swapnumbers
\newtheorem{theorem}{Theorem}[section]
\newtheorem{example}{Example}[section]
\newtheorem{conj}{Conjecture}[section]
\newtheorem{hypothesis}{Hypothesis}[section]
\newtheorem{problem}[theorem]{Problem}
\newtheorem{df}[theorem]{Definition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{prop}[theorem]{Proposition}
\newtheorem{corr}[theorem]{Corollary}
\newtheorem{remark}[theorem]{Remark}
\def\mathrm{Ch}{\mathrm{Ch}}
\def\mathbf{Z}{\mathbf{Z}}
\def\mathbf{Q}{\mathbf{Q}}
\def\mathbf{R}{\mathbf{R}}
\def\epsilon{\epsilon}
\def\varepsilon{\varepsilon}
\def\mathcal{O}{\mathcal{O}}
\def\mathrm{Tr}{\mathrm{Tr}}
\def\mathscr{M}{\mathscr{M}}
\def\mathscr{N}{\mathscr{N}}
\def\underline{k}{\underline{k}}
\def\underline{\eps}{\underline{\epsilon}}
\begin{document}
\title{Abelian Spiders}
\author{Frank Calegari and Zoey Guo}
\thanks{The authors were supported in part by NSF Grant
DMS-1404620.}
\maketitle
\section{Introduction}
Let~$\Gamma$ be a connected finite graph. Fix an integer~$k$, and let $v_1, \ldots, v_k$ be~$k$
(not necessarily distinct) vertices of~$\Gamma$. For any~$k$-tuple $\underline{k} = (r_1,\ldots,r_k)$ of non-negative integers, we define
a ($k$-)\emph{spider graph} $\Gamma_{\underline{k}}$ on~$\Gamma$ to be the graph obtained by adjoining a~$2$-valent tree
of length~$r_i$ to~$\Gamma$ at~$v_i$. We say a graph~$\Gamma$ is~\emph{abelian} if~$\mathbf{Q}(\lambda^2)$
is an abelian extension, where~$\lambda$ is the Perron--Frobenius eigenvalue of~$\Gamma$ (the
unique largest real eigenvalue of the adjacency matrix~$M_{\Gamma}$ of~$\Gamma$).
If~$\Gamma$ is one of the affine Dynkin diagrams, then~$\Gamma$ is abelian, and~$\lambda^2 = 4 \cos^2(2 \pi/N)$ for some integer~$N$. Conversely, if~$\lambda \le 2$, then~$\Gamma$ is an affine Dynkin diagram.
\begin{theorem} \label{theorem:one} Fix~$\Gamma$ and~$k$. There are only
finitely many abelian~$k$-spiders~$\Gamma_{\underline{k}}$ which are not Dynkin
diagrams.
There is an effective algorithm for determining all such spiders.
\end{theorem}
\begin{remark} \emph{If~$\Gamma$ is not already of the form~$A_n$ or~$D_n$,
then only finitely many of the spiders~$\Gamma_{\underline{k}}$ will be Dynkin diagrams.}
\end{remark}
One motivation for this paper is the application to subfactors, as in~\cite{CMS}.
One of the main results (Theorem~1.0.3) of~\cite{CMS} was a version of Theorem~\ref{theorem:one}
for~$1$-spiders. The paper~\cite{CMS} also contained a weaker result (Theorem~1.0.6) which
was sufficient for the application
to subfactors but had the advantage that the effective constants could be made explicit.
In contrast, Theorem~\ref{theorem:one} already comes
with computable effective constants, and moreover these constants will be small enough
that our results are ``effectively effective'' in many cases
(although there is certainly
some combinatorial explosion as~$k$ increases). In order for this to be so,
we have worked hard in this paper to make our results as tight as possible, even when weaker estimates
would certainly suffice to prove the main theorem. As
an application of Theorem~\ref{theorem:one} to the theory
of subfactors, we
prove the following result, conjectured by S.~Morrison~\cite{MorrisonNotes}. Let $\Gamma_{a,b}$ denote the ``Morrison spider,'' given as follows:
\begin{figure}[!ht] \label{fig:one}
\begin{center}
\includegraphics[width=60mm]{graph.png}
\end{center}
\end{figure}
\begin{theorem} \label{theorem:morrison} The spider~$\Gamma_{a,b}$ above is abelian only when $(a,b) = (0,0)$ or $(1,1)$.
\end{theorem}
\section{Acknowledgments}
We thank Scott Morrison for bringing to our attention the problem of understanding
the Perron--Frobenius eigenvalues arising in Theorem~\ref{theorem:morrison}.
We also thank Chris Smyth for bringing to our attention the papers~\cite{Zone,Ztwo}, where
calculations similar to Prop.~\ref{prop:list} are carried out.
\section{Estimates}
The first technical ingredient is the following inequality below, which is similar (but quite
a bit more complicated) to Lemma~4.2.3 of~\cite{CMS}. In fact, it turns out that inequalities
of a similar shape were first considered by Smyth in~1981~\cite{Smythtwo,Smyth}, where the intended
application was to generalizations of Siegel's theorem on lower bounds for the trace
of totally positive integers. The creation of such inequalities seems to be part science and part art.
Let $\mathrm{Ch}_{N}(x)$ denote the minimal polynomial of $(\zeta_N + \zeta^{-1}_N)^{2}$. The table below contains
explicit expressions for the~$\mathrm{Ch}_{N}(x)$ together with the value of~$\mathscr{M}(\zeta_N + \zeta^{-1}_N)$, where~$\mathscr{M}(\beta):= \frac{\mathrm{Tr}_{K/\mathbf{Q}}(\beta^2)}{[K:\mathbf{Q}]} $ is the normalized trace of~$\beta^2$. The coefficient~$a_N$ is used below in the definition of~$B(x)$.
The optimization of the coefficients~$a_N$ in the definition of~$B(x)$ was performed by simulated annealing.
\begin{center}
\begin{tabular}{clcc}
\hline
$N$ & $\mathrm{Ch}_{N}(x)$ & $\mathscr{M}(\zeta_N + \zeta^{-1}_N)$ & $a_N$ \\
$1$ & $x - 4$ & $4$ & $673$ \\
$3$ & $x - 1$ & $1$ & $6$ \\
$4$ & $x$ & $0$ & $4$ \\
$5$ & $x^2 - 3x + 1$ & $3/2$ & $2$ \\
$7$ & $x^3 - 5 x^2 + 6 x - 1$ & $5/3$ & $5$ \\
$8$ & $x - 2$ & $2$ & $157$ \\
$9$ & $x^3 - 6 x^2 + 9 x - 1$ & $2$ & $13$ \\
$12$ & $x - 3$ & $3$ & $578$ \\
$15$ & $x^4 - 9x^3 + 26x^2 - 24x + 1$ & $9/4$ & $43$ \\
$16$ & $x^2 - 4 x + 2$ & $2$& $49$ \\
$20$ & $x^2 - 5 x + 5$ & $5/2$ & $215$ \\
$21$ & $x^6 - 13x^5 + 64 x^4 - 146 x^3 + 148 x^2 - 48 x + 1$
& $13/6$ & $10$ \\
$24$ & $x^2 - 4 x + 1$ & $2$ & $25$ \\
$28$ & $x^3 - 7 x^2 + 14 x - 7$ & $7/3$ & $80$ \\
$44$ & $x^5 - 11 x^4 + 44 x^3 - 77 x^2 + 55 x - 11$ & $11/5$ & $24$ \\
$52$ & $x^6 - 13 x^5 + 65 x^4 - 156 x^3 + 182 x^2 - 91 x + 13$ & $13/6$ & $1$ \\
\end{tabular}
\end{center}
If~$N$ is not on this list, set~$a_N = 0$.
This list of polynomials includes every~$N$ where the inequality~$\mathscr{M}(\zeta_N + \zeta^{-1}_N) > 13/6$ is
satisfied, as well as a complete list
of all such polynomials for~$N < 11$.
\begin{df}
Define the function~$B(x)$ as follows:
$$B(x) = \frac{9}{4} - x - \frac{1}{1000} \sum a_N \log | \mathrm{Ch}_N(x)|.$$
\end{df}
The key property of $B(x)$ is the following estimate:
\begin{lemma} \label{lemma:sneak} For $x \in [0,4]$ where $B(x)$ is defined, $B(x) \ge 0$.
\end{lemma}
\begin{figure}[H] \label{fig:B}
\begin{center}
\includegraphics[width=160mm]{norm.png}
\end{center}
\caption{The graph of $ B(x)$ in $[0,4]$.}
\end{figure}
The derivative of~$B(x)$ lies in $\mathbf{Q}(x)$.
The minimum value of~$B(x)$ in~$[0,4]$ occurs at an algebraic number~$\alpha
\sim 0.00209304$ of degree~$40$, where~$B(x)$ obtains the value~$\sim 0.00599001$.
For~$x > 4$ (away from singularities), $B(x)$ is decreasing. One
has the estimate $\lim_{x \rightarrow \infty} B(x)/x = -1$.
\begin{theorem} \label{theorem:bounds} Let~$L$ be a non-negative
real number, and let $\beta$ be a totally real algebraic integer with $K = \mathbf{Q}(\beta^2)$ such that:
\begin{enumerate}
\item $\beta^2$ is not a singularity of~$B(x)$,
\item The largest conjugate $\ho{\beta}$ of~$\beta$ satisfies $\ho{\beta} < L$.
\item At most~$M$ conjugates of~$\beta^2$ lie outside the interval~$[0,4]$.
\end{enumerate}
Then,
$\displaystyle{\mathscr{M}(\beta) = \frac{\mathrm{Tr}_{K/\mathbf{Q}}(\beta^2)}{[K:\mathbf{Q}]} < \frac{14}{5}}$
if either $B(L^2) > 0$ or
$\displaystyle{[K:\mathbf{Q}] \ge \frac{20}{11} \cdot M \cdot |B(L^2)|}$.
\end{theorem}
\begin{proof} At most~$M$ conjugates of~$\beta^2$ lie outside the interval~$[0,4]$.
Consider the sum
$\sum B(\sigma \beta^2)$.
If~$\sigma \beta^2$ is not a singularity of~$B$, then the sum of each logarithmic term
is a negative rational number times the logarithm of the norm of an algebraic integer, and
is hence negative. If $D = [\mathbf{Q}(\beta^2):\mathbf{Q}]$, it follows that
$$\sum B(\sigma \beta^2) \le \frac{9}{4} \cdot D - D \mathscr{M}(\beta).$$
On the other hand, we have the estimate $B(x) \ge 0$ for $x \in [0,4]$, and
that~$B(x)$ is decreasing otherwise. Hence
$$\sum B(\sigma \beta^2) \ge M \cdot B(L^2).$$
(Note that $B(L^2) \le 0$ for $L > 2.0152$ or so).
Combining these estimates, we deduce that
$$\frac{9}{4} - \mathscr{M}(\beta) - \frac{M \cdot B(L^2)}{D} =
\frac{14}{5} - \mathscr{M}(\beta) - \frac{11}{20} - \frac{M \cdot B(L^2)}{D} \ge 0,$$
which is a contradiction as soon as either of the inequalities of
the statement are satisfied.
\end{proof}
\subsection{The spectrum of~\texorpdfstring{$\Gamma_{\underline{k}}$}{Gammak}}
\label{section:spectrum}
We begin by recording some basic properties of eigenvalues of graphs. A reference
for this section is~\cite{Salem}. The following Lemma is essentially Lemma~12 of~\cite{Salem}:
\begin{lemma} If~$r_i \ge 2$ for all~$i$, then the characteristic polynomial~$P_{\underline{k}}(x)$ of~$\Gamma_{\underline{k}}$
has the form:
$$\left(t - \frac{1}{t}\right)^k P_{\underline{k}}\left(x \right)= \sum_{\underline{\eps}}
t^{\sum \epsilon_i r_i} P_{\underline{\eps}}\left(x \right),$$
where~$x = t + t^{-1}$, the index $\underline{\eps}$ runs over~$k$-tuples $(\epsilon_1,\ldots,\epsilon_k)$ with $\epsilon_i \in \{1,-1\}$,
and
where the polynomials $P_{\underline{\eps}} \in \mathbf{Z}[x]$ do not depend on~$\underline{k}$.\end{lemma}
\medskip
Let~$Q(x) = F_{\epsilon}(x)$ where $\epsilon = (1,1,\ldots,1)$.
Let~$S$ denote the set of real roots of~$Q(x)$ in $(2,\infty)$, counted with multiplicity.
Say that a vector~$\underline{k}$ is large if \emph{all} the entries~$r_i$ are large.
\begin{lemma} \label{lemma:general} We have the following:
\begin{enumerate}
\item If $\underline{k}' > \underline{k}$ in the partial ordering, then the Perron Frobenius eigenvalue~$\lambda'$ of
$\Gamma_{\underline{k}'}$ is strictly larger than~$\lambda$.
\item $P_{\underline{k}}(x)$ has~$|S|$ real roots $> 2$ for sufficiently large~$\underline{k}$, and they are converging from below to~$S$.
\end{enumerate}
\end{lemma}
\begin{proof} The first claim follows from the interlacing Theorem. The second claim is
proved in~\cite{Salem}. The main point is that any root $> 1 + \varepsilon$ of $P_{\underline{k}}(t + t^{-1})$
will continue (by interlacing)
to be $> 1 + \varepsilon$ as $\underline{k}$ grows. Then, for sufficiently large~$\underline{k}$, Rouch\'{e}'s theorem will show that
the number of real roots~$>1 + \varepsilon$ of $P_{\underline{k}}(t + t^{-1})$ will
be equal to the number of real roots of~$Q(t + t^{-1})$.
\end{proof}
We immediately deduce:
\begin{lemma} \label{lemma:uniform} There exists constants~$M = M_{\Gamma}$ and~$L = L_{\Gamma}$ such that:
\begin{enumerate}
\item If~$\lambda$ is the Perron--Frobenius eigenvalue of~$\Gamma_{\underline{k}}$, then
$\lambda^2 - 2 \le L$.
\item At most~$M$ of the conjugates of $\lambda^2 -2$ lie outside the interval $[-2,2]$.
\end{enumerate}
\end{lemma}
In practice, these constants are often small and computable (indeed, often~$M$ is equal to one, as it
will be in our examples).
We have, moreover, the following easy upper bound for~$\lambda$:
\begin{lemma} Suppose that the largest valence of any vertex of~$\Gamma_{\underline{k}}$
is $v$. Then~$\lambda \le v$.
\end{lemma}
Combined with
Theorem~\ref{theorem:bounds} above, we deduce:
\begin{corr} For all sufficiently large~$\underline{k}$, the largest eigenvalue~$\lambda$ of
$\Gamma_{\underline{k}}$ satisfies~$\mathscr{M}(\lambda^2 - 2) < 14/5$.
\end{corr}
\begin{proof} Since~$\lambda^2 - 2$ is strictly increasing as~$\underline{k}$ increases, it suffices to show that
the degree of~$\lambda^2$ is not bounded. Yet all the conjugates of~$\lambda^2 - 2$ are bounded by~$L$,
and there are only a finite number of algebraic integers of fixed degree with this property, by
a well known lemma of Kronecker.
\end{proof}
We shall prove in Prop~\ref{prop:list} that if $\mathbf{Q}(\lambda^2)$ is abelian,
then~$\mathscr{M}(\lambda^2 -2 ) < 14/5$ implies either that $\lambda \le 2$ or~$\lambda^2 - 2$ is
one of a finite set of algebraic integers. This is is enough to prove that there are only finitely many abelian
spiders which are not Dynkin diagrams for sufficiently large~$\underline{k}$. On the other hand, if one the~$r_i$ is
bounded by a constant~$B$, then we can proceed by induction and consider the~$k-1$ spiders on the finitely many
graphs where a $2$-valent tree of length $r_i \le B$ is attached to~$\Gamma$ at~$v_i$.
This leads to a proof of Theorem~\ref{theorem:one}.
The problem is that
Kronecker's theorem, although ``explicit,'' is not really so explicit in practice (since it involves
checking a super-exponential set of polynomials).
Instead, we shall give a different argument which can be used in practice.
\begin{prop} \label{prop:degree} Suppose that each element of~$\underline{k}$ is at least~$n \ge 2$. There is a bound:
$$D:=[\mathbf{Q}(\lambda^2):\mathbf{Q}] \gg n,$$
where the implied constant depends only on~$\Gamma$ and is explicitly computable.
\end{prop}
\begin{proof} We may assume that~$n$ is large (in practice, what counts as ``large'' is usually
not prohibitive). Write~$\lambda = \rho + \rho^{-1}$. Certainly $[\mathbf{Q}(\lambda^2):\mathbf{Q}(\rho)] \le 4$,
so it suffices to give a linear lower bound on the degree of~$\rho$.
Let~$\rho_{\infty}$ denote the largest root of~$Q(t + t^{-1})$. We know that the values~$\rho$
are converging to~$\rho_{\infty}$;
the basic idea is to show that this convergence is exponentially fast, which,
together with the fact that the conjugates of~$\rho$ are constrained in absolute value, is enough to give the requisite
bound on the degree of~$\rho$.
Write~$\rho_{\infty} - \rho = \varepsilon$.
Since~$P_{\underline{k}}(\rho + \rho^{-1}) = 0$, we deduce that:
$$0 = P_{\underline{k}}\left(\rho + \rho^{-1} \right)= \sum_{\underline{\eps}}
\rho^{\sum \epsilon_i r_i} F_{\underline{\eps}}(\rho + \rho^{-1}).$$
Taking absolute values and applying the triangle inequality, we deduce that
$$
| Q(\rho + \rho^{-1}) | \ll \rho^{-2n},$$
where the constants can easily be made effective in any particular case (they involve
the supremum of the polynomials~$F_{\underline{\eps}}(t + t^{-1})$ for~$t$ in a neighbourhood of~$\rho_{\infty}$).
On the other hand, suppose that the root~$\rho_{\infty}$ of
$Q(x)$ has a multiplicity exactly~$m$. Then
there is an inequality
$|Q(\rho + \rho^{-1})| \gg \varepsilon^{m}$ for some explicitly computable
constant~$A > 0$ depending on the~$m$th derivative of~$Q$ at~$\rho_{\infty} + \rho^{-1}_{\infty}$.
Since~$\rho$ is converging to~$\rho_{\infty} > 1$, it satisfies~$\rho^{2/m} > \theta$ for some explicit~$\theta > 1$
which does not depend on~$n$.
It follows that, where as above the implicit constants can easily be evaluated
explicitly, we have the following inequality:
$$ |\rho_{\infty} - \rho | \ll \frac{1}{\theta^{n}}.$$
Let $R(t)$ be the minimal polynomial of~$\rho_{\infty}$.
The polynomial~$R(t)$ does not vanish on any conjugate
of~$\rho$ because~$\rho_{\infty} > |\sigma \rho|$
for all conjugates of~$\rho$ and~$R(t)$ is irreducible.
The polynomial~$R(t)$ is bounded on the ball $|t| \le \rho_{\infty}$ by some absolute constant~$C$.
Let~$D = [K:\mathbf{Q}]$ with~$K = \mathbf{Q}(\lambda^2)$, and let~$L = \mathbf{Q}(\rho)$.
Since~$[L:K] \le 4$, the degree of~$L/\mathbf{Q}$ is at most~$4D$.
Since $R(\rho) \ne 0$, we have
$$1 \le N_{L/\mathbf{Q}}(R(\rho)) \le
C^{4D-1} \cdot |\rho - \rho_{\infty}|
\le \frac{C^{4D-1}}{\theta^{n}}.$$
Taking logarithms and using the fact that~$\theta > 1$ leads to a linear lower bound in~$D$, as desired.
\end{proof}
Combining this result with
Theorem~\ref{theorem:bounds} above, we deduce:
\begin{corr} \label{corr:deduce} There exists an effectively computable constant~$n$ such
that for all $r_i \ge n$, either the largest eigenvalue~$\lambda$ of
$\Gamma_{\underline{k}}$ satisfies~$\mathscr{M}(\lambda^2 - 2) < 14/5$, or~$\Gamma_{\underline{k}}$ is an affine Dynkin
diagram.
\end{corr}
\begin{proof} The previous Lemma shows that we may find an explicit~$n$ so that the degree of~$\lambda^2$ is large.
The result then follows from~Theorem~\ref{theorem:bounds} and Lemma~\ref{lemma:uniform} once we have
an effective bound on~$n$ so that~$\lambda^2 - 2$ is not conjugate to a singularity of~$B(x)$. Note, however, that
all the singularities of~$B(x)$ are algebraic integers all of whose conjugates lie in $[0,4]$. If~$\lambda^2-2$ is such an integer,
then~$|\lambda| \le 2$ and~$\Gamma_{\underline{k}}$ is an (affine) Dynkin diagram.
\end{proof}
\section{Totally real cyclotomic integers with small \texorpdfstring{$\mathscr{M}$}{M}}
In this section, we shall improve on some estimates from~\cite{CMS}.
We make, however, the following preliminary remark.
Modifying the proof of Theorem~\ref{theorem:bounds} slightly, we see that
there exists a lower bound on~$D$ (depending on~$L$ and any~$\epsilon > 0$)
that guarantees the inequality $\mathscr{M}(\beta) \le 9/4 + \epsilon$.
However, this can be improved further.
The proof
of Lemma~\ref{lemma:sneak} has some slack can also be exploited, namely by replacing~$B(x)$
by~$B(x)$ by~$B(x) - \delta$ for small but non-zero~$\delta > 0$.
(As mentioned directly after the statement of Lemma~\ref{lemma:sneak},
one could take~$\delta$ to be anything less than approximately
$0.00599001$.)
This would allow us to modify the proof of~Theorem~\ref{theorem:bounds}
to give an explicit lower bound on~$D$ (in terms only of~$B(L)$) which would
guarantee that~$\mathscr{M}(\beta) < 9/4$.
We could then dispense
with~Prop.~\ref{prop:list} below entirely and use Lemma~9.0.1 of~\cite{CMS},
which classifies those~$\beta$ with~$\mathscr{M}(\beta) < 9/4$. However,
such an argument would lead to (significantly) worse bounds.
\medskip
We shall freely use many of the concepts from~\cite{Cassels} and~\cite{CMS}.
Recall that two algebraic cyclotomic integers are called equivalent
if their ratio is a root of unity, and that a cyclotomic integer~$\beta$ is minimal
if it has smallest conductor amongst all its equivalent forms. If~$\beta$ is totally real,
it is not always the case that a minimal equivalent cyclotomic integer is also totally real,
but this is almost true:
\begin{lemma} \label{lemma:real} If~$\beta$ is a minimal cyclotomic integer of odd conductor~$N$
which is equivalent to a cyclotomic integer, then, up to a root of unity in~$\mathbf{Q}(\zeta_{N})$,
either~$\beta$ or~$\beta \cdot \sqrt{-1}$ is totally real.
\end{lemma}
\begin{proof}
Suppose that~$\beta$ is minimal of conductor~$N$. Write $\gamma = \zeta \beta$, where~$\gamma$
is totally real.
If~$\zeta \in \mathbf{Q}(\zeta_{2N}) = \mathbf{Q}(\zeta_{N})$, then the result is trivial.
Hence there exists a prime~$p$ such that $2N$ is exactly divisible by $p^m$ and the order of~$\zeta$ is
exactly divisible by $p^{n}$ for some $n > m$.
Let~$\xi$ denote a primitive~$p^{n}$ root of unity.
There exists a Galois automorphism~$\sigma$ fixing~$\mathbf{Q}(\zeta_{N})$ and hence fixing~$\beta$ such that
$$\gamma/\sigma \gamma = \sigma \zeta/\zeta = \xi^{p^{m}}.$$
Since~$\gamma$ is totally real, the latter element must also be real, which forces
$p = 2$ and $\xi^{4} = 1$ (noting that $p^m = 2$ if~$p = 2$, since~$N$ is odd). The result follows.
\end{proof}
\begin{remark} \emph{Let~$\alpha$ be a cyclotomic integer. Let~$\mathscr{N}(\alpha)$ denote
the minimum number of roots of unity required to express~$\alpha$. If~$\alpha \in K = \mathbf{Q}(\zeta_N)$,
let $\mathscr{N}_K(\alpha)$ denote the minimum number of roots of unity in~$K$ required to express~$\alpha$.
We recall the following facts from~\cite{Cassels,CMS} for cyclotomic
integers~$\alpha$:
\begin{enumerate}
\item If $\mathscr{N}(\alpha) > 1$ is not a root of unity, then $\mathscr{M}(\alpha) \ge 3/2$. (\cite{Cassels}, Lemma~2)
\item If $\mathscr{N}(\alpha) > 1$, and~$\alpha$ is not a root of unity times a conjugate of~$1 + \zeta_5$,
then~$\mathscr{M}(\alpha) \ge 5/3$.
\item If $\mathscr{N}(\alpha) \ge 3$, then~$\mathscr{M}(\alpha) \ge 2$. (\cite{Cassels}, Lemma~3)
\end{enumerate}
}
\end{remark}
\begin{prop} \label{prop:list} Suppose that $\mathscr{M}(\beta) < 14/5$ and~$\beta$ is a totally
real cyclotomic integer. Suppose, moreover, that
$\beta$ is not the sum of at most two roots of unity. Then~$\ho{\beta}$ is one of the following numbers:
\begin{center}
{\small
\begin{tabular}{rrr}
$\beta \qquad \qquad \qquad$ & $\mathscr{M}(\beta) $ & $[\mathbf{Q}(\beta):\mathbf{Q}]$ \\
\hline
$ \displaystyle{ \frac{\sqrt{3} + \sqrt{7}}{2} = 2.188901\ldots}$ & $5/2$ & $4$ \\
$ 1 + 2 \cos(2 \pi/7) = 2.246979\ldots$ & $2$ & $3$ \\
$ \zeta_{12} + \zeta_{20} + \zeta^{17}_{20} = 2.404867 \ldots$ & $2$ & $8$ \\
$ 2 \cos(11\pi/42) + 2 \cos(13\pi/42) = 2.486985 \ldots$ & $ 8/3 $ & $12$ \\
$ 1 + 2 \cos(2 \pi/11) = 2.682507\ldots$ & $12/5$ & $5$ \\
$ 1 + 2 \cos(2 \pi/13) = 2.770912 \ldots$ & $5/2$ & $6$ \\
$ 1 + 2 \cos(2 \pi/17) = 2.864944 \ldots$ & $21/8$ & $8$ \\
$ 1 + 2 \cos(2 \pi/19) = 2.891634 \ldots$ & $8/3$ & $9$ \\
$ 2\cos(2 \pi/35) + 2\cos(12 \pi/35) = 4 \cos(\pi/7) \cos(\pi/5) = 2.915596 \ldots$ & $5/2 $ & $6$ \\
$ 1 + 2 \cos(2 \pi/23) = 2.925834 \ldots$ & $30/11$ & $11$ \\
$ 1 + 2 \cos(2 \pi/29) = 2.953241 \ldots$ & $39/14$ & $14$ \\
$ 1 + 2 \cos(2 \pi/30) = 2.956295 \ldots$ & $11/4$ & $4$ \\
$ 1 + 2 \cos(2 \pi/60) = 2.989043 \ldots$ & $11/4$ & $8$ \\
$ \zeta_{84}^{-9} + \zeta_{84}^{-7} + \zeta_{84}^{3} + \zeta_{84}^{15} = 3.056668 \ldots$ & $5/2 $ & $12$ \\
$ 2\cos(6\pi/55) + 2\cos(16\pi/55) = 4 \cos(\pi/11) \cos(\pi/5) = 3.104984 \ldots$ & $ 27/10 $ & $10$ \\
$ 2\cos(8\pi/65) + 2\cos(18\pi/65) = 4 \cos(\pi/13) \cos(\pi/5) = 3.142033 \ldots$ & $ 11/4 $ & $12$ \\
$ 2\cos(11 \pi/70) + 2\cos(17 \pi/70) = 3.206780\ldots$ & $8/3 $ & $24$ \\
$ 2\cos(37 \pi/210) + 2\cos(47 \pi/210) = 3.227019 \ldots$ & $ 11/4 $ & $24$ \\
$ 2\cos(\pi/42) + 2\cos(11 \pi/42) = 3.354753 \ldots$ & $8/3 $ & $12$ \\
\end{tabular}
}
\end{center}
\end{prop}
\begin{proof}
We may assume that $\mathscr{N}(\beta) \ge 3$.
Consider the case~$\mathscr{N}(\beta) = 3$.
By Theorem~4.0.3 of~\cite{CMS}, we may assume that, up to conjugation and sign, either
$\beta = 1 + \zeta + \zeta^{-i}$ for some root of unity~$\zeta$, or
$\beta = \zeta_{12} + \zeta_{20} + \zeta^{17}_{20}$.
The latter element is included on the list, the former elements
satsify~$\mathscr{M}(\beta) \le 14/5$ if and only if they are included in the statement of the theorem. Hence
we may assume that~$\mathscr{N}(\beta) \ge 4$.
Let us now weaken the assumption on~$\beta$ to assume merely that it is equivalent to a totally
real cyclotomic integer, and that~$\mathscr{N}(\beta) \ge 4$. This allows us to also assume that~$\beta$ is minimal, that is, it
lives in~$\mathbf{Q}(\zeta_N)$ where~$N$ is the conductor of~$\beta$, and no multiple of~$\beta$ by a root
of unity lives in a smaller cyclotomic field.
Recall (following~\cite{Cassels,CMS}) that can write
$$\beta = \sum_{S} \alpha_i \zeta^i,$$
where~$p^k \| N$, where $\zeta$ is a primitive $p^{k}$th root of unity,
where $\alpha_i \in \mathbf{Q}(\zeta_M)$, where $pM = N$, and where~$S$
is a subset of $\{0,1,\ldots,p-1\}$ of order which we denote by~$X$.
Note that when~$p \| N$, this expression is only unique up to translating each~$\alpha_i$
by the same constant.
Assume that~$p^2 | N$ for some~$p$. Then $\mathscr{M}(\beta) = \sum \mathscr{M}(\alpha_i)$ (\cite{CMS}, Lemma~5.2.1).
If $|S| = X \ge 3$, then $\mathscr{M}(\beta) \ge 3$.
If $X =1$, then $\beta = \alpha \zeta$, and we could divide by~$\zeta$, contradicting the minimality of~$\beta$.
If $X = 2$, then $\mathscr{M}(\beta) = \mathscr{M}(\alpha_1) + \mathscr{M}(\alpha_2)$.
The assumption $\mathscr{N}(\alpha_1) + \mathscr{N}(\alpha_2) > 3$ implies that
$\mathscr{M}(\beta) \ge 3/2 + 3/2 = 3$ or $\mathscr{M}(\beta) \ge 1 + 2 = 3$. This also contradicts our assumptions, and so~$N$ is squarefree. Recall this
implies the equality (Eq.~3.9 of~\cite{Cassels}):
$$\mathscr{M}(\beta) = (p-X) \sum \mathscr{M}(\alpha_i) + \sum \mathscr{M}(\alpha_i - \alpha_j),$$
where we assume that exactly~$X$ of the~$\alpha_i$ are non-zero.
\medskip
Suppose that $p|N$ for some $p > 7$. Since $\mathscr{M}(\beta) < 7/2 \le (p+3)/4$,
then by Lemma~1 of~\cite{Cassels} (as used in~\cite{CMS}), we may assume that
there are exactly
of $X \le (p-1)/2$ non-zero terms~$\alpha_i$ in the expansion of~$\beta$ above. If $X \ge 4$, then
we deduce that
$$(p-1) \mathscr{M}(\beta) \ge (p - X) X \ge 4(p-4).$$
This implies (for~$p > 7$) that~$\mathscr{M}(\beta) \ge 14/5$.
Suppose that~$X = 3$.
If $\alpha_i$ is a root of unity for each~$i$, then $\mathscr{N}(\beta) \le 3$, a contradiction. Hence at least
one~$\alpha_i$ is not a root of unity. If all the~$\alpha_i$ are not roots of unity then~$(p-1) \mathscr{M}(\beta) \ge (p-3)(3/2)$
which directly leads to a contradiction. Otherwise, there must be at least two pairs which are non-zero, and so
$$(p-1) \mathscr{M}(\beta) \ge (p-3)(1 + 1 + 3/2) + 2,$$
from which $\mathscr{M}(\beta) \ge 3$.
Hence we may assume that~$X = 2$, and in particular that
$$\beta = \alpha + \zeta \gamma,$$
where $\zeta$ is a primitive $p$th root of unity, $\alpha$ and $\gamma$ are cyclotomic integers in $\mathbf{Q}(\zeta_M)$ for
$M$ dividing~$N$ and prime to~$p$.
Since~$\mathscr{N}(\beta) > 3$, either $\alpha$ is a root of unity and $\mathscr{N}(\gamma) \ge 3$, or
$\alpha$ and~$\gamma$
are both not roots of unity. In the first case, $\mathscr{N}(\gamma - \alpha) \ge 2$ so~$\mathscr{M}(\gamma - \alpha) \ge 2$.
Hence
$$(p-1) \mathscr{M}(\beta) \ge (p-2)(1 + 2) + 2,$$
and so~$\mathscr{M}(\beta) \ge 29/10$.
In the second case, if~$\alpha \ne \gamma$, then
$$(p-1) \mathscr{M}(\beta) \ge (p-2)(3/2 + 3/2) + 1,$$
and~$\mathscr{M}(\beta) \ge 14/5$. If~$\alpha = \gamma$ and~$\mathscr{M}(\alpha) \ge 5/3$, then~$\mathscr{M}(\beta) \ge 3$.
So, after conjugation, we must have:
$$\beta = (1 + \zeta)(1 + \zeta_5).$$
In this case, we have~$\ho{\beta} = \ho{{1 + \zeta}} \cdot \ho{{1 + \zeta_5}}$. Note that~$\ho{{1 + \zeta}} = 2\cos(\pi/p)$.
If~$p > 13$, then we have~$\mathscr{M}(\beta) \ge 45/32$, so this leaves only~$p = 11$ and~$p = 13$, and these
cases are covered in the statement of the theorem.
This portion of the argument is the one which most strongly requires
the bound~$\mathscr{M}(\beta) < 14/5$ rather than~$\mathscr{M}(\beta) < 3$.
In particular, all the integers
$4 \cos(\pi/p) \cos(\pi/5)$ for a prime~$p > 5$ will satisfy this bound.
\begin{lemma} \label{lemma:exhaust}
If $\beta \in K = \mathbf{Q}(\zeta_{105})$ is a sum of~$4$ or~$5$ roots of unity in~$K$, and~$\beta$ is
equivalent to a totally real integer, then either~$\ho{\beta}$ is one of the exceptions listed in the statement of the theorem, or $\mathscr{M}(\beta) \ge 14/5$.
\end{lemma}
\begin{proof}
One proceeds by enumeration, after noting by Lemma~\ref{lemma:real} that~$\beta \in K$ is equivalent to
a totally real integer if and only if $\beta$ times some $420$th root of unity is real.
\end{proof}
We let~$p = 5$, and
write $\beta = \sum \alpha_i \zeta^i$ where $\zeta^5 = 1$. We have the following
by Lemmas~7.0.1 and~7.0.3 of~\cite{CMS}:
\begin{lemma} \label{lemma:useful} If~$\alpha \in L = \mathbf{Q}(\zeta_{21})$,
\begin{enumerate}
\item If $\mathscr{N}_L(\alpha) \ge 2$, then $\mathscr{M}(\alpha) \ge 5/3$.
\item If~$\mathscr{N}_L(\alpha) \ge 3$, then~$\mathscr{M}(\alpha) \ge 2$.
\item If~$\mathscr{N}_L(\alpha) \ge 4$, then~$\mathscr{M}(\alpha) \ge 5/2$.
\item If~$\mathscr{N}_L(\alpha) \ge 5$, then~$\mathscr{M}(\alpha) \ge 23/6$.
\end{enumerate}
\end{lemma}
Since we are not assuming that~$N$
is divisible by~$5$, we have to allow the possibility that~$X = 1$.
We consider various cases:
\begin{enumerate}\item If $X = 1$, then $\beta \in \mathbf{Q}(\zeta_{21})$.
By the Lemma~\ref{lemma:useful}, we may assume that~$\mathscr{N}_L(\beta) > 5$. Hence~$\mathscr{M}(\alpha) \ge 23/6$, which is a contradiction.
\item If $X = 2$, then we may write $\beta = \alpha + \gamma \zeta$ with $\alpha, \gamma \in \mathbf{Q}(\zeta_{21})$, and
we have the equality:
$$4 \mathscr{M}(\beta) = 3 \mathscr{M}(\alpha) + 3 \mathscr{M}(\gamma) + \mathscr{M}(\alpha - \gamma).$$
Since $\mathscr{N}_K(\beta) > 5$, we may assume that either $\mathscr{N}_L(\alpha), \mathscr{N}_L(\gamma) \ge 3$, or~$\mathscr{N}_L(\alpha) = 2$ and~$\mathscr{N}_L(\gamma) \ge 4$,
or~$\mathscr{N}_L(\alpha) = 1$ and~$\mathscr{N}_L(\gamma) \ge 5$. Using~Lemma~\ref{lemma:useful}, and the fact that
$\mathscr{N}_L(\alpha - \gamma) \ge \mathscr{N}_L(\gamma) - \mathscr{N}_L(\alpha)$, we find in each case that:
$$\mathscr{M}(\beta) \ge \frac{1}{4} (3 \cdot 2 + 3 \cdot 2) = 3,$$
$$\mathscr{M}(\beta) \ge \frac{1}{4} (3 \cdot 5/3 + 3 \cdot 5/2 + 5/3) = 85/24,$$
$$\mathscr{M}(\beta) \ge \frac{1}{4} (3 + 3 \cdot 5/2 + 5/2) = 13/4,$$
which all yield
contradictions.
\item If $X = 3$, then, as in the proof of the similar step in Lemma~9.0.1 of~\cite{CMS}, not all the~$\alpha_i$ can be the same (since otherwise we could reduce to the case~$X = 2$), and hence
at least two of the $\alpha_i - \alpha_j$ are non-zero. More generally, we have
$$4 \mathscr{M}(\beta) \ge 2 \sum \mathscr{M}(\alpha_i) + \sum \mathscr{M}(\alpha_i - \alpha_j).$$
The values~$(1,1,1)$, $(1,1,2)$, $(1,2,2)$, $(1,1,3)$, are ruled out as values of~$\{\mathscr{N}_L(\alpha_i)\}$ by Lemma~\ref{lemma:exhaust}.
This leaves the possibilities:
$$(1,1,>3), (1,2,>2), (1,>2,>2), (>1,>1,>1).$$
Considering each in turn and using Lemma~\ref{lemma:useful}, along with the fact that not all the~$\alpha_i$ are equal
in the final case, we have the four estimates:
$$\mathscr{M}(\beta) \ge \frac{1}{4} (2 \cdot 1 + 2 \cdot 1 + 2 \cdot 5/2 + 2 + 2) = 13/4,$$
$$\mathscr{M}(\beta) \ge \frac{1}{4} (2 \cdot 1 + 2 \cdot 5/3 + 2 \cdot 2 + 1 + 1 + 5/3) = 13/4,$$
$$\mathscr{M}(\beta) \ge \frac{1}{4} (2 \cdot 1 + 2 \cdot 2 + 2 \cdot 2 + 5/3 + 5/3) = 10/3,$$
$$\mathscr{M}(\beta) \ge \frac{1}{4} (2 \cdot 5/3 + 2 \cdot 5/3 + 2 \cdot 5/3 + 1 + 1) = 3,$$
which all lead to a contradiction.
\item If $X = 4$ or $X = 5$, we may reduce to $X \le 3$ exactly as in the proof of Lemma~9.0.1 of~\cite{CMS}.
\end{enumerate}
\end{proof}
\section{Proof of Theorem~\ref{theorem:one}}
Recall that the Perron--Frobenius eigenvalue~$\lambda$ of a graph~$\Gamma$
is $< 2$ (respectively, $\le 2$) if and only~$\Gamma$ is a Dynkin diagram (respectively,
affine Dynkin diagram). For topological reasons, only finitely many of the spiders~$\Gamma_{\underline{k}}$
are affine Dynkin diagrams.
Assume that infinitely many of the~$\Gamma_{\underline{k}}$ are abelian. We proceed
by induction on~$k$, the result for~$k = 0$ being trivial.
If there exist infinitely many such graphs with $r_1 \le M$, then we may reduce the problem
to~$k - 1$ replacing~$\Gamma$ by the finitely many $1$-spiders on~$\Gamma$ with a $2$-valent tree
of length $\le M$ attached to~$\Gamma$ at~$v_1$. Hence we may assume that all the $r_i$ are tending to infinity.
If the limit of the~$\lambda$ as~$\underline{k}$ increases is $\le 2$, then all the~$\Gamma_{\underline{k}}$ are Dynkin diagrams.
Hence we may assume that the limit of the largest eigenvalue~$\lambda$ is $> 2$.
By Prop.~\ref{prop:degree}, we obtain a lower bound on~$[\mathbf{Q}(\lambda)^2:\mathbf{Q}]$ which allows us (for sufficiently
large~$n$) to deduce as in Corr.~\ref{corr:deduce} that~$\mathscr{M}(\lambda^2 -2 ) < 14/5$. Since~$\lambda > 2$,
it is not the sum of two roots of unity. It follows
that~$\lambda> 2$ must be one of the finitely many exceptional numbers occurring in Prop.~\ref{prop:list}.
Yet these numbers have (explicitly) bounded degree, and so using the lower
bounds on~$[\mathbf{Q}(\lambda^2):\mathbf{Q}]$ in the proof of Prop.\ref{prop:degree}, these eigenvalues
can occur as~$\lambda^2 - 2$ for only finitely many~$\Gamma_{\underline{k}}$.
Hence we may explicitly compute such an~$n$ such that~$\Gamma_{\underline{k}}$ is not abelian when
each~$r_i \ge n$,
which completes the proof of Theorem~\ref{theorem:one}.
\section{Examples}
We shall consider two examples. Let~$\Gamma_{a,b}$ be the Morrison spider given in the introduction.
Let $P_{a,b}(x)$ denote the characteristic polynomial of~$\Gamma_{a,b}$. We find:
\begin{lemma} There is an equality
$$ \left(t - \frac{1}{t}\right)^2 P_{a,b} \left(t + \frac{1}{t} \right)
= F_{a,b}(t) + F_{a,b}(1/t),$$
where
$$F_{a,b}(t) = t^{a+b}
(t^{-2} + 2 + 2 t^2 - 2 t^4 - 2 t^6 - 2 t^8 + t^{10}) + t^{a - b} (t^{-6} - 2 + t^6) .$$
\end{lemma}
Let~${\rho_{\infty}} = 1.6826\ldots $ be the largest real root of $t^6 - 2 t^4 - 2 t^2 - 1 = 0$, which is
also a root of
$$t^{-2} + 2 + 2 t^2 - 2 t^4 - 2 t^6 - 2 t^8 + t^{10} = 0$$
(the other roots of this polynomial are cyclotomic).
Let~$\gamma = \left({\rho_{\infty}} + {\rho^{-1}_{\infty}} \right)^2 = 5.18438\ldots$ denote the largest real root of
$$x^3 - 6 x^2 + 5 x - 4 = 0.$$
The following is the specialization of Lemma~\ref{lemma:general}:
\begin{lemma}
The polynomial $P_{a,b}(x)$ has a unique pair of roots $(\lambda,-\lambda)$ of
absolute value $> 2$. As~$a$ and~$b$ strictly increase, the value of $\lambda$ strictly increases.
The limit as $a,b \rightarrow \infty$ of~$\lambda^2$ is~$\gamma$.
\end{lemma}
We now find an explicit exponential bound relating~$\lambda$ to~$\gamma$.
\begin{lemma} \label{lemma:estimate}
Let $\rho \in [3/2,{\rho_{\infty}})$ denote the largest root of $P_{a,b}(t+t^{-1})$, and assume $a,b \ge n \ge 10$. Then
$$|\rho - {\rho_{\infty}}| < \frac{1}{6} (1.682)^{-2n}.$$
\end{lemma}
\begin{remark} \emph{When we write a real number as a finite decimal, we refer to an exact
element of~$\mathbf{Z}[1/10]$. Although the inequalities below are quite tight, they still hold by some
comfortable margin of error. Certain numbers are chosen to make various ratios integral, purely
for presentation purposes.}
\end{remark}
\begin{proof} Write $\rho = {\rho_{\infty}} - \varepsilon$.
For~$a,b \ge n \ge 10$, we have the estimate $\rho \in [1.682,{\rho_{\infty}})$.
In this range, the following inequalities hold:
$$|\rho^{-2} + 2 + 2 \rho^2 - 2 \rho^4 - 2 \rho^6 - 2 \rho^8 + \rho^{10}| > 270 \cdot \varepsilon,$$
$$|\rho^2 + 2 + 2 \rho^{-2} - 2 \rho^{-4} - 2 \rho^{-6} - 2 \rho^{-8} + \rho^{-10}| < 6,$$
$$|\rho^6 - 2 + \rho^{-6}| < 21.$$
The first inequality is obtained by looking at the derivative of this rational function
in the interval~$[1.682,\rho_{\infty}]$; the other inequalities are easy.
Using the equality $P_{a,b}(\rho + \rho^{-1}) = 0$ together with the triangle inequality, we find that
$$\begin{aligned}
270 \cdot \varepsilon \cdot \rho^{2n} \le & \ |\rho^{-2} + 2 + 2 \rho^2 - 2 \rho^4 - 2 \rho^6 - 2 \rho^8 + \rho^{10}|
\cdot \rho^{2n} \\
\le & \ |\rho^6 - 2 + \rho^{-6}| + |\rho^2 + 2 + 2 \rho^{-2} - 2 \rho^{-4} - 2 \rho^{-6} - 2 \rho^{-8} + \rho^{-10}|
\cdot
\rho^{-2n} \\
\le & \
42 + 6 \cdot \rho^{-2n} \\
\le & \ 45. \end{aligned}$$
The result follows.
\end{proof}
\begin{lemma} \label{lemma:estimatetwo}
If $a,b \ge n \ge 10$, and~$\lambda$ is the Perron Frobenius eigenvalue
of~$\Gamma_{a,b}$, then
$$ |\lambda^6 - 6 \lambda^4 + 5 \lambda^2 - 4| \
\cdot |\lambda^2|^{29/1000} \cdot |\lambda^2 - 2|^{14/100}
\cdot |\lambda^2 - 3|^{471/1000} \cdot |\lambda^2 - 4|^{362/1000}
\cdot |\lambda^6 - 6 \lambda^4 + 9 \lambda^2 - 1|^{8/625}$$
is bounded above by $23 \cdot (1.682)^{-2n}$.
\end{lemma}
\begin{proof} The function is decreasing on the interval~$[1.618,\rho_{\infty}]$. Hence,
by interlacing, it suffices to consider the case $a=b=n$.
The result is then an elementary calculus exercise from Lemma~\ref{lemma:estimate}.
The main point is that if one replaces~$\lambda$ in the above expression by~$t + 1/t$, the
resulting expression has derivative
$< 138 = 6 \times 23$ in $[1.618,\rho_{\infty}]$ (for comparison, the exact value at~$\rho_{\infty}$ is
approximately $\sim 136.12$).
\end{proof}
We can now give a lower bound on the degree of~$\lambda^2$, following
the argument of Prop~\ref{prop:degree}.
\begin{prop} \label{prop:degreetwo} Suppose that $a,b \ge n \ge 10$. Then
$$D = [\mathbf{Q}(\lambda^2):\mathbf{Q}] > \frac{11}{25} \cdot n - \frac{1}{3}.$$
\end{prop}
\begin{proof} All proper conjugates $\sigma \lambda^2 \ne \lambda^2$
satisfy $0 < \sigma \lambda^2 < 4$. Hence (by calculus)
$$ |\lambda^6 - 6 \lambda^4 + 5 \lambda^2 - 4| \
\cdot |\lambda^2|^{29/1000} \cdot |\lambda^2 - 2|^{14/100}
\cdot |\lambda^2 - 3|^{471/1000} \cdot |\lambda^2 - 4|^{362/1000}
\cdot |\lambda^6 - 6 \lambda^4 + 9 \lambda^2 - 1|^{8/625}$$
is bounded above in this interval by $10.56$.
(In contrast to the proof of Prop~\ref{prop:degree}, we include here some extra factors of $\lambda^2 - m$ for small~$m$ to
mollify the first factor as much as possible.)
On the other hand, since $\lambda \ne {\rho_{\infty}}$ is an algebraic integer, if $K = \mathbf{Q}(\lambda^2)$,
the product of the expression above over all conjugates of~$\lambda$ (assuming it is non-zero)
is a product of positive rational powers of norms, and is thus $\ge 1$. Using
the inequality above for $\sigma \lambda^2 \ne \lambda^2$ and Lemma~\ref{lemma:estimatetwo}
for $\sigma \lambda^2 = \lambda^2$, it follows that
$$1
< 23 \cdot (1.682)^{-2n} \cdot (10.56)^{D- 1}.$$
If the degree $D$ is less than the value in the theorem, the RHS is less than one.
\end{proof}
We deduce:
\begin{prop}
Suppose that~$a,b \ge 56$. Then~$\Gamma_{a,b}$ is not abelian.
\end{prop}
\begin{proof} In the context of Theorem~\ref{theorem:bounds}, with
$\beta = \lambda^2 - 2$ we have~$M = 1$ and~$L = \gamma - 2$, where
$B((\gamma - 2)^2) \sim -13.1241\ldots$.
This yields the upper bound~$\mathscr{M}(\lambda^2 - 2) < 14/5$ as soon
as $D \ge 20|B|/11$, or when~$D \ge 24$. By Prop~\ref{prop:degreetwo}, we have~$D > 24$ as soon
as~$n \ge 56$. Hence, in this range,~$\lambda^2 -2$ must be one of the exceptions listed
in Prop.~\ref{prop:list}. On the other hand, for~$n$ in this range, we also have the estimate
$3.17438\ldots < \gamma -2 - 1/100 < \lambda^2 - 2 < \gamma - 2 = 3.18438 \ldots$, which certainly rules out all such
exceptions.
\end{proof}
To complete the proof of Theorem~\ref{theorem:morrison}, it suffices to consider the case when~$a \le 56$
or~$b \le 56$ (since the polynomial~$P_{a,b}(x)$ is symmetric in~$a$ and~$b$, we may assume the former).
However, we can now apply the algorithm of~\cite{CMS} to rule out the remaining cases (we thank Scott Morrison
for carrying out this computation). We could also
rule out the cases using the methods in this paper, however, we omit the details for reasons of space, and
because we include the relevant details in the case of~$3$-spiders below.
\section{\texorpdfstring{$3$}{3}-Spiders}
We consider the case when~$k = 3$ and~$\Gamma$ is a single point. Let the resulting~$3$-spider be denoted~$\Gamma_{a,b,c}$.
\begin{theorem} \label{theorem:3spider} The complete set of abelian~$3$-spiders is as follows:
\begin{enumerate}
\item Those that are Dynkin diagrams, equivalently, those with $\lambda^2 \le 4$:
\begin{figure}[H] \label{fig:two}
\begin{center}
\includegraphics[width=60mm]{Dynkin.png}
\end{center}
\end{figure}
\item Exactly three spiders with $\displaystyle{\lambda^2 = \frac{5 + \sqrt{13}}{2} = 4.302775\ldots}$
\begin{figure}[H] \label{fig:three}
\begin{center}
\includegraphics[width=60mm]{Haagerup.png}
\end{center}
\end{figure}
\item Exactly three spiders with $\displaystyle{\lambda^2 = \zeta^{11} + \zeta^{10} + \zeta^{3} + \zeta^2 + 2 = 4.377202 \ldots}$,
where $\zeta = \exp(2 \pi i/13)$:
\begin{figure}[H] \label{fig:four}
\begin{center}
\includegraphics[width=60mm]{BigHaagerup.png}
\end{center}
\end{figure}
\item Exactly three spiders with $\displaystyle{\lambda^2 =3 + \sqrt{2} = 4.414213 \ldots}$, namely:
\begin{figure}[H] \label{fig:five}
\begin{center}
\includegraphics[width=60mm]{Mystery.png}
\end{center}
\end{figure}
\end{enumerate}
\end{theorem}
\begin{remark} \emph{The first two non-Dynkin diagrams include graphs
which correspond to the Haagerup and extended Haagerup subfactors respectively
(namely, the~$(3,3,3)$ and~$(3,3,7)$ spiders). However, none of the final class of graphs
correspond to any subfactors, because the index is~$< 5$ and does not occur as an index
of any possible subfactor in the classification~\cite{class}.}
\end{remark}
By symmetry, we may assume that $a \le b \le c$.
Let $P_{a,b,c}(x)$ denote the characteristic polynomial of~$\Gamma_{a,b,c}$. Using Lemma~11 of~\cite{Salem}, one easily establishes the following equality:
\begin{lemma} There is an equality
$$\begin{aligned} P(t) \left(t - t^{-1}\right)^3 (-1)^{a+b+c-1} = & \ t^{a+b+c+4} - 2 t^{a+b+c+2} + t^{a+b-c} + t^{a+c-b} + t^{b+c-a} \\
& \ - t^{a-b-c} - t^{b-a-c} - t^{c-a-b}
+ 2 t^{-a-b-c-2} - t^{-a-b-c-4}. \end{aligned}$$
\end{lemma}
It is easy to identify the triples~$(a,b,c)$ such that~$\Gamma_{a,b,c}$ is a Dynkin diagram,
so we assume that the Perron--Frobenius eigenvalue~$\lambda$ of~$\Gamma$ is always strictly
larger than~$2$.
From Lemma~\ref{lemma:general}, we deduce that
the polynomial $P_{a,b,c}(x)$ has a unique pair of roots $(\lambda,-\lambda)$ of
absolute value $> 2$, and that the
limit as $a,b,c \rightarrow \infty$ of~$\lambda$ is~$3/\sqrt{2}$.
\begin{prop} Let $D = [\mathbf{Q}(\lambda^2):\mathbf{Q}]$. Then either
$\mathscr{M}(\lambda^2 - 2) < 14/5$ or~$D \le 12$.
\end{prop}
\begin{proof} It suffices to note that (taking~$\beta = \lambda^2 - 2$) that at most
one conjugate of~$\beta$ lies outside~$[-2,2]$, and so we
deduce the inequality on~$\mathscr{M}(\lambda^2 - 2)$
from
Theorem~\ref{theorem:bounds} providing
$$D \ge 2 \cdot \left| B \left(\frac{25}{4}\right) \right| = 12.904524\ldots$$
\end{proof}
Let us now make the running assumption that~$\mathscr{M}(\lambda^2 -2) \ge 14/5$; we shall
deal with the alternative below. It follows that we may assume that
$D = [\mathbf{Q}(\lambda^2):\mathbf{Q}]$ is at most~$12$.
\begin{lemma} With $a \le b \le c$, we have~$a \le 30$.
\end{lemma}
\begin{proof} Since it is useful to have a tight a bound as possible, instead of using the
trivial bound~$|2x-9| \le 9$ on $[0,4]$ we note that
$$|2x-9| \cdot |x|^{52/100} \cdot |x - 1|^{337/1000} \cdot |x-2|^{3/10} \cdot |x-3|^{13/100} < 5.58,$$
for~$x \in [0,4]$, a fact which is tedious but elementary to prove by calculus.
By giving a lower estimate for the derivative of this function in a neighbourhood of $9/2$, we also find that
$$|2\lambda^2-9| \cdot |\lambda^2|^{52/100} \cdot |\lambda^2 - 1|^{337/1000} \cdot |\lambda^2-2|^{3/10} \cdot |\lambda^2-3|^{13/100} < 4.63 \cdot |(2 \lambda^2 - 9)|,$$
for all~$\lambda$.
Taken together, we deduce that
$$1 \le (5.58)^{D-1} \cdot 4.63 \cdot
|(2 \lambda^2 - 9)|,$$
and hence, since~$D \le 12$,
$$\left| \lambda^2 - \frac{9}{2} \right| > 6.6132 \ldots \times 10^{-10}.$$
This inequality is violated as soon as~$a > 30$.
\end{proof}
\subsection{Fixed~\texorpdfstring{$a$}{a}, and
varying~\texorpdfstring{$b$}{b} and~\texorpdfstring{$c$}{c}}
In this section, we effectively consider the~$2$-spiders on the Dynkin diagram~$\Gamma = A_{n}$
with~$n = a+1$ and $v_1 = v_2$ a terminal point of~$\Gamma$. Hence, for this section,
the values of~$\rho_{\infty}$ reflects the appropriate root of the new polynomial~$Q(t+t^{-1})$
in this setting. Note, however, that we still know that~$\lambda^2$ has a unique conjugate
outside the range~$[0,4]$.
Suppose that~$a$ is fixed, and let~$c$ and~$b$ with $c \ge b \ge a$ vary without bound.
If one writes~$\lambda^2 - 2 = \rho^2 + \rho^{-2}$, then~$\rho^2$ is a Salem number, that is,
all the conjugates of~$\rho^2$ beside~$\rho^{-2}$ have modulus one. Since we
are assuming $D=[\mathbf{Q}(\lambda^2):\mathbf{Q}] \le 12$, we also have the inequality
$[\mathbf{Q}(\rho^2):\mathbf{Q}] \le 24$. As~$b$ and $c$ tend to infinity,~$\rho$ tends towards the (unique)
largest root~$\rho_{\infty}$ of the polynomial~$1 - 2t^{2a+2} + t^{2a+4}$, which
is the polynomial~$Q(t + t^{-1})$ (up to powers of~$t^{\pm 1}$) of~\S~\ref{section:spectrum}
\begin{lemma} \label{lemma:foursix}
We have an inequality:
$$|1 - 2 \rho^{2a +2} + \rho^{2a + 4}| > \frac{1}{4^{23}}.$$
\end{lemma}
\begin{proof} Since~$\rho_{\infty} > \rho$ is the only real root of this polynomial greater than one,
it follows that neither~$\rho$ nor any of its conjugates is a root of this polynomial.
For any non-trivial conjugate of~$\rho^2$, we have
the easy estimate $|1 - 2 \sigma \rho^{2a+2} + \sigma \rho^{2a + 4}| \le 4$, with a strict inequality
for the real root. Hence the result follows from the fact that the norm of $1 - 2 \rho^{2a+2} + \rho^{2a+4}$
from~$\mathbf{Q}(\rho^2)$ to~$\mathbf{Q}$
has absolute value at least one, and that the degree of~$\rho^2$ is at most~$24$.
\end{proof}
By interlacing, the root~$\rho$ increases with~$b$ and~$c$. Hence, by checking
for suitable choices of~$b$ and~$c$, we immediately deduce:
\begin{lemma} For each~$a$, we have the following upper bound on~$b = \min(b,c)$:
\begin{center}
\begin{tabular}{cc|cc|cc}
$a$ & $\min(b,c)$ & $a$ & $\min(b,c)$ & $a$ & $\min(b,c)$ \\
\hline
$1$ & $67$ & $11$ & $59$ & $21$ & $69$ \\
$2$ & $55$ & $12$ & $60$ & $22$ & $70$ \\
$3$ & $53$ & $13$ & $61$ & $23$ & $71$ \\
$4$ & $53$ & $14$ & $62$ & $24$ & $72$ \\
$5$ & $53$ & $15$ & $63$ & $25$ & $73$ \\
$6$ & $54$ & $16$ & $64$ & $26$ & $74$ \\
$7$ & $55$ & $17$ & $65$ & $27$ & $75$ \\
$8$ & $56$ & $18$ & $66$ & $28$ & $76$ \\
$9$ & $57$ & $19$ & $67$ & $29$ & $77$ \\
$10$ & $58$ & $20$ & $68$ & $30$ & $78$ \\
\end{tabular}
\end{center}
\end{lemma}
\subsection{Fixed~\texorpdfstring{$a$}{a} and
~\texorpdfstring{$b$}{b}, and varying~\texorpdfstring{$c$}{c}}
We have reduced to a finite number of pairs~$(a,b)$, and we could finish
with an appeal to~\cite{CMS}. Instead, however, we give a
a treatment similar to the case when~$a$ is fixed and~$b$ and~$c$ are varying.
As in the previous section, we assume~$c \ge b \ge a$, and redefine the polynomials~$Q(t+t^{-1})$
and~$\rho_{\infty}$ (for each~$(a,b)$) to be the corresponding values for these~$1$-spiders.
\begin{lemma}
We have an inequality:
$$| {\rho}^{2a+2b+4} - 2 {\rho}^{2a+2b+2} + {\rho}^{2b} + {\rho}^{2a} - 1| > \frac{1}{6^{23}}.$$
\end{lemma}
\begin{proof} The proof is the same as the proof of Lemma~\ref{lemma:foursix}; the polynomial
above is the minimal polynomial of~$\rho_{\infty}$.
\end{proof}
By interlacing and computing the values of~$\rho$ for various triples~$(a,b,c)$,
we deduce:
\begin{lemma} If~$\Gamma_{a,b,c}$ is abelian, then one of the following holds:
\begin{enumerate}
\item There are bounds~$a \le 30$, $b \le 78$,
and~$c \le 170$.
\item $\mathscr{M}(\lambda^2 - 2) < 14/5$.
\end{enumerate}
\end{lemma}
We now complete the proof of Theorem~\ref{theorem:3spider}. Suppose that~$\mathscr{M}(\lambda^2 -2) < 14/5$.
Then by Prop.~\ref{prop:list}, we deduce that either~$\lambda^2 - 2$ is a sum of two roots of unity
or less (from which it follows immediately that~$\Gamma_{a,b,c}$ is an (affine) Dynkin diagram, or
$\lambda^2 -2$ is one of the following numbers:
$$\alpha = \frac{\sqrt{3} + \sqrt{7}}{2}, \beta = 1 + 2 \cos(2 \pi/7), \gamma =
\zeta_{12} + \zeta_{20} + \zeta^{17}_{20},
\delta = 2 \cos(11\pi/42) + 2 \cos(13\pi/42),$$
where we use the fact that~$\beta^2 < 9/2$. The algebraic
numbers~$\alpha$ and~$\delta$ have conjugates $< 2$,
yet~$\lambda$ is totally real, so~$\lambda^2 - 2$ has no such conjugate.
In the second and third cases, we have
$$\beta \sim 2.060820\ldots \ \text{or} \ \gamma \sim 2.098777\ldots $$
We dispense with these possibilities (for $(a,b,c)$ outside the bounds in part one) by the following argument:
\begin{enumerate}
\item
If $a \ge 3$, then $\lambda > 2.074313\ldots > 2.060820\ldots$
\item If $a = 2$ and $b \ge 4$, then $\lambda > 2.074313\ldots > 2.060820\ldots$
\item If $a = 2$ and $b = 2$, then $\lambda < \sqrt{2+\sqrt{5}} < 2.060820\ldots$
\item If $a = 2$, $b = 3$, and $c \ge 5$, then $\lambda > 2.069782\ldots > 2.060820\ldots$
\item If $a =1$, then $\lambda < \sqrt{2+\sqrt{5}} < 2.060820\ldots$
\end{enumerate}
for the first case, and
\begin{enumerate}
\item
If $a \ge 4$, then $\lambda > 2.101002\ldots > 2.098777\ldots$
\item If $a = 2$ and $b \ge 5$, then $\lambda > 2.101002\ldots > 2.098777\ldots$
\item If $a = 2$ and $b \le 4$, then $\lambda < 2.084868\ldots < 2.060820\ldots$
\item If $a \le 2$, then $\lambda < 2.093555\ldots < 2.098777\ldots$
\end{enumerate}
in the second.
Finally, we check all the remaining polynomials to see which give rise to abelian extensions.
We say a few words about this computation. The first step consists of looping through the polynomials
(which have root~$\rho$)
and dividing through by the cyclotomic factors. If the remaining polynomial
is irreducible and of degree $\ge 48$, then we are done.
Degree considerations eliminated all polynomials with~$a \ge 12$ except some of the form
$(a,b,c)= (a,a+1,2a+3)$, $(a,a+2,a+2)$ or $(a,a,a)$. The polynomial was irreducible
except for a few exceptional cases, namely, $(a,b,c) = (2,6,20)$, and the triple
of graphs $(4,8,14)$, $(4,9,9)$, and $(5,5,8)$. The latter triple is somewhat interesting --- the value of~$\lambda^2 -2$
in each case is the largest real root of~$\theta^3 - 2 \theta^2 - 4 \theta + 7 =0$, whose splitting field is the Hilbert class field
of~$\mathbf{Q}(\sqrt{229})$.
The second check consisted of computing the corresponding minimal polynomial of~$\lambda^2 - 2$,
and then checking (using \texttt{polcompositum} in \text{gp/pari}) whether the field was Galois or not. Finally, it was checked whether any of the fields thus obtained were abelian or not (there were no false positives).
\subsection{Miscellaneous Applications}
All Salem numbers~$\rho$ are reciprocal. If~$\mathbf{Q}(\rho)$ is abelian, then since~$\rho$ is real,
it must be totally real, yet~$\rho$ (by definition) has a root of absolute value~$1$. Thus no
Salem number can generate an abelian extension. In light of this, the following definition
is perhaps not too confusing.
\begin{df} A Salem number~$\rho$ is of abelian type if~$\mathbf{Q}(\rho + \rho^{-1})$ is an abelian extension.
\end{df}
If~$K$ is any totally real field, then, because the image of the units~$\mathcal{O}^{\times}_K \otimes \mathbf{R}$ in
$K \otimes \mathbf{R}$ has co-dimension one (by the proof of Dirichlet's unit theorem), there exists a totally
positive unit~$\alpha \in \mathcal{O}_K$ such that $\alpha > 1$ in one complex embedding and $< 1$ in all
other complex embeddings. Replacing~$\alpha$ by a suitable power so that it is $>2$ and letting
~$\rho + \rho^{-1} = \alpha$, we find that~$\rho$ is a Salem number of abelian type if~$K$ is a (totally) real abelian
field. Hence there exist an abundance of Salem numbers of abelian type. However, we prove the following:
\begin{prop} The set of Salem numbers of abelian type is discrete in~$\mathbf{R}$.
\end{prop}
\begin{proof}
It suffices to show that the Salem numbers of abelian type less than a given bound~$L$ is finite.
Since the number of Salem numbers of bounded degree less than a given bound is finite, it
suffices to prove that the Salem numbers of abelian type less than~$L$ have bounded degree.
However, from Theorem~\ref{theorem:bounds}, for all Salem numbers of sufficiently large degree
(depending on~$L$) we have the bound~$\mathscr{M}(\rho + \rho^{-1}) < 14/5$.
If~$\rho$ is of abelian type,
then by Prop.~\ref{prop:list}, the element~$\rho$ lives in some finite set (if~$\rho$ is Salem,
then~$\rho + \rho^{-1} > 2$ is not a sum of two roots of unity).
\end{proof}
Note that, from the classification of the smallest totally real cyclotomic integers~\cite{CMS}, one sees that the smallest Salem number of abelian type is~$\theta = 1.635573\ldots$, the root of
$\theta^6 - 2 \theta^5 + 2 \theta^4 - 3 \theta^3 + 2 \theta^2 - 2 \theta + 1 = 0$.
One can make the previous proposition effective. Namely, suppose that $\rho > \rho'$
are two Salem numbers of abelian type. There is a bound~$B(x^2 + x^{-2} + 2) > - 11/10 x^2$ for all~$x > \theta$. Hence
$$\frac{20}{11} \cdot B(\rho^2 + \rho^{-2} + 2) \ge - 2 \rho^2,$$
and so either the degrees of~$\rho$ and~$\rho'$ are either bounded by $4 \rho^2$, or
the corresponding Salem numbers lie on the list in Prop.~\ref{prop:list}, in which case one can
check that the bound still holds.
In the former case, by estimating the norm of~$\rho - \rho'$, which has degree at most~$16 \rho^4$ and each conjugate has absolute value at most~$2 \rho$, we deduce that
\begin{prop} Let $\rho > \rho'$ be two Salem numbers of abelian type. Then
$$\rho - \rho' > \frac{1}{(2 \rho)^{16 \rho^4}}.$$
\end{prop}
Naturally enough, the same result (and proof) hold if one replaces Salem numbers by numbers~$\rho$
conjugate to~$\rho^{-1}$
with a uniformly bounded number of real roots $> 1$.
\bibliographystyle{amsalpha}
|
\section{Introduction}
\noindent
Microblogging services have become increasingly popular in recent years allowing users to share small snippets of text with the world. Users share a variety of information from personal opinions and thoughts to breaking news and product marketing messages. Of the many microblogging services available, Twitter has become the dominant service with estimates of 250 million registered users who post up to 500 million Tweets per day [12].
\\
\noindent
As Twitter grew in popularity, it became apparent that an organizational strategy was needed to organize the massive amounts of data. In 2009, Twitter introduced a user driven annotation called a hashtag [3]. Today, hashtags are used for many purposes such as finding posts of interest, organizing a group message, and determining the most popular topics of discussion. Users can apply hashtags to their posts by placing the number sign (\#) before a word they want to use as a hashtag. Users are free to create almost any hashtag they can imagine as long as it is a single alphanumeric string of text. The practical ramifications of this mean that hashtags containing more than one word are simply pushed together by removing spaces so these hashtags are not a single correctly spelled word. For example, a user who enjoys big data processing might Tweet something like "I heart \#bigdataprocessing".
\\
\noindent
Hashtags are simple to apply and well known to mainstream Twitter users. However, hashtags are only used in approximately 10\% of Tweets [4]. This limited usage greatly reduces the usefulness of the hashtag as a global organizational strategy. If hashtags could be recommended to users before they post their Tweet, we suspect hashtag usage would increase dramatically.
\\
\noindent
In this paper, we develop a method for recommending relevant hashtags to users in real-time. Hashtag recommendation has two main challenges that set it apart from traditional document tag recommendations. First, the content of a Tweet is very short and often includes abbreviations, misspellings, and incorrect grammar. The limited number of words in a Tweet makes traditional document classification techniques such as TF-IDF (Term Frequency and Inverse Document Frequency) ineffective because of the TF=1 challenge which means that words are rarely repeated in a Tweet [10]. This makes it difficult to determine which words are most important to the meaning of the Tweet. The second challenge is the sheer volume and speed at which data is produced. Tweets are produced constantly so the incomming Tweets need to be treated as an infinite length stream. Our approach to recommendation attempts to handle these challenges with a variety of preprocessing techniques and a recommendation model based on the popular classification algorithms K-Nearest Neighbor and Naive Bayes.
\section{Constraints}
\noindent
Three main constraints were taken into account when developing our recommendation system. First, our recommendations are based solely upon the content of the Tweet. We scoped the problem this way because the content of a Tweet must be the core factor of any hashtag recommendation system so it is essential to develop a solid approach using this factor alone before extending the recommender to consider other factors like the author's social network or previous posting history.
\\
\noindent
The second constraint is speed. For this appraoch to be useful in a real world application, we need to generate many recommendations very rapidly. Today's internet users are unwilling to wait for a web page for even a few seconds [14] so the usefulness would be greatly reduced if recommendations could not be produced in milliseconds. Also, as previously mentioned, Twitter has seen as many as 500 million posts per day, so a recommendation system needs to be able to scale to handle large volumes of input.
\\
\noindent
The third and final constraint is computational resource demands. The recommendation system needs to process and infinite length stream, run indefinitely, and scale gracefully as the stream progresses so demands on computional resources like CPU and memory need to remain relatively constant.
\section{Prior Work}
\noindent
There has been significant research on keyword extraction, document classification, and tag recommendation but far less as it relates to very short documents like microblogging posts. In this section, we briefly review the main directions for the studies related to this paper.
\\
\noindent
In some ways, hashtag recommendation is similar to a typical text classification problem and various algorithms from regression and classification analysis might be applied here [7][18]. Other studies focus on classification of short-text documents but commonly use only a limited number of features. Therefore they almost always degrade if we consider each hashtag as a separate class. The main reason for this is the obvious obstacles of maintaining and processing an exceedingly big number of classes [2][8].
\\
\noindent
Other related methods for hashtag recommendation are based on measuring the similarity of tweets. Zangerle[5] and Godin[6] have several recommendation approaches based on TF-IDF methods and feature vectors for each tweet. Also, several papers have been published on recommending hashtags by using external meta data sources such as WordNet and Wikipedia to build context which could be applied to the text of post [11][15]. A study by Kharibi[4] considers a tweet as a set of words and measures a relevance between the words by aggregating all relevance scores which are computed from a co-occurrence graph. Feng's research [16] develops and expands this approach both with variety of optimizations and improvements and by processing a user's personal information. In this paper, we apply concepts from these papers to develop our own hashtag recommendation approach.
\section{Our Approach}
\noindent
Our hashtag recommendation process can be broken down into a three main topics: preprocessing, classification, and computational resource management.
\subsection{Preprocessing}
\noindent
Preprocessing is extemely important in many big data applications because it provides a mechanism to filter, aggregate, and cleanse the dataset making the actual processing more accurate and efficient. Hashtag recommendation is one such application that benefits greatly from this approach. Tweet content is very noisy when compared to many other document datasets because character constraints force abbreviations and posts are written by millions of different people so there is no consistency between authors. Because of this, preprocessing Tweets is an essential first step to minimize the impact of this noise to prepare the data for accurate classification.
\\
\noindent
The first preprocessing step was to normalize the character set by converting all alpha characters to lower case and stripping out punctuation. Our classifiers work by finding terms that appear in multiple Tweets so this normalization step allows us to group words with the same meaning. For example, "Don't", "don't", and "dont" are all the same word but are written differently. This normalization will convert each of these words to "dont" so each occurrence will be grouped together making the distribution of terms more favorable for processing and more reflective of the intent of the authors.
\\
\noindent
For our classifiers to work most effectively, they need to base the classification on words with the most significance to the meaning of the Tweet. To maximize the likelihood of this, we removed words that have a high probability contributing little to the meaning of the Tweet. We removed 175 common English stop words like "a", "the", and "and" as well as common words found on Twitter like "rt" (i.e. retweet). We also removed any words that contained less than 3 characters and words made up entirely of numbers. We also exclude links from consideration in our classifiers. Because of the 140 character limit, links in Twitter are generally created using url shortening tools like bit.ly (https://bitly.com) which generates unique links for each user even if they point to the same destination page. Because of this, very little information can be gleaned from the url.
\\
\noindent
The final preprocessing step we investigated was stemming using the Porter Stemming algorithm [9] which removes prefixes and suffixes from each word in a Tweet. The intution was that this would improve the accuracy of the classifier because terms with a common stem usually have similar meanings [9]. In practice however, we found this to have neglible impact on the accuracy of our recommendations as well as a non-trivial decrease in recommendation speed. Because of this, we excluded this preprocessing step in our final algorithm. We did not investigate why stemming did not improve accurracy but our suspicion is that the many abbreviations, mispellings, and slang terms commonly used on Twitter interfere with the stemming algorithm. Further research is necessary here to determine if modifications to the stemming algorithm could improve results.
\subsection{Classifications}
\noindent
Our recommendation engine is based primarily on classification. We evaluated a variety of clustering and classification algorithms and selected K-Nearest Neighbor and Naive Bayes to base our recommendations on. These algorithms are straight forward to implement but more importantly, they are lazy evaluating algorithms. This means the underlying data of the classification model can be modified without needing to rebuild the entire model. By contrast, decision trees and rule based classifiers require the decision tree or rule sets to be regenerated if the model data changes.
\subsubsection{Naive Bayes}
\noindent
Naive Bayes classification uses Bayes' Theorem to determine the conditional probability of a class given an item's features. In this application, the class is the hashtag and the features are the words in a Tweet. Bayes Theorem makes it possible to calculate the probability of a hashtag given a set of words in a Tweet. If this is done for all known hashtags, the hashtags with the highest probability can be used as the recommendations.
\begin{theorem}[Bayes' Theorem] Multiple Features
\begin{eqnarray*} P(h|w_{1},w_{2},...,w_{n}) = \frac{P(w_{1},w_{2},...,w_{n}|h) * P(h)}{P(w_{1},w_{2},...,w_{n})} = P(h|w_{1}) * P(h|w_{2}) * ... * P(h|w_{n}) \end{eqnarray*}
\end{theorem}
\noindent
However, there are a few problems with this approach. In many cases, $P(h|w_{i})$ will be zero when we have not yet seen a Tweet containing word $w_{i}$. If this occurs, the entire $P(h|w_{1},w_{2},...,w_{n})$ goes to zero even if other words have high conditional probabilities. We investigated methods such as Laplace Smoothing which did improve our results, but we found that summing the conditional probabilities instead of using the product produced even better results. The intuition behind this is that in many cases a single word is a strong predictor of the hashtag. If most of the words in a Tweet are unrelated to a hashtag but one word is very related, the unrelated words drag down the overall probability of that hashtag when computing the product. By summing the conditional probabilities, we preserve the high weighting contributed by the high probability words. Summing also handles the issue where $P(h|w_{i})$ is zero because a word with a score of zero has no negative impact on $P(h|w_{1},w_{2},...,w_{n})$.
\\
\noindent
Below is our optimal Naive Bayes algorithm. Let $W$ be the list of words in the Tweet we are classifying. Let $T$ be the list of Tweets in our classification model. Let $n$ be the number of hashtag classifications to return.
\\
\begin{algorithm}
\begin{algorithmic}
\STATE $T_{s} = SIMILAR\_TWEETS(W, T)$
\STATE $H = T_{s}.hashtags$ //All hashtags in list of Tweets
\STATE $classifications$ = [ ]
\FOR {$h$ in $H$}
\STATE $score$ = 0
\FOR {$w$ in $W$}
\STATE $score$ += $(\frac{P(w|h) * P(h)}{P(w)} * WORD\_WEIGHT(w))$
\ENDFOR
\STATE $classifications.add(h, score)$
\ENDFOR
\RETURN $getHighestScores(classifications,n)$ //Return n highest scored hashtags
\end{algorithmic}
\caption{[$NB\_CLASSIFY({W, T, n})$]}
\end{algorithm}
\subsubsection{K-Nearest Neighbor}
\noindent
K-Nearest Neighbor classifiers find the K most similar data points to the item being classified and return the most common class found in those K items. The key to a successful K-Nearest Neighbor algorithm is the similarity measure to compare two items. There are many ways to do this from simple calculations like the Tanimoto Distance (number of words found in both documents) to weighting measures like TF-IDF (Term Frequency and Inverse Document Frequency). For the Tweet dataset, we found the most successful similarity measure to be the sum of the Term-Corpus Relevance (TCoR) [10] scores for all words found in both Tweets. TCoR is a weighting measure proposed by Timonen, et al. to measure how strong of a class predictor the word is across the entire dataset.
\\
\begin{definition}
\begin{algorithmic}
\STATE $TCoR(w) = \frac{\frac{1}{fl(w)} + \frac{1}{c_{w}}}{2}$
\end{algorithmic}
\end{definition}
\noindent
$fl(w)$ is the average number of words in Tweets containing the word and $c_{w}$ is the number of hashtags the word co-occurs with. The intuition for TCoR is that the fewer number of words in a Tweet, the more important an individual word is to the overall meaning of that Tweet and words occurring with a small number of distinct hashtags are more imformative when predicting a hashtag [10]. Timonen, et al. propose a second measure called Term-Category Relevance (TCaR) to combine with TCoR for the final weight. Interestingly, using TCaR reduced the accuracy of our recommendations by almost 50\% so we used TCoR alone to weight the occurrence of words.
\\
\noindent
Below is our optimal K Nearest Neighbor algorithm. Let $W$ be the list of words in the Tweet we are classifying. Let $T$ be the list of Tweets in our classification model. Let $K$ be the number of nearest neighbors to consider. Let $n$ be the number of hashtag classifications to return.
\\
\begin{algorithm}
\begin{algorithmic}
\STATE $T_{s} = SIMILAR\_TWEETS(W, T)$
\STATE $neighbors =$ [ ]
\FOR {$t$ in $T_{s}$}
\STATE $score = TCoR(w) * WORD\_WEIGHT(w)$
\STATE $neighbors.add([t, score])$
\ENDFOR
\STATE $nearest\_neighbors = getHighestScores(neighbors, K)$ //Get K tweets with highest score
\STATE $classifications =$ [ ] //Map of hashtag and count
\FOR {$t$ in $nearest\_neighbors$}
\FOR {$h$ in $t.hashtags$}
\STATE $classifications.add([h,1])$ //Add h with count of 1 (or increment count)
\ENDFOR
\ENDFOR
\RETURN $getHighestScores(classifications, n)$ //Return n hashtags with highest count
\end{algorithmic}
\caption{[$KNN\_CLASSIFY({W, T, K, n})$]}
\end{algorithm}
\subsubsection{Shared Classifier Functions}
\noindent
Both the Naive Bayes and K Nearest Neighbor classifiers share a few other enhancements to improve classification accuracy and speed. We could compute the classification by considering every Tweet in the dataset but that would be very slow as the dataset could contain millions of Tweets. Instead, we use $SIMILAR\_TWEETS(W, T)$ to quickly narrow down the full list of Tweets to just the most likely matches.
\\
\noindent
We do this by computing the Inverse Document Frequency (IDF) for all the words in $W$. IDF is a commonly used measure of how common a word is across a set of documents. It is an important measure because not all words contribute equally to the meaning of a document. A high IDF score indicates the word is rare across the document set which means it is likely important to the meaning of a document in which it is found.
\\
\begin{definition}
\begin{algorithmic}
\STATE $IDF(w) = log(\frac{|T|}{|w \in T|})$
\end{algorithmic}
\end{definition}
\noindent
We select the three words from $W$ with highest IDF scores as these are likely to be the three most important words in the Tweet we are classifying. We then find all Tweets in our dataset that contain at least one of those three words. This list of Tweets is just a fraction of the size of the entire Tweet dataset and likely contains Tweets most similar to the one we are classifying. Limiting our classification algorithms to investigate only this list of Tweets improves the speed of classifications and also helps to filter some of the noise created by high frequency words.
\\
\noindent
The second shared function is $WORD\_WEIGHT(w)$ which computes additional weighting for a given word. If a word starts with an '@' it is a mention, which means the Tweet is directed to a certain user. We apply a weight of 3 to all mentions because Tweets directed to the same user have a high likelihood of being similar and containing the same hashtag. If the word is not a mention, we give it a weight of 1 plus 0.1 for every letter in the word. This weights longer words more heavily than shorter words. The intuition is that longer words generally have a more specific meaning and are more likely to be important to the meaning of the Tweet.
\subsubsection{Hybrid Classifier}
\noindent
Our final classification model makes use of both the Naive Bayes and K Nearest Neighbors classifiers described above. We compute classifications using both algorithms and then combine the results with a weighting factor we determined experimentally. Each hashtag recommended by Naive Bayes recieves a weight of 0.4 and each hashtag recommended by K Nearest Neighbor receives a weight of 0.6. We sum up the weights and return the hashtags with the highest scores. Using both classifiers takes advantage of the strengths of each classifier which causes the most likely hashtags to move to the top of the list of recommendations.
\subsection{Computational Resource Management}
\noindent
Finally, we need to consider speed and computational resources. Tweets are produced in high volume, 24 hours a day, so being able to produce recommendations quickly with constant computational resources is extremely important. Traditionally, classification algorithms are given a training set of data up front to train the classifier and all future classifications are computed using that initial model. This method doesn't work for a stream of Tweets because hashtag usage changes very rapidly on Twitter as topics trend and fade away. Because of this it's very important to keep the classifier's model as current as possible.
\\
\noindent
We modified the classifiers to use the "sliding window" pattern which ensures the most recently seen Tweets are used in the model [4]. In our implementation, the classifier data model is a FIFO (first in first out) queue. First, the classifier is initialized with a predefined number of Tweets. Following initialization, every time a new Tweet arrives from the stream, the oldest Tweet is removed from the model and the new Tweet is added. Because we chose to use Naive Bayes and K Nearest Neighbor, which are lazy evaluating classifiers, adding and removing Tweets to and from the dataset is very fast because we do not need to regenerate any structures built on top of the model like a rule set or a decision tree.
\\
\noindent
However, because these are lazy evaluators, classification itself can take longer than algorithms that use complex prebuilt structures. Because of this, we do maintain some external structures that are easy to update when Tweets are added and removed from the model. Besides the ordered list of Tweets in the model itself, we keep two hash maps, one indexed by a word and the other indexed by a hashtag. Each of these hash maps store basic statistics about the word or hashtag which we use to compute the weights and probabilities needed in our classification algorithms. Because these are hash maps, we can find statistics about a given word or hashtag in constant time. This provides a nice compromise between lazy and eagar classifiers which allows for efficient updates to our data model as it changes but also supports fast computation of classifications.
\\
\noindent
In order to further improve the speed of our recommendations, we parallelized our classifiers. We observed significant speed improvements when running our recommendations with multiple threads and did not observe any decrease in recommendation accuracy. This is an important finding because it shows our classification can scale to handle larger data volumes. We did not extend this parallelization to multiple machines so this would be an interesting area to investigate further.
\\
\noindent
The final area we investigated to speed up classification was maintaining a seperate model for each language. This provides a great opportunity to parallelize the recommendation system even further. Each language could have it's own classifier and model which would function independently, and it would only need to process a fraction of the full volume of Tweets. For example, when an English Tweet is received it could be routed to the machine, or cluster of machines, that maintains the English model and performs English classifications. This system could be duplicated for each language. We proved out this architecture on a single machine and achieved classifications of comparable accuracy but found the language detection itself to be such a bottleneck that it drastically slowed down recommendations. It would be interesting to investigate this further to see if a faster language detection algorithm could be designed to take advantage of this further parallelization.
\section{Experimental Results}
\noindent
To evaluate our recommendation algorithm, we simulate how a system like this would function in a real world application. We begin by initializing the classifier by fully populating the sliding window dataset with Tweets with known hashtags. When initialization completes, we start producing recommendations for Tweets with known hashtags. For each Tweet, we recommend up to three hashtags. If one of those hashtags matches the actual hashtag of the Tweet, we consider this recommendation a success. Following each recommendation, we add that Tweet to the sliding window dataset and removed the oldest Tweet which maintains our sliding window.
\\
\noindent
We ran our evaluation with a sliding window size of 1 million Tweets and produced recommendations for an additional 1 million Tweets. For each Tweet, we chose to recommend up to three hashtags as this is a realistic number that could be shown to a user in a real application. The dataset we used for testing contained a random sampling of 35,575,057 Tweets collected from 4/14/13 to 4/21/13. Of that, 4,038,934 Tweets (11.3\%), contained hashtags.
\\
\begin{figure}[ht!]
\centering
\includegraphics[width=90mm]{chart.png}
\caption{Successful classification percentages for each classifier}
\label{success_chart}
\end{figure}
\noindent
The optimal Naive Bayes recommender successfully recommended hashtags for 30.22\% Tweets and the optimal K Nearest Neighbor recommender successfully recommended hashtags for 31.4\% of Tweets. When we combined these together in our Hybrid recommender we were able to achieve a successful recommendations for 32.71\% of the test Tweets. Figure 1 shows the success percentages for each classifier. We observed a precision of 0.2, recall of 0.27, and an F measure of 0.23.
\\
\noindent
This evaluation method is very conservative and underestimates the actual success rate of our recommendations for a few key reasons. First, each recommendation is considered successful only if the actual hashtag exactly matches one of the recommendations. In reality, most Tweets don't have one single correct hashtag but could have numerous hashtags that would make sense to a user and would be considered good recommendations by a human. However, due to resource and time constaints, it was not possible to have a real person evaluate the millions of recommendations made by our classifiers during these evaluations.
\\
\noindent
The second reason is that we exclude the known hashtag of the Tweet that we are making recommendations for. Including the hashtag unfairly biases the recommendation in cases where the hashtag is truely a tag. For example, consider a Tweet like "Just heard a funny joke \#laughoutloud". In this case, the user would never have included "\#laughoutloud" if they did not know about that hashtag. Because of this, we removed hashtags from the Tweet as we did not want to give our recommender an unfair advantage. However, this negatively impacts the recommendations in other cases where the hashtag is part of the content of the Tweet. For example, "Just heard a funny \#joke". In this case, the user would have likely typed in the word "joke" even if they did not know to use it as a hashtag. Because we are excluding hashtags, we would not use the word joke when producing recommendations for this Tweet so we are making this recommendation without an important word, making it far less likely for us to produce an accurate recommendation. Including the hashtag when making our recommendations nearly doubled our success rate.
\\
\noindent
The Hybrid recommender processed the 1 million test Tweets in just under 29 minutes at an average rate of 576 Tweets / Second. This equates to approximately 50 million Tweets per day. These tests were run on a standard laptop. In a real production application, much more powerful hardware would be available that could fully take advanatage of the parallelization capabilities discussed previously so we would expect even better performance.
\section{Conclusions}
\noindent
In this paper we discussed the challenges that come with Twitter hashtag recommendation and proposed solutions to overcome them. We combined these ideas to create a scalable recommender system based on Naive Bayes and K Nearest Neighbor classifiers. We used the sliding window pattern to maintain constant computational resource demands and used IDF to efficiently filter down the dataset to allow recommendations to be produced rapidly. To improve recommendation accuracy, we combined classifiers and used a variety of weighting measures to ensure these classifiers consider only the most significant words in each Tweet. Throughout the paper, we also proposed a few key areas to further extend our work to build upon the speed and accuracy we were able to achieve with this recommender system.
\\
\newpage
|
\section{Introduction: a Historical Tour D'Horizon}
In the last decades, deformation quantization evolved into a widely
accepted quantization scheme which, on one hand, provides deep
conceptual insights into the question of quantization and, on the
other hand, proved to be a reliably technique leading to explicit
understanding of many examples. It will be the aim of this review to
give some overview on the developments of deformation quantization
starting from the beginnings but also including some more recent
ideas.
The original formulations of deformation quantization by Bayen
et. al. aimed mainly at finite-dimensional classical mechanical
systems described by symplectic or Poisson manifolds
\cite{bayen.et.al:1978a} and axiomatize the heuristic quantization
formulas found earlier by Weyl, Groenewold, and Moyal \cite{weyl:1931a,
groenewold:1946a, moyal:1949a}. Berezin considered the more
particular case of bounded domains and Kähler manifolds
\cite{berezin:1975c, berezin:1975b, berezin:1975a}. Shortly after it
proved to be a valuable tool to approach also problems in quantum
field theories, see e.g. the early works of Dito \cite{dito:1990a,
dito:1992a, dito:1993a}.
Meanwhile, the question of existence and classification of deformation
quantizations, i.e. of star products, on symplectic manifolds was
settled: first DeWilde and Lecomte showed the existence of star
products on symplectic manifolds \cite{dewilde.lecomte:1983b} in 1983
after more particular classes \cite{dewilde.lecomte:1983a,
dewilde.lecomte:1983c} were considered. Remarkably, also in 1983 the
first genuine class of Poisson structures was shown to admit star
products, the linear Poisson structures on the dual of a Lie algebra,
by Gutt \cite{gutt:1983a} and Drinfel'd \cite{drinfeld:1983a}. In 1986
Fedosov gave a very explicit and constructive way to obtain star
products on a symplectic manifold by means of a symplectic connection
\cite{fedosov:1986a}, see also \cite{fedosov:1996a, fedosov:1994a} for
a more detailed version. His construction is still one of the
cornerstones in deformation quantization as it provides not only a
particularly nice construction allowing to adjust many special
features of star products depending on the underlying manifold like
e.g. separation of variables (Wick type) on Kähler manifolds
\cite{karabegov:2000a, karabegov:1996a, bordemann.waldmann:1997a,
neumaier:2003a} or star products on cotangent bundles
\cite{bordemann.neumaier.waldmann:1998a,
bordemann.neumaier.waldmann:1999a,
bordemann.neumaier.pflaum.waldmann:2003a}. Even beyond the
symplectic world, Fedosov's construction was used to globalize the
existence proofs of star products on Poisson manifolds
\cite{cattaneo.felder.tomassini:2002b, dolgushev:2005a}. Yet another
proof for the symplectic case was given
in\cite{omori.maeda.yoshioka:1991a}.
Even though the symplectic case was understood well, the question of
existence on Poisson manifolds kept its secrets till the advent of
Kontsevich's formality theorem, solving his formality conjecture
\cite{kontsevich:1997a, kontsevich:1997:pre, kontsevich:2003a}. To
give a adequate overview on Kontsevich's formality theorem would
clearly go beyond the scope of this short review. Here one can rely on
various other publications like e.g. \cite{cattaneo:2005a,
esposito:2015a}. In a nutshell, the formality theorem proves a very
general fact about smooth functions on a manifold from which it
follows that every (formal series of) Poisson structures can be
quantized into a star product, including a classification of star
products. Parallel to Kontsevich's groundbreaking result, the
classification of star products on symplectic manifolds was achieved
and compared by several groups \cite{bertelson.cahen.gutt:1997a,
nest.tsygan:1995a, deligne:1995a, gutt.rawnsley:1999a,
neumaier:2002a}. Shortly after Kontsevich, Tamarkin gave yet another
approach to the quantization problem on Poisson manifolds
\cite{tamarkin:1998b}, see also \cite{kontsevich:1999a,
kontsevich.soibelman:2000a}, based on the language of operads and
the usage of Drinfel'd associators. Starting with these formulations,
formality theory has evolved and entered large areas of contemporary
mathematics, see e.g. \cite{alekseev.meinrenken:2006a,
alekseev.torossain:2010a, alekseev.torossian:2012a,
dolgushev.tamarkin.tsygan:2007a,dolgushev.rubtsov:2009a,
kontsevich:2001a, kontsevich:1999a} to name just a few.
While deformation quantization undoubtedly gave many important
contribution to pure mathematics over the last decades, it is now
increasingly used in contemporary quantum physics as well: perhaps
starting with the works of Dütsch and Fredenhagen on the perturbative
formulations of algebraic quantum field theory
\cite{duetsch.fredenhagen:2001b, duetsch.fredenhagen:1999a,
duetsch.fredenhagen:2001a} it became clear that star products
provide the right tool to formulate quantum field theories in a
semiclassical way, i.e. as formal power series in $\hbar$. Now this
has been done in increasing generalities for various scenarios
including field theories on general globally hyperbolic spacetimes,
see e.g. \cite{brunetti.fredenhagen.verch:2003a,
brunetti.fredenhagen.ribeiro:2012a:pre, baer.ginoux.pfaeffle:2007a,
hollands.wald:2010a}.
Of course, from a physical point of view, deformation quantization can
not yet be the final answer as one always deals with formal power
series in the deformation parameter $\hbar$. A physically reasonable
quantum theory, however, requires of course convergence. Again, in the
very early works \cite{bayen.et.al:1978a} some special cases were
treated, namely the Weyl-Moyal product for which an integral formula
exists which allows for a reasonable analysis based on the Schwartz
space. The aims here are at least two-fold. On one hand one wants to
establish a reasonable spectral calculus for particular elements in
the star product algebra which allows to compute spectra in a
physically sensitive way. This can be done with the star exponential
formalism, which works in particular examples but lacks a general
framework. On the other hand, one can try to establish from the formal
star product a convergent version such that in the end one obtains a
$C^*$-algebra of quantum observables being a deformation, now in a
continuous way, of the classical functions on the phase space. This is
the point of view taken by strict deformation quantization, most
notably advocated by Rieffel \cite{rieffel:1989a, rieffel:1993a} and
Landsman \cite{landsman:1998a}, see also
\cite{bordemann.meinrenken.schlichenmaier:1991a,
cahen.gutt.rawnsley:1995a, cahen.gutt.rawnsley:1994a,
cahen.gutt.rawnsley:1993a, cahen.gutt.rawnsley:1990a} for the
particular case of quantizable Kähler manifolds and
\cite{natsume.nest.peter:2003a, natsume.nest:1999a, natsume:2000a} for
more general symplectic manifolds. Bieliavsky and coworkers found a
generalization of Rieffel's approach by passing from actions of the
abelian group $\mathbb{R}^d$ to more general Lie group actions
\cite{bieliavsky.detournay.spindel:2009a, bieliavsky:2002a,
bieliavsky.massar:2001b}. Having a $C^*$-algebra one has then the
full power of $C^*$-algebra techniques at hands which easily allows to
get a reasonable spectral calculus. However, constructing
$C^*$-algebraic quantizations is still very much in development: here
one has not yet a clear picture on the existence and classification of
the quantizations. In fact, one even has several competing definitions
of what one is looking for. It is one of the ongoing research projects
by several groups to understand the transition between formal and
strict quantizations in more detail.
Needless to say, that in the above historical survey we can barely
scratch on the surface of this vast topic: many aspects have not been
mentioned like the role played by symmetries and reduction, the
applications to concrete physical systems, various generalizations of
deformation quantization to other geometric brackets, alternative
approaches to various index theorems, relations to noncommutative
geometry, and many more. In the remaining part of this review we will
focus on two aspects of the theory: first, we discuss the role of
classification results beyond the notion of equivalence,
i.e. isomorphism. Here we are particularly interested in the
classification of star products up to Morita equivalence. Second, we
give a short outlook on star products in infinite dimensions and
problems arising there by investigating one particular example: the
Weyl algebra of a vector space with a (quite arbitrary) bilinear
form. Beside the purely algebraic construction we obtain a locally
convex algebraic deformation once we start in this category.
\section{From Poisson Manifolds to Star Products}
In this section we give a more detailed but still non-technical
motivation of the definition of star products and list some first
examples.
The set-up will be a finite-dimensional phase space which we model by
a symplectic or, more generally, a Poisson manifold $(M, \pi)$ where
$\pi \in \Secinfty(\Anti^2 TM)$ is a bivector field satisfying
\begin{equation}
\label{eq:Jacobi}
\Schouten{\pi, \pi} = 0.
\end{equation}
Here $\Schouten{\argument, \argument}$ is the Schouten bracket and the
condition is equivalent to the Jacobi identity for the Poisson bracket
\begin{equation}
\label{eq:PoissonBracket}
\{f, g\} = - \Schouten{\Schouten{f, \pi}, g} = \pi(\D f, \D g)
\end{equation}
determined by $\pi$ for functions $f, g \in \Cinfty(M)$. One can then
formulate classical Hamiltonian mechanics using $\pi$ and
$\{\argument, \argument\}$. For a gentle introduction to Poisson
geometry see \cite{waldmann:2007a} as well as \cite{vaisman:1994a,
dufour.zung:2005a, laurent-gengoux.pichereau.vanhaecke:2013a,
cannasdasilva.weinstein:1999a}. One has several important examples
of Poisson manifolds:
\begin{itemize}
\item Every symplectic manifold $(M, \omega)$ where $\omega \in
\Secinfty(\Anti^2T^*M)$ is a closed non-degenerate two-form, is a
Poisson manifold with $\pi = \omega^{-1}$. The Jacobi identity
\eqref{eq:Jacobi} corresponds then directly to $\D\omega = 0$.
\item Every cotangent bundle $T^*Q$ is a symplectic manifold in a
canonical way with an exact symplectic form $\omega = \D \theta$
where $\theta \in \Secinfty(T^*(T^*Q))$ is the canonical (or
tautological) one-form on $T^*Q$. This is the arena of geometric
mechanics.
\item Kähler manifolds are particularly nice examples of symplectic
manifolds as they possess a compatible Riemannian metric and a
compatible complex structure.
\item The dual $\lie{g}^*$ of a Lie algebra $\lie{g}$ is always a
Poisson manifold with a linear Poisson structure: the coefficient
functions of the tensor field $\pi$ are linear functions on
$\lie{g}^*$, explicitly given by
\begin{equation}
\label{eq:KKS}
\{f, g\}(x)
=
x_i c^i_{k\ell}
\frac{\partial f}{\partial x_k}
\frac{\partial g}{\partial x_\ell},
\end{equation}
where $x_1, \ldots, x_n$ are the linear coordinates on $\lie{g}^*$
and $c^i_{k\ell}$ are the corresponding structure constants of
$\lie{g}$. Since it vanishes at the origin, this is never
symplectic.
\item Remarkably and slightly less trivial is the observation that on
every manifold $M$ and every $p \in M$ there is a Poisson
structure $\pi$ with compact support but $\pi\at{p}$ has maximal
rank.
\end{itemize}
To motivate the definition of a star product we consider the most easy
example of the classical phase space $\mathbb{R}^2$ with canonical
coordinates $(q, p)$. Canonical quantization says that we have to map
the spacial coordinate $q$ to the position operator $Q$ acting on a
suitable domain in $L^2(\mathbb{R}, \D x)$ as multiplication
operator. Moreover, we have to assign the momentum coordinate $p$ to
the momentum operator $P = - \I \hbar \frac{\partial}{\partial q}$,
again defined on a suitable domain. Since we want to ignore
functional-analytic questions at the moment, we simply chose
$\Cinfty_0(\mathbb{R})$ as common domain for both operators. In a next
step we want to quantize polynomials in $q$ and $p$ as well. Here we
face the ordering problem as $pq = qp$ but $PQ \ne QP$. One simple
choice is the \emph{standard ordering}
\begin{equation}
\label{eq:StandardOrd}
q^np^m \; \mapsto \;
\stdrep(q^np^m) = Q^nP^m
=
(-\I\hbar)^m q^n\frac{\partial^m}{\partial q^m}
\end{equation}
for monomials and its linear extension to all polynomials. More
explicitly, this gives
\begin{equation}
\label{eq:StdRep}
\stdrep(f) = \sum_{r=0}^\infty \frac{1}{r!}
\left(\frac{\hbar}{\I}\right)^r
\frac{\partial^r f}{\partial p^r}\At{p=0}
\frac{\partial^r}{\partial q^r}.
\end{equation}
Now this formula still makes sense for smooth functions $f$ which are
polynomial in $p$, i.e. for $f \in \Cinfty(\mathbb{R})[p]$. The main
idea of deformation quantization is now to pull-back the operator
product: this is possible since the image of $\stdrep$ is the space of
all differential operators with smooth coefficients which therefore is
a (noncommutative) algebra. We define the \emph{standard-ordered star
product} by
\begin{equation}
\label{eq:starstd}
f \starstd g
= \stdrep^{-1}(\stdrep(f)\stdrep(g))
= \sum_{r=0}^\infty \frac{1}{r!}
\left(\frac{\hbar}{\I}\right)^r
\frac{\partial^r f}{\partial p^r}
\frac{\partial^r g}{\partial q^r}
\end{equation}
for $f, g \in \Cinfty(\mathbb{R})[p]$. While it is clear that
$\starstd$ is an associative product the behaviour with respect to the
complex conjugation is bad: we do not get a $^*$-involution,
$\cc{f \starstd g} \ne \cc{g} \starstd \cc{f}$, since
\begin{equation}
\label{eq:NeumaierOp}
\stdrep(f)^\dag = \stdrep(N^2 f)
\quad
\textrm{with}
\quad
N
=
\exp\left(
\frac{\hbar}{2\I} \frac{\partial^2}{\partial q \partial p}
\right),
\end{equation}
as a simple integration by parts shows. We can repair this unpleasant
feature by defining the Weyl ordering and the Weyl product by
\begin{equation}
\label{eq:WeylStuff}
\weylrep(f) = \stdrep(Nf)
\quad
\textrm{and}
\quad
f \starweyl g = N^{-1}(Nf \starstd Ng).
\end{equation}
Note that $N$ is indeed an invertible operator on
$\Cinfty(\mathbb{R})[p]$. Again, $\starweyl$ is associative. Then we
get
\begin{equation}
\label{eq:WeylFeatures}
\cc{f \starweyl g} = \cc{g} \starweyl \cc{f}
\quad
\textrm{and}
\quad
\weylrep(f \starweyl g) = \weylrep(f) \weylrep(g).
\end{equation}
For both products we can collect the terms of order $\hbar^r$ which
gives
\begin{equation}
\label{eq:StarProd}
f \star g = \sum_{r=0}^\infty \hbar^r C_r(f, g)
\end{equation}
with bidifferential operators $C_r$ of order $r$ in each argument. The
explicit formula for $\starweyl$ is slightly more complicated than the
one for $\starstd$ in \eqref{eq:starstd} but still easy to compute. We
have
\begin{equation}
\label{eq:StarFirstOrders}
f \star g = fg + \cdots
\quad
\textrm{and}
\quad
f \star g - g \star f = \I\hbar \{f, g\} + \cdots,
\end{equation}
where $+ \cdots$ means higher orders in $\hbar$. Also $f \star 1 = f =
1 \star f$. Note that the seemingly infinite series in
\eqref{eq:StarProd} is always finite as long as we take functions in
$\Cinfty(\mathbb{R})[p]$.
The idea is now to axiomatize these features for $\star$ in such a way
that it makes sense to speak of a star product on a general Poisson
manifold. The first obstacle is that on a generic manifold $M$ there
is nothing like functions which are polynomial in certain
coordinates. This is a chart-dependent characterization which one does
not want to use. But then already for $\starweyl$ and $\starstd$ one
encounters the problem that for general $f, g \in
\Cinfty(\mathbb{R}^2)$ the formulas \eqref{eq:starstd} and
\eqref{eq:WeylFeatures} will not make any sense: the series are indeed
infinite and since we can adjust the Taylor coefficients of a smooth
function in a rather nasty way, there is no hope for convergence.
The way out is to consider \emph{formal} star product in a first step,
i.e. formal power series in $\hbar$. This yields the definition of
star products \cite{bayen.et.al:1978a}:
\begin{defn}
A formal star product $\star$ on a Poisson manifold $(M, \pi)$ is
an associative $\mathbb{C}[[\hbar]]$-bilinear associative product
for $\Cinfty(M)[[\hbar]]$ such that
\begin{equation}
\label{eq:FormalStarProd}
f \star g = \sum_{r=0}^\infty \hbar^r C_r(f, g)
\end{equation}
with
\begin{enumerate}
\item $C_0(f, g) = fg$,
\item $C_1(f, g) - C_1(g, f) = \I\{f, g\}$,
\item $C_r(1, f) = 0 = C_r(f, 1)$ for $r \ge 1$,
\item $C_r$ is a bidifferential operator.
\end{enumerate}
\end{defn}
Already in the trivial example above we have see that there might be
more than one star product. The operator $N$ interpolates between them
and is invisible in classical physics: for $\hbar = 0$ the operator
$N$ becomes the identity. As a formal series of differential operator
starting with the identity it is invertible and implements an algebra
isomorphism. This is now taken as definition for equivalence of star
products: given two star products $\star$ and $\star'$ on a manifold,
a formal power series $T = \id + \sum_{r=1}^\infty \hbar^r T_r$ of
differential operators $T_r$ with $T1 = 1$ is called an equivalence
between $\star$ and $\star'$ if one has
\begin{equation}
\label{eq:Equivalence}
f \star' g = T^{-1}(Tf \star Tg).
\end{equation}
Note that $T$ is indeed invertible as a formal power series. Hence
this is an equivalence relation. Conversely, given such a $T$ and
$\star$ we get a new star product $\star'$ by \eqref{eq:Equivalence}.
After the general set-up we are now in the position to list some basic
examples of star products:
\begin{itemize}
\item The explicit formulas for $\starstd$ and $\starweyl$ immediately
generalize to higher dimensions yielding equivalent star products
on $\mathbb{R}^{2n}$ and hence also on every open subset of
$\mathbb{R}^{2n}$. Since by the Darboux Theorem every symplectic
manifold looks like an open subset of $\mathbb{R}^{2n}$
\emph{locally}, the question of existence of star products on
symplectic manifolds is a global problem.
\item For the linear Poisson structure \eqref{eq:KKS} on the dual
$\lie{g}^*$ of a Lie algebra $\lie{g}$ one gets a star product as
follows \cite{gutt:1983a}: First, we note that
$\Sym^\bullet(\lie{g}) = \Pol^\bullet(\lie{g}^*)$. Then the
PBW isomorphism
\begin{equation}
\label{eq:Symmetrizer}
\Sym^\bullet(\lie{g}) \ni \xi_1 \vee \cdots \vee \xi_k
\; \mapsto \;
\frac{(\I\hbar)^k}{k!} \sum_{\sigma \in S_k}
\xi_{\sigma(1)} \cdots \xi_{\sigma(k)}
\in \mathcal{U}(\lie{g})
\end{equation}
from the symmetric algebra over $\lie{g}$ into the universal
enveloping algebra allows to pull the product of
$\mathcal{U}(\lie{g})$ back to $\Sym^\bullet(\lie{g})$ and hence
to polynomials on $\lie{g}^*$. One can now show that after
interpreting $\hbar$ as a formal parameter one obtains indeed a
star product quantizing the linear Poisson bracket. This star
product is completely characterized by the feature that
\begin{equation}
\label{eq:BCH}
\exp(\hbar\xi) \star \exp(\hbar\eta)
=
\exp(\mathrm{BCH}(\hbar\xi, \hbar\eta))
\end{equation}
for $\xi, \eta \in \lie{g}$ with the Baker-Campbell-Hausdorff
series $\mathrm{BCH}$, see \cite{gutt:1983a,
bordemann.neumaier.waldmann:1998a}.
\item The next interesting example is perhaps the complex projective
space $\mathbb{CP}^n$ and its non-compact dual, the Poincaré disc
$\mathbb{D}_n$ with their canonical Kähler structures of constant
holomorphic sectional curvature. For these, star products were
considered by Moreno and Ortega-Navarro
\cite{moreno.ortega-navarro:1983b} who gave recursive formulas
using local coordinates. Cahen, Gutt, and Rawnsley
\cite{cahen.gutt.rawnsley:1995a, cahen.gutt.rawnsley:1994a,
cahen.gutt.rawnsley:1993a, cahen.gutt.rawnsley:1990a} discussed
this in their series of papers of quantization of Kähler manifolds
as one of the examples. The first explicit (non-recursive) formula
was found in \cite{bordemann.brischle.emmrich.waldmann:1996a,
bordemann.brischle.emmrich.waldmann:1996b} by a quantization of
phase space reduction and extended to complex Grassmannians in
\cite{schirmer:1997a:pre}. Ever since these star products have
been re-discovered by various authors.
\end{itemize}
We briefly comment on the general existence results: as already
mentioned, the symplectic case was settled in the early 80s. The
Poisson case follows from Kontsevich's formality theorem.
\begin{thm}[Kontsevich]
On every Poisson manifold there exist star products.
\end{thm}
The classification is slightly more difficult to describe: we consider
\emph{formal Poisson structures}
\begin{equation}
\label{eq:FormalPoisson}
\pi = \hbar \pi_1 + \hbar^2 \pi_2 + \cdots
\in \hbar \Secinfty(\Anti^2 TM)[[\hbar]]
\quad
\textrm{with}
\quad
\Schouten{\pi, \pi} = 0.
\end{equation}
Moreover, let $X = \hbar X_1 + \hbar^2 X_2 + \cdots \in
\hbar\Secinfty(TM)[[\hbar]]$ be a formal vector field, starting in
first order of $\hbar$. Then one calls $\exp(\Lie_X)$ a formal
diffeomorphism which defines an action
\begin{equation}
\label{eq:expLX}
\exp(\Lie_X)\colon
\Secinfty(\Anti^2 TM)[[\hbar]] \ni \nu
\; \mapsto \;
\nu + \Lie_X\nu + \frac{1}{2} \Lie_X^2 \nu + \cdots
\in
\Secinfty(\Anti^2 TM)[[\hbar]].
\end{equation}
Via the Baker-Campbell-Hausdorff series, the set of formal
diffeomorphisms becomes a group and \eqref{eq:expLX} is a group
action. Since $\Lie_X$ is a derivation of the Schouten bracket, it
follows that the action of $\exp(\Lie_X)$ preserves formal Poisson
structures. The space of orbits of formal Poisson structures modulo
this group action gives now the classification:
\begin{thm}[Kontsevich]
The set of equivalence classes of formal star products is in
bijection to the set of equivalence classes of formal Poisson
structures modulo formal diffeomorphisms.
\end{thm}
In general, both moduli spaces are extremely difficult to describe.
However, if the first order term $\pi_1$ in $\pi$ is symplectic, then
we have a much easier description which is in fact entirely
topological:
\begin{thm}[Bertelson, Cahen, Gutt, Nest, Tsygan, Deligne, \ldots]
On a symplectic manifold $(M, \omega)$ the equivalence classes of
star products are in bijection to the formal series in the second
deRham cohomology. In fact, one has a canonical surjective map
\begin{equation}
\label{eq:CharClass}
c\colon
\star \; \mapsto \;
c(\star) \in
\frac{[\omega]}{\I\hbar} +
\mathrm{H}^2_{\mathrm{dR}}(M, \mathbb{C})[[\hbar]]
\end{equation}
such that $\star$ and $\star'$ are equivalent iff $c(\star) =
c(\star')$.
\end{thm}
This map is now called the characteristic class of the symplectic star
product. In a sense which can be made very precise
\cite{bursztyn.dolgushev.waldmann:2012a}, the inverse of $c(\star)$
corresponds to Kontsevich's classification by formal Poisson tensors.
\section{Morita Classification}
\label{sec:Morita}
We come now to some more particular topics in deformation
quantization. In this section we discuss a coarser classification
result than the above classification up to equivalence.
The physical motivation to look for Morita theory is rather simple and
obvious: in quantum theory we can not solely rely on the observable
algebra as the only object of interest. Instead, we also need to have
a reasonable notion of states. While for $C^*$-algebras there is a
simple definition of a state as a normalized positive functional, in
deformation quantization we do not have $C^*$-algebras in a first
step. Surprisingly, the notion of positive functionals still makes
sense if interpreted in the sense of the ring-ordering of
$\mathbb{R}[[\hbar]]$ and it produces a physically reasonable
definition, see \cite{bordemann.waldmann:1998a}. However, the
requirements from quantum theory do not stop here: we also need a
super-position principle for states. Since positive functionals can
only be added convexly, we need to realize the positive functionals as
expectation value functionals for a $^*$-representation of the
observable algebra on some (pre) Hilbert space. Then we can take
complex linear combination of the corresponding vectors to implement
the super-position principle. This leads to the need to understand the
representation theory of the star product algebras, a program which
was investigated in great detail \cite{bursztyn.waldmann:2001b,
bursztyn.waldmann:2001a, bursztyn.waldmann:2004a,
bursztyn.waldmann:2005b, bursztyn.dolgushev.waldmann:2012a,
jansen.waldmann:2006a}, see also \cite{waldmann:2005b} for a review.
The main point is that replacing the ring of scalars from $\mathbb{R}$
to $\mathbb{R}[[\hbar]]$ and thus from $\mathbb{C}$ to
$\mathbb{C}[[\hbar]]$ works surprisingly well as long as we do not try
to implement analytic concepts: the non-archimedean order of
$\mathbb{R}[[\hbar]]$ forbids a reasonable analysis. However, the
concept of positivity is entirely algebraic and hence can be used and
employed in this framework as well.
In fact, one needs not to stop here: \emph{any} ordered ring
$\ring{R}$ instead of $\mathbb{R}$ will do the job and one can study
$^*$-algebras over $\ring{C} = \ring{R}(\I)$ and their
$^*$-representation theory on pre Hilbert modules over $\ring{C}$.
For many reasons it will also be advantageous to consider
representation spaces where the inner product is not taking values in
the scalars but in some \emph{auxiliary} $^*$-algebra $\mathcal{D}$.
\begin{ex}
Let $E \longrightarrow M$ be a complex vector bundle over a smooth
manifold $M$. Then $\Secinfty(E)$ is a $\Cinfty(M)$-module in the
usual way. A Hermitian fiber metric $h$ give now a sesquilinear
map
\begin{equation}
\label{eq:FiberMetric}
\SP{\argument, \argument}\colon
\Secinfty(E) \times \Secinfty(E) \longrightarrow \Cinfty(M)
\end{equation}
which is also $\Cinfty(M)$-linear in the second argument, i.e. we
have $\SP{s, tf} = \SP{s, t}f$ for all $s, t \in \Secinfty(E)$ and
$f \in \Cinfty(M)$. Moreover, the pointwise positivity of $h_p$ on
$E_p$ implies that the map
\begin{equation}
\label{eq:CompletePositive}
\SP{\argument, \argument}^{(n)}\colon
\Secinfty(E)^n \times \Secinfty(E)^n
\longrightarrow M_n(\Cinfty(M)) = \Cinfty(M, M_n(\mathbb{C}))
\end{equation}
is positive for all $n$ in the sense that the matrix-valued
function $\SP{S, S}^{(n)} \in \Cinfty(M, M_n(\mathbb{C}))$ yields
a positive matrix at all points of $M$ for all $S = (s_1, \ldots,
s_n) \in \Secinfty(E)^n$.
\end{ex}
Using this kind of \emph{complete positivity} for an inner product
yields the definition of a pre Hilbert right module over a
$^*$-algebra $\mathcal{D}$, where the inner product takes values in
$\mathcal{D}$. Then again, we can formulate what are
$^*$-representations of a $^*$-algebra $\mathcal{A}$ on such a pre
Hilbert right module over $\mathcal{D}$. Without further difficulties
this gives various categories of $^*$-representations of $^*$-algebras
on inner product modules or pre Hilbert modules over auxiliary
$^*$-algebras.
Having now a good notion of $^*$-representations of $^*$-algebras it
is a major task to understand the resulting categories for those
$^*$-algebras occurring in deformation quantization. Now from
$C^*$-algebra theory we anticipate that already with the full power of
functional-analytic techniques it will in general be impossible to
``understand'' the category of $^*$-representations completely, beside
rather trivial examples. The reason is that there will simply be too
many inequivalent such $^*$-representations and a decomposition theory
into irreducible ones is typically an extremely hard problem. In a
purely algebraic situation like for formal star product algebras,
things are even worse: here we expect even more inequivalent ones
which are just artifacts of the algebraic formulation. There are many
examples of inequivalent $^*$-representations which, after one
implements mild notions of convergence and hence of analytic aspects,
become equivalent. From a physical point of view such inequivalences
would then be negligible. However, it seems to be quite difficult to
decide this \emph{before} convergence is implemented, i.e. on the
algebraic side.
Is the whole program now useless, hopeless? The surprising news is
that one can indeed say something non-trivial about the
$^*$-representation theories of the star product algebras from
deformation quantization, and for $^*$-algebras in general. The idea
is that even if the $^*$-representation theory of a given $^*$-algebra
is horribly complicated and contains maybe unwanted
$^*$-representation, we can still \emph{compare} the whole
$^*$-representation theory of one $^*$-algebra to another $^*$-algebra
and ask whether they are equivalent as categories.
This is now the basic task of Morita theory. To get a first impression
we neglect the additional structure of ordered rings,
$^*$-involutions, and positivity and consider just associative
algebras over a common ring of scalars. For two such algebras
$\algebra{A}$ and $\algebra{B}$ we want to know whether their
categories of left modules are equivalent categories. Now there might
be many very strange functors implementing an equivalence and hence
one requires them to be compatible with direct sums of modules, which
is clearly a reasonable assumption. The prototype of such a functor is
then given by the tensor product with a $(\algebra{B},
\algebra{A})$-bimodule. Since the tensor product with $\algebra{A}$
itself is (for unital algebras) naturally isomorphic to the identity
functor and since the tensor product of bimodules is associative up to
a natural isomorphism, the question of equivalence of categories via
such tensor product functors becomes equivalent to the question of
\emph{invertible bimodules}: Here a $(\algebra{B},
\algebra{A})$-bimodule $\BEA$ is called invertible if there is a
$(\algebra{A}, \algebra{B})$-bimodule $\AEpB$ such that the tensor
product $\BEA \tensor[\algebra{A}] \AEpB$ is isomorphic to
$\algebra{B}$ and $\AEpB \tensor[\algebra{B}] \BEA$ is isomorphic to
$\algebra{B}$, always as bimodules.
The classical theorem of Morita now gives a complete and fairly easy
description of the possible bimodules with this property: $\BEA$ has
to be a finitely generated projective and full right
$\algebra{A}$-module and $\algebra{B}$ is isomorphic to
$\End_{\algebra{A}}(\EA)$ via the left module structure, see
e.g. \cite{lam:1999a}.
Now the question is how such bimodules look like for star product
algebras. Classically, the finitely generated projective modules over
$\Cinfty(M)$ are, up to isomorphism, just sections $\Secinfty(E)$ of a
vector bundle $E \longrightarrow M$, this is the famous Serre-Swan
theorem in its incarnation for differential geometry. As soon as the
fiber dimension is non-zero, the fullness condition is trivially
satisfied. Hence the only Morita equivalent algebras to $\Cinfty(M)$
are, again up to isomorphism, the sections $\Secinfty(\End(E))$ of
endomorphism bundles. The corresponding bimodule is then
$\Secinfty(E)$ on which both algebras act in the usual way. It
requires now a little argument to see that for star products, an
equivalence bimodule gives an equivalence bimodule in the classical
limit $\hbar = 0$, i.e. a vector bundle. Conversely, the sections of
every vector bundle can be deformed into a right module over the star
product algebra in a unique way up to isomorphism. Thus for star
products, we have to look for the corresponding module endomorphisms
of such deformed sections of vector bundles. Finally, in order to get
again a star product algebra, the endomorphisms of the deformed
sections have to be, in the classical limit, isomorphic to the
functions on a manifold again. This can only happen if the vector
bundle was actually a line bundle over the same manifold. Hence the
remaining task is to actually compute the star product of the algebra
acting from the left side when the star product for the algebra on the
right side is known. Here one has the following results:
\begin{thm}[Bursztyn, W. \cite{bursztyn.waldmann:2002a}]
Let $(M, \omega)$ and $(M', \omega')$ be a symplectic manifolds
and let $\star$, $\star'$ be two star products on $M$ and $M'$,
respectively. Then $\star$ and $\star'$ are Morita equivalent iff
there exists a symplectomorphism $\psi\colon M \longrightarrow M'$
such that
\begin{equation}
\label{eq:ClassesME}
\psi^*c(\star') - c(\star) \in 2\pi\I
\mathrm{H}^2_{\mathrm{dR}}(M, \mathbb{Z}).
\end{equation}
The difference of the above classes defines then a line bundle
which implements the Morita equivalence bimodule by deforming its
sections.
\end{thm}
This theorem has already an important physical interpretation: for
cotangent bundles $T^*Q$ the characteristic classes $c(\star)$ can be
interpreted as the classes of magnetic fields $B$ on the configuration
space $Q$. Then a quantization of a charged particle in the background
field of such a $B$ requires a star product with characteristic class
$c(\star)$. Compared to the trivial characteristic class, $c(\star) =
0$, the above theorem then tells that quantization with magnetic field
has the same representation theory iff the magnetic field satisfies
the integrality condition for a Dirac monopole. Thus we get a Morita
theoretic interpretation of the charge quantization for magnetic
monopoles which is now extremely robust against details of the
quantization procedure: the statement holds for all cotangent bundles
and for all equivalent star products with the given characteristic
class.
Also in the more general Poisson case the full classification is
known. Here the actual statement is slightly more technical as it
requires the Kontsevich class of the star products and a canonically
given action of the deRham cohomology on equivalence classes of formal
Poisson structures by gauge transformations. Then one obtains the
following statement, see also \cite{jurco.schupp.wess:2002a} for an
earlier heuristic argument based on noncommutative field theories:
\begin{thm}[Bursztyn, Dolgushev, W. \cite{bursztyn.dolgushev.waldmann:2012a}]
Star products on Poisson manifolds are Morita equivalent iff their
Kontsevich classes of formal Poisson tensors are gauge equivalent
by a $2\pi\I$-integral deRham class.
\end{thm}
\section{Beyond Formal Star Products}
\label{sec:Beyond}
Since formal star products are clearly not sufficient for physical
purposes, one has to go beyond formal power series. Here several
options are available: on one hand one can replace the formal series
in the star products by integral formulas. The formal series can then
be seen as the asymptotic expansions of the integral formulas in the
sense of Taylor series of smooth functions of $\hbar$, which are
typically not analytic: hence we can not expect convergence.
Nevertheless, the integral formulas allow for a good analytic
framework.
However, if one moves to field theories and hence to
infinite-dimensional systems, quantization becomes much more
complicated. Surprisingly, series formulas for star products can still
make sense in certain examples, quite unlike the integral formulas:
such integrals would consist of integrations over a
infinite-dimensional phase space. Hence we know that such things can
hardly exist in a mathematically sound way.
This motivates the second alternative, namely to investigate the
formal series in the star products directly without integral formulas
in the back. This might also be possible in infinite dimensions and
yields reasonable quantizations there. While this is a program far
from being understood, we can present here now one class of examples
with a particular physical relevance: the Weyl algebra.
Here we consider a real vector space $V$ with a bilinear map
$\Lambda\colon V \times V \longrightarrow \mathbb{C}$. Then we
consider the complexified symmetric algebra
$\Sym_{\mathbb{C}}^\bullet(V)$ of $V$ and interpret this as the
polynomials on the dual $V^*$. In finite dimensions this is correct,
in infinite dimensions the symmetric algebra is better to be
interpreted as the polynomials on the (not necessarily existing)
pre-dual. On $V^*$, there are simply much more polynomials than the
ones arising from $\Sym^\bullet_{\mathbb{C}}(V)$. Now we can extend
$\Lambda$ to a biderivation
\begin{equation}
\label{eq:PLambda}
P_\Lambda\colon
\Sym_{\mathbb{C}}^\bullet(V)
\tensor
\Sym_{\mathbb{C}}^\bullet(V)
\longrightarrow
\Sym_{\mathbb{C}}^\bullet(V)
\tensor
\Sym_{\mathbb{C}}^\bullet(V)
\end{equation}
in a unique way by enforcing the Leibniz rule in both tensor
factors. If we denote by $\mu\colon \Sym_{\mathbb{C}}^\bullet(V)
\tensor \Sym_{\mathbb{C}}^\bullet(V) \longrightarrow
\Sym_{\mathbb{C}}^\bullet(V)$ the symmetric tensor product, then
\begin{equation}
\label{eq:PoissonLambda}
\{a, b\}_\Lambda
=
\mu \circ (P_\Lambda (a \tensor b) - P_\Lambda(b \tensor a))
\end{equation}
is a Poisson bracket. In fact, this is the unique constant Poisson
bracket with the property that for linear elements $v, w \in V$ we
have $\{v, w\} = \Lambda(v, w) - \Lambda(w, v)$. Hence the
antisymmetric part of $\Lambda$ determines the bracket. However, we
will use the symmetric part for defining the star product. This will
allow to include also standard-orderings or other orderings like Wick
ordering from the beginning.
A star product quantizing this constant Poisson structure can then
be found easily. We set
\begin{equation}
\label{eq:Star}
a \star b = \mu \circ \exp(z P_\Lambda) (a \tensor b)
\end{equation}
where $z \in \mathbb{C}$ is the deformation parameter. For physical
applications we will have to set $z = \I\hbar$ later on. Note that
$\star$ is indeed well-defined since on elements in the symmetric
algebra, the operator $P_\Lambda$ lowers the degree by one in each
tensor factor.
In a next step we want to extend this product to more interesting
functions than the polynomial-like ones. The strategy is to look for a
topology which makes the product continuous and which allows for a
large completion of $\Sym^\bullet_{\mathbb{C}}(V)$. To start with, one
has to assume that $V$ is endowed with a topology itself. Hence let
$V$ be a locally convex Hausdorff space. In typical examples from
quantum mechanics, $V$ is the (dual of the) phase space and hence
finite dimensional, which makes the topology unique. In quantum field
theory, $V$ would be something like a test function space, i.e. either
the Schwartz space $\mathcal{S}(\mathbb{R}^d)$ or $\Cinfty_0(M)$ for a
manifold $M$, etc. In this case $V$ would be a Fréchet or LF space.
We use now the continuous seminorms of $V$ to extend them to tensor
powers $V^{\tensor k}$ for all $k \in \mathbb{N}$ by taking their
tensor powers: we equip $V^{\tensor k}$ with the $\pi$-topology
inherited from $V$. This means that for a continuous seminorm $p$ on
$V$ we consider $p^{\tensor k}$ on $V^{\tensor k}$ and take all such
seminorms to define a locally convex topology on $V^{\tensor
k}$. Viewing the symmetric tensor powers as a subspace, this induces
the $\pi$-topology also for $\Sym^\bullet_{\mathbb{C}}(V)$, simply by
restricting the seminorms $p^{\tensor k}$. For the whole symmetric
algebra we need to extend these seminorms we have on each symmetric
degree. This can be done in many inequivalent ways. Useful for our
purposes is the following construction. We fix a parameter $R \ge
\frac{1}{2}$ and define
\begin{equation}
\label{eq:SeminormS}
p_R(a) = \sum_{k=0}^\infty k!^R p^{\tensor k}(a_k)
\end{equation}
for every $a = \sum_{k=0}^\infty a_k$ with $a_k \in
\Sym^k_{\mathbb{C}}(V)$. Note that the sum is finite as long as we
take $a$ in the symmetric algebra. Now taking all those seminorms
$p_R$ for all continuous seminorms $p$ of $V$ induces a locally convex
topology on $V$. Clearly, this is again Hausdorff. Moreover, all
$\Sym^k_{\mathbb{C}}(V)$ are closed embedded subspaces in
$\Sym^\bullet_{\mathbb{C}}(V)$ with respect to this topology.
The remarkable property of this topology is now that a continuous
$\Lambda$ will induce a continuous star product \cite{waldmann:2014a}:
\begin{thm}
Let $\Lambda\colon V \times V \longrightarrow \mathbb{C}$ be a
continuous bilinear form on $V$. Then $\star$ is a continuous
associative product on $\Sym^\bullet_{\mathbb{C}}(V)$ with respect
to the locally convex topology induced by all the seminorms $p_R$
with $p$ being a continuous seminorm on $V$, as long as $R \ge
\frac{1}{2}$.
\end{thm}
The proof consists in an explicit estimate for $p_R(a \star b)$. Note
that the topology can \emph{not} be locally multiplicatively convex
since in the Weyl algebra we have elements satisfying canonical
commutation relations, thereby forbidding a submultiplicative
seminorm.
\begin{defn}[Locally convex Weyl algebra]
Let $\Lambda\colon V \times V \longrightarrow \mathbb{C}$ be a
continuous bilinear form on $V$. Then the completion of
$\Sym^\bullet_{\mathbb{C}}(V)$ with respect to the above locally
convex topology and with the canonical extension of $\star$ is
called the locally convex Weyl algebra $\mathcal{W}_R(V, \star)$.
\end{defn}
Thus we have found a framework where the Weyl star product actually
converges. Without proofs we list a few properties of this Weyl
algebra:
\begin{itemize}
\item The locally convex Weyl algebra $\mathcal{W}_R(V, \star)$ is a
locally convex unital associative algebra. The product $a \star b$
can be written as the absolutely convergent series
\begin{equation}
\label{eq:SeriesWeyl}
a \star b = \mu \circ \exp(z P_\Lambda) (a \tensor b).
\end{equation}
\item The product $\star$ depends holomorphically on $z \in
\mathbb{C}$.
\item For $\frac{1}{2} \le R < 1$ the locally convex Weyl algebra
$\mathcal{W}_R(V, \star)$ contains the exponential functions
$\E^{\alpha v}$ for all $v \in V$ and all $\alpha \in
\mathbb{C}$. They satisfy the usual Weyl relations. Note that not
only the unitary ones, i.e. for $\alpha$ imaginary, are contained
in the Weyl algebra, but all exponentials.
\item The locally convex Weyl algebra is nuclear iff $V$ is
nuclear. In all relevant examples in quantum theory this will be
the case. In this case we refer to the \emph{nuclear Weyl
algebra}.
\item If $V$ admits an absolute Schauder basis, then the symmetrized
tensor products of the basis vectors constitute an absolute
Schauder basis for the Weyl algebra, too. Again, in many
situations $V$ has such a basis.
\item The Weyl algebras for different $\Lambda$ on $V$ are isomorphic
if the antisymmetric parts of the bilinear forms coincide.
\item Evaluations at points in the topological dual $V'$ are
continuous linear functionals on $\mathcal{W}_R(V, \star)$. Hence
we still can view the elements of the completion as particular
functions on $V'$.
\item The translations by elements in $V'$ still act on
$\mathcal{W}_R(V, \star)$ by continuous automorphisms. If $R < 1$
these translations are inner automorphism as soon as the element
$\varphi \in V'$ is in the image of the musical map induced by
$\Lambda$.
\end{itemize}
We conclude this section now with a few comments on examples. First it
is clear that in finite dimensions we can take $V = \mathbb{R}^{2n}$
with the canonical Poisson bracket on the symmetric algebra. Then many
types of orderings can be incorporated in fixing the symmetric part of
$\Lambda$, while the antisymmetric part is given by the Poisson
bracket. Thus all the resulting star products allow for this analytic
framework. This includes examples known earlier in the literature, see
e.g. \cite{omori.maeda.miyazaki.yoshioka:2007a,
beiser.roemer.waldmann:2007a}. In this case we get a nuclear Weyl
algebra with an absolute Schauder basis.
More interesting is of course the infinite dimensional case. Here we
have to specify the space $V$ and the bilinear form $\Lambda$ more
carefully. In fact, the \emph{continuity} of $\Lambda$ becomes now a
strong conditions since bilinear maps in locally convex analysis tend
to be only separately continuous without being continuous. However,
there are several situations where we can either conclude the
continuity of a bilinear separately continuous map by abstract
arguments, like for Fréchet spaces. Or one can show directly that the
particular bilinear form one is interested in is continuous. We give
one of the most relevant examples for (quantum) field theory:
\begin{ex}
Let $M$ be a globally hyperbolic spacetime and let $D$ be a
normally hyperbolic differential operator acting on a real vector
bundle $E$ with fiber metric $h$. Moreover, we assume that $D$ is
a connection Laplacian for a metric connection with respect to $h$
plus some symmetric operator $B$ of order zero. In all relevant
examples this is easy to obtain. Then one has advanced and
retarded Green operators leading to the propagator $F_M$ acting on
test sections $\Secinfty_0(E^*)$. We take $V = \Secinfty_0(E^*)$
with its usual LF topology. Then
\begin{equation}
\label{eq:LambdaCov}
\Lambda(\varphi, \psi)
=
\int_M h^{-1}(F_M(\varphi), \psi) \mu_g
\end{equation}
is the bilinear form leading to the Peierls bracket on the
symmetric algebra $\Sym^\bullet(V)$. Here $\mu_g$ is the metric
density as usual. The kernel theorem then guarantees that
$\Lambda$ is continuous as needed. Thus we obtain a locally convex
and in fact nuclear Weyl algebra from this. Now $\Lambda$ is
highly degenerated. It follows that in the Poisson algebra there
are many Casimir elements. The kernel of $F_M$ generates a Poisson
ideal and also an ideal in the Weyl algebra, which coincides with
the vanishing ideal of the solution space. Hence dividing by this
(Poisson) ideal gives a Poisson algebra or Weyl algebra which can
be interpreted as the observables of the (quantum) field theory
determined by the wave equation $Du = 0$. It can then be shown
that for every Cauchy surface $\Sigma$ in $M$ there is a canonical
algebra isomorphism to the Weyl algebra build from the symplectic
Poisson algebra on the initial conditions on $\Sigma$. Details of
this construction can be found in \cite{waldmann:2014a}, see also
\cite{baer.ginoux.pfaeffle:2007a} for the background information
on the wave equation.
\end{ex}
\subsection*{Acknowledgment}
It is a pleasure to thank the organizers of the Regensburg conference
for their kind invitation and the fantastic organization of this
stimulating conference. Moreover, I would like to that Chiara Esposito
for helpful remarks on the manuscript.
|
\section{Introduction}
This paper takes a local approach to studying the evolution of
cellular automata, following on the global approach of ``PICARD''
\cite{pavlic2014self}.
\begin{quote}
\emph{Like a traditional one-dimensional CA, PICARD executions move
from one iteration to another by some rule. However, whereas
traditional CA's require the rule to be static and externally
specified, PICARD infers the iteration rule from the current state
of the CA itself.}
\sourceatright{\cite[pp. 1--2]{pavlic2014self}}
\end{quote}
PICARD's inferred rule is derived from the current state of the CA by
a global characteristics, such as the number of 1's in the CA's
current state (modulo 256), or the density $\rho$ of 1's (normalised
as $\rho/256$). These global criteria are similar to Van Valen's
theory of resource density as an ``incompressible gel'' \cite{van1973new}.
In the current paper we introduce the notion of a MetaCA, in which CA
rules are derived locally at each cell within the CA as it runs.
Examples appear in Figure \ref{metaca-taster}. Here, each colour
represents one of the 256 standard one-dimensional CA rules. States
evolve locally, according to globally-defined dynamics.
\begin{figure}
\includegraphics[width=\columnwidth]{metaca.png}
\caption{An illustration of MetaCA evolution\label{metaca-taster}}
\end{figure}
\newpage
\section{Background}
\subsection{Cellular Automata}
A crucial development in the history of CA research was the proof
\cite{cook2004universality} that certain CA rules are Turing complete
(in particular, Rule 110 in Wolfram's numbering system
\cite{wolfram1994cellular} enjoys this property).
Earlier classic works
\cite{langton1990computation,mitchell1993revisiting,packard1988adaptation}
exploring related ``edge of chaos'' effects. In
\cite{packard1988adaptation,mitchell1993revisiting,mitchell1994evolving},
genetic algorithms are used to search the space of CA rules via
crossover and mutation. This sort of evolution is global and is
connected with the CA rule by a derived parameter, ``Langton's
$\lambda$'' (cf. \cite{langton1990computation}). An overview of the
``EvCA'' programme is presented in \cite{hordijk2013evca}.
Closest to the work presented here is \cite{sipper1997evolution},
which introduces the paradigm of \emph{cellular programming}. As the
name indicates, this approach is a fusion of ideas from cellular
automata and genetic programming.
\begin{quote}
\emph{As opposed to the standard genetic algorithm, where a population
of independent problem solutions globally evolves, our approach
involves a grid of rules that coevolves locally.}
\sourceatright{\cite[p. 74]{sipper1997evolution}}
\end{quote}
In cellular programming, local evolution of the CA rule makes use of a
local ``fitness'' (\cite[pp. 79--81]{sipper1997evolution}), as the
systems are evolved to perform certain global computational tasks.
In the current effort, although we are interested in behaviour that
tends towards edge-of-chaos effects, system evolution is not directly
guided by a specific fitness criterion, but only by variations on the
``crossover'' mechanism.
One early application of cellular programming was to evolutionary game
theory, a field with natural parallels
(cf.~\cite{nowak1992evolutionary}). We will not consider game
theoretic approaches in this work, despite being inspired by the
social metaphors that are involved (e.g.~\cite{nowak2006five}). In
our thinking we often switch between conceptual/symbolic,
social/ethical, biological/genetic, and physical/geometric metaphors.
In this connection it is worth mentioning some recent work
\cite{goerg2012licors,goerg2012mixed} that continues in the earlier
tradition of the EvCA project (cf. \cite{hordijk2001upper}), making
use of a relativistic ``light cone'' analysis to identify structure in
CAs. The current paper does not pursue any detailed \emph{post hoc}
analysis of CA behaviour, although we plan to explore this further in
subsequent work. Finally, although not focusing on CAs per se,
\cite{hofstadter1995prolegomena} outlines a set of criteria for the
design of systems that exhibit ``emergent'' intelligence which helped
to motivate the present effort.
\subsection{Conceptual Blending} \label{sec:blend}
One of our inspirations for working with cellular automata is that we
are involved with a research project that studies computational
blending \cite{schorlemmer2014coinvent}, and cellular automata seem to
offer a very simple example of blending behaviour. That is, they
consider the value of neighbouring cells, and produce a result that
``combines'' these results (in some suitably abstract sense) in order
to produce the next generation. We were also inspired by the idea of
``blending'' ordered and chaotic behaviour to produce edge-of-chaos
effects.
We propose to exploit existing formalisms of blending (in the style of
Goguen \cite{gog05}) in the context of cellular automata to
investigate emergent and novel behaviours. The fundamental building
blocks used in calculating concept or theory blends are:
\begin{description}
\item[Input Concepts] are the concepts or theories which are understood have some degree of commonality (syntactic or semantic).
\item[Signature Morphism] is a definition of how symbols are mapped between theories or concepts.
\item[Generic Space] is the space which contains a theory which is common to both input theories.
\item[Blend] is the space computed by combining both theories. The computation is computed using a ``pushout'' from the underlying categorical semantics \cite{MossakowskiEA06}.
\end{description}
Once a blend has been computed, it may represent a concept which is in
some way inconsistent. Equally it may represent a concept which is in
some way incomplete. We can then either weaken an input theory, or
refine the blend:
\begin{description}
\item[Weakening] Given an inconsistent blend it is possible to weaken the input concept in order to produce a consistent blend. Weakening means removing symbols or axioms from the input concept.
\item[Refinement] Given a blend which represents a concept which is in some way incomplete, it is possible to refine the concept by adding symbols or axioms.
\end{description}
This paper presents several examples of simple concepts to which the
blending process applies. In general the notion of a signature
morphism allows input concepts expressed in different languages to be
blended. In this paper the examples shown have input concepts
expressed in the same language, and indeed have the same
specification. This means that the morphisms are not interesting and
the calculated pushout could be computed without utilising the full
machinery of category theory. Planned extensions will explore the
idea of combining rules for cellular automata which may have entirely
different techniques for expressing propagation. For this reason, we
target the Heterogeneous Tool Set (HETS) system
\cite{mossakowski2007heterogeneous} as an infrastructure for computing
blends. We describe our current approach to blending in the context
of cellular automata in Sections \ref{introducing-blending} and
\ref{2d-experiments-design}.
\section{Implementation}
\subsection{Generating Genotypes} \label{sec:geno}
Each elementary CA rule defines a mapping from all eight strings of
0's and 1's to the set \{0,1\}. Thus, for example the rule \textbf{01010100}
is defined as the following operation:
\begin{lstlisting}[mathescape]
0 0 0 $\mapsto$ $\mathbf{0}$
0 0 1 $\mapsto$ $\mathbf{1}$
0 1 0 $\mapsto$ $\mathbf{0}$
0 1 1 $\mapsto$ $\mathbf{1}$
1 0 0 $\mapsto$ $\mathbf{0}$
1 0 1 $\mapsto$ $\mathbf{1}$
1 1 0 $\mapsto$ $\mathbf{0}$
1 1 1 $\mapsto$ $\mathbf{0}$
\end{lstlisting}
There are 256 of these rules; the example above is Rule 84 in
Wolfram's standard enumeration of 1D CAs \cite{wolfram1994cellular}.
The basic concept of the MetaCA is to evolve a CA with 256 possible
states -- rather than the traditional two -- where each state now
corresponds to a ``CA rule''. Then we can then apply this rule to
decide the output for the next cell, depending also on the state of
the neighbouring cells. By positioning three CA rules next to each
other, we define a multiplication by applying the central rule bitwise
across the alleles.
For example, here is the result of ``multiplying'' $01101110\times
01010100\times 01010101$. In the context of such an operation, we
refer to the central term as the ``local rule,'' and we highlight it
in bold below.
\lstset{
xleftmargin=.1\columnwidth, xrightmargin=.01\columnwidth
}
\begin{lstlisting}[mathescape]
0 $\mathbf{0}$ 0 $0$ $\text{\emph{Apply local rule to ``000''}}$
1 $\mathbf{1}$ 1 $0$ $\text{\emph{Apply local rule to ``111}}$
1 $\mathbf{0}$ 0 $0$ $\text{\emph{Apply local rule to ``100''}}$
0 $\mathbf{1}$ 1 $\droparrow$ $1$ $\text{\emph{Apply local rule to ``011''}}$
1 $\mathbf{0}$ 0 $0$ $\text{\emph{Apply local rule to ``100''}}$
1 $\mathbf{1}$ 1 $0$ $\text{\emph{Apply local rule to ``111''}}$
1 $\mathbf{0}$ 0 $0$ $\text{\emph{Apply local rule to ``100''}}$
0 $\mathbf{0}$ 1 $1$ $\text{\emph{Apply local rule to ``001''}}$
\end{lstlisting}
Realised in a simulation with random starting conditions, the results
of this operation are not particularly impressive: they stabilise
early and do not produce any interesting patterns (Figure
\ref{barcode}).
\begin{figure}
\includegraphics[width=\columnwidth,trim = 135mm 177mm 0mm 0mm,clip=true]{paint-drips.png}
\caption{Under evolution according to the local rule without blending
dynamics, a barcode-like stable pattern forms
quickly\label{barcode}}
\end{figure}
\subsection{Introducing Blending} \label{introducing-blending}
The blending variant says to first compute the ``generic space'' by
noting the alleles where the two adjacent neighbours are the same, and
where they differ. Only when the generic space retains some ambiguity
(indicated by $\{0,1\}$) do we apply the local rule (again recorded
on the centre cell at left and highlighted in bold) in a bitwise
manner across each allele, to arrive at the final result.
\lstset{
xleftmargin=.05\columnwidth, xrightmargin=.01\columnwidth
}
\begin{lstlisting}[mathescape]
0 $\mathbf{0}$ 0 0 $\:0$ $\text{\emph{Neighbours are both 0}}$
1 $\mathbf{1}$ 1 1 $\:1$ $\text{\emph{Neighbours are both 1}}$
1 $\mathbf{0}$ 0 {0,1} $\boxed{0}$ $\text{\emph{Apply local rule to ``100''}}$
0 $\mathbf{1}$ 1 $\droparrow$ {0,1} $\droparrow$ $\boxed{1}$ $\text{\emph{Apply local rule to ``011''}}$
1 $\mathbf{0}$ 0 {0,1} $\boxed{0}$ $\text{\emph{Apply local rule to ``100''}}$
1 $\mathbf{1}$ 1 1 $\:1$ $\text{\emph{Neighbours are both 1}}$
1 $\mathbf{0}$ 0 {0,1} $\boxed{0}$ $\text{\emph{Apply local rule to ``100''}}$
0 $\mathbf{0}$ 1 {0,1} $\boxed{1}$ $\text{\emph{Apply local rule to ``001''}}$
\end{lstlisting}
For illustrative purposes, this blend has been formalised in the HETS
system by introducing CASL files to represent the 8 bit encodings
(Listing \ref{CASL-listing}, and corresponding development graph shown
in Figure \ref{fig:hetsblend}).
The computed blend is inconsistent as there is not a unique value
representing the output value of each function. In order to resolve
this we weaken the input rules in CASL by removing the function values
which cause conflict.
Note that purposes of efficiency, we have implemented our 1D
experiments in LISP rather than in \mbox{HETS}/\mbox{CASL}. We've put
the working code on
Github\footnote{\url{https://github.com/holtzermann17/metaca}}.
\begin{figure}[!ht]
\includegraphics[width=0.45\textwidth]{dol_dev.pdf}
\caption{The development graph for calculating a blend of 8 bit encodings\label{fig:hetsblend}}
\end{figure}
\subsection{2D Experiments} \label{2d-experiments-design}
In order to extend the ideas presented so far in the 1D case, let us
consider a variant of Conway's Game of Life \cite{conway}, in which a
global rule exists defining whether a square is alive or dead. We
extend this by introducing the notion of a local rule at each square
-- a genotype, which governs the propagation of the phenotype.
In Conway's Game of life, one can view the rules for propagation as
partitions on a finite interval $[0,8]$.
\begin{center}
\includegraphics[width=0.45\textwidth]{conway.pdf}
\end{center}
The number on the line corresponds to the number of alive neighbours
adjacent, in cardinal and inter-cardinal directions, to a given
square. If the square is dead then it becomes alive (labelled
reproduce) if the number of alive neighbours is exactly three. If
there are five or more alive neighbours the square dies from
overcrowding. If there are fewer than three alive neighbours the
square dies from underpopulation. In all other cases the square
maintains its status.
This can be generalised to partitions within a more finely grained
line, for example from 0 to 1000, one creates a genotype $(x,y,z)$:
\begin{center}
\includegraphics[width=0.45\textwidth]{2dgenotype.pdf}
\end{center}
We introduce the corresponding notion of a {\em weight} for each cell.
The {\em phenotype} of the cell is then a pair $(\text{\emph{alive}},
\text{\emph{weight}})$ which denotes whether the cell is alive, and
what weight is has. In this paper we always calculate a newly
propagated weight as the average of the neighbours' weights.
The notion of local propagation is introduced by allowing the
genotypes to be blended at each point where a cell remains or becomes
alive. As we have represented the genotype as a partitioned line, we
can, for example perform a blend where the partition is blended in
such a way as the minimise the lowest bound and maximise the highest
bound, and maximise the interval for reproduction. Given two
genotypes $(x_1, y_1, z_1)$ and $(x_2, y_2, z_2)$, the blend is
$(\mathrm{min} \{x_1,x_2\}, \mathrm{max} \{y_1,y_2\}, \mathrm{max}
\{z_1,z_2\})$
\begin{center}
\includegraphics[width=0.45\textwidth]{2dgenotypeblend.pdf}
\end{center}
Note that this is just one of several possible blending strategies,
which we refer to as a {\em union} blend, since it maximises the
partitions which pertain to survival. We consider alternative blends
in our experiments.
\section{Results}
\subsection{1D CAs} \label{1d-results}
One of the first things we noticed was that even though the blending
dynamic creates more interesting ``CA-like'' patterns than simple
evolution according to the local rule (as illustrated in Figure
\ref{metaca-taster}), it also forms stable bands after this
interesting initial period. In Figure \ref{flag}, this is illustrated
in a CA running with 500 cells over 500 generations. Figure
\ref{flag} also includes a phenotype (in black and white) which is
driven entirely by the genotype: that is, if the local genotype is
\boxed{\alpha\vphantom{b\gamma}}\boxed{\beta\vphantom{b\gamma}}\boxed{\gamma\vphantom{b\gamma}}
where $\alpha, \beta, \gamma \in \{0,1\}^8$
and the local phenotype is
\boxed{a\vphantom{b\gamma}}\boxed{b\vphantom{b\gamma}}\boxed{c\vphantom{b\gamma}}
where $a, b, c \in \{0,1\}$,
then the genotype evolves locally according to the meta-rule $\alpha
\times \beta \times \gamma$ (in the blending variant) while the
phenotype evolves by applying the local rule $\beta$ to the data
``$abc$.''
\begin{figure}
\includegraphics[width=\columnwidth]{flag.png}
\caption{Phenotype with behaviour determined by genotype\label{flag}}
\end{figure}
In the phenotype layer, we see a few bands with interesting patterns,
where the MetaCA at left has stabilised locally into one of the more
interesting CA rules. However, the long term evolution is not
particularly interesting: the structure observed in Figure
\ref{metaca-taster} disappears quickly.
We therefore decided to introduce random mutations to the genotype,
illustrated in Figures \ref{random-mutation}--\ref{seti}. With a high
mutation rate, both genotype and phenotype are almost reduced to
confetti. If we reduce the mutation rate sufficiently, some degree of
stability is preserved, and the vertically striped bands are
transformed into intermingling swaths of colour (Figure
\ref{lower-rate}). We also see areas with more finely-grained
structure in the phenotype layer.
In Figure \ref{seti}, the colour-coded genotype layer has been
replaced with a greyscale coding, and we see more clearly how the
phenotype behaviour follows that of the genotype. That is, genotypes
similar to Rule 0 (00000000) or Rule 256 (11111111) tend to produce 0
or 1, respectively, in the phenotype layer. Rules that output a blend
of 0's and 1's are mapped to grey shades. Several interesting rules
(Rule 110, Rule 30, Rule 90, Rule 184, and their reversals, bitwise
inverses, and inverted-reversals) are highlighted in colour. In
particular, Rule 110 variants are highlighted in red.
\begin{figure}
\includegraphics[width=\columnwidth]{big.png}
\caption{A high rate of mutation produces tantalising random structures \label{random-mutation}}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{lamp-down-low.png}
\caption{Throttling down the mutation rate preserves some of the large-scale stability while making room for variability \label{lower-rate}}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{seti.png}
\caption{The search for intelligent life in the computational universe \label{seti}}
\end{figure}
We observe that Rule 0 and Rule 256 behaviour tends to predominate.
Grey areas appear to be semi-stable. Red patches appear and
disappear, as if independent planets evolve intelligent life and are
then extinguished. With this physics, ``intelligent life'' seems
inevitable, but also inevitably short-lived. One would have to look
for another overall physics for intelligent behaviour to predominate.
A potential indication of the direction to look in is presented in
Figure \ref{reef}, which presents CAs generated by adjusting the
typical blending evolution pattern by an (erroneously-programmed)
mutation rule that only flips the first bit. We see that long-term
behaviour in the genotype flutters randomly between Rule 0 (00000000)
and Rule 128 (10000000). The short-term behaviour in the phenotype is
nevertheless quite interesting, exhibiting many of the familiar
lifelike edge-of-chaos patterns before ultimately succumbing to a
version of Newton's First Law.
\begin{figure}
\includegraphics[width=\columnwidth]{reef.png} \newline
\includegraphics[width=\columnwidth]{eoc.png}
\caption{A skewed mutation pattern \label{reef}}
\end{figure}
\subsection{2D CAs} \label{2d-results}
To see the behaviour of the union blend in action consider an
initially populated grid, where colours represent the weights of alive
cells:
\begin{center}
\includegraphics[width=0.45\textwidth]{initial2d.jpg}
\end{center}
For this example, we initially restrict the computation of the blend for a particular cell to take place when the cell is alive in the next iteration. Also we compute the blend of genotype for all neighbours, whether dead or alive.
After 300 iterations the colony has grown a small amount:
\begin{center}
\includegraphics[width=0.45\textwidth]{3002d.jpg}
\end{center}
Over time, the population continues to grow, with large patches of low-weight (black) cells:
\begin{center}
\includegraphics[width=0.45\textwidth]{30002d.jpg}
\end{center}
Finally some structure starts to appear in the clustering:
\begin{center}
\includegraphics[width=0.45\textwidth]{300002d.jpg}
\end{center}
The propagation that follows shows a population of cells which grows
slowly overtime. The majority of the members have low weight
(represented by black squares), but interspersed within the population
are chains of squares with high weight (represented by red squares)
adjacent to dead cells (white).
\subsubsection{Modified Blends}
So far we have only showed the union blend working on the genotype. However, it is possible to use different blending techniques:
\begin{itemize}
\item{Consider blending only the genotypes of alive neighbours, or all neighbours;}
\item{Consider only blending genotypes for cells which are alive after propagation;}
\item{Consider an \emph{intersection} blend, where the partition sizes for survival are minimised;}
\item{Consider an \emph{average} blend, where the values of each genotype $(x_i,y_i,z_i)$ are summed and divided by either the number of alive neighbours, or the total number of neighbours.}
\end{itemize}
As an example of different observed emergent behaviour consider a union blend where the blend is only computed from alive neighbours, and as before we compute only for cells which are alive at the next iteration. We start with an initial state:
\begin{center}
\includegraphics[width=0.45\textwidth]{init2.jpg}
\end{center}
and observe a changing, but relatively steady pattern (resembling the motion of a flame) which does not grow in size using the union blend:
\begin{center}
\includegraphics[width=0.45\textwidth]{init2_union.jpg}
\end{center}
\noindent where the weight characteristic of the phenotype of each
cell has fallen to very low.
Finally, consider applying instead an average blend under the same
initial conditions:
\begin{center}
\includegraphics[width=0.45\textwidth]{init2_average.jpg}
\end{center}
Then we see a less steady but more active growth, with populations moving in triangular shapes away from population centres, leaving very small but steady and inactive populations behind:
\begin{center}
\includegraphics[width=0.45\textwidth]{init2daverage++.jpg}
\end{center}
\noindent the quickly-moving populations do not have a convergent weight characteristic of the phenotype, as in the case with the union blend for the same initial conditions.
\FloatBarrier
\begin{listing}[H]
{\scriptsize
\begin{hetcasl}
\KW{library} \Id{metaca}\\
\\
\KW{logic} \SId{CASL}\\
\\
\SPEC \=\SIdIndex{METACABitencoding} \Ax{=}\\
\> \KW{free} \KW{type} \=\Id{Bit} \Ax{:}\Ax{:}\=\Ax{=} \Ax{0} \AltBar{} \Ax{1}\\
\> \SORT \Id{Triple}\\
\> \OPS \=\Id{t} \Ax{:} \=\Id{Bit} \Ax{\times} \Id{Bit} \Ax{\times} \Id{Bit} \Ax{\rightarrow} \Id{Triple};\\
\>\> \Id{bitop}\Ax{\_\_} \Ax{:} \=\Id{Triple} \Ax{\rightarrow} \Id{Bit}\\
\KW{end}\\
\\
\%\% How to calculate a blend given three 8-bit genotypes\\
\SPEC \=\SIdIndex{METACABitcalc} \Ax{=}\\
\> \SId{METACABitencoding}\\
\THEN \=\OP \=\Id{blend}\Ax{\_\_}\Ax{\_\_} \Ax{:} \=\Id{Triple} \Ax{\times} \Id{Triple} \Ax{\rightarrow} \Id{Bit}\\
\> \Ax{\forall} \=\Id{t1}, \Id{t2}, \Id{t3} \Ax{:} \Id{Triple} \\
\> \Ax{\bullet} \=\Id{bitop} \Id{t1} \Ax{=} \Id{bitop} \Id{t2} \Ax{\Rightarrow} \=\Id{blend} \Id{t1} \Id{t2} \Ax{=} \Id{bitop} \Id{t1}\\
\> \Ax{\bullet} \=\Ax{\neg} \=\Id{bitop} \Id{t1} \Ax{=} \Id{bitop} \Id{t2} \Ax{\Rightarrow} \=\Id{blend} \Id{t1} \Id{t2} \Ax{=} \Id{bitop} \Id{t3}\\
\KW{end}\\
\\
\SPEC \=\SIdIndex{LeftRule} \Ax{=}\\
\> \SId{METACABitencoding}\\
\THEN \=\Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{0}, \Ax{0}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{0}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{1}, \Ax{0}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{1}, \Ax{1}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{0}, \Ax{0}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{0}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{1}, \Ax{0}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{1}, \Ax{1}) \Ax{=} \Ax{0}\\
\KW{end}\\
\\
\SPEC \=\SIdIndex{RightRule} \Ax{=}\\
\> \SId{METACABitencoding}\\
\THEN \=\Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{0}, \Ax{0}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{0}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{1}, \Ax{0}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{1}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{0}, \Ax{0}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{0}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{1}, \Ax{0}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{1}, \Ax{1}) \Ax{=} \Ax{1}\\
\KW{end}\\
\\
\SPEC \=\SIdIndex{LocalRule} \Ax{=}\\
\> \SId{METACABitencoding}\\
\THEN \=\Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{0}, \Ax{0}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{0}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{1}, \Ax{0}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{1}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{0}, \Ax{0}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{0}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{1}, \Ax{0}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{1}, \Ax{1}) \Ax{=} \Ax{0}\\
\KW{end}\\
\\
\%\% Generic is common between left and right\\
\SPEC \=\SIdIndex{Generic} \Ax{=}\\
\> \SId{METACABitencoding}\\
\THEN \=\Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{0}, \Ax{0}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{0}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{0}, \Ax{1}) \Ax{=} \Ax{1}\\
\KW{end}\\
\\
\%\% Morphism from Generic to Left\\
\VIEW \=\SId{Left} \Ax{:} \=\SId{Generic} \KW{to} \SId{LeftRule}\\
\KW{end}\\
\\
\%\% Morphism from Generic to Right\\
\VIEW \=\SId{Right} \Ax{:} \=\SId{Generic} \KW{to} \SId{RightRule}\\
\KW{end}\\
\\
\%\% This will be inconsistent\\
\SPEC \=\SIdIndex{Blend} \Ax{=}\\
\> \KW{combine} \=\Id{Left}, \Id{Right}\\
\KW{end}\\
\\
\SPEC \=\SIdIndex{WeakenedLeftRule} \Ax{=}\\
\> \SId{METACABitencoding}\\
\THEN \=\Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{0}, \Ax{0}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{0}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{1}, \Ax{0}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{1}, \Ax{1}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{0}, \Ax{0}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{0}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{1}, \Ax{0}) \Ax{=} \Ax{1}\\
\KW{end}\\
\\
\SPEC \=\SIdIndex{WeakenedRightRule} \Ax{=}\\
\> \SId{METACABitencoding}\\
\THEN \=\Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{0}, \Ax{0}) \Ax{=} \Ax{0}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{0}, \Ax{0}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{0}, \Ax{1}) \Ax{=} \Ax{1}\\
\> \Ax{\bullet} \=\Id{bitop} \Id{t}(\=\Ax{1}, \Ax{1}, \Ax{1}) \Ax{=} \Ax{1}\\
\KW{end}\\
\VIEW \=\SId{WeakenedLeft} \Ax{:} \=\SId{Generic} \KW{to} \SId{WeakenedLeftRule}\\
\KW{end}\\
\VIEW \=\SId{WeakenedRight} \Ax{:} \=\SId{Generic} \KW{to} \SId{WeakenedRightRule}\\
\KW{end}\\
\\
\%\% A computed blend as new 8 bit encoding\\
\SPEC \=\SIdIndex{ConsistentBlend} \Ax{=}\\
\> \KW{combine} \=\Id{WeakenedLeft}, \Id{WeakenedRight}\\
\AND \SId{METACABitcalc}\\
\AND \SId{LocalRule}\\
\KW{end}
\end{hetcasl}
}
\caption{CASL source code listing calculating the running example $01101110\times 01010100\times 01010101$ via the blending meta-rule\label{CASL-listing}}
\end{listing}
\newpage
\begin{figure}[H]
\includegraphics[width=0.4\textwidth]{2dblend.pdf}
\caption{Blending different 2d genotypes}
\label{fig:2dblend}
\end{figure}
\begin{listing}[H]
{\scriptsize
\begin{hetcasl}
\KW{library} \Id{metaca2d}\\
\\
\KW{logic} \SId{CASL}\\
\\
\SPEC \=\SIdIndex{Nat} \Ax{=}\\
\> \SORT \Id{Nat}\\
\> \OP \=\Id{max} \Ax{:} \=\Id{Nat} \Ax{\times} \Id{Nat} \Ax{\rightarrow} \Id{Nat}\\
\> \OP \=\Id{min} \Ax{:} \=\Id{Nat} \Ax{\times} \Id{Nat} \Ax{\rightarrow} \Id{Nat}\\
\KW{end}\\
\\
\SPEC \=\SIdIndex{Colour} \Ax{=}\\
\> \SORT \Id{Colour}\\
\> \OP \=\Id{maxhue} \Ax{:} \=\Id{Colour} \Ax{\times} \Id{Colour} \Ax{\rightarrow} \Id{Colour}\\
\KW{end}\\
\\
\\
{{}\KW{\%\%} a 2\Ax{-}d cellular automaton with numerical Genotype}\\
\SPEC \=\SIdIndex{Numericalencoding} \Ax{=}\\
\> \SId{Nat}\\
\THEN \=\SORT \Id{NGenotype}\\
\> \OPS \=\Id{genotype} \Ax{:} \=\Id{Nat} \Ax{\times} \Id{Nat} \Ax{\times} \Id{Nat} \Ax{\rightarrow} \Id{NGenotype};\\
\>\> \Id{t} \Ax{:} \=\Id{Nat} \Ax{\times} \Id{Nat} \Ax{\times} \Id{Nat} \Ax{\rightarrow} \Id{NGenotype};\\
\>\> \Id{numblend} \Ax{:} \=\Id{NGenotype} \Ax{\times} \Id{NGenotype} \Ax{\rightarrow} \Id{NGenotype}\\
\> \Ax{\forall} \Id{g1}, \Id{g2} \Ax{:} \Id{NGenotype}; \=\Id{x1}, \Id{y1}, \Id{z1}, \Id{x2}, \Id{y2}, \Id{z2}, \Id{x3}, \Id{y3}, \Id{z3} \Ax{:} \Id{Nat} \\
\> \Ax{\bullet} \=\Id{g1} \Ax{=} \Id{t}(\=\Id{x1}, \Id{y1}, \Id{z1}) \Ax{\wedge} \=\Id{g2} \Ax{=} \Id{t}(\=\Id{x2}, \Id{y2}, \Id{z2}) \\
\>\> \Ax{\Rightarrow} \=\Id{numblend}(\=\Id{g1}, \Id{g2}) \\
\>\>\> \Ax{=} \Id{t}(\=\Id{min}(\=\Id{x1}, \Id{x2}), \Id{min}(\=\Id{y1}, \Id{y2}), \Id{max}(\=\Id{z1}, \Id{z2}))\\
\KW{end}\\
\\
\\
{{}\KW{\%\%} A colour CA Genotype}\\
\SPEC \=\SIdIndex{Colourencoding} \Ax{=}\\
\> \SId{Colour}\\
\THEN \=\SORT \=\Id{CGenotype} \Ax{=} \Id{Colour}\\
\> \OP \=\Id{hueblend} \Ax{:} \=\Id{CGenotype} \Ax{\times} \Id{CGenotype} \Ax{\rightarrow} \Id{CGenotype}\\
\> \Ax{\forall} \=\Id{g1}, \Id{g2} \Ax{:} \Id{CGenotype} \\
\> \Ax{\bullet} \=\Id{hueblend}(\=\Id{g1}, \Id{g2}) \Ax{=} \Id{maxhue}(\=\Id{g1} \Id{as} \Id{Colour}, \=\Id{g2} \Id{as} \Id{Colour})\\
\KW{end}\\
\\
\\
{{}\KW{\%\%} A generic space}\\
\SPEC \=\SIdIndex{Genencoding} \Ax{=}\\
\> \SORT \Id{S}\\
\> \SORT \Id{Genotype}\\
\> \OP \=\Id{blend} \Ax{:} \=\Id{Genotype} \Ax{\times} \Id{Genotype} \Ax{\rightarrow} \Id{Genotype}\\
\KW{end}\\
\\
\\
{{}\KW{\%\%} A signature morphism from Generic to Numerical}\\
\VIEW \=\SId{NumericalSM} \Ax{:} \\
\> \SId{Genencoding} \KW{to} \SId{Numericalencoding} \Ax{=} \\
\> \Id{S} \Ax{\mapsto} \Id{Nat}, \Id{Genotype} \Ax{\mapsto} \Id{NGenotype}, \=\Id{blend} \Ax{\mapsto} \Id{numblend}\\
\KW{end}\\
\\
\\
{{}\KW{\%\%} A signature morphism from Generic to Colour}\\
\VIEW \=\SId{ColourSM} \Ax{:} \\
\> \SId{Genencoding} \KW{to} \SId{Colourencoding} \Ax{=} \\
\> \Id{S} \Ax{\mapsto} \Id{Colour}, \Id{Genotype} \Ax{\mapsto} \Id{CGenotype}, \=\Id{blend} \Ax{\mapsto} \Id{hueblend}\\
\KW{end}\\
\\
\SPEC \=\SIdIndex{blend} \Ax{=}\\
\> \KW{combine} \=\Id{NumericalSM}, \Id{ColourSM}\\
\KW{end}
\end{hetcasl}
}
\caption{CASL source code using signature morphisms and pushout
calculation to blend genotypes with different
languages\label{CASL-blend-listing}}
\end{listing}
\FloatBarrier
\section{Discussion}
\subsection{Research Contribution}
The motivation for combining a notion of blending with cellular
automata was to investigate ways in which cellular automata could be
used to model processes, where propagation rules, or genotypes, were
locally defined. The main research contributions in the field of two
dimensional cellular automata are
\begin{itemize}
\item We built and implemented a framework where local propagation
experiments can be performed;
\item We used the HETS system to show that the notion of blending can
be used to invent new propagation rules for different genotypes;
\item We invented simply definable genotypes and blends of these
genotypes to show proof of concept;
\item Finally, we shared the results of simulations that illustrate
qualitative behaviour in one and two dimensional MetaCAs.
\end{itemize}
The primary limitation of this work is that our results are purely
observational at present. For example, the early experiments seemed
to provide visual evidence that blending is useful: Figure
\ref{metaca-taster} more interesting than Figure \ref{barcode}. The
robustness of our qualitative findings have been supported by
developing a range of different experiments, for example, some analogy
could be drawn between the ``grey areas'' observed in Figure
\ref{seti} for the 1D case and the red-and-white chains that develop
in the 2D case under union blending.
Our results confirm the basic finding of CA research: interesting
global behaviour can arise from simple rules governing local
interactions, with the added twist the rules can also arise locally.
The MetaCA setting seems to offer fertile ground for further
computational research into evolutionary and co-evolutionary effects.
\subsection{Social Interpretation}
One can view the propagation of cells and patterns in a 1D or 2D
MetaCA as a social process, and blending as a knowledge exchange. In
the 2D case, we can think of the generated diagrams as illustrations
of interactions between individuals with high knowledge, skill, or
social impact (high weight), and those with less (low weight). The
propagation in the ``union'' blend shows how large numbers of
individuals with low social impact outnumber those with high social
impact, but those with high social impact impose the emergent
structure and determine the growth of the group of individuals.
In a fundamental respect our blending rules seem to embody a
thought-provoking blend of two very different kinds of ``ethics.''
Specifically: blending seems to introduce a dynamic similar to Carol
Gilligan's \emph{ethic of care} \cite{gilligan1982different}, which
seeks to defend the relationships that obtain in a given situation.
Here this is manifested by the question ``Have my neighbours already
formed a consensus?'' This behaviour augments and extends the local
rule, which would correspond to Lawrence Kohlberg’s \emph{ethic of
justice} (cf. \cite{benhabib1985generalized}).
As we saw in Section \ref{1d-results}, we would have to work harder to
find meta-rules that give rise to an ``intelligent universe'' or in
which life (considered as symbolic computation) plays an obvious
negentropic role (\emph{apr\`es} Bergson \cite{bergson1912creative}).
One strategy that has not been developed here would be to make use of
a ``Baldwin effect'' \cite{baldwin-effect,weber2003evolution}, to use
``learning'' (considered as entropy) in the phenotype layer to drive
evolution. More specifically,
\boxed{0\vphantom{b\gamma}}\boxed{0\vphantom{b\gamma}}\boxed{0\vphantom{b\gamma}} $\mapsto$
\boxed{0\vphantom{b\gamma}} and
\boxed{1\vphantom{b\gamma}}\boxed{1\vphantom{b\gamma}}\boxed{1\vphantom{b\gamma}} $\mapsto$
\boxed{1\vphantom{b\gamma}} seem to be relatively uninteresting behaviours, but
they are also hard to resist under the blending dynamics as we've
defined them (compare Figures \ref{flag} and \ref{seti}). Can we find
ways to select against them?
\subsection{Planned extensions}
One observes that under our blending rule, the two non-entropic
behaviours listed above tend to selected for, not against, because
they are examples of the ``neighbours match'' condition. Indeed,
reviewing the essential features of blending in the 1D case, we can
use our basic principles:
\begin{quote}
``\emph{If neighbours match:} \emph{use their shared value as the result.}\\
\emph{If neighbours don't match:} \emph{use local logic to get the result.}''
\end{quote}
\noindent to define a 1D CA rule, if we interpret ``local logic'' to
mean ``substitute my own value as the result.'' Here's how we would
then define blending for triplets:
\lstset{
xleftmargin=.2\columnwidth, xrightmargin=.01\columnwidth
}
\begin{lstlisting}[mathescape]
0 0 0 $\mapsto$ 0 $\text{\emph{Neighbours match}}$
0 0 1 $\mapsto$ 0 $\text{\emph{Local logic}}$
0 1 0 $\mapsto$ 0 $\text{\emph{Neighbours match}}$
0 1 1 $\mapsto$ 1 $\text{\emph{Local logic}}$
1 0 0 $\mapsto$ 0 $\text{\emph{Local logic}}$
1 0 1 $\mapsto$ 1 $\text{\emph{Neighbours match}}$
1 1 0 $\mapsto$ 1 $\text{\emph{Local logic}}$
1 1 1 $\mapsto$ 1 $\text{\emph{Neighbours match}}$
\end{lstlisting}
This is Wolfram's Rule 23: and as it happens, its evolutionary
behaviour is not particularly interesting. Of course, for blending at
the genotype level, ``local logic'' can be determined by any CA. Even
so, when we use blending bitwise on alleles, we only ever run the
local logic on half of the cases, and moreover it always the same
half, determined by a ``censored'' version of Rule 23.
\begin{lstlisting}[mathescape]
0 0 0 $\mapsto$ 0 $\text{\emph{Neighbours match}}$
0 0 1 $\mapsto$ * $\text{\emph{Local logic}}$
0 1 0 $\mapsto$ 0 $\text{\emph{Neighbours match}}$
0 1 1 $\mapsto$ * $\text{\emph{Local logic}}$
1 0 0 $\mapsto$ * $\text{\emph{Local logic}}$
1 0 1 $\mapsto$ 1 $\text{\emph{Neighbours match}}$
1 1 0 $\mapsto$ * $\text{\emph{Local logic}}$
1 1 1 $\mapsto$ 1 $\text{\emph{Neighbours match}}$
\end{lstlisting}
Rather than using Censored Rule 23 as our template, we could instead
have the template determined by phenotype data, thereby inserting the
phenotype as a ``hidden layer'' in the computation.
The standard template could be understood to be generated by locking in
\boxed{0\vphantom{b\gamma}}\boxed{0\vphantom{b\gamma}}\boxed{0\vphantom{b\gamma}} $\mapsto$ \boxed{0\vphantom{b\gamma}}
along with a ``variation''\footnote{%
\boxed{0\vphantom{b\gamma}}\boxed{1\vphantom{b\gamma}}\boxed{0\vphantom{b\gamma}} =
\boxed{0\vphantom{b\gamma}}\boxed{0\vphantom{b\gamma}}\boxed{0\vphantom{b\gamma}} +
\boxed{\:\:\:\vphantom{b\gamma}}\boxed{1\vphantom{b\gamma}}\boxed{\:\:\:\vphantom{b\gamma}}}
\boxed{0\vphantom{b\gamma}}\boxed{1\vphantom{b\gamma}}\boxed{0\vphantom{b\gamma}} $\mapsto$
\boxed{0\vphantom{b\gamma}} and the bitwise inverses of these. A wider class of
templates could calculated from arbitrary phenotype data by the same
operations. What we would lose in abandoning the intuition associated
with local blending, we may be repaid through a much more abstract
but richer procedural blend, operating at the level of
genotype+phenotype evolution. At the very least, we can point to a
generic space, namely the locked-in local rule which would be carried
over (along with its variants) from the phenotype to the corresponding
alleles.
As a simple example of cross-domain blending consider a genotype
defined as in \S\ref{2d-experiments-design}, and another which is
defined by comparing the hue of just one neighbour. Their blend is a
richer theory combining elements from both genotypes. CASL code
expressing these concepts is given in Listing
\ref{CASL-blend-listing}, and the resulting categorical diagram can be
seen in Figure \ref{fig:2dblend}. Experimentation with more
sophisticated genotypes and blends is ongoing.
\subsection{Future work}
Coevolution has been understood to be relevant from both a
philosophical \cite{mead1932philosophy} and empirical perspective
\cite{van1973new}. Finding patterns that allow us to exploit Baldwin
effects to drive the co-evolution of genotype and phenotype in the
direction of intelligent behaviour is an interesting computational
project. The MetaCA domain may help to show how to systematise some
aspects of the search for the principles and techniques that underlie
broader computational intelligence.
Expanding on the relatively simple domain of CAs, we would like to use
HETS to formalise the mechanisms of social knowledge sharing and
problem solving in fields like mathematics. Indeed it may be possible
in the future to encode mathematical problems in a MetaCA or cellular
program and see how a group of agents can solve the problems as a
society. This would be informed by ongoing empirical analysis of real
problem-solving activities \cite{eca} developed in parallel to the
simulation work presented here.
\section{Conclusion}
This research was inspired by the aim to build an example of
computational blending that matched, to some extent, the way blending
might work in social settings. One person suggests an idea, and
another offers a variant of that, a third brings in another idea from
elsewhere and some combination is made. The next day, things head in
another direction completely. Our progress in this research project
has followed this sort of trajectory: from an initial critique of
blending theory (``it's not dynamic enough to be social!'') to some
tentative examples showing how large-scale system dynamics can be
driven by local behaviour in an emergent manner. Perhaps the most
interesting aspect of this research is the relationship between these
emergent dynamics and the meta-rules. Whereas previous CA research
has shown that complex global behaviour can be generated from a set of
simple, local rules, this project gives an enticing glimpse of a
future research programme that carries out a computational search for
those very rules (out of the many possible) that lead to system
behaviour we would recognise as ``intelligent.''
\section{Acknowledgements}
This research has been funded by the Future and Emerging Technologies
(FET) programme within the Seventh Framework Programme for Research of
the European Commission, under FET-Open Grant number 611553
(COINVENT).
We thank Timothy Teravainen, Raymond Puzio, and Cameron Smith for
helpful conversations and pointers to literature.
|
\section{Introduction}
An indefinite inner product in $\mathbb{C}^{n}$ is a conjugate
symmetric sesquilinear form $[x, y]$ together with the regularity
condition that $[x, y]=0$ $ \forall y \in \mathbb{C}^{n}$ holds only
when $x = 0$. Associated with any indefinite inner product, there
exists a unique invertible hermitian matrix $N\in
\mathbb{C}^{n\times n}$ (called a weight) such that $[x, y] =
\langle x, Ny\rangle, $ where $\langle.,.\rangle$ denotes the
Euclidean inner product on $\mathbb{C}^{n}$ and vice versa.
Motivated by the notion of Minkowski space (as studied by
physicists), we also make an additional assumption on N, namely,
$N^{2} = I$. It should be remarked that this assumption also allows
us to compare our results with the Euclidean case, apart from
allowing us to present the results with much algebraic ease.
Investigations of linear maps on indefinite inner product spaces
employ the usual multiplication of matrices which is induced by the
Euclidean inner product of vectors (See for instance \cite{jb}).
This causes a problem as there are two different values for the dot
product of vectors. To overcome this difficulty; Kamaraj,
Ramanathan and Sivakumar introduced a new matrix product called
indefinite matrix multiplication and investigated some of its
properties in \cite{krkkkcs}. More precisely, the indefinite
matrix product of two matrices A and B of sizes $m \times n$ and $n
\times l$ complex matrices, respectively, is defined to be the
matrix $A\circ B := ANB$. The adjoint of A, denoted by $A^{[*]}$,
is defined to be the matrix $NA^{*}M$, where N and M are weights in
the appropriate spaces. Many properties of this product are similar
to that of the usual matrix product (refer \cite{krkkkcs}).
Moreover, it not only rectifies the difficulty indicated earlier,
but also enables us to recover some interesting results in
indefinite inner product spaces in a manner analogous to that of the
Euclidean case. Kamaraj, Ramanathan and Sivakumar \cite{krkkkcs}
also shown that in the setting of indefinite inner product spaces,
Moore-Penrose inverses of certain matrices do not exist with respect
to the usual matrix product where as Moore-Penrose inverses of such
matrices exist with respect to the indefinite matrix product. Hence
they concluded that indefinite matrix product is more appropriate
than the usual matrix product.
The problem of nonnegative invertibility of matrices (or inverses of matrices leaving a cone invariant) was first studied by Collatz \cite{cz2} when he
applied a finite difference method for solving a class of two point
boundary value problems. This idea of nonnegative invertibility has
undergone
a plethora of generalizations over the years. We refer the reader \cite{bp2} (and the references cited there in)
for a detailed survey of these extensions.
In recent years, nonnegative invertibility of Gram matrices has received, a lot
of attention. This has been primarily motivated by applications in
convex optimization problems. In this connection, there is a well
known result that characterizes non negative invertibility of Gram
matrices in terms of obtuseness or acuteness of certain polyhedral
cones. (See for instance Lemma 1.6 in \cite{ce}).
Recently, Sivakumar \cite{s2gram} characterized Moore-Penrose
inverses of Gram operators leaving a cone invariant over Hilbert
spaces. In this paper, we follow the approach of Sivakumar
\cite{s2gram} and discuss the Moore-Penrose inverses of Gram
matrices leaving a cone invariant in an indefinite inner product
space using indefinite matrix product. As the indefinite matrix
product encompasses the Euclidean case as a particular example, it
follows that earlier results in the finite dimensional Euclidean
spaces, are easy corollaries of our main result.
The paper is organized as follows. In section 2, we introduce basic
notations, definitions and results. In section 3, we prove series of
lammas and derive the main theorem.
\section{Notations, Definitions and Preliminaries}
In this section, we introduce notations, definitions and basic results that will be used in
the rest of the paper.
Let $\langle.,.\rangle$ denote the usual Euclidean inner product in
$\mathbb{R}^{n}$. An indefinite inner product is denoted by
$[x,y]=\langle x,Ny\rangle$, where $N\in \mathbb{R}^{n\times n}$ and
$N=N^{-1}$. Such a matrix $N$ is called weight. A space with an
indefinite inner product is called an indefinite inner product
space. In the rest of the paper $\mathbb{R}^{m}, \mathbb{R}^{n}$
denote indefinite inner product spaces with weights $M,N$
respectively. Let $A,B$ be two real matrices of order $m \times n$
and $n \times l$ respectively, then the indefinite matrix product of
those matrices be denoted by $A\circ B$ and defined as $A\circ
B=ANB$, where $N$ is a weight matrix as defined earlier. For $A\in
\mathbb{R}^{m\times n}$, the adjoint $A^{[*]}$, of $A$ is defined
by $A^{[*]}=NA^{*}M$, where * denotes the transpose of $A$, $M$ and
$N$ are weights of order m and n respectively.
Let $K$ be a subset
of $\mathbb{R}^n$. Then $K$ is called cone if (i) $x, y\in
K\Rightarrow x+y\in K$ and (ii)$x\in K$, $\alpha\in \mathbb{R}$,
$\alpha\geq0\Rightarrow \alpha x\in K$. The dual of cone $K$ is
denoted by $K^{[*]}$ and is defined as $K^{[*]}=\{x\in
\mathbb{R}^n:[x, t]\geq0,~ \text{for all}~ t\in K\}$. Let
$K^{[*][*]}$ denote $({K^{[*]}})^{[*]}$. If $K=\mathbb{R}^n_{+}$
then $K^{[*]}=I\circ \mathbb{R}^n_{+}$ and $K^{[*][*]}=K$.
A cone $C$ is said to be acute if $[x, y]\geq0$ for all $x, y\in C$.
$C$ is said to be obtuse if $C^{[*]}\cap\{cl~span~C\}$ is acute. In
particular, let $C=A\circ I\circ K$ then we say that $C=\{A\circ
I\circ x:x\in K\}$ is obtuse if $(A\circ I\circ
K)^{[*]}\cap\mathcal{R}(A\circ I)$ is acute. According to Novikoff,
the acuteness of a cone $C$ in $\mathbb{R}^n$ is defined by the
inclusion $C\subseteq C^{*}$. We can easily verify this condition in
indefinite inner product spaces as $C\subseteq C^{[*]}$.
For $A\in \mathbb{R}^{m\times n}$, $A^{[*]}\circ A$
will be called the Gram matrix of $A$. For $A\in \mathbb{R}^{m\times n}$, the following equations are known to have unique solution \cite{krkkkcs}:\\
$A\circ X\circ A=A$, $X\circ A\circ X=X$, $(A\circ
X)^{[*]}=A\circ X$, $(X\circ A)^{[*]}=X\circ A$. Such an $X$ will
be denoted by $A^{[\dagger]}$. If the weight matrices in indefinite
inner product spaces are equal to identity then
$A^{[\dagger]}=A^{\dagger}$. We refer the reader \cite{bg} (and the
references cited there in)
for a detailed study of
$A^{\dagger}$.
Next, we collect some properties of $A^{[\dagger]}$. Some of these
have been proved in \cite{krkkkcs} and rest can be demonstrated
easily. The range space of $A$, $\mathcal{R}(A)$ is defined by
$\mathcal{R}(A)=\{y\in \mathbb{R}^m: y=A\circ x, x\in
\mathbb{R}^n\}$ and the null space of $A$, $\mathcal{N}(A)$ is
defined by $\mathcal{N}(A)=\{x\in\mathbb{R}^n: A\circ x=0\}$. For
$A\in \mathbb{R}^{m\times n}$,
$A^{[\dagger]}$ also satisfies the following properties:
$\mathcal{R}(A^{[*]})=\mathcal{R}(A^{[\dagger]})$,
$\mathcal{N}(A^{[*]}) =\mathcal{N}(A^{[\dagger]})$, $A\circ
A^{[\dagger]}=P_{\mathcal{R}(A)}$, $A^{[\dagger]}\circ
A=P_{\mathcal{R}(A^{[*]})}$. We also have $(A^{[*]}\circ
A)^{[\dagger]}\circ (A^{[*]}\circ A)=P_{\mathcal{R}(A^{[*]}\circ
A)^{[\dagger]}}=P_{\mathcal{R}(A^{[*]}\circ
A)^{[*]}}=P_{\mathcal{R}(A^{[*]})} =A^{[\dagger]}\circ A$.
\begin{lem}
Let $A\in \mathbb{C}^{m\times n}$. Then\\
(i)$A^{[*]}=A^{[*]}\circ A\circ A^{[\dagger]}=A^{[\dagger]}\circ A\circ A^{[*]}$\\
(ii)$A^{[\dagger]}=A^{[*]}\circ (A\circ A^{[*]})^{[\dagger]}=(A^{[*]}\circ A)^{[\dagger]}\circ A^{[*]}$\\
(iii)$A^{[\dagger]}\circ (A^{[\dagger]})^{[*]}=(A^{[*]}\circ A)^{[\dagger]}$\\
(iv)$(A\circ I)^{[\dagger]}=I\circ A^{[\dagger]}$\\
(v)$\mathcal{R}(A\circ A^{[\dagger]})=\mathcal{R}(A)$,
$\mathcal{R}(A^{[\dagger]}\circ A)=\mathcal{R}(A^{[*]})$,
$\mathcal{N}(A\circ A^{[\dagger]})=\mathcal{N}(A^{[*]})$,
$\mathcal{N}(A^{[\dagger]}\circ A)=\mathcal{N}(A)$ where
$\mathcal{R}(X)$ and $\mathcal{N}(X)$ denote the range and null
spaces of X respectively.
\end{lem}
We use the following lemma frequently in this paper.
\begin{lem}\label{lineq}
Let $A\in \mathbb{R}^{m\times n}$ and $b\in \mathbb{R}^m$. Then, the
linear equation $A\circ X=b$ has a solution iff $b\in \mathcal{R}(A)
$. In this case, the general solution is given by
$x=A^{[\dagger]}\circ b+z$ where $z\in \mathcal{N(A)}$.
\end{lem}
\section{Main Results}
For given $A\in \mathbb{R}^{m\times n}$, Ramanathan and Sivakumar
\cite{krksgram} derived a set of necessary and sufficient conditions
for a cone to be invariant under $(A^{[*]}\circ A)^{[\dagger]}.$
These conditions include pairwise acuteness (or pairwise obtuseness)
of certain cones. In this article, we avoid pairwise acuteness of
cones and characterize Moore-Penrose inverses of Gram matrices
leaving a cone invariant in the approach of Sivakumar \cite{s2gram}.
These results generalize the existing results of Sivakumar
\cite{s2gram} in the finite dimensional setting from Euclidean
spaces to indefinite inner product spaces. First we prove series of
lemmas that lead up to the main theorem (Theorem \ref{main}).
As mentioned earlier; $\mathbb{R}^m, \mathbb{R}^n$ denote
indefinite inner product spaces with weights $M,N$ respectively. Let
$A\in\mathbb{R}^{m\times n}$ be such that $I\circ A=A\circ I$ that
is $MA=AN$ and let $K$ be a closed cone in $\mathbb{R}^n$.
\begin{lem}\label{lem1}
$[A\circ x, y]=[x, A^{[*]}\circ y]$.
\end{lem}
\begin{proof}
$[A\circ x, y]=\langle A\circ x, My\rangle=\langle ANx,
My\rangle=\langle x, NA^{*}My\rangle=[x, A^{*}My]= [x, I\circ
(NA^{*}M)\circ I\circ y]=[x, I\circ A^{[*]}\circ I\circ y]=[x,
(I\circ A\circ I)^{[*]}\circ y]=[x, A^{[*]}\circ y]. $
\end{proof}
\begin{lem}\label{lemma2}
$u\in (A\circ I\circ K)^{[*]}\Rightarrow(A\circ I)^{[*]}\circ u\in {K}^{[*]}$.
\end{lem}
\begin{proof}
Let $u\in (A\circ I\circ K)^{[*]}$ and $r\in {K}$. Then $0\leq [u,
A\circ I\circ r]=[(A\circ I)^{[*]}\circ u, r]$, by Lemma
\ref{lem1}.
Thus $(A\circ I)^{[*]}\circ u\in K^{[*]}$.
\end{proof}
Next, we show that $K$ is invariant under $A^{[\dagger]}\circ A$ if
and only if $K^{[*]}$ is invariant under $A^{[\dagger]}\circ A$.
\begin{lem}\label{AKdualAK}
$A^{[\dagger]}\circ A\circ K\subseteq K \Leftrightarrow
A^{[\dagger]}\circ A\circ K^{[*]}\subseteq K^{[*]}$.
\end{lem}
\begin{proof}
Let $A^{[\dagger]}\circ A\circ K\subseteq K$, $y=A^{[\dagger]}\circ A\circ x$ with $x\in
K^{[*]}$,
$u\in K ~\text{and}~u^{1}=A^{[\dagger]}\circ A\circ u\in K$. Then
$[y, u]=[A^{[\dagger]}\circ A\circ x, u]=[x, (A^{[\dagger]}\circ A)^{[*]}\circ u]=[x, A^{[\dagger]}\circ A\circ u]=[x, u^{1}]\geq0$.
This shows that $y\in K^{[*]}$. Hence $A^{[\dagger]}\circ A\circ K^{[*]}\subseteq K^{[*]}$.
Similarly one can easily prove the converse part.
\end{proof}
In the next result, we determine the set $(A\circ I\circ K)^{[*]}$
in the presence of an additional condition.
\begin{thm}\label{dual of Aok}
$(A\circ I\circ K)^{[*]}\subseteq(A^{[\dagger]})^{[*]}\circ I\circ
K^{[*]}+\mathcal{N}((A\circ I)^{[*]})$.
If $K$ is invariant under $A^{[\dagger]}\circ A$, then equality holds.
\end{thm}
\begin{proof}
Let $y\in (A\circ I\circ K)^{[*]}$. Then by Lemma $\ref{lemma2}$, $z=(A\circ I)^{[*]}\circ y\in {K}^{[*]}. $
By Lemma \ref{lineq}, $y=((A\circ I)^{[*]})^{[\dagger]}\circ z+w$
for some $w\in \mathcal{N}((A\circ I)^{[*])}. $ Then $y\in ((A\circ
I)^{[*]})^{[\dagger]}\circ K^{[*]}+\mathcal{N}((A\circ I)^{[*]})=
(A^{[\dagger]})^{[*]}\circ I\circ K^{[*]}+\mathcal{N}((A\circ
I)^{[*]})$. This proves the first part.
Next, suppose that
$A^{[\dagger]}\circ A\circ K\subseteq K$. Let $u=u^{1}+u^{2}$, where
$u^{1}=(A^{[\dagger]})^{[*]}\circ I\circ l$ with $l\in K^{[*]}$ and
$u^{2}\in \mathcal{N}((A\circ I)^{[*]}). $ Let $v=A\circ I\circ t$,
$t\in K$ and set $t^{'}=A^{[\dagger]}\circ A\circ t\in K$. Then $[u,
v]=[u^{1}+u^{2}, v]=[u^{1}, v]+[u^{2}, v]=[u^{1}, A\circ I\circ
t]=[(A^{[\dagger]})^{[*]}\circ I\circ l, A\circ I\circ t]=[l,
t^{'}]\geq 0$, since $[u^{2}, v]=[u^{2}, A\circ I\circ t]=0$. Thus
$u\in (A\circ I\circ K)^{[*]}$.
\end{proof}
\begin{rem}\label{remarkI}
The following example shows that in the absence of the condition
$A^{[\dagger]}\circ A\circ K\subseteq K$, the reverse inclusion may
not hold in Theorem \ref{dual of Aok}.
Let $A = \begin{pmatrix} 1 & 0 & 0\\
0 & -1 & 1
\end{pmatrix}, $
$M=\begin{pmatrix} 1 & 0\\
0 & -1
\end{pmatrix} $ and
$N =\begin{pmatrix} 1 & 0 & 0\\
0 &-1 & 0\\
0 & 0 &-1
\end{pmatrix}$.
Then $A^{\dagger} =\displaystyle\frac{1}{2}\begin{pmatrix} 2 & 0 \\
0 & -1 \\
0 & 1
\end{pmatrix} $ and
$A^{[\dagger]}=NA^{\dagger}M =\displaystyle\frac{1}{2}\begin{pmatrix} 2 & 0 \\
0 & -1 \\
0 & 1 \\
\end{pmatrix}$.
Let $K=\mathbb{R}^{3}_{+}$ then $K^{[*]}=N\mathbb{R}^{3}_{+}$.
Suppose $x=(1,2,3)^t$. Then $A^{[\dagger]}\circ A\circ x=\displaystyle(1,\frac{-1}{2},\frac{1}{2})\notin K.$
So, $A^{[\dagger]}\circ A\circ K\nsubseteq K$. Also $A\circ I = \begin{pmatrix} 1 & 0 & 0\\
0 & 1 & -1
\end{pmatrix}$. So $\mathcal{N}((A\circ I)^{[*]})$ contins only the zero
vector. Let $y=(1,2,0)^t\in K$ then $y^{1}=Ny=(1,-2,0)\in
K^{[*]}$. Then $u=(A^{[\dagger]})^{[*]}\circ I\circ y^{1}=(1,1)^t\in
(A^{[\dagger]})^{[*]}\circ I\circ K^{[*]}$. But $u\notin (A\circ I\circ K)^{[*]}$,
since $[u, A\circ I\circ v]=\langle u, MAv \rangle<0$ for $v=(1,4,8)^t$.
\end{rem}
The next result is analogous to Theorem \ref{dual of Aok}. This will
be used later.
\begin{lem}\label{analogus}
$ ((A^{[\dagger]})^{[*]}\circ I\circ K^{[*]})^{[*]}\subseteq A\circ
I\circ K+\mathcal{N}((A\circ I)^{[*]})$. If $A^{[\dagger]}\circ
A\circ K\subseteq K$, then equality holds.
\end{lem}
\begin{proof}
The proof follows from Lemma \ref{AKdualAK} and Theorem \ref{dual of
Aok} by replacing $A$ by $((A^{[\dagger]})^{[*]}$ and $K$ by
$K^{[*]}$.
\end{proof}
\begin{rem}
Let A be given as in Remark \ref{remarkI}. Then $A^{[\dagger]}\circ
A\circ K\nsubseteq K$. Let $y=(2,5,8)^t\in K $ and set
$y^{1}=A\circ I\circ y=Ay=(2,3)^t\in A\circ I\circ K$. Let
$v=N(1,2,0)^t=(1,-2,0)^t\in K^{[*]}$ and
$z=(A^{[\dagger]})^{[*]}\circ I\circ v=
(1,1)^t\in(A^{[\dagger]})^{[*]}\circ I\circ K^{[*]}.$ Then
$[y^{1}, z]=\langle y^{1}, Mz \rangle=\langle (2, 3)^t, (1, -1)^t
\rangle<0$,
so that $y^{1}\notin ((A^{[\dagger]})^{[*]}\circ I\circ K^{[*]})^{[*]}$.
This shows that the condition $A^{[\dagger]}\circ A\circ K\subseteq K$ is essential for the reverse inclusion to hold in Lemma \ref{analogus}.
\end{rem}
\begin{lem}\label{intersection}
$(A\circ I\circ K)^{[*]}\cap \mathcal{R}(A\circ I)\subseteq
(A^{[\dagger]})^{[*]}oI\circ K^{[*]}$. If $A^{[\dagger]}\circ A\circ
K\subseteq K$, then equality holds in the above inclusion.
\end{lem}
\begin{proof}
Let $y=A\circ I\circ x\in (A\circ I\circ K)^{[*]}$. Then by Lemma
\ref{lemma2},
$(A\circ I)^{[*]}\circ y\in K^{[*]}.$ Also, $y=(A\circ I)\circ (A\circ I)^{[\dagger]}\circ y=((A\circ I)\circ (A\circ I)^{[\dagger]})^{[*]}\circ y
=((A\circ I)^{[\dagger]})^{[*]}\circ (A\circ I)^{[*]}\circ y
=(A^{[\dagger]})^{[*]}\circ I\circ (A\circ I)^{[*]}\circ y\in(A^{[\dagger]})^{[*]}\circ I\circ K^{[*]}$,
proving that $(A\circ I\circ K)^{[*]}\cap \mathcal{R}(A\circ I)\subseteq (A^{[\dagger]})^{[*]}\circ I\circ K^{[*]}$.
Conversely, suppose that $x\in (A^{[\dagger]})^{[*]}\circ I\circ K^{[*]}$. Then $x=((A\circ I)^{[\dagger]})^{[*]}\circ u$ for some $u\in K^{[*]}$.
This implies $x\in \mathcal{R}(A\circ I)$. Let $w\in K$, $v=A\circ I\circ w\in
A\circ I\circ K$ and $w^{1}=A^{[\dagger]}\circ A\circ w\in K$. Then we have $[x, v]=[(A^{[\dagger]})^{[*]}\circ I\circ u, A\circ I\circ w]=[u, A^{[\dagger]}\circ A\circ w]=[u,
w^{1}]\geq0$.
Thus $x\in (A\circ I\circ K)^{[*]}$.
\end{proof}
\begin{rem}
Let A be given as in Remark \ref{remarkI}. Then $A^{[\dagger]}\circ
A\circ K\nsubseteq K$. Let $y=(2,5,8)^t\in K$, and
$y^{1}=A\circ I\circ y=Ay=(2,3)^t\in A\circ I\circ K$. Let
$v=N(1,2,0)^t=(1,-2,0)^t\in K^{[*]}$ and
$z=(A^{[\dagger]})^{[*]}\circ I\circ v=(1,1)^t\in
(A^{[\dagger]})^{[*]}\circ I\circ K^{[*]}$. But
$[y^{1}, z]=\langle y^{1}, Mz \rangle=\langle (2, 3)^t,(1,-1)^t \rangle<0.$
Thus $z \notin (A\circ I\circ K)^{[*]}\cap \mathcal{R}(I\circ A)$.
Hence the condition $A^{[\dagger]}\circ A\circ K\subseteq K$ is necessary for the reverse inclusion to hold in Lemma \ref{intersection}.
\end{rem}
Next, we obtain an equivalent condition for the acuteness of the
cone $(A\circ I\circ K)^{[*]}\cap\mathcal{R}(A\circ I)$.
\begin{lem}\label{acute}
Let $A^{[\dagger]}\circ A\circ K\subseteq K$. Then $(A\circ I\circ
K)^{[*]}\cap\mathcal{R}(A\circ I)$ is acute $\Leftrightarrow(A\circ
I\circ k)^{[*]}\cap\mathcal{R}(A\circ I)\subseteq A\circ I\circ K$.
\end{lem}
\begin{proof}
Suppose that $L=(A\circ I\circ K)^{[*]}\cap\mathcal{R}(A\circ I)$ is
acute. Then $L\subseteq L^{[*]}$. By Lemma \ref{analogus} and Lemma
\ref{intersection}, it follows that $L^{[*]}=((A\circ I\circ
K)^{[*]}\cap\mathcal{R}(A\circ I))^{[*]}
=((A^{[\dagger]})^{[*]}\circ I\circ K^{[*]})^{[*]}=A\circ I\circ K+\mathcal{N}((A\circ I)^{[*]}).$
So, $(A\circ I\circ K)^{[*]}\cap\mathcal{R}(A\circ I)\subseteq
A\circ I\circ K+\mathcal{N}((A\circ I)^{[*]}).$ But, we have to
show that $(A\circ I\circ K)^{[*]}\cap\mathcal{R}(A\circ I)\subseteq
A\circ I\circ K.$ Let $x\in (A\circ I\circ
K)^{[*]}\cap\mathcal{R}(A\circ I)$. Then $x=A\circ I\circ u+z$, with
$u\in K$, $z\in \mathcal{N}((A\circ I)^{[*]})$. But since $x, A\circ
I\circ u\in \mathcal{R}(A\circ I)$, it follows that
$z\in \mathcal{R}(A\circ I)\cap \mathcal{N}((A\circ I)^{[*]})=\{0\}.$
Thus $x\in A\circ I\circ K$.
Conversely, let $x,y\in
(A\circ I\circ K)^{[*]}\cap\mathcal{R}(A\circ I)\subseteq A\circ
I\circ K$. Then $x=A\circ I\circ u,~u\in K.$ We also have $(A\circ
I)^{[*]}\circ y\in K^{[*]}$. Now, $[x,y]=[A\circ I\circ
u,y]=[u,(A\circ I)^{[*]}\circ y]\geq 0.$ Thus $(A\circ I\circ
k)^{[*]}\cap\mathcal{R}(A\circ I)$ is acute.
\end{proof}
We next obtain a necessary and sufficient condition for a cone to be
invariant under $(A^{[*]}\circ A)^{[\dagger]}$ (See Lemma
\ref{nonnegativity}).
\begin{lem}\label{lemma 11}
$(A^{[\dagger]})^{[*]}\circ I\circ K^{[*]}\subseteq A\circ I\circ
K+\mathcal{N}((A\circ I)^{[*]})$ $\Leftrightarrow (A^{[*]}\circ
A)^{[\dagger]}\circ K^{[*]}\subseteq K+\mathcal{N}(A\circ I)$
\end{lem}
\begin{proof}
For $x\in K^{[*]}$, let $y=(A^{[*]}\circ A)^{[\dagger]}\circ
x=((A\circ I)^{[*]}\circ (A\circ I))^{[\dagger]}\circ x$ $= (A\circ
I)^{[\dagger]}\circ \linebreak((A\circ I)^{[\dagger]})^{[*]}\circ
x.$ Then
\begin{align*}
A\circ I\circ y &=(A\circ I)\circ (A\circ I)^{[\dagger]}\circ ((A\circ I)^{[\dagger]})^{[*]}\circ x\\
&=((A\circ I)^{[\dagger]}\circ (A\circ I)\circ (A\circ I)^{[\dagger]})^{[*]}\circ x\\
&=((A\circ I)^{[\dagger]})^{[*]}\circ x \\
&=(A^{[\dagger]})^{[*]}\circ I\circ x\in
(A^{[\dagger]})^{[*]}\circ I\circ K^{[*]}\\
&\subseteq
A\circ I\circ K+\mathcal{N}((A\circ I)^{[*]})
\end{align*}
Therefore $A\circ I\circ y=A\circ I\circ v+w$, $v\in k$,
$w\in\mathcal{N}((A\circ I)^{[*]})$. So , $A\circ I\circ (y-v)\in
\mathcal{R}(A\circ I)\cap\mathcal{N}((A\circ I)^{[*]})=\{0\}$. Then
$A\circ I\circ (y-v)=0$. This implies, $ y-v=u\in\mathcal{N}(A\circ
I)$. Then $y=u+v$, $v\in K$, $u\in \mathcal{N}(A\circ I).$ This
shows that $(A^{[*]}\circ A)^{[\dagger]}\circ K^{[*]}\subseteq
K+\mathcal{N}(A\circ I)$.
Conversely, let $y=(A^{[\dagger]})^{[*]}\circ I\circ x,~x\in
K^{[*]}.$ Then $y=((A\circ I)^{[\dagger]})^{[*]}\circ x$ and
$(A\circ I)^{[\dagger]}\circ y=(A\circ I)^{[\dagger]}\circ ((A\circ
I)^{[\dagger]})^{[*]}\circ x=((A\circ I)^{[*]}\circ (A\circ
I))^{[\dagger]}\circ x$ $=(A^{[*]}\circ A)^{[\dagger]}\circ x=u+v$,
$u\in K,v\in \mathcal{N}(A\circ I).$ Then $y=((A\circ
I)^{[\dagger]})^{[\dagger]}\circ (u+v)+w$, $w\in \mathcal{N}((A\circ
I)^{[\dagger]})$ Then $y=A\circ I\circ u+w\in A\circ I\circ
K+\mathcal{N}((A\circ I)^{[*]})$.
\end{proof}
We also have stronger one-way implication, given below. The proof
follows from necessity part of Lemma \ref{lemma 11}.
\begin{lem}\label{lemma12}
$(A^{[\dagger]})^{[*]}\circ I\circ K^{[*]}\subseteq A\circ I\circ K$
$\Rightarrow (A^{[*]}\circ A)^{[\dagger]}\circ K^{[*]}\subseteq
K+\mathcal{N}(A\circ I).$
\end{lem}
\begin{lem}\label{lemma13}
$(A^{[*]}\circ A)^{[\dagger]}\circ K^{[*]}\subseteq
K+\mathcal{N}(A\circ I)$ $\Rightarrow K^{[*]}\cap \mathcal{R}(A\circ
I)^{[*]}\subseteq A^{[*]}\circ A\circ K+\mathcal{N}((A\circ I)$
\end{lem}
\begin{proof}
Let $y=(A\circ I)^{[*]}\circ x\in K^{[*]}$. Then $(A^{[*]}\circ
A)^{[\dagger]}\circ y=u+z$, $u\in K$, $z\in \mathcal{N}(A\circ I).$
From this $y=(A^{[*]}\circ A)\circ (u+z)+w$, $w\in
\mathcal{N}(A^{[*]}\circ A)^{[\dagger]}.$ Since $A^{[*]}\circ
A=(A\circ I)^{[*]}\circ (A\circ I)$ and $z\in \mathcal{N}(A\circ
I)$, we get $y=A^{[*]}\circ A\circ u+w\in A^{[*]}\circ A\circ
K+\mathcal{N}(A\circ I)$
\end{proof}
\begin{lem}\label{nonnegativity}
Suppose that $A^{[\dagger]}\circ A\circ K\subseteq K$. Then
$(A^{[*]}\circ A)^{[\dagger]}\circ K^{[*]}\subseteq
K+\mathcal{N}(A\circ I)\Leftrightarrow (A^{[*]}\circ
A)^{[\dagger]}\circ K^{[*]}\subseteq K$
\end{lem}
\begin{proof}
It is enough to show the necessity part. Let $x\in K^{[*]}$ and
$y=(A^{[*]}\circ A)^{[\dagger]}\circ x.$ Then $(A^{[*]}\circ
A)^{[\dagger]}\circ x=u+v$ where $u\in K$, $v\in \mathcal{N}(A\circ
I)$. This implies $x=(A^{[*]}\circ A)\circ (u+v)+w,~w\in
\mathcal{N}(A\circ I)$, so that $y=(A^{[*]}\circ A)^{[\dagger]}\circ
(A^{[*]}\circ A)\circ u=A^{[\dagger]}\circ A\circ u\in K$.
\end{proof}
\begin{rem}
Let $A$ be as given in Remark \ref{remarkI} and let
$K=\mathbb{R}^{3}_{+}$. Then $K^{[*]}=NK^{*}=N\mathbb{R}^{3}_{+}$,
$A\circ I = \begin{pmatrix} 1 & 0 & 0\\
0 & 1 & -1
\end{pmatrix}$,
$\mathcal{N}(A\circ I)=span\{(0,1,1)^t\}$ and
$\mathcal{R}(A\circ I)^{[*]}=\{(x,-y,y)^t: x,y\in \mathbb{R}\}$.
Also,
$ A^{[*]}\circ A=\begin{pmatrix} 1 & 0 & 0\\
0 &-1 & 1\\
0 & 1 &-1
\end{pmatrix}$. So,
$(A^{[*]}\circ A)^{\dagger}=\displaystyle\frac{1}{4} \begin{pmatrix} 4 & 0 & 0\\
0 &-1 & 1\\
0 & 1 &-1
\end{pmatrix}.$
Let $x^{1}=(x,y,z)^t\in K^*$,
then $(A^{[*]}\circ A)^{[\dagger]}\circ Nx^{1}=\displaystyle\frac{1}{4}\begin{pmatrix} 4 x\\
-y+z \\
y-z
\end{pmatrix}\in (A^{[*]}\circ A)^{[\dagger]}\circ K^{[*]}$.
\newline
Since
$(4x,-y+z,y-z)^t=(4x,b,c)^t-\left(\displaystyle\frac{b+c}{2}\right)(0,1,1)^t$
where
$b,c\geq0$ such that $\displaystyle\frac{b-c}{2}=-y+z$,nn $(4x,-y+z,y-z)^t\in K+\mathcal{N}(A\circ I).$
Thus $(A^{[*]}\circ A)^{[\dagger]}\circ K^{[*]}\subseteq
K+\mathcal{N}(A\circ I)$.
But for $x^{1}= (1,2,3)\in K^{*}$, $Nx^{1}\in K^{[*]}$
and $(A^{[*]}\circ A)^{[\dagger]}\circ Nx^{1}=\displaystyle\frac{1}{4}(4,1,-1)^t\notin K$
Thus $(A^{[*]}\circ A)^{[\dagger]}\circ K^{[*]}\nsubseteq K$. Hence we
can conclude that in the absence of the condition $A^{[\dagger]}\circ A\circ K\subseteq
K$, Lemma \ref{nonnegativity} may not be true.
\end{rem}
We are now in a position to prove the main result of this article.
\begin{thm}\label{main}(Main Result)
Let $A\in \mathbb{R}^{m\times n}$ with $\mathcal{R}(A\circ I)$
closed, K be a closed in $\mathbb{R}^{n}$ with $A^{[\dagger]}\circ
A\circ K\subseteq K$. Let $C=A\circ I\circ K$ and
$D=(A^{[\dagger]})^{[*]}\circ I\circ K^{[*]}.$ Then the following
conditions are equivalent:
\newline
(i) $D$ is acute.
\newline
(ii) $(A^{[*]}\circ A)^{[\dagger]}\circ K^{[*]}\subseteq
K+\mathcal{N}(A\circ I)$.
\newline
(iii) $C$ is obtuse.
\end{thm}
\begin{proof}
$(i)\Rightarrow(ii)$:
\newline
Suppose $D$ is acute then by definition, $D\subseteq D^{[*]}.$ By
Lemma \ref{analogus}, $D^{[*]}=A\circ I\circ K+\mathcal{N}(A\circ
I)^{[*]}.$ Thus $D\subseteq A\circ I\circ K+\mathcal{N}(A\circ
I)^{[*]}.$ Now, by Lemma \ref{lemma 11}, we obtain $(A^{[*]}\circ
A)^{[\dagger]}\circ K^{[*]}\subseteq
K+\mathcal{N}(A\circ I)$.\\
$(ii)\Rightarrow(i)$:
\newline
Suppose
$(A^{[*]}\circ A)^{[\dagger]}\circ K^{[*]}\subseteq
K+\mathcal{N}(A\circ I).$ By Lemma \ref{lemma 11}, $D\subseteq
A\circ I\circ K+\mathcal{N}((A\circ I)^{[*]})$. But by Lemma
\ref{analogus}, $A\circ I\circ K+\mathcal{N}((A\circ
I)^{[*]})=D^{[*]}.$ So, $D\subseteq D^{[*]}. $ Hence $D$ is acute.
\newline$(ii)\Rightarrow(iii)$ Suppose
$(A^{[*]}\circ A)^{[\dagger]}\circ K^{[*]}\subseteq
K+\mathcal{N}(A\circ I).$ Note that $C=A\circ I\circ K$ is obtuse if
$C^{[*]}\cap\mathcal{R}(A\circ I)$ is acute. By Lemma \ref{acute},
it is enough to show that $C^{[*]}\cap\mathcal{R}(A\circ I)\subseteq
C.$
Let $y\in C^{[*]}\cap\mathcal{R}(A\circ I).$ Then $y=A\circ I\circ
x$ and by Lemma \ref{lemma2}, $(A\circ I)^{[*]}\circ y\in K^{[*]}$.
So, $(A\circ I)^{[*]}\circ y\in K^{[*]}\cap \mathcal{R}(A\circ
I)^{[*]}.$ By Lemma \ref{lemma13}, $(A\circ I)^{[*]}\circ
y=A^{[*]}\circ A\circ u+z$ with $u\in K$, $z\in \mathcal{N}(A\circ
I)$. Since $A^{[*]}\circ A=(A\circ I)^{[*]}\circ (A\circ I)$, it
follows that $(A\circ I)^{[*]}\circ y,A^{[*]}\circ A\circ u\in
\mathcal{R}(A\circ I)^{[*]}$. Thus $z\in \mathcal{R}(A\circ
I)^{[*]}\cap \mathcal{N}(A\circ I)=\{0\}$. This implies $z=0$. Then
$(A\circ I)^{[*]}\circ y=A^{[*]}\circ A\circ u$. From this,
\begin{align*}
y&=((A\circ I)^{[\dagger]})^{[*]}\circ ((A\circ I)^{[*]}\circ A\circ I\circ u)+w\\
&=((A\circ I)\circ (A\circ I)^{[\dagger]})^{[*]}\circ (A\circ I)\circ u+w\\
&= (A\circ I)\circ (A\circ I)^{[\dagger]}\circ (A\circ I)\circ u+w\\
&=(A\circ I)\circ u+w,
\end{align*}
where $w\in \mathcal{N}((A\circ I)^{[*]})$.
\newline Since $y\in
\mathcal{R}(A\circ I)$, it follows that $w\in \mathcal{R}(A\circ I)\cap
\mathcal{N}(A\circ I)^{[*]})
=\{0\}$. Thus $y\in A\circ I\circ K=C$.\\
$(iii)\Rightarrow(ii)$: \newline Let $C=A\circ I\circ K$ be obtuse.
Then by definition, $C^{[*]}\cap\mathcal{R}(A\circ I)\subseteq C$.
By Lemma \ref{intersection}, $(A^{[\dagger]})^{[*]}\circ I\circ
K^{[*]}\subseteq C$. Now by Lemma \ref{lemma12}, $(A^{[*]}\circ
A)^{[\dagger]}\circ K^{[*]}\subseteq K+\mathcal{N}(A\circ I). $
\end{proof}
\begin{rem}\hfill
\begin{enumerate}
\item [(i)]
The following exmaple illustrates Thorem \ref{main}.
Let $A=\begin{pmatrix} 1 & 0 & 1\\
1 &0 & 1
\end{pmatrix}$,
$M=\begin{pmatrix} 0 & 1\\
1 &0 \\
\end{pmatrix}$,
$N=\begin{pmatrix} 0 & 0 & 1\\
0 &1 & 0\\
1 &0 & 0
\end{pmatrix}$ and $K=\mathbb{R}^{3}_{+}$.
Then $A^{\dagger}=\displaystyle\frac{1}{4}\begin{pmatrix} 1&1\\
0 & 0\\
1 & 1
\end{pmatrix}$,
$A^{[\dagger]}=NA^{\dagger}M=\displaystyle{\frac{1}{4}\begin{pmatrix} 1&1\\
0 & 0\\
1 & 1
\end{pmatrix}}$ and
$K^{[*]}=N\mathbb{R}^{3}_{+}$.
Note that for $x^{1}=(x,y,z)^t\in K$, $A^{[\dagger]}\circ A\circ
x^{1}=A^{[\dagger]}Ax^{1}=\displaystyle\frac{1}{2} (x+z,0,x+z)^t\in
K$.
Thus $A^{[\dagger]}0A\circ K\subseteq K$.
And
$(A^{[*]}\circ A)^{\dagger}=\displaystyle{\frac{1}{16}}\begin{pmatrix} 2 & 0 & 2\\
0 & 0 & 0\\
2 & 0 & 2
\end{pmatrix}$. Therefore
$(A^{[*]}\circ A)^{[\dagger]}\circ K^{[*]}=N(A^{[*]}\circ A)^{\dagger}NK^{[*]}\subseteq
K$. Also one can easily verify that $C=A\circ I\circ K$ is obtuse and
$D=(A^{[\dagger]})^{[*]}\circ I\circ K^{[*]}$ is acute.
\item [(ii)] Here, we show by an example that in the absense of the condition $A\circ I=I\circ A$, Theorem \ref{main} may not hold.
Let $ A=\begin{pmatrix} 0 & 1\\
0 & 1
\end{pmatrix}$,
$ M=\begin{pmatrix} 0 & 1\\
1 & 0 \\
\end{pmatrix}=N$. Then clearly $A\circ I\neq I\circ A.$
Let $K=\{(x,0):x\geq0 \}$ then $K^{*}=\{(x,y):x\geq0,y\in \mathbb{
R}\}$ and $K^{[*]}=\{(y,x):x\geq0,~y\in \mathbb{R}\}$. Also,
$A^{\dagger}=\displaystyle\frac{1}{2}\begin{pmatrix} 0 & 0\\
1 & 1 \\
\end{pmatrix}$ and
$A^{[\dagger]}=\displaystyle\frac{1}{2}\begin{pmatrix} 1 & 1\\
0 & 0
\end{pmatrix}$.
Clearly $A^{[\dagger]}\circ A\circ K\subseteq K$ and
$D=\left\{(\frac{x}{2},\frac{x}{2}): x\geq 0\right\}$ is acute but
$(A^{[*]}\circ A)^{[\dagger]}\circ K^{[*]}\nsubseteq K$
where $(A^{[*]}\circ A)^{[\dagger]}=\displaystyle\frac{1}{4}\begin{pmatrix} 0 & 2\\
0 & 0 \\
\end{pmatrix}.$
\item [(iii)] For given $A\in \mathbb{R}^{m\times n}$, Ramanathan and Sivakumar
\cite{krksgram} derived a set of necessary and sufficient conditions
for a cone to be invariant under $(A^{[*]}\circ A)^{[\dagger]}$ in
terms of pairwise acuteness of cones $D$ and $I\circ D$ in
indefinite inner product space. We would like to remark here that
pairwise acuteness of $D$ and $I\circ D$ is same as acuteness of the
cone $D$ in usual inner product space.
\end{enumerate}
\end{rem}
\vspace{.5cm} \noindent \textbf{Acknowledgements}: We thank
Prof.K.C. Sivakumar for his valuable comments and suggestions to
improve this article.
|
\section*{Executive Summary}
The discovery of planets around other stars is revolutionizing our
notions of planet formation and is poised to do the same for planetary
climate. Studying \emph{transiting} planets is complementary to eventual studies of directly-imaged planets: 1) we can readily measure the mass and radius of transiting planets, linking atmospheric properties to bulk composition and formation, 2) many transiting planets are strongly irradiated and exhibit novel atmospheric physics, and 3) the most common temperate terrestrial planets orbit close to red dwarf stars and are difficult to image directly. We have only been able to comprehensively
characterize the atmospheres of a handful of transiting planets,
because most orbit faint stars. The Transiting Exoplanet Survey Satellite (TESS) will discover
transiting planets orbiting the brightest stars, enabling, in
principle, an atmospheric survey of $10^{2}$--$10^{3}$
bright hot Jupiters and warm sub-Neptunes. Uniform
observations of such a statistically significant sample would provide
leverage to understand---and learn from---the diversity of short-period
planets, and would identify the minority of truly special planets
worthy of more intensive follow-up.
We argue that the best way to maximize the scientific returns of TESS is
to adopt a triage approach. A space mission consisting of a $\sim$1 m
telescope with an optical--NIR spectrograph could
measure molecular absorption for non-terrestrial planets discovered by TESS,
as well as eclipses and phase variations for the hottest jovians. Such a mission could observe up to $10^{3}$
transits per year, thus enabling it to survey a large
fraction of the bright ($J<11$) hot-Jupiters and warm sub-Neptunes TESS
is expected to find. The James Webb Space Telescope (JWST) could be used to perform detailed atmospheric characterization of the most interesting transiting targets (transit, eclipse, and---when possible---phase-resolved spectroscopy). TESS is also expected to discover
a few temperate terrestrial planets transiting nearby M-Dwarfs.
Characterizing these worlds will be time-intensive: JWST will need months to provide tantalizing constraints on the presence of an atmosphere, planetary rotational state, clouds, and greenhouse gases. Future flagship missions should be designed to provide better constraints on the habitability of M-Dwarf temperate terrestrial planets.
\newpage
\section{Context}
The study of exoplanet atmospheres has exploded in the past decade. In 2013, the Exoplanet Exploration Analysis Group (ExoPAG) created---with approval from NASA's Astrophysics Subcommittee---a tenth Study Analysis Group (SAG-X) to consider what NASA could do in the next decade to better understand the atmospheres of transiting planets. SAG-X had open membership and involved three presentations to the exoplanet community. The first presentation, outlining the challenges and opportunities of studying transiting exoplanet atmospheres, was made at the ExoPAG~8 meeting preceding the Oct.\ 2013 Division of Planetary Sciences meeting in Denver, CO. We held a mini-workshop on the capabilities of the James Webb Space Telescope for characterizing transiting exoplanets at ExoPAG~9 preceding the Jan.\ 2014 American Astronomical Society meeting in Washington, DC. Interested members of the community helped draft the current document over the course of the 2014 calendar year and we presented it at ExoPAG 11, preceding the Jan.\ 2015 AAS meeting in Seattle, WA. Multiple drafts of this report have been circulated to the ExoPAG membership and the exoplanet community at large and we have done our best to implement feedback. This document is therefore a consensus view of what can and should be done in the field of transiting exoplanet atmospheres in the next decade.
\section{Planetary Science from the Top--Down}
\textbf{Our knowledge of Earth and the Solar System planets will always exceed our knowledge of any \emph{individual} exoplanet, but the diversity of exoplanets enables the \emph{statistical} study of planets to crack difficult problems in planetary science.}
What started as a trickle in the mid 1990's is now a torrent, with over one thousand extrasolar planets currently known, and thousands of candidates awaiting confirmation. The study of exoplanets has already revolutionized our view of planet formation, and will soon do the same to our understanding of planetary atmospheres and interiors. The diversity of exoplanets gives us the leverage to crack hard problems in planetary science: cloud formation, atmospheric circulation, plate tectonics, etc. However, the characterization of exoplanets presents a challenge familiar to astronomers: our targets are so distant that we only see them as unresolved dots.
Many aspects of planetary science are currently accessible for exoplanets, or soon will be. Since we observe exoplanetary systems from the outside, the easiest aspect to constrain is the architecture of planetary systems, and indeed our theories of planet formation are currently being revolutionized by our growing knowledge of planetary demographics and architecture. Transiting planets are crucial to our understanding of planet formation, because they are the only planets for which we can hope to know the orbital architecture, bulk density, \emph{and} atmospheric composition.
The study of individual planets is likewise progressing from the top--down: first the exospheres, then the atmospheres, and the surfaces last. It is therefore difficult to make definitive statements about the surface conditions of exoplanets, and their interiors will only be known to us through their bulk density, and surface character. The study of planetary atmospheres, however, is poised to be revolutionized by observations of ``exoclimes.'' The first such measurements constrained planetary-scale temperature structure and composition, while the second generation of measurements leveraged the orbital motion of the planet to infer the horizontal temperature structure of the planet. Planned instruments will enable 3D measurements of atmospheric composition, temperature structure, and winds.
\section{Planetary Climate}
\textbf{A planet's albedo determines how much radiation it absorbs, atmospheric composition dictates how energy trickles up via radiation and convection, while atmospheric (and oceanic) dynamics determine how heat is transported from regions receiving more sunlight to those receiving less.}
The primary aspect of planetary climate is temperature, averaged over time, and often space. A detailed calculation of climate involves radiative transfer, fluid dynamics, chemistry and, in the case of an inhabited planet, biology. Stripped to its essentials, however, climate describes the interaction of star light with a planetary atmosphere (Figure~\ref{climate_cartoon}).
\begin{figure}[!htb]
\begin{center}
\vspace{-0.1cm}
\includegraphics[width=40mm, angle=90]{climate_cartoon.ps}
\caption{Planetary climate is determined by stellar radiation interacting with a planetary atmosphere. Incoming radiation (solid yellow lines) is either reflected (dashed yellow lines) or absorbed. The warm air emits at longer wavelengths than the incoming light. Longwave radiation is readily absorbed by greenhouse molecules in the atmosphere, resulting in radiative diffusion (small red arrows). The inefficient upward heat transport means that the lower parts of the atmosphere tend to be hotter than the overlying regions, and in practice most planetary atmospheres convect over some pressure range (blue arrows). The cooler, thinner upper regions of the atmosphere emit thermal radiation (dashed red lines) that balances the absorbed shortwave radiation. \label{climate_cartoon}}
\vspace{-0.1cm}
\end{center}
\end{figure}
A simple energy balance model, shown schematically in the left panel of Figure~\ref{climate_flow}, can therefore be used to predict planetary climate. Unfortunately, despite over a century of research into Earth's climate, there are currently no comprehensive, predictive theories for cloud formation (planetary albedo), volatile cycling (greenhouse gas abundances), or wind speeds (heat transport). The empirical approach to this challenge is to acquire observations for many different planets in the hopes of uncovering the principles of climate (right panel of Figure~\ref{climate_flow}); the first step in that direction has been the study of Solar System worlds over the last half-century.
\begin{figure*}[!htb]
\begin{center}
\vspace{-0.1cm}
\includegraphics[width=50mm, angle=-90]{climate_forward.ps} \includegraphics[width=50mm, angle=-90]{climate_backward.ps}
\caption{\emph{Left:} It is possible to predict a planet's climate given its albedo (reflectiveness), atmospheric opacity (greenhouse gas abundances), and heat transport (wind speeds and ocean circulation). Unfortunately, these critical inputs cannot be predicted in general. \emph{Right:} Fortunately, it is possible to measure planetary albedo, atmospheric composition, and heat transport, even for exoplanets (in many cases, there are multiple independent means of constraining atmospheric properties; we list only a few in the interest of clarity). This makes it possible to empirically determine a planet's climate, even in the absence of a fully predictive theory of planetary climate. In the long run, such observations may reveal some of the underlying principles governing cloud formation, volatile cycling, and large-scale circulation. \label{climate_flow}}
\vspace{-0.1cm}
\end{center}
\end{figure*}
The logical next step, already underway, is to characterize the atmospheres of extrasolar planets. Exoplanets are more diverse than the planets orbiting our Sun, and hence provide more leverage for testing theories. On the other hand, many exoplanets are similar to those in our Solar System, providing crucial Rosetta stones: temperate terrestrial planets that rotate slowly allow us to empirically verify the effects of Coriolis forces on atmospheric circulation, super-Jupiters let us test the effects of surface gravity on Jovian atmospheres, etc.
While the most detailed atmospheric studies will always focus on Earth, and \emph{in situ} measurements will be limited to the Solar System, the vast majority of planets are extrasolar. This means that the most extreme worlds, and those most like the Earth, are exoplanets. Only by studying these planets can we hope to develop comprehensive theories of climate. To paraphrase Kipling (1891): \emph{What should they know of Earth who only Earth know?}
\section{Transiting Planets}
\textbf{Transiting planets are representative of planets on short-period orbits. Exoplanets that transit Sun-like stars tend to be hot, while those transiting low-mass stars are merely warm, and represent the majority of temperate terrestrial planets. Transit spectroscopy probes atmospheric opacity, while measurements of thermal emission constrain vertical and horizontal temperature structure. High cadence, high-precision emission spectroscopy throughout a planet's orbit enables 3D mapping of its atmosphere.}
A transiting planet is a planet that passes in between its host star and the observer. Any planet that orbits a star may be a transiting planet, but transiting planets are essentially synonymous with short-period planets, since the probability that a planet transits is the ratio of the stellar radius to the planet--star separation. This makes the transit method an improbable way to study Solar System analogs.
Direct imaging is a promising approach to studying the atmospheres of Solar System analogs, but close-in planets are difficult to study in this way, because any means of blocking/nulling the starlight is liable to block the planet, too. We instead study the combined light of the planetary system, so the dominant source of photons (and photon noise) is the star rather than the planet. The signal-to-noise ratio for measurements of planetary light is therefore proportional to the planetary flux, rather than its square-root; the usual biases in favor of bigger and brighter targets are especially true for transiting planets.
Characterizing transiting planets would be of merely theoretical interest, except that many exoplanets orbit much closer to their stars than in our Solar System. Some of these unfamiliar worlds are actually rare, like hot Jupiters \citep[][]{Wright_2012}, while others represent a common outcome of planet formation, as with packed systems of warm sub-Neptunes \citep[][]{lissauer2011closely}.
A hot Jupiter is a jovian planet that orbits its host star with a period of $\lesssim$1~week.\footnote{Hot Jupiters should not be confused with \emph{young Jupiters}, the typical targets of current direct-imaging surveys, which are also hot.} Such planets only exist around $\sim1$\% of Sun-like stars \citep[][]{Wright_2012}, but have a transit probability of 10\%, so there is a transiting hot Jupiter for every few thousand FGK stars. These extreme worlds experience strong radiative forcing resulting in day--night temperature contrasts of hundreds to thousands of K \citep{Showman_2002}. The high atmospheric temperatures thermally ionize alkali metals, effectively coupling atmospheric dynamics to planetary magnetic fields \citep{Perna_2010a, Perna_2010b}. The radiative environment of hot Jupiters often results in mass loss, and some are thought to be undergoing Roche lobe overflow onto their host star \citep{Li_2010}.
The most common currently detectable transiting planets in the Galaxy are sub-Neptunes in tight (10--100 days) orbits around their host stars \citep{Howard_2012, Fressin_2013}. These warm sub-Neptunes have masses dominated by rock and ice, but covered in thick H+He atmospheres, and are sufficiently cool (500--1000~K) that photochemistry often trumps thermochemical equilibrium \citep{Moses_2014}. The \emph{a priori} transit probability of such planets is low, but they are intrinsically common, occurring around roughly half of stars. As such, they form the bulk of the \emph{Kepler} crop. The densities of these planets suggest they are largely made of ice, and/or have substantial H+He atmospheres, hence their name. Our experience from the Solar System suggests that these planets likely have high atmospheric metallicities \citep{Fortney_2013}. In short, these planets are expected to be less extreme than hot Jupiters in terms of radiative forcing, but probably more interesting in terms of chemistry.
Temperate terrestrial planets transiting M-dwarf stars are often touted as the poor-astronomer's Earth analog, since they are easier to detect and characterize than a true Earth twin. Based on what we currently know, however, M-Dwarf planets are \emph{the most common habitable worlds}. That is because: 1.\ rocky planets are much more common in the temperate zones of M-Dwarfs \citep{Dressing_2013, Morton_2013} than in the temperate zones of Sun-like stars \citep{Petigura_2013, Foreman-Mackey_2014, Farr_2015}, 2. small stars are more common than big stars \citep[e.g.,][]{Bochanski_2010}, 3.\ the tidally-locked nature of these planets is not a challenge to climate \citep{Joshi_1997, Merlis_2010, Edson_2011} and may double the width of the habitable zone \citep{Yang_2013},
4.\ the red stellar radiation results in a weaker ice-albedo feedback and hence stabler climate \citep{Joshi_2012, Shields_2013, Shields_2014}, and 5.\ the slow main sequence evolution of M-Dwarfs means that a geological thermostat is not strictly necessary to maintain habitable conditions for billions of years \citep{Kasting_1993}. \emph{Studying temperate terrestrial planets around red dwarfs is our best shot at understanding habitability writ large.}
\subsection{Transit}
When a planet passes in front of its host star, it blocks a fraction of the star's light equal to the planet/star area ratio. Inferring the stellar radius from its color and surface gravity, it is possible to convert this relative size into a physical dimension.
On top of the opaque planetary disk there is an annulus of partially transparent atmosphere that filters starlight (Figure~\ref{transit_cartoon}). The spectrum of a planet in transit therefore contains an imprint of scattering and absorption that occurs in the upper atmosphere near the planet's day-night terminator. Even if the bulk composition can be taken as a given (e.g., H+He for jovian worlds), it is still difficult to nail down the abundances of trace gases, and high-altitude hazes can wash out spectral features \citep{Burrows_2014}: even if the sky looked clear to an inhabitant, it may very well be opaque to the grazing rays relevant for transit spectroscopy. Nonetheless, transmission spectroscopy is a powerful characterization tool that can only be applied to transiting planets.
\begin{figure}[!htb]
\begin{center}
\vspace{-0.1cm}
\includegraphics[width=60mm, angle=90]{Transit_Cartoon_v3.ps}
\caption{A planet whose orbit is nearly edge-on will \emph{transit} in front of its star. This is more likely to occur for planets that orbit close to their star, so transiting planets are often synonymous with short-period planets. The amount of light blocked during transit tells us the planet's size, while the transmission spectrum during transit is sensitive to scattering and absorption in the planet's upper atmosphere. During transit we see the planet's nightside, but the planet's orbital motion eventually brings the dayside into view. The dayside reflects sunlight and is often hotter than the nightside, causing variations in brightness known as \emph{phases}. At the top of the orbit, the star will \emph{eclipse} the planet, allowing us to measure the brightness of the star without contamination from the planet; a difference measurement yields the planet's dayside brightness. Reflected light eclipse measurements are sensitive to the planetary albedo and reflected light spectroscopy constrains the nature of scattering. Thermal eclipse measurements can be used to estimate the planet's dayside temperature, while thermal spectroscopy is sensitive to atmospheric composition and vertical temperature profile. \citep[from][]{Cowan_2014} \label{transit_cartoon}}
\vspace{-0.1cm}
\end{center}
\end{figure}
\subsection{Eclipse}
It is possible to use eclipses of the planet by its star to isolate planetary light. A planet that passes directly in front of its host star usually passes directly behind it half an orbit later. The brightness of the planetary system immediately before and after occultation is compared to the brightness during eclipse, and the difference is a measure of the planet's dayside brightness.
We may then convert the eclipse measurement into an estimate of the planet's geometric albedo or dayside brightness temperature, depending on whether the instrument is sensitive to visible or thermal radiation. Spectrally resolved eclipse measurements can constrain atmospheric scattering, composition, and vertical temperature profile \citep{Burrows_2014}.
High-resolution emission spectra are able to resolve molecular lines (as opposed to bands), providing two novel capabilities: performing Doppler measurements on the planet itself, and probing the compositions of cloudy worlds. For transiting planets the system inclination and planetary mass are known, but Doppler measurements might be used to infer high-altitude wind velocities \citep{Snellen_2010} and planetary rotation \citep[][]{Snellen_2014}, but these are arguably accessible via thermal phase curve and eclipse mapping measurements \citep{Rauscher_2014}, or transit morphology \citep{Carter_2010}. Clouds, on the other hand, have been most problematic in grazing transit geometry \citep{2014Natur.505...69K, 2014Natur.505...66K}. Although many emission spectra have so far been featureless, this is likely due to coarse spectral resolution, large measurement uncertainties, and roughly isothermal atmospheres \citep[][and references therein]{Hansen_2014}.
The eclipse depth is sensitive to the hemisphere-averaged properties of a planet, while the very beginning and end of an eclipse offer a means of resolving the planet's dayside: as the planet disappears behind its star and reappears (ingress and egress, respectively), the star's edge scans across the planet. It is possible to invert these raster scans to construct a coarse two-dimensional map of the planet's dayside. The first-order effect is the phase offset of the eclipse due to a zonally-advected hot-spot \citep{Williams_2006} and was first detected by \cite{Agol_2010}. The eclipse timing offset is largely degenerate with orbital eccentricity, but may be teased out if it is chromatic. The detailed morphology of ingress and egress provides a smaller but more robust signal about the 2D flux distribution of the planet's dayside \citep{Rauscher2007}, observed by \cite{Majeau2012} and \cite{deWit2012}.
\subsection{Phases}
Horizontal and temporal differences in planetary temperature produce time variations in thermal emission. For example, the dayside of a slowly-rotating planet might appear warmer and hence brighter than its nightside. Most known exoplanets have curiously eccentric orbits and therefore experience significant seasons due to the changing star-planet separation. Rocky planets are expected to form with randomly oriented spin axes, which leads to obliquity seasons. Disentangling the diurnal cycle, eccentricity seasons and obliquity seasons based on thermal phase variations is a work in progress \citep{Cowan_Voigt_Abbot_2012}.
Tides damp obliquity, slow planetary rotation, and damp orbital eccentricity \citep[e.g.,][]{Heller_2011}, so most short period planets do not experience seasons, but probably have permanent day and night hemispheres. The day--night temperature contrast is therefore an indirect measure of atmospheric heat transport \citep{Cowan2007, Cowan_2012}. High-precision thermal phase curves can be inverted to construct coarse longitudinal temperature maps of short-period exoplanets \citep{Knutson2007, Cowan_Agol_2008}, and even contain indirect information about the planet's latitudinal flux distribution \citep{Cowan_2013}. Full-orbit observations of emission spectra enable spatially-resolved inferences of temperature structure and composition \citep{Stevenson_2014}.
With full-phase observations at high-cadence, high signal-to-noise, and high spectral resolution, it should be possible to constrain the 3-dimensional composition and temperature of an exoplanet's atmosphere. This would allow, for the first time, realistic initialization and/or testing of a
general-circulation model with active chemistry (as opposed to fixed chemistry). \emph{This is the
holy grail of atmospheric characterization, and a top priority for
future exoplanet observations.}
\section{Lessons Learned}
\textbf{The principle lessons learned from the first dozen years of exoplanet atmospheric observations are: (1) short-period planets are not all alike, nor are they a one-parameter family, (2) observations are usually systematics-limited, making out-of-occultation data critical to modeling detector behavior, and dictating that repeatability is the only reliable test of accuracy.}
\subsection{Short-Period Planets}\label{lessons_planets}
Observations of exoplanet atmospheres have so far been weakly constraining, but they have provided a few robust surprises \citep[for more complete reviews, see][]{Burrows_2014b, Burrows_2014, Bailey_2014, Heng_2015}.
Alkali metals have been detected in the atmospheres of some hot Jupiters \citep{Charbonneau_2002}, but are obscured by Rayleigh-scattering hazes in others \citep{Pont_2013}; a Rayleigh scattering slope has also been seen in the transmission spectrum of a warm ice giant \citep{Biddle_2014}. Hazes are present on a warm sub-Neptune \citep{2014Natur.505...69K} and a hot Neptune \citep{2014Natur.505...66K}, but absent on another \citep{Fraine_2014}. Some hot Jupiters are slowly evaporating \citep{2004ApJ...604L..69V}, while there is questionable evidence that one planet is overflowing its Roche lobe and accreting onto its star \citep{Fossati_2010, Cowan_2012}.
Water vapor absorption has now been securely detected in hot Jupiters with the WFC3 instrument on HST using both transit spectroscopy \citep{Deming_2013, Huitson_2013, Wakeford_2013, Mandell_2013} and eclipse spectroscopy \citep{Kreidberg_2014b}, and one hot Neptune shows evidence of water as well \citep{Fraine_2014}. However, claims of molecular absorption in transit and eclipse measurements with other instruments remain controversial. Early detections in NICMOS spectroscopy data \citep{Swain_2008, Swain_2009a, Swain_2009b} have been called into question \citep{Gibson_2011, Gibson_2012, Crouzet_2012}, while analyses of multi-band eclipse photometry with \emph{Spitzer} showing evidence of temperature inversions \citep{Knutson_2008}, disequilibrium chemistry \citep{Stevenson_2010, Swain_2010} and super-Solar C/O \citep{Madhusudhan_2011} have been disputed by subsequent studies \citep{Beaulieu_2011, Mandell_2011, Crossfield_2012, Zellem_2014, Diamond-Lowe_2014}.\footnote{Even when the data reduction scheme is taken at face value, disagreements may occur based on the assumptions that go into spectral retrieval. For example, while \citet{Madhusudhan2011} forces plausible chemistry, \citet{Line2013b} and \citet{Line2014} do not. Since eclipse spectroscopy represents disk-averaged emission from regions with different temperature structures and possibly different chemistries, it is not obvious which strategy is more sound. A pragmatic solution might be to report the best-fit ``reasonable chemistry'' solution, but with uncertainty estimates that are ``chemistry agnostic.''} As the acquisition, reduction, and analysis of eclipse measurements have improved, the broadband emission measurements of hot Jupiters have trended toward featureless Planck spectra \citep{Hansen_2014}.
Dayside broadband emission measurements of hot Jupiters suggest that the hottest planets cannot effectively transport heat to their nightside \citep{Cowan_2011b}, and this trend has been corroborated by thermal phase measurements of two planets \citep[][]{Cowan_2012, Maxted2013}. Thermal phase and eclipse maps are consistent with an equatorial hotspot \citep{Majeau2012, deWit2012} and suggest eastward equatorial winds at a variety of depths \citep{Knutson2007, Knutson2009, Knutson2012} with a characteristic advective time comparable to the radiative timescale \citep{Agol_2010}. Thermal phase measurements of hot Jupiters on eccentric orbits suggest radiative times less than a day and equatorial super-rotation \citep{Lewis2013, Wong_2014}.
The measured geometric albedos of transiting planets are generally small \citep{Rowe_2008, Kipping_2011} and blue \citep{Evans_2013} but with notable exceptions \citep{Demory_2011}. Moreover, the one hot Jupiter with high albedo appears to have spatially inhomogeneous clouds \citep{Demory_2013}.
\subsection{\emph{Spitzer}}
By operating in the mid-infrared without
the confounding telluric IR background, the Spitzer Space Telescope
overcame a key barrier and measured the first exoplanetary photons in
the Fall of 2004. Teams led by \citet{Charbonneau2005} and
\citet{Deming2005} observed secondary eclipses of TrES-1 and HD
209458b, respectively, submitting their independent reports
simultaneously and participating in a joint NASA press release in
April 2005. \emph{Spitzer} remained the observational tool of choice for most of the subsequent decade, and it is useful to summarize what we have learned during this time.
In its cryogenic mission (2003 -- 2009), \emph{Spitzer} had six
photometric channels useful for exoplanet characterization, centered
at 3.6, 4.5, 5.8, 8, 16, and 24 \micron. Channels ranged from 1--3~\micron~wide. \emph{Spitzer}'s InfraRed Spectrograph had a
low-resolution, 5.3--14~\micron~mode useful for exoplanets.
Unfortunately, the rate of transiting-planet detection was low until
around 2010, when the cryogen was already gone. In Warm \emph{Spitzer} only
the 3.6 and 4.5 \micron~photometric channels remain operational; spectroscopy
is unavailable. Thus, only HD~209458b and HD~189733b were ever
observed spectroscopically; only they and GJ 436b were observed at 24
\micron; and only GJ~436b, TrES-1, TrES-4, HD~189733b, and
HD~149026b were observed at 16 \micron. However, numerous planets
have measurements in the four shortest bands, and even more have published 3.6 and 4.5 \micron~eclipse depths \citep{Hansen_2014}.
With an aperture of just 85~cm, \emph{Spitzer} was designed for 10\% absolute
and 1\% relative photometry, an order of magnitude poorer precision
than required for exoplanet studies. The race to achieve results thus
became a contest of systematic correction methods. Bayesian sampling (e.g., Markov Chain Monte Carlo)
replaced model fitting by simple $\chi\sp{2}$ minimization, Bayesian
methods for comparing models with different numbers of free parameters
became the norm, and a variety of numerical approaches modeled
\emph{Spitzer}'s systematics. Although known for a decade or more, improved
image centering and photometric extraction techniques also finally
entered common use. Ultimately, contrast uncertainties better than
0.01\% for single eclipses became possible. While an impressive
improvement over \emph{Spitzer}'s design, the 0.01\% uncertainty still left
the broadband measurements at S/N of only $\sim10$ for most observed
planets. The 1 -- 3 \micron~width and small number of photometric channels
further confounded spectroscopic retrieval.
Most \emph{Spitzer} exoplanet observations consist of unrepeated eclipse observations obtained with different observing schemes and analyzed by disparate groups using a variety of evolving reduction pipelines. The poor signal-to-noise is mostly a testament to the relative faintness of currently known transiting planets; NASA's TESS mission will address this problem. The observations, however, would have greatly benefited from instruments designed to stare at bright point sources, and the same observations would have had much more scientific value had they been obtained in a uniform way and reduced with a uniform, open-source pipeline. \emph{The \emph{Spitzer} experience ultimately justifies the
expense of a purpose-built mission to perform a survey of exoplanet atmospheres.}
\emph{Spitzer}'s after-launch data analysis efforts attempted
to characterize the varying sensitivity across the faces of individual
pixels and to fit its temporal response curves. However, pre-launch
laboratory calibration measurements at better than the 0.01\% level
could have done a better job, as could calibrating long stares at
point sources. For example, the prominent, time-dependent sensitivity
``ramp'' at 5.8, 8, and 16 \micron~was unknown prior to launch
because lab calibrations on bright sources lasted only a short time,
rather than the hours-long timescale of the ramp. Learning from this
experience, instruments that reduce intrapixel sensitivity variations,
that have long-term stability, and that point consistently are now
being designed and proposed for dedicated missions. Even JWST, which
was designed before some key lessons were learned from \emph{Spitzer}, is
being calibrated with enhanced emphasis on observing modes suitable
for exoplanet studies.
\section{Exoplanet Observatories Through 2025}
\textbf{Many ground- and spaced-based observatories can be used to study the atmospheres of hot Jupiters and warm sub-Neptunes in the next decade. For terrestrial planets, especially the temperate variety, the choice of instruments is much more limited: only the Extremely Large Telescopes and the James Webb Space Telescope will be capable of atmospheric observations.}
\subsection{4m telescopes}
4-meter class telescopes are an underused resource in the atmospheric characterization of exoplanets. Ground-based, optical multi-object spectrographs typically capture light from 0.4--1.0 microns and are sensitive to the presence of H$_2$O, alkali metals such as Na and K, and metal hydrides in cloud-free atmospheres. Additionally, in atmospheres with clouds or hazes, we can use spectral information to determine the size distribution and altitude of the cloud/haze particles.
In order to achieve the equivalent precision and be competitive with larger telescopes, these smaller telescopes must acquire multiple transits of a single target. For comparison, four transits of a single target with a 4-meter telescope, such as Mayall or Blanco, have the equivalent photon-limited precision of two transits with an 8-meter telescope, such as Gemini. However, acquiring numerous transit observations with the smaller telescope will mitigate residual atmospheric effects that limit our precision with ground-based observations, plus these observations will likely have a higher duty cycle, thereby outperforming the larger telescope. With a sufficient number of transits, 4-meter class telescopes can contribute exciting, cutting-edge research like with larger telescopes, but at a fraction of the cost.
High precision photometry in the near-infrared of exoplanet occultations provides a direct measurement of thermal emission from hot planets, with single-eclipse precisions of $2\times10^{-4}$ now all but routine \citep{Croll_2014}. \emph{Wide-field NIR photometry with 4m class telescopes is a viable means of achieving high precision photometry for gas and ice giants orbiting relatively bright stars}. The wide field of view allows for a significant number of reference stars for differential photometry, which is critical for minimizing effects of Earth's atmosphere (typically the limiting factor for high precision photometry in the near-infrared from the ground).
Despite the advantages of these types of facilities, only a handful exist: WIRCam on the 3.6m CFHT, WFCAM on the 3.8m UKIRT, NEWFIRM on the KPNO 4m Mayall Telescope, and the Spartan IR Camera on the 4.1m SOAR. Of these, only WIRCam on CFHT and WFCAM on UKIRT have been demonstrated for exoplanet studies (e.g., Croll et al. 2011, Col\'on \& Gaidos 2013).
\subsection{10m telescopes}
As demonstrated by \cite{Colon_2009}, large ground-based telescopes are capable of contributing significantly to photometric follow-up efforts for both small, long-period planets discovered by missions like Kepler and for larger, Jupiter-size planets. Indeed, the first ground-based detection of sodium in an exoplanet atmosphere came from the Hobby-Eberly Telescope \citep{Redfield_2008}. Current 10m-class facilities include the Keck telescopes, the Gran Telescopio Canarias (GTC), Hobby-Eberly Telescope, Southern African Large Telescope, Subaru, Very Large Telescope, Large Binocular Telescope, and Gemini. These facilities offer a number of instruments suitable for exoplanet atmospheric characterization. For example, the primary instrument on the GTC suitable for exoplanet observations is the Optical System for Imaging and low Resolution Integrated Spectroscopy \citep[OSIRIS]{Cepa_2000,Cepa_2003}, which offers a moderately sized field of view of $7.8 \times 7.8$~arcmin. OSIRIS has the capability for standard long-slit spectroscopy, and it also offers a unique tunable filter imaging mode, allowing the user to specify custom optical bandpasses with FWHM = 1.2--2.0 nm. In this mode, observations can be conducted in multiple tunings nearly-simultaneously, thus allowing for narrow-band spectrophotometry. \cite{bean:2010} used multi-object spectroscopy on the VLT to obtain transit spectra of GJ~1214b, with an epoch-to-epoch white-light accuracy of $\sim2\times10^{-4}$. Other up-and-coming facilities include GMOS on Gemini and MOSFIRE on Keck. \emph{Thanks to their large aperture 10~m telescopes are capable of the high precision spectroscopy or spectrophotometry required to measure the small signals from exoplanet atmospheres, e.g., alkali metals in hot Jupiter atmospheres \citep{Sing_2011} or methane in warm sub-Neptunes \citep{Wilson_2014}}.
\subsection{Extremely Large Telescopes}
By 2025, we expect 1--3 of the extremely large, ground-based
telescopes (ELTs) currently being planned (GMT, TMT, E-ELT) to be
operational. These facilities may provide several possible benefits:
their larger apertures will enable studies of larger numbers of
transiting planets around fainter stars, with techniques used today on
8--10~m telescopes; if systematic effects can be controlled, the
larger apertures may also allow planets in brighter systems to be
studied at higher precision; finally, by relieving some pressure on
existing 8--10~m telescopes it may be possible to use these older
facilities for large-scale survey science precluded today due to the
limited availability of observing time.
The ELTs will be optimized to observe faint targets, often using
adaptive optics -- implying relatively narrow fields of view. Thus
fewer nearby systems will be accessible to the differential photometric and
spectroscopic techniques so popular today, which use multiple
comparison stars to remove telluric and instrumental
noise \citep{bean:2010,gibson:2013,crossfield:2013b}. Today, these
efforts perform within factors of 2--3 of the spectroscopic photon
noise limit, but there is no consensus as to whether the extra noise
comes from detectors, instruments, telescopes, or the Earth's
atmosphere. \emph{For ELT multi-object transit spectroscopy to succeed,
these noise sources must be identified and mitigation strategies
incorporated into the new instruments currently being designed.}
High-dispersion spectroscopy may be the niche in which ELTs can best
make significant progress. Observations at high dispersion
($\lambda/\Delta \lambda \gtrsim 20,000$) can measure unique molecular
signatures, thermal profiles, global atmospheric circulation, and
orbital motion and has already been used to characterize the
atmospheres of several transiting planets
\citep{snellen:2010,crossfield:2011,rodler:2012,birkby:2013}. Such observations
also come within 20\% of the photon noise limit \citep{brogi:2014},
which indicates the technique is less susceptible to the systematic
effects that limit multi-object observations. Thus, ELT
high-dispersion spectroscopy may be especially well-suited to high-precision
atmospheric studies of transiting planets.
\emph{High dispersion spectroscopy with the ELTs will be also key to detecting atmospheric
signatures of transiting Earth-like planets hosted by nearby stars, to be discovered by TESS and
PLATO} \citep[see][]{snellen:2013,rodler:2014}.
\subsection{SOFIA}
SOFIA combines a number of advantages for extremely precise time-domain optical and near-infrared spectrophotometric observations using its HIPO \citep{2004SPIE.5492..592D} and FLITECAM \citep{2006SPIE.6269E.168M} instruments, in particular when used in simultaneous FLIPO mode \citep{2010PASP..122.1020A,2012SPIE.8446E..18NS}.
SOFIA is able to avoid most of the perturbing variations of atmospheric trace gases that are the main source of systematic noise for ground-based observations at shorter NIR wavelengths, as these telluric molecules are also the species of interest in the exoplanetary atmospheres. The SOFIA telescope, operating at much lower temperatures (240~K) than ground-based telescopes, reduces thermal background contributions that are the dominant noise source for transit observations at wavelengths longer than 3 microns. SOFIA can observe time-critical events, such as planetary occultations, under optimized conditions, as demonstrated by an observation of a Pluto occultation in June 2011 \citep{2013AJ....146...83P}. SOFIA's initial forays into transit observations have yielded precisions within a factor of 2 of the Poisson limit (Angerhausen et al.\ submitted). \emph{In contrast to space telescopes it is possible to update SOFIA with state-of-the-art instrumentation}, e.g.\ with a dedicated 2nd generation exoplanet instrument \citep[such as NIMBUS,][]{2012SPIE.8446E..7BM}.
\subsection{Hubble Space Telescope}
Planetary emission peaks in the near-to-mid-infrared, while stellar
emission falls dramatically longward of its peak in the visible to
near-infrared. In addition,
prominent molecular rotational-vibrational bands for numerous abundant
molecules, such as water, methane, carbon monoxide, and carbon
dioxide, occur in the near- and mid-infrared. The IR has thus become
the most productive spectral region for exoplanet characterization,
despite the disadvantages of poorer technology, orders-of-magnitude
worse thermal background, and fewer photons compared to the visible
range.
HST/WFC3 can be used for NIR transit spectroscopy of any planet with large scale-height atmosphere, and emission spectroscopy for the hottest planets. It has proven capable of 30~ppm spectrophotometry \citep[in fifteen 7-pixel spectral bins;][]{2014Natur.505...69K}, sufficient to detect molecular features (\S\ref{lessons_planets}). While the limited wavelength range prohibits detailed atmospheric retrieval using HST measurements alone, it is sufficient to determine which exoplanets are hazy. It is also possible to stitch together multiple HST orbits to obtain continuous phase measurements of an exoplanet, hence constraining its global albedo and heat transport \citep{Stevenson_2014}.
HST/COS can be used to perform reflection spectroscopy of the shortest-period giant planets \citep{Evans_2013} as well as to characterize the UV environment of temperate terrestrial planets, and hence photochemistry of water, carbon dioxide, methane, oxygen, and ozone \citep{Linsky_2013, France_2014}. \emph{As the only UV instrument available in the forseeable future, a high priority for HST must be to characterize the stellar UV spectrum of temperate terrestrial transiting planets.} This will be critical in assessing the habitability of these worlds \citep{Wordsworth_2013} and the eventual detection of biosignatures \citep{Tian_2014}. Since HST and its UW spectroscopic instruments, COS and STIS, have a limited lifetime ($\sim$2019), \emph{the need to obtain UV spectra of exoplanet host stars and transmission spectra of planets may justify a future UV mission}.
\subsection{K2, TESS, CHEOPS, PLATO}
\emph{These space-based transit search telescopes can measure white-light geometric albedos and scattering phase functions for hot planets, but with unknown contamination from the planet's thermal emission} \citep[][]{Cowan_2011b, Esteves_2014}. Even a simple red vs.\ blue bandpass, as was available on CoRoT, would greatly alleviate this problem. Alternatively, white-light optical measurements can be combined with thermal measurements obtained with other observatories in order to obtain geometric albedo estimates \citep[e.g., \emph{Spitzer};][]{Demory_2013}.
The geometric albedo constrains atmospheric scattering, and can in principle be converted into a Bond albedo to inform climate calculations, but only by making assumptions about the planet's reflectance spectrum, spatial homogeneity, and scattering phase function.
\subsection{Spitzer Space Telescope}
The \emph{Warm Spitzer} mission has two functioning IRAC channels at 3.6 and 4.5 \micron, with a demonstrated precision of $<10^{-4}$ with intensive campaigns. These are suitable for broadband transit and eclipse measurements, but on its own, \emph{Spitzer} only provides radius and temperature measurements for transiting planets. On the other hand, \emph{Spitzer's ability to continuously stare at a planetary system is unique}. Moreover, the mid-IR wavelengths observable with \emph{Spitzer} help anchor the continuum level of transit and eclipse spectra \citep{Fraine_2013}.
\subsection{James Webb Space Telescope}
The James Webb Space Telescope (JWST) has four instruments, all of which can be used to study transiting exoplanets: Near-Infrared Camera (NIRCam), Near-Infrared Spectrograph (NIRSpec), Near Infrared Imager and Slitless Spectrograph (NIRISS), and Mid-Infrared Instrument (MIRI). For detailed discussion of the instruments and modes best suited to exoplanet studies, see \cite{Beichman_2014}. One of the key takeaways from that report is that although JWST has sensitivity from 0.6--28~\micron, at least four passes with different instruments/modes will be necessary to obtain a 1--12~\micron~spectrum for nearly all systems (faint targets can be observed in 2 passes, provided they don't saturate NIRSpec near 1--2~$\mu$m). \emph{JWST will be the most powerful space observatory for transiting planets in the coming decade, and the only one able to characterize the climate of temperate, terrestrial planets}.
\section{Benchmark Planets}\label{benchmarks}
\textbf{JWST and ELTs will easily characterize the atmospheres of nearby hot Jupiters, hot super-Earths, and warm sub-Neptunes. However, atmospheric measurements of temperate rocky planets transiting nearby M-Dwarfs will be much harder: (1) detection of reflected light or detection of the Planck peak of planetary emission is impossible in a single occultation, (2) transit spectral features in atmospheres free of clouds/hazes are barely detectable in a single occultation with JWST; such measurements could be feasible with ELTs if they approach photon-counting precision in the optical--NIR, and (3) planetary emission on the Rayleigh-Jeans tail should be barely detectable with JWST/MIRI in a single occultation. Since JWST cannot achieve a robust atmospheric measurement of an Earth-analog in a single occultation in the photon-counting limit, characterizing such planets will require campaigns lasting weeks to months.}
We consider four fiducial transiting planetary systems (Table~\ref{parameters}): a hot Jupiter, a hot super-Earth, a warm sub-Neptune, and a temperate super-Earth. The two ``hot'' planets orbit K-dwarfs, while the cooler planets orbit M-dwarfs.
\begin{deluxetable}{lcccrrrrr}
\tabletypesize{\scriptsize}
\tablecaption{Fiducial Transiting Planet Parameters \label{parameters}}
\tablewidth{0pt}
\tablehead{
\colhead{Planet Type} &\colhead{Stellar}& \colhead{Stellar} & \colhead{Planetary}& \colhead{\emph{Transit}}& \colhead{$a/R*$}&\colhead{\emph{Dayside}}& \colhead{Mean Mol.}&\colhead{Planet}\\
&\colhead{Temp.}& \colhead{Radius} & \colhead{Radius}& \colhead{\emph{Depth}$^{a}$}&&\colhead{\emph{Temp.}$^a$}& \colhead{Mass}&\colhead{Gravity}
}
\startdata
Hot Jupiter &5000~K& $0.75R_\odot$ & 1.10~$R_J$& $2.2\times10^{-2}$&5& 1787~K& 2~$\mu$&20~m/s$^2$\\
Hot Super-Earth &5000~K& $0.75R_\odot$& 1.50~$R_\oplus$ &$2.9\times10^{-4}$&3&2308~K& 140~$\mu$&12~m/s$^2$\\
Warm Sub-Neptune &3000~K& $0.20R_\odot$& 0.24~$R_J$& $1.4\times10^{-2}$&15&619~K&2~$\mu$ &9~m/s$^2$\\
Temperate Super-Earth &3000~K& $0.20R_\odot$& 1.50~$R_\oplus$ &$4.1\times10^{-3}$&90&253~K& 28~$\mu$&12~m/s$^2$
\enddata
\tablecomments{$^a$Derived quantity.}
\end{deluxetable}
\subsection{Signal}
The reflected light contrast is simply $[F_p/F_*]_{\rm ref} = A_g(R_p/a)^2$, where $A_g$ is the geometric albedo (assumed to be 0.3 for all four planets), $R_p$ is the planetary radius, and $a$ is the semi-major axis.
The amplitude of transit spectral features is approximated as $2R_p N_H H/R_*^2$, where the stellar radius is $R_*$, the atmospheric scale height of the planet is $H = k_B T/(\mu g)$, atmospheric temperature is $T$, atmospheric mean molecular mass is $\mu$, and surface gravity is $g$. The number of scale heights probed, $N_H$, is approximately 4 \citep{Griffith_2014}. We set the atmospheric temperature to the dayside effective temperature, described below.
The dayside emitting temperature depends on the incident stellar flux, the fraction that is reflected away, and the fraction that is transported to the planet's nightside. Adopting the parametrization of \cite{Cowan_2011b}, we have $T_d = T_* \sqrt{R_*/a} (1-A_B)^{1/4} (2/3 - 5\varepsilon/12)^{1/4}$, where $T_*$ is the stellar effective temperature, and we assume a Bond albedo of $A_B=0.3$ and heat recirculation efficiency of $\varepsilon=0.2$ for all planets.
If we treat the planet and its host star as blackbodies, then the thermal contrast ratio between them is proportional to the ratio of Planck functions: $[F_p/F_*]_{\rm therm} = (R_p/R_*)^2 B(\lambda, T_d)/B(\lambda, T_*)$. The peak of the planetary emission is given by Wien's Law, $\lambda_{\rm peak} = 2898/T_d$, while in the Rayleigh-Jeans limit ($\lambda \to \infty$) the thermal contrast is simply $(R_p/R_*)^2 (T_d/T_*)$.
\begin{deluxetable}{lrrrr}
\tabletypesize{\scriptsize}
\tablecaption{Atmospheric Signals \label{values}}
\tablewidth{0pt}
\tablehead{
\colhead{Planet Type} & \colhead{Reflected} &\colhead{Transit} & \colhead{Planet Peak} & \colhead{Rayleigh-Jeans}\\
& \colhead{Contrast} &\colhead{Feature} & \colhead{Contrast}& \colhead{Contrast}
}
\startdata
Hot Jupiter & \boldmath{$2.6\times10^{-4}$} & \boldmath{$8.3\times10^{-4}$} &\boldmath{$7.4\times10^{-4}$}&\boldmath{$7.7\times10^{-3}$}\\
Hot Super-Earth & \boldmath{$9.8\times10^{-6}$} &\boldmath{$3.0\times10^{-6}$}& \boldmath{$1.8\times10^{-5}$}& \boldmath{$1.4\times10^{-4}$}\\
Warm Sub-Neptune & $1.9\times10^{-5}$ &\boldmath{$2.0\times10^{-3}$}& $1.8\times10^{-4}$& \boldmath{$3.0\times10^{-3}$}\\
Temperate Super-Earth & {\color{red}$1.5\times10^{-7}$} &$2.3\times10^{-5}$& {\color{red}$1.5\times10^{-5}$}& $3.5\times10^{-4}$
\enddata
\tablecomments{Fonts denote the difficulty of measuring these signals with JWST in the white-light \emph{photon-counting limit} (cf.\ Table~\ref{precisions}): \boldmath{bold} indicates signal measured at $>10\sigma$ in a one hour integration, while {\color{red}red} denotes $<1\sigma$.}
\end{deluxetable}
The amplitudes of atmospheric signatures for the four planets are listed in Table~\ref{values}. If one simply considers the amplitudes of the atmospheric signals, the easiest measurement is the thermal secondary eclipse, which constrains a planet's emitting temperature, followed by emission spectroscopy to constrain vertical temperature structure.\footnote{A typical temperature contrast of $0.5^{1/4}\approx0.8$ between the infrared photosphere and the stratosphere produces molecular emission features of order unity.} For hot Jupiters and hot Earths, thermal phase variations---constraining horizontal temperature structure---are similar in amplitude to eclipses, but require much longer observations. With the exception of the warm sub-Neptune, transit spectroscopy---yielding atmospheric scale height and opacity---is a smaller signal, and in practice requires higher spectral resolution. The smallest signal is typically reflected light, which informs atmospheric scattering and clouds.
It is worth discussing what habitability constraints such measurements would provide for the handful of transiting temperate terrestrials TESS is expected to discover within the decade. The detection of planetary emission places a lower limit on the surface temperature of the planet (modulo the greenhouse effect), while broadband thermal phase measurements can be used to infer a planetary atmosphere \citep{Seager_2009}, constrain its mean-molecular weight and opacity \citep{Menou_2012, Kataria_2014} and surface pressure \citep{Koll_2015}, and might indicate the presence of water clouds \citep{Yang_2013}. Measuring a transit spectral feature, on the other hand, will establish that the planet has an atmosphere and signal the presence (but not abundance) of a molecule at low pressures \citep{Burrows_2014}. With a full transit spectrum from the optical through the infrared it would be possible to uniquely determine the composition of such a planet's upper atmosphere \citep{Benneke_2012}.
\subsection{Noise}
While the atmospheric signal depends on a combination of planetary and stellar parameters, the overall flux---and hence noise---is dominated by the star. We compute the number of photons collected by an instrument, accounting for limited bandpass and imperfect system throughput:
\begin{equation}
N_{\rm phot} = \frac{\pi \tau \Delta t}{hc}\left(\frac{R_* D}{2d}\right)^2 \int_{\lambda_1}^{\lambda_2} B(\lambda, T_*) \lambda d\lambda ,
\end{equation}
where $\lambda_1$--$\lambda_2$ is the bandpass, $d$ is the distance to the star (fixed to 20~pc), $D$ is the telescope diameter, $\tau$ is the system throughput (photon conversion efficiency = electrons out per photon in), and $\Delta t$ is the integration time (set to 1~hour). The telescope and instrument parameters used are listed in the top half of Table~\ref{precisions}. In the poisson limit for large numbers of photons, the precision is simply $1/\sqrt{N_{\rm phot}}$, as listed in the bottom half of Table~\ref{precisions}.
It is important to note that the numbers listed in Table~\ref{precisions} are \emph{white light} precisions and therefore only directly apply to measurements of reflected light, thermal emission, or broadband spectroscopy. The precision for spectroscopy is less impressive. For example, if one wanted to obtain a 1.0--4.0~$\mu$m spectrum with 0.1$\mu$m resolution using JWST/NIRSpec, then the expected precision would be $\sqrt{30}\approx 5.5$ times worse than the values shown in the table.
The amplitude of spectral features for a temperate terrestrial planet transiting in front of a nearby M-Dwarf is therefore comparable to the single-epoch photon-counting precision. In the absence of a systematic noise floor, 100 transits of such a planet could yield 10$\sigma$ detections of greenhouse gases. Spending a total of one month of JWST time to characterize the atmosphere of a potentially habitable world is compelling, but the observations would have to be spread out over nearly a decade for a planet in a month-long orbit (this scheduling problem is alleviated for planets in the habitable zones of later M-Dwarfs, which have shorter orbital periods).
Table~\ref{precisions} suggests that detecting thermal emission from temperate terrestrials is a better option, but one should be cautious here, too: at longer wavelengths, the dominant source of noise may not be photon counting, but the warm detector background. Moreover, staring continuously at an M-Dwarf to monitor the phase variations of a temperate planet poses technical challenges that have not yet been resolved. \emph{Any atmospheric measurement of a temperate terrestrial planet will be difficult for JWST, and any robust atmospheric signature will require at least a month of observing time.}
\subsection{Comparison to Empirical Precisions and Other S/N Estimates}\label{comparison}
Since these precision estimates are idealized, it is worth considering a few concrete examples to make sure they are approximately correct.
\cite{2014Natur.505...69K} report 30~ppm ($3\times10^{-5}$) precision for transits of GJ1214b, at the photon-counting limit. This planet is closer than our nominal warm sub-Neptune (13 instead of 20~pc), the authors observed 12 transits lasting 0.9 hours each, and divide the HST/WFC3 G141 band into 22 spectral bins. The eclipse depth is differential measurement so the uncertainty in the out-of eclipse baseline contributes to the overall uncertainty, a $\sqrt{2}$ penalty. Combining these adjustments, we would naively expect a photon-limited precision of $(1.6\times10^{-5})(13/20)\sqrt{22/(12\times0.9)}\sqrt{2} = 2\times10^{-5}$.
\cite{Knutson2012} report precisions of $5\times10^{-5}$ for a single eclipse of HD~189733b measured with \emph{Spitzer}/IRAC ch1, a factor of a few worse than the Poisson limit. Since this planet is nearly identical to our notional hot Jupiter, we can compare to our predicted photon limit for the same instrument (Table~\ref{precisions}). The eclipse duration is $1.8$ hours, and we again account for the $\sqrt{2}$ penalty for a differential measurement with equal number of photons in and out of transit. These two effects roughly offset, so Spitzer/IRAC should be able to measure a single eclipse of this planet with a precision of approximately $2\times10^{-5}$, in the photon-counting limit.
\cite{Crossfield_2012} report a precision of $3\times 10^{-4}$ for three eclipses of HD~209458b observed with \emph{Spitzer}/MIPS at 24~$\mu$m. We begin with the value of $8.4\times 10^{-5}$ from Table~\ref{precisions}. Ignoring the somewhat hotter star (6000 rather than 5000~K) but accounting for the more distant system (50~pc), $3\times3$~hour eclipses, and the $\sqrt{2}$ penalty for a differential measurement, we obtain a photon-counting precision of $9.9\times10^{-4}$, a factor of three better than the empirical accuracy obtained by \cite{Crossfield_2012}.
In short, the values listed in Table~\ref{precisions} accurately reflect the best-case scenario of photon counting precision.
We can also compare our JWST precision estimates to those made by other groups. Detailed simulations of JWST/NIRSpec indicate that it will only be possible to detect water absorption in transit spectra for planets that are hydrogen-rich and/or hotter than 400~K (N.~Batalha, private communication). In other words, \emph{the only way that JWST can easily perform transit spectroscopy of temperate terrestrial planets is to stretch the definition of the latter}. Quantitatively, they find that observing 40 transits might enable the $15\sigma$ detection of water vapor absorption for a bona fide temperate terrestrial planet. If we begin with our JWST/NIRCam F115W estimate, and account for the approximately 120~hrs of in-eclipse time, $\sim$20 resolution elements, and 3~pc system distance, we obtain a precision of $1.3\times10^{-5}\times\sqrt{20/120}\times(3/20) = 8\times10^{-7}$, or a 14$\sigma$ detection, in excellent agreement with the estimate of Batalha et al.
\cite{Deming_2009} estimate that JWST/MIRI can obtain a 3.2$\sigma$ detection of a HZ super-Earth's Planck peak emission at 13~pc in 4$\times$90 min eclipses. If we take our JWST/MIRI F1000W estimate and account for the 6~hrs of integration, and the closer planet, we obtain an estimate of $2.6\times10^{-5}\times(13/20)\times(1/\sqrt{6}) = 7\times 10^{-6}$. Since the signal at this wavelength is $1.5\times10^{-5}$, we would have predicted a 2$\sigma$ detection. This is excellent agreement given the somewhat different assumptions about stellar and planetary properties.
\cite{Yang_2013} predict a 49$\sigma$ detection of planetary emission in a 1-day integration of a $2R_\oplus$ super-Earth orbiting a 3000~K M-dwarf 20~pc away, but assuming one integrates 10--28 micron, which would take multiple passes with JWST/MIRI. Starting with our JWST/MIRI F2550W precision estimate and accounting for the longer integration time yields $9.7\times10^{-5}/\sqrt{24} = 2\times10^{-5}$. This would be an 18$\sigma$ detection of planetary emission according to Table~\ref{values}, but the planetary flux at 25~$\mu$m is a few times smaller than in the Rayleigh-Jeans limit. \emph{We therefore find that measuring thermal emission with JWST/MIRI is the easiest atmospheric measurement one can make for temperate terrestrial planets, but it is harder than stated in \cite{Yang_2013}.}
One may fear that detector systematics will stymie our ability to detect the subtle signatures of temperate terrestrial planets. This is a legitimate worry, since many of the instruments on JWST have heritage in instruments on HST and \emph{Spitzer}. It is therefore encouraging that the current broadband precision record holders with NASA's great observatories are comparable to the amplitude of atmospheric signatures on transiting temperate terrestrials: $5\times10^{-6}$ in the NIR \citep{Stevenson_2014}, $5\times10^{-5}$ at 3.6~$\mu$m \citep{Knutson2012}, and $3\times10^{-4}$ at 24~$\mu$m \citep{Crossfield_2012}. \emph{Robustly detecting the atmospheres of temperate terrestrial exoplanets requires less than an order-of-magnitude improvement in our detector models} \cite[this is in line with the claim that the transit spectral signature of a temperate terrestrial planet could be measured with an intensive HST campaign;][]{2014Natur.505...69K}.
\section{How to Spend JWST Time}
\textbf{During a 5~year lifetime, JWST could determine the vertical temperature structure of 150 transiting giant planets, map the atmospheres of 25 giants, \emph{or} constrain the habitability of 3 temperate terrestrial planets. A balanced portfolio of these three categories would maximize science returns.}
As a ballpark estimate based on experience with the Hubble and Spitzer Space Telescopes, we anticipate that 25\% of JWST time will be devoted to transiting planets. Multiplying this by the estimated 70\% duty cycle of the observatory and a 5 year mission, yields 320 days (should the mission last 10~years, there would be approximately 600~days to devote to transiting exoplanets). How should we spend it? We summarize a few endmember observing portfolios in Table~\ref{portfolios} and discuss the topic in more detail below \citep[for a qualitative discussion see][]{Heng_2015}. We focus only on the most time-consuming observations; there will undoubtedly be rare opportunities that will be pursued (e.g., temperate transiting gas giants).
\begin{deluxetable}{lccc}
\tabletypesize{\scriptsize}
\tablecaption{JWST Transiting Planet Observing Portfolios \label{portfolios}}
\tablewidth{0pt}
\tablehead{
\colhead{Portfolio Name} & \colhead{Number of Targets} & \colhead{Duration per Target} & \colhead{Total Time}}
\startdata
Atmospheric Structure & 150 & 2 days & 300 days\\
Atmospheric Mapping & 25 & 12 days & 300 days\\
Temperate Terrestrials & 3 & 100 days & 300 days
\enddata
\tablecomments{Endmember portfolios for transiting planet science with JWST. Linear combinations of these are possible, and probably more scientifically productive: e.g., $70\times2=140$~days of structure, $5\times12=60$~days of mapping, and 1 terrestrial (100~days) also adds up to 300~days.}
\end{deluxetable}
\subsection{Hot Jupiters and Warm Sub-Neptunes}
At six hours per transit (2 out + 2 in + 2 out) $\times$ 4 instruments (the JWST Science Working Group should codify these modes), it takes roughly
24 hrs to get a 1--10~$\mu$m transit spectrum. Eclipse
spectroscopy will take just as long \citep[a single pass is also sufficient for eclipse mapping of the brightest targets;][]{Rauscher2007}. \emph{It would therefore take 2 days to get a baseline atmospheric characterization of a transiting exoplanet with JWST}. If the planet's atmosphere is well mixed, one can roughly think of the transit spectra as constraining the atmospheric opacity, which allows the dayside emission spectroscopy to more robustly constrain the temperature structure. If one performed a systematic survey using JWST, it would be possible to obtain 1--10~$\mu$m transmission and eclipse spectra for 150 short-period giant planets over the course of a 5~yr mission.
Full phase coverage constrains the planet's energy budget, heat transport, radiative response, and enables spatially-resolved chemical mapping \citep{Stevenson_2014}. If one considers planets with a typical orbital period of 3 days, it takes 12 days to obtain a full phase-resolved spectrum of a single planet. A systematic survey using all of JWST's transiting planet time could perform such detailed observations for 25 short-period planets.
It should be noted that even a partial planetary spectrum with a single JWST instrument would be much more valuable than a pair of broadband \emph{Spitzer} measurements (currently the typical state of knowledge for hot Jupiters): the high signal-to-noise ratio and spectral resolution should enable meaningful inferences about atmospheric structure and composition.
\subsection{Hot Earths}
TESS is expected to discover 4--6 hot Earths orbiting V=9--11 stars (Sanchis-Ojeda, priv.\ comm.). \cite{Samuel_2014} estimates that NIRSpec observations of CoRoT-7b (V=12) would take 70~hrs, while a V=8 target would only take 3~hrs; our S/N estimates in Table~\ref{values} are in broad agreement. Given the short orbital periods, it is both expedient and scientifically important to obtain full phase curves. Devoting 4 days each with 3 instruments (too faint for MIRI) for all 5 bright targets adds up to 60 days for exo-geology.
\subsection{Temperate Terrestrial Planets}
Given the planetary demographics of \emph{Kepler}, TESS should deliver 4 temperate terrestrial planets ($0.5R_\oplus<R<1.5R_\oplus$ and $0.5S_\oplus<S<1.5_\oplus$, where $R$ and $S$ are planetary radius and insolation, respectively) that transit bright M-dwarfs in the JWST continuous viewing zone (P.~Sullivan, private communication).\footnote{\cite{Gaidos_2014}, on the other hand, estimate that TESS only has a 1.3\% chance of detecting a HZ planet around one of the $\sim3000$ \emph{brightest} late-K and early-M stars (CONCH-SHELL stars).} TESS host stars are early-M so the HZ corresponds to month-long orbits. HZ worlds discovered by the MEarth survey would be better because the host stars might be cooler, the transit depths more favorable, $F_p/F_*$ larger, and the HZ orbits shorter. But at 1.9--2.6$R_\oplus$ in radius, these planets are likely temperate sub-Neptunes (Z.~Berta-Thompson, private communication).
Given the numbers in Tables~\ref{values} and \ref{precisions}, it would take a few months to obtain high-precision JWST/NIRSpec transit spectra and JWST/MIRI phase curves for a single nearby temperate terrestrial planet. Using all 300~days of anticipated JWST transiting planet time, we could obtain transit spectra and thermal phase curves for 3 terrestrial HZ worlds orbiting nearby M-Dwarfs.
\subsection{Target-Limited vs.\ Time-Limited Regimes}
In the target-limited regime, there are precious few bright targets to observe, so they are observed with all possible instruments (Figure~\ref{target_selection_now}). This describes the current state of transiting planet atmosphere observations: the vast majority of atmospheric observations have focused on the few brightest targets \citep[$\sim 50$ planets have had their thermal emission measured with \emph{Spitzer}, but these data only serve to make estimates on the planets' dayside emitting temperature;][]{Cowan_2011b, Hansen_2014}.
In the time-limited regime, on the other hand, there is an embarrassment of riches. The challenge is to determine which targets are most interesting for follow-up work. There are two reasons to believe that we will soon be entering the time-limited regime for hot Jupiter and warm sub-Neptune characterization. Firstly, the large collecting area of the JWST ensures that current borderline targets will yield high-precision observations. Secondly and more importantly, TESS will provide thousands of targets that are as bright as our current darling planets.
\begin{figure}[!htb]
\begin{center}
\vspace{-0.1cm}
\includegraphics[width=70mm]{target_selection_now.eps}
\caption{\emph{Target-Limited Regime:} Past and current ground-based radial velocity and transit surveys have combined to discover and characterize hundreds of transiting exoplanets, only a dozen of which are bright enough to warrant detailed follow-up with the Spitzer and Hubble Space Telescopes. In effect, any bright transiting planet is the subject of detailed atmospheric follow-up. \label{target_selection_now}}
\vspace{-0.1cm}
\end{center}
\end{figure}
Planets that are harder to detect, because of their small size and/or long orbit, will remain in the target-limited regime even with the advent of TESS. It is a safe bet that JWST will study all the bright exemplars. Moreover, the small size and cool temperatures of such rare worlds make it likely that JWST would be the only space observatory capable of characterizing these worlds, which will only strengthen the case for studying them with the flagship telescope (ELTs may perform transit spectroscopy on such cool planets, but will not be able to detect their thermal radiation).
\subsection{A Straw-Man Schedule for JWST}
Temperate terrestrial planets discovered by TESS will probably not be confirmed until JWST's second year of observations. It therefore seems likely that the first two years of transiting exoplanet time will be spent characterizing the atmospheres of currently known (or soon to be discovered) hot Jupiters and warm sub-Neptunes. \emph{In principle, this could yield credible atmospheric abundances and vertical temperature structure for about 50 short-period giants.} Our community should be wary of observing more planets at the price of less complete spectral coverage and poor signal-to-noise; this is one of the important lessons learned from \emph{Spitzer}. As observers and instrument specialists learn about the telescope's abilities and limitations, it may also be possible to perform spectrally resolved phase mapping of a half-dozen choice planets.
Once there are confirmed temperate terrestrial planets transiting bright M-Dwarfs in the JWST continuous viewing zone ($\sim$2020), a large fraction of the telescope's transiting planet time could and should be devoted to characterizing their atmospheres. \emph{These measurements will be time-intensive and unlikely to conclusively establish or refute the habitability of these worlds, but would lay the ground-work for future flagship missions.}
\section{Maximizing Exoplanet Science}
The next decade of transiting planet characterization is bound to be more exciting than the last. The challenge will be to ensure that the science yield keeps pace with the excitement.
\subsection{The Case for a Dedicated Exoplanet Atmosphere Survey Telescope}
\textbf{JWST will have insufficient time to study the hundreds--thousands of hot Jupiters and warm sub-Neptunes discovered by TESS, and these observations will be heterogeneous both in terms of spectral coverage, observing mode, and data reduction. The best way to capitalize on the TESS discoveries is with a dedicated exoplanet atmosphere survey. A meter-class space telescope with an optical---NIR spectrograph could perform a survey of the brightest non-terrestrial TESS planets in 1--2 years.}
There is a compelling portfolio of transiting exoplanet science that can be achieved with JWST, but two important caveats: (1) if the characterization of temperate terrestrial planets is feasible, it will likely take up much of the exoplanet time allotment in the 2020's, and (2) there is currently no intelligible plan for choosing which of the thousands of bright TESS giant planets to characterize.\footnote{It is also possible that the very nearest TESS targets will be too bright for JWST to observe, or at least to observe efficiently. For example, various NIRSpec and NIRISS modes have saturation limits of $J$=7--5 \citep{Beichman_2014}.} We discuss each of these problems in turn below.
If one accepts the hypothesis that target-limited planets will be observed at any cost, then observations of potentially habitable worlds could take most of the $\sim640$~days of transiting planet time we have hypothesized the exoplanet community might obtain over a ten year mission. In particular, all estimates indicate that even a basic characterization of a temperate terrestrial planet will take of order 100~days. One could easily spend the entirety of JWST transiting planet time following up the 4 most habitable TESS planets. The only scenario in which JWST does \emph{not} spend hundreds of days staring at cool rocky planets is if its instruments suffer from such severe systematics that such observations are deemed a waste of time; this scenario seems unlikely.
There is only enough transiting exoplanet time in a 5~year JWST mission to fully characterize 25 short-period giant planets (full-orbit observations with full spectral coverage). It is not obvious how to partition this time between hot Jupiters and warm sub-Neptunes, between circular and eccentric orbits, or between host star spectral type in order to maximize the scientific returns. If \emph{Spitzer} is any indication, the temptation will be to use the majority of the telescope time on the smaller, cooler targets. The question of target selection is further complicated by the existence of certain rare classes of objects. For example, a dozen hot Jupiters have precise \emph{Kepler} eclipse measurements \citep{Heng_2013}, but only one has an appreciable albedo indicative of clouds \citep[Kepler-7b;][]{Demory_2013}. Likewise, many transit spectra are washed out by atmospheric hazes \citep{2014Natur.505...66K, 2014Natur.505...69K}, but some are haze-free \citep{Fraine_2014}; it would be wasteful to obtain countless featureless spectra with the premier space observatory. The scientific value of JWST observations would be maximized by focusing on archetypal planets rather than simply going after the brightest targets.
\emph{In short, there is a clear need for a dedicated telescope to perform a uniform survey of hundreds of bright transiting giant planets in order to allow JWST to focus on (1) the smaller, cooler planets that it is uniquely able to characterize, and (2) the dozen archetypal giant planets worthy of detailed study (Figure~\ref{target_selection_later}).}
\begin{figure}[!htb]
\begin{center}
\vspace{-0.1cm}
\includegraphics[width=70mm]{target_selection_later.eps}
\caption{\emph{Time-Limited Regime:} A combination of ground-based RV monitoring and space-based transit surveys (primarily TESS, but also Cheops, and eventually PLATO), will deliver thousands of transiting planets comparably bright to the current best targets. There will be insufficient time to perform atmospheric follow-up observations with JWST and ELTs. However, these targets will be sufficiently bright that a modest telescope could obtain compelling atmospheric constraints. Moreover, the duration of transit and eclipse observations are limited by planetary dynamics, so a serial transit survey (one system at a time) could observe $10^3$ planets per year. This motivates a dedicated exoplanet atmosphere survey telescope to obtain uniform observations of the brightest transiting planets, and to determine the few non-terrestrial planets most worthy of follow-up observations. Since the targets would be bright and have large scale heights, such a mission could be relatively modest: $\sim$1~m diameter primary mirror coupled to a optical--NIR spectrograph. \label{target_selection_later}}
\vspace{-0.1cm}
\end{center}
\end{figure}
Based on \S\ref{benchmarks}, \emph{a $\sim1$~m space telescope equipped with an optical---NIR spectrograph should be able to measure---in a single occultation---the molecular absorption features of bright ($J<11$) TESS sub-Neptunes, and to detect the reflected light and thermal emission of the very brightest ($J<9$) hot Jupiters.} We expect such a mission to cost \$500M, to within a factor of 2. Alternatively, a scaled-down version of this survey could be performed using a probe-class ($\lesssim$ \$1B) direct-imaging mission without its coronagraph/starshade.
There is a trade-off between the capabilities of such a survey instrument and its timeliness. In order for it to act as a triage for JWST sources, the mission would need to fly as soon as possible and would nessecarily be relatively modest. A more ambitious mission, on the other hand, would likely fly towards the twilight years of the JWST mission, but would itself be able to provide detailed characterization of hot Jupiters and warm sub-Neptunes.
An easier question to address is how many transiting planets such a survey telescope could/should observe. First of all, the shortest segment of time that makes sense for observing a transiting planet is 3$\times$ the transit duration, or 6 hours, for typical short-period planet with 2-hour transits. Even telescopes with very large collecting areas must observe planets for at least this amount of time. Secondly, bright transiting targets are few and far between, dictating that an atmospheric characterization mission will necessarily study them serially rather than in parallel. As a result, a transit characterization mission can observe up to $365\times4\approx1000$~planets per year, regardless of telescope diameter. By lucky coincidence, this number is well-matched to the number of bright ($J<11$) transiting planets that TESS is expected to find.
For the shortest-period planets, eclipse and phase measurements could additionally constrain reflected light and thermal emission. Depending on the telescope diameter and the brightness of target stars, some systems might benefit from multiple observations to beat down the photon noise. Repeat measurements would also enable robust inference of uncertainties, although a dedicated mission may be so well calibrated that uncertainties can be adequately estimated based on a single event. These sorts of considerations ---not to mention scheduling constraints--- would realistically limit a dedicated mission to observing hundreds of transiting exoplanets in one year, roughly lining up with the expected number of TESS candidates with $J<10$, not to mention transiting planets discovered by the K2 \citep{Barclay_2014}, CHEOPS \citep{Broeg_2013}, and PLATO \citep{Rauer_2014}, or the countless ground-based wide-field transit searches.
\subsection{The Case for M-Dwarfs}
\textbf{JWST will be the best observatory for constraining the atmospheres of temperate terrestrial planets orbiting nearby M-Dwarfs, but will only be able to obtain very basic atmospheric constraints: establishing the presence of an atmosphere, measuring its emitting temperature, detecting greenhouse gases, and---if orbital phase monitoring is possible---constraining the planet's rotational state, surface pressure, and the presence of clouds. These constraints will be tantalizing, but in order to properly characterize these worlds, future flagship missions will have to be designed with M-Dwarf planets in mind.}
It is generally accepted that the next big step in exoplanet atmospheric characterization after JWST will be a direct-imaging mission to search for and characterize the atmospheres of Earth-like planets orbiting Sun-like stars \citep[e.g., ATLAST, LUVOIR;][]{Kouveliotou_2014}. In order to minimize risk and maximize scientific return, such a flagship mission must also be able to directly image habitable planets orbiting nearby M-Dwarfs.
The frequency of temperate terrestrial planets is already known to be high for M-Dwarfs \citep{Dressing_2013, Morton_2013, Dressing_2015}, and the TESS mission is predicted to find a few nearby transiting examples. On the other hand, not a single Earth analog has yet been detected in orbit around a G-Dwarf, and the frequency of such planets therefore requires extrapolation to unexplored regions of parameter space \citep{Petigura_2013, Foreman-Mackey_2014, Farr_2015}. It might be possible to interpolate over (rather than extrapolate to) the habitable zone of Sun-like stars by combining the microlensing statistics of WFIRST with the transiting statistics of \emph{Kepler} in order to obtain a more robust frequency by the late 2020's. In order to find \emph{nearby} Earth twins, we will need 10~cm/s radial velocity surveys, a dedicated astrometry mission, or an ambitious direct-imaging survey. It therefore appears likely that a next-generation direct-imaging flagship mission will be designed before we know the frequency of terrestrial planets in the habitable zones of Sun-like stars, let alone the location of the nearest examples. In order to minimize the risk of such a blind search, it would be prudent to design the telescope such that it could characterize temperate planets orbiting \emph{low}-mass stars. Such planets will be discovered in the next decade, and their atmospheric characterization will have begun in the JWST era.
\section{Summary}
\textbf{JWST will completely characterize a few dozen hot Jupiters and sub-Neptunes, and obtain rudimentary constraints on the atmospheres of Earth-like worlds. This leads us to two conclusions: (1) a dedicated mission is needed to perform an atmospheric survey of TESS planets. Most of these planets are hotter and larger than Earth, with atmospheres made of lighter molecules, and therefore relatively easy to observe: a 1~m space telescope equipped with an optical---NIR spectrograph could perform a uniform atmospheric survey for hundreds of the brightest TESS planets; (2) the detailed characterization of temperate terrestrial planets orbiting M-Dwarfs should be an explicit goal of future flagship missions.}
\acknowledgements{NBC acknowledges the generous hospitality of l'Institut de Plan\'etologie et d'Astrophysique de Grenoble, where he wrote much of this report.}
\begin{deluxetable}{lcccccccc}
\rotate
\tabletypesize{\scriptsize}
\tablecaption{White Light Photon-Counting Precision of 1~hr Integrations on a Target at 20~pc$^a$ \label{precisions}}
\tablewidth{0pt}
\tablehead{
&\colhead{JWST/NIRSpec} & \colhead{HST/WFC3}& \colhead{JWST/NIRSpec} & \colhead{Spitzer/IRAC} & \colhead{JWST/NIRCam} & \colhead{JWST/MIRI} & \colhead{Spitzer/MIPS}&\colhead{JWST/MIRI}\\
&\colhead{0.6--1.0~$\mu$m} & \colhead{G141} & \colhead{1.0--4.0~$\mu$m} & \colhead{ch1} & \colhead{F356W} & \colhead{F1000W} & \colhead{24~$\mu$m}&\colhead{F2550W}
}
\startdata
$D$& 6.5~m& 2.4~m& 6.5~m& 0.85~m&6.5~m& 6.5~m&0.85~m&6.5~m\\
$\lambda_1$& 0.6~$\mu$m& 1.1~$\mu$m& 1.0~$\mu$m& 3.225~$\mu$m& 3.12~$\mu$m& 9~$\mu$m& 21.65~$\mu$m &23.5~$\mu$m\\
$\lambda_2$& 1.0~$\mu$m& 1.7~$\mu$m& 4.0~$\mu$m& 3.975~$\mu$m& 4.01~$\mu$m& 11~$\mu$m& 26.35~$\mu$m& 27.5~$\mu$m\\
$\tau$& 0.30 & 0.40 & 0.40 & 0.40 & 0.40 & 0.36 & 0.45 & 0.18\\
& & & & & & &\\
5000~K & $8.8\times10^{-7}$ & $2.2\times10^{-6}$& $5.4\times10^{-7}$& $1.6\times10^{-5}$&$1.8\times10^{-6}$&$5.3\times10^{-6}$&$8.4\times10^{-5}$&$2.1\times10^{-5}$\\
3000~K & $1.1\times10^{-5}$ & $1.8\times10^{-5}$& $3.8\times10^{-6}$ &$8.9\times10^{-5}$&$1.0\times10^{-5}$& $2.7\times10^{-5}$& $4.2\times10^{-4}$& $1.0\times10^{-4}$\\
\enddata
\tablecomments{$^a$These white light precisions only directly apply to measurements of reflected light, thermal emission, or broadband spectroscopy. The precision for bona fide spectroscopy is less impressive. For example, if one wanted to obtain a 1.0--4.0~$\mu$m spectrum with 0.1$\mu$m resolution using JWST/NIRSpec ($R\approx20$), then the expected precision would be $\sqrt{30}\approx 5.5$ times worse than the values shown. These photon-counting estimates include imperfect instrument throughput, but not read-noise, dark current, sky background, nor any detector or astrophysical systematics. The precisions for HST and Spitzer are computed the same way and are close to current empirical precisions (\S\ref{comparison}). Nonetheless, it is possible than MIRI observations will be limited by warm telescope background rather than photon counting, since JWST will be passively cooled.}
\end{deluxetable}
|
\section{Introduction}
\IEEEPARstart For binary linear codes, the decoding performance of
belief propagation (BP)-based iterative decoding is dominated by
stopping sets over the binary erasure channel (BEC). Stopping sets were firstly introduced for the
analysis of low-density parity-check (LDPC) codes over BECs \cite{Di-Proietti-TIT-02-06}.
It was shown that the iterative decoder failed to decode to a
codeword if and only if the set of erasure positions was a superset
of some stopping set in the Tanner graph during decoding. In
particular, the number and the size of stopping sets is important
for determining the performance of iterative decoders. Stopping
sets in a small size for the BEC can lead
to small Hamming distance.
The success of stopping sets in analyzing LDPC codes has created a paradigm for researchers to analyze the other codes. For example, Rosnes and Ytrehus introduced the concept of stopping sets to
analyze turbo decoding and proposed turbo stopping set \cite{Rosnes-Ytrehus-TIT-07-11}.
Abdel-Ghaffar and Weber derived an equation based on the number of
stopping sets for a full-rank parity-check matrix of the Hamming
code\cite{Abdel-Weber-TIT-07-09}. Tuvi examined the stopping redundancy
Reed-Muller codes \cite{Tuvi-TIT-06-11}. Wadayama presented the stopping set
of redundant random ensembles \cite{Wadayama-TIT-08-11}.
Recently,
much attention has been given to a class
of error-control codes, Fountain codes, due to their excellent performance, especially in erasure channels and the simplicity of encoding and decoding.
Three typical examples of rateless codes were developed based upon the Fountain codes: Luby Transform (LT) codes\cite{Luby-FOCS-02}, Raptor codes\cite{Shokrollahi-TIT-06-06}, and Online codes. As LT codes own the basic structure of Fountain code family, many studies on error analysis were conducted based on LT codes.
For instance, the error analysis reported in \cite{Karp-Luby-CISIT-04} gave a basic result depending on the exact calculation of the error probability. The works in \cite{Abdulhussein-Oka-TCL-08-06} and \cite{Orozco-Yousefi-QBSC-10} respectively developed stopping criterions so as to detect the earlier decoding termination with a lower cost.
Although the error-control mechanism in Fountain codes facilitated error analysis, two major factors in Fountain codes on the BEC still affect decoding performance. One is the essential codes structure, on which stopping set analysis operates. The other is the effect from the channel characteristic, which has not been effectively resolved yet. Current finite length analysis nonetheless still focused on the former problem -- the error-prone structures of codes. It is much more difficult to give a precise estimation of error-prone patterns.
As Fountain code family belongs to nonsystematic codes, which are different from the existing families like LDPC and Turbo, the conventional stopping set is not applicable. To overcome such a problem, in this study, we focus on the performance analysis when output nodes are erased. We introduce uncorrectable set in Fountain codes in order
to analyze the decoding performance of Fountain codes over the BEC. Furthermore, we also provide the concept of uncorrectable set and analyze
the probability of bit erasure of Fountain codes over the BEC in average
The rest of this paper is organized as follows. Section II briefs the LT code. Section III then describes the Foutain uncorrectable set. Next, Section IV shows the probability of bit erasure followed by the integrated performance analyze in Section V. Conclusions are finally drawn in Section VI.
\section{Preliminaries}\label{sec_2}
\subsection{Principle of LT codes }
Fountain codes include three typical classes: Luby Transform (LT)
codes, Raptor codes, and Online codes. Among these, LT codes is the
basic to construct other families. LT code retains good performance
of random linear fountain code, while drastically reduces the
complexities both in encoding and decoding process. During encoding, LT
divides the uncoded message into $k$ blocks with roughly equal
length. The degree $d$ $(1\leq d\leq k)$ of the next packet is
is randomly chosen. Accordingly, $d$ input symbols are chosen
uniformly at random.
Let $G$ denote a generation matrix for a length given LT code. The
encoding can be represented by:
\begin{equation}
t_i=\sum_{j=1}^{k}x_j\cdot G_{ji}
\end{equation}where $n$ is the code length, $k$ is the length of the input symbol, $t_i$
denotes the $i$th of encoded symbol, $x_j$ denotes the $j$th of encoding symbol. Without loss of generality, this paper considers the symbol is binary.
\subsection{The graph representation of LT codes}
The parity-check matrix $H$ can also be represented by a bipartite
graph $\mathcal {G}=(\mathcal {V}\cup \mathcal {C},\mathcal {E})$, where the set of variable nodes $\mathcal {V}$
represents the codeword symbol and the set of check nodes $C$
represents the set of parity-check constraints satisfied by the
codeword bits, and edges $\mathcal {E}\subset \{(v,c)|v\in \mathcal {V}, c\in \mathcal {C}\}$. First,
let us briefly review conventional stopping sets in LDPC codes. The
concept of stopping sets is proposed based on Tanner graph. A
stopping set $S$ in a code is a subset of the variable nodes in a
Tanner graph for $\mathbb{C}$ such that all the neighbors of $S$ are
connected to $S$ at least twice.
For a given matrix $G_{k, n}$, let $X=(x_1,x_2,...,x_k)$ denote the
encoding symbols. Let $T = (t_1,t_2,...,t_n)$ denote the codeword.
Then, $X\cdot G_{k, n}=T$. In general case, the relation $G_{k,n}H^{\textrm{T}}=0$ is adopted to computer the parity-check matrix $H$.
For binary linear systematic code, parity-check matrix $H$ of LDPC
is obtained according to $G_{k,n}H^T=0$. The matrix $H$ can verify
the estimation value of $X=(x_1,x_2,...,x_k)$ because LDPC is
systematic code; $T$ can be represented by
$T=(x_1,x_2,...,x_k,p_1,p_2,...,p_{n-k})$, where
$p_1,p_2,...,p_{n-k}$ denotes the parity bits. Thus, the encoding
bits $X$ are included in the transmitted bits $T$ and are sent to
the receiver.
However, LT codes are nonsystematic codes which only transmit parity
symbols. $T$ can be represented by $T=(p_1,p_2,...,p_{n-k})$. The
transmitted symbols do not include the encoding symbols $X$. Then, the
matrix $H$ deduced from $G_{k,n}H^T=0$ only verifies the transmitted
symbols $T$ but not to verify the encoding bits $X$. For the sake of
clarity, here we only concern the validity of encoding symbols
$X=(x_1,x_2,...,x_k)$ without caring for the transmitted symbols
$T=(t_1,t_2,...,t_n)$. Therefore, the conventional solution on
parity-check matrix $H$ must be changed in order to suitable to LT
codes.
We propose a method which can create the parity-check matrix of LT
codes: Since the transmitted bits are either lost or correct when
the code transmits on BEC, the all received bits are correct. Let
$P=(p_1,p_2,...,p_r)$ represent the received bits. The partitions of
matrix $G_{k,n}$ corresponding to $P=(p_1,p_2,...,p_r)$ make up of
the matrix $G_{k,r}$. There is,
\begin{equation}
X\cdot G_{k, r}=P.
\end{equation}
Let $\mathcal {G}\{k,\lambda (v),\rho (d)\}$ denote a Fountain code
ensemble, where $k$ is input symbol length, $\lambda (v)$ is the
degree distribution of input node, and $\rho (d)$ is the degree
distribution of output node. From the above analysis, the matrix
$G_{k,r}$ plays the role in the parity-check matrix which can verify the
encoding bits $X=(x_1,x_2,...,x_k)$ in Fountain codes. Hence, for a
particular code $G\in \mathcal {G}$ can also be represented by a
bipartite graph $\mathcal {G}=\{\mathcal{V}\cup \mathcal{C},\mathcal{E}\}$, where the set of variable nodes
$\mathcal{V}$ represents $k$ input nodes, corresponding to the input symbols.
The set of check nodes $\mathcal{C}$ represents the set of parity-check
constraints satisfied by the input symbols, corresponding to the
output symbols,and edges $\mathcal{E}\subset \{(v,c)|v\in \mathcal{V}, c\in \mathcal{C}\}$.
\section{Fountain Uncorrectable Set}
In this section, we analyze the decoding performance of Fountain
code over BEC. It is known that the length of output symbols directly reflects
the performance of iterative decoding algorithms. According to the
above analysis, we build the Tanner graph of $G_{k,n}$, as Shown in
Fig.1. Circular nodes correspond to the input symbols, and the
rectangular nodes correspond to the output symbols. There exists an
edge between the input symbol and output symbol if and only if
$a_{ij}=1$, where $a_{ij}$ denotes the element of generator matrix
in the $i$th row and $j$th column.
We define Fountain uncorrectable set as follows.
\textbf{Definition 1.} An uncorrectable set $\mathcal {U}$ in Fountain codes represents
a subset $\mathcal {V}$ of information nodes. The nodes directly
connected to $\mathcal {V}$ will be erased.
As shown in Fig.1, the different line type expresses an uncorrectable set. For the code in Fig.1, if only $c_1$ is erased, the maximal uncorrectable set is $\mathcal {U}=\{{\O}\}$. If $c_1, c_2$, and $c_3$ are deleted, it means that the connected $v_1$ and $v_4$ cannot be decoded successfully. Accordingly, the uncorrectable set is $\mathcal {U}=\{v_1,v_4\}$.
\textbf{Properties.} An uncorrectable set has the following properties:
1) The union of uncorrectable sets is also an uncorrectable set.
2) Each erasure pattern contains a unquie maximal uncorrectable set which
might be an empty set.
\section{Symbol Erasure Probability}
For a particular code $G$ in a given ensemble $\mathcal
{G}(k,\lambda (v),\rho (d))$, let $P_b(G,\varepsilon)$ denote the
expected bit erasure probability if $G$ is used to transmit over a
BEC with erasure probability $\varepsilon$. Let $E_{\mathcal
{G}(k,\lambda (v),\rho (d))}[P_b(G,\varepsilon)]$ denote the probability of corresponding ensemble average bit erasure. Assuming the
number of erasure bits is $|e|$, where $e$ denotes the pattern of
erasure. There are $E(e)$ output node sockets in some arbitrary but
fixed way with elements from the set $e$. Similarly, there are also
input node sockets in some arbitrary but fixed way with elements
from the set $V(e)$. The element of $V(e)$ cannot be recovered. As
shown in Fig.2, the rectangular nodes with black correspond to the
$|e|$ lost output symbols. Circular nodes with black correspond to
the $V$ input symbols cannot be recovered because output symbols
incident upon them are all lost. Circular nodes with gray correspond
to the input symbols may be recovered because output symbols
incident upon them are not all lost.
\begin{figure}
\noindent\begin{minipage}{\linewidth}
\begin{minipage}[t]{.48\linewidth}
\centering\includegraphics[width=1.5in]{Fig1.eps}\caption{Fountain
uncorrectable set}\label{fig1}
\end{minipage}
\begin{minipage}[t]{.48\linewidth}
\centering\includegraphics[width=1.4in]{Fig2.eps} \caption{There are
$|e|$ output nodes lost, which lead to $V$ input nodes undecodable}
\end{minipage}
\end{minipage}
\vspace{-1em}
\end{figure}
When $|e|$ output nodes are lost, the edges incident upon them are
also lost. The following the number of edges connected to $|e|$
output nodes lost is computed.
\textbf{Theorem 1.} {\em The probability of the number of edges with
$L$ connected to the set $e$ in Fountain codes $\mathcal
{G}\{k,\lambda (v),\rho (d)\}$ is:
\begin{equation}\label{Theorem1}
\mathrm{coef}\bigg(\prod^{d_{\max}}_{i=1}(1+xz^i)^{\rho_in},x^{|e|} z^L
\bigg)\bigg/{n \choose |e|}
\end{equation}
}where $\mathrm{coef}(f(x),x^i)$ denotes the coefficient of $x^i$ in the
polynomial $f(x)$, and $d_{\max}$ denotes the maximal degree of output
nodes. Since
$\mathrm{coef}\big(\prod^{d_{\max}}_{i=1}(1+xz^i)^{\rho_in},x^{|e|} z^L
\big)$ is the number of sets with $|e|$ output nodes and $L$ edges
incident upon them, the total numbers
for selecting the pattern of erasure set $e$ are $\binom{n}{|e|}$. Combing the above
equation, then the edge distribution connected to $e$ is
(\ref{Theorem1}).
Now, we consider the number of input nodes incident upon $E=|E(e)|$.
\textbf{Theorem 2. } {\em The average bit erasure probability for
Fountain ensembles $\mathcal {G}\{k,\lambda (v),\rho (d)\}$ when
transmitting over a BEC with erasure probability $\varepsilon$ is
\begin{equation}\label{Theorem2}
\begin{split}
&E_{\mathcal
{G}(k,\lambda (v),\rho (d))}[P_b(G,\varepsilon)]\\
&=\sum_{|e|}\dbinom{n}{|e|}\varepsilon^{|e|}(1-\varepsilon)^{n-|e|}\\
&\times\sum_{L=1}^{E}\mathrm{coef}\big(\prod^{d_{\max}}_{i=1}(1+xz^i)^{\rho_i
n},x^{|e|} z^L\big)\bigg/{n \choose |e|}\\
&\times\sum^k_{V=1}\frac{V}{k}\times P(e,L,V)\\
\end{split}
\end{equation}
}where $P(e,L,V)$ denotes the probability of the uncorrectable set when the maximum size of uncorrectable edges reaches $L$, and the maximal size of uncorrectable set is equal to $V$.
\begin{proof} Note that for Fountain ensembles $\mathcal {G}\{k,\lambda
(v),\rho (d)\}$, if all edges incident upon an input node belong
to the edges connected to $e$, the uncorrectable set of this input node
is lost. Hence, this input node cannot be recovered.
Assume that the set of $V$ nodes connected with $L$ edges forms the maximal uncorrectable set. Hence, the number of sets with $V$ input nodes and $L$ $(0\leq L\leq
E)$ edges incident upon them is
\begin{equation}
M_1(k,L,V)=\sum_{l\leq L}\mathrm{coef}\bigg(\prod^{v_{\max}}_{j=1}(1+yz^j)^{\lambda _j
k},y^V z^L \bigg)(l)!.
\end{equation}
Let $T(k,n)$ denote the number of the all maps with $k$ input nodes
connected to $n$ output nodes. It is
\begin{equation}
T(k,n)=(\sum^{v_{\max}}_{j=1}j\times \lambda _j \times k)!.
\end{equation}
Let $\mathcal {U}$ be an uncorrectable set if it contains a nonempty subset of the variable nodes such that any regular check node $\verb"c"$, which is connected to $\mathcal {U}$, is connected to $\mathcal {U}$ at least twice. Obviously, there is $\mathcal {U}\subseteq \mathcal {V}$, where $\mathcal {V}$ is the set of variable set.
Let $\mathcal {L}$ be the set that any check node, which is connected to $\mathcal {L}$ but not to $\mathcal {U}$, is connected to $\mathcal {L}$ at least twice. There is $\mathcal {L}\subseteq \mathcal {V}\setminus \mathcal {U}$. Let $\mathcal {K}$ be the maximal uncorrectable set. If every check node that is connected to $\mathcal {L}$ but not to $\mathcal {U}$ at least twice, there is $\mathcal {L}=\mathcal {K}\setminus \mathcal {U}$. If $\mathcal {V}\setminus \mathcal {U}$ does not contain a subset $ \mathcal {L}$ with the property that every check node with is connected to $\mathcal {L}$ at least twice.
Define the functions $Q(k,L,V)$, $N(k,L,V)$ and $M(k,L,V)$ by the recursions
\begin{equation}
Q(k,L,V):=\sum_{V>0} M(k,L,V)
\end{equation}
\begin{equation}
N(k,L,V):=T(k,n)-Q(k,L,V)
\end{equation}
\begin{flalign}
M(k,L,V):=M_1(k,L,V)\cdot N(k-V, E-L,0)
\end{flalign}where $M(k,L,V)$ is the number of maximal uncorrectable set $V$ with $E$ erasure edges. $N(k-V, E-L,0)$ denote the number which the remaining $k-V$ variable nodes with the remaining $E-L$ edges does not contain the uncorrectable set. And there are $k-V$ variable nodes in $\mathcal {V}\setminus \mathcal {U}$ and there are $E-L$ check nodes which are not neighbors of $\mathcal {U}$. We have
\begin{flalign}\nonumber
M(k,L,V)=&\Bigg(\sum_{l\leq L}\mathrm{coef}\bigg(\prod^{v_{\max}}_{j=1}(1+yz^j)^{\lambda _j
k},y^V z^L \bigg)(l)!\Bigg)\\
&\cdot N(k-V, E-L,0).
\end{flalign}
Then, the probability which the maximal uncorrectable set is equal to $V$ is $P(e,L,V)=\frac{M(k,L,V)}{T(k,n)}$.
It is easy to see that the probability is $\dbinom{n}{|e|}\varepsilon^{|e|}(1-\varepsilon)^{n-|e|}$ that pattern erasure is $e$.
The probability that $L$ edges are connected to the $e$ is
$\mathrm{coef}\big(\prod^{d_{\max}}_{i=1}(1+xz^i)^{\rho_i
n},x^{|e|} z^L\big)\bigg/{n \choose |e|}$.
Consequently, (\ref{Theorem2}) holds.
\end{proof}
In particular, when the degree of input node for Fountain codes is
uniformity randomly distribution,the parity matrix has constant row
weight $r$. The next theorem gives the bit erasure probability of
constant row weight ensemble.
\textbf{Theorem 3. } {\em The probability of averaged bit erasure for
Fountain ensembles $\mathcal {G}\{k,r,\rho (d)\}$ when transmitting
over a BEC with erasure probability $\varepsilon$ is}
\vspace{-0.5em}
\begin{equation}\label{Theorem2.2}
\begin{split}
&E_{\mathcal
{G}(k,r,\rho (d))}[P_b(G,\varepsilon)]\\
&=\sum_{|e|}\dbinom{n}{|e|}\varepsilon^{|e|}(1-\varepsilon)^{n-|e|}\\
&\times\sum_{L=1}^{E}\mathrm {coef}\big(\prod^{d_{\max}}_{i=1}(1+xz^i)^{\rho_i
n},x^{|e|} z^L\big)\bigg/{n \choose |e|}\\
&\times\sum^k_{V=1}\frac{V}{k}\times P(e,L,V).\\
\end{split}
\end{equation}
\normalsize \vspace{-1em}
\begin{proof} For Fountain ensembles $\mathcal {G}\{k,r,\rho (d)\}$,
similarly, if all edges incident upon a input node belong to the
edges connected to $e$, the stopping set of this input node is lost,
hence, this input node cannot be recovered.
Assuming the set of $V$ nodes connected with $L$ edges is the maximal uncorrectable set. Hence, the number of sets with $V$ input nodes and $L$ $(0\leq L\leq
E)$ edges incident upon them is
\vspace{-1em}
\begin{flalign}
M_1(k,L,V)=&\sum_{l\leq L}\mathrm{coef}\big((1+yz^r)^k,y^V z^l
\big)(l)!.
\end{flalign}
\normalsize \vspace{-1em}
Like the proof in Theorem 2, we have
\begin{flalign}
P(e,L,V)=\frac{M(k,L,V)}{T(k,n)}
\end{flalign}
where
\begin{flalign}\nonumber
M(k,L,V)=&\Bigg(\sum_{l\leq L}\mathrm{coef}\bigg((1+yz^r)^{k},y^V z^l \bigg)(l)!\Bigg)\\
&\cdot N(k-V, E-L,0).
\end{flalign}
Hence, (11) holds.
\end{proof}
\section{Integrated Performance Analysis of Fountain Codes for BEC}
The performance of Fountain codes for BEC depends on two aspects. One is the essential codes structure, on which stopping set analysis operates. The other is the effect from the channel characteristic, which can be analyzed through the proposed uncorrectable set.
From the Theorem 6 in \cite{Shokrollahi-TIT-06-06}, the probability that $\mathcal
{G}$ has a maximal stopping set of size $s$ is at most
\begin{equation}\label{e:stopping}
S(k,\mathcal {E},s)=\dbinom{k}{s}\sum_{z=0}^n A_s(z,0)\Biggl( 1- \sum_d \rho_d \frac{\binom{n-z}{d}}{\binom{n}{d}} \Biggr)^{k-s}
\end{equation}
where $A_s(z,0)$ denotes the probability that a given subset $\Theta$ of size $s$ of the message nodes is a stopping set, given that $\Theta$ is a stopping set with $z$ check nodes of degree zero and 0 check nodes of degree one.
Eq.(\ref{e:stopping}) represents the decoding error probability due to the structure. From the above analysis, the whole decoding error set includes: (1) the uncorrectable set due to erasure, and (2) the received symbols which form a stopping set. Consequently, the final decoding error probability is
\begin{flalign}\label{e:total}\nonumber
&E_{\mathcal{G}(k,r,\rho (d))}[P_b(G,\varepsilon)]+ E_{\mathcal
{G}(k,r,\rho (d))}[S(k-V,|\mathcal {E}|-L,s)]\\\nonumber
=&\sum_{|e|}\dbinom{n}{|e|}\varepsilon^{|e|}(1-\varepsilon)^{n-|e|}\\\nonumber
&\times\sum_{L=1}^{E}\mathrm {coef}\big(\prod^{d_{\max}}_{i=1}(1+xz^i)^{\rho_i
n},x^{|e|} z^L\big)\bigg/{n \choose |e|}\\\nonumber
&\times\sum^k_{V=1}\frac{V}{k}\times P(e,L,V)\\
+&\sum_s s\times\dbinom{k}{s}\sum_{z=0}^n A_s(z,0)\Biggl( 1- \sum_d \rho_d \frac{\binom{n-z}{d}}{\binom{n}{d}} \Biggr)
\end{flalign}
where $E_{\mathcal {G}(k,r,\rho (d))}[S(k-V,|\mathcal {E}|-L,s)]$ denotes the expectation of that the $|\mathcal {E}|-L$ received nodes and $k-V$ information nodes have a maximal stopping set of size $s$.
\section{Conclusions}
This paper proposes the concept of uncorrectable sets for Fountain codes as the conventional stopping set cannot completely model the performance of the Fountain codes, especially in BECs. The advantage of the proposed mechanism is that it allows the transmission system to analyze the performance of codes when output nodes are erased. The probability of averaged bit erasure over BEC is analyzed. It can help us design efficient codes according to channel states. In the future research, we will design an algorithm with low complexity to rapidly estimate the decoding error probability.
\bibliographystyle{IEEEtran
|
\section{Introduction}
The MiMeS (Magnetism in Massive Stars) project showed that about 7\% of single
OB stars are magnetic \citep{neiner2011,wade2014}. In addition, a similar
proportion of A stars are known to be magnetic \citep{wolff1968,power2007}. The
magnetic fields of OBA stars have simple configurations, stable mainly oblique
dipoles, and their strengths range from $\sim$100 to $\sim$30000 G. More details
on the properties of magnetic fields in hot stars are presented in Grunhut et
al. (these proceedings). Therefore, there seems to be a common origin of
magnetic fields in all hot (OBA) stars. This origin, however, has remained
unknown for a long time.
In cool stars, including the Sun, magnetic fields are generated and sustained by
a dynamo in the convective envelope \citep[e.g.][]{charbonneau2010,brun2004}. As
a consequence, the magnetic fields of cool stars are highly dynamic, and exhibit
variability on a very wide range of timescales. The internal structure of the
star, its rotation, and its accretion state can strongly influence the dynamo,
and ultimately set the broad properties of the magnetic field. The
dynamo-generated magnetic field, in turn, drives the mass loss and angular
momentum loss through magnetised winds and coronal mass ejection processes
\citep[e.g.][]{matt2012,reville2015}. Therefore, a complex interplay exists
between magnetic fields and rotation of cool stars during their whole evolution.
These properties are not observed in hot stars. They do not have a thick outer
convective envelope and a Sun-like dynamo can thus not develop. The origin of
their magnetic field must be found elsewhere.
\section{Dynamo fields?}
\subsection{Dynamo field in the convective core}
A hot star consists of a convective core, a radiative envelope, and a very thin
convective layer just below the surface. Like in the external convective
envelope of low-mass and solar-type stars, a dynamo takes place in the
convective core of intermediate-mass and massive stars. It generates and
sustains a magnetic field, because of the combined action of differential
rotation and turbulent helical flows \citep[i.e. an $\alpha-\Omega$ dynamo
action; e.g.][]{brun2005}. However, the time needed for this field to reach the
surface and become visible is longer than the lifetime of the star
\citep{charbonneau2001, MGC2003}. Moreover, an $\alpha-\Omega$ dynamo would lead
to a correlation between the magnetic field properties and stellar rotation,
which is not observed. Therefore, even if such a core dynamo field exists, it is
not the one that we observe at the surface of hot stars.
\subsection{Dynamo field in the radiative envelope}
Over the last decade, various groups investigated the possibility of creating a
dynamo in the radiative envelope of hot stars
\citep[e.g.][]{spruit2002,zahn2007,arlt2011,rudiger2012,Jouve2014}. Like in
convective regions, the $\Omega$ effect, i.e. differential rotation, transforms
an initial axisymmetric poloidal field into an axisymmetric toroidal field.
Tayler's and other MHD instabilities, that can develop in radiation regions,
then transform this field into a field with a non-axisymmetric component
\citep[][]{Tayler1973,Markey1973,Brun2007}. To maintain the magnetic field, it
is then necessary to close the dynamo loop, by regenerating an axisymmetric
field from the non-axisymmetric field. \cite{spruit2002} proposed to regenerate
the axisymmetric toroidal field, while \cite{braithwaite2006} proposed to
regenerate the poloidal field. For this, both used the shear, but
\cite{zahn2007} showed that axisymmetric fields cannot be regenerated by the
shear alone. Instead, \cite{zahn2007} proposed to close the loop thanks to the
electromotrice force of the instability \citep[see also][]{rudiger2012}. While
this seems to work theoretically, numerical simulations have shown that this
dynamo is not excited or maintained.
Moreover, if an $\alpha-\Omega$ dynamo existed in the radiative envelope, a
correlation would exist between the rotation and the magnetic field properties.
Such a correlation is not observed in OB stars \citep{wade2014}. Consequently,
the possible production of a dynamo field in the radiative envelope of hot stars
must be rejected.
\subsection{Dynamo field in the sub-surface convection layer}
Hot stars have a very thin convective layer just below their surface.
\cite{cantiello2011} showed that a dynamo may develop in this layer. However,
the fields produced this way are of the order of 5 to 50 G for B stars, which is
much weaker than the magnetic fields observed at the surface of these stars.
Moreover, a magnetic field produced by sub-surface convection would likely have
a small-scale and time-dependent structure, while the observed fields are mostly
dipolar and stable. As a result, even if such a sub-surface dynamo field may
exist, it does not correspond to the ones observed at the surface of hot stars.
\section{Fossil fields}
\subsection{Fossil origin of magnetism in hot stars}
During the formation of a hot star, the magnetic field present in the molecular
cloud can get trapped in the star as the cloud collapses. Fossil magnetic fields
are descendants from this seed field \citep{Mestel1999}. During the early stage
of the life of the star, when it is fully convective, this seed field can get
enhanced and sustained by a dynamo. As the radiative core forms and the
convective turbulence disappears in the center of the star, this dynamo field
relaxes onto a large-scale mixed (poloidal+toroidal) stable (possibly oblique)
dipole. Such relaxation processes have been observed in numerical simulations
\citep{braithwaite2004,braithwaitenordlund2006}. Moreover, theoretical work
demonstrated that this mechanism results from a selective decay of the energy of
the system and of ideal MHD invariants such as magnetic helicity
\citep{duez2010}. Finally, to stay stable on long timescale, the field must have
a given ratio of the relative energies contained in its toroidal and poloidal
components \citep{braithwaite2009,duezbm2010}. The external convection zone then
disappears and the star becomes fully radiative with the dipolar fossil field
emerging at its surface. It is possible that the appearance of the convective
core, just before reaching the ZAMS, produces a (extra) tilt of the dipole and
explains why oblique dipoles are observed in basically all hot stars. Indeed,
from ASH 3D MHD simulations, \cite{featherstone2009} showed that the interaction
of a core dynamo with a fossil field in the envelope produces several effects:
it strengthens the core field, it makes the rotation of the envelope more rigid,
and it changes the orientation of the fossil field in the envelope.
Figure~\ref{fossildiagram} shows a diagram of the evolution of fossil fields.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\textwidth]{Neiner_fig.eps}
\caption{Diagram showing the pre-main sequence (PMS) from the birthline to the
ZAMS, with various evolutionary tracks from 1.2 to 8 M$_\odot$ shown with black
solid lines. The PMS is divided in 4 parts indicated with colours, showing 4
stages of the evolution of the structure of the stars and of their fossil magnetic field.}
\label{fossildiagram}
\end{center}
\end{figure}
In addition, \cite{alecian2013} showed that Herbig Ae/Be stars, which are the
precursors of the magnetic Ap/Bp stars, host magnetic fields with a similar
occurence rate and configuration to main sequence hot stars. This indicates that
the fields observed in hot stars are already present at the PMS phase.
Therefore, it is now well established that the magnetic fields of hot stars are
of fossil origin. However, the exact details of the creation and evolution of
these fields, from the molecular cloud to the main sequence, require further
investigations, in particular on the influence of rotation and stellar formation
conditions.
\subsection{Impact of rotation on a fossil field}
Recent theoretical calculations showed that rotation modifies the internal
distribution of the magnetic flux and the stability of fossil magnetic
fields, while it does not modify their surface geometric configuration. Indeed,
fossil fields relax onto mixed dipolar configurations, no matter how fast the
star rotates \citep{emeriau2014}. However, as demonstrated by
\cite{braithwaite2013}, the time needed to reach equilibrium increases with
rotation. Therefore, it is probably more difficult for a rapidly rotating star
to reach a stable dipolar configuration.
This would explain, in particular, why magnetic fields have not been directly
detected in classical Be stars, which rotate close to their breakup velocity
\citep[see][]{wade2014be}. However, rapidly rotating magnetic hot stars do
exist. This is the case, for example, of HR\,7355 \citep{oksala2010} and
HR\,5907 \citep{grunhut2012}. More investigations on the impact of rapid
rotation are thus needed.
When equilibrium is not reached, the star can still host a magnetic field, but
this field will most likely be very weak and on small scales
\citep{auriere2007,lignieres2014,braithwaite2013}. This kind of ultra-weak field
has been observed in Vega \citep{lignieres2009,petit2010} and in a few Am stars
\citep[e.g.][see also Blazere et al., these proceedings]{petit2011}.
\section{Lack of magnetic fields in hot binaries}
BinaMIcS \citep[Binarity and Magnetic Interactions in various classes of
Stars,][]{neiner2013,alecian2014} is an ongoing project that exploits binarity
to yield new constraints on the physical processes at work in hot and cool
magnetic stars. It rests on two large programs of observations with the ESPaDOnS
spectropolarimeter at CFHT in Hawaii and its twin Narval at TBL in France.
BinaMIcS aims at studying the role of magnetism during stellar formation,
magnetospheric star-star (and star-planet) interactions, the impact of tidal
flows on fossil and dynamo fields, its impact on mass and angular momentum
transfer, etc.
In the frame of BinaMIcS, a large survey of magnetism in hot spectroscopic
binary systems with 2 spectra (SB2) has been undertaken. Out of $\sim$200
observed SB2, including at least one star (and most of the time two stars) with
spectral type O, B or A in each system, none were found to host a magnetic
field, while the detection threshold was similar to the one used in the MiMeS
project on single hot stars. This lack of detections in $\sim$400 stars with
BinaMIcS, compared to the $\sim$7\% detection rate in $\sim$500 single stars
with MiMeS, is thus statistically significant: magnetism is less present in hot
binaries than in single hot stars.
This lack of magnetic stars in hot binaries might be related to results obtained
from simulations of star formation by \cite{commercon2011}. They found that the
more magnetic the medium is, the less fragmentation of dense cores there is. In
other words, when the medium is magnetic, it is more difficult to form binaries.
As a consequence, forming a binary with a fossil field is unlikely.
Nevertheless, 6 SB2 systems hosting a magnetic OBA star are known to exist.
These are HD\,5550, HD\,37017, HD\,37061 (NU Ori), HD\,47129 (Plaskett's star),
HD\,98088, and HD\,136504 ($\epsilon$\,Lup). It is possible that magnetic hot
binaries still form sometimes. However, for these 6 systems, only one of the two
components is known to be magnetic, which is puzzling if the stars were formed
simultaneously. A possible explanation is that these binaries were formed in a
later stage of stellar evolution, e.g. by capture, from a magnetic single hot
star and non-magnetic hot star.
\section{Conclusions}
The magnetic fields of single hot stars is of fossil origin, i.e. they are the
descendants of the seed field from the molecular cloud from which the stars were
formed. They are found in $\sim$7\% of all single OBA stars and are mainly
dipolar. Rapid rotation makes it more difficult for fossil fields to reach this
dipolar equilibrium, and this may explain the lack of field detections in
classical Be stars, even though a few examples of rapidly rotating magnetic hot
stars exist.
A few magnetic hot binaries also exist, but magnetism is much less present in
hot binaries than in single hot stars. This might be related to stellar
formation issues: it is more difficult to fragment dense cores when the medium
is magnetic.
These results provide constraints and challenges for formation theories and
simulations to understand the magnetic properties of upper-main sequence stars,
which are very different from those of low-mass and solar-type stars.
\bibliographystyle{iau307}
|
\section{Introduction}
\label{sec_Intro}
In the past decade, with the advance of computation technology and the accessibility of real-world large-scale network data, the exploration and analysis of large-scale network attributes have received tremendous attention in network science \cite{Lewis08} as they disclosed the mysterious masks in nature as well as man-made engineered systems and contrive to answer the fundamental networking problems such as network formulation, dependency, resilience and evolution.
Such networks, consisting of numerous nodes and intricate interconnections embedded with heterogeneous network structures in the graph-theoretic point of view, are renowned as \textit{complex networks} \cite{Albert02,Newman03,Wang03}.
Owing to large-scale network size, extreme volume of empirical network data, and potentially biased network sampling techniques \cite{Kurant11}, explicit analysis on the network structure turns out to be computationally infeasible and theoretically intractable. Consequently, collective network attributes instead of exact network topology are preferable for complex network analysis, and the developed measurement metrics (e.g., clustering coefficient and network centrality) play an essential role in network science and they have been applied to aid the design of communication systems \cite{Cui10}. Among all the network attributes, the degree distribution of the entire network is one of the most salient feature that specifies the link characteristics since the degree distribution is defined as the probability distribution of the number of links of an arbitrarily selected node in the complex network, and it can be specified by a few network parameters.
What is of our particular interest in network science is the study of network resilience \cite{Albert00} (i.e., the extent of network tolerance to node removals) because of its kin relation and assessment to network robustness and connectivity in many networked engineering systems \cite{Menth09,Smith11,CPY12,CPY13GlobalSIP,CPY14ComMag}. Typical examples include but are not limited to denial-of-service (DoS) attacks and jamming attacks. In particular, the U.S. Department of Energy (DOE) has identified attack resistance to be one of the seven major properties required for the operation of smart grid \cite{DOE2007}.
From the bird's-eye view of the entire network, the giant connected component vanishes and the entire network is disintegrated into several small components when the fraction of the removed nodes exceeds certain critical value, which is known as the critical phenomenon of percolation theory in statistic physics \cite{Callaway00}. More importantly, this critical phenomenon can be well mapped to the network robustness and connectivity of many practical networked engineering systems, owing to the network resilience protocols that the network retains its operations as long as a majority of nodes remain its functionality (i.e., most of the nodes are still connected). Throughout this paper, the critical phenomenon for network disruption caused by node removals are used to evaluate the performance of the proposed network defense mechanism and we denote the critical value for network disruption as the \textit{percolation-based connectivity}.
Our physical model is built upon the structure of many practical networked engineering systems where a data fusion is responsible for data inference and decision making as illustrated in Fig. \ref{fig_system_1}.
Although a vast amount of research has been done in analyzing intrinsic network resilience in complex networks and devising efficient intrusion/anomaly detection techniques in practical networked engineering systems separately, a complete and interdisciplinary network robustness analysis including both the intrinsic network resilience as well as the embedded attack detection capability is still poorly understood. In this paper, a sequential defense mechanism is first proposed in complex networks where each node performs individual attack detection and sequentially reports binary attack status (i.e., under attack or not) to the data fusion center as shown in Fig. \ref{fig_system_1}. The data fusion center then sequentially infers the presence of network attacks based on the feedback and makes a final decision when sufficient information has been collected. This mechanism is particularly applicable to networking paradigms with enormous number of nodes and stringent data transmission resources. It is also worth mentioning that the proposed sequential defense mechanism is quite distinct from the traditional data fusion scheme \cite{Varshney96} due to the fact that the network attack may not be a common event to all the nodes in the network as illustrated in Fig. \ref{fig_system_1}. In other words, an intelligent adversary can target at some crucial nodes instead of launching attacks on the entire network to efficiently disrupt the network and reduce the risks of being detected, which therefore hinders the attack inference precision and poses severe threats on the network robustness.
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig_system_1}
\caption{Structure of practical engineering system. A data fusion center is responsible for data inference and decision making based on the feedback data from the network. The solid lines represent localized connections (e.g., physical links in a power grid) and the dashed lines represent delocalized connections (e.g., friends in a social network) in the complex network.
(a) Traditional data fusion scheme. Each node feedbacks its observation on a common event (e.g., channel vacancy or temperature) to the data fusion center for hypothesis test. (b) Intelligent targeted attack. Red solid arrows point to the targeted nodes.
An intelligent adversary leverages the network topology to target the most vulnerable nodes to disrupt the entire network. As this targeted attack is not a common event to all the nodes, most of the nodes are unaware of the attack and therefore it is more difficult to be detected. Consequently, intelligent targeted attack hinders the attack inference precision and poses severe threats on the network robustness.}
\label{fig_system_1}
\end{figure}
The performance of the proposed sequential defense mechanism is evaluated under random and intentional attacks, as random attack plays an identical role of temporal node disfunction and intentional attack refers to malicious attack caused by an adversary.
We provide a parametric plug-in model for performance evaluation on the sequential defense mechanism, and we implement our mechanism in both canonical complex network models and empirical network data to validate its reliability and effectiveness. In addition to analyzing the critical value to sustain percolation-based connectivity via statistic physics approaches \cite{Callaway00}, we would like to point out that our defense mechanism is a general framework which does not depend on any underlaying complex network models but is applicable to any network with arbitrary network structures, provided that the critical value of the network can be realized at hand. The results show that our defense mechanism greatly enhances the network robustness and provides reliable protection against fatal attacks, even in the complex networks with fragile network structure and weak detection capability, which also offer new insights toward network robustness enhancement and robust network design.
The rest of this paper is organized as follows. The related works are summarized in Sec. \ref{sec_Related}. Preliminaries on the percolation-based connectivity and the canonical complex network models are introduced in Sec. \ref{sec_Preliminary}. The system model and sequential defense mechanism are elucidated in Sec. \ref{sec_system}. The critical values under random and intentional attacks are analyzed in Sec. \ref{sec_attack}. The analysis on the sequential defense mechanism is derived in Sec. \ref{sec_performance_analysis}. The performance evaluation of the proposed defense mechanism in canonical complex network models and empirical network data are shown in Sec. \ref{sec_performance}. Sec. \ref{sec_robust} provides discussions for robust network design. Finally, Sec. \ref{sec_con} concludes this paper.
\section{Related Works}
\label{sec_Related}
The intrinsic topological vulnerabilities regarding different network structures under random and intentional attacks were first introduced in \cite{Albert00}. Compared with random attack, intentional attack is shown to be quite effective in disintegrating the entire network by removing a relatively small fraction of nodes with the highest degree in the network. As many real-world networks are observed to possess a heavy-tailed degree distribution, such as the webpage links in World Wide Web (WWW) \cite{Barabasi99}, router maps in Internet \cite{Faloutsos99} and contacts in email networks \cite{Ebel02}, the existence of nodes with a relatively large number of links render such networks particularly vulnerable to intentional attack. Moreover,
it has been demonstrated in \cite{Xiao08} that intentional attack is the most effective attack strategy to disrupt the entire network when the network topology is known by the adversary, which suggests intentional attack to be an ever-increasing threat on the network robustness of many networked engineering systems.
With the aid of statistic physics and percolation theory \cite{Callaway00}, the critical values (i.e., the fraction of removed nodes) for a complex network to sustain random and intentional attacks prior to network disruption are investigated in \cite{Cohen00} and \cite{Cohen01}, respectively, which offer analytically tractable tools for network robustness assessment. Please note that most of the existing research on network robustness against attacks mainly focus on intrinsic topological vulnerabilities while the impacts of implementing network defense mechanisms on the network robustness are still poorly understood.
A naive perfect node protection scheme is proposed in \cite{Xiao11} to prevent a subset of nodes in the network from being attacked, which can be shown as a degenerate case of our proposed model.
A two-player, zero-sum attack and defense game is introduce in \cite{CPY11,CPY12,CPY14JIOT} to alleviate the damage caused by intentional attack by acquiring attack status from each node for attack inference and defense reaction, and the outcome of the game equilibrium is used to evaluate the network robustness. However, this mechanism is not suitable in networked systems with an enormous number of nodes and stringent data transmission resources as frequent data transmissions may deteriorate the system performance and inevitably incur excessive energy consumption.
To provide efficient defense for complex networks, a sequential hypothesis test approach \cite{Wald} is proposed to identify the attack while acquiring as little information from the network as possible. The data fusion center acquires the reports from each node in descending degree order, and therefore it is able to spare the transmissions of the unreported nodes once the process of sequential test terminates, which balances the goals of promptness and accuracy for attack inference.
\section{Preliminaries on Complex Networks}
\label{sec_Preliminary}
\subsection{Percolation-based Connectivity in Complex Networks}
In the realms of network science, the degree (the number of links of a node) distribution plays an essential role in characterizing the collective topological features. With the advance of computation capability and the accessibility of large-scale network data, the long-believed totally random link connections \cite{ER59} have been overthrown by the extraordinary and ubiquitous degree distributions found in a variety of research areas, such as the power-law distribution in the Internet router-level topological maps \cite{Faloutsos99} and the small world phenomenon in social networks \cite{Watts98}. We denote the degree distribution of a complex network by $P(k)$, where $k \in [k_{min},k_{max}]$ and $k_{min}$ ($k_{max}$) is the smallest (largest) degree of the complex network.
From the bird's eye view, the network attack can be mapped to the node removal in the corresponding network graph (all links attached to the removed node are removed as well), and the network is said to be connected in percolation sense if the giant component (the connected component that includes a majority of nodes) still exists after node removal, which we refer to as the percolation-based connectivity. The physical interpretation of the percolation-based connectivity is that owing to the network resilience protocols \cite{Smith11}, the network can continue its main operations under temporal node disfunction as long as most of the nodes are still connected.
According to the seminal work in~\cite{Molloy95}, given the degree distribution $P(k)$ of an arbitrary network, a giant component containing the majority of the nodes exists in the network if $P(k)$ satisfies the criterion $\sum_k k(k-2)P(k)>0$, which is equivalent to the condition
\begin{align}
\label{eqn_condition}
\tau \triangleq \frac{\mathbb{E}[\mathbf{K}^2]}{\mathbb{E}[\mathbf{K}]}>2,
\end{align}
where $\mathbf{K} \in [k_{min}, k_{max}]$ is the random variable representing the degree of a randomly selected node.
With the aid of percolation theory, the critical phenomenon of network disruption occurs if more than $q_c$ fraction of nodes are removed from the network, where the critical value $q_c$ can be estimated when the remaining degree distribution satisfies the criterion $\tau_c=2$. In other words, the complex network transitions from the connected phase to the disconnected phase in percolation sense once more than $q_c$ fraction of nodes are removed. Throughout this paper, the critical value $q_c$ is used to evaluate the network robustness under different network structures and attack schemes. Please note that in the case of small-scale networks, the critical value can be obtained by performing exhaustive node removal experiments (i.e., searching over all possible node removal strategies) instead of using statistic physics approaches (i.e., estimating $q_c$ by degree distribution).
\subsection{Canonical Complex Network Models}
In this paragraph, we introduce three canonical complex network models that serve as the platforms for performance evaluation of the proposed defense mechanism.
\begin{itemize}
\item \textbf{ER network}. In an ER network \cite{ER59}, a link between any arbitrarily selected node pair is present with probability $p_{ER}$. If the network size is large enough, the degree distribution approaches to the Poisson distribution
$P(k)=e^{-\widehat{k}} \frac{\widehat{k}^k}{k!}$, where $\widehat{k}=N \cdot p_{ER}$ is the mean degree of the network and $N$ is the number of nodes in the network.
\item \textbf{Power-law network}. A power-law network possesses a skewed degree distribution $P(k) \sim k^{-\alpha}$, where $\alpha>0$ is the skewness parameter. The heavy tail of the degree distribution suggests the existence of the hub nodes that few nodes have relatively high degree compared with most of the nodes in the network, which well explains the connectivity of the WWW \cite{Barabasi99} or the Internet router maps \cite{Faloutsos99}.
\item \textbf{Exponential network}. An exponential network has its degree distribution $P(k) \sim \frac{1}{\beta}e^{-\frac{k}{\beta}}$, where $\beta>0$ is the mean degree of the network in the large scale network limit.
It is demonstrated in \cite{Sole08} that the degree distribution of the power grid can be characterized by the exponential distribution, both in the national power grid scale and the European power grid scale.
\end{itemize}
\section{System Model}
\label{sec_system}
\subsection{Network Model and Sequential Defense Mechanism}
\label{subsec_system_network_model}
Without loss of generality, we consider the complex network model consisting of $N$ nodes characterized by its degree distribution $P(k)$ and the corresponding critical value $q_c$ to sustain network connectivity against attacks. Each node is equipped with certain detection capability, for instance, intrusion detection techniques \cite{Mukherjee94} or malicious activity filtering \cite{Androulidaki09} for attack inference. The $N$ nodes are sorted in descending degree order, i.e., $k_1 \geq k_2, \geq \ldots \geq k_N$. Let $H_1$ denote the hypothesis that the attack occurs in the complex network (either on one node or several nodes simultaneously), and $H_0$ denote the alternative hypothesis of a null attack (i.e., there is no attack in the network). Based on the nodal detection, every node sequentially reports its binary hypothesis testing decision to the data fusion center in descending degree order since intuitively the removal of nodes with higher degree results in more severe damage to the network robustness.
It is worth mentioning that although enabling local information exchange or cooperative detection among nodes may enhance the attack inference precision, these approaches inevitably increase the computation and data transmission overheads. Throughout this paper, we will concentrate on the degree-based sequential defense mechanism owing to its feasibility and simplicity.
In practice, these local decisions can be transmitted in the header of data packets, or a node is regarded as being attacked if it fails to reply to the periodic beaconing from the data fusion center.
If the attack is confirmed by the data fusion center, network defense schemes such as node quarantine or system renewal will be launched to alleviate the damage, otherwise it keeps surveillance on the collected information.
The advantages of reporting binary attack status for sequential defense in complex networks reside in the feasibility of data transmission and computation complexity in the large-scale networked systems. The enormous network size (e.g., Internet routers or wireless sensors) render simultaneous data transmissions infeasible, especially for wireless networks with scarce radio resources. Moreover, owing to the large network size and limited computational power, analyzing the collected information from all nodes incurs tremendous computation overheads and it may fail to provide timely defense. Consequently, sequential hypothesis test with minimum (one-bit) feedback information is an essential must for attack inference in complex networks because of its least additional communication overheads and timely defense. In other words, the sequential hypothesis test terminates once sufficient information is collected and a final decision is made by the data fusion center so that the system can spare the transmissions of the unreported nodes.
Let $x_i$ denote the attack status reported by the $i$th node. $x_i=1$ when the attack on the $i$th node is detected and $x_i=0$ for null attack on the $i$th node. We assume that the detection capability of each node is identical with probability of detecting an attack $P_D$ and probability of false alarm $P_F$. Each node performs independent hypothesis test such that the joint probability distribution of the first $m \leq N$ reports when $H_j$ is true can be represented as $P(x_1,\ldots,x_m|H_j)=\prod_{i=1}^{m} P(x_i|H_j)$.
In general, we assume $P_D \geq P_F$, otherwise the sequential hypothesis test should be altered for attack inference.
\subsection{Sequential Probability Ratio Test (SPRT)}
\label{subsec_system_SPRT}
Using sequential analysis \cite{Wald}, let $P_{jm}=\prod_{i=1}^{m} P(x_i|H_j)$ denote the probability of obtaining a report sample ($x_1,\ldots,x_m$), the sequential hypothesis test is carried out by performing the probability ratio test with two specified parameters $A$ and $B$. After receiving the report from the $m$th node, if $\frac{P_{1m}}{P_{0m}} \geq A~\left(\frac{P_{1m}}{P_{0m}} \leq B \right)$, then the data fusion center declares the presence of a (null) attack, otherwise it keeps surveillance on the next report for attack inference. Moreover, for purposes of practical computation, it is much more convenient to perform sequential probability ratio test by computing the logarithm of the ratio $\frac{P_{1m}}{P_{0m}}$ instead of the ratio itself as the product of individual tests can be decomposed into sum of the log likelihood ratios.
Let $z_i=\log \frac{P(x_i|H_1)}{P(x_i|H_0)}$ denote the log likelihood ratio of the $i$th report and $\Lambda_m=\sum_{i=1}^{m} z_i= \sum_{i=1}^{m} \log \frac{P(x_i|H_1)}{P(x_i|H_0)}$ denote the cumulative value of the first $m$ reports for hypothesis testing. Consequently, the sequential hypothesis test is terminated with hypothesis $H_1$ ($H_0$) if $\Lambda_m \geq \log A$ ($\Lambda_m \leq \log B$), otherwise the process is continued by taking an additional report. These two parameters $A$ and $B$ can be determined by setting $A=\frac{1-\theta}{\delta}$ and $B=\frac{\theta}{1-\delta}$, where $\delta=P(\textnormal{say}~H_1~\textnormal{when}~H_0~\textnormal{is~true})$ and $\theta=P(\textnormal{say}~H_0~\textnormal{when}~H_1~\textnormal{is~true})$ are the required false alarm and miss detection probabilities at the system level.
\section{Critical Values under Random and Intentional Attacks}
\label{sec_attack}
Incorporating the topological vulnerabilities of the complex network, the critical value $q_c$ to sustain percolation-based connectivity under random and intentional attacks are analyzed with respect to distinct canonical complex network models. For comprehensive analysis and clear reading, only the results are displayed in this section and the mathematical derivations are placed in the appendices.
\subsection{Random Attack}
Random attack on the $q$ fraction of nodes in the network plays an identical role of random node removal. Given the original network degree $\mathbf{K_0}$ of a randomly selected node, the critical value for random attack becomes
\begin{align}
\label{eqn_random_attack_thre}
q_c^{ran}=1-\frac{1}{\tau_0 -1},
\end{align}
where $\tau_0 \triangleq \frac{\mathbb{E}[\mathbf{K_0}^2]}{\mathbb{E}[\mathbf{K_0}]}$ is calculated from the original degree distribution.
For ER network, $q_{c}^{ran-ER}=1-\frac{1}{\widehat{k}}$.
For power-law network, $q_{c}^{ran-POW}=1-\left(\frac{2-\alpha}{3-\alpha}\frac{k_{1}^{3-\alpha}- k_{N}^{3-\alpha}}{k_{1}^{2-\alpha}- k_{N}^{2-\alpha}}-1\right)^{-1}$.
For exponential network, $q_c^{ran-EXP}=1-\left({\frac{k_N^2+2 k_N \beta+2 \beta^2}{k_N+\beta}-1}\right)^{-1}$.
Detailed derivations can be found in Appendix \ref{appex_critical_random}.
\subsection{Intentional Attack}
As demonstrated in \cite{Cohen01}, removing $q$ fraction of nodes with the highest degree in the network is equivalent to randomly removing $\widetilde{q}$ fraction of nodes in the remaining network with new cutoff degree $\widetilde{k}_{max}<k_{max}$. With the continuous degree approximation and the relation $\sum_{k_{max}}^{\infty} P(k)=\int_{k_{max}}^{\infty} P(k)dk=\frac{1}{N}$,
the new cutoff degree $\widetilde{k}_{max}$ can be evaluated from
\begin{align}
\label{eqn_intentional_removal}
\int_{k=\widetilde{k}_{max}}^{k_{max}} P(k) dk=\int_{k=\widetilde{k}_{max}}^{\infty} P(k)dk -\frac{1}{N}=q.
\end{align}
Moreover, $\widetilde{q}$ can be interpreted as the link deletion probability of a randomly selected link leading to a deleted node, which equals the ratio of the number of links belonging to the deleted nodes to the number of links \cite{Cohen01,Newman01}, i.e., $\widetilde{q}=\sum_{k=\widetilde{k}_{max}}^{k_{max}} \frac{k P(k)}{\mathbb{E}[\mathbf{K_0}]}$. By specifying the relations between link deletion probability and targeted node removal in (\ref{eqn_intentional_removal}), the critical value $q_c$ under intentional attack can be obtained by evaluating the critical link deletion probability $\widetilde{q}_c$ with the ubiquitous criterion for percolation-based connectivity in (\ref{eqn_random_attack_thre}).
For ER network, $q_c^{int-ER}=\frac{1}{N}- e^{-\widehat{k}} \frac{\widehat{k}^{\widetilde{k}_{max}-1}}{\left( \widetilde{k}_{max}-1 \right)!}+1-\frac{1}{\widehat{k}}$.
For power-law network, $q_c^{int-POW}=\left(\frac{\widetilde{k}_{max}}{k_N}\right)^{1-\alpha}$.
For exponential network, the critical value can be obtained by solving
\begin{align}
\label{eqn_intentional_EXP_main}
&\left[ 1-\ln \left( q_c^{int-EXP}+ \frac{1}{N}\right) \right]\left(q_c^{int-EXP}+\frac{1}{N}\right) \nonumber \\
&+\frac{k_N+\beta}{k_N^2+2 k_N \beta+2 \beta^2-k_N-\beta}-1=0.
\end{align}
Detailed derivations can be found in Appendix \ref{appex_critical_intentional}.
\subsection{Unified Notations for Attack Schemes and SPRT}
Since each node sequentially reports its one-bit detection result to the data fusion center for attack inference,
let $a_i$ denote the probability of attacking $i$th node, we introduce the unified notations for the aforementioned attack schemes as
\begin{align}
\label{eqn_attack_coefficient_formulation}
P(x_i|H_j)=
\left\{
\begin{array}{ll}
\mathcal{B}(a_i \cdot P_D),~&\textnormal{if}~j=1, \\
\mathcal{B}(P_F),~&\textnormal{if}~j=0,
\end{array}
\right.
\end{align}
where $\mathcal{B}(p)$ is the Bernoulli trial with probability of success ($x_i=1$) equals $p$.
Incorporating the attack schemes, we have
\begin{align}
\label{eqn_attack_coefficient}
a_i^{ran}&=q^{ran},~\forall~i;~ \\
a_i^{int}&=\mathbf{1}_{i \leq \lceil Nq^{int} \rceil}+\frac{P_F}{P_D}\mathbf{1}_{i > \lceil Nq^{int} \rceil} \nonumber \\
&=(1-\frac{P_F}{P_D}) \mathbf{1}_{i \leq \lceil Nq^{int} \rceil}+\frac{P_F}{P_D},
\end{align}
where $\mathbf{1}_E$ denotes the indicator function of the event $E$ and $\lceil x \rceil$ is the smallest integer that exceeds $x$.
\section{Performance Analysis of Sequential Defense in Complex Networks}
\label{sec_performance_analysis}
Given the specified system parameters ($\delta,\theta$), we are interested in the effectiveness and the performance of the proposed sequential defense mechanism against random and intentional attacks in complex networks. Furthermore, knowing the critical value $q_c$, the data fusion center is required to infer the presence of the attack prior to the network disruption. In other words, a final decision has to be made according to the first $M_c=\lceil N \cdot q_c \rceil$ reports for practical implementation purposes, which we refer to as the worst case scenario. Upon the reception of the $M_c$th report, if a final decision has not been reached, the data fusion center declares the presence of attack when $0<\Lambda_{M_c}<\log A$ and declares a null attack when $\log B<\Lambda_{M_c} \leq 0$.
Let $M_j$ denote the expected number of reports required for hypothesis testing when $H_j$ is true.
The proposed sequential defense mechanism is regarded as effective against attacks in the complex network if $M_1 \leq M_c$, i.e., the number of reports required for attack inference is less than the threshold of network disruption, otherwise the defense is in vain since it fails to provide timely defense reaction. We derive the closed-form expressions of $M_1$ for random and intentional attacks, and we prove that for intentional attack, taking additional reports from $m > M_c$ nodes does not improve the performance of the sequential defense mechanism.
\subsection{Random Attack}
\label{subsec_performance_random}
For random attack, with (\ref{eqn_attack_coefficient_formulation}) we have
\begin{align}
P_{1m}&=\prod_{i=1}^m P(x_i|H_1) \nonumber \\
&=(q^{ran}P_D)^{d_m}(1-q^{ran}P_D)^{m-d_m}; \\
P_{0m}&=\prod_{i=1}^m P(x_i|H_0) \nonumber \\
&=(P_F)^{d_m}(1-P_F)^{m-d_m},
\end{align}
where $d_m$ is the number of ones in the first $m$ reports. Simple calculation on $\Lambda_m$ yields
\begin{align}
\label{eqn_performance_random}
\Lambda_m=d_m \log \frac{q^{ran}P_D}{P_F}+(m-d_m) \log \frac{1-q^{ran}P_D}{1-P_F}.
\end{align}
Following the process of SPRT in Sec. \ref{subsec_system_SPRT}, the sequential defense criterion for random attack becomes
\begin{align}
\left\{
\begin{array}{ll}
\textnormal{say}~H_1,~\textnormal{if}~d_m \geq \frac{\log A}{\log\frac{q^{ran}P_D}{P_F}-\log\frac{1-q^{ran}P_D}{1-P_F}} \nonumber \\
~~~~~~~~~~~~~~~~~~+m \frac{\log\frac{1-P_F}{1-q^{ran}P_D}}{\log \frac{q^{ran}P_D}{P_F}-\log \frac{1-q^{ran}P_D}{1-P_F}}, \\
\textnormal{say}~H_0,~\textnormal{if}~d_m \leq \frac{\log B}{\log\frac{q^{ran}P_D}{P_F}-\log\frac{1-q^{ran}P_D}{1-P_F}} \nonumber \\
~~~~~~~~~~~~~~~~~~+m \frac{\log\frac{1-P_F}{1-q^{ran}P_D}}{\log \frac{q^{ran}P_D}{P_F}-\log \frac{1-q^{ran}P_D}{1-P_F}}, \\
\textnormal{keep surveillance},~\textnormal{otherwise}.
\end{array}
\right.
\end{align}
The expected number of reports to identify random attack when $H_1$ is true is
\begin{align}
\label{eqn_performance_random_M1}
M_1^{ran}&=\frac{\theta \log B+(1-\theta) \log A}{\mathbb{E}[z_i|H_1]} \nonumber \\
&=\frac{\theta \log \frac{\theta}{1-\delta}+(1-\theta) \log \frac{1-\theta}{\delta}}{q^{ran}P_D\log \frac{q^{ran}P_D}{P_F} + (1-q^{ran}P_D)\log \frac{1-q^{ran}P_D}{1-P_F}}.
\end{align}
For the worst case scenario, if $m$ is large enough, from central limit theorem we obtain the lower bounds of the probability that the SPRT will terminate by declaring attack or null attack with $m \leq M_c$ reports as \cite{Wald}
\begin{align}
\label{eqn_performance_random_worst_1}
&P(\textnormal{declare~attack})=P(\Lambda_m \geq \log A) \geq 1-\Phi(y_1(M_c)); \\
\label{eqn_performance_random_worst_2}
&P(\textnormal{declare~null~attack})=P(\Lambda_m \leq \log B) \geq \Phi(y_2(M_c)),
\end{align}
where $\Phi(x)$ is the cumulative density function (CDF) of a standard normal distribution, and
\begin{align}
y_1(M_c)&=\frac{\log A-M_c \mathbb{E}[z_i|H_1]}{\sqrt{M_c}\sigma(z_i|H_1)}; \\
y_2(M_c)&=\frac{\log B - M_c \mathbb{E}[z_i|H_0]}{\sqrt{M_c} \sigma(z_i|H_0)};\\
\mathbb{E}[z_i|H_0]&=P_F \log \frac{q^{ran}P_D}{P_F}+(1-P_F) \log \frac{1-q^{ran}P_D}{1-P_F}; \\
\sigma(z_i|H_1)&=\sqrt{q^{ran}P_D(1-q^{ran}P_D)} \log \frac{q^{ran}P_D(1-P_F)}{P_F(1-q^{ran}P_D)};\\
\sigma(z_i|H_0)&=\sqrt{P_F(1-P_F)} \log \frac{q^{ran}P_D(1-P_F)}{P_F(1-q^{ran}P_D)},
\end{align}
where
$\sigma(z_i|H_1)$ and $\sigma(z_i|H_0)$ are the standard deviation of $z_i$ under $H_1$ and $H_0$, respectively.
Moreover, when a final decision needs to be made upon the reception of $M_c$th report, the system level false alarm and miss detection probabilities ($\delta,\theta$) when taking $M_c$ reports are bounded by \cite{Wald}
\begin{align}
\label{eqn_performance_random_worst_3}
\delta(M_c)&\leq \delta+\Phi(y_3(M_c))-\Phi(y_4(M_c)); \\
\label{eqn_performance_random_worst_4}
\theta(M_c)&\leq \theta+\Phi(y_5(M_c))-\Phi(y_6(M_c)),
\end{align}
where
\begin{align}
y_3(M_c)&=\frac{\log A - M_c \mathbb{E}[z_i|H_0]}{\sqrt{M_c}\sigma(z_i|H_0)};\\ y_4(M_c)&=-\sqrt{M_c}\frac{\mathbb{E}[z_i|H_0]}{\sigma(z_i|H_0)};\\ y_5(M_c)&=-\sqrt{M_c}\frac{\mathbb{E}[z_i|H_1]}{\sigma(z_i|H_1)};\\
y_6(M_c)&=\frac{\log B - M_c \mathbb{E}[z_i|H_1]}{\sqrt{M_c}\sigma(z_i|H_1)}.
\end{align}
The aforementioned equations are well-known results from \cite{Wald} applied by the specified parameters $q^{ran}$, $P_F$, $P_D$ and $M_c$. Interested readers are referred to \cite{Wald} for more details.
\subsection{Intentional Attack}
\label{subsec_performance_intentional}
Let $d_m$ denote the number of nodes reporting attack for the first $M=\lceil Nq^{int} \rceil$ reports and $d_{m^{\prime}}$ denote the number of nodes reporting attack starting from the $m^{\prime}$th node ($m^{\prime}>M$). With (\ref{eqn_attack_coefficient}), we obtain
\begin{align}
\label{eqn_performance_intentional_product}
P_{1m}&=P_D^{d_m}\left[(1-P_D)^{m-d_m}\mathbf{1}_{m \leq M} + (1-P_D)^{M-d_m}\mathbf{1}_{m > M}\right] \nonumber \\
&~~\cdot \left[ \mathbf{1}_{m \leq M} + P_F^{d_{m^{\prime}}}(1-P_F)^{m-M-d_{m^{\prime}}} \mathbf{1}_{m>M}\right] \nonumber \\
&=P_D^{d_m}\left[(1-P_D)^{m-d_m}\mathbf{1}_{m \leq M} + (1-P_D)^{M-d_m} \right. \nonumber \\
&~~\cdot \left. P_F^{d_{m^{\prime}}}(1-P_F)^{m-M-d_{m^{\prime}}} \mathbf{1}_{m>M} \right], \\
P_{0m}&=P_F^{d_m}\left[(1-P_F)^{m-d_m}\mathbf{1}_{m \leq M} + (1-P_F)^{M-d_m} \right. \nonumber \\
&~~\cdot \left.P_F^{d_{m^{\prime}}}(1-P_F)^{m-M-d_{m^{\prime}}} \mathbf{1}_{m>M} \right].
\end{align}
The cumulative log likelihood ratio becomes
\begin{align}
\label{eqn_performance_intentional_log_likelihood}
\Lambda_m&=d_m \log \frac{P_D}{P_F}+\left[(m-d_m) \log \frac{1-P_D}{1-P_F}\right]\mathbf{1}_{m \leq M} \nonumber \\
&~~+\left[(M-d_m)\log\frac{1-P_D}{1-P_F}+d_{m^{\prime}} \log \frac{P_F}{P_F} \right. \nonumber \\
&~~+ \left.(m-M-d_{m^{\prime}}) \log \frac{1-P_F}{1-P_F}\right] \mathbf{1}_{m>M} \nonumber \\
&=d_m \log \frac{P_D}{P_F}+\left[(m-d_m) \log \frac{1-P_D}{1-P_F}\right]\mathbf{1}_{m \leq M} \nonumber \\
&~~+ (M-d_m)\left[\log\frac{1-P_D}{1-P_F}\right]\mathbf{1}_{m>M} \nonumber \\
&=d_m \left( \log \frac{P_D}{P_F}-\log \frac{1-P_D}{1-P_F}\right) \nonumber \\
&~~+\left(m \mathbf{1}_{m \leq M}+ M \mathbf{1}_{m > M} \right)\log \frac{1-P_D}{1-P_F},
\end{align}
which suggests that taking additional reports starting from the $m^{\prime}$th node does not help to improve the performance of the sequential defense mechanism as intuitively intentional attack targets only on the first $M$ nodes. The sequential defense criterion for intentional attack is
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig_qc_mean_deg_2}
\caption{Vulnerabilities of different canonical complex network models under random and intentional attacks. Intentional attack is much more effective in disintegrating a complex network compared with random attack. Although power-law networks are resilient to random attack, they are very vulnerable to intentional attack due to the existence of hub nodes with relatively high degree.}
\label{fig_qc_mean_deg_2}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig_M1_ran_q_3}
\caption{Number of reports required for attack detection ($M_1^{ran}$) with respect to $q^{ran}$ under random attack. $M_1^{ran}$ is shown to be a decreasing function of $P_D$ due to better precision in attack inference. $M_1^{ran}$ increases with $P_F$ to distinguish between attack and null attack.}
\label{fig_M1_ran_q_3}
\end{figure}
\begin{align}
\left\{
\begin{array}{ll}
\textnormal{say}~H_1,~\textnormal{if}~d_m \geq \frac{\log A}{\log\frac{P_D}{P_F}-\log\frac{1-P_D}{1-P_F}} \nonumber \\
~~~~~~~~~~~~~~~~~~+\left(m \mathbf{1}_{m \leq M} + M \mathbf{1}_{m > M} \right) \frac{\log\frac{1-P_F}{1-P_D}}{\log \frac{P_D}{P_F}-\log \frac{1-P_D}{1-P_F}}, \\
\textnormal{say}~H_0,~\textnormal{if}~d_m \leq \frac{\log B}{\log\frac{P_D}{P_F}-\log\frac{1-P_D}{1-P_F}} \nonumber \\
~~~~~~~~~~~~~~~~~~+\left(m \mathbf{1}_{m \leq M} + M \mathbf{1}_{m > M} \right) \frac{\log\frac{1-P_F}{1-P_D}}{\log \frac{P_D}{P_F}-\log \frac{1-P_D}{1-P_F}}, \\
\textnormal{keep surveillance},~\textnormal{otherwise}.
\end{array}
\right.
\end{align}
The expected number of reports required to identify intentional attack when $H_1$ is true is
\begin{align}
\label{eqn_performance_intentional_M1}
M_1^{int}=\frac{\theta \log \frac{\theta}{1-\delta}+(1-\theta) \log \frac{1-\theta}{\delta}}{P_D\log \frac{P_D}{P_F} + (1-P_D)\log \frac{1-P_D}{1-P_F}}.
\end{align}
Consequently, the sequential defense mechanism loses its appeals if $M_1>M_c$ since the reports received are insufficient for attack inference before the adversary disrupts the entire network as proved in (\ref{eqn_performance_intentional_log_likelihood}).
Moreover, it is easy to show that the performance of worst case scenario ($M=M_c$) for intentional attack is identical to that of random attack by substituting $M_c=\lceil Nq_c^{int} \rceil$ and $a_i^{int}$ into (\ref{eqn_performance_random_worst_1}), (\ref{eqn_performance_random_worst_2}), (\ref{eqn_performance_random_worst_3}) and (\ref{eqn_performance_random_worst_4}).
\section{Performance Evaluation}
\label{sec_performance}
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig_M1_int_PD_4}
\caption{Number of reports required for attack detection ($M_1^{int}$) with respect to $P_D$ under intentional attack. Compared with random attack, the proposed sequential defense mechanism requires only a few number of reports to target intentional attack, even in the low detection probability regime.}
\label{fig_M1_int_PD_4}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig_ER_3D_5}
\caption{Performance of sequential defense against random attack in ER networks with $P_F=0.001$.}
\label{fig_ER_3D_5}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig_POW_3D_6}
\caption{Performance of sequential defense against random attack in power-law networks with $P_F=0.001$.}
\label{fig_POW_3D_6}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig_EXP_3D_7}
\caption{Performance of sequential defense against random attack in exponential networks with $P_F=0.001$.}
\label{fig_EXP_3D_7}
\end{figure}
In this section, the proposed sequential defense mechanism is employed on canonical complex network models as well as empirical network data to evaluate the system performance and offer new insights on robust network design. The system parameters are set to be $N=10000$, $k_1=1000$, $k_N=1$, $\delta=0.01$ and $\theta=0.001$ without additional specifications.
\subsection{Critical Values of Canonical Complex Network Models}
For fair comparisons between different canonical complex network models, we set the original mean degree to be identical such that $\widehat{k}=c_1 \cdot \frac{k_1^{2-\alpha}-k_N^{2-\alpha}}{2-\alpha}=k_N+\beta$ and accordingly extract the network parameters $\widehat{k}$, $\alpha$ and $\beta$ for ER, power-law and exponential networks. As shown in Fig. \ref{fig_qc_mean_deg_2}, the critical value $q_c$ approaches to $0$ as the mean degree decreases
to $1$ for all canonical complex network models since intuitively a network is prone to disruption if every node has only one link in average. On the other hand, the critical value
increases with the mean degree as every node is able to connect to more nodes in the network in order to strengthen the network connectivity. Compared with random attack, intentional attack is shown to be more effective in disintegrating a network by sabotaging a small fraction of nodes with the highest degree. Moreover, despite the fact that the power-law network is resilient to random attack, the inherently skewed degree distribution render it quite vulnerable to intentional attack due to the existence of hub nodes with relatively high degree, which reveal the bottleneck of network robustness against intelligent attacks.
\subsection{Performance of Sequential Defense Mechanism}
By employing the proposed sequential defense mechanism in the complex networks, we select the number of reports required to identify an attack $(M_1)$ as the performance measure for timely and efficient defense. As shown in Fig. \ref{fig_M1_ran_q_3}, $M_1^{ran}$ is shown to be a decreasing function of $P_D$ because of better precision in attack inference, and $M_1^{ran}$ increases with $P_F$ in order to distinguish attack and null attack. In addition, since $M_1^{ran}$ is also a decreasing function of $q^{ran}$, the optimal attack strategy for an intelligent adversary to disrupt the complex network would be choosing $q^{ran}=q_c^{ran}$ in order to disrupt the network while minimizing the risks of being detected. The performance of sequential defense mechanism against intentional attack is shown in Fig. \ref{fig_M1_int_PD_4}. Similar to random attack, $M_1^{int}$ increases with $P_F$ to validate the presence of attack.
Compared with random attack, the proposed sequential defense mechanism requires only a few number of reports to target intentional attack, even in the low detection probability regime.
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig_prob_q_8}
\caption{Probability of acceptance and rejection under worst case scenario with $P_F=0.001$. The probability of acceptance can be interpreted as the precision for attack inference, and the probability of rejection can be interpreted as the probability for an adversary to disrupt a network.}
\label{fig_prob_q_8}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig_delta_q_9}
\caption{System level false alarm probability under worst case scenario with $P_F=0.001$ and $\delta=0.01$.}
\label{fig_delta_q_9}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig_theta_q_10}
\caption{System level miss detection probability under worst case scenario with $P_F=0.001$ and $\theta=0.001$.}
\label{fig_theta_q_10}
\end{figure}
To gain clear insights on the performance of the proposed sequential defense mechanism, we plot $M_1^{ran}$ and its contours with respect to the network parameters and $P_D$ for ER, power-law and exponential networks in Fig. \ref{fig_ER_3D_5}, Fig. \ref{fig_POW_3D_6} and Fig. \ref{fig_EXP_3D_7}, respectively. The network parameters are associated with the critical values under random attack as discussed in Sec. \ref{sec_attack}. For ER and exponential networks, $M_1^{ran}$ increases with the decrease of $\widehat{k}$ ($\beta$) and $P_D$ as the mean degree is proportional to $\widehat{k}$ ($\beta$) and low $P_D$ hinders the process of SPRT. For power-law networks, more skewed degree distribution (larger $\alpha$) incurs larger $M_1^{ran}$ since the network is prone to disruption as $\alpha$ increases \cite{Cohen00}.
\subsection{Reliability of Sequential Defense Mechanism}
To validate the reliability of the proposed sequential defense mechanism, the performance of worst case scenario is investigated with respect to the critical values to sustain network connectivity. In view of practical implementations, an attack decision has to be made upon the reception of $M_c=\lceil N \cdot q_c \rceil$ reports. The probability of acceptance (declaring attack) and the probability of rejection (declaring null attack) are displayed in Fig. \ref{fig_prob_q_8}. It is observed that the proposed sequential defense mechanism achieves high accuracy as the probability of acceptance (probability of rejection) approaches to $1$ ($0$) at extremely small critical values, and higher $P_D$ enhances the accuracy for attack inference, which validate that the proposed sequential defense mechanism is able to identify the attack with high precision. More importantly, given a critical value of a complex network, the probability of acceptance can be interpreted as the precision of identifying an attack prior to the network disruption, and the probability of rejection can be interpreted as the probability for an adversary to disrupt a complex network.
The system level parameters ($\delta(q_c),\theta(q_c)$) of the worst case scenario are demonstrated in Fig. \ref{fig_delta_q_9} and Fig. \ref{fig_theta_q_10}, respectively. These parameters converge to the desired system level parameters ($\delta,\theta$) at extremely small critical values, suggesting that the proposed sequential defense mechanism offers reliable and effective approaches against random and intentional attacks in complex networks.
\begin{figure}[t]
\centering
\includegraphics[width=3.5in]{fig_empirical_random_11}
\caption{Performance of sequential defense mechanism under random attack with empirical network data. The critical values are $(q_c^{ran},M_c)=(0.9909,322780)$, $(q_c^{ran},M_c)=(0.9673,6000)$ and $(q_c^{ran},M_c)=(0.629,764)$ for the WWW, Internet and EU power grid, respectively.}
\label{fig_empirical_random_11}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig_empirical_intentional_12}
\caption{Performance of sequential defense mechanism under intentional attack with empirical network data. The critical values are $(q_c^{int},M_c)=(0.067,21824)$, $(q_c^{int},M_c)=(0.03,187)$ and $(q_c^{int},M_c)=(0.275,766)$ for the WWW, Internet and EU power grid, respectively.}
\label{fig_empirical_intentional_12}
\end{figure}
\subsection{Empirical Network Data}
As a demonstration, the proposed sequential defense mechanism are implemented in real-world large-scale networks with network parameters extracted from empirical network data collected in \cite{Albert00,Sole08}. The WWW contains $325729$ nodes (webpages) and $1798353$ links with $\mathbb{E}[\mathbf{K}_0]=4.6$. The Internet router-level map contains $6209$ nodes (routers) and $12200$ links with $\mathbb{E}[\mathbf{K}_0]=3.4$. The EU power grid contains $2783$ nodes (power stations) and $3762$ links with $\mathbb{E}[\mathbf{K}_0]=3.4$. The WWW and the Internet are power-law networks with network parameters $\alpha=2.1$ and $\alpha=2.5$, respectively. The EU power grid is an exponential network with network parameter $\beta=1.63$. As shown in Fig. \ref{fig_empirical_random_11}, the number of reports required to identify random attack ($M_1^{ran}$) is lower than the threshold $M_c$, even in the case of weak detection capability (low $P_D$). On the other hand, $M_1^{ran}$ increases with $P_F$ as the data fusion center requires more reports to distinguish between attack and null attack
when the false alarm probability increases. Fig. \ref{fig_empirical_intentional_12} displays the performance of the sequential defense mechanism with respect to $P_D$ and $P_F$ under intentional attack. A surge increase of $M_1^{int}$ is observed with the decrease of $P_D$ and the increase of $P_F$, suggesting that the defense configurations have to be adjusted according to network characteristics in order to guarantee robust and reliable operations of the entire system, especially for the networks which are particularly vulnerable to intentional attack.
\subsection{US Power Grid}
We implement the proposed sequential defense mechanism on the US power grid topology collected in \cite{Watts98}. In addition to degree and random attacks, we also consider betweenness attack, where betweenness of a node is defined as the fraction of all shortest paths passing through the node among all shortest paths between each node pair in the network \cite{Freeman77}. As shown in Fig. \ref{fig_Power_Grid_13}, the network resilience is evaluated in terms of the largest component size when a subset of nodes is removed from the network. Given the expected number of reports required for attack detection $M_1$, if an adversary attacks less than $M_1$ nodes in the network, then the attack will not be detected, which we refer to as the undetectable region. As shown in Fig. \ref{fig_Power_Grid_13}, $M_1$ decreases as $P_D$ increases, and it is shown to be relatively small compared with the network size. Notably, in the undetectable region, most of the nodes are still connected, even with small $P_D$. The results indicate that the proposed sequential defense mechanism is quite effective in attack detection and the network suffers slight connectivity loss in the undetectable region. Note that the perfect protection defense strategy proposed in \cite{Xiao11} is a degenerate case of our proposed mechanism when $P_D \rightarrow 1$ and $P_F \rightarrow 0$. It results in extremely small $M_1$ and suggests that a network can be robust to attacks if perfect protection is plausible.
\begin{figure}[t]
\centering
\includegraphics[width=3.5in]{fig_Power_Grid_13}
\caption{Performance of sequential defense mechanism on US power grid topology \cite{Watts98} under different attack schemes with $P_F=0.005$. The power grid topology contains $4941$ nodes (power stations) and $6594$ edges (power lines). For random attack, the results are averaged over $100$ realizations. The expected number of reports ($M_1$) needed for attack detection is relatively small and it decreases as $P_D$ increases. The proposed sequential defense mechanism is quite effective in the sense the network suffers slight connectivity loss when the number of attacked nodes is less than $M_1$ (i.e., the undetectable region), even for small $P_D$.}
\label{fig_Power_Grid_13}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=3in]{fig_robust_network_14}
\caption{Operation curves of $P_D$ and $P_F$ with respect to a specified network disruption threshold $M_c$ under intentional attack. The feasible operation region are composed of the feasible parameters ($P_D$,$P_F$) such that $M_c \geq M_1^{int}$ in order to guarantee network robustness.}
\label{fig_robust_network_14}
\end{figure}
\section{Discussions toward Robust Network Design}
\label{sec_robust}
Based on the performance analysis of a complex network empowered with the proposed sequential defense mechanism, we provide some insights on robust network design against attacks in complex networks. To guarantee that the process of SPRT terminates before an adversary paralyzes the entire system, the baseline requirement for the network disruption threshold is $M_c \geq \max \left\{ M_1^{ran},M_1^{int} \right\}$, which ensures that the data fusion center can acquire sufficient information for attack inference and make immediate reactions against the attacks prior to the network disruption. In other words, in the network operator's point of view, one has to enhance the critical values of a network under attacks to maintain the operations of the defense mechanism in the system, which can be achieved via different approaches in consideration of the network configurations and implementation costs. Consequently, this paper offers analytically tractable tools for robust network design and network defense performance assessment. Potential approaches to network robustness enhancement are discussed as follows.
\begin{itemize}
\item \textbf{Link Addition}.
As shown in Fig. \ref{fig_qc_mean_deg_2}, adding more links in the network (i.e., increasing the mean degree) strengthens the network connectivity and thereby offers more protection against attacks. Although link addition is a straightforward solution \cite{Ghosh06,Yehezkel12}, the major drawback of link addition is that it may decrease the system revenue if the costs for link constructions are high, such as the transportation systems.
\item \textbf{Topology Adjustment}.
As the network resilience varies from network parameters, the critical value of a network can be modified by topology adjustment while keeping the number of links in the network (i.e., the mean degree) unchanged \cite{Moreira09}, which is especially suitable for networks connected by logical configurations (e.g., the WWW). For an example, as demonstrated in Fig. \ref{fig_qc_mean_deg_2}, adjusting a power-law network to an exponential network enhances the resilience against intentional attack at the cost of decreasing the resilience against random attack, which offers tradeoffs between $M_1^{ran}$ and $M_1^{int}$.
\item \textbf{Detection Capability Enhancement}.
In cases that link addition and topology adjustment are infeasible and thereby the critical values can not be modified, one has to enhance the detection capability to provide reliable network defense against attacks. Take sequential defense against intentional attack as a motivating example, the sequential defense mechanism is able to target intentional attack if the network disruption threshold $M_c$ is no less than $M_1^{int}$. Applying this criterion to (\ref{eqn_performance_intentional_M1}), the feasible parameters $P_D$ and $P_F$ for sequential defense mechanism need to satisfy the inequality
\begin{align}
\label{eqn_robust_network}
&&P_D\log \frac{P_D}{P_F} + (1-P_D)\log \frac{1-P_D}{1-P_F} \nonumber \\
&& \geq \frac{\theta \log \frac{\theta}{1-\delta}+(1-\theta) \log \frac{1-\theta}{\delta}}{M_c}.
\end{align}
The operation curves when the equality in (\ref{eqn_robust_network}) holds given a specified network disruption threshold ($M_c$) are shown in Fig. \ref{fig_robust_network_14}, which can be interpreted as the minimum detection probability ($P_D$) required to perform sequential defense with respect to a false alarm probability $P_F$ and $M_c$. The feasible operation region is composed of the parameters ($P_D$,$P_F$) satisfying the inequality in (\ref{eqn_robust_network}), and the increase of $M_c$ enlarges the feasible operation region since the data fusion center can acquire more reports for attack inference prior to network disruption, even in the low $P_D$ regime.
\end{itemize}
\section{conclusion}
\label{sec_con}
In this paper, a sequential defense mechanism based on sequential hypothesis test is proposed in complex networks with an aim of enhancing the network robustness of networked engineering systems. This mechanism provides timely and efficient defense against random and intentional attacks by sequentially acquiring binary attack status of each node in descending degree order. The data collection process terminates once a final decision has been made by the data fusion center, which is particularly preferable in networking paradigms with stringent data transmission resources. Therefore the low computation complexity and sequential transmission schemes render this defense mechanism compatible to practical networked engineering systems. A parametric plug-in model is proposed to evaluate the performance of the proposed sequential defense mechanism.
By implementing this mechanism on the canonical complex network models as well as the empirical network data extracted from the WWW, the Internet, the EU power grid, and the US power grid topology, the results validate the effectiveness and reliability of this mechanism against fatal attacks. These attacks can be identified with high precision with limited binary attack status reported from a small subset of nodes in the network
and thereby immediate defense reactions can be performed prior to the network disruption, even in the weak topological vulnerability and low detection capability regime. Based on the performance analysis and network configurations, several approaches including link addition, topology adjustment and detection capability enhancement are elucidated to guarantee robust operations of the entire system.
Consequently, this paper provides profound theoretic framework of sequential defense in complex networks and offers new insights on robust network design in complex networks.
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
The applied potential governs the thermodynamics and kinetics of lithium
ion battery (LIB) interfacial processes. Li$^+$ insertion into graphite
anodes to form LiC$_6$ is completed at 0.1~V vs.~Li$^+$/Li(s),
or $\sim$$-2.9$~V vs.~standard hydrogen potential. Before this low voltage
is reached, commonly used battery electrolytes containing ethylene carbonate
(EC), cosolvents, and Li$^+$/PF$_6^-$ salt already decompose
at 0.7-0.8~V vs.~Li$^+$/Li(s). Fortunately, the growth of self-limiting
films (called ``solid-electrolyte interphase,'' or SEI) formed via
electron-injection-induced sacrificial electrolyte degradation passivates
and stabilizes the anode.\cite{review0,review} SEI considerations are
also relevant for new anode materials like silicon. To intepret measurements
and to help devise better artificial SEI/passivation layers,\cite{dillon}
there is a need to use electronic structure computational tools (e.g.,
Density Functional Theory, DFT) to predict the voltage dependence of
liquid-solid interfacial processes. Such a capability will have significant
impact for studying not just LIB,\cite{tateyama,greeley} but also
lithium-air batteries,\cite{bryantsev} water-splitting
processes,\cite{selloni,sprik12} and broad areas relevant to fuel cells,
catalysis,\cite{otani08,casewestern,rossmeisl13,gross} and
electrodeposition.\cite{schmickler,zavadil} In this work, we validate a
recently devised potential calibration scheme,\cite{voltage} apply it to
interfaces between liquid EC and oxidized edge planes of LiC$_6$, and explore
the possibility of electrochemical decomposition of the counter-ion (PF$_6^-$)
used in commercial LIB electrolytes. In the electronic supporting
information (S.I.), we document the significant error that can arise if the
liquid electolyte is omitted in voltage estimates at interfaces.
For non-redox active systems like non-Faradaic supercapacitors, the voltage
difference between two electrodes arises from their different surface charges
mediated by electric double layers in the liquid regino.\cite{borodin} In
contrast, on complex LIB electrode surfaces, what a certain ``applied voltage''
means at the atomic level has not been sufficiently conceptualized, partly due
to the difficulty in probing details at such lengthscales.\cite{harris} On
the theory side, potential calibration has been challenging in periodic
boundary conditions, condensed-phase DFT simulations that depict liquid-solid
interfaces.\cite{sprik12,otani08,casewestern,rossmeisl13,gross}
DFT calculations are performed at constant number of electrons, not constant
voltage. Each DFT simulation cell is associated with one electrode/Fermi
level and is incompatible with a second reference electrode. Furthermore, most
DFT electrochemistry calculations are conducted at T=0~K, which precludes
explicit treatment of liquid solvents, dissolved salts, and, in the majority of
cases, charged interfaces. Several interesting recent advances have coupled
the Poisson-Boltzmann equation or related approaches to DFT at the expense of
introducing vacuum-solid or -liquid interfaces into the
model.\cite{otani08,casewestern} (The vacuum layer in effect serves as the
reference electrode.) Purely condensed phase simulations with no vacuum
region require other methods.
We recently estimated what will be called the ``anode potential''
(${\cal V}$) of inert LiC$_6$ basal planes at finite temperature.\cite{voltage}
The justification is briefly and heuristically described here. Our approach
seeks to mimic LIB experimental processes, where Li-deintercalation occurs
via transfer of Li$^+$ from LiC$_6$, through the liquid-solid interface, to
the liquid electrolyte, and ultimately into the Li metal counter electrode
not explicitly depicted in the simulation. $e^-$ flows in the same direction,
but through the external circuit. Experimentally, it is known that these
charge transfer processes occur at the onset potential of 0.1~V
vs.~Li$^+$/Li(s).
We model this half-cell reaction at the onset of LiC$_6$ delithiation,
\begin{equation}
({\rm LiC}_6)_n \rightarrow ({\rm Li}_{(1-1/n)}{\rm C}_6)_n^-
+ {\rm Li}^+({\rm solv}). \label{eq1}
\end{equation}
While our model does not include counter electrodes, and the excess $e^-$
is left on the anode in the simulation (with finite surface area correction,
see below), Eq.~\ref{eq1} has effectively completed the $e^-$ circuit. This
is because, at equilibrium and in the absence of load-induced voltage
drop, the Fermi level ($E_{\rm F}$) of the Li metal ``counter electrode''
in our thought experiment must be lowered by 0.1~V to coincide with the
$E_{\rm F}$ of LiC$_6$. Under these conditions, the excess $e^-$ on
Li$_{1-\delta}$C$_6$ can start to flow to Li(s). By reaching equilibrium
(tuning the free energy change ($\Delta G_t$) of the Eq.~\ref{eq1} to zero)
via varying the surface charge ($\sigma$) on the electrode surface, we arrive
at the experimentally known half-cell voltage for this reaction. This is the
reference point that can be pegged to measurements. Away from this
${\cal V}$=0.1~V vs.~Li$^+$/Li(s) fixed point, anode voltages are clearly
related to the free energy of monovalent Li$^+$ transfer between LiC$_6$ and
EC liquid ($\Delta G_t$), $\Delta {\cal V}=-\Delta G_t/|e|$, provided that the
interior Li atoms are frozen and not allowed to leak into the electrolyte
(i.e., there is no redox reaction), which is the case in our
simulations.\cite{voltage} The $\sigma$-${\cal V}$ relationship associated
with our frozen-Li basal plane LiC$_6$ model electrode is reminiscent of
those in non-Faradaic supercapacitors.
The {\it ab initio} molecular dynamics (AIMD, or DFT/MD) technique is
used to calculate $\Delta G_t$.\cite{voltage} $\Delta G_t$ calculations
deal with physical ions, not infinitesimal/theoretical test charges, and
therefore circumvent formal/unmeasurable concepts like the difference between
``Volta'' and ``Galvani'' definitions of the potential.\cite{sprik12,pratt92}
Further justification, including a thought experiment on explicitly
including the interface between the liquid electrolyte and the Li metal
reference electrode, and comparison between our approach and
related methods found in the aqueous computational electrochemistry
literature, are given in Ref.~\onlinecite{voltage}.
The present work focuses on LiC$_6$ edge planes through which Li$^+$ can
intercalate and deintercalate. Edge planes are far more technologically
relevant and complex than the proof-of-principle pristine basal plane
considered previously.\cite{voltage} But the same theoretical method can be
used to examine the edge plane voltage. Indeed, Ref.~\onlinecite{voltage}
(Fig.~2) implies that this Li$^+$ transfer protocol can in priciple directly
compare the voltage on any two electrodes. This is because a sufficiently
thick liquid electrolyte region intervening between a basal and an edge
plane can chemically and electrostatically screen them from each other. So
the difference in equilibrium free energies of Li$^+$ transfer from each
electrodes to the electrolyte, divided by $|e|$ should be proportional
to their voltage difference. Since LiC$_6$ is an electronic conductor,
at equilibrium $e^-$ flows between surfaces to make the voltage the
same on basal and edge planes. On pristine edge planes, the all-important
$\Delta G_t$=0 point is determined not only by the electronic charge
($\sigma$) compensated by mobile Li$^+$ in the liquid electrolyte, but also
the fraction of Li occupying the edge sites ($n_{\rm Li}$). At a fixed,
applied potential, these two quantities should adjust themselves to minimize
the free energy of the system. For the illustrative purpose of this
work, we have fixed $\sigma$=0 and only varied $n_{\rm Li}$, and ${\cal V}$
can be considered the instantaneous voltage before the edge Li$^+$ content
can change.
Our predicted edge plane voltage as $n_{\rm Li}$ varies will be corroborated
using explicit electron transfer from fluoroethylene carbonate radical anion
markers. Using AIMD simulations with appropriately calibrated potentials,
PF$_6^-$ is shown to exhibit concerted $e^-$ transfer and bond-breaking
reactions at low voltages, suggesting that under such conditions electrochemical
decomposition may need to be considered. This is significant because it is
widely accepted that PF$_6^-$ decomposes only thermally or due to reaction
with trace water.\cite{kostecki_pf6,plak_pf6} Finally, the simulation cells
used in this work provide information about the electronic orbital
alignment at explicit electrode/electrolyte interfaces. We show that excess
electrons reside in localized states-in-the-gap in the organic carbonate
liquid region, not in delocalized states at the conduction band minimum. The
band structure is {\it not} semiconductor-like (band-state-like) as widely
assumed in the literature.\cite{goodenough}
\section{Method}
Our simplified electrode model consists of a LiC$_6$ strip with all C=O
termination.\cite{ec,mccreery} Four neighboring C=O bonds form a pocket
where Li$^+$ can reside. The Li$^+$ surface density ($n_{\rm Li}$) is unity
if all such pockets are occupied. Figure~\ref{fig1} depicts the periodically
replicated 29.74$\times$14.97$\times$15.06~\AA$^3$
interfacial simulation cell containing a Li$_x$C$_{192}$O$_{48}$ anode slab
and 32~EC molecules. Interior atoms in the anode, with LiC$_6$ stochiometry,
are frozen while C, O, and Li atoms at the edges are allowed to move.
$n_{\rm Li}<1$ electrode configurations are obtained by randomly, and as
uniformly as possible, removing Li atoms from the two LiC$_6$ edge planes.
\begin{table}\centering
\begin{tabular}{ l c c c r r r } \hline
& $n_{\rm Li}$ & $N$(Li) & $\lambda$ & $t_{\rm max}$ &
$\langle dH(\lambda)/d\lambda \rangle_\lambda$ & $\Delta G_t$ \\ \hline
A& 0.417 & 0 & 0.211 & 43.9 & +5.45$\pm$0.11 & \\
B& 0.417 & 0 & 0.789 & 44.4 & -6.63$\pm$0.04 & -1.14 \\ \hline
C& 0.500 & 0 & 0.211 & 52.1 & +5.86$\pm$0.10 & \\
D& 0.500 & 0 & 0.789 & 40.8 & -6.31$\pm$0.10 & -0.78 \\ \hline
E& 0.583 & 0 & 0.211 & 26.1 & +6.59$\pm$0.11 & \\
F& 0.583 & 0 & 0.789 & 20.9 & -6.23$\pm$0.11 & -0.36 \\ \hline
G& 0.417 & 1 & 0.211 & 25.0 & +5.80$\pm$0.05 & \\
H& 0.417 & 1 & 0.789 & 25.2 & -6.61$\pm$0.07 & -0.94 \\ \hline
\end{tabular}
\caption[]
{\label{table1} \noindent
Details of AIMD trajectories for $\Delta G_t$ calculations. $N$(Li) is
the number of mobile Li$^+$ in the liquid region, and $\lambda$ is the
net charge of the Li$^{\lambda +}$ ion frozen in the middle of the liquid.
$t_{\rm max}$ is the total trajectory duration in picoseconds, and include
the first 1~ps equilibration discarded when collecting statistics. The
exceptions are C \&~D, where $t_{\rm max}$ includes the discarded first
10~ps. Integrands and $\Delta G_t$ are in eV; the latter is obtained by
averaging the two integrands, and includes a $-0.39$~eV entropic correction
and a $-0.15$~eV correction for using a 2-point treatment of Li$^+$
solvation.\cite{voltage} To convert ($-\Delta G_t$) to ${\cal V}$, add
0.1~V for Li$^+$/Li(s) and 0.1~V for the ``half $e^-$ rule" (see text).
}
\end{table}
AIMD trajectories and a two-point
thermodynamic integration (T.I.) formula with corrections/extrapolations are
used to compute $\Delta G_t$ (Table~\ref{table1}). These simulations apply
the VASP code\cite{vasp} with PAW pseudopotentials\cite{paw} and the DFT/PBE
functional.\cite{pbe} An energy cut-off of 400~eV, 10$^{-6}$~eV wavefunction
convergence, and $\Gamma$-point sampling are enforced. Spot checks show
that 1$\times$2$\times$2 $k$-point sampling changes the integrands in
$\Delta G_t$ by less than $\sim$0.05~eV, similar to basal plane
cases.\cite{voltage} A thermostat keeps the trajectories at an elevated
T=450~K to improve statistics and prevent EC crystallization. Tritium masses
are substituted for proton masses to permit 1~fs time steps. Compared to
the previous work,\cite{voltage} the predicted $\Delta G_t$ is shifted by
$-0.17$~V to correct the prior neglect of quantum nuclear effect inside
bulk LiC$_6$ and the inadvertent use of T=450~K when adding translational
and vibrational entropies to compare with experiments performed at T=300~K.
The trajectories are initiated using Monte Carlo (MC) simulations in which
anode atoms are frozen in DFT/PBE-optimized configurations. LIB salt
concentration is typically 1.0~M, the static dielectric constant is large
(Debye length $\sim$3~\AA), and electrode surfaces should be screened
from each other even in a small simulation cell. Electrical double
layers should be well-equilibrated to the extent that the simple classical
force fields used are accurate. More MC details are described in the S.I.
\section{Results}
\subsection{Controlling Potential at LiC$_6$ Edge Planes}
Instead of mapping the entire two-dimensional potential
${\cal V}(n_{\rm Li},\sigma)$, we focus on $\sigma=0$. In
Fig.~\ref{fig1}a, the linearly extrapolated ${\cal V}(\sigma=0,n_{\rm Li})$
reaches the LiC$_6$ experimental plateau voltage of 0.1~V vs.~Li$^+$/Li(s) at
$n_{\rm Li}$$\sim$0.69. If $n_{\rm Li}$$>$0.69, $\sigma >0$ would be needed
to raise ${\cal V}(\sigma,n_{\rm Li})$ back to the green line and achieve
the experimental potential associated with LiC$_6$. This merely means that
some of the edge Li must then be considered Li$^+$ ions --- not atoms ---
compensated with mobile PF$_6^-$ further away in the electrolyte.\cite{conceit}
Note that a ``half-electron rule'' vertical shift has been included to convert
$\Delta G_t$ (Table~\ref{table1}) to ${\cal V}(\sigma=0,n_{\rm Li})$
(Fig.~\ref{fig1}a). T.I.~calculations involve moving a Li$^+$ from LiC$_6$ to
the middle of the solvent region of a charge-neutral simulation cell, leaving
an $e^-$ behind. Along the T.I.~path, an average of half an $e^-$, or
$\sigma$=$-|e|$/($4A)$, exists on the two electrode surfaces, where $A$ is
the lateral surface area. We assume this excess charge is uniformly
distributed on the conducting electrode surfaces, in accordance with classical
electrostatic predictions; see the S.I. of Ref.~\onlinecite{voltage} for
analysis of instantaneous charge distributions. The effect of this average
$\sigma$ is estimated by finite difference and subtracted from
Fig.~\ref{fig1}a, as follows. We inject one mobile Li$^+$, compensating an
excess $e^-$ on the electrode, and recompute $\Delta G_t$ (Table~\ref{table1},
trajectories G-H). The simulation cells remain charge-neutral.
$\Delta {\cal V}$ is found to be $-0.20$~V ($\delta \sigma/ \delta
{\cal V}$=$17.8$~$\mu$C/(cm$^2 V)$) after adding the one $e^-$. This is
smaller in magnitude than that in basal plane simulation cells.\cite{voltage}
To undo this surface charging effect, a $+0.1$~V correction is thus applied.
The shift vanishes at large $A$, and represents an extrapolation to infinite
system size. We stress that experimentalists can impose a potential without
knowing details about the surfaces, but DFT calculations work differently;
$\sigma$ and $n_{\rm Li}$ need to be adjusted to arrive at the desired voltage.
In the absence of the liquid electrolyte, the potential at zero surface
charge is directly related to the work function of the electrode in vacuum.
In the S.I., we report work functions and show that the potential in vacuum,
predicted as a function of $n_{\rm Li}$, is significantly modified by the
inclusion of the liquid electrolyte in the main text. This observation
dovetails with predictions that work functions of metals can vary by $\sim$1~V
when their surfaces are covered with a monolayer of water\cite{gross} or
organic molecules.\cite{brocks}
\begin{figure}
\centerline{\hbox{ (a) \epsfxsize=3.00in \epsfbox{fig1a.ps} }}
\centerline{\hbox{ (b) \epsfxsize=3.00in \epsfbox{fig1b.ps} }}
\caption[]
{\label{fig1} \noindent
(a) ${\cal V}$ at zero electronic charge as a function of edge plane Li
content. The dashed line is a linear extrapolation. (b) A snapshot of
the edge plane interfacial simulation cell. C, O, H, and Li are grey,
red, white, and blue, respectively.
}
\end{figure}
Although linear extrapolation is not expected to hold over the entire
$n_{\rm Li}$ range, ${\cal V}(n_{\rm Li},\sigma=0)$ appears to extrapolate
to a value substantially below 0~V vs.~Li$^+$/Li(s) at $n_{\rm Li}$=1.
Negative potentials relative to Li$^+$/Li(s) are below the operating conditions
of LIB anodes. This strongly suggests that $\sigma$=0 and $n_{\rm Li}$=1
yield an overpotential for electrolyte decomposition. In the literature, C=O
edge AIMD simulations have been reported at $n_{\rm Li}$=1 and $\sigma$=0, and
EC molecules are found to decompose in picosecond via two different 2-$e^-$
mechanisms, releasing CO and C$_2$H$_4$ gases, respectively.\cite{ec} What
are the potential dependences of these two competing processes? Recently,
it has been predicted that the 2-$e^-$ CO-releasing route has a far lower
reaction barrier than C$_2$H$_4$ generation in bulk liquid electrolyte
regions.\cite{cpl} It is therefore unlikely that C$_2$H$_4$ should be
a dominant 2-$e^-$ product unless there is an overpotential. Consistent
with this deduction, Fig.~\ref{fig1} indeed suggests the Ref.~\onlinecite{ec}
system, where $n_{\rm Li}=1$, is at ${\cal V}<0$~V. If this were not the
case, mostly CO products are expected. Note that two-electron reduction can
yield C$_2$H$_4$ gas if $e^-$ are added sequentially, not simultaneously,
separated by milliseconds.\cite{cpl} This timescale is far larger than
our AIMD trajectory durations.
\subsection{Validating Predicted Potential: Electron Transfer}
The predicted potential should not depend on whether Li$^+$ or $e^-$ moves
across the interface. Next we demonstrate that ${\cal V}(\sigma=0,n_{\rm Li})$,
calibrated using Li$^+$ transfer above, is also consistent with $e^-$ transfer.
In the middle of the liquid region is placed a fluoroethylene carbonate
(FEC$^-$) radical anion (Fig.~\ref{fig2}), which is an effective electrolyte
additive molecule for improving SEI on anode surfaces.\cite{fec} In
charge-neutral FEC (Fig.~\ref{fig2}c), the carbonyl carbon (C$_{\rm C}$) is
coplanar with the three O~atoms. In contrast, FEC$^-$ is bent
(Fig.~\ref{fig2}d), with C$_{\rm C}$ now $sp^3$ hybridized. This leads to a
large ``reorganization energy'' in the Marcus theory sense,\cite{marcus}
discussed below. We define the C$_{\rm C}$ out-of-plane displacement $R$
as the scalar product between (i) the normalized vector product connecting
the three O-atoms, and (ii) $({\bf R_{\rm C}}-{\bf R_{\rm Oave}})$, where
``Oave'' is the mean position of the three oxygen atoms.
$R$=0.0~\AA\,~and~0.4~\AA\, in the optimized FEC and FEC$^-$ geometries.
Monitoring $R$ grants easy access to FEC charge states. To initiate AIMD
simulations, we first conduct MC using classical force fields (see the S.I.),
then run AIMD for 1.0~ps while freezing all FEC$^-$ atoms, and finally remove
FEC constraints at ``$t$=0'' of AIMD trajectories. The time evolution of
$R$ are shown in Fig.~\ref{fig2}.
\begin{figure}
\centerline{\hbox{ \epsfxsize=3.00in \epsfbox{fig2abc.ps} }}
\centerline{\hbox{ (d) \epsfxsize=1.50in \epsfbox{fig2d.ps} }
\hbox{ (e) \epsfxsize=1.50in \epsfbox{fig2e.ps} }}
\caption[]
{\label{fig2} \noindent
(a)-(c) $R$ as a function of time at various anode potentials ${\cal V}$.
Different line shapes denote trajectories with different initial
configurations. (d)\&(e) Optimized FEC and FEC$^-$ molecules. Out-of-plane
coordinates $R$ are illustrated. Purple circles represent F~atoms.
}
\end{figure}
The reduction potential of FEC in bulk liquid regions is predicted to be
$\Phi$=0.58~eV when it is not coordinated to Li$^+$ (S.I.). If
${\cal V}(\sigma=0,n_{\rm Li})$$<$$\Phi$, the excess $e^-$ should stay on
FEC$^-$, ultimately leading to a second reduction of FEC$^-$ and rapid
decomposition.\cite{fec,note3} In the opposite case, the excess $e^-$
should be transferred to the electrode. The S.I.~shows that this expectation
is always satisfied with different initial configurations on less
electrochemically active basal plane surfaces. At edge planes, this test is
successful, but less than 100~\% of the time. When ${\cal V}$$\sim$0.18~V
($n_{\rm Li}$=0.667), lower than $\Phi$=0.58~V, FEC$^-$ persists in all
trials until it absorbs a second $e^-$ and decomposes (Fig.~\ref{fig2}a).
This dovetails with our expectation. When ${\cal V}$=0.56~V, very close to
$\Phi$, FEC$^-$ is stable for hundreds of femtoseconds until it decomposes in
two out of four trials; in the remaining two cases FEC$^-$ loses its electron
(Fig.~\ref{fig2}b). Given our statistical uncertainties, this 50/50 split
in the outcome is reasonable. When $\Phi$$<$${\cal V}$=0.98~V, FEC$^-$
should give up its excess $e^-$. This is observed 3 out of 4 times
(Fig.~\ref{fig2}c).
On the whole, Fig.~\ref{fig2} demonstrates that ${\cal V}$ calibrated using
Li$^+$ transfer also correctly governs $e^-$ transfer. The one glaring
``error'' at ${\cal V}$=0.98~V is apparently due to overly rapid C-O bond
breaking in FEC$^-$ predicted using the PBE functional. This occurs within
100~fs (arrow in Fig.~\ref{fig2}c; the predicted MP2 barriers for breaking
this bond in EC$^-$ and FEC$^-$ are consistent with slower reaction
rates.\cite{cpl,fec}). Afterwards, the FEC$^-$ ring cannot be reformed in
AIMD timescale even if the anode potential favors it. It is also worth
pointing out that DFT/PBE allows unphysical splitting of an excess $e^-$
between the electrode and the redox center, artificially accelerating $e^-$
transfer rate,\cite{wtyang} and may make Fig.~\ref{fig2} unduly sensitive to
initial configurations. Instantaneous fluctuations in the potential
experienced by redox centers are physical and real, but $e^-$ transfer occurs
over a finite timescale that partially averages out the fluctuations. Despite
this, the overall trend of ${\cal V}$ and $\Phi$ correspondence is correctly
predicted: as ${\cal V}$ decreases, FEC$^-$ retains its excess $e^-$ more
readily.
\subsection{Excess Electrons Form Localized States in the Gap}
Figure~\ref{fig3} depicts spatially-demarcated DFT Kohn Sham spin-orbital
levels at $t$=0~ps in two trajectories taken from Fig.~\ref{fig2}b~\&~c.
In panel (a) (${\cal V}$=0.56~V), the highest occupied molecular orbital
(HOMO) is localized on FEC$^-$ --- reminiscent of a polaron in solid state
physics --- just below the Fermi level ($E_{\rm F}$). FEC$^-$ is found to
decompose from this configuration. In panel (b), $V$=0.98~V, higher than
$\Phi$=0.58~V; the localized orbital resides above $E_{\rm F}$, and FEC$^-$
loses its excess $e^-$ rapidly in this trajectory. Fig.~\ref{fig3}
illustrates that, for organic carbonate-based liquid electrolytes, the
thermodynamic onset of electrolyte electrochemical reduction and SEI formation
occurs when $E_{\rm F}$ coincides with the organic carbonate {\it localized}
orbital found in the gap between liquid HOMO and liquid LUMO (lowest
unoccupied molecular obrital or liquid conduction band minimum).\cite{note4}
SEI formation does {\it not} begin, thermodynamically speaking, when the
electrolyte LUMO coincides with the anode Fermi level, as has been widely
assumed in the literature.\cite{goodenough} The LUMO lies above the organic
carbonate localized orbital and exhibits substantial statistical fluctuations;
however it may influence $e^-$ transfer kinetics if electron transfers
from the anode, through a substantial liquid layer, to FEC in the bulk
liquid electrolyte region, instead of towards FEC that diffuses near
the anode surface.
The above discussion adopts the solid state physics language often used in the
battery community. It is important and of great interest to
reconcile our study with molecular electrochemistry terminology\cite{saveant}
less often featured in battery studies. According to Marcus
theory,\cite{marcus} $e^-$ injection into FEC is accompanied with
reorganization (free) energies ($\lambda$). For FEC$^-$, $\lambda$ contains
a large intramolecular component, and is not solely due to ``outer shell''
solvation effects. The ``polaronic shift'' of the HOMO of FEC$^-$, from
above the liquid conduction band edge if FEC$^-$ were flat, to within the
liquid gap due to FEC$^-$ geometry change and dielectric solvation, is a
non-trivial manifestation of this $\lambda$. Quantitatively, the vertical
electron affinity should be at a value $\lambda$ above the molecular
reduction potential $\Phi$.\cite{adriaanse} Unfortuately, it is difficult
to compute $\lambda$ in simulation cells with electrodes which are electron
conductors. For example, the liquid electrolyte LUMO at frozen liquid
geometry tend to reside above the Fermi level. As a result, injecting
an $e^-$ to the system to calculate the vertical electron affinity immediately
populates the Fermi level of the electrode, not the the electrolyte, unless
constrained DFT methods are used in the electrolyte region. In this sense,
the systems considered in this work may differ from electrodes with a
significant band gap, like TiO$_2$.\cite{cheng2014} Thus all AIMD simulations
in this work report adiabatic free energy changes and redox potentials, not
vertical excitations that include $\lambda$. To estimate $\lambda$, we have
applied a localized basis set and dielectric continuum approach similar to
Ref.~\onlinecite{cpl}. We find that the total $\lambda$ for FEC is 3.2~eV.
This is somewhat larger than that of the structurally similar EC.\cite{cpl}
Our predicted $\lambda$ can be compared with future optical measurements in
organic solvents analogous to those in aqueous media.\cite{adriaanse}
Incidentally, the wide separation between the localized excess $e^-$ orbital
in FEC$^-$ and the liquid electrolyte HOMO minimizes hybridization between
the localized state and solvent orbitals, which has been shown to be
important for accurate DFT treatment of anions.\cite{adriaanse,galli}
\begin{figure}
\centerline{\hbox{ \epsfxsize=3.00in \epsfbox{fig3.ps} }}
\caption[]
{\label{fig3} \noindent
Instantaneous Kohn-Sham orbital decomposed on to atoms at their
$x$-coordinates at $t$=0~ps in Fig.~\ref{fig2}b\& c. $0<x<13.4$~\AA\,
denotes the metallic LiC$_6$ region; outside that range resides the liquid
electrolyte which has a wide band gap. The Fermi level is at $E$=0.0~eV.
Panels a~\&~b depict the majority and minority spin channels of on trajectory
each in Fig.~\ref{fig2}a and Fig.~\ref{fig2}b. The excess $e^-$ ``polaron''
orbital of FEC resides in the majority spin channel, near $E_{\rm F}$, in
the electrolyte region.
}
\end{figure}
\subsection{PF$_6^-$ Electrochemical Reduction}
Finally, this potential-calibration technique is used to investigate possible
PF$_6^-$ reductive decompostion on anode surfaces as ${\cal V}$ varies. A
recent AIMD study has predicted rapid reductive decomposition of PF$_6^-$ into
LiF and PF$_n^{q-}$ fragments, $n\leq$4, at liquid EC/graphite edge
interfaces.\cite{kent} Seemingly consistent with this prediction, simple
cluster-based calculations suggests that the LiPF$_6$ ``molecule'' has
a reduction potential of $1.46$~V (S.I.), far above LiC$_6$ potentials during
battery charging. However, such a PF$_6^-$ reduction
signature has not been observed in cyclic voltametry. Experimentally, the
SEI formed in LiPF$_6$-based electrolytes is known to contain LiF from
PF$_6^-$ breakdown, but it has been widely accepted that
PF$_6^-$ decomposes thermally or due to reaction with trace water over a
period of hours,\cite{kostecki_pf6,plak_pf6} not electrochemically in seconds.
To reconcile these observations, we note that, unlike FEC reduction, $e^-$
transfer to PF$_6^-$ occurs in concert with P-F bond breaking. This is
reminiscent of alkyl halide reduction,\cite{saveant} which is clearly voltage
dependent. Previous PF$_6^-$ modeling work\cite{kent} has not specified its
anode potential. In Fig.~\ref{fig4}, we apply AIMD potential-of-mean-force
($\Delta W(R)$) techniques to estimate the free energy barrier ($\Delta G^*$)
of PF$_6^-$ decomposition at two different potentials.
The trajectories include a PF$_6^-$ pre-equilibrated at each graphite edge,
charge-balanced by Li$^+$ in the electrolyte. The 3-atom reaction coordinate is
$R'$=$|{\bf R_{\rm P}}-{\bf R}_{\rm F}|$$-$$|{\bf R_{\rm Li}}-{\bf R_{\rm F}}|$;
P, F, and Li atoms are chosen such that they are in position to react at
$t$=0~ps. Increasing $R'$ is correlated with F$^-$ transfer from P to an
edge Li. Umbrella sampling potentials of the form $A(R'-R_o)^2/2$ are
enforced, where $B$=4~to~10~eV and $R_o$ span the range between the reactant
and the transition state. See the S.I.~for details.
Figure~\ref{fig4}a shows that, at $\cal V$=0.56~V ($n_{\rm Li}$=0.583),
$\Delta G^*$ is at least 0.9~eV. As soon as one P-F bond breaks completely,
$e^-$ is transferred, other F$^-$'s detach from the P-atom
spontaneously (Fig.~\ref{fig4}d), and these irreversible steps render a
quasi-equilibrium sampling of $\Delta W(R)$ in the barrier top region
impossible. In contrast, at ${\cal V}$=$-0.21$~V ($n_{\rm Li}$=0.75), the
barrier appears not much higher than 0.2~eV. Fig.~\ref{fig4} indicates
that PF$_6^-$ electrochemical reduction may occur during the initial stage
of SEI formation if the anode is at sufficiently low potentials, although
this process faces competition from solvent reductive decomposition.
\begin{figure}
\centerline{\hbox{ (a) \hspace*{0.02in} \epsfxsize=1.80in \epsfbox{fig4a.ps} }
\hbox{ (b) \epsfxsize=1.80in \epsfbox{fig4b.ps} }}
\centerline{\hbox{ (c) \epsfxsize=1.80in \epsfbox{fig4c.ps} }
\hbox{ (d) \epsfxsize=1.80in \epsfbox{fig4d.ps} }}
\caption[]
{\label{fig4} \noindent
(a) Potentials-of-mean-force (PMF) of PF$_6$ electrochemical decomposition, as
functions $R'$ at estimated 0.56~V (red,
$n_{\rm Li}$=0.583) and $-$0.21~V (blue, $n_{\rm Li=0.750}$), respectively.
The solid line portions depict actual data. The barriers must reside in
the dashed line regions because PMF calculations there lead to spontaneous
PF$_6^-$ decomposition in picosecond timescale. (b) Reduced
Li$^+$:F$^-$:PF$_5^-$ fragment from static, localized basis-set calculations.
(c)\& (d): AIMD snapshots in PMF trajectories at $V$=0.56~V, $R'$$\sim$$-0.21$
and $R'$$\sim$$+0.5$\AA, respectively. The rightmost PF$_6^-$ has decomposed
in panel (d); recall that periodic boundary conditions are used.
}
\end{figure}
\section*{Conclusions}
In conclusion, we have calibrated the anode potential
(${\cal V}(\sigma=0,n_{\rm Li})$) of lithium intercalated graphite edge planes
at zero surface electronic charge ($\sigma$=0) as a function of the edge
Li content ($n_{\rm Li}$) by computing the free energy of Li$^+$ transfer
between electrode and liquid electrolyte. The estimated ${\cal V}$ is
shown to be reasonable by correlating with observed electron transfer from
reduced fluoroethylene carbonate (FEC$^-$) radical anions inserted into the
liquid region. Electrochemical reduction of PF$_6^-$ at the pristine edge
plane is shown to be viable at low potentials and to exhibit
potential-dependent kinetics. This reduction pathway may need to be
considered during SEI formation, in addition to thermal/impurity water-induced
PF$_6^-$ decomposition routes widely accepted in the literature. In the
future, optimization of the free energy with respect to all surface
parameters ($N_{\rm Li}$ and $\sigma$) will be performed, and our method
will be used to study the dynamics of Li$^+$ insertion into
passivated anodes as a function of the applied potential.
\section*{Acknowledgement}
We thank Kevin Zavadil and Jun Cheng for interesting discussions. This work
was supported by Nanostructures for Electrical Energy Storage (NEES), an
Energy Frontier Research Center funded by the U.S.~Department of Energy,
Office of Science, Office of Basic Energy Sciences under Award Number
DESC0001160. Sandia National Laboratories is a multiprogram laboratory
managed and operated by Sandia Corporation, a wholly owned subsidiary of
Lockheed Martin Corporation, for the U.S.~Deparment of Energy's National
Nuclear Security Administration under contract DE-AC04-94AL85000.
$\dagger$ Electronic supporting Information available. See DOI:TBA.
|
\section{Introduction}
\label{introduction}
Flash memory has become a leading storage media thanks to its many
excellent features such as random access and high storage
density. However, it also faces significant reliability and endurance
challenges. In flash memory, programming cells with lower charge
levels to higher levels can be done efficiently, while the opposite
requires erasing the whole block containing millions of cells. Block
erasure degrades cell quality, and current flash memory can survive
only a small number of block erasures. To mitigate the reliability and
the endurance issues, this paper studies write-once memory (WOM) codes
that combine erasure-free information rewriting and error correction.
WOM was first studied by Rivest and Shamir~\cite{RivSha82}. In the
model of WOM, new information is written by only increasing cell
levels. Compared to traditional flash, WOM-coded flash achieves higher
reliability when the same amonut of information is written, or writes
more information using the same number of program/erase (P/E) cycles.
We illustrate these benefits using Fig.~\ref{fig:wom-error}, where we
show the bit error rates (BERs) of the first write and the next
rewrite measured for the scheme of this paper in a 16nm flash chip.
When using the standard setting for error correcting codes (ECCs),
flash memory can survive $14000$ P/E cycles without an ECC decoding
failure. Using a code constructed in this paper that allows user to
write $35\%$ more information, we only need $10370$ P/E cycles to
write the information. Notice that the raw BER at $10370$ P/E cycles
is much lower than that at $14000$ P/E cycles, hence ECC decoding will
have much lower failure rate, which leads to higher reliability. On
the other hand, if we use WOM until ECC fails at $14000$ P/E cycles,
the total amount information that is written requires $18900$ P/E
cycles to write in traditional flash. WOM codes can also be used for \emph{scrubbing} the memory. In this use, the memory is read periodically, to correct errors that were introduced over time. The errors are corrected using an ECC, and the corrected data is written back using a WOM code (see~\cite{LiJiangBru14}).
Many WOM constructions were
proposed recently. Codes with higher rates were
discovered~\cite{JiaLanSchBru13}\cite{YaaKaySieVarWol12}, and codes
that achieve capacity have also been
found~\cite{BurStr13}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.6\linewidth, angle=270]{wom-rber.eps}
\end{center}
\vspace{-0.3cm}
\caption{\label{fig:wom-error} The raw BERs when using the proposed
rewriting scheme.}
\vspace{-0.7cm}
\end{figure}
In this paper, we propose an alternative construction of WOM
codes. Our scheme differs from the WOM codes mentioned above mainly in
two aspects. First, we focus on a specific rewriting model with two
writes, where only the second write uses WOM codes. Such rewriting
scheme has no code rate loss in the first write, and recent
experimental study has demonstrated its effectiveness on improving the
performance of solid state drives~\cite{YadYaaSch15}. Note that, the
model of this rewriting scheme is not only an instance of the general
WOM model~\cite{Hee85}, but also an instance of the model studied by
Gelfand and Pinsker~\cite{GelPin80}. Second, our construction is based
on binary erasure quantization with low-density-generator-matrix
(LDGM) codes. The encoding is performed by iterative quantization
studied by Martinian and Yedidia~\cite{MarYed03}, which is a
message-passing algorithm similar to the decoding of
low-density-parity-check (LDPC) codes. As LDPC codes have been widely
adopted by commercial flash memory controllers, the hardware
architectures of message-passing algorithms have been well understood
and highly optimized in practice. Therefore, our codes are
implementation-friendly for practitioners. Extensive simulations show
that the rewriting performance of our scheme compares favorably with
that of the capacity-achieving polar WOM code~\cite{BurStr13} in the
rate region where a low rewriting failure rate is desired. For
instance, we show that our code allows user to write $40\%$ more
information by rewriting with very high success probability. We note
that the iterative quantization algorithm of~\cite{MarYed03} was used
in~\cite{ChaMarWor06} in a different way for the problem of
information embedding, which share some similarity with our model.
Moreover, our code construction is extended with error correction. The
need for error correction is observed in our experiments. As shown in
Fig.~\ref{fig:wom-error}, the BERs of both writes increase rapidly
with the number of block erasures.
Constructions of error-correcting WOM codes have been studied in
recent literature. Error-correcting WOM codes have been proposed in~\cite{EngLiKliLanJiaBru14a}\cite{GabSha12}\cite{JiaLiEngLanBru13}\cite{YaaSieVarWol12}\cite{ZemCoh91}. Different
from the existing constructions above, we use conjugate code pairs studied in the context of quantum error
correction~\cite{Ham06}. As an example, we construct LDGM WOM codes whose
codewords also belong to BCH codes. Therefore, our codes allows to use
any decoding algorithm of BCH codes. The latter have been implemented
in most commercial flash memory controllers. We also present two additional approaches to add error correction, and compare their performance.
\section{Rewriting and Erasure Quantization}
\label{sec:rewrite_quant}
\subsection{Rewriting Model}
We consider a model that allows two writes on a block of $n$ cells. A cell has a binary state chosen from $\{0,1\}$, with the rewriting constraint that state $1$ can be written to state $0$, but not vice versa. All cells are initially set to be in state 1, and so there is no writing constraint for the first write.
A vector is denoted by a bold symbol, such as ${\boldsymbol s}=(s_1,s_2,\dots,s_n)$.
The state of the $n$ cells after the first write is denoted by the vector ${\boldsymbol s}$.
We focus only on the second write, and we assume that after the first write, the state of the cells is i.i.d., where for each $i$, $\Pr\{s_i=1\} = \beta$. We note that the special case of $\beta=1/2$ is of practical importance, since it approximates the state after a normal page programming in flash memory\footnote{In flash memory, the message to be written can be assumed to be random due to data compression and data randomization used in memory controllers.}.
The second write is concerned with how to store a message $\bm{m} \in \mathbb{F}_2^k$ by changing ${\boldsymbol s}$ to a new state $\bm{x}$ such that 1) the rewriting constraint is satisfied, and 2) $\bm{x}$ represents $\bm{m}$.
This is achieved by the encoding operation of a rewriting code, defined formally in the following.
\begin{definition}
A \emph{rewriting code} $C_R$ is a collection of disjoint subsets of
$\mathbb{F}_2^n$.
\end{definition}
Each element of $C_R$ corresponds to a different message. Consider $M \in C_R$ that corresponds to a message $\bm{m}$, then for all $\bm{x} \in M$, we say that $\bm{x}$ is \emph{labeled} by $\bm{m}$. The
decoding function maps the set of labeled vectors into their labels,
which are also the messages. To encode a message ${\boldsymbol m}$ given a state ${\boldsymbol s}$, the encoder needs to find a vector ${\boldsymbol x}$ with
label ${\boldsymbol m}$ that can be written over ${\boldsymbol s}$. If the encoder does not
find such vector ${\boldsymbol x}$, it declares a failure. The rewriting rate of
$C_R$ is defined by $R_{\mathrm{WOM}}=k/n$. The rewriting capacity, which characterizes the maximum amount of information that
can be stored per cell in the second write, is known to be $\beta$ bits~\cite{Hee85}.
We are interested in rewriting codes with rates close to the capacity,
together with efficient encoding algorithms with low failure probability. The
main observation in the design of the proposed rewriting scheme of
this paper is that the rewriting problem is related to
the problem of binary erasure quantization (BEQ), introduced in the next subsection.
\subsection{Binary Erasure Quantization}
The BEQ problem is concerned with the quantization of a binary \emph{source sequence} ${\boldsymbol s}'$, for which some bits are erased. Formally, $\bm{s}' \in \{0,1,*\}^n$, where $*$ represents erasures. ${\boldsymbol s}'$ needs to be quantized (compressed) such that every non-erased symbol of ${\boldsymbol s}'$ will maintain its value in the reconstructed vector. A reconstructed vector with such property is said to have \emph{no distortion} from ${\boldsymbol s}'$. In this paper we use linear BEQ codes, defined as follows:
\begin{definition}
A linear BEQ code $C_Q$ is a subspace of $\mathbb{F}_2^n$. Each ${\boldsymbol c} \in C_Q$ is called a codeword of $C_Q$. The dimension of $C_Q$ is denoted by $r$.
\end{definition}
Each codeword of $C_Q$ is labeled by a different $r$-bits sequence ${\boldsymbol u}$.
Given a BEQ code $C_Q$ and a source sequence ${\boldsymbol s}'$, a quantization algorithm $Q$ is invoked to find a label ${\boldsymbol u}$ whose codeword ${\boldsymbol c}\in C_Q$ has \emph{no distortion} from ${\boldsymbol s}'$. If such label is found, it is denoted by ${\boldsymbol u}=Q({\boldsymbol s}')$, and is considered as the compressed vector. Otherwise, a quantization failure is declared, and $Q({\boldsymbol s}')=Failure$.
The reconstruction uses a generator matrix $G_Q$ of $C_Q$ to obtain the codeword ${\boldsymbol c}={\boldsymbol u} G_Q$.
\subsection{Reduction from Rewriting to Erasure Quantization}
\label{subset:red}
In this subsection we show that the problem of rewriting can be efficiently reduced to that of BEQ.
Let $C_Q$ be a linear quantization code, and let $H$ be a parity-check matrix of $C_Q$.
\begin{construction}
A rewriting code $C_R$ is constructed as the collection of all cosets of $C_Q$ in $\mathbb{F}_2^n$.
A decoding function for $C_R$ is defined by a parity check matrix $H$ of $C_Q$, such that a vector ${\boldsymbol x}\in\mathbb{F}_2^n$ is decoded into its syndrome
\begin{equation}
\label{eq:decode}
\text{DEC}_H({\boldsymbol x})={\boldsymbol x} H^T.
\end{equation}
\end{construction}
Since the dimension of $C_Q$ is $r$, it has $2^{n-r}$ cosets. Therefore the rate of $C_R$ is $R_{\text{WOM}}=\frac{n-r}{n}$, implying that $k=n-r$. We define some notation before introducing the reduction algorithm.
Let $(H^{-1})^T$ be a left inverse for $H^T$, meaning that $(H^{-1})^T H^T$ is the $k\times k$ identity matrix.
Define a function $BEC: \{0,1\}^n \times \{0,1\}^n \to \{0,1,*\}^n$ as:
\begin{align*}
BEC(\bm{w},\bm{v})_i = \left \{ \begin{array}{cc}
w_i & \text{if }v_i=0\\
* & \text{if }v_i=1
\end{array} \right. \ \ \ , \forall i=1,...,n
\end{align*}
$BEC(\bm{w},\bm{v})$ realizes a binary erasure channel that erases entries in $\bm{w}$ whose corresponding entries in $\bm{v}$ equal 1. We are now ready to introduce the encoding algorithm for the rewriting problem.
\begin{algorithm}
\caption{${\boldsymbol x} = ENC(G_{Q}, \bm{m},\bm{s}$): Encoding for Rewriting}
\label{alg:encoding}
\begin{algorithmic}[1]
\STATE ${\boldsymbol z} \leftarrow {\boldsymbol m} (H^{-1})^T $
\STATE ${\boldsymbol s}' \leftarrow BEC({\boldsymbol z},{\boldsymbol s})$
\STATE $\bm{u} \leftarrow Q({\boldsymbol s}')$
\IF{$\bm{u}$ = FAILURE}
\RETURN FAILURE
\ELSE
\RETURN ${\boldsymbol x} \leftarrow \bm{u}G_Q + {\boldsymbol z}$
\ENDIF
\end{algorithmic}
\end{algorithm}
\begin{theorem}\label{th:cordec}
Algorithm \ref{alg:encoding} either declares a failure or returns a vector $\bm{x}$ such that $\bm{x}$ is rewritable over $\bm{s}$ and $\bm{x}H^T = \bm{m}$.
\end{theorem}
\begin{IEEEproof}
Suppose failure is not declared and $\bm{x}$ is returned by Algorithm \ref{alg:encoding}. We first prove that $\bm{x}$ is rewritable over $\bm{s}$. Consider $i$ such that $s_i =0$. Then it follows from the definition of $BEC$ that $s'_i = z_i$. Remember that $Q({\boldsymbol s}')$ returns a label ${\boldsymbol u}$ such that ${\boldsymbol c}={\boldsymbol u} G_Q$ has no-distortion from ${\boldsymbol s}'$. Therefore, $c_i=s'_i=z_i$, and $x_i = c_i + z_i = z_i + z_i =0=s'_i$.
So ${\boldsymbol x}$ can be written over ${\boldsymbol s}$. To prove the second statement of the theorem,
notice that
\begin{align*}
\bm{x}H^T &= (\bm{u}G_Q + \bm{z})H^T = \bm{u}G_QH^T + \bm{m}(H^{-1})^TH^T \\
&= \bm{m}(H^{-1})^TH^T = \bm{m}.
\end{align*}
\vspace{-0.2cm}
\end{IEEEproof}
\section{Rewriting with Message Passing}
\label{sec:mp}
In this section we discuss how to choose a quantization code $C_Q$ and quantization algorithm $Q$ to obtain a rewriting scheme of good performance. Our approach is to use the iterative quantization scheme of Martinian and Yedidia~\cite{MarYed03}, where $C_Q$ is an LDGM code, and $Q$ is a message-passing algorithm.
This approach is particularly relevant for flash memories, since the hardware architecture of message-passing algorithms is well understood and highly optimized in flash controllers.
The algorithm $Q$ can be implemented by a sequential or parallel scheduling, as described in~\cite[Section 3.4.2]{MarYed03}. For concreteness, we consider the sequential algorithm denoted by \textsf{ERASURE-QUANTIZE} in~\cite{MarYed03}. Since the performance of \textsf{ERASURE-QUANTIZE} depends on the chosen generator matrix, we abuse notation and denote it by $Q(G_Q,{\boldsymbol s}')$.
Algorithm $Q(G_Q,{\boldsymbol s}')$ is presented in Appendix~\ref{app:algo}, for completeness.
Finally, we need to describe how to choose a generator matrix $G_Q$ that work well together with Algorithm $Q$.
We show next that a matrix $G_Q$ with good rewriting performance can be chosen to be a \emph{parity-check matrix} that performs well in message-passing decoding of erasure channels. This connection follows from the connection between rewriting and quantization, together with a connection between quantization and erasure decoding, shown in~\cite{MarYed03}. These connections imply that we can use the rich theory and understanding of the design of parity-check matrices in iterative erasure decoding, to construct good generating matrices for rewriting schemes. To make the statement precise, we consider the standard iterative erasure-decoding algorithm denoted by \textsf{ERASURE-DECODE}$(H,{\boldsymbol y})$ in~\cite{MarYed03}, where $H$ is an LDPC matrix and ${\boldsymbol y}$ is the output of a binary erasure channel.
\begin{theorem}
\label{th:ldpc}
For all ${\boldsymbol m}\in\mathbb{F}_2^k$ and ${\boldsymbol z}',{\boldsymbol s}\in\mathbb{F}_2^n$, $\text{ENC}(G_Q,{\boldsymbol m},{\boldsymbol s})$ fails if and only if $\textsf{ERASURE-DECODE}(G_Q,\text{BEC}(\bm{z}',{\boldsymbol s}+\bm{1}_n))$ fails, where $\bm{1}_n$ is the all-one vector of length $n$.
\end{theorem}
The proof of Theorem \ref{th:ldpc} is available in Appendix \ref{app:A}.
The running time of the encoding algorithm ENC is analyzed formally in the following theorem.
\begin{theorem}
\label{th:comp}
The algorithm $\text{ENC}(G_Q,{\boldsymbol m},{\boldsymbol s})$ runs in time $\mathcal{O}(nd)$ where $n$ is the length of ${\boldsymbol s}$ and $d$ is the maximum degree of the Tanner graph of $G_Q$.
\end{theorem}
The proof of Theorem \ref{th:comp} is available in Appendix \ref{app:B}.
Theorems~\ref{th:ldpc} and~\ref{th:comp}, together with the analysis and design of irregular LDPC codes that achieve the capacity of the binary erasure channel~\cite{OswSho02}, imply the following capacity-achieving results.
\begin{corollary}
\label{cor:capacity}
There exists a sequence of rewriting codes which can be efficiently encoded by Algorithm~\ref{alg:encoding} and efficiently decoded by Equation \emph{(}\ref{eq:decode}\emph{)} that achieves the capacity of the rewriting model $\beta$.
\end{corollary}
The proof of Corollary~\ref{cor:capacity} is available in Appendix~\ref{app:capacity}.
The finite-length performance of our rewriting scheme is evaluated
using extensive simulation with the choice of $\beta=0.5$ and $G_Q$ to
be the parity-check matrix of a Mackay code~\cite{Mac99}. The
rewriting failure rates of our codes with lengths $n = 8000$ and
$16000$ that are relevant to flash applications are compared with
those of the polar WOM codes of lengths $2^{13}$ and
$2^{14}$~\cite{BurStr13}. Fig.~\ref{fig:ldgm-vs-polar} shows the
rewriting failure rates of both codes at different rewriting rate,
where each point is calculated from $10^5$ experiments. Remember that
the capacity of the model is $0.5$. The results suggest that our
scheme achieves a decent rewriting rate (e.g. 0.39) with low failure
rate (e.g. $< 10^{-4}$). Moreover, our codes provide significantly
lower failure rates than polar WOM codes when the rewriting rate is
smaller, because of the good performance in the waterfall region of
message-passing algorithm.
\begin{figure}
\begin{center}
\includegraphics[width=0.6\linewidth, angle=270]{ldgm-vs-polar.eps}
\end{center}
\vspace{-0.2cm}
\caption{Rewriting failure rates of polar and LDGM WOM codes.}
\label{fig:ldgm-vs-polar}
\vspace{-0.6cm}
\end{figure}
\section{Error-Correcting Rewriting Codes}\label{sec:con}
The construction of error-correcting rewriting codes is based on a pair of linear codes $(C_1,C_Q)$, that satisfies the condition $C_1\supseteq C_Q$, meaning that each codeword of $C_Q$ is also a codeword of $C_1$. Define $C_2$ to be the dual of $C_Q$, denoted by $C_2=C_Q^{\perp}$. A pair of linear codes $(C_1,C_2)$, that satisfies $C_1\supseteq C_2^{\perp}$ is called \emph{a conjugate code pair}, and it is useful in quantum error correction and cryptography~\cite{Ham06}. For the flash memory application, we let $C_1$ be an error-correction code, while $C_2^{\perp}=C_Q$ is a BEQ code. The main idea in the construction of error-correcting rewriting codes is to label \emph{only} the codewords of $C_1$, according to their membership in the cosets of $C_Q$. The construction is defined formally as follows:
\begin{construction}
\label{con:conj}
For ${\boldsymbol c}\in C_1$, let ${\boldsymbol c}+C_Q$ be the coset of $C_Q$ in $C_1$ that contains ${\boldsymbol c}$.
Then the error-correcting rewriting code is constructed to be the collection of cosets of $C_Q$ in $C_1$.
\end{construction}
Next we define the matrices $(H^{-1})^T$ and $H^T$ to be used in encoding and decoding.
Let $G_1$ and $G_Q$ be generator matrices of the codes $C_1$ and $C_Q$, respectively, such that each row of $G_Q$ is also a row of $G_1$. Since $C_1$ contains $C_Q$, such matrix pair always exists. Define $(H^{-1})^T$ to be constructed by the rows of $G_1$ that are \emph{not} rows of $G_Q$. Let $H^T$ be a right inverse of $(H^{-1})^T$.
The encoding is performed according to Algorithm~\ref{alg:encoding}, with the matrix $(H^{-1})^T$ defined above. Note that in Step 1, ${\boldsymbol z}$ is a codeword of $C_1$, since each row of $(H^{-1})^T$ is also a row of $G_1$. In addition, in Step 7, ${\boldsymbol u} G_Q$ is also a codeword of $C_1$ (unless $Q(G_Q,{\boldsymbol s}')$ fails), since $C_Q$ is contained in $C_1$. Therefore, ${\boldsymbol x}={\boldsymbol u} G_Q+{\boldsymbol z}$ is a codeword of $C_1$. The decoding can begin by the recovery of ${\boldsymbol x}$ from its noisy version, using the decoder of $C_1$. The message ${\boldsymbol m}$ can then be recovered by the product ${\boldsymbol x} H^T$.
A similar framework was described in~\cite{JacCalSor12}, which proposed a construction of a repetition code contained in a Hamming code, with a Viterbi encoding algorithm. In this paper we make the connection to the quantum coding literature, which allows us to construct stronger codes.
\subsection{Conjugate Codes Construction}
\label{sub:con_construction}
We look for a conjugate pair $(C_1,C_2)$ such that $C_1$ is a good error-correcting code, while $C_2^{\perp}$ is a good LDGM quantization code. Theorem~\ref{th:ldpc} implies that $C_2$ needs to be an LDPC code with a good performance over a binary erasure channel (under message passing decoding).
Constructions of conjugate code pairs in which $C_2$ is an LDPC code are studied in~\cite{HagIma07}\cite{IofMez07}\cite{SarKlaRot09}. Sarvepalli \emph{et al.}~\cite{SarKlaRot09} construct a pair of codes such that $C_1$ is a BCH code and $C_2$ is a Euclidean geometry LDPC code, which is particularly useful for our purpose. This is because BCH codes are used extensively for error correction in flash memories.
Below we first briefly review the construction of Euclidean geometry LDPC codes and then discuss the application of the results in~\cite{SarKlaRot09} to our settings.
Denote by $\text{EG}(m,p^s)$ the Euclidean finite geometry over $\mathbb{F}_{p^s}$ consisting of $p^{ms}$ points. Note that this geometry is equivalent to the vector space $\mathbb{F}_{p^s}^m$. A $\mu$-dimensional subspace of $\mathbb{F}_{p^s}^m$ or its coset is called a \emph{$\mu$-flat}. Let $J$ be the number of $\mu$-flats that do not contain the origin, and let $\alpha_1, ... \alpha_{p^{sm} - 1}$ be the points of $\text{EG}(m,p^s)$ excluding the origin. Construct a $J \times p^{sm}-1$ matrix $H_{EG}$ in the way that its $(i,j)$-th entry equals 1 if the $i$-th $\mu$-flat contains $\alpha_j$, and equals 0 otherwise. $H_{EG}$ is the parity check matrix of the (Type-I) Euclidean geometry LDPC code $C_{EG}(m,\mu,s,p)$. $C_{EG}(m,\mu,s,p)$ is a cyclic code and by analyzing the roots of its generator polynomial, the following result is obtained~\cite{SarKlaRot09}.
\begin{prop}\label{propconj}
$C^\perp_{EG}(m,\mu,s,p)$ is contained in a BCH code of design distance $\delta=p^{\mu s} -1$.
\end{prop}
Hence we may choose $C_2$ to be $C_{EG}(m,\mu,s,p)$ and $C_1$ to be a BCH code with distance equal to or smaller than $\delta$. Some possible code constructions are shown in Table~\ref{womerrcode}. Their encoding performance, with respect to the probability $\beta$ that a cell in the state is writable, is shown in Fig.~\ref{ach}. Note from Fig.~\ref{ach} that a code with smaller rewriting rate achieves a fixed failure rate at a smaller value of $\beta$. In particular, the codes corresponding to the top three rows of Table~\ref{womerrcode} achieve very small failure rate at $\beta=0.5$, the point of practical interest. These results also show that the slope of the figures becomes sharper when the length of the codes increases, as expected. Out of the three codes that can be rewritten with $\beta=0.5$, $C_{EG}(3,1,3,2)$ poses the best rate and error-correction capability.
\begin{table}
\caption{Error-correcting Rewriting Codes Constructed from pairs of conjugate BCH and EG-LDPC Codes.}
\vspace{-0.8cm}
\label{womerrcode}
\center
\begin{tabular}{c|c|c|c}
$(m,\mu,s,p)$&$C_1[n,k,\delta]$&$C_2[n,k]$&Rewriting Rate\\
\hline
(4,1,2,2)&[255,247,3]&[255,21]&0.0510\\
(3,1,2,2)&[65,57,3]&[65,13]&0.1111\\
(3,1,3,2)&[511,484,7]&[511,139]&0.2192\\
(3,1,4,2)&[4095,4011,15]&[4095,1377]&0.3158
\end{tabular}
\vspace{-0.5cm}
\end{table}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.3\textwidth, angle=270]{EGPlot.eps}
\caption{Encoding performance of the codes in Table \ref{womerrcode}.}\label{ach}
\end{center}
\vspace{-0.4cm}
\end{figure}
\section{Alternative Approaches for Error Correction}
In this section we present two alternative approaches to combine rewriting codes with error correction.
\subsection{Concatenated Codes}
In this scheme, we concatenate an LDGM rewriting code with a
systematic error-correcting code. The outer code is an LDGM rewriting
code without error-correction capability, as in Section
~\ref{sec:mp}. The systematic ECC is used as the inner code. The
concatenated scheme is used in the second write. The scheme requires
the first write to \emph{reserve} some bits to store the redundancy of
the ECC in the second write.
In the second write, the encoder begins by finding a vector ${\boldsymbol x}$
that can be written over the current state. After ${\boldsymbol x}$ is written,
the systematic ECC calculates the redundancy bits required to protect
${\boldsymbol x}$ from errors. The redundancy bits are then written into the
reserved cells. The decoding of the second write begins by recovering
${\boldsymbol x}$ using the systematic ECC and its redundancy bits. After ${\boldsymbol x}$
is recovered, the decoder of the rewriting code recovers the stored
message from ${\boldsymbol x}$.
We note that reserving bits for the second write have a negative effect on the performance of the system, since it reduces the total amount of information that could be stored in the memory on a given time. Therefore, the next subsection extends the concatenation scheme using a chaining technique, with the aim of reducing the number of bits required to be reserved for the second write.
\subsection{Code Chaining}
The chaining approach is inspired by a similar construction in polar coding~\cite{MonHasUrbSas14}.
The idea is to chain several code blocks of short length. In the following we use a specific example to demonstrate the idea.
We use a BCH code for error correction, since its performance can be easily calculated. We note, however, that LDPC codes may be used in practice, such that the circuit modules may be shared with the rewriting code, to reduce the required area. The performance of LDPC code in the considered parameters is similar to that of BCH codes.
A typical BCH code used in flash memory has the parameters $[8191,7671,81]$, where the length is $8191$, the dimension is $7671$, and the minimum distance is $81$. If this code is used in a concatenated scheme for the second write, the first write needs to reserve $8191-7671=520$ bits for redundancy.
To reduce the amount of required reserved bits, we consider the chaining of $8$ systematic BCH codes with the parameters $[1023,863,33]$. The encoding is performed sequentially, beginning with the rewriting encoding that finds a vector ${\boldsymbol x}_1$ of $863$ bits. The vector ${\boldsymbol x}_1$ represents a message ${\boldsymbol m}_1$ of $310$ bits, according to an $[863,310]$-LDGM rewriting code. Once ${\boldsymbol x}_1$ is found, the BCH encoder finds $1023-863=160$ redundancy bits to protect ${\boldsymbol x}_1$, as in the concatenated scheme.
The encoder then ``chains'' the redundancy bits forward, by encoding them, together with $150$ new information bits, into another block of $863$ bits, using the $[863,310]$-LDGM code. Let ${\boldsymbol m}_2$ denote the vector of $310$ bits encoded into the second block. ${\boldsymbol m}_2$ contains the $160$ redundancy bits of ${\boldsymbol x}_1$, together with the additional $150$ information bits. Note that once ${\boldsymbol m}_2$ is decoded, the redundancy bit of ${\boldsymbol x}_1$ are available, allowing the recovery ${\boldsymbol x}_1$, and then ${\boldsymbol m}_1$. The encoding continues in this fashion $8$ times, to write over a total of $8$ blocks, each containing $863$ cells. The $160$ redundant bits used to protect $\bm{x}_8$ are stored in the reserved cells. The decoding is done in the reverse order, where each decoded vector contains the redundancy bits of the previous block.
\subsection{Comparison}
We compare the different error-correction approaches, and discuss their trade-offs. The first code in the comparison is a conjugate code pair, described in Section~\ref{sec:con}. We use a conjugation of a $[511,484,7]$-BCH code containing a $[511,372]$-LDGM code, dual to the $(3,1,3,2)$-Euclidean geometry LDPC code in Table~\ref{womerrcode}. The second code in the comparison is a concatenation of an outer $[7671,2915]$-LDGM Mackay rewriting code with an inner $[8191,7671,81]$-BCH code. The third code is a chaining of $8$ blocks of $[863,310]$-LDGM Mackay codes, each concatenated with a $[1023,863,33]$-BCH code. We compare the decoding BER $P_D$, the fraction $\alpha$ of bits required to be reserved, and the rewriting rate $R_{\text{WOM}}$ of the codes. The encoding failure rate of each of the three codes for $\beta=0.5$ is below $10^{-3}$. $P_{D}$ is estimated with a standard flash memory assumption of a raw BER of $1.3\times 10^{-3}$. To achieve a comparable code length, the conjugated code is assumed to be used 16 times in parallel, with a total length of $511\times 16=8176$.
The comparison is summarized in Table~\ref{tab:comparison}.
Flash systems require $P_{D}$ below $10^{-15}$. We see in Table~\ref{tab:comparison} that conjugated code still do not satisfy the reliability requirement. We also see that concatenated codes that satisfy the reliability requirement need a large fraction of reserved space. The chained code reduces the fraction of reserved space to $2\%$, with a rate penalty in the second write.
\begin{table}
\caption{Error-correcting rewriting codes of length~$\approx 8200$.}
\vspace{-0.3cm}
\label{tab:comparison}
\center
\begin{tabular}{l | l |c|c}
Code&$P_D$&$\alpha$&$R_{\text{WOM}}$\\
\hline
Conjugated&$10^{-5}$&$0\%$&$0.21$\\
Concatenated&$10^{-16}$&$6.3\%$&$0.35$\\
Chained&$10^{-16}$&$2\%$&$0.19$
\end{tabular}
\vspace{-0.4cm}
\end{table}
\IEEEpeerreviewmaketitle
\allowdisplaybreaks
\bibliographystyle{IEEEtranS}
|
\section{Introduction}
Over the past few years large cosmological simulations have been performed to determine the properties of dark matter halos, including density profiles, shapes and accretion histories (see e.g. \citealt{Springel05,Klypin2011,Bryan13}). These properties are of particular interest, as forming galaxies depend on the structural properties of the halos in which they are embedded.
During hierarchical growth, halos acquire a density profile with a near universal shape, that can be described by a simple formula known as the `NFW profile' (\citealt{NFW97}, hereafter NFW). The NFW density profile is described by just two parameters, halo mass, $M$, and concentration, $c$. A halo's concentration is defined as the ratio of the virial radius, $R_{\rm{vir}}$, and the scale radius, $r_{-2}$, which is defined as the radius where the logarithmic density slope is $-2$. Thus, given the NFW profile, only a relation between concentration and halo mass (hereafter, the $c-M$ relation) is needed to fully specify halo structure at fixed mass. Therefore, numerous studies have been undertaken to improve the $c-M$ calibration.
Despite its importance, there is still no solid agreement on the dependence of halo concentration on halo mass and redshift. A small change in the adopted cosmology can have important effects on the structure of dark matter halos (\citealt{Maccio08}), and on their mass accretion histories (\citealt{Zhao09}). For example, the mean concentrations of dwarf-scale dark matter halos change by a factor of 1.5 between the various Wilkinson Microwave Anisotropy Probe (WMAP) cosmologies (\citealt{Spergel03,Spergel07}). The Planck cosmology (\citealt{Planck}) has higher matter density, $\Omega_{\rm{m}}$, and higher power spectrum normalization, $\sigma_{8}$, compared to the cosmological parameters of the year 5 data release of WMAP (WMAP5;~\citealt{Komatsu}). The Planck cosmology therefore suggests that halos assemble earlier and are more concentrated (c.f. $c-M$ relations from \citealt{Dutton14} and \citealt{Duffy08}).
However, cosmology may not be the primary reason for the differences in the concentration$-$mass relations found by various authors. Recent works that adopt the same cosmology still find different $c-M$ relations (compare for example \citealt{Dutton14} and \citealt{Diemer14b}, or \citealt{Klypin2011} and \citealt{Prada}). \citet{Dutton14} found that the $c-M$ relation is well described by a power law, but flattens at high redshift and exhibits a positive slope at $z>4$. In contrast, \citet{Diemer14b} found a strong upturn in the high-mass end of the $c-M$ relation at all redshifts. The disparity between these studies could be due to the dynamical state of the selected dark matter halos. For example, \citet{Ludlow12} showed that massive halos that are substantially out of equilibrium are more likely to be found at a transient stage of high concentration, thus explaining the puzzling upturn in the high-mass end of the $c-M$ relation. Indeed, they reported that the upturn disappears when only dynamically-relaxed systems are considered. However, \citet{Klypin14} argued that the virial criterion used by \citet{Ludlow12} to select relaxed systems is incorrect, as it needs to include effects of the surface pressure and external forces. \citet{Klypin14} modified the virial criterion and ended up selecting massive halos that had previously been considered as unrelaxed. As a result, \citet{Klypin14} obtained an upturn in the $c-M$ relation of their relaxed halo sample and concluded that the upturn is a real feature of the $c-M$ relation. They explained that as extremely massive halos have more radial infall velocities, infalling mass penetrates deeper within the inner halo, thus increasing the concentration and producing the upturn.
The main goal of this work is to derive a physically motivated model for the $c-M$ relation of relaxed halos based on the dark matter halo accretion history. By relating the concentration to the halo accretion history, we find that the $c-M$ relation does not show any upturn or strong flattening. We then study the $c-M$ relation in detail using simulations and selecting relaxed halos without using the virial criterion, and investigate whether recently accreted particles are able to reach the inner parts of the halo and thus increase the concentration.
Our $c-M$ model relies on the fact that concentrations depend on the evolutionary stage of halos when they were formed. Several works have suggested that halo formation can be described as an `inside out' process, where a bound core (of a certain fraction of the halo mass today) collapses, followed by the gradual addition of material at the cosmological accretion rate (\citealt{Manrique,Wang09,Dalal,Salvador}). In this framework, the halo concentration should depend on the epoch at which a certain fraction of the halo mass was assembled. As a result, various authors (\citealt{Bullock,Wechsler,Zhao03,Ludlow14}) have provided models that relate $c$ to the halo mass history. For instance, \citet{Zhao03} showed that when the mass accretion rate of a halo slows down at low redshift, its scale radius, $r_{-2}$, remains approximately constant, and hence that concentration scales with the virial radius. On the other hand, in the regime of a high mass accretion rate (at high redshift), the scale radius scales approximately as the virial radius and thus $c$ remains constant.
The connection between a halo's mass accretion history (hereafter MAH) and its concentration, $c$, is therefore obtained through its `formation' time. The halo formation (or assembly) time is traditionally defined as the point in time when the halo mass reached a fraction of the total mass today. Low-mass halos typically assemble earlier, when the Universe was denser, than high-mass halos do. As a result, low-mass halos are more concentrated. Clearly, if concentration correlates with formation time, and formation time depends on the mass variance, $\sigma$ (because $\sigma$ describes the halo MAH, see the analytic model for the MAH from \citealt{PaperI}), then it is expected that $c$ correlates with $\sigma$ and hence with the peak height, $\nu$, defined as $\nu=1.686/\sigma$. This is indeed what several works have found (e.g. \citealt{Zhao09,Prada,Ludlow14,Dutton14}). We showed in \citet{PaperII} that the physical origin of the $c-\sigma$ (or $c-\nu$) relation is the halo MAH.
Recently, in \citet{PaperI} and \citet{PaperII} (hereafter Paper I and Paper II, respectively), we provided two models for the MAH of halos, an analytic model and a semi-analytic model. The semi-analytic model uses a functional form for the MAH, that is motivated by extended Press-Schechter (EPS) theory, and links the MAH to halo structure through two empirical relations obtained from simulations. The analytic model is fully derived from the EPS formalism and thus does not require calibration against any simulation data.
In the analytic model, the halo MAH is described in terms of the {\it{rms}} of the density perturbation field, $\sigma$, as ${M(z)=M_{0}(1+z)^{af(M_{0})}e^{-f(M_{0})z}}$, where $M_{0}$ refers to the present halo mass, $a$ depends on cosmology, and ${f(M_{0})\sim1/\sigma(M_{0})}$. This expression illustrates that as $\sigma$ decreases with halo mass, the function $f(M_{0})$ increases, causing the exponential in $M(z)$ to dominate. As a result, high-mass halos accrete faster than low-mass ones, due to their low value of $\sigma$. As low $\sigma$ implies large peak height, the EPS formalism predicts that density perturbations with large $\nu$ experience an accelerated collapse phase relative to the average, and grow faster in time.
In this work we present a semi-analytic, physically motivated model for dark matter halo concentration as a function of halo mass, redshift and cosmology. The semi-analytic model, which builds on that of \citet{Ludlow14}, uses the analytic model for the halo MAH provided in Paper I, as well as an extension of the empirical relation between concentration and formation time obtained through fits to simulations provided in Paper II. As a result, the semi-analytic model for halo concentrations shows how the $c-M$ relation is expected to evolve based on the hierarchical growth of halos.
This paper is organized as follows. We begin in Section~\ref{Simulations} with a description of the set of cosmological simulations used in this work. In Section~\ref{Halo_history_modelling}, we describe the analytic MAH model provided in Paper I and extend it to high redshift. In Section~\ref{cM_model}, we define halo formation time and build an empirical relation between formation time and concentration through fits to simulation data. Next, we describe the semi-analytic model for halo concentrations that combines the analytic model for the MAH and the empirical relation described previously. We analyse the evolution of concentration that predicts the semi-analytic model in Section~\ref{c_evolution}. In Section~\ref{DM_annihilation_signal} we discuss the impact of the results of our semi-analytic model for halo concentration on the signal from dark matter annihilation. In Section~\ref{Discussion} we discuss the main assumptions the semi-analytic model relies on. Finally, we summarize and conclude in Section~\ref{Conclusion}.
\section{Simulations}\label{Simulations}
Throughout this work we compare our analytic results to the output from numerical simulations. We use a set of cosmological dark matter only (DMONLY) simulations from the OWLS project (\citealt{Schaye}). These simulations were run with a significantly extended version of the N-Body Tree-PM, SPH code gadget3 (last described in \citealt{Springel05}). The initial conditions were generated with CMBFAST (version 4.1; \citealt{Seljak}) and evolved to redshift $z = 127$, where the simulations were started, using the \citet{ZelDovich} approximation from an initial glass-like state (\citealt{White96}). In order to assess the numerical convergence, we use simulations of different box sizes (ranging from 25 to 400 $\, h^{-1}{\rm Mpc}$) and particle numbers (ranging from $128^{3}$ to $512^{3}$). The simulation names contain strings of the form LxxxNyyy, where xxx is the simulation box size in comoving $\, h^{-1}{\rm Mpc}$ and yyy is the cube root of the number of particles. Our DMONLY simulations assume the WMAP5 cosmology. However, to investigate the dependence on the adopted cosmology, we use an extra set of five dark matter only simulations ($100 \, h^{-1}{\rm Mpc}$ box size and $512^{3}$ dark matter particles) which assume values for the cosmological parameters derived from different releases of the WMAP and the Planck missions. See the tables in Appendix \ref{Simulations_appendix} for the sets of cosmological parameters adopted in the different simulations, as well as the main numerical parameters of the runs such as comoving box size, number of dark matter particles, dark matter particle mass, comoving gravitational softening and maximum physical softening.
\section{Halo mass accretion history}\label{Halo_history_modelling}
We begin this section by briefly reviewing the analytic model for the MAH derived from the EPS formalism in Paper I, and showing how the MAH depends on cosmology and on the initial peak of the primordial density field. In Section \ref{MAHmodel_highz} we extend it to estimate the halo MAH tracked from an arbitrary redshift. Readers only interested in the $c-M$ relation model can skip directly to Section~\ref{cM_model}.
\subsection{Analytic model for the halo mass history}\label{MAH_model}
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=0.48\textwidth]{Fig1.ps}\\
\caption{Halo MAH of a $10^{12}\,{\rm M_{\odot}}$ halo (coloured lines) obtained from the model given by eqs. (\ref{MzLCDM_highz})-(\ref{zf_LCDM}), and by assuming various cosmologies as indicated in the legend. The grey lines correspond to MAH obtained from DMONLY simulations that assume the Planck and WMAP5 cosmologies. In the top right corner, we plot $\sigma$ versus halo mass, to show that the change in $\sigma$ under different cosmologies drives the change in the MAH.}
\label{Mz12_LCDM}
\end{center}
\end{figure}
In Paper I, we used simple analytic arguments based on the EPS formalism and the analytic formulation of \citet{Neistein}, to show that the `shape' of the MAH is determined by the growth factor of the initial density pertubations. The halo MAH is well described by an exponential in the high-redshift regime, but it slows to a power law at low redshift, because the growth of density perturbations is halted in the dark energy dominated era due to the accelerated expansion of the Universe. Therefore, we showed that the expression
\begin{equation}\label{MzLCDM}
M(z)_{\Lambda\rm{CDM}}=M_{0}(1+z)^{\alpha}e^{\beta z},
\end{equation}
\noindent accurately captures the median halo MAH, where $M_{0}$ refers to halo mass today, and $\alpha$ and $\beta$ are parameters that depend on $M_{0}$, cosmology and the linear power spectrum. In the case of an Einstein de Sitter (EdS) cosmology ($\Omega_{\Lambda}=0$ and $\Omega_{\rm{m}}=1$) or an open universe ($\Omega_{\Lambda}=0$ and $\Omega_{\rm{m}}<1$), there is no acceleration in the expansion of the Universe at low redshift. Then the halo mass history is simply described by an exponential as ${M(z)_{\rm{EdS}}=M_{0}e^{\beta z}}$, where $\beta=-1.686(2/\pi)^{1/2}f(M_{0})$. For a complete description of the model, see Paper I.
We find that the MAH model can be used to calculate halo mass histories in cosmologies other than WMAP5, and that the differences are mainly driven by the changes in $\sigma_{8}$ and $\Omega_{\rm{m}}$. We show this in Fig. \ref{Mz12_LCDM}, where the halo MAH of a $10^{12}\,{\rm M_{\odot}}$ halo (coloured lines) was estimated for the various cosmologies, as indicated in the legend. In the top right corner of Fig. \ref{Mz12_LCDM}, we plot $\sigma$ versus halo mass, to show how the change in $\sigma$ drives the change in the MAH. The exception is the Planck cosmology, which has a relatively low $\sigma_{8}$ but a large $\Omega_{\rm{m}}=0.317$, which raises $M(z)$ close to the WMAP1 $M(z)$.
The overplotted grey lines in Fig.~\ref{Mz12_LCDM} correspond to the MAH obtained from DMONLY simulations that assume the Planck and WMAP5 cosmologies. In this case, we compute the MAH of the main subhalo (that is not embedded inside a larger halo) of Friends-of-Friends (FoF) groups (\citealt{Davis}), by tracking the virial mass of the main progenitor at each prior output redshift. Halo virial masses and radii were determined using a spherical overdensity routine within the SUBFIND algorithm (\citealt{Springel}) centred on the main subhalo of FoF halos. Throughout this work we define the halo mass as the total mass within the radius $r_{200}$ for which the mean internal density is 200 times the critical density. For a more detailed description of the method used to create merger trees, see Paper II.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=0.8\textwidth]{Fig2.ps}\\
\caption{Median MAHs for halos of $10^{11}\,{\rm M_{\odot}}$ starting from various redshifts. In both panels the grey solid lines correspond to the analytic model described in Section~\ref{MAHmodel_highz}. The coloured curves in the left panel are the MAHs obtained from the DMONLY simulations WMAP5$_{-}$L025N512 and WMAP5$_{-}$L050N512. The mass histories are computed by calculating the median value and the $1\sigma$ error bars are determined by bootstrap resampling the halos from the merger tree at a given output redshift. The coloured dot dashed curves in the right panel are the MAHs obtained from the \citet{vandenBosch14} model.}
\label{MAH_z0}
\end{center}
\end{figure*}
\subsection{Analytic model for the MAH: high redshift prediction}\label{MAHmodel_highz}
The model presented in Paper I is suitable for estimating halo MAHs that are tracked from ${z=0}$. In this section we extend this analytic model to estimate MAHs of halos of the same halo mass that are tracked from arbitrary redshifts $z_{i}$. This is shown in Fig. \ref{MAH_z0}, where the MAHs of $10^{11}\,{\rm M_{\odot}}$ halos are obtained from DMONLY simulations (coloured curves). The curves show the mean MAH of halos of the same mass ($10^{11}\,{\rm M_{\odot}}$ in this case) that begin at $z_{i}=0$ (blue curve), $1$ (dark green curve), $2$ (green curve), $3$ (orange curve) and $4$ (red curve). High-redshift MAHs are dominated by large accretion rates and characterized by a pure exponential.
We generalize the analytic model so that it describes the MAHs from any $z_{i}$ redshift. Expression (\ref{MzLCDM}) can be rewritten as
\begin{equation}\label{MzLCDM_highz}
\tilde{M}(z,M(z_{i}),z_{i})= M(z_{i})(1+z-z_{i})^{\tilde{\alpha}}e^{\tilde{\beta}(z-z_{i})},
\end{equation}
\noindent where $\tilde{M}(z,z_{i})$ denotes the MAH of a halo with mass $M(z_{i})$ at redshift $z_{i}$. In the above expression, $z>z_{i}$ and the parameters $\tilde{\alpha}$ and $\tilde{\beta}$ depend on ${M(z_{i})}$ and redshift $z_{i}$
\begin{eqnarray}\label{alpha_tilde}
\tilde{\alpha} &=& \left[\frac{1.686(2/\pi)^{1/2}}{D(z_{i})^{2}}\frac{dD}{dz}|_{z=z_{i}}+1\right]f(M(z_{i})),\\\label{beta_tilde}
\tilde{\beta} &=& -f(M(z_{i})),\\\label{fM}
f(M(z_{i})) &=& [\sigma^{2}(M(z_{i})/q)-\sigma^{2}(M(z_{i}))]^{-1/2},\\\label{sigma_def}
\sigma^{2}(R) &=& \frac{1}{2\pi^{2}}\int_{0}^{\infty}P(k)\hat{W}^{2}(k;R)k^{2}dk,\\\label{q_LCDM}
q&=&4.137\times z_{\rm{f}}^{-0.9476},\\\label{zf_LCDM}
z_{\rm{f}}&=&-0.0064(\log_{10}M_{0})^{2}+0.0237(\log_{10}M_{0})\\\nonumber
& & +1.8837,
\end{eqnarray}
\noindent where $D(z)$ is the linear growth factor, $P(k)$ the linear power spectrum, $\hat{W}(k;R)$ the Fourier transform of a top hat window function and $R$ defines $\sigma$ in a sphere of mass $M=(4\pi/3)\rho_{\rm{m},0}R^{3}$, where $\rho_{\rm{m},0}$ is the mean background density today. We use the approximation of \citet{Eisenstein} to compute $P(k)$, normalized such that $\sigma(8h^{-1}\rm{Mpc})=\sigma_{8}$. As a result, $f(M_{0})$ depends on the power spectrum and halo mass. It can be seen from eqs.~(\ref{alpha_tilde}) and (\ref{beta_tilde}) that at large $z_{i}$, $\tilde{\alpha}\rightarrow 0$ due to $D(z_{i})\propto 1.686(2/\pi)^{1/2}/(1+z_{i})$ for $z_{i}\gg 1$, indicating that the MAH is mainly described by an exponential. Table \ref{Notation} provides a summary of the nomenclature adopted throughout this work.
The above equations introduce an analytic halo MAH model directly derived from EPS theory that does not require calibration against any simulation data (see Paper I for more details). The numerical values given in eqs. (\ref{q_LCDM}) and (\ref{zf_LCDM}) were determined by assuming the WMAP5 cosmology (${\Omega_{\rm{m}}=0.258}, {\Omega_{\Lambda}=0.742}, {h=0.72}, {n_{s}=0.963}, {\sigma_{8}=0.796}$).
In the left panel of Fig.~\ref{MAH_z0}, we compare the model given by eqs.~(\ref{alpha_tilde})-(\ref{zf_LCDM}) to various MAHs obtained from a set of DMONLY simulations. Our analytic model is shown by grey solid lines, where we have taken ${M(z_{i})=10^{11}\,{\rm M_{\odot}}}$. The coloured curves in the left panel correspond to the MAHs obtained from the DMONLY simulations WMAP5$_{-}$L025N512 and WMAP5$_{-}$L050N512. We find very good agreement between the simulation outputs and the analytic model at all redshifts. The simulation outputs from the boxes $L=25\, h^{-1}{\rm Mpc}$ and $L=50\, h^{-1}{\rm Mpc}$ converge up to $z=5$. At higher $z$, the outputs from the $L=25\, h^{-1}{\rm Mpc}$ simulation underestimate $M(z)$ because the box size limits the maximum sizes of the structures that can form at each redshift.
In the right panel of Fig.~\ref{MAH_z0}, we compare our extended analytic model with the \citet{vandenBosch14} model. \citet{vandenBosch14} extracted halo mass histories from the Bolshoi simulations (\citealt{Klypin2011}) and extended them below the numerical resolution limit using EPS merger trees. Once they had obtained the MAH curves for a large range of redshifts and halo masses, they made use of a semi-analytic model to transform the (average or median) MAHs, based on the Bolshoi cosmology, to other cosmologies. Using their publicly available code, we calculate the mass history curves for the WMAP5 cosmology for comparison with our results. We find that there is some discrepancy at high-redshift for all the curves. The \citet{vandenBosch14} MAH model seems to over predict the halo mass at $z>5$, most likely as a consequence of the different halo definitions, and subtle differences in the definition of the main progenitor (van den Bosch, private communication). Overall, there is very good agreement between the most recent accretion history study in the literature and our model, as well as with the simulation outputs. In Section~\ref{cM_model} we will make use of our analytic MAH model to calculate concentrations.
Using the extended MAH model for high redshift, we can calculate the accretion rate of a halo at redshift $z$. We differentiate eq.~(\ref{MzLCDM_highz}) with respect to time and replace $dz/dt$ by $-H_{0}[\Omega_{\rm{m}}(1+z)^{5}+\Omega_{\Lambda}(1+z)^{2}]^{1/2}$, to obtain
\begin{eqnarray}\nonumber
\frac{d\tilde{M}(z,M(z_{i}),z_{i})}{dt} &=& 71.6{\rm{M}}_{\sun}{\rm{yr}}^{-1} \left(\frac{\tilde{M}(z,M(z_{i}),z_{i})}{10^{12}\rm{M}_{\sun}}\right) \\\nonumber
&\times & \left(\frac{h}{0.7}\right)[-\tilde{\alpha}/(1+z-z{i})-\tilde{\beta}]\\\label{dMdt}
&\times & (1+z)[\Omega_{\rm{m}}(1+z)^{3}+\Omega_{\Lambda}]^{1/2},
\end{eqnarray}
\noindent where $\tilde{\alpha}$ and $\tilde{\beta}$ are given by eqs.~(\ref{alpha_tilde}) and (\ref{beta_tilde}), respectively. Note that the above formula will give the accretion rate at redshift $z$ of a halo that has mass $M(z_{i})$ at redshift $z_{i}$, and mass $\tilde{M}(z,M(z_{i}),z_{i})$ at redshift $z$.
\section{Concentration$-$ mass relation}\label{cM_model}
A theoretical understanding of the physical connection between concentration (the parameter that characterizes the internal structure of NFW dark matter halos) and the initial conditions of the density field, is essential for the physical interpretation of relations like $c-\nu$ (concentration$-$peak height) or $c-M$, that have been calibrated using cosmological simulations (e.g. \citealt{Bullock,Neto,Gao,Maccio07,Duffy08,Ludlow13,Dutton14,Diemer14b}).
It has previously been shown that concentration is determined by the halo MAH, and that the MAH depends on the power spectrum and the adopted cosmological parameters (\citealt{Wechsler,Zhao03,Ludlow13,Ludlow14}). In this section we show, through analytic and numerical modelling, how the concentration of dark matter halos depends on cosmology and the power spectrum of density perturbations. Our results imply that the halo MAH is the physical link between concentration and peak height.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=0.75\textwidth]{Fig3.ps}
\end{center}
\caption{{\it{Left panel:}} mean density within the NFW scaled radius, $\langle\rho\rangle(<r_{-2})$, at $z_{i}$ against the critical density of the universe at the formation time, $\rho_{\rm{crit}}(z_{-2})$. Each dot in the panel corresponds to an individual relaxed halo identified at $z_{i}$ and coloured by mass according to the colour bar at the top of the plot. The open symbols show the median values of the sample at $z_{i}$ as indicated by the legend in logarithmic mass bins of width $\delta\log_{10}M=0.4$. The solid line shows the best linear fit to the $\rho_{\rm{crit}}(z_{-2})-\langle\rho\rangle(<r_{-2})$ relation. {\it{Right panel:}} formation redshift against concentration. The solid lines show the $c-z_{-2}$ relations given by equation (\ref{zf_relation}) for various $z_{i}$ as indicated in the legend of the left panel. The open symbols correspond to the median values of the samples in logarithmic mass bins of width $\delta\log_{10}M=0.4$ and are colour coded by $z_{i}$. The grey areas show the scatter in $z_{-2}$.}
\label{formation_time}
\end{figure*}
\subsection{Formation redshift}\label{zf_definition}
As discussed in the Introduction, halo MAHs can be used to estimate halo concentrations. Halo concentrations reflect the mean density of the Universe at the formation redshift (NFW;\citealt{Wechsler,Zhao03,Zhao09,Ludlow13}). Therefore, the essential link between a halo's MAH and its internal structure is the formation redshift. For a halo with mass $M(z_{i})$ at redshift $z_{i}$, we define the formation redshift to be $z_{-2}$, the redshift at which the mass of the main progenitor equals the mass enclosed within the scale radius at $z=z_{i}$,
\begin{equation}\label{z2}
z_{-2}=z[\tilde{M}(z_{-2},M(z_{i}),z_{i})=M_{r}(r_{-2},z_{i})],
\end{equation}
\noindent (\citealt{Ludlow13}). Here $\tilde{M}(z_{-2},M(z_{i}),z_{i})$ is the mass at $z_{-2}$ of a halo with mass $M(z_{i})$ at $z_{i}$, and we denote the mass enclosed within $r$, $M(<r)$, as $M_{r}$. For an NFW profile the internal mass $M_{r}(r_{-2},z_{i})$ is related to the total halo mass as
\begin{equation}\label{M2}
M_{r}(r_{-2},z_{i})=M(z_{i})\frac{Y(1)}{Y(c[M(z_{i}),z_{i}])},
\end{equation}
\noindent where $Y(u)=\ln(1+u)-u/(1+u)$, $c[M(z_{i}),z_{i}]$ is the concentration at $z_{i}$ and $M(z_{i})$ is the total halo mass at $z_{i}$. In cases where we identify halos at $z_{i}=0$ and track their mass histories, we calculate $z_{-2}$ by setting $M(z_{-2})$ equal to the mass enclosed within $r_{-2}$ today. In cases where we identify halos at $z_{i}>0$, we first calculate $r_{-2}$ and $M_{r}(r_{-2},z_{i})$ at the particular redshift $z_{i}$, and then find $z_{-2}$ by tracking the MAH (for $z>z_{i}$) and equating $\tilde{M}(z_{-2},M(z_{i}),z_{i})$ to $M_{r}(r_{-2},z_{i})$. See Table~\ref{Notation} for a summary of the nomenclature.
\citet{Ludlow13,Ludlow14} and Paper II showed that $z_{-2}$ correlates strongly with $c$, and in Paper II we demonstrated that the scatter in $z_{-2}$ and in the halo MAH predicts the scatter in $c$. In this section we explore how the formation time $-$ concentration relation varies for halos identified at various redshifts.
We computed density profiles and MAHs for halos identified at redshifts $z_{i} = 0, 1, 2, 3$ and $4$. The density profiles were computed by fitting the NFW density profile,
\begin{equation}\label{NFW}
\rho(r,z_{i})=\frac{\rho_{\rm{crit}}(z_{i})\delta_{\rm{c}}}{(cr/r_{200})[1+(cr/r_{200})]^{2}},
\end{equation}
\noindent for each individual halo. In the above equation, ${\rho_{\rm{crit}}(z_{i})=3H^{2}(z_{i})/8\pi G}$ is the critical density of the universe, $\delta_{\rm{c}}$ is a dimensionless parameter related to the concentration $c=r_{200}/r_{-2}$ by $\delta_{\rm{c}}=\frac{200}{3}\frac{c^{3}}{Y(c)}$ and $r_{200}$ is the virial radius.
\begin{table}
\centering
\caption{Notation reference.}
\label{Notation}
\begin{tabular}{ll}
\hline
Notation & Definition \\ \hline\hline
$M(z_{i})$ & Total halo mass at $z_{i}$,\\
& defined as $M_{200}$\\
$r_{200}$ or $r_{200}[M(z_{i}),z_{i}]$ & Virial radius at $z_{i}$ of a halo of\\
& total mass $M(z_{i})$\\
$r_{-2}$ or $r_{-2}[M(z_{i}),z_{i}]$ & NFW scale radius at $z_{i}$\\
$c$ or $c[M(z_{i}),z_{i}]$ & NFW concentration at $z_{i}$\\
$M_{r}(r,z_{i})$ & $M(<r)$, mass enclosed within $r$\\
& at $z_{i}$ of a halo of total mass $M(z_{i})$\\
$M_{r}(r_{-2},z_{i})$ & Mass enclosed within $r_{-2}$ at $z_{i}$\\
$\tilde{M}(z,M(z_{i}),z_{i})$ & Mass at $z$ of a halo with mass\\
& $M(z_{i})$ at $z_{i}$\\
$z_{-2}$ & Formation redshift, when equating\\
& $\tilde{M}(z_{-2},M(z_{i}),z_{i})$ to $M_{r}(r_{-2},z_{i})$\\
$\langle\rho\rangle(<r_{-2},z_{i})$ & Mean density within $r_{-2}$ at $z_{i}$\\
$\rho_{\rm{crit},0}$ & Critical density today\\
$\rho_{\rm{crit}}(z_{i})$ & Critical density at $z_{i}$\\
\hline
\end{tabular}
\end{table}
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=0.95\textwidth]{Fig4.ps}
\end{center}
\caption{Concentration$-$mass relations at $z=0$ (left panel), $z=1$ (middle panel) and $z=2$ (right panel). The dots in the panels correspond to individual, relaxed halos illustrating the scatter in the relation. The simulations assume the WMAP5 cosmological parameters and have box sizes of 400, 200, 100, 50 and 25 $\, h^{-1}{\rm Mpc}$, as indicated. Because of resolution limits only halos in the mass ranges indicated in the top legend were used from a particular simulation. The open symbols show the median $c-M$ values in logarithmic mass bins of width $\delta\log_{10}M=0.4$. The solid line shows the prediction of the $c-M$ model obtained from the halo MAH as described in Section~\ref{cM_model}.}
\label{3plots}
\end{figure*}
We begin by fitting NFW profiles to all halos at $z_{i}$ that contain at least $10^{4}$ dark matter particles within the virial radius. Throughout this work, we define the virial radius as $r_{200}$, the radius for which the mean internal density is 200 times the critical density. Then, for each halo, all particles in the range $-1.25\le \log_{10}(r/r_{200})\le 0$ are binned radially in equally spaced logarithmic bins of size $\Delta\log_{10}r=-0.078$. The density profile is then fitted to these bins by performing a least-squares minimization of the difference between the logarithmic densities of the model and the data, assuming equal weighting. The corresponding mean enclosed mass, $M_{r}(r_{-2},z_{i})$, and mean inner density at $r_{-2}$, $\langle\rho\rangle(<r_{-2},z_{i})$, are found by interpolating along the cumulative mass and density profiles from $r=0$ to $r_{-2}=r_{200}/c$, where $c$ is the concentration from the fit of the NFW halo. Then we generate merger trees for these halos and by interpolation we determine the redshift $z_{-2}$ at which $\tilde{M}(z_{-2},z_{i})=M_{r}(r_{-2},z_{i})$.
In order to obtain robust estimates and to test whether the $c-M$ relation includes an upturn in the median concentrations of massive halos (\citealt{Prada,Dutton14,Diemer14b}), we only consider `relaxed' halos. We define relaxed halos as those halos for which the separation between the most bound particle and the centre of mass of the FoF halos is smaller than $0.07R_{\rm{vir}}$ (following \citealt{Maccio07}, \citealt{Neto} and \citealt{Duffy08}), where $R_{\rm{vir}}$ is the radius within which the mean density is $\Delta$, as given by \citealt{BryanNorman}, times the critical density. Our relaxed sample contains 2425 halos at $z=0$, 726 halos at $z=1$, 226 halos at $z=2$ and 78 and 20 halos at $z=3$ and $z=4$, respectively.
The left panel of Fig. \ref{formation_time} shows the mean density within the NFW scale radius, $r_{-2}$, at redshift $z_{i}$. The median values of $\langle\rho\rangle(<r_{-2},z_{i})$ follow the best-fitting relation
\begin{equation}\label{rho2}
\langle\rho\rangle(<r_{-2},z_{i})=200\frac{c[M(z_{i}),z_{i}]^{3}Y(1)}{Y(c[M(z_{i}),z_{i}])}\rho_{\rm{crit}}(z_{i}),
\end{equation}
\noindent expressed as a function of the critical density of the Universe at $z_{-2}$,
\begin{equation}\label{rho_crit}
\rho_{\rm{crit}}(z_{-2})=\rho_{\rm{crit,0}}[\Omega_{\rm{m}}(1+z_{-2})^{3}+\Omega_{\Lambda}],
\end{equation}
\noindent where $\rho_{\rm{crit,0}}=3H^{2}(z=0)/8\pi G$. Note that densities along both the $x-$ and $y-$axes are expressed in units of the critical density at $z_{i}$. Each dot in the panel corresponds to an individual relaxed halo identified at $z_{i}$ and coloured by mass according to the colour bar at the top of the plot. The open symbols show the median value of the sample in logarithmic mass bins of width $\delta\log_{10}M=0.4$ and are coloured by $z_{i}$ as indicated in the legend\footnote{Note that it is possible for individual halos to appear multiple times in Fig.~\ref{formation_time} (left panel). For example, a $10^{13}\,{\rm M_{\odot}}$ halo at $z=0$, has a total mass of $\sim 10^{12.2}\,{\rm M_{\odot}}$ at $z=2$, therefore the halo will be included in the $\rho_{\rm{crit}}(z_{-2})-\langle\rho\rangle(<r_{-2},z_{i})$ relation at $z_{i}=0$ but also at $z_{i}=2$.}.
At each redshift $z_{i}$, the $\rho_{\rm{crit}}(z_{-2})-\langle\rho\rangle(<r_{-2},z_{i})$ correlation clearly shows that halos which collapsed earlier have denser cores.
We perform a least-squares minimization of the quantity $\Delta^{2} = \frac{1}{N}\sum_{j=1}^{N}[\langle\rho_{j}\rangle(<r_{-2},z_{i})-F(\rho_{\rm{crit},j}(z_{-2}),A)]$, where $j$ goes from 1 to the number of dark matter halos, N, at $z_{i}$ and $F(\rho_{\rm{crit},j}(z_{-2}),A)=A\times \rho_{\rm{crit},j}(z_{-2})$, to obtain the constant of proportionality, $A$. The solid line corresponds to the best-fit to the $\rho_{\rm{crit}}(z_{-2})-\langle\rho\rangle(<r_{-2},z_{i})$ relation, and we find (in agreement with Ludlow et al. 2014) that the average relation
\begin{equation}\label{rho_formation}
\frac{\langle\rho\rangle(<r_{-2},z_{i})}{\rho_{\rm{crit}}(z_{i})} = A\times \frac{\rho_{\rm{crit}}(z_{-2})}{\rho_{\rm{crit}}(z_{i})}
\end{equation}
\noindent is maintained through time with $A=887\pm 36$, where the $1\sigma$ error was obtained from the least-squares fit.
Using eqs. (\ref{rho2}) and (\ref{rho_crit}) we can rewrite this relation as
\begin{equation}\label{zf_relation}
\frac{c[M(z_{i}),z_{i}]^{3}Y(1)}{Y(c[M(z_{i}),z_{i}])} = \frac{A}{200}\frac{[\Omega_{\rm{m}}(1+z_{-2})^{3}+\Omega_{\Lambda}]}{[\Omega_{\rm{m}}(1+z_{i})^{3}+\Omega_{\Lambda}]},
\end{equation}
\noindent The right panel of Fig. \ref{formation_time} shows the ${c-z_{-2}}$ relation (solid lines) given by eq.~(\ref{zf_relation}) for various $z_{i}$. The open symbols correspond to the median values of the sample in logarithmic mass bins of width $\delta\log_{10}M=0.4$. The grey areas show the scatter in $z_{-2}$.
\subsection{Semi-analytic model for halo concentration}\label{halo_concentration}
In this section we describe the semi-analytic model for halo concentration as a function of halo mass and redshift. This model combines the analytic model for the halo MAH given by eqs. (\ref{MzLCDM_highz}-\ref{zf_LCDM}) and the empirical relation between $z_{-2}$ and $c$ given by eq. (\ref{zf_relation}).
We begin by calculating $\tilde{M}(z_{-2},M(z_{i}),z_{i})$ from eq.~(\ref{MzLCDM_highz}), and use the equality
\begin{equation}
\frac{\tilde{M}(z_{-2},M(z_{i}),z_{i})}{M(z_{i})}=\frac{M_{r}(r_{-2},z_{i})}{M(z_{i})}=\frac{Y(1)}{Y(c[M(z_{i}),z_{i}])},
\end{equation}
\noindent which follows from eqs. (\ref{z2}) and (\ref{M2}) and is valid under the assumption that the halo density profile follows the NFW profile, to obtain
\begin{equation}\label{eq_c1}
\frac{Y(1)}{Y(c[M(z_{i}),z_{i}])}=(1+z_{-2}-z_{i})^{\tilde{\alpha}}e^{\tilde{\beta}(z_{-2}-z_{i})},
\end{equation}
\noindent where $\tilde{\alpha}$ and $\tilde{\beta}$ are given by eqs. (\ref{alpha_tilde}) and (\ref{beta_tilde}), respectively. Next, we combine eqs.~(\ref{zf_relation}) and (\ref{eq_c1}) to obtain the concentration, $c[M(z_{i}),z_{i}]$, of a halo of total mass $M(z_{i})$ at $z_{i}$. We remind the reader that throughout this work the adopted halo mass definition is $M_{200}$, and the concentrations are therefore defined as $c=c_{200}$.
Fig. \ref{3plots} shows the concentration$-$mass relation at $z=0$ (left panel), at $z=1$ (middle panel), and at $z=2$ (right panel). The dots in the panels correspond to individual relaxed halos identified in the simulations at $z_{i}=0,1$ and $2$, whereas the open symbols correspond to the median values in logarithmic mass bins of width $\delta\log_{10}M=0.4$. The solid line shows the $c-M$ relation that results from the semi-analytic model described above. We find excellent agreement between the median values from the simulations and the $c-M$ relation predicted by the semi-analytic model at all redshifts.
So far we have adopted the WMAP5 cosmology. In Appendix \ref{cosmo_dependence}, we discuss the dependence of our $c-M$ relation model on cosmology and extend it to make it suitable for any values of the cosmological parameters.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=\textwidth]{Fig5.ps}
\end{center}
\caption{Concentration $-$ mass relation at $z=0, 1, 2, 3$ and $4$ under the WMAP5 cosmology for the relaxed halo sample (left panel) and the full sample (middle panel). The open symbols indicate the median concentrations in logarithmic mass bins of width $\delta\log_{10}M=0.25$ at $z=0$ and $z=1$, and $\delta\log_{10}M=0.30$, $0.40$ and $0.50$ at $z=2$, $3$ and $4$, respectively. Only bins containing at least 10 halos are shown. The error bars show the $1\sigma$ confidence limits. The dashed lines correspond to the best-fitting power laws to the open symbols. In the left panel, the solid grey line shows the $c-M$ relation predicted by the semi-analytic model. The right panel shows the fraction of relaxed halos, with respect to the complete sample, for each mass bin and redshift. The inclusion of unrelaxed halos results in a flattening of, or even an upturn in, the $c-M$ relation at high redshift.}
\label{relaxed_comparison}
\end{figure*}
\subsection{Impact of relaxedness on the $c-M$ relation}
Several recent studies (\citealt{Klypin2011,Prada,Dutton14,Diemer14b}) have found that the $c-M$ relation flattens at high redshift and exhibits an `upturn' at the high-mass end, meaning that the concentration increases with halo mass for the most massive halos. In this section we investigate whether this interesting behaviour is seen in our semi-analytic model or in the simulation outputs.
Our model does not predict an upturn. The model relates $c$ to the MAH via the formation redshift, $z_{-2}$ (see Fig. \ref{formation_time}, right panel), which decreases with halo mass, meaning that more massive halos are less concentrated because they formed more recently. If $c$ were to increase with halo mass, then high-mass halos would have to form earlier than low-mass ones, at a point when the Universe was denser. This behaviour is neither seen in our simulations (see Fig. \ref{Mz12_LCDM}, coloured lines), as we only consider relaxed systems, nor predicted by EPS theory, because it would be antihierarchical for such systems.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=\textwidth]{Fig6.ps}
\end{center}
\caption{Comparison of the $c-M$ relation predicted by our model (solid lines) with the relations of \citet{vandenBosch14} (left-panel), \citet{Diemer14b} (middle-panel) and \citet{Dutton14} (right-panel). Note that \citet{Dutton14} consider only relaxed halos for their best-fitting model and fit their model in the halo mass range $10^{10}-10^{15}\,{\rm M_{\odot}}$, whereas \citet{Diemer14b} and \citet{vandenBosch14} use their full halo sample and fit their model in halo mass range $10^{10}-10^{15}\,{\rm M_{\odot}}$ and $10^{11}-10^{15}\,{\rm M_{\odot}}$, respectively.}
\label{comparison}
\end{figure*}
To investigate further, we use the simulation outputs to calculate concentrations by fitting NFW profiles to halos that are resolved with at least $10^{4}$ particles within the virial radius, and for which the convergence radius\footnote{The convergence radius is defined such that the two$-$body dynamical relaxation time-scale of the particles is similar to the age of the universe. For more details see \citet{Power03} or \citet{Duffy08}.} (\citealt{Power03}) is smaller than the minimum fit radius of 0.05 times the virial radius. In addition, we consider two halo samples. A relaxed halo sample\footnote{As proposed by \citet{Neto}, relaxed halos are defined as those halos for which the separation between the most bound particle and the centre of mass of the FoF halo is smaller than 0.07 times the virial radius.} and a full halo sample. When considering only relaxed halos, as we have done so far, we find that we restrict our halo sample to around $80\%$ of the total at $z=0$, $65\%$ at $z=1$, $55\%$ at $z=2$, $50\%$ at $z=3$ and $43\%$ at $z=4$.
Fig.~\ref{relaxed_comparison} shows the $c-M$ relation (at various $z$) of the relaxed sample (left panel) and the full sample (middle panel). These panels show the median value of the concentration (open symbols) in logarithmic mass bins of width $\delta\log_{10}M=0.25$ at $z=0$ and $1$, and $\delta\log_{10}M=0.30$, $0.40$ and $0.50$ at $z=2$, $3$ and $4$, respectively. We increase the bin size with redshift so that each bin at a fixed mass contains on average approximately the same number of halos. For each bin the $1\sigma$ error bars were determined by bootstrap resampling the halos. Only bins containing at least 10 halos are shown. The dashed lines correspond to the best-fitting power laws to the open symbols. In addition, the left panel shows the $c-M$ relations predicted by the semi-analytic model in the solid grey lines. The middle panel shows a strong flattening and upturn in the $c-M$ relation at high $z$, in agreement with \citet{Munoz} and \citet{Prada}. However, this upturn is not seen for the relaxed sample. Thus, we conclude that the previously seen upturn results from the inclusion of unrelaxed halos, in agreement with \citet{Ludlow12}. We show the fraction of relaxed halos (with respect to the total sample) for each mass bin and redshift in the right panel of Fig.~\ref{relaxed_comparison}. We find that the relaxed fraction tends to decrease towards high mass and redshift.
Our results suggest that the dynamical state of dark matter halos should be considered when analysing the parameters that describe the halo internal structure, because the density profiles of unrelaxed halos are poorly captured by the NFW fitting formula (e.g. \citealt{Neto}). Because halo concentrations are clearly affected by transient departures from equilibrium, we only consider relaxed halos in the remainder of this work.
\subsection{Comparison with previous studies}
In this section we compare the $c-M$ relations of the most recent studies on dark matter halo concentrations, \citet{vandenBosch14} (hereafter, vdB14), \citet{Diemer14b} (hereafter, DK14), \citet{Dutton14} (hereafter, DM14) and \citet{Ludlow14} (hereafter, L14), with the model presented in this work.
vdB14 used the $c-M$ relation of \citet{Zhao09} (obtained from fits of a full halo sample from numerical simulations) and adjusted the parameters by fitting it to the $c-M$ relation of the full halo sample from the Bolshoi simulations. vdB14 assumed the Bolshoi cosmology (consistent with WMAP7, \citealt{Komatsu11}), but they made use of a semi-analytic model to scale their model to any cosmology. We assume the Planck cosmology and use the publicly available code of vdB14 to calculate their $c-M$ relation. DK14 obtained a concentration model given by a best-fit seven parameter function of peak height ($\nu$) and slope of the linear power spectrum. They considered their full halo sample and extended their model to make it suitable for any cosmology. Finally, DM14 followed the evolution of the concentration of relaxed dark matter halos from a series of $N$-body simulations that assumed the Planck cosmology. DM14 fitted a power-law to the $c-M$ relation and restricted their analysis to relaxed halos only.
The left panel of Fig.~\ref{comparison} shows a comparison of our $c-M$ model (solid lines) to the model of vdB14. To compare with vdB14, we predicted the concentrations using the analytic expression for the MAH assuming the Planck cosmology (shown in Fig.~\ref{Mz12_LCDM}) and a $z_{-2}-c$ relation with a constant of proportionality of 850 instead of the value 887 used for the WMAP5 cosmology (see Appendix~\ref{cosmo_dependence} for a discussion of the cosmology dependence of our model). We find broad agreement with the relation of vdB14 only at $z=1$ and $2$. In their work, \citet{vandenBosch14} used the \citet{Zhao09} model which assumes that $c$ never drops below 4 at high redshift.
The middle-panel of Fig.~\ref{comparison} shows the DK14 $c-M$ relation calculated assuming the Planck cosmology. As they included their entire sample of halos for their $c-\nu$ relations, they obtained an upturn at the high-mass end at all redshifts. We find that our model predicts concentrations that are a factor of 1.2 larger just before the high-mass upturn. Finally, the right-panel of Fig.~\ref{comparison} shows reasonable agreement between our model and the DM14 $c-M$ relation for $z=0, 1, 2$ and $3$ although the results diverge at low masses. In their work, DM14 fitted a power-law, $c\propto M^{\alpha}$, to the $c-M$ relation at all redshifts, and found that the slope, $\alpha$, increases from $-0.1$ at $z=0$, to $0.03$ at $z=5$. As they restricted their halo sample to relaxed halos, they did not obtain a significant upturn at the high-mass end of the $c-M$ relation.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=0.8\textwidth]{Fig7.ps}
\end{center}
\caption{Predicted concentration $-$ mass relation for the WMAP5 cosmology over a wide range of halo masses (${\log_{10}M/\,{\rm M_{\odot}}=[-2 ,16]}$) and redshifts (${z=0-20}$). The solid lines correspond to our $c-M$ model obtained from the halo MAH as described in Section~\ref{cM_model}. The lines are coloured as a function of redshift as indicated. The dashed lines correspond to power-law fits to the high-mass $c-M$ relation.}
\label{cvsM_big}
\end{figure*}
Fig.~\ref{comparison} shows that the physically motivated model presented in this work yields $c-M$ relations that are generally in agreement with previous results. However, the important improvement with respect to previous works is that we are presenting a physical analytic model that can then be extrapolated to very low-masses, and is suitable for any cosmology.
The model for dark matter halo concentrations presented in this work strongly relies on the relation $\langle\rho(r_{-2})\rangle-\rho_{\rm{crit}}(z_{-2})$, which supports the idea that halos grow inside-out. This relation was introduced in \cite{Ludlow13} and explored in L14, who recently presented a related model for the concentration-mass relation. In their work, L14 used the average MAHs from \cite{van} and \cite{Zhao09} that begin at $z_{i}=0$\footnote{L14 $c-M$ model used MAHs from \cite{van} and \cite{Zhao09} to show specific examples on how to construct a $c(M, z)$ relation for a given MAH, but any MAH model can be used.}. They fitted the halo MAHs, written as $M(\rho_{\rm{crit}})$, with the NFW profile expressed in terms of the enclosed density. They looked for a correlation between the concentration parameter $c_{\rm{MAH}}$, that results from an NFW fit to the halo MAH, and the concentration parameter of the halos density profile, $c_{\rm{NFW}}$, and used the best-fitting relation to predict $c_{\rm{NFW}}$ from $c_{\rm{MAH}}$. L14 and this work use the same formation redshift definition to connect concentrations with halo MAHs. L14 used the $\langle\rho(r_{-2})\rangle-\rho_{\rm{crit}}(z_{-2})$ relation to find the $c_{\rm{MAH}}-c_{\rm{NFW}}$ relation, whereas in this work we used the analytic MAH model to define formation redshifts and used the $\langle\rho(r_{-2})\rangle- \rho_{\rm{crit}}(z_{-2})$ relation to predict concentrations. Although there is good agreement between L14 and our $c-M$ relation at $z=0$, there are differences in the relations at high$-z$, e.g. a factor of 1.2 difference between the concentrations of a $10^{10}h^{-1}\rm{M}_{\odot}$ halo at $z=2$ ($c\sim 5.25$ versus $c_{\rm{L14}}\sim 6.3$), and a factor of 1.58 for a $10^{5}h^{-1}\rm{M}_{\odot}$ halo at $z=2$ ($c\sim 7.95$ versus $c_{\rm{L14}}\sim 12.58$), for the WMAP5 cosmology. Those differences are mainly due to the different MAH models. Since the $\langle\rho(r_{-2})\rangle-\rho_{\rm{crit}}(z_{-2})$ relation is essentially equivalent to the $c_{\rm{MAH}}-c_{\rm{NFW}}$ relation, we expect L14 and our semi-analytic model to give consistent results if the same MAH model is used. We believe however that we have improved upon the L14 $c-M$ model by combining the $\langle\rho(r_{-2},z_{i})\rangle-\rho_{\rm{crit}}(z_{-2},z_{i})$ relation with an analytical MAH model, $M(z,z_{i})$, that begins at any redshift $z_{i}$, and allows a detailed analysis of the redshift dependence of the $c(M,z)$ relation for relaxed halos. Another important difference is the tentative evidence for a cosmology dependence in the $\langle\rho(r_{-2},z_{i})\rangle-\rho_{\rm{crit}}(z_{-2},z_{i})$ relation (for a discussion see Appendix~\ref{cosmo_dependence}).
\subsection{Extrapolation to low halo masses and high redshifts}
Because our semi-analytic model for halo concentration is physical, rather than a purely empirical fit to the simulation results, we can use it to extrapolate beyond the mass and redshift ranges spanned by our simulations, assuming that the $z_{-2}-c$ relation given by eq. (\ref{zf_relation}) holds. Fig.~\ref{cvsM_big} shows the predicted concentration-mass relation for a wide range of halo masses (${\log_{10}M/\,{\rm M_{\odot}}=[-2 ,16]}$) and redshifts (${z=0-20}$). The dashed lines correspond to the high-mass power-law $c-M$ relations at low redshift. These are included to aid the comparison of the slopes of the $c-M$ relation in the high- and low-mass regimes. There is a clear `break' in the $z=0$ $c-M$ relation. For $M>10^{12}\,{\rm M_{\odot}}$ concentration scales as $c\propto M^{-0.083}$, whereas at $M<10^{9}\,{\rm M_{\odot}}$ it scales as $c\propto M^{-0.036}$. The change of slope around these halo masses is substantial up to $z=3-4$. However, at $z>4$ there is no `break' in the $c-M$ relation. In Section~\ref{c_evolution}, we provide a tentative explanation for the physical origin of the break in the $c-M$ relation.
We provide fitting functions for the $c-M$ relation in the high-$z$ and low-$z$ regimes. The following expression is suitable for the low-redshift regime ($z\le 4$) and at all halo masses,
\begin{eqnarray}\label{c_lowz}
\log_{10} c &=& \alpha+\beta\log_{10}(M/\,{\rm M_{\odot}})[1+\gamma(\log_{10}M/\,{\rm M_{\odot}})^{2}],\\\nonumber
\alpha &=& 1.62774-0.2458(1+z)+0.01716(1+z)^{2},\\\nonumber
\beta &=& 1.66079+0.00359(1+z)-1.6901(1+z)^{0.00417},\\\nonumber
\gamma &=& -0.02049+0.0253(1+z)^{-0.1044}.
\end{eqnarray}
In the high-redshift regime the $c-M$ relation can be fitted using only two parameters. The following expression is suitable for $z>4$ and at all halo masses,
\begin{eqnarray}\label{c_highz}
\log_{10} c &=& \alpha+\beta\log_{10}(M/\,{\rm M_{\odot}}),\\\nonumber
\alpha &=& 1.226-0.1009(1+z)+0.00378(1+z)^{2},\\\nonumber
\beta &=& 0.008634-0.08814(1+z)^{-0.58816}.
\end{eqnarray}
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=0.46\textwidth]{Fig8.ps}\\
\end{center}
\caption{{\it{Top panel}}: evolution of the concentrations of halos that at $z=0$ have masses of $M_{0}=10^{6}, 10^{8}, 10^{10}, 10^{12}$ and $10^{14}\,{\rm M_{\odot}}$, as indicated in the legends. {\it{Bottom panel}}: MAHs of halos of the same masses as in the top panel.}
\label{cMH_figa}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=0.46\textwidth]{Fig9.ps}\\
\end{center}
\caption{Dashed curves show the concentration$-$mass relation at various redshifts $z$. The solid and dashed lines show the $c-M$ evolution of halos whose $c(z)$ and $M(z)$ are shown in the top and bottom panels of Fig.~\ref{cMH_figa}, respectively. All results were obtained using the physical model described in Sections \ref{Halo_history_modelling} and \ref{cM_model}.}
\label{cMH_figb}
\end{figure}
The above fitting functions have been calculated assuming the WMAP5 cosmology. Appendix~\ref{cosmo_dependence} provides a series of best-fitting relations for the Planck cosmology, as well as a short discussion of the cosmology dependence of the $c-M$ relation presented in this work. In addition, Appendix~\ref{script_details} provides a description of a simple code (available for download in IDL and Python) that computes concentrations, MAHs and accretion rates as a function of redshift for any cosmology.
The $c-M$ model presented in this work predicts a concentration of $c=3$ for microhalos of $10^{-7}\,{\rm M_{\odot}}$ at $z=31$, in agreement with simulations of microhalo formation from \citet{Anderhalden} and \citet{Ishiyama}. In their work, \citet{Anderhalden} compared the empirical $c-M$ relations from \citet{Bullock} and \citet{Maccio08} to their simulation outputs, and concluded that extrapolating simple power-law approximations to typical microhalo scales results in an overestimation of $c$ by up to a factor of 10 at low$-z$. We also find large differences between extrapolations of the high-mass power-law fits to low-masses and the predictions of our physical model, as can be seen by comparing the dashed and solid lines in Fig.~\ref{cvsM_big}. This impacts calculations of the dark matter (DM) annihilation signal boost (see e.g. \citealt{Sanchez}) and the power from DM annihilation (see e.g. \citealt{Mack}), which make extensive use of $c-M$ relations at various redshifts. In Section~\ref{DM_annihilation_signal} we analyse this issue in more detail by calculating the power from DM annihilation, assuming either the \citet{Duffy08} $c-M$ relation (an empirical power-law fit for high halo masses) or the $c-M$ relation from this work.
\section{Evolution of the concentration}\label{c_evolution}
In this section we use our semi-analytic model to investigate the evolution of concentration and the effects that determine the slope of the $c-M$ relation. The top panel of Fig. \ref{cMH_figa} shows the evolution of the concentration of halos that have masses of $M_{0}=10^{6}, 10^{8}, 10^{10}, 10^{12}$ and $10^{14}\,{\rm M_{\odot}}$ at $z=0$. The bottom panel shows the corresponding halo MAHs normalized to the final halo mass at $z=0$ ($M_{0}$). We computed $c(z)$ and $M(z)$ following the models described in Sections~\ref{cM_model} and \ref{MAHmodel_highz}, respectively.
In Paper I, we used EPS theory to show that the MAH of all halos can be described by the expression ${M(z)=M_{0}(1+z)^{\alpha}e^{\beta z}}$, where the exponential is due to the fast growth at high$-z$ and the power-law due to the slow growth at low$-z$. In addition, the parameters $\alpha$ and $\beta$ depend on halo mass. As a result, MAHs of halos larger than $10^{11}\,{\rm M_{\odot}}$ are mainly characterized by an exponential growth, whereas lower-mass halos exhibit a MAH closer to a power law, as can be seen in the bottom panel of Fig.~\ref{cMH_figa}.
Comparing the coloured curves in the top and bottom panels of Fig.~\ref{cMH_figa}, we see an interesting relation between the evolution of $c$ and the corresponding MAH. Dark matter halos with a small growth rate are {\it{appear}} to contract, and so their concentrations grow rapidly. This can be understood as follows. At low redshift, during the dark energy dominated epoch, $M(z)$ of low-mass halos is characterized by a power law (\citealt{PaperI}). During this epoch, there is a drop in the accretion and merger rates of small halos, and the halo mass increases due to the evolution of the reference density used in the spherical overdensity definition of the halo ($\rho_{\rm{crit}}(z)$ in this case). This so-called {\it{pseudo-evolution}} of the halo mass is thus driven by the halo mass definition rather than the accretion of new material (see \citealt{Diemer13} and references therein). The pseudo-evolution of the halo mass gives the impression that concentrations are increasing because of contraction of the bound cores (\citealt{Tasitsiomi,Zhao03,LuYu,Li08,vandenBosch14}), when in fact the core radius remain constant. Indeed, the evolution of the $c-M$ relation has been shown to be sensitive to the definition of halo mass (e.g. \citealt{Duffy08}). If we assume that $r_{-2}$ is constant in the redshift range $z=0-1$, then the increase in $r_{200}$ due to the drop in $\rho_{\rm{crit}}$ gives the approximate increase in the concentration values. We find $\frac{c(z=0)}{c(z=1)}=\frac{r_{200}(z=0)}{r_{200}(z=1)}=\left(\frac{\rho_{\rm{crit}}(z=1)}{\rho_{\rm{crit}}(z=0)}\right)^{1/3}\sim 1.4$, in agreement with the increase in concentration of a $M_{0}=10^{6}\,{\rm M_{\odot}}$ halo.
In the high-redshift regime ($z\gg 1$, matter-dominated epoch), the halo MAH is mainly characterized by exponential growth. During this time, concentrations grow by a factor of 2 (from $z=8$ to $z=2$) for a $M_{0}=10^{6}\,{\rm M_{\odot}}$ halo, decreasing to a factor of 1.08 (from $z=8$ to $z=2$) for a $M_{0}=10^{14}\,{\rm M_{\odot}}$ halo. The pseudo-evolution of the halo mass is negligible in comparison with the high accretion rates, and the core radius increases simultaneously with the virial radius, hence the concentration hardly grows.
In the case where the halo mass history is characterized by exponential growth at all $z$, representing the situation of a universe with no dark energy but $\Omega_{\rm{m}}\le 1$, we find that concentrations do not reach such large values at $z=0$. We thus conclude that the evolution of the concentration is indirectly affected by the accelerated expansion of the Universe through the MAH and the halo mass definition.
Next, we analyse how the evolution of the concentration determines the change in slope of the $c-M$ relation. Fig.~\ref{cMH_figb} shows $c-M$ relations at various redshifts (dashed lines), and the $c-M$ evolution of halos with $M_{0}=10^{6}, 10^{8}, 10^{10}, 10^{12}$ and $10^{14}\,{\rm M_{\odot}}$ (solid lines). From this figure, we see that the `break' in the low-redshift $c-M$ relation that occurs at $M\sim 10^{11}\,{\rm M_{\odot}}$ is produced by the change in the halo MAH.
As mentioned, $M(z)$ changes from being dominated by exponential growth for high-mass halos, to power-law growth for low-mass halos. It is natural to ask why the break in $M(z)$ (and consequently in $c-M$) occurs at $\sim 10^{11}\,{\rm M_{\odot}}$. The answer is given by the rms of the linear theory density perturbation field, $\sigma$, that determines at which halo mass $M(z)$ changes from power-law dominated to exponential dominated. Since $\sigma\gg1$ at low masses ($\ll 10^{11}\,{\rm M_{\odot}}$) and $\sigma\ll 1$ at high masses ($\gg 10^{11}\,{\rm M_{\odot}}$), the low values of $\sigma$ at high masses increase the value of the $\beta$ parameter in the exponential function of the $M(z)$ model, and thus $M(z)$ is mostly dominated by the exponential growth. As a result, higher mass halos increase their mass faster, their inner cores increase with the virial radius, and their concentrations do not grow as rapidly. The different growth rate of the concentrations produced by the change in the halos MAH, creates the `break' in the $c-M$ relation.
Therefore, the break can be understood as being produced by the varying power in the density perturbations through the halo MAH, where the MAH of low-mass halos at $z<1$ is mostly driven by pseudo-evolution. The break is less prominent at $z>1$, because at higher redshifts $M(z)$ is mostly exponential for all halo masses (see eq.~\ref{alpha_tilde}, $\tilde{\alpha}\rightarrow 0$ for $z_{i}>0$ due to the growth factor and Fig.~\ref{MAH_z0}).
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=\textwidth]{Fig11.ps}\\
\caption{{\it{Left panel:}} Concentration-mass relation at various $z$ ($z=0-10$, top to bottom) from this work (blue solid lines) and from Duffy et al. (2008) (red dashed lines). At $z=10$ and at a mass-scale of $1\,{\rm M_{\odot}}$ concentrations of Duffy et al. (2008) are a factor of 10 larger than concentrations predicted by this work, and a factor of 40 for a $10^{-9}\,{\rm M_{\odot}}$ halo. {\it{Middle panel:}} DM annihilation power per hydrogen nucleus as a function of redshift. The dashed line corresponds to the smooth component of the power and the dot-dashed lines to the structure component. The solid lines show the sum of the two components (structure+smooth). The red lines correspond to the power assuming the \citet{Duffy08} $c-M$ relation, whereas the blue lines correspond to the power assuming the $c-M$ relation from this work. {\it{Right panel:}} as the middle panel, but showing the effective DM energy density as a function of redshift.}
\label{DM}
\end{center}
\end{figure*}
\section{Implications for the dark matter annihilation signal}\label{DM_annihilation_signal}
Dark matter (DM) particles are predicted to self-annihilate into Standard Model particles, thus injecting energy into the surrounding medium (e.g. \citealt{Furnaletto2006}). In this section we calculate the DM annihilation rate per unit volume produced by a smooth density field of DM (dominant before structure formation) and by cosmic structures (halos and microhalos). We separate DM into a smooth and structure component because the spatial distribution of mass is almost completely smooth at very early cosmic times. Later gravitational instability causes overdensities to grow, until micro DM halos form. We follow \citet{Cirelli} and \citet{Mack} in this calculation, and obtain the DM energy density and mean power from DM annihilation, assuming the physical $c-M$ relation from this work and from extrapolations of the fits to simulations from \citet{Duffy08}. Below we briefly describe the calculation of the DM annihilation rate produced by cosmic structures.
The DM annihilation rate per unit volume results from the sum of two parts, a structure contribution and a smooth contribution. The smooth contribution, dominant before structure formation, $z\gtrsim 100$, can be written as
$$R^{\rm{smooth}}(z)= \frac{\langle\sigma\nu\rangle}{2m_{\chi}^{2}}\rho^{2}_{\rm{DM,0}}(1+z)^{6}.$$
\noindent Where $m_{\chi}$ is the mass of the DM particle, $\langle\sigma\nu\rangle$ the self-annihilation cross-section, and $\rho_{\rm{DM,0}}$ the smooth DM density today, $\rho_{\rm{DM,0}}=\Omega_{\rm{DM}}\rho_{\rm{crit}}$.
The DM annihilation rate per unit volume due to halos, $R$, is given by
\begin{equation}\label{integral}
R(z)=\frac{\langle\sigma v \rangle}{2m_{\chi}^{2}}\int dM\frac{dn}{dM}(z,M)(1+z)^{3}\int dr 4\pi r^{2}\rho^{2}(r,M).
\end{equation}
\noindent Here $m_{\chi}$ is the mass of the dark matter particle and $\langle\sigma v \rangle$ is the average annihilation cross section, which we assume to be $100$ GeV and $10^{-26}\rm{cm}^{3}\rm{s}^{-1}$, respectively (e.g. \citealt{Aprile2012}). For the halo mass function, $\frac{dn}{dM}(z,M)$, we adopt the expression from \citet{Reed}. For $\rho(r,M)$ we use the NFW density profile. We use $M_{200}$ as the halo mass definition.
Next, we calculate the effective DM density from structure formation, defined as
\begin{eqnarray}
\rho_{\rm{DM}}^{\rm{eff}}(z) &\equiv&\rho_{\rm{DM},0}(1+z)^{3}R_{i}(z),\\
R_{i}(z) &\equiv & \left(\frac{2m_{\chi}^{2}}{\langle\sigma v \rangle}R(z)\right)^{1/2},
\end{eqnarray}
\noindent where $\rho_{\rm{DM},0}$ is the average DM density today, ${\rho_{\rm{DM},0}=\Omega_{\rm{DM}}\rho_{\rm{crit},0}}$, with $\Omega_{\rm{DM}}h^{2}=0.11$. In this calculation we assume the WMAP5 cosmology in order to facilitate a comparison with models using the extrapolation of the Duffy et al. (2008) power-law fit to the $c-M$ relation predicted by WMAP5 $N$-body simulations.
Similarly, we calculate the averaged volume power, per hydrogen nucleus, produced from DM annihilation events as
\begin{equation}
P(z)=2m_{\chi}c^{2}\frac{R(z)}{n_{\rm{H}}(z)},
\end{equation}
\noindent with $n_{\rm{H}}(z)= \Omega_{\rm{b}}\rho_{\rm{crit},0}(1-Y_{\rm{p}})(1+z)^{3}/m_{\rm{H}}$,\footnote{$\Omega_{\rm{b}}$ is the present day baryon density parameter, $Y_{\rm{p}}=0.24$ the primordial mass fraction of helium and $m_{\rm{H}}$ the proton mass.} the number density of hydrogen.
\subsection{Implications}
Several of the models that have been used to predict the DM annihilation signal (see e.g. \citealt{Lavalle,Pieri,Pinzke}) have extrapolated $c-M$ relations, obtained from power-law fits to simulation results, to mass far below the resolution limit of the simulations. These power-law extrapolations assign huge concentrations to the smallest halos, thus increasing the DM annihilation power. In this section we explore how our physically motivated $c-M$ relation, which flattens towards low-masses at low-redshift, affects the DM annihilation power when comparing it to the power calculated using an extrapolation of the \citet{Duffy08} power-law $c-M$ relation (the results are similar for other published power-law fits to the $c-M$ relation from simulations).
The left panel of Fig.~\ref{DM} shows a comparison between the concentration-mass relations at various redshifts ($z=0-10$) from this work (blue solid lines) and of \citet{Duffy08} (red dashed lines). For $z=0$ there is good agreement at the high-mass end between both relations. However, at $z=10$ and at a mass-scale of $1\,{\rm M_{\odot}}$, the concentrations of \citet{Duffy08} are a factor of 10 larger than the concentrations predicted by this work. For mass-scales of $10^{-9}\,{\rm M_{\odot}}$ the difference is a factor of 40. In the middle and right panels of Fig.~\ref{DM} we investigate the effects of these different $c-M$ relations, showing the DM annihilation power per hydrogen nucleus and the effective DM energy density, respectively, as a function of redshift. In the middle panel, the dashed line corresponds to the smooth DM component of the power whereas the dot-dashed lines correspond to the structure component. The solid line shows the sum of the two components (structure+smooth). In each case the red lines correspond to the power assuming the \citet{Duffy08} $c-M$ relation, whereas the blue lines correspond to the power assuming the $c-M$ relation from this work. The change in $c-M$ affects the normalization of the power as well as the redshift at which structures begin to dominate. Lower concentrations result in lower central densities. Since the annihilation rate per unit volume, $R$, scales as $\rho^{2}$, it is clear that $R$ should decrease accordingly. At $z=0$, the DM annihilation power that assumes the $c-M$ relation predicted by this work is two orders of magnitude lower than the power obtained by extrapolating the \citet{Duffy08} $c-M$ relation. In addition, the higher concentrations predicted by the \citet{Duffy08} $c-M$ relation imply that halos dominate the power over the smooth DM density component at higher redshifts. Adopting the $c-M$ relation from this work results in the power from structures starting to dominate at $z\approx 50$ rather than at $z\approx 85$ (in agreement with \citealt{Mack} and \citealt{Ng2014}). This lower redshift of structure formation dominating over the smooth component could have important implications for searches of the `Dark Ages' by radio telescopes (e.g. \citealt{Pritchard}).
\section{Discussion}\label{Discussion}
\begin{figure}
\centering
\includegraphics[angle=0,width=0.48\textwidth]{Fig12.ps}
\caption{Distribution of particles within $r_{200}$. We show the particle distribution of two different halo samples, the first sample contains halos of $10^{15}\,{\rm M_{\odot}}$ identified at $z_{i}=0$ (solid lines) and the second sample contains halos of $10^{14}\,{\rm M_{\odot}}$ identified at $z_{i}=1$ (dashed lines). The solid grey line corresponds to a NFW density profile with concentration $c=4$ and the vertical dashed line to the corresponding $r_{-2}$ radius. The different colours indicate the redshift ($z_{\rm{accr}}$) during which the particles were accreted on to the halos (note that the same colour corresponds to different redshift ranges for different $z_{i}$).}
\label{discussion_plot}
\end{figure}
Dark matter halo concentrations have recently been the subject of extensive analysis due to the controversial reports of an upturn at the high-mass end of the relation (\citealt{Munoz,Prada,Klypin14,Diemer14b}). However, the semi-analytic model for dark matter halo concentrations presented in this work does not predict such an upturn. In this section, we review the main assumptions that the model relies on and discuss the plausibility of the existence of the upturn.
First, the model assumes that the halo density profile is described by the NFW profile at all times. Although it is known that the Einasto profile (\citealt{Einasto}) is more accurate than the NFW profile (\citealt{Gao08}), it has an extra `shape' parameter that complicates the fitting procedure and affects the concentration. Also, the residuals from the systematic deviations from the NFW shape are generally smaller than $10\%$ and the NFW concentrations only differ by $10-20\%$ from Einasto fit and the velocity profile fit (as recently discussed by \citealt{Dutton14} and \citealt{Klypin14}). We thus conclude that using the NFW profile to predict densities is not a major determinant in the model.
Secondly, the model depends on the calibration of the $\rho_{\rm{crit}}(z_{-2})-\langle\rho\rangle(<r_{-2})$ relation, which implies that halo formation is an `inside out' process, where the central part of a dark matter halo (contained within $r_{-2}$) forms first, and later accretion and mergers increase the mass and size of the halo without adding much material to its inner regions (\citealt{Huss,Wang09}). We test the assumption of `halo formation' of our model. We analyse the distribution of particles within $r_{200}$, and differentiate the particles according to the period of time during which they were accreted. Fig.~\ref{discussion_plot} shows the radial distribution of particles of two different halo samples. The first sample contains halos of $10^{15}\,{\rm M_{\odot}}$ identified at $z_{i}=0$ (which are formed at $z_{-2}\approx 1$) and the second sample contains halos of $10^{14}\,{\rm M_{\odot}}$ identified at $z_{i}=1$ (formed at $z_{-2}\approx 2$). We analysed halos of different masses but focused on massive halos, because their large radial velocities makes them more likely to contain recently accreted particles in their inner regions, and they are located in the upturn in the $c-M$ relation.
Fig.~\ref{discussion_plot} shows the NFW density profile (in grey solid line) and the $r_{-2}$ radius (in vertical dashed line) for a concentration of $c=4$. The figure also shows the distribution of particles at $z_{i}=0$ (solid coloured lines) and at $z_{i}=1$ (dashed coloured lines). The different colours indicate the redshift ($z_{\rm{accr}}$) during which the particles were accreted on to the halos. The blue lines show that recently accreted particles are distributed around $r_{200}$ and that only a tiny fraction ($<2\%$ of the total, in the two cases), reside in the inner parts of the halo. In the case of the distribution of particles at $z_{i}=0$, $2.5\%$ of the total particles are in regions within $r_{-2}$ after being accreted during $z_{\rm{accr}}=0.25-0.5$, and $4.5\%$ during $z_{\rm{accr}}=0.5-1$. The same behaviour is observed in halos of different masses identified at higher redshifts. We find that $8\%$ of all particles accreted after the halo has formed are in the centre, not enough to significantly alter the mass within $r_{-2}$ so as to increase concentration. We then find the halo formation assumption that the model relies on to be valid.
Finally, in the calibration of the $\rho_{\rm{crit}}(z_{-2})-\langle\rho\rangle(<r_{-2})$ relation, we only consider relaxed halos. The selection conditions generally used to differentiate relaxed halos from unrelaxed (\citealt{Maccio07,Neto}) have recently been revisited by \citet{Klypin14}. These conditions include the virial parameters ($2K/|W|-1$, where $K$ and $W$ are the kinetic and potential energies), the offset parameter $X_{\rm{off}}$ (distance between the potential minimum and the center of mass), and the spin parameter. In their work, \citet{Klypin14} argued that the virial equilibrium condition is too simplistic and needs to include the effects of the surface pressure and external forces. They applied these corrections to the virial parameters and selected halos that had previously been rejected. As a result, they obtained an upturn in the high-mass end of the $c-M$ relation and claimed that the large concentration of massive halos is due to their infall velocities, which are more radial and result in deeper penetration of infalling mass into the halo that reaches the inner parts. In this work we selected relaxed halos using only the condition that $X_{\rm{off}}<0.07$ following \citet{Duffy08} and \citet{Neto}, who found that this simple criterion resulted in the removal of the vast majority of unrelaxed haloes. We did not use any additional criteria and did not find any upturn at high halo masses, but concluded that the strong flattening of the $c-M$ relation at high redshift is due to unrelaxed halos (Fig.~\ref{relaxed_comparison}). We cannot say we disagree with \citet{Klypin14} regarding the relaxation conditions, because our simulations do not have sufficiently large box sizes to model a large sample of the massive highly unrelaxed halos that likely `shape' the upturn.
\section{Conclusion}\label{Conclusion}
In this paper, we have linked the concentration of a halo to its MAH. We extended the analytic framework presented in Paper I to show that the halo mass history $\tilde{M}(z,M(z_{i}),z_{i})$ of a halo with mass $M(z_{i})$ at $z_{i}$ can be described by
$$\tilde{M}(z,M(z_{i}),z_{i})= M(z_{i})(1+z-z_{i})^{\tilde{\alpha}}e^{\tilde{\beta}(z-z_{i})},$$
\noindent where $\tilde{\alpha}$ and $\tilde{\beta}$ are parameters that depend on $M(z_{i})$ and cosmology. We have compared the above formula to simulation outputs and the most recent empirical mass history model from the literature (\citealt{vandenBosch14}) and found generally reasonable agreement.
Building on the work by \citet{Ludlow14}, we presented a semi-analytic model for halo concentration as a function of halo mass and redshift. The model connects the analytic model for the MAH from Paper I to halo concentration through an empirical relation between concentration and formation redshift, obtained through fits to simulation data in Paper II. The formation redshift definition of \citet{Ludlow13} and \citealt{Ludlow14}, $z_{-2}$, defined as the redshift at which the mass of the main progenitor equals the mass enclosed within the scale radius of the NFW density profile at $z=z_{i}$, results in an inner halo structure that reflects the background density of the Universe at the time when the halo formed.
The resulting $c-M$ relations were tested using $N$-body simulations, and compared to the most recent empirical $c-M$ relations from the literature (\citealt{vandenBosch14,Dutton14,Diemer14b}). The `upturn' at high masses seen by some studies (\citealt{Klypin2011,Prada,Dutton14,Diemer14b}) is not reproduced by our physically derived model which, however, only applies to relaxed halos. We analysed the $c-M$ relations obtained from the numerical simulations, where we differentiated between relaxed and unrelaxed halos. We found that the upturn is due to the inclusion of unrelaxed halos, supporting the previous claim of \citet{Ludlow12}.
We applied our model to a large range of halo mass (${\log_{10}M/\,{\rm M_{\odot}}=[-2 ,16]}$) and redshift (${z=0-20}$), and provided fits to the $c-M$ relations as well as numerical routines\footnote{Available at \href{https://bitbucket.org/astroduff/commah}{\it{https://bitbucket.org/astroduff/commah}} and \href{http://astro.physics.unimelb.edu.au/Research/Public-Data-Releases/COMMAH}{\it{http://astro.physics.unimelb.edu.au/}} in Research/Public-Data-Releases/COMMAH. See Appendix \ref{script_details} for a short description.} to compute concentrations and MAHs as a function of halo mass, redshift and cosmology. We caution the reader that baryonic processes will almost certainly increase the scatter in the $c-M$ relation and will modify the inner density profile (e.g. \citealt{Duffy10,Governato12,Teyssier13,Schaller}).
Our model predicts a change in the slope of the $c-M$ relation at $z=0-3$ and a `break' in the $z\sim 0$ $c-M$ relation at a mass of $\sim 10^{11}\,{\rm M_{\odot}}$. We analysed the evolution of concentration and found that it increases more rapidly during the dark energy era, when the accretion rates of dark matter halos decrease due to the accelerated expansion of the Universe. We found that the break at a halo mass $\sim 10^{11}\,{\rm M_{\odot}}$ results from the change of the functional form of $M(z)$, that goes from being dominated by a power-law (for low-mass halos) to an exponential (for high-mass halos). This change in $M(z)$ is driven by the rms of the linear density perturbation field. Halos with mass $M\gg 10^{11}\,{\rm M_{\odot}}$, have $\sigma\ll 1$, are characterized by an exponential growth, and have lower concentrations as a result. Whereas halos with mass $M\ll 10^{11}\,{\rm M_{\odot}}$, have large $\sigma$, are characterized by a power-law growth. In this last case, there is a pseudo-evolution in the halo masses (i.e. mass growth due to the definition of the halo in terms of an overdensity criterion, \citealt{Diemer13}) and the core radius remains approximately constant, causing the concentrations to grow. The different growth rate of the concentrations at low and high mass produces the break in the $z\sim 0$ $c-M$ relation. This break is not so evident for $z>1$, because at higher redshifts $M(z)$ is mostly exponential for all halo masses ($\tilde{\alpha}\rightarrow 0$ for $z_{i}>0$ due to the growth factor), causing all concentrations to grow at approximately the same rate (as seen in Fig.~\ref{cMH_figa}).
Finally, we addressed the impact of the $c-M$ relation presented in this work on predictions for the dark matter annihilation signal. We calculated the DM annihilation rate from cosmic structures and compared the results obtained by extrapolating the \citet{Duffy08} power-law fit (which is similar to other published fits to the results of simulations) to the rate obtained by using the $c-M$ relation predicted by our model. We found that the power from DM annihilation at $z=0$ is two orders of magnitude lower than the power obtained by extrapolating the Duffy et al. (2008) $c-M$ relation (in agreement with \citealt{Mack,Ng2014}).
\section*{Acknowledgements}
We are grateful to the OWLS team for their help with the simulations. We would also like to thank the referee, Aaron Ludlow, for many insightful comments and suggestions that greatly improved the manuscript. JSBW is supported by an Australian Research Council Laureate Fellowship. JS acknowledges support by the European Research Council under the European Union's Seventh Framework Programme (FP7/2007-2013)/ERC Grant agreement 278594-GasAroundGalaxies.
|
\section{Introduction}
\label{intro}
Let $S=S_{g}$ denote the closed orientable surface of genus $g$, let $\mbox{Mod}^*(S)$ denote the associated \textit{extended mapping class group}, and let $\mathcal{I}(S)$ be the set of isotopy classes of essential simple closed curves on $S$. There is a natural action of $\mbox{Mod}^*(S)$ on $\mathcal{I}(S)$, and analyzing orbits of various finite subsets of $\mathcal{I}(S)$ has proven to be a fruitful way of probing the algebra and geometry of $\mbox{Mod}^*(S)$. For example, the $\mbox{Mod}^*(S)$-orbits of pants decompositions have been used to estimate the Weil-Petersson diameter of the thick part of Moduli space \cite{cavendish-parlier}.
The main focus of this paper is to explicitly construct many distinct $\mbox{Mod}^*(S)$-orbits of collections of curves with intersection properties that are reminiscent of pants decompositions. We show:
\begin{customthm}{1}
\label{MainThm} For $g\ge 2$, there exist at least $2^{g-3}/(g-1)$ and at most $(4g^2+2g)!$ distinct $\mbox{Mod}^*(S)$-orbits of collections of $2g+1$ simple closed curves pairwise intersecting once.
\end{customthm}
Malestein-Rivin-Theran have shown that any collection of curves pairwise intersecting once has cardinality at most $2g+1$. Thus we think of such collections as analogous to pants decompositions in the following way: pants decompositions are the maximal cliques of the \textit{curve graph}, $\mathcal{C}(S)$, of $S$ $-$ the graph with vertex set $\mathcal{I}(S)$, and edges between two isotopy classes that can be realized disjointly on $S$. Similarly, the curve systems in Theorem \ref{MainThm} are the largest cliques of the \textit{Schaller} or \textit{systole graph}, $\mathcal{SC}(S)$, of $S$, whose vertices correspond to the subset of $\mathcal{I}$ of non-separating simple closed curves, and whose edges correspond to pairs of curves intersecting exactly once. Note that, unlike pants decompositions, there do exist non-maximum collections of curves pairwise intersecting once that are nonetheless maximal with respect to inclusion.
We remark that the number of $\mbox{Mod}^*(S)$-inequivalent pants decompositions grows at least factorially in $g$ (see \cite{bollobas} for an asymptotically precise count), and we conjecture that the same is true for the types of curve systems considered here. However, though $\mathcal{SC}(S)$ and $\mathcal{C}(S)$ are $\mbox{Mod}^*(S)$-equivariantly quasi-isometric, it is not necessarily the case that the corresponding growth rates of $\mbox{Mod}^*(S)$-inequivalent maximal cliques are comparable. This question requires a more detailed understanding of the specific nature of $\mathcal{SC}(S)$.
Schaller has shown that the automorphism group of $\mathcal{SC}(S)$ is the (extended) mapping class group \cite{schmutz-schaller}, and thus as a corollary to Theorem \ref{MainThm} we obtain:
\begin{corollary*} \label{Clique} The number of maximum cardinality cliques of $\mathcal{SC}(S)$, inequivalent under the action of $\mbox{Aut}(\mathcal{SC}(S))$, grows at least exponentially in $g$.
\end{corollary*}
By a $k$-system, we mean any subset $\Gamma \subset \mathcal{I}(S)$ consisting of curves pairwise intersecting at most $k$ times. A \textit{complete} $k$-system is a $k$-system in which any two curves intersect \textit{exactly} $k$ times. Thus the main focus of this paper is the study of complete $1$-systems of maximum possible size, or \textit{maximum complete 1-systems}.
Malestein-Rivin-Theran showed that such $1$-systems are unique up to the action of $\mbox{Mod}^*(S)$ for $g=1,2$, and they asked if this uniqueness persists for higher genera. The first author answered this question by subsequently constructing two distinct orbits of complete $1$-systems of size $2g+1$ on $S$, for all $g \geq 3$. Thus we view Theorem \ref{MainThm} as a further demonstration of the non-uniqueness of maximum complete $1$-systems.
Note that `complete'-ness, for the 1-systems we consider, is a significant simplifying assumption. Though there has been substantial recent progress towards estimating the size of maximum 1-systems \cite{przytycki}, even asymptotically precise counts are not currently available. While it would be interesting to examine the number of $\mbox{Mod}^*(S)$-orbits of maximum 1-systems, the absence of any examples when $g\ge 3$ makes this seem difficult\footnote{\cite{m-r-t} calculate that there are two $\mbox{Mod}^*(S)$-orbits of maximum 1-systems for $g=2$.}.
Our method of distinguishing $\mbox{Mod}^*(S)$-orbits of a curve system $\Gamma$ is to analyze the \textit{dual cube complex} $C(\Gamma)$ to $\Gamma$, a complex built from cubes of various dimensions which encodes the combinatorics of the intersections between curves. This invariant is a useful way of organizing topological information about the $\mbox{Mod}^*(S)$-orbit of $\Gamma$. Along the way in our analysis, we show:
\begin{customthm}{2} \label{CubeChar} Let $\Lambda_{1}, \Lambda_{2}$ be any two collections of curves which fill a closed surface $S$. Then $\Lambda_{1}$ and $\Lambda_{2}$ are equivalent under the action of the extended mapping class group if and only if there is an isomorphism of cube complexes $C(\Lambda_{1}) \cong C(\Lambda_{2})$. The induced set map from $\Lambda_1$ to $\Lambda_2$ corresponds to the induced map between hyperplanes of $C({\Lambda_1})$ and hyperplanes of $C({\Lambda_2}$).
\end{customthm}
Thus, the reader may view the main result as a construction of many non-isomorphic cube complexes, each dual to a maximum complete $1$-system. In particular, we show:
\begin{proposition*} \label{CubeReal} For any even $k \in [\lfloor g/2\rfloor,g]$, there exists a complete $1$-system of size $2g+1$ on $S_{g}$ whose dual cube complex has dimension $k$.
\end{proposition*}
It is interesting to consider whether or not the dimension of the cube complex dual to a maximum complete 1-system grows with the genus. At the moment this problem seems difficult. As a first step, one might determine whether there exists a maximum complete $1$-system $\Gamma$ whose dual cube complex is $2$-dimensional:
\begin{question} \label{2d} Does there exist a maximum complete $1$-system whose dual cube complex is $2$-dimensional?
\end{question}
In this case, the quotient of the dual cube complex $C(\Gamma)$ by the action of $\pi_{1}(S)$ produces a square-tiled copy of $S$, with at least $4$ squares around each vertex. We conjecture that the answer to Question \ref{2d} is no.
In general, $C(\Gamma)$ can be a very complicated combinatorial object; indeed, one may interpret this as a consequence of Theorem \ref{CubeChar}, since $\mbox{Mod}^*(S)$-orbits of curve systems can be difficult to distinguish (cf.~\cite{levitt-vogtmann}). In order to leverage $C(\Gamma)$ to useful information about $\mbox{Mod}^*(S)$, we make use of the following simplifying theorem for cube complexes dual to $1$-systems:
\begin{customthm}{3} \label{3-to-n} Suppose $\Gamma= \left\{\gamma_{1},...,\gamma_{n}\right\}$ is any $1$-system on an orientable surface $S$, possibly with boundary. Then the dimension of $C(\Gamma)$ is $n$ if and only if the dimension of $C(\Gamma')$ is $3$, for $\Gamma'$ any triple of curves in $\Gamma$.
\end{customthm}
\begin{remark} We note that one direction of Theorem \ref{3-to-n} is immediate: if the dimension of the entire cube complex is $n$, then any three curves must correspond to a $3$-cube in the dual cube complex. However, as Figure \ref{necessity1system} in Section \ref{qualitative info section} demonstrates, the converse is false if the assumption of being a $1$-system is dropped.
\end{remark}
Our main construction requires $g$ to be odd, and we extend the conclusion of Theorem \ref{MainThm} and Proposition \ref{CubeReal} to even $g$ via the following simple process (see \S\ref{stabilizing} for a slightly more careful description):
Beginning with a complete $1$-system $\Gamma$ of size $2g+1$ on $S_{g}$ for $g= 2k+1$, excise a pair of small open disks which are locally on opposite sides of some $\gamma \in \Gamma$, and glue on an annulus $A$ along the resulting boundary circles. Note that $\Gamma$ is still a complete $1$-system on $S_{g+1}$, and we extend $\Gamma$ to a collection of $2g+3$ curves by adding $\gamma', \gamma''$, defined as follows: both $\gamma', \gamma''$ run parallel to $\gamma$ in the complement of the new annulus $A$. Within $A$, both $\gamma', \gamma''$ run from one boundary component of $A$ to the other, intersecting once in the interior of $A$.
Therefore $\gamma', \gamma''$ intersect each other exactly once, within $A$, and each intersects all of the original elements of $\Gamma$ exactly once because $\gamma$ does. If $\Gamma$ on $S_{g}$ is obtained from a complete $1$-system $\Gamma'$ on $S_{g-1}$ as described above, we call $\Gamma$ a \textit{stabilization} of $\Gamma'$.
It is natural to ask whether or not \textit{every} complete $1$-system on $S_{g}$ of size $2g+1$ is obtained from one on $S_{g-1}$ of size $2g-1$ via this process:
\begin{question} \label{Stab} Let $\Gamma$ be a complete $1$-system on $S_{g}$ of size $2g+1$. Is it always the case that $\Gamma$ is a stabilization of some complete $1$-system on $S_{g-1}$?
\end{question}
We observe that each complete $1$-system we construct is indeed a stabilization of a complete $1$-system on a lower genus surface. Furthermore, we note that a positive answer to Question \ref{Stab} implies a negative answer to Question \ref{2d}: Lemmas \ref{tool1} and \ref{stabilized 3-cubes general 1} imply that the dimension of the cube complex dual to any complete $1$-system obtained via stabilization is at least three.
\textbf{Organization of paper.} In \S\ref{background} and \S\ref{cube cplx construction section} we outline some preliminary notions regarding the mapping class group and Sageev's construction of dual cube complexes. In \S\ref{cube complex characterizes section}, we prove Theorem \ref{CubeChar}; in \S\ref{qualitative info section}, we prove Theorem \ref{3-to-n}; in \S\ref{construction section}, we outline the main construction of our complete $1$-systems; in \S\ref{using dual cube complex section}, \S\ref{polygon section}, and \S\ref{stabilizing} we prove that these complete $1$-systems are indeed inequivalent up to the action of the extended mapping class group, completing the proof of Theorem \ref{MainThm}.
\textbf{Acknowledgements.} Both authors thank David Dumas for numerous helpful conversations and suggestions. The second author thanks as well Peter Shalen for pointing out the relevance of Sageev's invariant, and Marc Culler and Daniel Groves for their patience and time.
\section{background}
\label{background}
Let $\Gamma=\{\gamma_{1},\ldots,\gamma_{n}\}$ be a collection of free homotopy classes of closed curves on $S$. Recall that a collection of curves form a \emph{bigon} if there is an embedded disk in $S$ whose boundary is the union of two arcs of the curves. A \emph{minimal position realization} of $\Gamma$ is a set $\lambda=\{\eta_{1},\ldots,\eta_{n}\}$ such that:
\begin{enumerate}[(i)]
\item Each $\eta_{i}:S^{1}\to S$ is a smooth immersion in the free homotopy class $\gamma_{i}$.
\item The union $\bigcup_{i}\eta_{i}(S^{1})$ forms no bigons.
\item The immersed submanifolds $\{\eta_{1}(S^{1}),\ldots,\eta_{n}(S^{1})\}$ intersect only at transverse double points.
\end{enumerate}
We will refer to minimal position realizations simply as \emph{realizations}. In everything that follows, we suppose that $\lambda=\{\eta_{1},\ldots,\eta_{n}\}$ is a realization of $\Gamma$. Condition (ii) above implies that $\lambda$ minimizes the sum of the pairwise geometric intersection numbers of the curves in $\Gamma$ \cite[Prop.~1.7, p.~31]{farb-margalit}. See \cite[Ch.~1]{farb-margalit} for background on curves on surfaces.
Let $\Lambda$ indicate the set of the lifts of elements of $\lambda$ to the universal cover $\widetilde{S}$, so that the elements of $\Lambda$ are curves in $\widetilde{S}$. The union of the curves in $\Lambda$ may be considered as an embedded graph $G\subset\widetilde{S}$. Condition (iii) above guarantees that each vertex of this graph has valence four, and condition (ii) implies that every pair of curves in $\Lambda$ intersect at most once (see \cite[Lemma 1.8, p.~30]{farb-margalit}).
When the curves in $\Gamma$ are disjoint, then the dual graph to the lifts $\Lambda$ admits an isometric action of $\pi_1 S$, and the quotient graph is an invariant for $\Gamma$ that can be used to distinguish mapping class group orbits. However, in general the dual graph to $\lambda$ in $S$ is not an invariant of $\Lambda$, as different realizations may yield non-isomorphic graphs. For example, the presence of a triangle in the complement of $\lambda$ allows a Reidemeister type III move, creating a new realization but changing the isomorphism types of the dual graph, as shown in Figure \ref{reidemeister}.
\begin{figure}[h]
\begin{minipage}[]{.45\linewidth}
\centering
\vspace{.2cm}
\includegraphics[width=5cm]{reidemeister1.pdf}
\vspace{.4cm}
\end{minipage}
\hspace{.5cm}
\begin{minipage}[]{.45\linewidth}
\centering
\vspace{.2cm}
\includegraphics[width=5cm]{reidemeister2.pdf}
\vspace{.4cm}
\end{minipage}
\caption{Changing the realization $\lambda$ by a Reidemeister type III move changes the isomorphism type of the dual graph.}
\label{reidemeister}
\end{figure}
While the dual graph depends essentially on the realization $\lambda$, we now describe the construction of a related cube complex `dual' to $\Lambda$ which is independent of realization. Originally due to Sageev, this produces an isometric action of $\pi_{1}S$ on a finite-dimensional cube complex, which we denote by $\widetilde{C(\lambda)}$, with quotient $C(\Lambda):=\widetilde{C(\lambda)}/\pi_{1}S$, a cube complex with finitely many maximal cubes. Though it will be unnecessary in this work, when $\Lambda$ is filling $C(\Lambda)$ is non-positively curved, and this construction can be placed in a considerably more general context (see \cite{sageev}, \cite{sageev2}, and \cite{chatterji-niblo}).
Recall that a \emph{cube} (of dimension $n$) is the cell complex $[-1,1]^{n}$, and a \emph{cube complex} is the quotient of disjoint cubes by a gluing map which is a Euclidean isometry on each cell. In such a complex, a \emph{maximal cube} is a cube which is not contained in a higher dimensional cube. The \emph{dimension} of a cube complex is the supremum of the dimensions of its cubes. A \emph{square complex} is a two-dimensional cube complex. See \cite{sageev2} for an introduction to cube complexes.
The \emph{local hyperplanes} of an $n$-cube are the intersections of $[-1,1]^{n}$ with coordinate planes. We introduce an equivalence relation on local hyperplanes in a cube complex: Two local hyperplanes are \emph{hyperplane equivalent} if they intersect along the $1$-skeleton of the cube complex, and a \emph{hyperplane} of a cube complex is the hyperplane equivalence class of a local hyperplane. An isomorphism of cube complexes is a homeomorphism that is a cellular Euclidean isometry, which thus preserves hyperplanes. We will denote the hyperplanes of a cube complex $C$ by $\mathcal{H}_{C}$. Note that $1$-dimensional cube complexes are graphs and trees, whose hyperplanes (and local hyperplanes) are midpoints of edges.
\section{Sageev's construction}
\label{cube cplx construction section}
Fix a choice of realization $\lambda$ of $\Gamma$, with lifts $\Lambda$ as before. We now describe the construction of a cube complex $\widetilde{C(\lambda)}$ that is independent of the choice of realization. Each element $\gamma\in\Lambda$ separates $\widetilde{S}\setminus\gamma$ into two connected components. Fix a choice of identification of these two half-spaces with $\{1,-1\}$ for each $\gamma\in\Lambda$. We will refer to a choice of one of these two as a \emph{labeling} of $\gamma$, and a labeling of $\Lambda$ is a labeling for each of the curves of $\Lambda$. Identifying $2$ with $\{1,-1\}$, we say a labeling $v\in2^{\Lambda}$ is \emph{admissible} if the half-spaces $v(\alpha)$ and $v(\beta)$ intersect for every pair of curves $\alpha,\beta\in\Lambda$.
Let $\mathcal{V}_{0}\subset2^{\Lambda}$ denote the collection of admissible labelings. Note that for each connected region $U\subset\widetilde{S}\setminus\Lambda$, there is an admissible labeling $v_{\lambda}(U)\in\mathcal{V}_{0}$ defined as follows: For each curve in $\Lambda$, $v_{\lambda}(U)$ chooses the half-space containing $U$. Let $\mathcal{V}_1'$ be the graph whose vertex set is $\mathcal{V}_{0}$, where two labelings are joined by an edge when they differ on exactly one element of $\Lambda$. Hence there is an element of $\Lambda$ associated to each edge of $\mathcal{V}_1'$.
Choose a connected region $U\subset\widetilde{S}\setminus\Lambda$, and let $\mathcal{V}_1$ denote the connected component of $v_{\lambda}(U)$ in $\mathcal{V}_1'$. (The choice of connected region $U$ is evidently not essential). When lifts $\gamma_1 , \ldots , \gamma_n \in \Lambda$ pairwise intersect, it is straightforward to check that $\mathcal{V}_1$ contains an embedded copy of the 1-skeleton of $[-1,1]^n$. We obtain the cube complex $\widetilde{C(\lambda)}$ by adding in the interior of any cube whose $1$-skeleton is contained in $\mathcal{V}_{1}$; the links of the resulting cube complex are all flag simplicial complexes (see \cite{chatterji-niblo} for a more detailed exposition).
The action of $\pi_{1}S$ on $\Lambda$ induces a permutation action of $\pi_{1}S$ on $2^{\Lambda}$: For $g\in\pi_{1}S$ and admissible labeling $v\in\mathcal{V}_0$, the labeling $g\cdot v\in\mathcal{V}_0$ is given by $(g\cdot v)(\gamma)=g^{-1}\cdot v(g\cdot\gamma)$ for each $\gamma\in\Lambda$. Since the action on $\widetilde{S}$ is by deck transformations, it is straightforward to check that the action of $\pi_{1}S$ preserves the admissible set $\mathcal{V}_0$ and induces an action by graph automorphisms on the $1$-skeleton $\mathcal{V}_1$. This action extends to $\widetilde{C(\lambda)}$ by definition.
The complex $\widetilde{C(\lambda)}$ with an action of $\pi_1 S$ is independent of the realization $\lambda$: a choice of half space for a lift $\gamma\in\Lambda$ determines, and is determined by, a choice of complement of $\partial \gamma \subset S^1=\partial \pi_1 S$. By [F-M, bigons], choices of half-spaces for two lifts $\gamma,\gamma'\in\Lambda$ intersect if and only if the endpoints of $\gamma$ link with those of $\gamma'$ on $S^1$. The latter is independent of the choice of realization. We thus denote $\widetilde{C(\lambda)}$ by $\widetilde{C(\Gamma)}$ when convenient.
Moreover, it is immediate that $\widetilde{C(\Gamma)}$ is non-positively curved. A direct argument shows that any loop in $\widetilde{C(\Gamma)}$ must backtrack, which implies that the complex is also simply-connected, and thus CAT(0). We collect the relevant information about $\widetilde{C(\Gamma)}$:
\begin{theorem}[Sageev]
\label{sageev's}
{Suppose that $\Gamma$ is a filling collection of curves. Then the cube complex $\widetilde{C(\Gamma)}$ is $CAT(0)$, and the action of $\pi_{1}S$ is free, properly discontinuous, and cocompact. Given realization $\lambda$, there is a $\pi_{1}S$-equivariant incidence-preserving identification of $\mathcal{H}_{\widetilde{C(\Gamma)}}$ with the lifts $\Lambda$.}
\end{theorem}
In what follows, we do not return to the combinatorial definition of $C(\Gamma)$. We will invoke the correspondence in Theorem \ref{sageev's} often, which we may briefly refer to as the \emph{curves to hyperplanes correspondence}.
\section{Mapping class group orbits of collections of curves}
\label{cube complex characterizes section}
We now characterize a filling curve system from the isomorphism type of its dual cube complex. Recall the \emph{extended mapping class group} $\mbox{Mod}^{*}(S)=\mathrm{Diff}(S)/\mathrm{Diff}_0(S)$. We recall Theorem \ref{CubeChar}:
\begin{customthm}{2}
{Two filling curve systems $\Lambda_1$ and $\Lambda_2$ are equivalent under the action of $\mbox{Mod}^{*}(S)$ if and only if there is an isomorphism of cube complexes $C({\Lambda_1})\cong C({\Lambda_2})$. The induced set map from $\Lambda_1$ to $\Lambda_2$ corresponds to the induced map between hyperplanes of $C({\Lambda_1})$ and hyperplanes of $C({\Lambda_2}$).}
\end{customthm}
\begin{proof}
{One direction is straightforward: Suppose $\phi\cdot\Lambda_1=\Lambda_2$ for $\phi\in \mbox{Mod}^{*}(S)$. Choose a realization $\lambda_{1}$ for $\Lambda_{1}$, and a homeomorphism $\phi'$ realizing $\phi$. In this case $\phi'$ induces a $\pi_{1}S$-equivariant isomorphism $\widetilde{\phi'}:\widetilde{C(\lambda_1)}\cong\widetilde{C(\phi\cdot\lambda_1)}$, where the induced map of hyperplanes corresponds to the set map induced by $\phi'$ from $\lambda_1$ to $\phi'\cdot\lambda_1$. Since $\phi'\cdot\lambda_{1}$ is a realization of $\Lambda_{2}$, we have $C(\Lambda_{1})\cong C(\lambda_{1})$ and $C(\Lambda_{2})\cong C(\phi\cdot\lambda_{1})$. The result follows.
On the other hand, suppose $\Phi:C(\Lambda_1)\cong C(\Lambda_2)$ is an isomorphism of cube complexes. Choose realizations $\lambda_{1}$ and $\lambda_{2}$ for $\Lambda_{1}$ and $\Lambda_{2}$. By Theorem \ref{sageev's}, we have that $\widetilde{C(\Lambda_{i})}$ is simply-connected and the action of $\pi_{1}S$ is free and properly discontinuous, for each $i=1,2$. Thus $\widetilde{C(\Lambda_{i})}$ is the universal cover of $C(\Lambda_{i})$ and there are isomorphisms $\pi_{1}S\cong\pi_{1}C(\Lambda_{i})$ for each $i$. Composing these isomorphisms with $\Phi_{*}$, we find an automorphism $\phi_{*}:\pi_{1}S\to\pi_{1}S$. By construction, we may lift $\Phi$ to a $\phi_{*}$-equivariant isomorphism of cube complexes $\widetilde{\Phi}:\widetilde{C(\Lambda_{1})}\to\widetilde{C(\Lambda_{2})}$, in the sense that $$\widetilde{\Phi}(g\cdot x)=\phi_{*}(g)\cdot\widetilde{\Phi}(x)$$
for each $x\in\widetilde{C(\Lambda_{1})}$ and $g\in\pi_{1}S$. This induces a corresponding equivariant map on hyperplanes, which in turn induces a correspondence of the collections of conjugacy classes of stabilizers of hyperplanes of $\widetilde{C(\Lambda_{1})}$ with those of $\widetilde{C(\Lambda_{2})}$.
The curves to hyperplanes correspondence guarantees that there is an identification of the collection of conjugacy classes of hyperplane stabilizers of $\pi_{1}S$ acting on $\widetilde{C(\Lambda_{i})}$ with the collection of conjugacy classes of stabilizers of curves in $\widetilde{\lambda_{i}}$, and so we arrive at a correspondence of the collection of conjugacy classes of stabilizers of curves in $\widetilde{\lambda_{1}}$ with those of $\widetilde{\lambda_{2}}$. The conjugacy classes of the stabilizers of curves in $\widetilde{\lambda_{i}}$ are naturally identified with the collection of conjugacy classes of $\Lambda_{i}$ in $\pi_{1}S$, from which it follows that the automorphism $\phi_{*}$ of $\pi_{1}S$ takes the conjugacy classes determined by $\Lambda_{1}$ to those of $\Lambda_{2}$. By the Dehn-Nielsen-Baer Theorem \cite[Ch.~8, Thm.~8.1]{farb-margalit}, there is $\phi\in \mbox{Mod}^{*}(S)$ inducing $\phi_{*}$, so that $\phi\cdot\Lambda_{1}=\Lambda_{2}$.}
\end{proof}
\section{Recognizing $n$-cubes dual to a 1-system}
\label{qualitative info section}
There remains the problem of recognizing quantitative information about $C(\Lambda)$ from a given set of curves $\Lambda$. In this section, we prove Theorem \ref{3-to-n}, giving a criterion for recognizing the dimensions of cubes in the complex dual to a 1-system.
A realization of a collection of curves forms a \emph{triangle} if there is an embedded disk on $S$ whose boundary components are three arcs of the curves, and so that these arcs intersect pairwise exactly once on the boundary of the disk. A collection of homotopy classes of curves forms a triangle if there is a realization of the curves that forms a triangle.
\begin{lemma}
\label{triangles}
{If a collection of homotopy classes of closed curves forms a triangle then every realization of the curves forms a triangle.}
\end{lemma}
\begin{proof}
{By \cite[Thm.~1]{graaf-schrijver} any two realizations are homotopic through isotopies and finitely many Reidemeister type III moves. Neither of these changes the existence of a triangle in the complement.}
\end{proof}
\begin{lemma}
\label{tool1}
{If $\Gamma=\{\gamma_{1},\gamma_{2},\gamma_{3}\}$ is a realization of simple closed curves on $S$, then $\dim C(\Gamma)=3$ if and only if $S\setminus\Gamma$ has a connected component that is a triangle.}
\end{lemma}
\begin{proof}
{If $\dim C(\Gamma)=3$ and $\Gamma$ is in minimal position, then the curves to hyperplanes correspondence guarantees that there are three mutually intersecting lifts of $\Gamma$ in $\widetilde{S}$. As $\Gamma$ is a realization, the intersections of these lifts are not concurrent, and there is a triangle $T$ in their complement. Since $T$ is compact, the intersection $T\cap\widetilde{\Gamma}$ consists of finitely many arcs. As each such arc doesn't form a bigon with the boundary of the triangle, the complement of the arc has a triangular component. Thus there is a triangular component $T'\subset T\setminus\widetilde{\Gamma}$ (cf. with \emph{innermost} bigon, \cite[p.~31]{farb-margalit}).
Let $\pi:\widetilde{S}\to S$ be the covering map. If $int(T')$ does not embed under $\pi$, then there is an element $g\in\pi_{1}S$ so that $int(T')\cap g\cdot int(T')\ne\emptyset$. In this case, by the Jordan Curve Theorem there is an intersection $p\in\partial T'\cap g\cdot\partial T'$. Since $\partial T'$ consists of arcs from lifts of $\Gamma$, this violates $T'\subset T\setminus\widetilde{\Gamma}$. Thus $int(T')$ embeds in $S$, and we have found a triangle in the complement of $\Gamma$.
Conversely, suppose $\Gamma$ forms a triangle $T$. Lift this topological disk to $\widetilde{S}$, and observe that the arcs of the boundary are contained in curves that pairwise intersect. By the curves to hyperplanes correspondence, there is a $3$-cube corresponding to this collection. Using the correspondence again, a $4$-cube would yield four lifts of curves in $\Gamma$ that pairwise intersect. Since $|\Gamma|=3$, two of these lifts are in the same $\pi_{1}S$-orbit. As each of the $\gamma_{i}$ are simple, this is impossible.}
\end{proof}
Recall that a \emph{1-system} is a collection of homotopy classes of simple closed curves whose pairwise geometric intersection number is at most one. It is interesting to note that our proof of Theorem \ref{3-to-n} below makes essential use of the orientation of $S$.
\begin{customthm}{3}
{Suppose $\Gamma=\{\gamma_{1},\ldots,\gamma_{n}\}$ is a $1$-system on $S$. Then the dimension of $C(\Gamma)$ is $n$ if and only if the dimension of $C(\Gamma')$ is three, for every triple $\Gamma'\subset\Gamma$.}
\end{customthm}
One direction is straightforward: If $\dim C(\Gamma)=n$, then the correspondence of curves to hyperplanes guarantees that there is a set of $n$ lifts of the curves of $\Gamma$ which mutually intersect. In this case, every trio of these lifts mutually intersect and form a $3$-cube in the complex dual to that trio. Towards the other direction, using Lemma \ref{tool1} we may assume that each trio forms a triangle.
We define an \emph{almost-realization} of a curve system to be a minimal position realization (see \S\ref{background}) with one slight change: the words `only at transverse double points' in condition (iii) should be replaced by `transversally'. (The terms \emph{pairwise minimal position realization} would be more descriptive, but less economical). Note that \cite[Prop.~1.7, p.~31]{farb-margalit} still guarantees that almost-realizations minimize pairwise geometric intersection numbers, as with realizations.
\begin{proposition}
\label{tool2claim}
{If $\{\gamma_{1},\ldots,\gamma_{n}\}$ is a $1$-system so that every trio $\{\gamma_{i},\gamma_{j},\gamma_{k}\}$ forms a triangle, then there is an almost-realization $\{\alpha_{1},\ldots,\alpha_{n}\}$ of $\{\gamma_{1},\ldots,\gamma_{n}\}$ so that the curves $\alpha_{1},\ldots,\alpha_{n}$ have a single common intersection point.}
\end{proposition}
\begin{figure}
\centering
\includegraphics[height=6cm]{necessity1system.pdf}
\caption{Four curves for which every trio forms a triangle, but so that there is no almost-realization with a single fourfold intersection point.}
\label{necessity1system}
\end{figure}
Note that it is essential that the curves form a $1$-system, as Figure \ref{necessity1system} exhibits four curves so that every trio forms a triangle, but the conclusion of Proposition \ref{tool2claim} fails. Note as well that, if the curves form a $1$-system and every trio forms triangles, then each pair intersects exactly once, which we assume below. We first prove Theorem \ref{3-to-n}, assuming Proposition \ref{tool2claim}.
\begin{proof}[Proof of Theorem \ref{3-to-n}]
{By Proposition \ref{tool2claim}, the curves $\gamma_{1},\ldots,\gamma_{n}$ have an almost-realization as $\alpha_{1},\ldots,\alpha_{n}$, so that the $\alpha_{i}$ have a single common intersection point. Choose a lift of this point to $\widetilde{S}$, and consider the lifts of the $\alpha_{i}$ passing through this point. Since the $\alpha_{i}$ form an almost-realization, there are no bigons formed by the curves, and the chosen lifts pairwise intersect in exactly one point. This implies that in any realization of the curves $\gamma_{1},\ldots,\gamma_{n}$ these lifts will pairwise intersect. By the curves to hyperplanes correspondence, there is an $n$-cube dual to these curves.}
\end{proof}
In the proof of Proposition \ref{tool2claim}, we will need a straightforward adaptation of \cite[Prop.~1.7, p.~31]{farb-margalit}:
\begin{lemma}
\label{homotopies to only one}
{If $\{\alpha_{1},\ldots,\alpha_{n}\}$ is an almost-realization of a collection of simple closed curves, then for any other simple closed curve $\beta$, there is a curve $\beta'$, homotopic to $\beta$, so that $\{\alpha_{1},\ldots,\alpha_{n},\beta'\}$ is an almost-realization.}
\end{lemma}
\begin{proof}
{By \cite[Prop.~1.7, p.~31]{farb-margalit}, two curves are in minimal position if and only if they form no bigons. One half of this proof shows that if a bigon is formed \cite[p.~32]{farb-margalit}, one may homotope one of these curves across the bigon while the other remains fixed. Choose any realization $\beta_{0}$ of $\beta$. If $\beta_{0}$ is in minimal position with $\alpha_{1}$, then let $\beta_{1}=\beta_{0}$. If $\beta_{0}$ is not in minimal position with $\alpha_{1}$, find a bigon that $\beta_{0}$ forms with $\alpha_{1}$. Use this bigon to find a curve $\beta_{0}'$, homotopic to $\beta_{0}$, so that $|\beta_{0}'\cap\alpha_{1}|<|\beta_{0}\cap\alpha_{1}|$.
It is crucial to observe that this step may be done so that $|\beta_{0}'\cap\alpha_{j}|\le|\beta_{0}\cap\alpha_{j}|$, for each $j$: The new curve $\beta_{0}'=\beta_{+}\cup\beta_{-}$ is composed of two arcs, where $\beta_{+}$ follows $\beta_{0}$ and $\beta_{-}$ follows $\alpha_{1}$. If an arc of $\alpha_{j}$ intersects $\beta_{0}'$ in the arc $\beta_{+}$, then there is a corresponding point of intersection of $\alpha_{j}$ with $\beta_{0}$. If an arc of $\alpha_{j}$ intersects $\beta_{0}'$ in the arc $\beta_{-}$, then this arc of $\alpha_{j}$ enters the bigon formed by $\beta_{0}$ and $\alpha_{1}$ along the side contained in $\alpha_{1}$. In this case, $\alpha_{j}$ must exit the bigon through the side contained in $\beta_{0}$, since $\alpha_{1}$ and $\alpha_{j}$ form no bigons by hypothesis (see \ref{bigon1} and \ref{bigon2}). Thus there is again a point of intersection of $\beta_{0}$ with this arc of $\alpha_{j}$ that corresponds to the intersection of $\alpha_{j}$ with $\beta_{0}'$. We conclude that $|\beta_{0}'\cap\alpha_{j}|\le|\beta_{0}\cap\alpha_{j}|$, for $j\ne1$.
\begin{figure}[h]
\begin{minipage}[]{.45\linewidth}
\centering
\vspace{.2cm}
\begin{lpic}[clean]{bigon1(,1.7cm)}
\Large
\lbl[]{98,6;$\beta_{0}$}
\lbl[]{-8,14;$\beta_{0}'$}
\lbl[]{14,-8;$\alpha_{1}$}
\lbl[]{73,38;$\alpha_{j}$}
\end{lpic}
\vspace{.4cm}
\caption{A point in $\alpha_{j}\cap\beta_{0}'$ corresponds to a point in $\alpha_{j}\cap\beta_{0}$.}
\label{bigon1}
\end{minipage}
\hspace{.5cm}
\begin{minipage}[]{.45\linewidth}
\centering
\vspace{.2cm}
\begin{lpic}[clean]{bigon2(,1.5cm)}
\Large
\lbl[]{98,6;$\beta_{0}$}
\lbl[]{-8,14;$\beta_{0}'$}
\lbl[]{14,-8;$\alpha_{1}$}
\lbl[]{79,37;$\alpha_{j}$}
\end{lpic}
\vspace{.4cm}
\caption{`New' intersections of $\beta_{0}'$ with $\alpha_{j}$, violating the assumption that $\alpha_{1}$ and $\alpha_{j}$ are in minimal position.}
\label{bigon2}
\end{minipage}
\end{figure}
We apply these finitely many homotopies across bigons to $\beta_{0}$, one for each of the bigons formed by $\beta_{0}$ and $\alpha_{1}$. The result is a curve $\beta_{1}$, homotopic to $\beta$, so that $\beta_{1}$ and $\alpha_{1}$ are in minimal position, and so that $|\beta_{1}\cap\alpha_{j}|\le|\beta_{0}\cap\alpha_{j}|$, for $j\ne1$. Do this one-by-one for each $\alpha_{k}$, and the result is a curve $\beta'$, homotopic to $\beta$, so that $\beta$ is in minimal position with $\alpha_{j}$ for each $j=1,\ldots,n$. Thus $\{\alpha_{1},\ldots,\alpha_{n},\beta'\}$ is an almost-realization.}\end{proof}
\begin{proof}[Proof of Proposition \ref{tool2claim}]{We proceed by induction on $n$. For $n=3$, choose a realization of the three curves that forms a triangle. Note that since the curves form a $1$-system, there is one boundary arc of the triangle contained in each of the curves. Fixing the curves outside of a disk containing the triangle, homotope one of the curves across the triangle so that it transversally crosses the other two curves at their intersection point. This produces an almost-realization of the curves, since no bigons have been created.
Assume now that we have $n+1$ curves $\{\gamma_{1},\ldots,\gamma_{n+1}\}$ so that every trio forms a triangle. By the inductive hypothesis, there is an almost-realization $\{\alpha_{1},\ldots,\alpha_{n}\}$ of $\{\gamma_{1},\ldots,\gamma_{n}\}$ so that the $\alpha_{i}$ all have a unique common intersection point. Our strategy of proof will be to first choose a curve in the homotopy class of $\gamma_{n+1}$ that is in minimal position with the $\alpha_{i}$, and then to use the intersection properties of the curves -- namely the fact that each pair intersects exactly once, and that every trio forms triangles -- to `weave' this curve around the other curves, achieving the desired arrangement. Note that in this process the curves $\alpha_{i}$, for $i=1,\ldots,n$, remain unchanged.
Lemma \ref{homotopies to only one} implies that we may find a curve $\beta$, homotopic to $\gamma_{n+1}$, so that $\{\alpha_{1},\ldots,\alpha_{n},\beta\}$ is an almost-realization. After possibly renaming the curves $\alpha_{1},\ldots,\alpha_{n}$, we may assume that these curves are arranged around their unique common intersection in the counter-clockwise order $\alpha_{1},\ldots,\alpha_{n}$. Let $p_{0}$ indicate the common intersection point of the $\alpha_{i}$, and choose a disk neighborhood $C$ of $p_{0}$, embedded on $S$, so that $\alpha_{i}\cap C$ is connected for each $i=1,\ldots,n$.
Let $p_{i}:=\beta\cap\alpha_{i}$. Replacing $\beta$ with a homotopic curve if necessary, we may assume that all of the $p_{i}$ are contained in the arcs $\alpha_{i}\cap C$, and that $\beta\cap C$ consists only of arcs that intersect an arc $\alpha_{i}\cap C$. We now have a picture where $\beta$ weaves in and out of $C$, intersecting each of the arcs $\alpha_{i}\cap C$ precisely once. Outside of $C$ the curves $\{\alpha_{1},\ldots,\alpha_{n},\beta\}$ are disjoint. The argument that follows applies various homotopies to $\beta$, maintaining the fact that $\beta$ and $\alpha_{i}$ are in minimal position, for each $i=1,\ldots,n$. Note that as we apply such homotopies, the $p_{i}$ move as well.
In fact, for each $p_{i}$ there is a homotopy of the curve $\beta$ that slides the intersection $p_{i}$ along $\alpha_{i}$, out of $C$, until $p_{i}$ ``reappears'' on the other side of $p_{0}$ on $\alpha_{i}\cap C$. Since the curves $\alpha_{i}$ do not intersect outside of $C$, this homotopy leaves the collection $\{\alpha_{1},\ldots,\alpha_{n},\beta\}$, pairwise, in minimal position. This homotopy of $\beta$ will be exploited in the following argument, where we will refer to it as the `slide move applied to $p_{i}$'.
\begin{figure}
\centering
\begin{lpic}{greatPictureTriangles(,6cm)}
\lbl[]{0,20;$\beta$}
\end{lpic}
\caption{A `localized realization' of $\beta$: $\beta\cap C$ is connected, and $p_{i}\in C$.}
\label{localized realization}
\end{figure}
Our goal is to apply a homotopy to $\beta$ to achieve a special formation, in which intersections of $\beta$ with $\alpha_{i}$ have been `localized' to $C$: We will say a curve is a \emph{localized realization} of $\beta$ if it is homotopic to $\beta$, $\beta\cap C$ is connected, and $p_{i}\in C$, for each $i$. See Figure \ref{localized realization} for an illustrated example. The bulk of the proof of the inductive step concerns existence of a localized realization of $\beta$. We will proceed by analyzing the triangles formed by $\beta$ with $\alpha_{i}$ and $\alpha_{i+1}$, as $i$ goes from $1$ to $n-1$. At each step, we either find a localized realization of $\beta$, or we apply a homotopy to $\beta$ so that the number of connected components of $\beta\cap C$ is at most $n-i$.
Given vertices $a_{1},a_{2},a_{3}$ of the triangle $T$, we will say that $T$ `realizes $(a_{1},a_{2},a_{3})$' if the counter-clockwise orientation of $\partial T$ induces the cyclic order $(a_{1},a_{2},a_{3})$. We will refer to a sufficiently small neighborhood of $a_{i}$ as a `corner' of $T$, or `the corner at $a_{i}$' when the points are distinct, where the sufficiency is fulfilled when the intersection of $T$ with this neighborhood is connected and disjoint from the side opposite the vertex.
Suppose that two curves $\eta$ and $\delta$ in minimal position, together with a third curve, form a triangle $T$, so that $T$ has a corner at the intersection point $p\in\eta\cap\delta$. In this case, the complement of $\eta\cup\delta$ in a small enough neighborhood of $p$ consists of four components. Note that these four components correspond to the four possibilities for the placement of a corner of $T$ at $p$. (When $\eta$ and $\delta$ are simple curves it is straightforward to check that these possibilities are mutually exclusive). This is exploited repeatedly below, for the placement of corners of triangles at $p_{0}$.
By hypothesis, and by Lemma \ref{triangles}, there is a triangular component $T_{1}$ of $S\setminus\{\alpha_{1},\alpha_{2},\beta\}$. Because the curves form a $1$-system, the vertices of $T_{1}$ are necessarily given by $p_{0}$, $p_{1}$, and $p_{2}$. Of the four ways for there to be a corner formed by $\alpha_{1}$ and $\alpha_{2}$ at $p_{0}$, two of them -- namely, the choice of corner so that $T$ realizes $(p_{0},p_{2},p_{1})$ (see Figure \ref{triangle 1}) -- would achieve a localized realization: The composition, if necessary, of slide moves applied to $p_{1}$ and $p_{2}$ with a homotopy of $\beta$ across the triangle $T_{2}$ would form a localized realization. We thus assume that $T$ realizes $(p_{0},p_{1},p_{2})$, in which case, applying slide moves to $p_{1}$ and $p_{2}$ if necessary, we have the situation pictured in Figure \ref{triangle 2}. Applying further slide moves when necessary, we may now assume that the $p_{i}$, for $i>2$, all lie on the same side of $\alpha_{1}\cap C$ inside $C$, as in Figure \ref{triangle 3}.
\begin{figure}
\begin{minipage}[]{.45\linewidth}
\centering
\vspace{.5cm}
\begin{lpic}[clean]{inductiveTriangles2(,6cm)}
\Large
\lbl[]{-2,88;$C$}
\lbl[]{70,66;$p_{0}$}
\lbl[]{48,28;$p_{1}$}
\lbl[]{35,70;$p_{2}$}
\lbl[]{100,5;$\alpha_{2}$}
\lbl[]{100,110;$\alpha_{1}$}
\lbl[]{35,120;$\beta$}
\lbl[]{-8,45;$\alpha_{n}$}
\end{lpic}
\caption{The triangle $T_{1}$ realizes $(p_{0},p_{2},p_{1})$, forming a localized realization of $\beta$.}
\label{triangle 1}
\end{minipage}
\hspace{.5cm}
\begin{minipage}[]{.45\linewidth}
\centering
\begin{lpic}[clean]{inductiveTriangles1(,5.8cm)}
\Large
\lbl[]{116,88;$C$}
\lbl[]{55,70;$p_{0}$}
\lbl[]{45,28;$p_{1}$}
\lbl[]{65,28;$p_{2}$}
\lbl[]{18,112;$\alpha_{2}$}
\lbl[]{100,110;$\alpha_{1}$}
\lbl[]{5,12;$\beta$}
\lbl[]{125,68;$\alpha_{n}$}
\end{lpic}
\caption{The triangle $T_{1}$ realizes $(p_{0},p_{1},p_{2})$, where no localized realization is yet assured.}
\label{triangle 2}
\end{minipage}
\end{figure}
We now consider a triangular component $T_{2}$ of $S\setminus\{\alpha_{2},\alpha_{3},\beta\}$, with vertices $p_{0}$, $p_{2}$, and $p_{3}$. As before, there are four possible placements of the corner of $T_{2}$ at $p_{0}$. Two of them correspond to a situation in which $T_{2}$ realizes $(p_{0},p_{3},p_{2})$, again allowing a homotopy of $\beta$ that forms a localized realization. Assuming then that $T_{2}$ realizes $(p_{0},p_{2},p_{3})$, we now describe why there is only one possible placement of the corner of $T_{2}$ at $p_{0}$.
\begin{figure}
\begin{minipage}[]{.45\linewidth}
\vspace{.5cm}
\centering
\begin{lpic}[clean]{inductiveTriangles3(6cm)}
\Large
\lbl[]{-2,88;$C$}
\lbl[]{51,72;$p_{0}$}
\lbl[]{32,39;$p_{1}$}
\lbl[]{48,23;$p_{2}$}
\lbl[]{77,41;$p_{3}$}
\lbl[]{100,110;$\alpha_{1}$}
\lbl[]{55,122;$\alpha_{2}$}
\lbl[]{18,110;$\alpha_{3}$}
\lbl[]{5,12;$\beta$}
\lbl[]{-8,45;$\alpha_{n}$}
\end{lpic}
\caption{Given that the triangle $T_{1}$ realizes $(p_{0},p_{1},p_{2})$, the arcs $\beta\cap C$ may appear as pictured.}
\label{triangle 3}
\end{minipage}
\hspace{.5cm}
\begin{minipage}[]{.45\linewidth}
\vspace{.8cm}
\centering
\begin{lpic}[clean]{inductiveTriangles4(6cm)}
\Large
\lbl[]{120,90;$C$}
\lbl[]{49,80;$p_{0}$}
\lbl[]{33,42;$p_{1}$}
\lbl[]{62,35;$p_{2}$}
\lbl[]{77,47;$p_{3}$}
\lbl[]{100,115;$\alpha_{1}$}
\lbl[]{56,129;$\alpha_{2}$}
\lbl[]{15,118;$\alpha_{3}$}
\lbl[]{4,13;$\beta$}
\lbl[]{125,71;$\alpha_{n}$}
\end{lpic}
\caption{Given triangle $T_{1}$ realized as $(p_{0},p_{1},p_{2})$, the darkly shaded triangle $T_{2}$ cannot have a corner at $p_{0}$ as pictured.}
\label{triangle 4}
\end{minipage}
\end{figure}
First suppose the corner of $T_{2}$ at $p_{0}$ is placed as pictured in Figure \ref{triangle 4}. That is, suppose $T_{2}$ does not share the side between $p_{0}$ and $p_{2}$, contained in the arc $\alpha_{2}\cap C$, with $T_{1}$. Because $T_{2}$ realizes $(p_{0},p_{2},p_{3})$, the counter-clockwise orientation of $\partial T_{2}$ turns left from $\alpha_{2}$ to $\beta$ at $p_{2}$. There are two ways to make a left turn from $\alpha_{2}$ to $\beta$ at $p_{2}$. Of them, the one so that $T_{2}$ does not share the side between $p_{0}$ and $p_{2}$ with $T_{1}$ contains the arc of the curve $\beta$ that lies between the vertices $p_{2}$ and $p_{3}$ and containing $p_{1}$. Consequently, one side of the arc $\alpha_{1}\cap C$ leaves $C$ inside $T_{2}$, while the other leaves $C$ outside of $T_{2}$. Since $\alpha_{1}$ may only intersect the sides of $T_{2}$ inside $C$, this is impossible.
This leaves only one possibility for the placement of the corner of $T_{2}$ at $p_{0}$ in which $T_{2}$ realizes $(p_{0},p_{2},p_{3})$ (see Figure \ref{triangle 5}). This case allows us to apply a homotopy to $\beta$ supported in the triangle $T_{2}$, leaving the intersections $p_{i}$ inside $C$, and ensuring that the number of connected components of $\beta\cap C$ is at most $n-2$ (see Figure \ref{triangle 6}).
Similarly, at the $k$th step, the triangular component $T_{k}$ of $S\setminus\{\alpha_{k},\alpha_{k+1},\beta\}$ either provides a homotopy of $\beta$ that achieves a localized realization, or provides a homotopy of $\beta$, supported inside $C$, that ensures that the number of connected components of $\beta\cap C$ is at most $n-k$. When $k=n-1$, we have ensured the existence of a localized realization of $\beta$.
\begin{figure}
\begin{minipage}[]{.45\linewidth}
\vspace{.5cm}
\centering
\begin{lpic}[clean]{inductiveTriangles5(6cm)}
\Large
\lbl[]{-2,88;$C$}
\lbl[]{49,80;$p_{0}$}
\lbl[]{33,42;$p_{1}$}
\lbl[]{46,18;$p_{2}$}
\lbl[]{77,42;$p_{3}$}
\lbl[]{100,115;$\alpha_{1}$}
\lbl[]{56,129;$\alpha_{2}$}
\lbl[]{15,118;$\alpha_{3}$}
\lbl[]{4,13;$\beta$}
\lbl[]{-8,45;$\alpha_{n}$}
\end{lpic}
\caption{The remaining case where triangle $T_{2}$ realizes $(p_{0},p_{2},p_{3})$.}
\label{triangle 5}
\end{minipage}
\hspace{.7cm}
\begin{minipage}[]{.45\linewidth}
\vspace{.5cm}
\centering
\begin{lpic}[clean]{inductiveTriangles6(6cm)}
\Large
\lbl[]{120,90;$C$}
\lbl[]{49,80;$p_{0}$}
\lbl[]{33,42;$p_{1}$}
\lbl[]{46,18;$p_{2}$}
\lbl[]{77,42;$p_{3}$}
\lbl[]{100,115;$\alpha_{1}$}
\lbl[]{56,125;$\alpha_{2}$}
\lbl[]{15,118;$\alpha_{3}$}
\lbl[]{4,13;$\beta$}
\lbl[]{125,68;$\alpha_{n}$}
\end{lpic}
\caption{Applying a homotopy to $\beta$ using $T_{2}$, getting `closer' to a localized realization.}
\label{triangle 6}
\end{minipage}
\end{figure}
Finally, after replacing $\beta$ with its localized realization, we assume that $\beta\cap C$ is connected, and that $p_{i}\in C$ for each $i$. Note that this implies that $\beta\cap\partial C$ consists of exactly two points, which we denote $p_{-}$ and $p_{+}$. A straightforward application of the Jordan Curve Theorem ensures that $p_{-}$ and $p_{+}$ are in diametrically opposed components of $\partial C \setminus \{\alpha_1,\ldots,\alpha_n\}$, and we may apply a homotopy to $\beta$ making it into a diameter passing through $p_{0}$, completing the inductive step.}
\end{proof}
A word of caution: It is not generally true that if $\Gamma'\subset\Gamma$ then the complex $C(\Gamma')$ is a subcomplex of $C(\Gamma)$ (see Figure \ref{3 simple curves}). The following corollary is a weaker version of such a statement that will suffice for our application. Given a collection of curves $\Gamma$, we will say that a subset of $n$ curves $\Gamma'\subset\Gamma$ \emph{form an n-cube} in $C(\Gamma)$ if there are $n$ hyperplanes corresponding to the curves of $\Gamma'$ intersecting in an $n$-cube of $C(\Gamma)$.
\begin{corollary}
\label{3-to-n subset}
{If $\Gamma$ is a $1$-system of curves, and $\{\gamma_{1},\ldots,\gamma_{n}\}=\Gamma'\subset\Gamma$, then the curves of $\Gamma'$ form an $n$-cube in $C(\Gamma)$ if and only if every triple of curves from $\Gamma'$ form a $3$-cube.}
\end{corollary}
\begin{figure}[h]
\vspace{.5cm}
\begin{lpic}[clean]{3SimpleCurves(6cm)}
\Large
\lbl{-12,35;$\alpha$}
\lbl{78,70;$\beta$}
\lbl{162,35;$\gamma$}
\end{lpic}
\caption{The complex $C(\{\alpha,\gamma\})$ is not a subcomplex of $C(\{\alpha,\beta,\gamma\})$.}
\label{3 simple curves}
\end{figure}
\begin{proof}
{Using the curves to hyperplanes correspondence, the curves of $\Gamma'$ form an $n$-cube in $C(\Gamma)$ if and only if there is a choice of lifts $\{\widetilde{\gamma_{1}},\ldots,\widetilde{\gamma_{n}}\}$ so that these lifts pairwise intersect, which in turn occurs if and only if $\dim C(\Gamma')=n$, at which point we apply Theorem \ref{3-to-n}.}
\end{proof}
\section{A family of maximum complete 1-systems}
\label{construction section}
\begin{figure}[h]
\begin{lpic}[clean]{generatorsHandle(,6.5cm)}
\large
\lbl{120,55;$x$}
\lbl{80,370;$r_{1}$}
\lbl{45,210;$s_{1}$}
\end{lpic}
\caption{The generators of $\pi_{1}(S,x)$ in one handle.}
\label{generators handle1}
\end{figure}
We now construct many maximum complete 1-systems on a surface $S$ of any odd genus $g=2n+1$. Consider the genus $g$ surface as the unit sphere in $\mathbb{R}^{3}$ with $g$ handles attached at evenly spaced disks centered on an equator, making the order $g$ homeomorphism $\sigma$ that cyclically permutes the handles apparent. Let $x$ be a point fixed by this homeomorphism. Consider the presentation $$\pi_{1}(S,x)=\left\langle r_{i},s_{i} \left| \; \prod_{i=1}^{g}[r_{i},s_{i}] \right. \right\rangle,$$ where the generators $r_{1}$ and $s_{1}$ are as pictured in \ref{generators handle1}, and $r_{i}=\sigma(r_{i-1})$ and $s_{i}=\sigma(s_{i-1})$ for $i=2,\ldots,g$.
Let $\alpha_{1}$, $\beta_{1}$, and $\delta$ be given by
\begin{enumerate}
\item $\alpha_{1}=\left[r_{n+2}r_{n+3}\ldots r_{2n+1}s_{1}\right]$,
\item $\beta_{1}=\left[r_{n+2}r_{n+3}\ldots r_{2n+1}s_{1}^{-1}\displaystyle\prod_{i=1}^{n+1}[s_{i},r_{i}]\right]$, and
\item $\delta=\left[r_{1}r_{2}\ldots r_{2n+1}\right]$.
\end{enumerate}
The orbit of $\alpha_{1}$ under $\sigma$ gives $g$ curves which we denote by $\{\alpha_{1},\alpha_{3},\ldots,\alpha_{2g-1}\}$. Similarly, we denote the $\sigma$-orbit of $\beta_{1}$ by $\{\beta_{1},\beta_{3},\ldots,\beta_{2g-1}\}$. We complete these collections to sequences $\{\alpha_{i}\}_{i=1}^{2g}$ and $\{\beta_{i}\}_{i=1}^{2g}$ by defining $\alpha_{2i}=\tau_{i}(\alpha_{2i-1})$ and $\beta_{2i}=\tau_{i}^{-1}(\beta_{2i-1})$ for $i=1,\ldots,g$, and where $\tau_{i}$ is the right Dehn twist around $r_{i}$. See Figures \ref{alphaCurve}, \ref{betaCurve}, \ref{moreAlphaBeta}, and \ref{deltaCurve} for illustrative examples.
Grouping these curves together, we will refer to $A=\{\alpha_{1},\ldots,\alpha_{2g}\}$ as the set of `up' curves and $B=\{\beta_{1},\ldots,\beta_{2g}\}$ as the set of `down' curves. We will refer to the pair of up curves (resp.~down curves) $\alpha_{2i-1}$ and $\alpha_{2i}$ (resp.~$\beta_{2i-1}$ and $\beta_{2i}$) as `partners', for $i=1,\ldots,g$.
\begin{figure}[h]
\begin{minipage}[]{.48\linewidth}
\centering
\includegraphics[width=5.5cm]{alphaCurve.pdf}
\caption{The curves $\alpha_{1},\alpha_{2}\in A$}
\label{alphaCurve}
\end{minipage}
\begin{minipage}[]{.48\linewidth}
\centering
\includegraphics[width=5.5cm]{betaCurve.pdf}
\caption{The curves $\beta_{1},\beta_{2}\in B$}
\label{betaCurve}
\end{minipage}
\end{figure}
\begin{figure}[h]
\begin{minipage}[]{.48\linewidth}
\centering
\includegraphics[width=5.5cm]{moreAlphaBeta.pdf}
\caption{The curves $\alpha_{3},\alpha_{4}\in A$ and $\beta_{5},\beta_{6}\in B$}
\label{moreAlphaBeta}
\end{minipage}
\begin{minipage}[]{.48\linewidth}
\centering
\includegraphics[width=5.5cm]{deltaCurve.pdf}
\caption{The curve $\delta$}
\label{deltaCurve}
\end{minipage}
\end{figure}
It is immediate that, $\alpha_{i}\cap\alpha_{j}=\beta_{i}\cap\beta_{j}=1$ for $i\ne j$. When $\alpha_{i}$ and $\alpha_{j}$ are not partners, then $\alpha_{i}\cap\beta_{j}=1$. This calculation makes it clear that we may form many maximum complete $1$-systems: For each $i=1,\ldots,g$, choose one of the two pairs of partners from $\{\alpha_{2i-1},\alpha_{2i}\}$ and $\{\beta_{2i-1},\beta_{2i}\}$. Together with the $\delta$ curve, this forms $2g+1$ curves that pairwise intersect once. This is maximum by \cite[Thm.~1.4]{m-r-t}.
\begin{figure}[h]
\centering
\includegraphics[height=8.5cm]{Gamma1,1,1,1,1.pdf}
\caption{The system of curves $\Gamma(1,1,1,1,1)$}
\label{Gamma(1,1,1,1,1)}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[height=8.5cm]{Gamma1,-1,1,1,-1.pdf}
\caption{The system of curves $\Gamma(1,-1,1,1,-1)$}
\label{Gamma(1,-1,1,1,-1)}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[height=8.5cm]{Gamma1,1,1,1,1,1,1.pdf}
\caption{The system of curves $\Gamma(1,1,1,1,1,1,1)$}
\label{Gamma(1,1,1,1,1,1,1)}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[height=8.5cm]{Gamma1,-1,1,-1,-1,1,-1.pdf}
\caption{The system of curves $\Gamma(1,-1,1,-1,-1,1,-1)$}
\label{Gamma(1,-1,1,-1,-1,1,-1)}
\end{figure}
More precisely, for $\epsilon\in\{1,-1\}^{g}$, let $A(\epsilon)=\{\alpha_{2i-1},\alpha_{2i}|\epsilon_{i}=1\}$ and $B(\epsilon)=\{\beta_{2i-1},\beta_{2i}|\epsilon_{i}=-1\}$, and let $\Gamma(\epsilon)=\{\delta\}\cup A(\epsilon)\cup B(\epsilon)$. Several examples in genus 5 and 7 are shown in Figures \ref{Gamma(1,1,1,1,1)}, \ref{Gamma(1,-1,1,1,-1)}, \ref{Gamma(1,1,1,1,1,1,1)}, and \ref{Gamma(1,-1,1,-1,-1,1,-1)}.
\begin{lemma}
\label{examples mc1s}
{For each $\epsilon\in\{-1,1\}^{g}$, the collection of curves $\Gamma(\epsilon)$ forms a maximum complete 1-system.}
\end{lemma}
There is an action of the dihedral group $D_{g}$ on $\{1,-1\}^{g}$ given by letting the generators act by a $g$-cycle and a reversal of the list, respectively. Letting $\mathbb{Z}/2\mathbb{Z}$ act by taking $\epsilon$ to $-\epsilon$, we obtain an action of $\mathbb{Z}/2\mathbb{Z}\oplus D_{g}$ on $\{1,-1\}^{g}$, and there is naturally an action of $\mbox{Mod}^{*}(S)$ on systems of conjugacy classes of curves on $S$. The following proposition, whose proof occupies the bulk of our analysis in \S\ref{using dual cube complex section} and \S\ref{polygon section}, implies that the maximum complete 1-systems from Lemma \ref{examples mc1s} represent many distinct orbits.
\begin{proposition}
\label{distinguishing orbits}
{If $\Gamma(\epsilon)$ and $\Gamma(\epsilon')$ are in the same $\mbox{Mod}^{*}(S)$-orbit, then $\epsilon$ and $\epsilon'$ are in the same $(\mathbb{Z}/2\mathbb{Z}\oplus D_{g})$-orbit in $\{1,-1\}^{g}$.}
\end{proposition}
We will also require the simple observation:
\begin{lemma}
\label{mc1s filling}
{A maximum complete $1$-system is filling.}
\end{lemma}
\begin{proof}
{Suppose not. Then there is a simple closed curve $\alpha$ disjoint from the curves in our $1$-system. Cut open $S$ along $\alpha$, and cap off the two resulting boundary components created with disks. Note that the resulting surface may be disconnected. In any case, the set of $2g+1$ curves obtained forms a maximum complete $1$-system of curves on a surface of genus $g'\le g-1$, contradicting \cite[Thm.~1.4]{m-r-t}.}
\end{proof}
In fact, one can show that any $2g$ curves from a maximum complete 1-system are filling, but we will not require this stronger statement.
Let $N(g)$ indicate the number of $\mbox{Mod}^{*}(S)$-orbits among maximum complete 1-systems. A simple argument involving the square complex dual to a realization of curves on $S$ provides an upper bound for $N(g)$ below. For $g$ odd, Proposition \ref{distinguishing orbits} allows using the $\Gamma(\epsilon)$ to provide a lower bound for $N(g)$. For $g$ even, we will use the stabilization procedure described in detail in \S\ref{stabilizing} to obtain lower bounds for $N(g)$. Given a choice of realization for a maximum complete 1-system $\lambda$, and an arc $\alpha$ intersecting each of the curves in $\lambda$ once, stabilization produces a maximum complete 1-system on a surface of genus $g+1$.
When necessary below, we identify $\Gamma(\epsilon)$ with fixed choices of realization for each such collection. In \S\ref{stabilizing} we prove:
\begin{proposition}
\label{stabilizing Gammas}
{There is a choice of arc $\alpha$ so that the stabilizations of $\Gamma(\epsilon)$ and $\Gamma(\epsilon')$ along $\alpha$ are in the same $\mbox{Mod}^*(S_{g+1})$-orbit if and only if $\Gamma(\epsilon)$ and $\Gamma(\epsilon')$ are in the same $\mbox{Mod}^*(S_{g})$-orbit.}
\end{proposition}
We restate and prove Theorem \ref{MainThm}:
\begin{customthm}{1}
{We have the bounds $$(4g^{2}+2g)! \ge N(g) \ge \frac{2^{g-1}}{4(g-1)}.$$}
\end{customthm}
\begin{proof}
{The lower bound follows from Proposition \ref{distinguishing orbits} when $g$ is odd, and from Proposition \ref{stabilizing Gammas} when $g$ is even.
Towards the upper bound, consider the set of isomorphism classes of square complexes $\mathcal{S}_{\lambda}$ that are dual to realizations $\lambda$ of maximum complete $1$-systems. For filling systems of curves, the dual square complex is isomorphic to the surface $S$, and the hyperplanes of the square complex are in the homotopy classes of the curves one started with. Lemma \ref{mc1s filling} now guarantees that an isomorphism of square complexes $\mathcal{S}_{\lambda}\cong\mathcal{S}_{\lambda'}$ yields a homeomorphism of $S$ taking $\lambda$ to $\lambda'$. Thus there is a well-defined map from the set of isomorphism classes of square complexes dual to maximum complete $1$-systems to the set of $\mbox{Mod}^{*}(S)$-orbits of maximum complete $1$-systems. This map is evidently surjective, so that an upper bound for the number of possible square complexes dual to a realization of a maximum complete 1-system produces an upper bound for $N(g)$.
For each realization $\lambda$ of a maximum complete $1$-system, each of the curves in $\lambda$ passes through exactly $2g$ squares of $\mathcal{S}_{\lambda}$. We may thus view $\mathcal{S}_{\lambda}$ as the quotient of the disjoint union of $2g+1$ annuli, each of which is built from $2g$ squares, where the quotient map identifies squares in pairs. There are at most $\binom{2g(2g+1)}{2,\ldots,2}$ pairings, and each pair of matched squares has two possible identifications, giving at most
$$2^{g(2g+1)}\binom{2g(2g+1)}{2,\ldots,2}=2^{g(2g+1)}\cdot\frac{\left(2g(2g+1)\right)!}{2^{g(2g+1)}}=\left(2g(2g+1)\right)!$$
square complexes $\mathcal{S}_{\lambda}$.}
\end{proof}
\section{Restricting mapping class group orbits via $C(\Gamma(\epsilon))$}
\label{using dual cube complex section}
This section is the first step towards the proof of Proposition \ref{distinguishing orbits}.
\begin{proposition}
\label{MCG-orbit restrictions}
{If $\phi\in \mbox{Mod}^{*}(S)$ satisfies $\phi\cdot\Gamma(\epsilon)=\Gamma(\epsilon')$, then
\begin{enumerate}
\item $\phi \cdot \delta =\delta$, and
\item $\phi\cdot \left\{A(\epsilon),B(\epsilon)\right\}=\left\{A(\epsilon'),B(\epsilon')\right\}$.
\end{enumerate}
Moreover, the map $\phi$ sends partner curves to partner curves.}
\end{proposition}
In other words, either $\phi$ preserves the sets of up and down curves, or it exchanges them. The proof of this proposition will follow from a coarse picture of the cube complex $C(\Gamma(\epsilon))$.
\begin{lemma}
\label{3cubes}
{In the complex $C(A\cup B\cup\{\delta\})$, the triples that form $3$-cubes are the following:
\begin{enumerate}
\item $\{\alpha_{i},\alpha_{j},\alpha_{k}\}$ or $\{\beta_{i},\beta_{j},\beta_{k}\}$, for distinct $i,j,k\in\{1,\ldots,2g\}$.
\item $\{\alpha_{2i-1},\alpha_{2i},\delta\}$ or $\{\beta_{2i-1},\beta_{2i},\delta\}$, for $i\in\{1,\ldots,g\}$.
\item $\{\alpha_{2j-1},\alpha_{2j},\beta_{i}\}$ or $\{\beta_{2j-1},\beta_{2j},\alpha_{i}\}$, for $i\in\{1,\ldots,2g\}$ and $j\in\{1,\ldots,g\}$.
\item $\{\alpha_{i},\beta_{j},\delta\}$ for $i,j\in\{1,\ldots,2g\}$.
\end{enumerate}
}
\end{lemma}
\begin{proof}
{Using Lemmas \ref{triangles} and \ref{tool1}, we determine whether a triple of curves forms a $3$-cube by choosing a realization of the curves, and observing whether there is a triangular component of the complement. For each of the curves, we fix choices of realizations as in Figures \ref{alphaCurve}, \ref{betaCurve}, and \ref{deltaCurve}.
If $\delta$ is one of the three curves, we arrange the possible ways to choose the other two curves according to whether the curves are chosen as `up' or `down' (i.e.~from $A$ or $B$): If both of the other curves are up, then there is a triangle in the complement of the trio if and only if the other two curves were partners. If one of the curves is up and one is down, there is such a triangle. The other cases are similar.
On the other hand, if $\delta$ is not one of the three curves: If all of the curves are up, there is such a triangle. If two of the curves are up and one is down, there is a triangle in their complement if and only if the two up curves are partners. The other cases are similar.}
\end{proof}
We proceed with an examination of hyperplanes of maximal cubes in the cases for $\epsilon\in\{1,-1\}^{g}$ where $|A(\epsilon)|,|B(\epsilon)|>1$.
\begin{lemma}
\label{max cubes}
{When $|A(\epsilon)|,|B(\epsilon)|>1$, the sets of hyperplanes of maximal cubes of $C(\Gamma(\epsilon))$ correspond to one of the following lists of curves:
\begin{enumerate}
\item The $2|\epsilon^{-1}(1)|$ curves $A(\epsilon)$.
\item The $2|\epsilon^{-1}(-1)|$ curves $B(\epsilon)$.
\item The 5 curves $\{\alpha_{2i-1},\alpha_{2i},\beta_{2j-1},\beta_{2j},\delta\}$, for $i,j\in\{1,\ldots,g\}$ such that $\alpha_{2i}\in A(\epsilon)$ and $\beta_{2j}\in B(\epsilon)$.
\end{enumerate}}
\end{lemma}
\begin{proof}
{Using Corollary \ref{3-to-n subset}, in order to check whether a subset of curves from $\Gamma(\epsilon)$ forms an $n$-cube, it is enough to check whether every triple forms a $3$-cube. By Lemma \ref{3cubes} we have a complete list of such 3-cubes.
The curves $A(\epsilon)$ and $B(\epsilon)$ form cubes of dimensions $2|\epsilon^{-1}(1)|$ and $2|\epsilon^{-1}(-1)|$, respectively, by Lemma \ref{3cubes} and Corollary \ref{3-to-n subset}. If one adds a down curve $\beta_{i}$ to $A(\epsilon)$, then a pair of up curves that are not partners will not form a $3$-cube with this down curve $\beta_{i}$, by Lemma \ref{3cubes}. If one adds $\delta$ to $A(\epsilon)$, then again a pair of up curves that are not partners will not form a $3$-cube with $\delta$. The analogous statements hold for $B(\epsilon)$. By Corollary \ref{3-to-n subset}, the cubes of dimension $2|\epsilon^{-1}(1)|$ and $2|\epsilon^{-1}(-1)|$ containing these sets of hyperplanes, respectively, must be maximal. The same analysis shows that a maximal cube containing $\delta$ must contain a pair of partner up curves and a pair of partner down curves.}
\end{proof}
The cases in which either of $|A(\epsilon)|$ or $|B(\epsilon)|$ are less than or equal to $1$ are quite similar, so we list the relevant information without proof.
\begin{lemma}
\label{max cubes 2}
{When $|B(\epsilon)|=0$ (resp.~$|A(\epsilon)|=0$), the sets of hyperplanes of maximal cubes of $C(\Gamma(\epsilon))$ correspond to one of the following lists of curves:
\begin{enumerate}
\item The $2g$ curves $A(\epsilon)$ (resp.~$B(\epsilon)$).
\item The 3 curves $\{\alpha_{2i-1},\alpha_{2i},\delta\}$ (resp.~$\{\beta_{2i-1},\beta_{2i},\delta\}$), for $i\in\{1,\ldots,g\}$.
\end{enumerate}
When $|B(\epsilon)|=1$ (resp.~$|A(\epsilon)|=1$), the sets of hyperplanes of maximal cubes of $C(\Gamma(\epsilon))$ correspond to one of the following lists of curves:
\begin{enumerate}
\item The $2g-2$ curves $A(\epsilon)$ (resp.~$B(\epsilon)$).
\item The 5 curves $\{\alpha_{2i-1},\alpha_{2i},\beta_{2j-1},\beta_{2j},\delta\}$, for $i,j\in\{1,\ldots,g\}$ such that $\alpha_{2i}\in A(\epsilon)$ and $\beta_{2j}\in B(\epsilon)$.\hfill \qedsymbol
\end{enumerate}}
\end{lemma}
\begin{proof}[Proof of Proposition \ref{MCG-orbit restrictions}]
{The simple observation we exploit is that cube complex isomorphisms must send maximal cubes to maximal cubes.
Suppose $|A(\epsilon)|,|B(\epsilon)|>1$. By Lemma \ref{max cubes} and Lemma \ref{max cubes 2}, the maximal cubes that the hyperplane corresponding to $\delta$ passes through are all $5$-dimensional, while the maximal cubes that the hyperplane corresponding to $\alpha_{i}$ (resp.~$\beta_{i}$) passes through, for any $i\in\{1,\ldots,2g\}$, include one of the two even-dimensional maximal cubes corresponding to $A(\epsilon)$ and $B(\epsilon)$. Thus any isomorphism of cube complexes $\Phi:C(\Gamma(\epsilon))\cong C(\Gamma(\epsilon'))$ must take the hyperplane corresponding to $\delta$ to itself. By Theorem \ref{CubeChar}, the corresponding mapping class $\phi$ fixes $\delta$. Similarly, the even-dimensional maximal cubes whose hyperplanes correspond to $A(\epsilon)$ and $B(\epsilon)$ must be sent to the pair of even-dimensional maximal cubes whose hyperplanes correspond to $A(\epsilon')$ and $B(\epsilon')$. The remaining cases are similar.
Finally, a pair of partner curves are simultaneously up or down. By Lemma \ref{3cubes}, they form a 3-cube with $\delta$ while a pair of non-partner curves that are simultaneously up or down do not. It follows that $\phi$ sends partner curves to partner curves.}\end{proof}
While we may conclude that the pair of numbers $|A(\epsilon)|$ and $|B(\epsilon)|$ is equal to $|A(\epsilon')|$ and $|B(\epsilon')|$ if $\Gamma(\epsilon)$ and $\Gamma(\epsilon')$ are $\mbox{Mod}^{*}(S)$-equivalent, we are not yet able to prove Proposition \ref{distinguishing orbits}. We turn to finer invariants of $\Gamma(\epsilon)$.
\section{The labeled polygon $P(\epsilon)$ associated to $\Gamma(\epsilon)$}
\label{polygon section}
Towards the proof of Proposition \ref{distinguishing orbits}, we introduce a more detailed invariant. The information inherent to the invariant we produce is easily packaged as a polygon. The essential tool to building this invariant is the ordering induced on intersection points of a realization of an oriented curve.
However, while an oriented curve in a realization of a curve system determines an ordering of its intersection points, this ordering may not be an invariant of the collection of homotopy classes; the presence of 3-cubes implies the existence of a Reidemeister move that will make it possible to switch the ordering of intersection points. We state the following only for complete 1-systems for ease in exposition, but with slightly more detail a more general statement could be made.
\begin{lemma}
\label{ordering sans 3cubes}
{Let $\{\gamma_1,\ldots,\gamma_k,\gamma,\gamma'\}$ be a complete 1-system of curves, and let $\vv{\gamma}$ be a choice of orientation of $\gamma$. Suppose that $\gamma$ does not form 3-cubes with any pair $\gamma_i$ and $\gamma_j$. Then the cyclic ordering of $\gamma_1,\ldots,\gamma_k$ induced by $\vv{\gamma}$ is invariant of the choice of realization. In this setting, the choice of $\gamma_j$ induces a well-defined ordering of $\{\gamma_i : i\ne j\}$. Moreover, if $\gamma$ and $\gamma'$ form 3-cubes with $\gamma_i$, for all $i$, then the cyclic orderings of $\gamma_1,\ldots,\gamma_k$ induced by the two orientations of $\gamma'$ coincide with those of the two orientations of $\gamma$.}
\end{lemma}
\begin{proof}
{Choose a realization of the curve system $\{\gamma_1,\ldots,\gamma_k,\gamma,\gamma'\}$. One obtains a cyclic ordering of $\{\gamma_1,\ldots,\gamma_k\}$ induced by $\vv{\gamma}$. This ordering is invariant of the chosen realization, since any realization can be obtained from any other realization by applying a sequence of Reidemeister moves, and births or deaths of monogons or bigons \cite[Lemma 5.6]{goldman2}.
Finally, if $\gamma$ and $\gamma'$ form 3-cubes with each of $\gamma_i$ and $\gamma_j$, but the orderings determined by the orientations of $\gamma$ and $\gamma'$ differ on the pair $\{\gamma_i,\gamma_j\}$, then it follows that there would be a triangle formed by $\gamma$, $\gamma_i$, and $\gamma_j$.}
\end{proof}
We refer to maximal sets of non-partner up (resp.~down) curves as \emph{full}, and we fix choices of full sets of up and down curves $U$ and $D$, respectively. In our setting, by Lemma \ref{3cubes} the curve $\alpha_i$ does not form 3-cubes with any pair of non-partner down curves. Thus, choosing an orientation $\vv{\alpha_i}$, we may apply Lemma \ref{ordering sans 3cubes} to $\vv{\alpha_i}$ with $D$. We conclude that a choice of orientation $\vv{\alpha_i}$ induces a cyclic order to any full set of down curves.
Moreover, as long as $|A(\epsilon)|>2$, the cyclic order of $D$ induced by $\vv{\alpha_i}$ can be upgraded, canonically, into a bona fide ordering: If $|A(\epsilon)|>2$, there is an up curve $\alpha_j$ which is not a partner of $\alpha_i$. By Lemma \ref{3cubes}, the curve $\alpha_i$ does not form a 3-cube with $\alpha_j$ and $\beta_l$, for any $l$, so that Lemma \ref{ordering sans 3cubes} applies to $\vv{\alpha_i}$ and the union of $D$ with $\alpha_j$. Moreover, it is evident that this order does not depend on the choice of $\alpha_j$ among up curves that are not partners of $\alpha_i$. We conclude that a choice of orientation $\vv{\alpha_i}$ induces an ordering of $D$, which we refer to as the $\vv{\alpha_i}$-ordering of $D$.
We choose an almost realization of $\Gamma(\epsilon)$ such that the curves in $U$ intersect at a single point $p$, and curves in $D$ intersect at a single point $p'$. Let $c$ be a small circle centered at $p$, and $c'$ a small circle centered at $p'$. An orientation of $c$ (resp. $c'$) induces a cyclic ordering on the finite set of points $U \cap c$ (resp. $D \cap c')$. Assuming that both $U$ and $D$ are non-empty, using the orientation of the surface $S$, we equip $c$ with a clockwise orientation, and $c'$ with a counter-clockwise orientation. This induces a cyclic ordering on the finite collection of points $U \cap c$ and on $D \cap c'$.
\begin{remark} We note the cyclic ordering on $U \cap c$ (and on $D \cap c'$) described above is an invariant of the $\mbox{Mod}^*(S)$-orbit of $\Gamma(\epsilon)$. This follows from the fact that it corresponds to the cyclic order on the set of endpoints of lifts to $\mathbb{H}^{2}$ of geodesics in $U$ on some hyperbolic surface, induced from an orientation of $\partial \mathbb{H}^{2}$, and therefore this ordering is detected by the dual cube complex to $\Gamma(\epsilon)$. Thus, the definition of this cyclic ordering on $U \cap c$ does not depend on our particular choice of almost realization for $\Gamma(\epsilon)$.
\end{remark}
Given a choice of orientation of a curve $\gamma\in U$ (resp.~$D$), the two points of $\gamma\cap c$ (resp.~$\gamma\cap c'$) are each equipped with an arrow that either points away from, or towards $p$ (resp.~$p'$). We refer to arrows pointing towards $p$ (resp.~$p'$) as \emph{inward pointing} and the others as \emph{outward pointing}. Given a set of choices of orientations for each of the curves in $U$, the set of inward and outward pointing arrows partitions the set of points $U\cap c$ into two sets. Moreover, there is an involution $\iota$ of $U \cap c$ that exchanges these two sets, for any set of choices of orientations for the curves in $U$: Given $v\in U \cap c$, $\iota(v)$ is the only other point of $U \cap c$ on the same curve of $\Gamma(\epsilon)$ as $v$.
\begin{lemma} \label{Cohere} There exists a choice of orientations $\{\vv{\alpha_i}\}$ for the curves in $U$ such that the $\vv{\alpha_i}$-ordering and the $\vv{\alpha_j}$-ordering on $D$ are cyclically equivalent.
\end{lemma}
\begin{proof} Choose orientations for $\alpha_i$ and $\alpha_j$ as in Figure \ref{orderingPic}. The down curves $D$ are partitioned by this choice into four sets $D_1,D_2,D_3,D_4$, and with the chosen orientations on $\alpha_i$ and $\alpha_j$, the $\vv{\alpha_i}$- and the $\vv{\alpha_j}$-orderings of $D$ are cyclically equivalent.
\end{proof}
\begin{figure}[]
\centering
\Large
\begin{lpic}{prop75drawing(10cm)}
\lbl[]{126,128;$\vv{u_1}$}
\lbl[]{160,100;$\vv{u_2}$}
\lbl[]{151,136;$D_2$}
\lbl[]{78,125;$D_3$}
\lbl[]{82,66;$D_4$}
\lbl[]{143,62;$D_1$}
\end{lpic}
\caption{The oriented up curve $u_1$ induces the ordering $(D_1,D_2,D_3,D_4)$ of the down curves, while $u_2$ induces $(D_4,D_1,D_2,D_3)$.}
\label{orderingPic}
\end{figure}
A set of choices of orientations of the curves in $U$ is \textit{coherent} if it satisfies the conclusion of Lemma \ref{Cohere}. The orientations chosen in Lemma \ref{Cohere} (see Figure \ref{orderingPic}) are a convenient choice, and we refer to these choices of orientations of up curves as the \emph{standard orientations} for curves in $U$. It will be useful to have a chosen orientation of the down curves as well. Note the orientation-reversing involution of the surface that exchanges each up curve $\alpha_i$ with the down curve $\beta_i$. An orientation of a down curve is \emph{standard} if the image under this involution is a standardly oriented up curve.
\begin{remark} \label{CoherentUnique} We note that there are exactly two coherent orientations on $U$; this follows from the fact that a coherent orientation is completely determined by choosing the orientation on one curve of $U$. We will refer to the coherent set of choices of orientations for $U$ that is not the standard one as \emph{non-standard}.
\end{remark}
\begin{definition} The \textit{labeled polygon} $P(\epsilon)$ associated to $\Gamma(\epsilon)$ is a $2|U|$-gon, satisfying:
\begin{enumerate}
\item The vertices are labeled from elements of $U \cap c$, in the cyclic ordering induced by the orientation of $c$; thus $P(\epsilon)$ comes equipped with a preferred orientation of its boundary. By placing an outward pointing arrow at each point of $U\cap c$, the vertices of $P(\epsilon)$ determine orderings of $D$. By Remark \ref{CoherentUnique}, these orderings partition the vertices into two sets $P_1$ and $P_2$, exchanged by $\iota$, that correspond to the two cyclic equivalence classes of orderings of $D$ determined by possible coherent choices of orientations for $U$. For each $l=1,2$, we let $R_l(\epsilon)$ denote the polygon formed by the cyclically ordered vertices in $P_l$. Note that the edges of $R_l(\epsilon)$ are diagonals of $P(\epsilon)$;
\item Each edge of $R_l(\epsilon)$ is decorated with a pair of integers $\mathcal{M}_l(e)$and $\mathcal{N}_l(e)$, defined as follows:
Let $e= (\vv{\alpha_{i}}, \vv{\alpha_{j}})$ be an edge of $R_l(\epsilon)$ with initial and terminal vertices corresponding to the oriented curves $\vv{\alpha_{i}}$ and $\vv{\alpha_{j}}$ in $U$, respectively. We set $\mathcal{M}_l(e)$ equal to the number of vertices of $P(\epsilon)$ between $\vv{\alpha_{i}}$ and $\vv{\alpha_{j}}$. If $\beta_{l_{1}},\ldots,\beta_{l_{m}}$ denotes the $\vv{\alpha_{i}}$-ordering of $D$, then by construction there exists some $r$ such that the $\vv{\alpha_{j}}$-ordering on $D$ is $$\beta_{l_{m-r+1}}, \ldots, \beta_{l_m},\beta_{l_{1}},\ldots,\beta_{l_{m-r}}.$$
We define $\mathcal{N}_l(e)= r$.
\end{enumerate}
\end{definition}
\begin{lemma} \label{WellDef} $P(\epsilon)$ is a well-defined invariant of $\Gamma(\epsilon)$.
\end{lemma}
\begin{proof} By Remark \ref{CoherentUnique}, the standard and non-standard pair of choices of orientations for the curves in $U$ are the only coherent such choices. Thus the partition of the vertices of $P(\epsilon)$ into the two sets $P_1$ and $P_2$ is well-defined. The labels $\mathcal{M}_l(e)$ and $\mathcal{N}_l(e)$ are evidently invariant under a different set of choices of full sets $U$ and $D$.
\end{proof}
\begin{lemma} \label{sumsN} For each of the labeled polygons $P(\epsilon)$ we have
\begin{align*}
\sum_e \mathcal{N}_{l_1}(e) & = |D|, \text{ and} \\
\sum_e \mathcal{N}_{l_2}(e) & = |D| \cdot (|U| - 1),
\end{align*}
where $\{l_1,l_2\} = \{1,2\}$.
\end{lemma}
\begin{proof}
One of the polygons has vertices that determine the standard orientations of the curves in $U$, while the other has vertices that determine the non-standard ones. By examining Figure \ref{orderingPic}, one sees that the sum for the polygon with standard orientations determined at its vertices is the first sum above, while the sum for the other polygon is the second sum above.
\end{proof}
After relabeling, we assume from now on that $R_1(\epsilon)$ refers to the polygon whose vertices are points of $U\cap c$ with outward pointing arrows in the standard orientations of the curves in $U$. See Figure \ref{labeledPolygons} for examples corresponding to $\epsilon=(1,1,1-1,-1,1,-1)$ and $\epsilon=(-1,1,1,-1,-1,1,1)$.
\begin{figure}[h]
\begin{minipage}[]{\linewidth}
\centering
\begin{lpic}{labelledPolygon1(13cm)}
\lbl[]{296.5,150;$(1,1)$}
\lbl[]{367.5,158.5;$(0,1)$}
\lbl[]{397.5,121;$(1,0)$}
\lbl[]{347,78;$(2,1)$}
\end{lpic}
\caption*{$\epsilon=(1,1,1-1,-1,1,-1)$}
\end{minipage}
\vspace{.3cm}\\
\begin{minipage}[]{\linewidth}
\centering
\begin{lpic}{labelledPolygon2(13cm)}
\lbl[]{292,125;$(1,0)$}
\lbl[]{340,168;$(0,2)$}
\lbl[]{395,121;$(1,0)$}
\lbl[]{340,74;$(2,1)$}
\end{lpic}
\caption*{$\epsilon=(-1,1,1,-1,-1,1,1)$}
\end{minipage}
\caption{Full sets of curves from the curve system $\Gamma(\epsilon)$, the `small circle' $c$, the polygon $P(\epsilon)$, and the polygon $R_1(\epsilon)$. Edge labels are written as the ordered pair $(\mathcal{M}_1(e),\mathcal{N}_1(e))$.}
\label{labeledPolygons}
\end{figure}
\begin{definition} An \textit{isomorphism of labeled polygons} from $P(\epsilon)$ to $P(\epsilon')$ is a permutation $\psi \in S_{2|U|}$, where $S_{n}$ denotes the symmetric group on $n$ symbols, that induces a label-preserving isomorphism between the labeled $1$-skeletons of $P(\epsilon) \cup R_1(\epsilon) \cup R_2(\epsilon)$ and $P(\epsilon')\cup R_1(\epsilon') \cup R_2(\epsilon')$.
\end{definition}
By Proposition \ref{MCG-orbit restrictions}, a homeomorphism of the surface that takes $\Gamma(\epsilon)$ to $\Gamma(\epsilon')$ takes $A(\epsilon)$ to either $A(\epsilon')$ or $B(\epsilon')$. In the first case, by construction, the homeomorphism induces an isomorphism of labeled polygons $P(\epsilon)\cong P(\epsilon')$. Note that it is a consequence of Lemma \ref{sumsN} that we may conclude that, if $\psi: P(\epsilon)\cong P(\epsilon')$ is an isomorphism of labeled polygons, then $\psi$ takes $R_1(\epsilon)$ to $R_1(\epsilon')$ and $R_2(\epsilon)$ to $R_2(\epsilon')$.
\begin{remark}
\label{preservesR1}
As a consequence, any homeomorphism taking $\Gamma(\epsilon)$ to $\Gamma(\epsilon')$ that takes $A(\epsilon)$ to $A(\epsilon')$ must preserve the standard orientations of curves.
\end{remark}
\begin{proposition} \label{Iso} If $\psi: P(\epsilon) \cong P(\epsilon')$ is an isomorphism of labeled polygons, then $\epsilon$ and $\epsilon'$ are in the same $(\mathbb{Z}/2\mathbb{Z} \oplus D_{g})$-orbit of $\left\{-1,1\right\}^{g}$.
\end{proposition}
\begin{proof}
Our strategy below is to make various choices for labels and orientations of curves that make the structure of $P(\epsilon)$ more transparent--by Lemma \ref{WellDef} these choices are allowed. With the structure of $P(\epsilon)$ clear, the result will follow easily. Roughly speaking, the constructions below are careful treatments of `how the curves in $\Gamma(\epsilon)$ look,' when they are drawn conveniently on the surface. In what follows, we fix the almost realization described at the beginning of this section, and fix as well the standard orientations of each of the curves.
We first establish a cyclic ordering on all the curves in $U \cup D$. Consider the collection of curves $\omega_{1},\ldots,\omega_{g}$ in Figure \ref{transversals}, realized in minimal position with $\Gamma(\epsilon)$. Note that for each $i$, the curve $\omega_{i}$ intersects exactly one curve in $U \cup D$. Given $\gamma \in U \cup D$, the \textit{transversal} to $\gamma$ is the unique curve in $\left\{\omega_{1},\ldots, \omega_{g} \right\}$ intersecting $\gamma$. Thus there is a cyclic ordering on $U \cup D$ associated to the cyclic ordering $(1,\ldots,g)$ on the indices of the $\omega_{i}$ curves.
\begin{figure}[]
\centering
\Large
\begin{lpic}{OmegaCurves(3.5in)}
\lbl[]{105,135;$\omega_1$}
\lbl[]{130,125;$\omega_2$}
\lbl[]{120,60;$\omega_{\frac{g-1}{2}}$}
\lbl[]{85,60;$\omega_{\frac{g+1}{2}}$}
\lbl[]{75,125;$\omega_g$}
\end{lpic}
\caption{The transversals $\omega_1,\ldots,\omega_g$.}
\label{transversals}
\end{figure}
We will now describe a cyclic order on $(U \cap c) \sqcup (D \cap c')$. For $\gamma \in U \cup D$, suppose $\gamma \in U$, and let $\omega$ denote the transversal to $\gamma$. Then $c$ separates $\gamma$ into two sub-arcs, both bounded by the two points $s_{1}(\gamma), s_{2}(\gamma)$ of $\gamma \cap c \subset U \cap c$. Let $\Lambda(\gamma)$ denote the sub-arc whose interior is not contained in the disk bounded by $c$. Then $\Lambda(\gamma)$ is subdivided further into two sub-arcs, which we denote $S_{1}(\gamma)$ and $S_{2}(\gamma)$; for each $i=1,2$, the arc $S_{i}(\gamma)$ is the sub-arc of $\Lambda(\gamma)$ bounded by $s_{i}(\gamma)$ and by $\gamma \cap \omega$ (see Figure \ref{S1S2arcs}). We remark that the arcs $S_1(\gamma)$ and $S_2(\gamma)$ are distinguished by the choice that $S_{1}(\gamma)$ contains none of the intersection points with $D$ in the chosen realization of $\Gamma(\epsilon)$. Similarly, if $\gamma \in D$ then $\Lambda(\gamma)$ is the sub-arc of $\gamma$ not contained within $c'$, and $\Lambda(\gamma)$ is subdivided into $S_1(\gamma)$ and $S_2(\gamma)$, where $S_{1}(\gamma)$ is the sub-arc of $\Lambda(\gamma)$ containing no intersections with $U$.
\begin{figure}[]
\centering
\Large
\begin{lpic}{S1S2(4.in)}
\lbl[]{126,126;$S_1(\gamma)$}
\lbl[]{88,77;$s_2(\gamma)$}
\lbl[]{88,111;$s_1(\gamma)$}
\lbl[]{85,95;$c$}
\lbl[]{102,180;$\omega$}
\end{lpic}
\caption{The intersections $s_1(\gamma)$ and $s_2(\gamma)$ of $\gamma \in U$ with the small circle $c$, and the sub-arc $S_1(\gamma)$ of $\gamma$. (The arc $S_2(\gamma)$ is not pictured).}
\label{S1S2arcs}
\vspace{.2cm}
\end{figure}
\begin{figure}[]
\centering
\Large
\begin{lpic}{cycUD(4in)}
\lbl[]{-20,113;$s_1(\gamma)$}
\lbl[]{195,20;$s_1(\gamma'')$}
\lbl[]{225,80;$s_2(\gamma')$}
\lbl[]{0,45;$\gamma''$}
\lbl[]{90,-10;$\gamma$}
\lbl[]{180,178;$\gamma'$}
\end{lpic}
\vspace{.4cm}
\caption{In the cyclic order on $(U \cap c) \sqcup (D \cap c')$, $s_{2}(\gamma')$ immediately follows $s_{1}(\gamma)$, and $s_{1}(\gamma'')$ immediately follows $s_{2}(\gamma')$.}
\label{cyclic order}
\end{figure}
Starting at some point $v$ of $U \cap c$, let $\gamma \in U \cup D$ denote the up curve associated to $v$, and let $\omega_{i}$ denote the transversal to $\gamma$. The point of $(U \cap c) \sqcup (D \cap c')$ immediately following $v$ is obtained as follows: We assume that $v= s_{1}(\gamma)$. Let $\gamma'$ be the curve obtained from $\gamma$ by moving $(g+1)/2$ around the cyclic order on $U \cup D$. Thus the transversal to $\gamma'$ is $\omega_{i+ (g+1)/2}$, where the addition is interpreted modulo $g$. Then the point of $(U \cap c) \sqcup (D \cap c')$ immediately following $v$ is $s_{2}(\gamma')$.
The next point is obtained in a similar fashion: starting at $s_{2}(\gamma')$ we move to the curve whose index in the cyclic order on $U \cup D$ is $(g+1)/2$ from that of $\gamma'$, in the clockwise direction. Letting $\gamma''$ denote this curve, the next point in the cyclic order on $(U \cap c) \sqcup (D \cap c')$ is $s_{1}(\gamma'')$. The ordering is defined by iterating this procedure: add $(g+1)/2$ to the cyclic index, and alternate between $s_{1}$ and $s_{2}$. See Figure \ref{cyclic order}.
Thus far, we have established a cyclic ordering on $U \cup D$, and an associated cyclic ordering on $(U \cap c) \sqcup (D \cap c)$; we also recall that the curves in $U$ and $D$ are equipped with standard orientations. Note that these choices are compatible in the following sense: Recall that the standard orientations of the curves induce inward and outward pointing arrows on the points in $(U\cap c) \sqcup (D\cap c)$. In the established cyclic ordering of $(U\cap c) \sqcup (D\cap c)$, each inward pointing arrow is followed by an outward pointing arrow, and likewise each outward pointing arrow is followed by an inward pointing arrow.
\begin{lemma} \label{Picture} For an edge $e=e(\vv{\alpha_{i}}, \vv{\alpha_{j}})$ of $R_1(\epsilon)$, the integer $\mathcal{M}_1(e)$ (resp.~$\mathcal{N}_1(e)$) is equal to the number of outwardly pointing arrows of $U$ (resp.~$D$) which follow $\iota(\vv{\alpha_{i}})$ and which precede $\iota(\vv{\alpha_{j}})$.
\end{lemma}
\begin{proof}
With the given choices made, the integer $\mathcal{M}_1(e)$ is equal to the number of inward pointing arrows that follow $\vv{\alpha_i}$ and precede $\vv{\alpha_j}$. Apply the involution $\iota$ and the claim follows for $\mathcal{M}_1(e)$. For $\mathcal{N}_1(e)$, the edge $e$ partitions the down curves into four pieces $D_1,D_2,D_3,D_4$, as in Figure \ref{orderingPic}. The label $\mathcal{N}_1(e)$ of the edge $e$ is equal to $|D_4|$, which is also the number of outwardly pointing arrows between $\iota(\vv{\alpha_{i}})$ and $\iota(\vv{\alpha_{j}})$.
\end{proof}
Henceforth, by an \textit{interval} of $(U \cap c) \sqcup (D \cap c')$, we mean the subset which follows a particular point in $(U \cap c) \sqcup (D \cap c')$ and which precedes some other point, as in the statement of Lemma \ref{Picture}. Evidently, given an interval $I$, $\iota(I)$ is an interval bounded by the image of the end points of $I$ under $\iota$.
To finish the proof of Proposition \ref{Iso}, it suffices to show that the orbit of $\epsilon$ can be constructed from information about $P(\epsilon)$ which is preserved under isomorphism of labeled polygons. The number of $1$'s in $\epsilon$ is equal to $|A(\epsilon)|/2$, which is equal to half of the number of edges of $P(\epsilon)$. Thus all that remains is to determine how to interleave the $|B(\epsilon)|/2$ necessary $-1$'s to obtain $\epsilon$. The following lemma completes the proof of Proposition \ref{Iso}.
\end{proof}
\begin{lemma} \label{Alt} If $e=(\vv{\alpha_{i}}, \vv{\alpha_{j}})$ is an edge of $R_1(\epsilon)$, then the number of $-1$'s between $\epsilon_i$ and $\epsilon_j$ is equal to $\mathcal{M}_1(e)+\mathcal{N}_1(e)-1$.
\end{lemma}
\begin{figure}[h]
\centering
\large
\begin{lpic}{Alternating(4in)}
\lbl[]{250,120;There must be}
\lbl[]{250,108; outwardly pointing}
\lbl[]{250,97;arrows alternating}
\lbl[]{250,87;with these}
\Large
\lbl[]{95,-10;$\vv{\alpha_i}$}
\lbl[]{60,195;$\vv{\alpha_j}$}
\end{lpic}
\vspace{.5cm}
\caption{Given standard orientations of the curves in $U$ and $D$, the arrows at the points of $(U\cap c) \sqcup (D\cap c')$ alternate between inward and outward pointing.}
\label{alternatingPic}
\end{figure}
\begin{proof} Using the standard orientations for the curves in $U\cup D$ again, the consecutive elements of $(U \cap c) \sqcup (D \cap c')$ are equipped with opposite pointing arrows. Each $-1$ between the $1$'s associated to $\vv{\alpha_{i}}$ and $\vv{\alpha_{j}}$ corresponds to a down curve whose outwardly-pointing arrow lies between the outwardly-pointing arrows for $\vv{\alpha_{i}}$ and $\vv{\alpha_{j}}$. Since $e$ is an edge of $R_1(\epsilon)$, there are no outwardly pointing arrows for up curves between the outwardly pointing arrows of $\vv{\alpha_{i}}$ and $\vv{\alpha_{j}}$.
By construction, the inward and outward pointing arrows alternate in $(U\cap c) \sqcup (D\cap c')$. Thus the number of outward pointing arrows that are between the outward pointing arrows at $\vv{\alpha_{i}}$ and $\vv{\alpha_{j}}$ is equal to one less than the number of inward pointing arrows in this same interval $I$ (see Figure \ref{alternatingPic}). Each inward pointing arrow in this interval corresponds to an outward pointing arrow in $\iota(I)$.
In turn, each outward pointing arrow in $\iota(I)$ is associated to either an up curve or a down curve-- and is therefore accounted for by either $\mathcal{M}_1(e)$ or $\mathcal{N}_1(e)$ by Lemma \ref{Picture}. This completes the proof of Lemma \ref{Alt}.
\end{proof}
We are now ready to prove Proposition \ref{distinguishing orbits}:
\begin{proof}[Proof of Proposition \ref{distinguishing orbits}]
{Let $\phi\in \mbox{Mod}^*(S)$ have $\phi\cdot \Gamma(\epsilon)= \Gamma(\epsilon')$. By Proposition \ref{MCG-orbit restrictions}, we have either $\phi \cdot A(\epsilon) =A(\epsilon')$ or $\phi \cdot A(\epsilon) = B(\epsilon')$. After composing if necessary with the reflection of the surface that exchanges all up curves $A$ with all down curves $B$ (preserving the $(\mathbb{Z}/2\mathbb{Z}\oplus D_{g})$-orbit of $\epsilon'$), we assume that $\phi\cdot A(\epsilon) = A(\epsilon')$. In this case, $\phi$ induces an isomorphism of the labeled polygons $P(\epsilon) \cong P(\epsilon')$. By Proposition \ref{Iso}, this implies that $\epsilon$ and $\epsilon'$ are in the same $(\mathbb{Z}/2\mathbb{Z}\oplus D_{g})$-orbit.}
\end{proof}
\section{Stabilizing 1-systems}
\label{stabilizing}
We return to stabilization. Suppose $\lambda$ is an almost realization of a maximal complete 1-system on the surface $S_g$ of genus $g$. Given any arc $\alpha$ on $S_g$, we can delete tiny disks at its endpoints (in the complement of any of the curves of $\lambda$), and glue in an annulus. The result is a surface $S_{g+1}$ of genus $g+1$, and a complete 1-system naturally in correspondence with $\Gamma$. We refer to this system of curves by $\Gamma$ as well, the distinction being clear from context. Note that $\Gamma$ is not maximum in $S_{g+1}$, as it consists of two too few curves.
When $\alpha$ intersects each of the curves of $\lambda$ once (on $S_g$), there are two readily available curves to add: Concatenate $\alpha$ with an arc that crosses the annulus to form a new simple closed curve $\alpha'$. By construction, $\alpha'$ intersects each of the curves of $\Gamma$ once. Moreover, the Dehn twist of $\alpha'$ around the core curve of the annulus, which we denote $\alpha''$, intersects each of the curves of $\Gamma\cup\{\alpha'\}$ once. Thus, given an arc $\alpha$ that intersects each of the curves of a realization of a maximum complete 1-system $\lambda$ once, we produce the \emph{stabilization} of $\lambda$ along the arc $\alpha$, denoted $\mathrm{stab}(\lambda,\alpha):=\Gamma\cup\{\alpha',\alpha''\}$, a maximum complete 1-system on $S_{g+1}$. The following is immediate:
\begin{lemma}
\label{stabilized 3-cubes general 1}
{For each $\gamma\in\Gamma$, the trio $\{\alpha',\alpha'',\gamma\}$ in $\mathrm{stab}(\lambda,\alpha)$ forms a triangle. If three curves don't form a triangle in $\Gamma$, then the corresponding curves don't form a triangle in $\mathrm{stab}(\lambda,\alpha)$.}
\end{lemma}
There is a special case of this stabilization procedure: Choose a curve $\gamma\subset\lambda$, and fix an identification of $\gamma$ with $S^1=[0,1]/0\sim1$, and the image of 0 with $p\in\gamma$. Identify an $\epsilon$-neighborhood of $\gamma$ with the annulus $S^1\times(0,1)$, with coordinates chosen, with $\epsilon$ small enough, such that the intersections of the other curves in $\lambda\setminus\gamma$ with the annulus each consist of a single vertical arc. In these coordinates, let $\alpha$ be the arc $$\left\{ \left( t , \frac{1}{2}-t \right) : t\in[0,1] \right\}.$$
We will refer to the stabilization $\mathrm{stab}(\lambda,\alpha)$ by $\mathrm{stab}(\lambda,p,\gamma)$, which we identify with $\Gamma\cup\{\gamma',\gamma''\}$. Note that the stabilization obtained may depend on the realization $\lambda$ and on the point $p\in\gamma$ chosen, in the sense that it is possible that $\mathrm{stab}(\lambda,p,\gamma)$ and $\mathrm{stab}(\eta,q,\gamma)$ are $\mbox{Mod}^*(S_{g+1})$-inequivalent, either for $p\ne q$ or for non-isotopic realizations $\lambda \not\simeq \eta$.
In the case of this special version of stabilization, we note that there is a one-holed torus subsurface $\Sigma\subset S_{g+1}$ so that the curves in $\Gamma \setminus \{ \gamma \}$ are disjoint from $\Sigma$, and so that the curves $\gamma$, $\gamma'$, and $\gamma''$ are homotopic as properly embedded arcs in the surface with boundary $S_{g+1} \setminus \Sigma$.
\begin{lemma}
\label{stabilized 3-cubes general 2}
{For each $\beta \in \Gamma \setminus \{\gamma\}$, the trios $\{\beta,\gamma,\gamma'\},\{\beta,\gamma,\gamma''\}\subset \mathrm{stab}(\lambda,p,\gamma)$ form triangles in $S_{g+1}$.}
\end{lemma}
Note that Lemmas \ref{tool1} and \ref{stabilized 3-cubes general 1} imply that any maximum complete 1-system obtained via stabilization has dual cube complex of dimension at least three, and Lemma \ref{stabilized 3-cubes general 2}, with Theorem \ref{3-to-n} then implies that any obtained via the special case above has dimension at least four.
\begin{remark} All of the maximum complete 1-systems in our construction can be obtained by a sequence of the more specialized stabilizing procedure, applied successively to the canonical example in genus 2. In particular, they all have dual cube complexes of dimension at least three. Question \ref{Stab} in \S\ref{intro} presents itself.
\end{remark}
We fix choices of almost realizations for $\Gamma(\epsilon)$ as in \S\ref{polygon section}, which we denote as well by $\Gamma(\epsilon)$, the distinction being clear from context.
\begin{lemma}
\label{stabilized 3-cubes}
{For any choice of $p\in\delta$, any trio from $A(\epsilon)\subset\mathrm{stab}(\Gamma(\epsilon),p,\delta)$ (resp.~$B(\epsilon)\subset\mathrm{stab}(\Gamma(\epsilon),p,\delta)$) forms a triangle in $S_{g+1}$.}
\end{lemma}
\begin{proof}
{In the chosen realizations, the triangles formed among the curves of $A(\epsilon)$ are all disjoint from the stabilizing arcs (which track $\delta$ very closely), and from the disks chosen in the stabilization process. Thus the triangles persist.}
\end{proof}
\begin{proposition}
\label{MCG-orbit restrictions stabilized}
{If $\phi\in \mbox{Mod}^{*}(S_{g+1})$ satisfies $$\phi\cdot\mathrm{stab}(\Gamma(\epsilon),p,\delta)=\mathrm{stab}(\Gamma(\epsilon'),p,\delta), \text{ then}$$
\begin{enumerate}
\item $\phi\cdot \{\delta,\delta',\delta''\}=\{\delta,\delta',\delta''\}$, and
\item $\phi\cdot \left\{A(\epsilon),B(\epsilon)\right\}=\left\{A(\epsilon'),B(\epsilon')\right\}$.
\end{enumerate}}
\end{proposition}
\begin{proof}
{By Lemma \ref{stabilized 3-cubes}, the hyperplanes corresponding to curves of $A(\epsilon)$ (resp.~$B(\epsilon)$) pass through maximal cubes of dimension $|A(\epsilon)|$ (resp.~$|B(\epsilon)|$), which is even. The hyperplanes corresponding to $\delta$, $\delta'$, and $\delta''$ pass through maximal cubes of odd dimension (e.g.~the cube corresponding to the curves $\delta$, $\delta'$, $\delta''$, $\alpha_{2i-1}$, $\alpha_{2i}$, $\beta_{2j-1}$, and $\beta_{2j}$, when $|A(\epsilon)|,|B(\epsilon)|\ge1$). The conclusion follows as in the proof of Proposition \ref{MCG-orbit restrictions}.}
\end{proof}
We are now able to prove Proposition \ref{stabilizing Gammas}:
\begin{proof}[Proof of Proposition \ref{stabilizing Gammas}]
{Suppose $\phi\in \mbox{Mod}^*(S_{g+1})$ sends $\mathrm{stab}(\Gamma(\epsilon),p,\delta)$ to $\mathrm{stab}(\Gamma(\epsilon'),p,\delta)$, so that by Proposition \ref{MCG-orbit restrictions stabilized} we have that $\phi$ preserves the set $\{\delta,\delta',\delta''\}$.
Consider the one-holed torus subsurface $\Sigma \subset S_{g+1}$ that arises in the course of the stabilizations $\mathrm{stab}(\Gamma(\epsilon),p,\delta)$. Consider the mapping class $\tilde{\phi}$ induced on $S_g$ by deleting $\Sigma$ and identifying the resulting boundary component to a point. Since the curves of $\Gamma(\epsilon)\setminus \{\delta\}$ (and likewise $\Gamma(\epsilon')\setminus\{\delta\}$) can be made disjoint from $\Sigma$, the induced map $\tilde{\phi}$ takes the induced curves $\Gamma(\epsilon)\setminus \{\delta\}$ to $\Gamma(\epsilon')\setminus\{\delta\}$. Moreover, in the resulting surface $\delta,\delta',$ and $\delta''$ go to the same homotopy class of curve, namely $\delta \subset S_g$. Since $\phi\cdot \{\delta,\delta',\delta''\}=\{\delta,\delta',\delta''\}$, we find that $\tilde{\phi}\cdot \delta = \delta$, so that $\Gamma(\epsilon)$ and $\Gamma(\epsilon')$ are equivalent under the action of $\mbox{Mod}^*(S_g)$.}
\end{proof}
|
\section{Introduction}
\label{introduction}
Detecting the presence and position of objects in an image is a fundamental task in
computer vision. For example, tracking humans in video or performing scene understanding
on a still image requires the ability to reason about the number and position of
objects. While great progress has been made in recent years in terms of feature sets,
the basic training procedure has remained the same.
In this procedure, a set of positive and negative image windows are selected from
training images. Then a binary classifier is trained on these windows.
Lastly, the classifier is tested on images containing no targets of interest, and
false alarm windows are identified and added into the training set. The
classifier is then retrained and, optionally, this process is iterated.
This approach does not make efficient use of the available training data since it trains on
only a subset of image windows. Additionally, windows partially overlapping an
object are a common source of false alarms. This training procedure makes it
difficult to directly incorporate these examples into the training set since these
windows are neither fully a false alarm or a true detection. Most importantly,
the accuracy of the
object detection system as a whole, is not optimized. Instead, the accuracy of a binary
classifier on the subsampled training set is used as a proxy.
In this work, we show how to address all of these issues. In particular, we
will show how to design an optimizer that runs over all windows and optimizes
the performance of an object detection system in terms of the number of missed
detections and false alarms in the final system output. Moreover, our
formulation leads to a convex
optimization and we provide an algorithm which finds the globally optimal
set of parameters. Finally, we test our method on three publicly available
datasets and show that it substantially improves the accuracy of the learned detectors.
Strikingly, we find that a single rigid HOG filter can outperform a state-of-the-art
deformable part model if the HOG filter is learned via MMOD.
\section{Related Work}
In their seminal work, Dalal and Triggs introduced the Histogram of Oriented Gradients (HOG)
feature for detecting pedestrians within a sliding window framework \cite{hog}.
Subsequent object detection research has focused primarily on finding improved representations.
Many recent approaches include features for part-based-modeling, methods for combining local features,
or dimensionality reduction \cite{huang, wu, schwartz, duan, tuzel}.
All these methods employ some form of binary classifier trained on positive and negative
image windows.
In contrast, Blaschko and Lampert's research into structured output regression is the most similar
to our own \cite{Blaschko08}. As with our approach, they use a structural support
vector machine formulation, which allows them to train on all window locations.
However, their training procedure assumes an image contains either
0 or 1 objects. While in the present work, we show how to treat
object detection in the general setting where an image may contain any number of objects.
\section{Problem Definition}
In what follows, we will use $r$ to denote a rectangular area of an image. Additionally,
let $\mathcal{R}$ denote the set of all rectangular areas scanned by our object detection
system. To incorporate the common non-maximum suppression practice, we define a valid
labeling of an image as a subset of $\mathcal{R}$ such that each
element of the labeling ``does not overlap'' with each other. We use the following popular definition
of ``does not overlap'': rectangles $r_1$ and $r_2$ do not
overlap if the ratio of their intersection area to total area covered is less than 0.5. That is,
\begin{equation}
\label{eq:overlap}
\frac{Area(r_1 \cap r_2)}{Area(r_1 \cup r_2)} < 0.5.
\end{equation}
Finally, we use $\mathcal{Y}$ to denote the set of all valid labelings.
Then, given an image $x$ and a window scoring function $f$, we can define the object detection
procedure as
\begin{equation}
\label{eq:detector}
y^* = \operatornamewithlimits{argmax}_{y \in \mathcal{Y}}\sum_{r \in y} f(x,r).
\end{equation}
That is, find the set of sliding window positions which have the largest scores
but simultaneously do not overlap. This is typically accomplished with the
greedy peak sorting method shown in Algorithm~\ref{alg:detector}.
An ideal learning algorithm would find the window scoring function which
jointly minimized the number of false alarms
and missed detections produced when used in Algorithm~\ref{alg:detector}.
It should be noted that solving Equation~(\ref{eq:detector}) exactly is not computationally feasible.
Thus, this algorithm does not always find the optimal solution to (\ref{eq:detector}). An
example which leads to suboptimal results is shown in Figure~\ref{bad-overlap}.
However, as we will see, this suboptimal behavior does not lead to difficulties.
Moreover, in the next section, we give an optimization algorithm capable of
finding an appropriate window scoring function for use with Algorithm~\ref{alg:detector}.
\begin{algorithm}[tb]
\caption{Object Detection}
\label{alg:detector}
\begin{algorithmic}[1]
\REQUIRE image $x$, window scoring function $f$
\STATE $\mathcal{D}$ := all rectangles $r \in \mathcal{R}$ such that $f(x,r) > 0$
\STATE Sort $\mathcal{D}$ such that $\mathcal{D}_1 \ge \mathcal{D}_2 \ge \mathcal{D}_3 \ge ...$
\STATE $y^* := \{\}$
\FOR{$i=1$ {\bfseries to} $|\mathcal{D}|$}
\IF{$\mathcal{D}_i$ does not overlap any rectangle in $y^*$}
\STATE $y^* := y^* \cup \{\mathcal{D}_i\}$
\ENDIF
\ENDFOR
\STATE {\bfseries Return:} $y^*$, The detected object positions.
\end{algorithmic}
\end{algorithm}
\begin{figure}[ht]
\begin{center}
\centerline{\includegraphics[scale=0.35]{bad_overlap.pdf}}
\caption{Three sliding windows and their $f$ scores. Assume non-max suppression rejects any rectangles which
touch. Then the optimal detector would select the two outside rectangles, giving a total score of 12, while a greedy detector selects the
center rectangle for a total score of only 7.}
\label{bad-overlap}
\end{center}
\end{figure}
\section{Max-Margin Object Detection}
In this work, we consider only window scoring functions which are linear in their parameters. In particular, we
use functions of the form
\begin{equation}
f(x,r) = \langle w, \phi(x,r) \rangle
\end{equation}
where $\phi$ extracts a feature vector from the sliding window location $r$ in image $x$, and $w$ is a
parameter vector. If we denote the sum of window scores for a set of rectangles, $y$, as $F(x,y)$, then Equation~(\ref{eq:detector})
becomes
\begin{equation}
y^* = \operatornamewithlimits{argmax}_{y \in \mathcal{Y}} F(x,y) = \operatornamewithlimits{argmax}_{y \in \mathcal{Y}}\sum_{r \in y} \langle w, \phi(x,r)\rangle.
\end{equation}
Then we seek a parameter vector $w$ which leads to the fewest possible detection mistakes. That
is, given a randomly selected image and label pair $(x_i,y_i) \in \mathcal{X}\times\mathcal{Y}$,
we would like the score for the correct labeling of $x_i$ to be
larger than the scores for all the incorrect labelings. Therefore,
\begin{equation}
F(x_i, y_i) > \max_{y \neq y_i} F(x_i, y)
\end{equation}
should be satisfied as often as possible.
\subsection{The Objective Function for Max-Margin Object Detection}
Our algorithm takes a set of images $\{x_1, x_2, ..., x_n\} \subset \mathcal{X}$
and associated labels $\{y_1, y_2, ..., y_n\} \subset \mathcal{Y}$ and
attempts to find a $w$ such that the detector makes the correct prediction on each training
sample. We take a max-margin approach \cite{Joachims/etal/09b} and require that the label for each training
sample is correctly predicted with a large margin. This leads to the following
convex optimization problem:
\begin{align}
\label{eq:hard-margin}
\min_{w} & \quad \frac{1}{2}||w||^2 & \\
\nonumber\text{s.t.} & \quad F(x_i,y_i) \geq \max_{y\in\mathcal{Y}} \left[ F(x_i,y) + \triangle(y,y_i)\right], &\forall i
\end{align}
Where $\triangle(y,y_i)$ denotes the loss for predicting a labeling of $y$ when the true
labeling is $y_i$. In particular, we define the loss as
\begin{align}
\triangle(y,y_i) = & L_{miss} \cdot (\text{\# of missed detections}) + \\
\nonumber & L_{fa} \cdot (\text{\# of false alarms})
\end{align}
where $L_{miss}$ and $L_{fa}$ control the relative importance of achieving high
recall and high precision, respectively.
Equation~(\ref{eq:hard-margin}) is a hard-margin formulation of our learning problem.
Since real world data is often noisy, not perfectly separable, or contains outliers, we
extend this into the soft-margin setting. In doing so, we arrive at the defining optimization
for Max-Margin Object Detection (MMOD)
\begin{align}
\label{eq:soft-margin}
\min_{w,\xi} \quad & \frac{1}{2}||w||^2 + \frac{C}{n}\sum_{i=1}^n \xi_i & \\
\nonumber\text{s.t.} \quad & F(x_i,y_i) \geq \max_{y\in\mathcal{Y}} \left[ F(x_i,y) + \triangle(y,y_i)\right] - \xi_i, &\forall i \\
\nonumber & \xi_i \geq 0, \quad \forall i
\end{align}
In this setting, C is analogous to the usual support vector machine parameter and controls the trade-off
between trying to fit the training data or obtain a large margin.
Insight into this formulation can be gained by noting that
each $\xi_i$ is an upper bound on the loss incurred by training example $(x_i,y_i)$.
This can be seen as follows (let $g(x) = \operatornamewithlimits{argmax}_{y\in\mathcal{Y}}F(x,y)$)
\begin{align}
\label{xi1}\xi_i & \geq \max_{y\in\mathcal{Y}} \left[F(x_i,y) + \triangle(y,y_i) \right] - F(x_i,y_i) \\
\label{xi2}\xi_i & \geq \left[F(x_i, g(x_i)) + \triangle(g(x_i),y_i) \right] - F(x_i,y_i) \\
\label{xi3}\xi_i & \geq \triangle(g(x_i),y_i)
\end{align}
In the step from (\ref{xi1}) to (\ref{xi2}) we replace the max over $\mathcal{Y}$ with a particular
element, $g(x_i)$. Therefore, the inequality continues to hold. In going from (\ref{xi2}) to (\ref{xi3})
we note that $F(x_i,g(x_i)) - F(x_i,y_i) \geq 0$ since $g(x_i)$ is by definition the element of $\mathcal{Y}$
which maximizes $F(x_i,\cdot)$.
Therefore, the MMOD objective function defined by Equation~(\ref{eq:soft-margin}) is a convex upper bound
on the average loss per training image
\begin{equation}
\frac{C}{n}\sum_{i=1}^n \triangle(\operatornamewithlimits{argmax}_{y\in\mathcal{Y}}F(x_i,y),y_i).
\end{equation}
This means that, for example, if $\xi_i$ from Equation~(\ref{eq:soft-margin}) is driven to zero then
the detector is guaranteed to produce the correct output from the corresponding training example.
This type of max-margin approach has been used successfully in a number of other
domains. An example is the Hidden Markov SVM \cite{Altun03hiddenmarkov}, which gives
state-of-the-art results on sequence labeling tasks. Other examples include
multiclass SVMs and methods for learning probabilistic context free grammars \cite{Joachims/etal/09b}.
\subsection{Solving the MMOD Optimization Problem }
We use the cutting plane method \cite{joachims_svm_struct, teo_bmrm} to solve the
Max-Margin Object Detection optimization problem defined by Equation~(\ref{eq:soft-margin}). Note that
MMOD is equivalent to the following unconstrained problem
\begin{equation}
\min_w J(w) = \frac{1}{2}||w||^2 + R_{emp}(w)
\end{equation}
where $R_{emp}(w)$ is
\begin{equation}
\frac{C}{n}\sum_{i=1}^n\max_{y\in\mathcal{Y}} \left[F(x_i,y) + \triangle(y,y_i) - F(x_i,y_i) \right].
\end{equation}
Further, note that $R_{emp}$ is a convex function of $w$ and therefore is lower bounded by any tangent plane.
The cutting plane method exploits this to find the minimizer of $J$. It does this by building a
progressively more accurate lower bounding approximation constructed from tangent planes. Each step
of the algorithm finds a new $w$ minimizing this approximation. Then it obtains the tangent plane to $R_{emp}$ at $w$,
and incorporates this new plane into the lower bounding function, tightening the approximation.
A sketch of the procedure is shown in Figure~\ref{fig:cpa}.
\begin{figure}[ht]
\begin{center}
\centerline{\includegraphics[width=\columnwidth]{cpa.pdf}}
\caption{The red curve is lower bounded by its tangent planes. Adding the tangent
plane depicted by the green line tightens the lower bound further.}
\label{fig:cpa}
\end{center}
\end{figure}
Let $\partial R_{emp}(w_t)$ denote the subgradient of $R_{emp}$ at a point $w_t$.
Then a tangent plane to $R_{emp}$ at $w_t$ is given by
\begin{equation}
\langle w, a \rangle + b
\end{equation}
where
\begin{align}
&a \in \partial R_{emp}(w_t) \\
&b = R_{emp}(w_t) - \langle w_t,a\rangle.
\end{align}
Given these considerations, the lower bounding approximation we use is
\begin{equation}
\frac{1}{2}||w||^2 + R_{emp}(w) \geq \frac{1}{2}||w||^2 + \max_{(a,b)\in P}\left[\langle w,a\rangle + b\right]
\end{equation}
where $P$ is the set of lower bounding planes, i.e. the ``cutting planes''.
\begin{algorithm}[tb]
\caption{MMOD Optimizer}
\label{alg:cpa}
\begin{algorithmic}[1]
\REQUIRE $\varepsilon \geq 0$
\STATE $w_0 := 0$, $t := 0$, $P := \{\}$
\REPEAT
\STATE $t := t + 1$
\STATE Compute plane tangent to $R_{emp}(w_{t-1})$, select $a_t \in\partial R_{emp}(w_{t-1})$
and $b_t := R_{emp}(w_{t-1}) - \langle w_{t-1},a_t\rangle$
\STATE $P_t := P_{t-1} \cup \{(a_t,b_t)\}$
\STATE Let $K_t(w) = \frac{1}{2}||w||^2 + \max_{(a_i,b_i)\in P_t}[\langle w,a_i\rangle + b_i]$
\STATE $w_t := \operatornamewithlimits{argmin}_w K_t(w)$
\UNTIL{$\frac{1}{2}||w_t||^2 + R_{emp}(w_t) - K_t(w_t) \leq \varepsilon$}
\STATE {\bfseries Return:} $w_t$
\end{algorithmic}
\end{algorithm}
Pseudocode for this method is shown in Algorithm~\ref{alg:cpa}. It
executes until the gap between the true MMOD objective function and the
lower bound is less than $\varepsilon$. This guarantees convergence
to the optimal $w^*$ to within $\varepsilon$. That is, we will have
\begin{equation}
|J(w^*) - J(w_t)| < \varepsilon
\end{equation}
upon termination of Algorithm~\ref{alg:cpa}.
\subsubsection{Solving the Quadratic Programming Subproblem}
A key step of Algorithm~\ref{alg:cpa} is solving the argmin on
step 7. This subproblem can be written as a quadratic program and solved efficiently
using standard methods. Therefore, in this section we derive a simple quadratic
program solver for this problem. We begin by writing step 7 as a quadratic program and obtain
\begin{align}
\min_{w,\xi} \quad & \frac{1}{2}||w||^2 + \xi \\
\nonumber\text{s.t.} \quad & \xi \geq \langle w,a_i\rangle + b_i, \quad \forall (a_i,b_i) \in P.
\end{align}
The set of variables being optimized, $w$, will typically have many more dimensions than the
number of constraints in the above problem. Therefore, it is more efficient to solve the dual
problem. To do this, note that the Lagrangian is
\begin{equation}
L(w,\xi,\lambda) = \frac{1}{2} ||w||^2 + \xi - \sum_{i=1}^{|P|}\lambda_i(\xi - \langle w,a_i\rangle - b_i).
\end{equation}
and so the dual \cite{fletcher} of the quadratic program is
\begin{align}
\max_{w,\xi,\lambda} \quad & L(w,\xi,\lambda) \\
\nonumber\text{s.t.} \quad & \bigtriangledown_w L(w,\xi,\lambda) = 0, \\
\nonumber & \bigtriangledown_\xi L(w,\xi,\lambda) = 0, \\
\nonumber & \lambda_i \geq 0, \quad \forall i
\end{align}
After a little algebra, the dual reduces to the following quadratic program,
\begin{align}
\label{eq:dual-simple}
\max_{\lambda} \quad & \lambda^Tb - \frac{1}{2}\lambda^T Q \lambda \\
\nonumber\text{s.t.} \quad & \lambda_i \geq 0, \sum_{i=1}^{|P|} \lambda_i = 1
\end{align}
where $\lambda$ and $b$ are column vectors of the variables $\lambda_i$ and $b_i$ respectively
and $Q_{ij} = \langle a_i,a_j\rangle$.
\begin{algorithm}[tb]
\caption{Quadratic Program Solver for Equation~(\ref{eq:dual-simple})}
\label{alg:qp-solver}
\begin{algorithmic}[1]
\REQUIRE $Q,b,\lambda,\varepsilon_{qp} \geq 0$
\STATE $\tau = 10^{-10}$
\REPEAT
\STATE $\bigtriangledown := Q\lambda - b$
\STATE $big := -\infty$
\STATE $little := \infty$
\STATE $l := 0$
\STATE $b := 0$
\FOR{$i=1$ {\bfseries to} $|\mathcal{\bigtriangledown}|$}
\IF {$\bigtriangledown_i > big$ {\bfseries and} $\lambda_i > 0$ }
\STATE $big := \bigtriangledown_i$
\STATE $b := i$
\ENDIF
\IF {$\bigtriangledown_i < little$ }
\STATE $little := \bigtriangledown_i$
\STATE $l := i$
\ENDIF
\ENDFOR
\STATE $gap := \lambda^T\bigtriangledown - little$
\STATE $z := \lambda_b + \lambda_l$
\STATE $x := \max(\tau, Q_{bb} + Q_{ll} - 2Q_{bl})$
\STATE $\lambda_b := \lambda_b - (big-little)/x$
\STATE $\lambda_l := \lambda_l + (big-little)/x$
\IF {$\lambda_b < 0$}
\STATE $\lambda_b := 0$
\STATE $\lambda_l := z$
\ENDIF
\UNTIL{ $gap \leq \varepsilon_{qp}$}
\STATE {\bfseries Return:} $\lambda$
\end{algorithmic}
\end{algorithm}
We use a simplified variant of Platt's sequential minimal optimization method to
solve the dual quadratic program of Equation~(\ref{eq:dual-simple}) \cite{platt.smo}. Algorithm~\ref{alg:qp-solver} contains the pseudocode. In each
iteration, the pair of Lagrange multipliers ($\lambda_b$, $\lambda_l$) which most violate the KKT conditions
are selected (lines 6-13). Then the selected pair is jointly optimized (lines 15-21).
The solver terminates when the duality gap is less than a threshold.
Upon solving for the optimal $\lambda^*$, the $w_t$ needed by step 7 of
Algorithm~\ref{alg:cpa} is given by
\begin{align}
w_t &= -\sum_{i=1}^{|P|} \lambda_i^* a_i.
\end{align}
The value of $\min_w K(w)$ needed for the test for convergence
can be conveniently computed as
\begin{align}
{\lambda^*}^Tb - \frac{1}{2}||w_t||^2.
\end{align}
Additionally, there are a number of non-essential but useful implementation tricks. In particular,
the starting $\lambda$ should be initialized using the $\lambda$ from the previous
iteration of the MMOD optimizer. Also, cutting planes typically become inactive
after a small number of iterations and can be safely removed. A cutting
plane is inactive if its associated Lagrange multiplier is 0. Our implementation
removes a cutting plane if it has been inactive for 20 iterations.
\subsubsection{Computing $R_{emp}$ and its Subgradient }
\begin{algorithm}[t]
\caption{Loss Augmented Detection}
\label{alg:loss_detector}
\begin{algorithmic}[1]
\REQUIRE image $x$, true object positions $y$, weight vector $w$, $L_{miss}$, $L_{fa}$
\STATE $\mathcal{D}$ := all rectangles $r \in \mathcal{R}$ such that $\langle w, \phi(x,r)\rangle + L_{fa} > 0$
\STATE Sort $\mathcal{D}$ such that $\mathcal{D}_1 \ge \mathcal{D}_2 \ge \mathcal{D}_3 \ge ...$
\STATE $s_r := 0, h_r := false, \quad \forall r \in y$
\FOR{$i=1$ {\bfseries to} $|\mathcal{D}|$}
\IF{$\mathcal{D}_i$ does not overlap $\{\mathcal{D}_{i-1}, \mathcal{D}_{i-2}, ... \}$}
\IF{$\mathcal{D}_i$ matches an element of $y$}
\STATE $r := $ best matching element of $y$
\IF{$h_r = false$}
\STATE $s_r := \langle w, \phi(x,\mathcal{D}_i)\rangle$
\STATE $h_r := true$
\ELSE
\STATE $s_r := s_r + \langle w, \phi(x,\mathcal{D}_i)\rangle + L_{fa}$
\ENDIF
\ENDIF
\ENDIF
\ENDFOR
\STATE $y^* := \{\}$
\FOR{$i=1$ {\bfseries to} $|\mathcal{D}|$}
\IF{$\mathcal{D}_i$ does not overlap $y^*$}
\IF{$\mathcal{D}_i$ matches an element of $y$}
\STATE $r := $ best matching element of $y$
\IF{$s_r > L_{miss}$}
\STATE $y^* := y^* \cup \{\mathcal{D}_i\}$
\ENDIF
\ELSE
\STATE $y^* := y^* \cup \{\mathcal{D}_i\}$
\ENDIF
\ENDIF
\ENDFOR
\STATE {\bfseries Return:} $y^*$, The detected object positions.
\end{algorithmic}
\end{algorithm}
The final component of our algorithm is a method for computing $R_{emp}$ and
an element of its subgradient. Recall that $F(x,y)$ and $R_{emp}$ are
\begin{align}
F(x,y) &= \sum_{r\in y} \langle w, \phi(x,r) \rangle \\
R_{emp}(w) &= \frac{C}{n}\sum_{i=1}^n\max_{y\in\mathcal{Y}} \left[F(x_i,y) + \triangle(y,y_i) - F(x_i,y_i) \right].
\end{align}
Then an element of the subgradient of $R_{emp}$ is
\begin{equation}
\partial R_{emp}(w) = \frac{C}{n}\sum_{i=1}^n \left[ \sum_{r\in{y_i^*}}\phi(x_i,r)-\sum_{r\in{y_i}}\phi(x_i,r) \right]
\end{equation}
where
\begin{equation}
\label{eq:loss_detector}
y_i^* = \operatornamewithlimits{argmax}_{y\in\mathcal{Y}} \left[\triangle(y,y_i) + \sum_{r\in y}\langle w, \phi(x_i,r) \rangle \right].
\end{equation}
Our method for computing $y_i^*$ is shown in Algorithm~\ref{alg:loss_detector}. It is a
modification of the normal object detection procedure from Algorithm~\ref{alg:detector}
to solve Equation~(\ref{eq:loss_detector}) rather than (\ref{eq:detector}).
Therefore, the task of Algorithm~\ref{alg:loss_detector} is to find the set of rectangles
which jointly maximize the total detection score and loss.
There are two cases to consider. First, if a rectangle does not hit any truth rectangles, then
it contributes positively to the argmax in Equation~\ref{eq:loss_detector} whenever its score plus the loss per false alarm ($L_{fa}$)
is positive. Second, if a rectangle hits a truth rectangle then we reason as
follows: if we reject the first rectangle which matches a truth rectangle
then, since the rectangles are sorted in descending order of score, we will
reject all others which match it as well. This outcome results in a single
value of $L_{miss}$. Alternatively, if we accept the first rectangle which
matches a truth rectangle then we gain its detection score. Additionally, we
may also obtain additional scores from subsequent duplicate detections, each of
which contributes the value of its window scoring function plus $L_{fa}$.
Therefore, Algorithm~\ref{alg:loss_detector} computes the total score for the
accept case and checks it against $L_{miss}.$ It then selects the result with
the largest value. In the pseudocode, these scores are accumulated in the
$s_r$ variables.
This algorithm is greedy and thus may fail to find the optimal $y_i^*$ according to
Equation~(\ref{eq:loss_detector}). However, it is greedy in much the same way
as the detection method of Algorithm~\ref{alg:detector}. Moreover, since our goal from the
outset is to find a set of parameters which makes Algorithm~\ref{alg:detector} perform
well, we should use a training procedure which respects the properties of the
algorithm being optimized. For example, if the correct output in the case of
Figure~\ref{bad-overlap} was to select the two boxes on the sides, then Algorithm~\ref{alg:detector}
would make a mistake while a method which was optimal would not. Therefore,
it is important for the learning procedure to account for this and learn that in such
situations, if Algorithm~\ref{alg:detector} is to produce the correct output, the side rectangles need a larger
score than the middle rectangle.
Ultimately, it is only necessary for Algorithm~\ref{alg:loss_detector} to give a value of $R_{emp}$
which upper bounds the average loss per training image. In our experiments, we always observed this to be
the case.
\section{Experimental Results}
To test the effectiveness of MMOD, we evaluate it on the TU Darmstadt cows \cite{tu_cows},
INRIA pedestrians \cite{hog}, and FDDB \cite{fddb} datasets.
When evaluating on the first two datasets, we use the same feature extraction
($\phi$) and parameter settings ($C$, $\varepsilon$, $\varepsilon_{qp}$,
$L_{fa}$, and $L_{miss}$), which are set as follows: $C = 25$, $\varepsilon = 0.15 C$, $L_{fa} = 1$, and $L_{miss} = 2$.
This value of $\varepsilon$ means the optimization runs until the potential improvement
in average loss per training example is less than $0.15$. For the QP subproblem, we
set $\varepsilon_{qp} = \min(0.01, 0.1(\frac{1}{2}||w_t||^2 + R_{emp}(w_t) - K_t(w_t)))$ to allow
the accuracy with which we solve the subproblem to vary as the overall optimization progresses.
For feature extraction, we use the popular spatial pyramid bag-of-visual-words model \cite{spatial_pyramid}.
In our implementation, each window is divided into a 6x6 grid. Within
each grid location we extract a 2,048 bin histogram of visual-words.
The visual-word histograms are computed by extracting 36-dimensional HOG \cite{hog} descriptors from each pixel location,
determining which histogram bin the feature is closest too, and adding 1 to that visual-word's bin count.
Next, the visual-word histograms are concatenated to form the feature vector for the sliding
window. Finally, we add a constant term which serves as a threshold for detection. Therefore,
$\phi$ produces 73,729 dimensional feature vectors.
The local HOG descriptors are 36 dimensional and are extracted from 10x10 grayscale pixel blocks
of four 5x5 pixel cells. Each cell contains 9 unsigned orientation bins. Bilinear
interpolation is used for assigning votes to orientation bins but not for spatial
bins.
To determine which visual-word a HOG descriptor corresponds to, many researchers
compute its distance to an exemplar for each bin and assign the vector to the
nearest bin. However, this is computationally expensive, so we use a fast approximate method
to determine bin assignment. In particular, we use a random projection based locality
sensitive hash \cite{lsh}.
This is accomplished using 11 random planes. A HOG vector
is hashed by recording the bit pattern describing which side of each plane it falls on. This
11-bit number then indicates the visual-word's bin.
Finally, the sliding window classification can be implemented efficiently using a set of integral images.
We also scan the sliding window over every location in an image pyramid which downsamples each
layer by $4/5$. To decide if two detections overlap for the purposes of non-max suppression we
use Equation~(\ref{eq:overlap}). Similarly, we use Equation~(\ref{eq:overlap}) to determine
if a detection hits a truth box. Finally, all experiments were run on a single desktop workstation.
\subsection{TU Darmstadt Cows}
We performed 10-fold cross-validation on the TU Darmstadt cows \cite{tu_cows}
dataset and obtained perfect detection results with no false alarms. The best
previous results on this dataset achieve an accuracy of 98.2\% at equal error rate \cite{Blaschko08}.
The dataset contains 112 images, each containing a side-view of a cow.
In this test the sliding window was 174 pixels wide and 90 tall.
Training on the entire cows dataset finishes in 49 iterations and takes 70 seconds.
\subsection{INRIA Pedestrians}
We also tested MMOD on the INRIA pedestrian dataset and followed the testing
methodology used by Dalal and Triggs \cite{hog}.
This dataset has 2,416 cropped images of people for training as well as 912 negative images.
For testing it has 1,132 people images and 300 negative images.
\begin{figure}[ht]
\begin{center}
\scalebox{0.75}{
\centerline{\includegraphics[width=\columnwidth]{drawing_loss2_pascal.pdf}}}
\caption{The y axis measures the miss
rate on people images while the x axis shows FPPW obtained when scanning the detector over
negative images. Our method improves both miss rate and false positives per window compared
to previous methods on the INRIA dataset.}
\label{inria-results}
\end{center}
\end{figure}
The negative testing images have an average of 199,834 pixels per image. We scan our detector
over an image pyramid which downsamples at a rate of 4/5 and stop when the smallest pyramid
layer contains 17,000 pixels. Therefore, MMOD scans approximately 930,000 windows per negative
image on the testing data.
We use a sliding window 56 pixels wide and 120 tall. The entire optimization takes
116 minutes and runs for 220 iterations.
Our results are compared to
previous methods in Figure~\ref{inria-results}. The detection tradeoff curve shows
that our method achieves superior performance even though we use a basic bag-of-visual-word feature set
while more recent work has invested heavily in improved feature
representations. Therefore, we attribute the increased improvement to our
training procedure.
\subsection{FDDB}
Finally, we evaluate our method on the Face Detection Data Set and Benchmark (FDDB) challenge.
This challenging dataset contains images of human faces in multiple poses captured in indoor and outdoor settings.
To test MMOD, we used it to learn a basic HOG sliding window classifier. Therefore the feature extractor ($\phi$) takes in
a window and outputs a HOG vector describing the entire window as was done in Dalal and Triggs's seminal paper\cite{hog}.
To illustrate the learned model, the HOG filter resulting from the first FDDB fold is visualized in Figure~\ref{hog-filter}.
\begin{figure}[ht]
\begin{center}
\scalebox{1.00}{
\centerline{\includegraphics[scale=0.5]{fddb_hog_detector.png}}}
\caption{The HOG filter learned via MMOD from the first fold of the FDDB dataset. The filters
from other folds look nearly identical.}
\label{hog-filter}
\end{center}
\end{figure}
During learning, the parameters were set as follows: $C = 50$, $\varepsilon = 0.01 C$, $L_{fa} = 1$, and $L_{miss} = 1$.
We also upsampled each image by a factor of two so that smaller faces could be detected. Since our HOG filter box is 80x80 pixels
in size this upsampling allows us to detect images that are larger than about 40x40 pixels in size. Additionally, we mirrored the dataset, effectively
doubling the number of training images. This leads to training sizes of about 5000 images per fold and our optimizer requires
approximately 25 minutes per fold.
\begin{figure}[ht]
\begin{center}
\scalebox{0.90}{
\centerline{\includegraphics[width=\columnwidth]{fig_fddb.png}}}
\caption{
A comparison between our HOG filter learned via MMOD and three other techniques, including another HOG
filter method learned using traditional means. The MMOD procedure results in a much more accurate HOG filter.
}
\label{fddb-results}
\end{center}
\end{figure}
To perform detection, this single HOG filter is scanned over the image at each level of an image pyramid and any
windows which pass a threshold test are output after non-max suppression is performed. A ROC curve that compares this learned HOG filter against other
methods is created by sweeping this threshold and can be seen in Figure~\ref{fddb-results}. To create the ROC curve
we followed the FDDB evaluation protocol of performing 10 fold cross-validation and combining the results in a single
ROC curve using the provided FDDB evaluation software.
Example images with detection outputs are also shown in Figure~\ref{fddb-pics}.
In Figure~\ref{fddb-results} we see that the HOG filter learned via MMOD substantially
outperforms a HOG filter learned with the typical linear SVM ``hard negative mining''
approach\cite{kostingerrobust} as well as the classic Viola Jones
method \cite{viola2001rapid}. Moreover, our single HOG filter learned via MMOD
gives a slightly better accuracy than the complex deformable part model of Zhu
\cite{zhu2012face}.
\begin{figure*}[ht]
\begin{center}
\scalebox{0.175}{
\minibox{
\includegraphics[scale=1.99]{fddb1.png}
\includegraphics[scale=1.07]{fddb2.png}
\includegraphics[scale=0.80]{fddb3.png}
}
}
\caption{Example images from the FDDB dataset. The red boxes show the detections from HOG filters
learned using MMOD. The HOG filters were not trained on the images shown. }
\label{fddb-pics}
\end{center}
\end{figure*}
\section{Conclusion}
We introduced a new method for learning to detect objects in images. In
particular, our method leads to a convex optimization and we provided an
efficient algorithm for its solution. We tested our
approach on three publicly available datasets, the INRIA person dataset, TU
Darmstadt cows, and FDDB using two feature representations. On all
datasets, using MMOD to find the parameters of the detector lead to
substantial improvements.
Our results on FDDB are most striking as we showed that
a single rigid HOG filter can beat a state-of-the-art deformable part model
when the HOG filter is learned via MMOD. We attribute our success to
the learning method's ability to make full use of the data. In particular, on
FDDB, our method can efficiently make use of all 300 million sliding window
positions during training. Moreover, MMOD optimizes the overall accuracy of
the entire detector, taking into account information which is typically ignored
when training a detector. This includes windows which partially overlap target
windows as well as the non-maximum suppression strategy used in the final
detector.
Our method currently uses a linear window scoring function. Future
research will focus on extending this method to use more-complex
scoring functions, possibly by using kernels. The work of Yu and Joachims
is a good starting point \cite{Yu/Joachims/08b}. Additionally, while
our approach was introduced for 2D sliding window problems, it
may also be useful for 1D sliding window detection applications, such as
those appearing in the speech and natural language processing domains.
Finally, to encourage future research, we have made a
careful and thoroughly documented implementation of our method available
as part of the open source dlib\footnote{Software available at http://dlib.net/ml.html\#structural\_object\_detection\_trainer} machine learning toolbox \cite{dlib09}.
{\small
\bibliographystyle{ieee}
|
\section{Introduction}
\label{sec:intro}
The importance of constraint optimization is outlined by the impact of its application in a
wide range of
domains, such as supply-chain management (e.g., \cite{supply1,gaudreault:09}), roster scheduling (e.g., \cite{nurse1,burke:04}), combinatorial auctions (e.g., \cite{sandholm:02}), bioinformatics (e.g., \cite{allouche:14,campeottoPDFP13,fiorettoDP15}),
multi-agent systems (e.g., \cite{dovier1}) and probabilistic reasoning (e.g, \cite{Pearl:88}).
In \emph{Constraint Satisfaction Problems (CSPs),} the goal is to find a value assignment for a set of variables that satisfies a set of constraints~\cite{apt:03,handbook}. The assignments satisfying the problem constraints are called \emph{solutions}. In \emph{Weighted Constraint Satisfaction Problems (WCSPs)} the goal is that of finding an optimal solution, given a set of preferences expressed by means of cost functions \cite{shapiro:81,schiex:95,Bistarelli:97}.
\rev{
When the problems involve uncertainty, we recur to the notion of \emph{belief} or \emph{Bayesian} networks (BNs) \cite{Pearl:88}. This framework aims at modeling natural phenomena and exogenous uncertainty through probabilistic reasoning. In BNs the goal is that of answering queries given partial beliefs under conditions uncertainty.
Common tasks over BNs include finding the \emph{most probable explanation} (MPE),
also known as \emph{maximum a posteriori hypothesis} (MAP), that is finding the assignment with largest probability to all unobserved variables given some observed variables
\cite{Dechter:03}. Belief networks are widely applied to a variety of applications, such as, diagnosis~\cite{lerner2000bayesian} and linkage analysis~\cite{fishelson:02,friedman:00}.
}
When resources are distributed among a set of autonomous agents and communication among the agents is restricted, WCSPs take the form of \emph{Distributed Constraint Optimization Problems (DCOPs)}~\cite{modi:05,petcu:05,yeoh:12}.
In this context, agents coordinate their value assignments to minimize the overall sum of resulting constraint costs.
DCOPs have been employed to model various distributed optimization problems, such as meeting scheduling~\cite{maheswaran:04a,yeoh:10}, resource allocation~\cite{farinelli:08,zivan:15}, power network management problems~\cite{kumar:09,naps13,fioretto:AAMAS-17b}, and coordination of appliances in smart homes~\cite{fioretto:AAMAS-17a,rust:16}. We will refer to WCSPs and DCOPs as \emph{discrete optimization problems}.
Algorithms to solve discrete optimization problems can be classified as \emph{exact} and \emph{approximated}.
Exact algorithms are guaranteed to find optimal solutions. However, since solving WCSPs and DCOPs is NP-hard \cite{handbook}, optimally solving these problems results in prohibitive runtime and/or use of resources, such as memory or network load.
In contrast, approximated algorithms trade solution optimality for shorter runtime and a more efficient use of the available resources.
Furthermore, discrete optimization algorithms can adopt two main paradigms: \emph{search} or \emph{inference}.
\emph{Search-based methods} rely on the use of non-deterministic branching rules to explore different value assignments to variables. These rules are applied recursively until all problem variables are assigned.
This process defines a search tree (typically traversed in a depth-first fashion), which has the advantage of requiring only polynomial space. However, the practical efficiency of these methods relies on their ability to prune redundant or
sub-optimal subtrees.
\emph{Inference-based methods} are inspired from dynamic programming (DP) techniques. These methods apply a sequence of transformations
to reduce the problem size at each step while preserving its semantics. A well known inference-based approach is \emph{Bucket Elimination} (BE) \cite{Dechter:99}. BE
iterates over the variables of the problem, reducing the size of the problem at each step, by replacing a variable and its related cost functions with a single new function, derived by optimizing over the possible values of the eliminated variable. The \emph{Dynamic Programming Optimization Protocol (DPOP)}~\cite{petcu:05} is one of the most efficient inference-based DCOP solvers, and it can be seen as a distributed version of BE, where agents exchange newly introduced cost functions via messages.
The importance of inference-based approaches arises in several optimization fields including constraint programming~\cite{apt:03,handbook}. For example, several \emph{propagators} adopt DP-based techniques to establish constraint consistency. For instance, {\bf (1)}~the \emph{knapsack} constraint propagator proposed by Trick applies DP techniques to establish arc consistency on the constraint~\cite{trick:03}; {\bf (2)}~the propagator for the \emph{regular} constraint establishes arc consistency using a specific digraph representation of the DFA, which has similarities to dynamic programming~\cite{pesant:04}; {\bf (3)}~the \emph{context free grammar} constraint makes use of a propagator based on the CYK parser that uses DP to enforce generalized arc consistency~\cite{quimper:06}.
The main drawback of inference-based methods, including BE and DPOP, is that each transformation may introduce cost functions with large arities, requiring exponential time and space in a key structural parameter of a problem, called \emph{induced width}.
While inference-based approaches may not always be appropriate to solve discrete optimization problems, as their time and space requirements may be prohibitive, they may be very effective in problems with particular structures, such as problems where their underlying \rev{primal graphs} have small induced widths or distributed problems where the number of messages is crucial for performance, despite the size of the messages.
Additionally, approximated inference methods can be effectively used to derive lower bounds, which are important components of branch and bound algorithms, as they can be used to prune parts of the search space by detecting \emph{dominated} solutions---i.e., solutions whose cost can provably not be lower than the best cost found so far. \emph{Mini-Bucket Elimination (MBE)} is an approximated variant of BE that can be used for this purpose.
Recent developments on external-memory algorithms have shown that the use of large secondary data storage can be effective to extend the applicability of memory intensive approaches \cite{Edelkamp2004,lim2010scaling,kask2012beem,sturtevant:13}. However, the computational solving runtime remains a bottleneck.
To contrast this background, we note that the structure exploited by inference-based approaches in constructing solutions makes it suitable to exploit a novel class of massively parallel platforms that are based on the \emph{Single Instruction Multiple Thread} (SIMT) paradigm---where multiple threads may concurrently operate on different data, but are all executing the same instruction at the same time.
The SIMT-based paradigm is widely used in modern \emph{Graphical Processing Units (GPUs)} for general purpose parallel computing. Several libraries and programming environments (e.g., the \emph{Compute Unified Device Architecture} (CUDA)) have been made available to allow programmers to
exploit the parallel computing power of GPUs.
In this paper, we propose the design and implementation of both
an exact and an approximated inference-based algorithm that exploits parallel computation using GPUs to solve WCSPs and DCOPs.
Our proposal aims at employing GPU hardware to speed up the inference process, thus providing
an alternative way to enhance the performance of inference-based discrete optimization approaches.
This paper makes the following contributions:
{\bf (1)}~We propose a novel design and implementation of a centralized and a distributed exact inference-based algorithm, inspired by BE and DPOP, to optimally solve WCSPs and DCOPs, which harnesses the computational power offered by parallel platforms based on GPUs;
{\bf (2)}~We introduce an approximated version of the GPU-based inference-based algorithm, inspired by MBE;
{\bf (3)}~We report an extensive empirical analysis that shows significant improvements in performance with respect to the
sequential CPU-based algorithms, reporting an average speedup of two order of magnitude; and
{\bf (4)}~We show the generality of our approach through empirical evaluations on three different GPU architectures, all providing significant speedups.
\rev{While the description of the techniques proposed in this paper focus on discrete optimization tasks, they also applies to other key problems in probabilistic graphical models, such as, MPE.}
\section{Background: Weighted Constraint Satisfaction Problems}
\label{sec:cop}\label{sec:be}
A \emph{weighted constraint satisfaction problem} (WCSP)~\cite{larrosa:02,shapiro:81} is a tuple $\langle \setf{X}, \setf{D}, \setf{C} \rangle$, where
$\setf{X} = \{x_1, \ldots, x_{n}\}$ is a finite set of variables,
$\setf{D} = \{D_{x_1}, \ldots, D_{x_n}\}$ is a set of finite domains for the variables in $\setf{X}$, with $D_{x_i}$ being the set of possible values for the variable $x_i$, and
$\setf{C}$ is a set of \emph{weighted constraints} (or \emph{cost functions}).
A weighted constraint $f_i \in \setf{C}$ is a function that maps tuples defined on the set of variables relevant to $f_i$ into \rev{$\mathbb{R}_+ \cup \{\infty\}$}, where $\infty$ is a special value denoting that a given combination of values is not allowed.
The set of variables relevant to $f_i$ is referred to as the \emph{scope} of $f_i$, and denoted as $\scope{i} \subseteq \setf{X}$.
Formally, \rev{$f_i : \bigtimes_{x_j \in \scope{i}} D_{x_j} \to \mathbb{R}_+ \cup \{\infty\}$}.\footnote{For simplicity, we assume that tuples of variables are built according to a predefined ordering.}
A \emph{solution} is a value assignment for a subset $\rho$ of variables from $\setf{X}$ that is consistent with their respective domains; i.e., it is a partial function $\theta: \setf{X} \rightarrow \bigcup_{i=1}^n D_{x_i}$ such that, for each $x_j \in \setf{X}$, if $\theta(x_j)$ is defined (i.e., $x_j\in\rho$), then $\theta(x_j) \in D_{x_j}$.
The \bemph{cost} of an assignment $\rho$ is the sum of the evaluation of the constraints involving all the variables in $\rho$.
A solution is \emph{complete} if it assigns a value to each variable in $\setf{X}$ and has finite cost (i.e., different
from $\infty$).
We will use the notation $\sigma$ to denote a complete solution, and, for a set of variables $\setf{V} = \{x_{i_1}, \ldots, x_{i_h}\} \subseteq \setf{X}$, $\sigma_\setf{V} = \tuple{\sigma(x_{i_1}), \ldots, \sigma(x_{i_h}) }$ is the projection of
$\sigma$ to the variables in $\setf{V}$, where $i_1 < \cdots < i_h$.
The goal of a WCSP is to find a complete solution $\sigma^*$ with minimal cost, i.e.,
\begin{equation}
\sigma^* \!=\! \operatornamewithlimits{argmin}_{\sigma \in \Sigma}
\sum_{f_i \in \setf{C}} f_i( \sigma_{\scope{i}} ),
\end{equation}
where $\Sigma$ is the \emph{state space}, defined as the set of all possible complete solutions.
\rev{
Given a WCSP $P$, $G_P \!=\! (\setf{X},E_{\setf{C}})$ is the \bemph{primal graph} of $P$, where $\{x,y\} \in E_{\setf{C}}$ iff $\exists f_i \in \setf{C}$ such that $\{x,y\} \subseteq \scope{i}$.
}
Given an ordering $o$ on $\setf{X}$, we say that a variable $x_i$ has lower \bemph{priority} w.r.t.~a variable $x_j$, denoted $x_i \prec_o x_j$, if $x_i$ precedes $x_j$ in $o$.
\begin{definition}[Induced Graph, Induced Width \cite{Dechter:03}]
Given the \rev{primal} graph $G_P$ and an ordering $o$ on its nodes, the \emph{induced graph} $G_P^*$ on $o$ is the graph obtained from $G_P$ by connecting nodes, processed in descending order of priority, to all their preceding neighbors.
\rev{
Processing a node $x_i$ results in the addition of
edges connecting pairs of preceding neighbors of $x_i$.
}
Given a graph and an ordering of its nodes, the \emph{width} of a node is the number of edges connecting it to its preceding nodes in the ordering. The induced width $w_o^*$ of $G_P$ is maximum width over all nodes of $G_P^*$ along the ordering $o$.
\end{definition}
\begin{figure}[!tbh]
\centering
\includegraphics[width=0.95\textwidth]{Constraints17-WCPS_b}
\caption{Example of a WCSP :
\textbf{(a):} Primal graph.
\textbf{(b):} Cost functions.
\textbf{(c):} A possible induced graph.
\textbf{(d):} A possible pseudo-tree.
\label{fig:wcsp}}
\end{figure}
\begin{example}
\label{ex:wcsp}
Fig.~\ref{fig:wcsp}(a) illustrates the \rev{primal graph} of a simple WCSP instance with 4 binary variables, $x_1$, $x_2$, $ x_3$, and $x_4$, and $5$
\rev{constraints, $f(x_1, x_2)$, $f(x_1, x_4)$, $f(x_2, x_3)$, $f(x_2, x_4)$, $f(x_3, x_4)$.}
Fig.~\ref{fig:wcsp}(b) illustrates the constraints costs of the WCSP, which associate a cost value for each combination of values for the variables in the scope of the constraints.
Fig.~\ref{fig:wcsp}(c) shows the induced graph $G_P^*$ obtained along the ordering $o = \langle x_1, x_2, x_3, x_4 \rangle$.
Its induced width is $3$.
\end{example}
\begin{definition}[Pseudo-tree \cite{Dechter:13}]
Given a \rev{primal graph} $G_P$ \rev{and an ordering $o$ on its nodes}, a \emph{DFS} \emph{pseudo-tree} arrangement for $G_P$ is a \rev{rooted directed} tree $T \!=\! \langle \setf{X}, E_T\rangle$ of $G_P$ such that if $f_i \!\in\! \setf{C}$ and $\{x,y\} \subseteq \scope{i}$, then $x$ and $y$ appear in the same branch of $T$.
\rev{The root of $T$ is the node associated to the variable with lower priority in $o$.}
Edges of $G_P$ that are \emph{in} (resp. \emph{out} of) $E_T$ are called \emph{tree edges} (resp. \emph{backedges}). The tree edges connect parent-child nodes, while backedges connect a node with its \emph{pseudo-parents} and its \emph{pseudo-children}.
\end{definition}
\begin{example}
\label{ex:wcsp-pt}
Fig.~\ref{fig:wcsp}(d) shows one possible pseudo-tree $T \!=\! \langle \setf{X}, E_T\rangle$ associated to the \rev{primal graph} shown in Fig.~\ref{fig:wcsp}(a), with
\rev{$E_T = \{f(x_1, x_2), f(x_2, x_3), f(x_3, x_4)\}$, and order $o = \langle x_1, x_2, x_3, x_4 \rangle$.}
\rev{The node labeled $x_1$ is the root node; it has a pseudo-child, node $x_4$. The node labeled $x_4$ has two pseudo-parents nodes: $x_2$ and $x_1$.}
The solid lines describe tree edges, while the dotted lines represent backedges.
\end{example}
\begin{definition}[Projection]
The \emph{projection} of a cost function $f_i$ on a set of variables $\setf{V} \subseteq \scope{i}$ is a new cost function
\rev{$ \fproj{i}{\setf{V}} : \setf{V} \to \mathbb{R}_+ \cup \{\infty\}$}, such that for each possible assignment
$\theta \in \bigtimes_{x_j \in \setf{V}} D_{x_j}$,
$\fproj{i}{\setf{V}} (\theta) = \displaystyle \min_{\sigma \in \Sigma, \sigma_\setf{V}=\theta} f_i(\sigma_{\scope{i}})$.\footnote{For simplicity, we also use $\theta$ to represent the tuple
$\langle \theta(x_{i_1}),\dots, \theta(x_{i_h})\rangle$ where $\{x_{i_1},\dots, x_{i_h}\}$ is the domain of
$\theta$.}
\end{definition}
In other words, $\fproj{i}{\setf{V}}$ is constructed from the tuples of $f_i$, removing the values of the variable that do not appear in $\setf{V}$ and removing duplicate values by keeping the minimum cost of the original tuples in $f_i$.
\begin{definition}[Concatenation]
Let us consider two assignments $\theta'$, defined for variables $V$, and $\theta''$, defined for
variables $W$, such that for each $x\in V\cap W$ we have that $\theta'(x) = \theta''(x)$.
Their \emph{concatenation} is an assignment $\theta' \cdot \theta''$ defined for
$V\cup W$, such as for each $x\in V$ (resp. $x\in W$) we have that
$\theta'\cdot\theta''(x) = \theta'(x)$ (resp. $\theta'\cdot\theta''(x) = \theta''(x)$).
\end{definition}
We define two operations on cost functions:
\bitemize
\item The \bemph{aggregation} of two functions $f_i$ and $f_j$, is a function \rev{$f_i + f_j : \scope{i} \cup \scope{j} \to \mathbb{R}_+ \cup \{\infty\}$}, such that $\forall \theta' \in \bigtimes_{x_k \in \scope{i}} D_{x_k}$ and $\forall \theta'' \in \bigtimes_{x_k \in \scope{j}} D_{x_k}$, if $\theta'\cdot\theta''$ is defined, then we have that
$$(f_i + f_j) (\theta' \cdot \theta'') \!=\! f_i(\theta') + f_j(\theta'').$$
\item The \bemph{elimination} of a variable $x_j \in \scope{i}$ from a function $f_i$, denoted as $\pi_{-x_j}(f_i)$, produces a new function with scope $\scope{i}\setminus\{x_j\}$, and defined as the projection of $f_i$ on $\scope{i}\setminus\{x_j\}$, i.e., $$\pi_{-x_j}(f_i) \!=\! \fproj{i}{\scope{i}\smallsetminus\{x_j\}}.$$
\end{list}
\rev{While the aggregation and elimination operators are defined on summation and minimization, respectively, for discrete optimization problems, several tasks in belief networks can be solved by using variants of the aggregation and elimination operators~\cite{Dechter:03}.}
\begin{algorithm}[!htbp]
\tcc{Variable Elimination Phase}
\For{$i\leftarrow n$ \textnormal{\textbf{downto}} $1$}{
$B_i \leftarrow \{ f_j \in \setf{C} \: | \: x_i \in \scope{j} \land i = \min \{k \: | \: x_k \in \scope{j}\} \}$\;
$\hat{f}_i \leftarrow \pi_{-x_i}\Big( \sum_{f_j \in B_i} f_j \Big)$\;
$\setf{X} \leftarrow \setf{X} \setminus \{x_i\}$\;
$\setf{C} \leftarrow (\setf{C} \setminus B_i) \cup \{\hat{f}_i\}$\;
}
\tcc{Value Assignment Phase}
\For{$i\leftarrow 1$ \textnormal{\textbf{to}} $n$}{
$x_i \leftarrow d_i$ s.t.~$d_i \in D_{x_i}$ and $d_i$ is the best extension of $x_1, \ldots, x_{i-1}$ w.r.t.~$B_i$\;
}
\Return{$\hat{f}_1$}\;
\caption{\textsc{Bucket Elimination}}
\label{alg:be}
\end{algorithm}
\subsection{Bucket Elimination}
\label{sec:be}
\emph{Bucket Elimination} (BE) \cite{Dechter:99,Dechter:03} is a complete inference algorithm
that can be used to find the optimal solutions of a WCSP. Algorithm~\ref{alg:be} illustrates its pseudocode.
BE operates in the following two phases:
\bitemize
\item \bemph{Variable Elimination Phase}. BE operates from the highest to lowest priority variable. When operating on variable $x_i$, it creates a bucket $B_i$, which is the set of all cost functions that involve $x_i$ as the highest priority variable in their scope (line~2). The algorithm then computes a new cost function $\hat{f}_i$ by aggregating the functions in $B_i$ and eliminating $x_i$ (line~3). Thus, $x_i$ can be removed from the set of variables $\setf{X}$ to be processed (line~4) and the new function $\hat{f}_i$ replaces in $\setf{C}$ all the cost functions that appear in $B_i$ (line~5).
\rev{Thus, a bucket $B_i$ contains both the original WCSP functions as well as the functions placed in it during the variable elimination process.}
In this work, we refer to the $\hat{f}_i$ functions as the \bemph{bucket functions}.
\item \bemph{Value Assignment Phase}. Once the variable with the lowest priority has been processed, the algorithm considers variables in increasing order of priority. For each variable $x_i$, it generates an optimal assignment by selecting a value $d_i \in D_{x_i}$ that minimizes the cost of the functions in $B_i$ given the assignments of all the other variables appearing in the scope of the functions in $B_i$.
\end{list}
As a byproduct, and without additional overhead, BE can compute the number of \rev{consistent} solutions of the problem (see \cite{Dechter:03}, for details).
The time and space complexity of BE is exponential on the induced width of the underlying \rev{primal graph}, which captures the maximum arity of the $\hat{f}_i$ functions (line~3).
\begin{example}
In our WCSP example of Fig.~\ref{fig:wcsp}, during the \emph{Variable Elimination Phase}, BE operates, in order, on the variables $x_4$, $x_3$, $x_2$, and $x_1$.
When $x_4$ is processed, \rev{the bucket $B_4 \!=\! \{f(x_1, x_4), f(x_2, x_4), f(x_3, x_4)\}$} is generated, and highlighted in Fig.~\ref{fig:be}(a)(top) by red edges. The resulting bucket function $\hat{f}_4$ is shown in Fig.~\ref{fig:be}(a)(bottom), where the rightmost column shows the values for $x_4$ after its elimination.
BE, hence, updates the sets $\setf{X} \!=\! \{x_1,x_2,x_3\}$ and \rev{$\setf{C} \!=\! \{f(x_1, x_2), f(x_2,x_3), \hat{f}_4\}$}, as shown in the \rev{primal graph} of Fig.~\ref{fig:be}(b)(top), where the function $\hat{f}_4$ is displayed as a dotted line.
When $x_3$ is processed, \rev{$B_3 \!=\! \{f(x_2, x_3), \hat{f}_4\}$},
and $\hat{f}_3$ is shown in Fig.~\ref{fig:be}(b)(bottom).
Thus, $\setf{X} \!=\! \{x_1,x_2\}$ and \rev{$\setf{C} \!=\! \{f(x_1, x_2), \hat{f}_3\}$}, as shown in Fig.~\ref{fig:be}(c)(top).
Next, $x_2$ is processed, and \rev{$B_2 \!=\! \{f(x_1, x_2), \hat{f}_3\}$}, and $\hat{f}_2$ is illustrated in Fig.~\ref{fig:be}(c)(bottom).
Thus, $\setf{X} \!=\! \{x_1\}$ and $\setf{C} \!=\! \{\hat{f}_2\}$, as shown in Fig.~\ref{fig:be}(d)(top).
Lastly, the algorithm processes $x_1$, sets $B_1 \!=\! \{\hat{f}_2\}$, and $\hat{f}_1$ is minimized when
$x_1 \!=\! 0$, as shown in Fig.~\ref{fig:be}(d)(bottom).
\begin{figure}[!tb]
\centering
\includegraphics[width=1.00\textwidth]{Constraint17_BE_2b}
\caption{Bucket Elimination steps for the WCSP of Fig.~\ref{fig:wcsp} .
\label{fig:be}}
\end{figure}
Next, BE starts the \emph{Value Assignment Phase}, which operates, in order, on the variables $x_1$, $x_2$, $x_3$, and $x_4$. First, it selects the value that minimizes $\hat{f}_1$, ($x_1 \!=\! 0$). Thus, it processes $x_2$, and selects the value $x_2 \!=\! 1$, as it minimizes $\hat{f}_2$ when $x_1 \!=\! 0$, as illustrated in Fig.~\ref{fig:be}(c)(bottom).
Similarly, when BE processes $x_3$, it selects the value $x_3 \!=\! 0$, as it minimizes $\hat{f}_3$ when $x_1 \!=\! 0$ and $x_2 \!=\! 1$, illustrated in Fig.~\ref{fig:be}(b)(bottom).
Finally, BE processes the last variable $x_4$ and assigns it the value $1$, since it minimizes $\hat{f}_4$ when $x_1\!=\! 0, x_2\!=\! 1$, and $x_3\!=\! 0$, illustrated in Fig.~\ref{fig:be}(a)(bottom). Thus, $\sigma^* \!=\! \tuple{0,1,0,1}$ is an optimal solution to the problem, with cost 4.
\end{example}
\begin{algorithm}[!t]
\tcc{Variable Elimination Phase}
\For{$i\leftarrow n$ \textnormal{\textbf{downto}} $1$}{
$B_i \leftarrow \{ f_j \in \setf{C} \: | \: x_i \in \scope{j} \land i = \min \{k \: | \: x_k \in \scope{j}\} \}$\;
Let $\{ B_{i_1}, \ldots, B_{i_m}\}$ be a partition of $B_i$ s.t.
${\left| \bigcup_{f_j \in B_{i_k}} \hspace{-0pt} \scope{j} \right| \leq z}$, for each $k = 1, \ldots, m$\;
%
\ForEach{ $k \in \{1, \ldots, m\}$} {
$\hat{f}_{i_k} \leftarrow \pi_{-x_i}\Big( \sum_{f_j \in B_{i_k}} f_j \Big)$\;
$\setf{C} \leftarrow (\setf{C} \setminus B_{i_k}) \cup \{\hat{f}_{i_k}\}$\;
}
$\setf{X} \leftarrow \setf{X} \setminus \{x_i\}$\;
}
\tcc{Value Assignment Phase}
\For{$i\leftarrow 1$ \textnormal{\textbf{to}} $n$} {
$x_i \leftarrow d_i$ s.t.~$d_i \in D_{x_i}$ and $d_i$ is the best extension of $x_1, \ldots, x_{i-1}$ w.r.t.~$B_i$\;
}
\Return{$\hat{f}_1$}\;
\caption{\textsc{Mini-Bucket Elimination}($z$)}
\label{alg:mbe}
\end{algorithm}
\subsection{Mini-Buckets}
\label{sec:mbe}
The memory complexity and time complexity of BE depend on the arity of the functions $\hat{f}$ produced during the variable elimination step. Such requirements can quickly become infeasible for problems with large induced widths.
To overcome this limitation, Dechter and Rish proposed an incomplete version of the Bucket Elimination~\cite{dechter:03b}. The \emph{Mini-Bucket Elimination} (MBE) is an approximated version of the BE that allows one to bound the arity of the functions $\hat{f}_i$ generated during the Variable Elimination Phase.
Its pseudocode is illustrated in Algorithm~\ref{alg:mbe}.
Similarly to BE, MBE operates in the following two phases:
\bitemize
\item \bemph{Variable Elimination Phase}. As in BE, during the variable elimination phase,
MBE operates on the problem variables in decreasing order of priority. However, rather than creating a single bucket function $\hat{f}_i$ whose scope is the union of the scope of each function in the bucket $B_i$, it partitions $B_i$ in a set of $m$ ``mini"-buckets $\{B_{i_1}, \ldots, B_{i_m}\}$, such that the size of the scope of the bucket function $\hat{f}_{i_k}$, obtained by aggregating the functions in $B_{i_k}$, is bounded by a parameter $z$, for each $k \in \{1,\ldots,m\}$ (line~3).
Thus, MBE considers each mini-bucket independently, and computes $m$ new bucket functions $\hat{f}_{i_k}$, by aggregating the functions in $B_{i_k}$ and eliminating $x_i$ (line~5). These functions replace in \setf{C} all the functions that appear in $B_{i_k}$ (line~6), and the set of variables is updated as in BE (line~7).
\item \bemph{Value Assignment Phase}. This phase
is analogous to that of BE (lines~8--9).
\end{list}
Consider the elimination step for a variable $x_i \in \setf{X}$. Since:
$$
\sum_{k=1}^m \left[ \pi_{-x_i}\Big( \sum_{f_j \in B_{i_k}} f_j \Big) \right]
\leq
\pi_{-x_i}\Big( \sum_{f_j \in B_{i}} f_j \Big)
$$
eliminating $x_i$ using mini-buckets produces a lower bound on the optimal cost for the bucket $B_i$.
Thus, MBE produces a lower bound on the optimal solution cost. Running the Value Assignment Phase might hence return a sub-optimal solution, whose evaluation will be an upper bound on the optimal solution cost.
The time and space complexity of MBE is exponential on the maximal arity of the aggregated functions in the mini-buckets (line~13), and thus it is bounded by the parameter $z$.
\begin{example}
Consider the WCSP of Fig.~\ref{fig:wcsp} solved with MBE using $z=1$.
As in BE, during the \emph{Variable Elimination Phase} MBE operates, in order, on the variables $x_4, x_3, x_2$, and $x_1$.
When $x_4$ is processed, the bucket \rev{$B_4 \!=\! \{f(x_1, x_4), f(x_2, x_4), f(x_3, x_4)\}$}---illustrated by the red edges in Fig.~\ref{fig:be}(a) top---would result in aggregated bucket function whose arity is $3$, and thus exceeds the maximal arity allowed.
Thus, MBE creates a partition $\{ B_{4_1}, B_{4_2}, B_{4_3} \}$ for $B_4$, whose sets consists of the functions, respectively, \rev{$f(x_1, x_4), f(x_2, x_4)$, and $f(x_3, x_4)$}.
The resulting functions $\hat{f}_{4_1}$, $\hat{f}_{4_2}$, and $\hat{f}_{4_3}$ have arity $1$, as illustrated in Fig.~\ref{fig:be}(a) bottom.
Then, MBE updates the sets $\setf{X}$ to $\{x_1,x_2,x_3\}$ and $\setf{C}$ to
\rev{$\{f(x_1, x_2), f(x_2, x_3), \hat{f}_{4_1}, \hat{f}_{4_2}, \hat{f}_{4_3}\}$}, as shown in the \rev{primal graph} of Fig.~\ref{fig:be}(b) top.
When $x_3$ is processed, \rev{$B_3 \!=\! \{f(x_2, x_3), \hat{f}_{4_3}\}$}, marked red in Fig.~\ref{fig:be}(b) top, and the mini-bucket $B_{3_1} \!=\! B_3$. The resulting bucket function $\hat{f}_{3_1}$ is shown in Fig.~\ref{fig:be}(b) bottom.
Thus, $\setf{X} \!=\! \{x_1,x_2\}$ and \rev{$\setf{C} \!=\! \{f(x_1, x_2), \hat{f}_{4_1}, \hat{f}_{4_2}\hat{f}_{3_1}\}$}.
Next, $x_2$ is processed; \rev{$B_2 \!=\! B_{2_1} \!=\! \{f(x_1, x_2), \hat{f}_{4_2}, \hat{f}_{3_1}\}$},
and $\hat{f}_{2_1}$ is illustrated in Fig.~\ref{fig:be}(c) bottom.
Thus, $\setf{X} \!=\! \{x_1\}$ and $\setf{C} \!=\! \{\hat{f}_{4_1}, \hat{f}_{2_1}\}$.
Lastly, the algorithm processes $x_1$, sets $B_1 \!=\! B_{1_1} \!=\! \{\hat{f}_{4_1}, \hat{f}_{2_1}\}$, and $\hat{f}_{1_1}$ is minimized when
$x_1 \!=\! 1$, as shown in Fig.~\ref{fig:be}(d) bottom.
\begin{figure}[!t]
\centering
\includegraphics[width=1.00\textwidth]{Constraint17_MiniBEb}
\caption{Mini-Bucket Elimination steps for the WCSP of Fig.~\ref{fig:wcsp} :
\label{fig:minibe}}
\end{figure}
The \emph{Value Assignment Phase} is analogous to the process carried by BE, except that when processing variable $x_4$ MBE assigns it the value $1$, since it minimizes $\hat{f}_{4_1} + \hat{f}_{4_2} + \hat{f}_{4_3}$ when $x_1\!=\! 0, x_2\!=\! 1$, and $x_3\!=\! 0$ (Fig.~\ref{fig:be}(a) bottom). Thus, $\sigma^* \!=\! \tuple{0,1,0,1}$ is the reported solution to the problem, with a lower bound cost of 2.
\end{example}
\section{\rev{Background: Belief Networks and Most Probable Explanation}}
\label{sec:bn}
A \bemph{belief network} (BN) \cite{Pearl:88} is a tuple $\langle \setf{X}, \setf{D}, \setf{P} \rangle$, where
$\setf{X} = \{x_1, \ldots, x_n\}$ is a set of ordered variables defined over finite domains
$\setf{D} = \{ D_{x_1}, \ldots, D_{x_n} \}$, with $o$ an ordering of the variables in $\setf{X}$, and
$\setf{P}$ is a set of \emph{conditional probability tables (CPTs)}.
A CPT $f_i = \{ \Pr(x_i | \textsl{pa}_i) \}$ of $\setf{P}$ denotes the join probability of $x_i$ with respect to the variables in $\textsl{pa}_i$, and $\textsl{pa}_i \subseteq \{ x_j \in \setf{X} | x_i \prec_o x_j \}$ is the set of variables with higher priority of $x_i$ in the ordering $o$, also called \emph{parent} variables of $x_i$.
%
A BN $B$ represents the probability distribution over the variables in $\setf{X}$:
$$
\Pr\!_{B} (\sigma) = \prod_{i=1}^n \Pr(x_i | \textsl{pa}_i),
$$
where $\sigma$ is a complete assignment for the variables in $\setf{X}$.
The \emph{scope} of a CPT $f_i \in \setf{P}$ is the set $\scope{i} = \{x_i\} \cup \textsl{pa}_i$. An \bemph{evidence set} $\sigma_E$ is an assigned subset of variables $E \subseteq \setf{X}$.
A BN $B$ is represented through a directed acyclic graph $G_B = (\setf{X}, E_P)$, where $(y, x_i) \in E_P$ iff $\exists f_i \in \setf{P}$ such that $y \in \textsl{pa}_i$. In other words, $E_P$ is the set of all directed arcs from each parent variable of $x_i$ to $x_i$, for every $x_i \in \setf{X}$.
The primal graph of a BN is called \bemph{moral graph} and it connects any two variables appearing in the same CPT.
\begin{figure}[!thb]
\centering
\includegraphics[width=0.75\textwidth]{Constraints17-BNa}
\caption{Example of a Belief Network :
\textbf{(a):} Belief Network: $Pr(x_4, x_3, x_2, x_1) = Pr(x_4|x_1) Pr(x_3| x_1, x_2) Pr(x_2) Pr(x_1)$.
\textbf{(b):} conditional probability tables.
\textbf{(c):} Its moral graph.
\label{fig:bn}}
\end{figure}
\begin{example}
Fig.~\ref{fig:bn}(a) illustrates a simple belief network with $4$ binary variables: $x_1, x_2, x_3$ and $x_4$, representing respectively the observations for a patient to have \emph{flu}, \emph{allergy}, \emph{fever}, and \emph{sinus infection}, and $4$ CPTs, each associated to a node and its parent nodes. For example, the CPT table illustrated in Fig.~\ref{fig:bn}(b) (top), describes the probability the patient has fever given that she/he does has have \emph{flu}. The conditional probability for $x_4=0$ is implied since the probabilities need to sum up to 1. Similarly, the CPT table of Fig.~\ref{fig:bn}(b) (bottom) describes the probabilities the patient has sinus infection for each combination of outcomes for flu and allergy.
The BN represents the joint probability distribution:
\begin{align*}
\forall x_1, x_2, x_3, x_4, \quad & Pr(x_4, x_3, x_2, x_1)
= Pr(x_4|x_1) Pr(x_3| x_1, x_2) Pr(x_2) Pr(x_1).
\end{align*}
Its moral graph is illustrated in Fig.~\ref{fig:bn}(c).
\end{example}
\smallskip
One of the main tasks posed over belief networks is that of finding the \bemph{maximum probably explanation} (MPE).
Given a BN $B$, and an evidence set $E$, finding the MPE correspond to finding a complete assignment for the variables of $B$ that has the maximal probability given the evidence $E$. More formally, the goal of the MPE is that of finding a complete assignment $\sigma^*$ such that:
\begin{equation}
\sigma^* = \operatornamewithlimits{argmax}_{\sigma \in \Sigma}
\prod_{f_i \in \setf{P}} \Pr(x_i | \sigma_{\textsl{pa}_i}, E)
\end{equation}
where $\sigma_{\textsl{pa}_i}$ is the projection of $\sigma$ to the variables in $\textsl{pa}_i$.
This problems can be solved with small variations of the Bucket Elimination algorithm presented in section \ref{sec:be}. Bucket Elimination can be adapted to solve MPE tasks on belief networks where the \emph{$\min$ operator} in the projection within the elimination operation is substituted by \emph{$\max$ operator} and the \emph{summation} in the aggregation operator is substituted by the \emph{product} (for more details we refer the reader to~\cite{Dechter:03}).
\iffalse
\subsection{Maximum a posteriori hypothesis (MAP)}
Given a BN $B$, an evidence set $E$, and a set of hypothesized variables $\setf{A} = \subseteq \setf{X}$, finding a maximum a posteriori hypothesis (\textsl{MAP}) correspond to finding an assignment $\sigma_A^*$ for the variable in $\setf{A}$ such that:
\begin{equation}
\sigma_A^* = \operatornamewithlimits{argmax}_{\sigma_A \in \Sigma_A} \sum_{\setf{X} \setminus \setf{A}}
\prod_{f_i \in \setf{P}} (x_i | \sigma_{\textsl{pa}_i}, E)
\end{equation}
where $\Sigma_A$ is set of all possible partial solution over the variables in $\setf{A}$.
In oder words, finding a MAP correspond to finding the partial assignment for the variables in $\setf{A}$ having the largest likelihood given the evidence set $E$.
MAP tasks can be solved with an adaptation of Bucket Elimination where the \emph{$\min$ operator} in the projection within the elimination operation is substituted by the \emph{$\max$ operator}, if the processed variable is one among the hypothesized variables (i.e., it is in $\setf{A}$), or by the \emph{summation} if the processed variables is in $\setf{X} \setminus \setf{A}$; Similarly as for MPE tasks, the \emph{summation} in the aggregation operator is substituted by the \emph{product} \cite{Dechter:03}.
\fi
\section{Background: Distributed Constraint Optimization Problems (DCOPs)}
\label{sec:dcop}
In a \emph{Distributed Constraint Optimization Problem} (DCOP)~\cite{modi:05,petcu:05,yeoh:12}, the variables, domains, and cost functions of a WCSP are distributed among a collection of \emph{agents}. A DCOP is defined as $ \langle \setf{X}, \setf{D}, \setf{C}, \setf{A}, \alpha \rangle$, where $\setf{X}, \setf{D}$, and $\setf{C}$ are defined as in a WCSP, $\setf{A} \!=\! \{a_1, \ldots, a_p\}$ is a set of \emph{agents}, and $\alpha: \setf{X} \rightarrow \setf{A}$ maps each variable to one agent. Following common conventions, we
assume that $\alpha$ is a bijection: Each agent controls exactly one variable. Thus, we will use the terms ``variable'' and ``agent'' interchangeably and assume that $\alpha(x_i) \!=\! a_i$.
In DCOPs, solutions are defined as for WCSPs, and many solution approaches emulate those proposed in the WCSPs literature. For example, ADOPT~\cite{modi:05} is a distributed version of \emph{Iterative Deepening Depth First Search}, and DPOP~\cite{petcu:05} is a distributed version of BE. The main difference is in the way the information is shared among agents. Typically, a DCOP agent knows exclusively its domain and the functions involving its variable. It can communicate exclusively with its neighbors (i.e., agents directly connected to it in the \rev{primal graph}\footnote{The \emph{\rev{primal graph}} of a DCOP is equivalent to that of the corresponding WCSP.}), and the exchange of information takes the form of messages.
Given a DCOP $P$, and a DFS pseudo-tree $T$ for the \rev{primal graph} $G_P$, we use $N(a_i) \!=\! \{ a_j \!\in\! \setf{A} \: | \: \{x_i,x_j\} \!\in\! E_{\setf{C}} \}$ to denote the \bemph{neighbors} of agent $a_i$;
and $sep(a_i)$ to denote the \bemph{separator} of agent $a_i$, which is the set of ancestor agents that are constrained (i.e.,~they are linked in $G_P$) with agent $a_i$ or with one of its descendant agents in the pseudo-tree $T$.
\begin{example}
\label{ex:dcop}
Fig.~\ref{fig:wcsp}(a--b) illustrate an example of a DCOP instance with $4$ agents, $a_i$ $(i\in \{1\ldots,4\})$, each controlling one variable, $x_i$. The problem variables, domains and constraints are analogous to those of the WCSP of Example \ref{ex:wcsp}.
\rev{Fig.~\ref{fig:wcsp}(d)} shows one possible pseudo-tree for the DCOP instance, where the agents $a_1$ and $a_2$ have one pseudo-child: $a_4$. The dotted lines represent backedges.
\end{example}
\subsection{Dynamic Programming Optimization Protocol (DPOP)}
\label{sec:dpop}
DPOP~\cite{petcu:05} is a dynamic programming based DCOP algorithm that is composed of three phases:
\beitemize
\item \bemph{Pseudo-tree Generation Phase}. In this phase the agents coordinate the construction of a pseudo-tree, realized through existing distributed pseudo-tree construction algorithms~\cite{hamadi:98}.
\item \bemph{\mbox{\it UTIL}\ Propagation Phase}. Each agent, starting from the leaves of the pseudo-tree, computes the optimal sum of costs in its subtree for each value combination of variables in its separator set. The agent does so by aggregating the costs of its functions with the variables in its separator and the costs in the \mbox{\it UTIL}\ messages received from its child agents, and then eliminating its own variable.
\item \bemph{\mbox{\it VALUE}\ Propagation Phase}: Each agent, starting from the root of the pseudo-tree, determines the optimal value for its variable. The root agent does so by choosing the value of its variable from its \mbox{\it UTIL}\ computations---selecting the value with the minimal cost. It sends the selected value to its children in a \mbox{\it VALUE}\ message. Each agent, upon receiving a \mbox{\it VALUE}\ message, determines the value for its variable that results in the minimum cost given the variable assignments (of the agents in its separator) indicated in the \mbox{\it VALUE}\ message. Once determined, such assignment is further propagated to the children via \mbox{\it VALUE}\ messages.
\end{list}
\begin{example}
In our example problem, after coordinating to construct the pseudo-tree (Fig.~\ref{fig:wcsp}(d)), agent $a_4$, being the leaf of the pseudo-tree, starts the \mbox{\it UTIL}\ propagation phase, by computing the optimal cost for each value combination of variables $x_1$, $x_2$, and $x_3$ (Fig.~\ref{fig:be}(a)(bottom)), and sending the costs to its parent agent $a_3$ in a \mbox{\it UTIL}\ message. Upon receiving the \mbox{\it UTIL}\ messages from each of its children, agents $a_3$ and $a_2$ follow an analogous process.
When the root agent $a_1$ receives the \mbox{\it UTIL}\ message from each of its children, it computes the minimum cost of the entire problem, and starts the \mbox{\it VALUE}\ propagation phase. It selects the value for its variable that minimizes the problem cost (Fig.~\ref{fig:be}(d)(bottom)) and sends this value down to the pseudo-tree to its child, $a_3$, in a \mbox{\it VALUE}\ message. Upon receiving a \mbox{\it VALUE}\ message from its parent, each agents follows the same process.
\end{example}
\rev{The time and the space complexities of DPOP are dominated by the \mbox{\it UTIL}\ \emph{Propagation Phase}, which is exponential in the size of the largest separator set $sep(a_i)$ for all $a_i \!\in\! \setf{A}$.}
The other two phases require a polynomial number of linear sized messages (in the number of variables of the problem), and the complexity of the local operations is at most linear in the size of the domain \cite{petcu:05}.
Observe that the \mbox{\it UTIL}\ \emph{Propagation Phase} of DPOP emulates the \emph{Variable Elimination Phase} of BE in a distributed context~\cite{brito:10}.
Given a pseudo-tree and its ordering $o$, the \mbox{\it UTIL}\ message generated by each DPOP agent $a_i$ is equivalent to the aggregated and projected function $\hat{f}_i$ in BE when $x_i$ is processed according to the ordering $o$.
\subsection{Approximate Distributed Pseudotree Optimization}
\label{sec:ADPOP}
Analogously to how DPOP emulates BE in the distributed context, the \emph{Approximate Distributed Pseudotree Optimization} (ADPOP) algorithm emulates MBE to solve DCOPs \cite{petcu:05c}. ADPOP has the same three phases as DPOP, and given a pseudo-tree and its ordering $o$, the content of the \mbox{\it UTIL}\ messages generated by each ADPOP agent $a_i$ is equivalent to the bucket functions $\hat{f}_{i_j}$ ($j \in \{ 1, \ldots, i_m\}$) in MBE when $x_i$ is processed according to the ordering $o$.
The complexity of ADPOP is exponential in the input parameter $z$, while its \mbox{\it VALUE}\ \emph{Propagation Phase} has the same order complexity of the \mbox{\it VALUE}\ \emph{Propagation Phase} in DPOP.
\section{Background: Graphical Processing Units (GPUs)}
\label{sec:gpu}
Modern \emph{Graphics Processing Units} (GPUs) are massive parallel architectures, offering thousands of computing cores and a rich memory hierarchy to support graphical processing (e.g., DirectX and OpenGL APIs). NVIDIA's \emph{Compute Unified Device Architecture} (CUDA)~\cite{CUDAbook} aims at enabling the use of the multiple cores of a graphic card to accelerate \emph{general purpose} (non-graphical) applications by providing programming models and APIs that enable the full programmability of the GPU.
The computational model supported by CUDA is \emph{Single-Instruction Multiple-Threads} (SIMT), where multiple threads perform the same operation on multiple data points simultaneously.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{cudafig_2}
\includegraphics[width=0.45\textwidth]{cudafig_1}
\caption{Fermi Hardware Architecture (left) and CUDA Logical Architecture (right)}
\label{fig:cuda}
\end{figure}
A GPU is constituted by a series of \emph{Streaming MultiProcessors} (SMs), whose number depends on the specific characteristics of each class of GPU. For example, the Fermi architecture provides 16 SMs, as illustrated in Fig.~\ref{fig:cuda}(left). Each SM contains a number of computing cores, each of which incorporate an ALU and a floating-point processing unit.
Fig.~\ref{fig:cuda}(right) shows a typical CUDA logical architecture. A CUDA program is a C/C++ program that includes parts meant for execution on the CPU (referred to as the \bemph{host}) and parts meant for parallel execution on the GPU (referred as the \bemph{device}). A parallel computation is described by a collection of \bemph{GPU kernels}, where each kernel is a function to be executed by several \bemph{threads}.
When mapping a kernel to a specific GPU,
CUDA schedules groups of threads (\bemph{blocks}) on the SMs. In turn, each SM partitions the threads within a block in \bemph{warps}\footnote{A warp is typically composed of 32 threads.} for execution, which represents the smallest work unit on the device.
Each thread instantiated by a kernel can be identified by a unique, sequential, identifier ($T_{id}$), which allows to differentiate both the data read by each thread and code to be executed.
\subsection{Memory Organization}
GPU and CPU are, in general, separate hardware units with physically distinct memory types connected by a system bus. Thus, in order for the device to execute some computation invoked by the host and to return the results back to the caller, a data flow needs to be enforced from the host memory to the device memory and vice versa.
The device memory architecture is quite different from that of the host, in that it is organized in several levels differing to each other for both physical and logical characteristics
Each thread can utilize a small number of \emph{registers},\footnote{In modern devices, each SM allots 64KB for registers space.} which have thread lifetime and visibility. Threads in a block can communicate by reading and writing a common area of memory, called \bemph{shared memory}. The total amount of shared memory per block is typically 48KB.
Communication between blocks and communication between the blocks and the host is realized through a large \bemph{global memory}.
The data stored in the global memory has global visibility and lifetime. Thus, it is visible to all threads within the application (including the host), and lasts for the duration of the host allocation.
Apart from lifetime and visibility, different memory types have also different dimensions, bandwidths, and access times.
Registers have the fastest access memory, typically consuming zero clock cycles per instruction, while the global memory is the slowest but largest memory accessible by the device, with access times ranging from 300 to 600 clock cycles.
The shared memory is partitioned into 32 logical banks, each serving exactly one request per cycle.
Shared memory has an extremely small access latency, provided that multiple thread memory accesses are mapped to different memory banks.
\subsection{Bottlenecks and Common Optimization Practices}
\begin{figure}[t]
\centering
\includegraphics[width=0.99\textwidth]{Constraint16-coaleshed}
\caption{Coalesced (left) and scattered (right) data access patterns.}
\label{fig:coaleshed}
\end{figure}
While it is relatively simple to develop correct GPU programs (e.g., by incrementally modifying an existing sequential program), it is nevertheless challenging to design an efficient solution. Several factors are critical in gaining performance. In this section, we discuss a few common practice that are important for the design of efficient CUDA programs.
Memory bandwidth is widely considered to be an important bottleneck for the performance of GPU applications.
Accessing global memory is relatively slow compared to accessing shared memory in a CUDA kernel. However, even if not cached, global accesses covering a contiguous 128 Bytes data are fetched at once. Thus, most of the global memory access latency can be hidden if the GPU kernel employs a \emph{coalesced} memory access pattern.
Fig.~\ref{fig:coaleshed}(left) illustrates an example of coalesced memory access pattern, in which aligned threads in a warp accesses aligned entries in a memory \emph{segment},
which results in a single transaction. Thus, coalesced memory accesses allow the device to reduce the number of fetches to global memory for every thread in a warp. In contrast, when threads adopt a \emph{scattered} data accesses (Fig.~\ref{fig:coaleshed}(right)), the device serializes the memory transaction, drastically increasing its access latency.
Data transfers between the host and device memory is performed through a system bus, which translates to slow transactions. Thus, in general, it is convenient to store the data onto the device memory. Additionally, batching small memory transfers into a large one will reduce most of the per-transfer processing overhead~\cite{CUDAbook}.
The organization of the data in data structures and data access patterns play a fundamental role in the efficiency of the GPU computations.
Due to the computational model employed by the GPU, it is important that each thread in a warp executes the same branch of execution. When this condition is not satisfied (e.g., two threads execute different branches of a conditional construct), the degree of concurrency typically decreases, as the execution of threads performing separate control flows can be serialized. This is referred to as \emph{branch divergence}, a phenomenon that has been intensely analyzed within the \emph{High Performance Computing} (HPC) community~\cite{HanA11,ChakrounMMB13,DiamosAMKWY11}.
\section{GPU-based (Distributed) Bucket Elimination (GPU-(D)BE)}
\label{sec:alg}
Our \emph{GPU-based (Distributed) Bucket Elimination} (GpuBE) framework, extends BE and MBE (DPOP and ADPOP, respectively) by exploiting GPU parallelism within the \emph{aggregation} and \emph{elimination} operations.
These operations are responsible for the creation of the functions $\hat{f}_i$ in BE and $\hat{f}_{i_k}$ in MBE (lines~3 and 5 of Algorithms~1 and 2, respectively) and the \mbox{\it UTIL}\ tables in DPOP and ADPOP (\mbox{\it UTIL}\ \emph{Propagation Phase}), and they dominate the complexity of the algorithms.
Thus, we focus on the details of the design and the implementation relevant to such operations.
The key observation that allows us to parallelize these operations is that the computation of the cost for each value combination in a bucket function is independent of the computation in the other combinations. The use of a GPU architecture allows us to exploit such independence, by concurrently exploring several value combinations of the bucket function, computed by the aggregation operator, as well as concurrently eliminating out variables.
Due to the equivalence of BE (resp. MBE) and DPOP (resp. ADPOP), we will refer to the \emph{bucket} functions $\hat{f}$ and \emph{\mbox{\it UTIL}\ tables} resulted by the aggregation and elimination operations of Algorithms \ref{alg:be} and \ref{alg:mbe}, as well as variables and agents, interchangeably.
\subsection{GPU Data Structures}
In order to fully capitalize on the parallel computational power of GPUs, the data structures need to be designed in such a way to limit the amount of information exchanged between the CPU host and the GPU device, minimizing the accesses to the (slow) device global memory, while ensuring that the data access pattern enforced is coalesced.
To do so, we store into the device global memory exclusively the minimal information required to compute the bucket functions, which are communicated to the GPU once at the beginning of the computation of each bucket or mini-bucket.
This allows the GPU kernels to communicate with the CPU host exclusively to exchange the results of the aggregation and elimination processes.
We introduce the following concept:
\begin{definition}[Bucket-table]
A \emph{bucket-table} is a $4$-tuple, $T =\langle \setf{S}, \setf{R}, \chi, \prec \rangle$, where:
\bitemize
\item $\setf{S} \subseteq \setf{X}$, is a list of variables denoting the \emph{scope} of $T$.
\item $\setf{R}$ is a list of tuples of values, each tuple having length $|S|$. Each element in this list (called \emph{row} of $T$) specifies an assignment of values for the variables in $\setf{S}$ that is consistent with their domains. We denote with $\setf{R}[i]$ the tuple of values corresponding to the $i$-th row in $\setf{R}$, for $i = \{1, \ldots, |\setf{R}|\}$.
\item $\chi$ is a list of length $|R|$ of cost values corresponding to the costs of the assignments in $\setf{R}$.
In particular, the element $\chi[i]$ represents the cost of the assignment $\setf{R}[i]$ for the variables in $\setf{S}$, with $i = \{1, \ldots, |\setf{R}|\}$.
\item $\prec$ denotes an ordering relation used to sort the variables in the list $\setf{S}$. In turn, the value assignments, and cost values, in each row of $\setf{R}$ and $\chi$, respectively, obey to the same ordering.
\end{list}
\end{definition}
As a technical note, a bucket table $T$ is mapped onto the GPU device to store exclusively the cost values $\chi$, not the associated variables values. We assume that the rows of $\setf{R}$ are sorted in lexicographic order---thus,
the $i$-th entry $\chi[i]$ is associated with the $i$-th permutation $\setf{R}[i]$ of the variable values in $\setf{S}$, in lexicographic order. This strategy allows us to employ a simple perfect hashing to efficiently associate row numbers with variables' values. We will elaborate on this topic in Section \ref{sec:aggregation}.
Additionally, all the data stored on the GPU global memory is organized in mono-dimensional arrays, so as to facilitate \emph{coalesced memory accesses}.
\begin{algorithm}[htbp]
\small{\caption{\textsc{GpuBE}(z)}\label{alg:gpu-dbe}}
\SetNlSty{textbf}{}{~~}
\tcc{Variable Ordering Phase (Pseudo-Tree Construction)}
$\setf{\bar{X}} \leftarrow $ Sort $\setf{X}$ w.r.t.~$\prec_{T}$ ordering\;
\tcc{Variable Elimination Phase}
\For{$i\leftarrow n$ \textnormal{\textbf{downto}} $1$, with $x_i \in \setf{\bar{X}}$} {
$B_i \leftarrow$ \textsc{Cpu{::}ConstructBucket($\setf{C}, x_i, z$)}\;
Let $\{ B_{i_1}, \ldots, B_{i_m}\}$ be a partition of $B_i$ s.t.
${\left| \bigcup_{f_j \in B_{i_k}} \hspace{-0pt} \scope{j} \right| \leq z}$, for each $k = 1, \ldots, m$\;
%
\ForEach{ $k \in \{1, \ldots, m\}$} {
$T_{i_k} = \langle B_{i_k}, \setf{R}_{i_k}, \chi_{i_k}, \prec_T \rangle \paral{} \textsc{Gpu{::}Reserve}( |\setf{R}_{i_k}| )$\;
\SetNlSty{textbf}{}{~~}
\ForEach{$f_j \in B_{i_k}$} {
\SetNlSty{textbf}{}{*}
$T_j = \langle \scope{f_j}, \setf{R}_{j}, \chi_{j}, \prec_T \rangle \paral{D \leftarrow H} \textsc{Gpu{::}Reserve}( |\setf{R}_j| )$\;
\SetNlSty{textbf}{}{*}
$T_{i_k} \paral{} \textsc{Gpu{::}Aggregate}( T_{i_k}, T_j )$\;
}
$\hat{f}_{i_k} \paral{H \leftarrow D} \textsc{Gpu{::}Eliminate}( T_{i_k}, x_i )$\;
}
}
\tcc{Variable Assignment Phase}
\For{$i \leftarrow 1$ \textnormal{\textbf{to}} $n$, with $x_i \in \setf{\bar{X}}$} {
$x_i \leftarrow$ \textsc{Cpu{::}FindBestAssignment($x_1, \ldots, x_{i-1}$)}
}
\Return{$\hat{f}_1$}\;
\end{algorithm}
\subsection{Algorithm Overview}
\label{sec:algorithm}
Algorithm~\ref{alg:gpu-dbe} illustrates the pseudocode of GpuBE, where $z$ is an input parameter denoting the maximal mini-bucket size to be processed.
We use the following notations:
\begin{itemize}
\item Starred line numbers denote those instructions that are executed concurrently by both the CPU and the GPU.
\item The symbols $\leftarrow$ and $\paral{}$ denote sequential and parallel (i.e., multiple GPU threads) operations, respectively.
\item If a parallel operation requires a copy from host (device) to device (host), we write $\!\paral{D \leftarrow H}$ ($\!\paral{H \leftarrow D}$).
Host to device (device to host) memory transfers are performed immediately before (after) the execution of the GPU kernel.
\end{itemize}
GpuBE is composed of three phases:
\textbf{(1)} \emph{Variable Ordering},
\textbf{(2)} \emph{Variable Elimination}, and
\textbf{(3)} \emph{Variable Assignment}.
Let us consider $N(x_i) \!=\! \{ x_j \!\in\! \setf{X} \: | \: \{x_i,x_j\} \!\in\! E_{\setf{C}} \}$,
defined analogously as for the agents' case.
During the first phase (line~1), the problem variables are sorted
according to a pseudo-tree ordering relation; in particular, we apply the following
heuristics in the construction of the pseudo-tree:
$x_i \prec_T x_j$ \textit{iff} $|N(x_i)| < |N(x_j)|$, for every $x_i, x_j \in \setf{X}$.
For the distributed case, this phase is identical to that of (A)DPOP, where the agents coordinate the construction of a pseudo-tree, using an off-the-shelf message-passing algorithm \cite{hamadi:98}.
In the second phase, the algorithm processes each variable, in descending order, according to the relation $\prec_T$, and proceeds as in (M)BE:
\bitemize
\item The function $\textsc{Cpu::ConstructBucket}$ constructs the bucket $B_i$ as illustrated in Algorithm 1, line~2. The algorithm proceeds in creating a partition of this bucket, if
required (i.e., if $z < w^*$).
This phase differs slightly in the distributed case, where each agent, upon receiving a new bucket function from its descendant agents, inserts it into its bucket set $B_i$.
\item For each mini-bucket $B_{i_k}$ ($k=1,\ldots,m$), GpuBE determines and reserves the amount of global memory to be assigned to each associated bucket-table $T_{i_k}$ (line~6).
After the \textsc{Gpu::Reserve} function is invoked, a space sufficient to store the
bucket-table is allocated, and its cost values $\chi_{i_k}$ are initialized to $0$.
\item Thus, GpuBE aggregates the bucket-table $T_j$ associated to each function $f_j$ in the mini-bucket with the bucket-table $T_{i_k}$ (lines~7--9). To do so, it first creates a bucket-table $T_j$ that encodes the cost values of the bucket function $f_j$, reordering them, if necessary, according to the order on its scope specified by the pseudo-tree relation $\prec_T$ (line~8). This procedure requires a memory transfer from the CPU host to the GPU device global memory. Then, it adds the values $\chi_j$ of the aggregating bucket-table $T_j$ into the corresponding entries of the bucket-table $T_{i_k}$ (line~9).
We will further discuss the details of this function, as well as the other kernel functions, in the next sections.
\item Finally, the algorithm invokes a GPU call to eliminate the variable $x_i$ from the bucket-table ${T}_{i_k}$, thereby constructing the bucket function $\hat{f}_{i_k}$, which is, finally, copied back to the CPU host memory (line~10).
\end{list}
In the distributed case, each agent processes lines~3--6 in parallel without prior coordination.
Starting from the leaves of the pseudo-tree, the agents build their \mbox{\it UTIL}\ messages containing the bucket functions (lines~5--10), and send them to their parents.
Thus, each agent waits to receive the \mbox{\it UTIL}\ messages from all of its children before performing the aggregation and elimination operations (lines~7--9 and line~10, respectively) for each mini-bucket.
By the end of this phase (line~10), the root agent knows the overall cost for each values of its variable $x_i$. Thus, it chooses the value that results in the minimum cost, and it starts the third phase by sending to each child agent the value of its variable $x_i$.
In the centralized case, when space is not a concern, there is no need of copying the bucket tables back to the host, after the variable elimination step (line~10). Thus, two memory transfer transactions are avoided for each variable being processed.
In the third phase, the algorithm proceeds analogously to as done in (M)BE. For the distributed case, the agents select the values for their variables that minimize their bucket functions costs, given the assignments of their ancestor agents, and send them in \mbox{\it VALUE}\ messages to their children. These operations are repeated by every agent receiving a \mbox{\it VALUE}\ message until the leaf agents are reached.
While we described the case in which the underlying problem \rev{primal graph} is connected, our implementation allows us to handle disconnected graphs. This is done by solving the sub-problems in each connected subgraph independently from other subproblems,
and retrieving the problem cost by aggregating the costs stored in the root of each pseudo-tree associated to the connected graphs.
\subsection{GPU-based Constraint Aggregation}
\label{sec:aggregation}
We now describe the implementation of the constraint aggregation GPU kernel.
This operation, takes as input two bucket-tables: $T_{i_k}$ and $T_j$, and aggregates the cost values in $\chi_j$ to those of $\chi_{i_k}$ for all the corresponding assignments of the shared variables in the scope of the two bucket-tables. We refer to $T_{i_k}$ and $T_j$ as to the \emph{output} and \emph{input} bucket-tables, respectively.
Consider the example in Fig.~\ref{fig:aggregate}, the cost values $\chi_j$ of the input bucket-table $T_j$ (right) are aggregated to the cost values $\chi_{i_k}$ of the output bucket-table $T_{i_k}$ (left)---which where initialized to $0$. The rows of the two tables with identical value assignments for the shared variables $x_2$ and $x_3$ are shaded with the same color.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.60\textwidth]{Constraints-aggregate}
\caption{Example of aggregation of two tables on GPU.
\label{fig:aggregate}}
\end{figure}
To optimize performance of the GPU operations and to avoid unnecessary data transfer to/from the GPU global memory, we only transfer the list of cost values $\chi$ for each bucket-table that need to be aggregated, and employ a simple perfect hashing
function to efficiently associate row numbers with variables' values.
This allows us to compute the indices of the cost vector of the input bucket-table relying exclusively on the information of the thread ID and, thus, avoiding accessing the scope $\setf{S}$ and assignment vectors $\setf{R}$ of the input and output bucket-tables.
We now discuss how this process can be efficiently handled on the GPU kernels.
Let $T^\textsl{out} = \langle \setf{S}^\textsl{out}, \setf{R}^\textsl{out}, \chi^\textsl{out}, \prec^\textsl{out} \rangle$ be the output bucket-table, whose scope is
$\setf{S}^\textsl{out} = \{x_1^\textsl{out}, \ldots, x_m^\textsl{out}\}$.
Let $T^\textsl{in} = \langle \setf{S}^\textsl{in}, \setf{R}^\textsl{in}, \chi^\textsl{in}, \prec^\textsl{in} \rangle$ be the input bucket-table, whose scope is $\setf{S}^\textsl{in} = \{ x_1^\textsl{in}, \ldots, x_s^\textsl{in} \}$, and such that $\setf{S}^\textsl{in} \sqsubseteq \setf{S}^\textsl{out}$,
where $A \sqsubseteq B$ denotes that A is a subsequence of B, and with $s \leq m$. Additionally, let $x_m^\textsl{out} = x_s^\textsl{in}$, i.e., the last variable of the input and output bucket-table scopes coincides. The latter is the variable to be eliminated; We will explain this design choice in the next section, where we will discuss the variable elimination process on a GPU.
Finally, let $\phi_\textsl{out}: \mathbb{N} \to \mathbb{N}$ be a mapping from input bucket-table scope variables indexes to output bucket-table scope variable indexes, such that
$\phi_\textsl{out}(i) = j$ \textit{iff} $x_i^\textsl{in} = x_j^\textsl{out}$.
For instance, in our example of Fig.~\ref{fig:aggregate}, $\phi_\textsl{out}(0) = 1$, as the variable $\setf{S}_j[0] = \setf{S}_{i_k}[1] = x_2$.
Hence, given a row index $r_{\textsl{out}}$ for the output bucket-table $\chi^\textsl{out}$, the corresponding row index $r_\textsl{in}$ associated to the input bucket-table cost array $\chi^\textsl{in}$ is given by:
\begin{equation}
\small{
r_{\textsl{in}} =
\sum_{k=1}^{s-1} \left[
\underbrace{
\left(\prod_{j = k+1}^{s} |D_{x_j^\textsl{in}}| \right)
}_{\varf{mul}[k]}
\!\cdot\!
\left( \lfloor
\frac{r_\textsl{out}}
{\displaystyle\underbrace{\prod_{j = \phi_\textsl{out}(k) + 1}^{m} |D_{x_j^\textsl{out}}|}_{\varf{div}[k]} }
\rfloor
\!\!\!\!\mod\!
\underbrace{|D_{x_k^\textsl{in}}|}_{\varf{mod}[k]}
\right)
\right]
+
r_\textsl{out}\!\!\!\! \mod \! \underbrace{|D_{x_s^\textsl{in}}|}_{\varf{mod}[s]}
}
\label{eq:hash}
\end{equation}
Each term in the summation of Equation~\eqref{eq:hash} represents the contribution of the $k$-th variable's value in $\setf{R}^\textsl{out}[r_\textsl{out}]$, as an offset to the index $r_\textsl{in}$ in the array $\setf{R}^\textsl{in}$.
The vectors $\varf{mul}$, $\varf{div}$, and $\varf{mod}$ are data structures employed to compute efficiently the $r_\textsl{in}$ indices on the GPU. The values $\varf{mul}[k]$, $\varf{div}[k]$, and $\varf{mod}[k]$ (and $\varf{mod}[s]$) can be efficiently computed in $O(s)$, $O(n)$, and $O(1)$, respectively, for each $k=\{1,\ldots,s-1\}$, and copied onto the GPU global memory with one copy transaction---we allocate them as a single mono-dimensional array.
In order to exploit the highest degree of parallelism offered by the GPU device, we
\textbf{(1)} map one GPU thread $T_{id}$ to one element of the output bucket-table $r_\textsl{out}$ and
\textbf{(2)} adopt the ordering relation $\prec_T$ for each input and output bucket-table processed.
Adopting such techniques allows each thread to be responsible of performing exactly two reads and one write from/to the GPU global memory. Additionally, the ordering relation enforced on the bucket-tables allows us to exploit the locality of data and to encourage coalesced data accesses.
As illustrated in Fig.~\ref{fig:aggregate}, this paradigm allows threads (whose IDs are identified in red by their $T_{id}$'s) to operate on contiguous chunks of data and, thus, minimizes the number of actual read (from the input bucket-table, on the right) and write (onto the output bucket-table, on the left) operations from/to the global memory performed by a group of threads with a single data transaction.\footnote{Accesses to the GPU global memory are cached into cache lines of $128$ Bytes, and can be fetched by all requiring threads in a warp.}
\begin{procedure}[!t]
$r_{i_k} \leftarrow $ the thread's entry ID ($T_{id}$)\;
$r_{j} \leftarrow 0$ \tcc{holds the value of the index entry of $\chi_j$}
$s \leftarrow |\setf{S}_j|$\;
$\tuple{ \varf{mul}, \varf{div}, \varf{mod}} \leftarrow \textsc{CopyToSharedMemory}()$\;
\For{$\ell \leftarrow (1 \ldots s \!-\! 1)$} {
$r_{j} \leftarrow r_j + \varf{mul}[\ell] \cdot
\big( \lfloor \frac{r_{i_k}}{\varf{div}[\ell]} \rfloor) \% \varf{mod}[\ell]$ \big)\;
}
$r_{j} \leftarrow r_{j} + \big( r_{i_k} \% \varf{mod}[s] \big)$\;
$\chi_{i_k}[r_{i_k}] \leftarrow \chi_{i_k}[r_{i_k}] + \chi_j[r_j]$\;
{\caption{{Gpu}{::}Aggregate($T_{i_k}, T_j$)} \label{alg:aggregation}}
\end{procedure}
The constraint aggregation GPU kernel is described in Procedure \ref{alg:aggregation}, which is computed in parallel by a number of threads equal to the number of rows of the output bucket-table.
Each thread identifies its row index $r_{i_k}$ within the output bucket-table cost values array $\chi_{i_r}$ based on its thread ID (line~1),
and it initializes a variable that will contain the input bucket-table row index to 0 (line~2). It then copies into the shared memory the static entities $\varf{mul}, \varf{div}$, and $\varf{mod}$ associated to the aggregation of the the bucket-tables being processed (line~4).
A further inspection to the \ref{alg:aggregation} procedure reveals how it makes use of the auxiliary data structures above to efficiently implement the \emph{hash function} of equation~\eqref{eq:hash}, and retrieve the entry index of the input bucket-table associated to the variables value permutation of the output bucket-table $\setf{R}_{i_k}[r_{i_k}]$ (lines~5--7).
Finally, the instruction in line~8 aggregates the corresponding input bucket-table value to the output bucket-table $\chi_{i_k}[r_{i_k}]$.
Note that this algorithm highly fits the SIMT paradigm adopted by GPUs; the thread ID and the auxiliary $\varf{mul}, \varf{div}$, and $\varf{mod}$ arrays are used to retrieve and update all the data necessary to compute the output bucket-table.
Additionally, the accesses to the global memory are minimized, as the auxiliary arrays are copied into the shared memory.
We illustrate the above process in the following example.
\begin{example}
Consider the operation of aggregating the input bucket-table $T_{j}$ with the bucket-table $T_{i_k}$ of Fig.~\ref{fig:aggregate} corresponding, respectively, to the bucket-table representing the constraint $f_{23}$ and the bucket-table $\hat{f}_3$ (before eliminating the variable $x_3$) in Fig.~\ref{fig:be}(b).
With the Equation~\eqref{eq:hash} notation, $s = 2$, $m=3$
and, thus, the index $k$ of the summation ranges from 1 to $s-1= 1$. Therefore:
\begin{align*}
\varf{mul}[0] &= \textstyle \prod_{j=2}^2 |D_{x_j}| = 2
& \varf{div}[0] &= \textstyle \prod_{j=\phi_\textsl{out}(1)+1= 2}^2 |D_{x_j}| = 2 \\
\varf{mod}[0] &= \textstyle |D_{x_2}| = 2
& \varf{mod}[1] &= \textstyle |D_{x_3}| = 2
\end{align*}
Therefore, the mapping from the thread IDs (or, equivalently, the output bucket-table row indices $r_{i_k}$) to the input bucket-table row indices $r_j$ is:
\begin{align*}
T_{id} &= 0 \quad\Rightarrow \quad r_j = 2 \cdot (\textstyle \lfloor \frac{0}{2} \rfloor ) + 0 \!\!\!\!\mod\! 2 = 0 \\
T_{id} &= 1 \quad\Rightarrow \quad r_j = 2 \cdot (\textstyle \lfloor \frac{1}{2} \rfloor ) + 1 \!\!\!\!\mod\! 2 = 1 \\
T_{id} &= 2 \quad\Rightarrow \quad r_j = 2 \cdot (\textstyle \lfloor \frac{2}{2} \rfloor ) + 2 \!\!\!\!\mod\! 2 = 2 \\
T_{id} &= 3 \quad\Rightarrow \quad r_j = 2 \cdot (\textstyle \lfloor \frac{3}{2} \rfloor ) + 3 \!\!\!\!\mod\! 2 = 3 \\
T_{id} &= 4 \quad\Rightarrow \quad r_j = 2 \cdot (\textstyle \lfloor \frac{4}{2} \rfloor ) + 4 \!\!\!\!\mod\! 2 = 0 \\
T_{id} &= 5 \quad\Rightarrow \quad r_j = 2 \cdot (\textstyle \lfloor \frac{5}{2} \rfloor ) + 5 \!\!\!\!\mod\! 2 = 1 \\
T_{id} &= 6 \quad\Rightarrow \quad r_j = 2 \cdot (\textstyle \lfloor \frac{6}{2} \rfloor ) + 6 \!\!\!\!\mod\! 2 = 2 \\
T_{id} &= 7 \quad\Rightarrow \quad r_j = 2 \cdot (\textstyle \lfloor \frac{7}{2} \rfloor ) + 7 \!\!\!\!\mod\! 2 = 3
\end{align*}
\end{example}
As a technical detail, the bucket-tables are created and processed so that the variables in their scope are sorted according to the order $\prec_T$.
This means that the variables with the highest priority appear first in the scope list, while the variable to be eliminated always appear last.
We will see, in the next section, that such detail allows us to efficiently encode the elimination operation on the GPU.
\begin{figure}[!h]
\centering
\includegraphics[width=0.75\textwidth]{Constraint16-host_dev}
\caption{Concurrent computation between host and device.\label{fig:hostdev}}
\end{figure}
To fully capitalize on the use of the GPU, we exploit an additional level of parallelism, achieved by running GPU kernels and CPU computations concurrently (lines~8--9 of Algorithm \ref{alg:gpu-dbe}). This is possible when the $T_j$ bucket-tables can be partitioned in multiple chunks.
Fig.~\ref{fig:hostdev} illustrates this operation. After transferring the first bucket-table chunk ($T_j$ ${}_{\#1}$) into the device memory, the process starts the execution of the \ref{alg:aggregation}() kernel, which operates on this portion of the bucket table (called \emph{Kernel} ${}_{\#1}$ in Fig.~\ref{fig:hostdev}).
Thus, the control immediately returns to the CPU host, which enforces the next data transfer onto the device memory, through a call to a \textsc{Gpu{::}Reserve}($T_j$ ${}_{\#2}$).
A host-device synchronization point is imposed after each memory transfer (except the first one), to ensure that no overlapping \ref{alg:aggregation}() GPU kernels are enforced.
\subsection{GPU-based Variable Elimination}
\label{sec:elimination}
We now describe the implementation of the variable elimination GPU kernel. This operation takes as input a bucket-table $T_{i_k}$ and a variable $x_i \in \setf{S}_{i_k}$ and removes this variable from the bucket-table's scope, optimizing over its cost rows. As a result, the output bucket-table rows list the unique assignments for the value combinations of $\setf{S}_{i_k} \setminus \{x_{i}\}$ in the input bucket-table $\setf{R}_{i_k}$ which minimizes the costs values for each $d \in D_{x_i}$.
\begin{figure}[!t]
\centering
\includegraphics[width=0.70\textwidth]{Constraints-eliminate}
\caption{Example of aggregation of two tables on GPU.
\label{fig:eliminate}}
\end{figure}
Fig.~\ref{fig:eliminate} illustrates this process, where the variable $x_3$ is eliminated from the bucket-table $T_{i_k}$. The column being eliminated is highlighted yellow in the input bucket-table. The different row colors identify the unique assignments for the remaining variables $x_1, x_2$, and exposes the high degree of parallelization that is associated to such operation.
To exploit this level of parallelization, we adopt a paradigm similar to that employed in the aggregation operation on GPU, where each thread is responsible of the computation of a single output element.
\begin{procedure}[!thb]
\small{\caption{Gpu{::}Eliminate(${T}_{i_k}, x_i$) \label{alg:eliminate}} }
$r_{i_k} \leftarrow $ the thread's entry ID ($T_{id}$)\;
$r_{j} \leftarrow r_{i_k} \cdot |D_{x_i}|$ \tcc{holds the value of the index entry of $\chi_{i_k}$}
$c^* \leftarrow \chi_{i_k}[r_j]$\;
\For{$\ell \leftarrow (1 \ldots | D_{x_i} | \!-\! 1)$} {
$c^* \leftarrow \min\{c^*, \chi_{i_k}[r_j + \ell] \}$ \;
}
$\chi_{i_k}[r_{i_k}] \leftarrow c^*$\;
\end{procedure}
The variable elimination GPU kernel is described in Procedure \ref{alg:eliminate}, which is computed in parallel by a number of threads equal to the number of rows of the output bucket-table. Each thread identifies its row index $r_{i_k}$ within the output bucket-table cost values $\chi_{i_k}$ (line~1), given its thread ID. It hence sets an input row index $r_j$ to the value of the first $\chi_{i_k}$ input bucket-table row to analyze (line~1), and it stores in $c^*$ its associated cost.
Note that, as the variable to eliminate is listed last in the scope of the bucket-table, it is possible to retrieve each unique assignment for the projected output bucket table, simply by offsetting $r_{i_k}$ by the size of $D_{x_i}$. Additionally, all elements listed in $\chi_{i_k}[r_j], \ldots, \chi_{i_k}[r_j + |D_{x_i}|]$ differ exclusively on the value assignment to the variable $x_i$ (see Fig.~\ref{fig:eliminate}).
Thus, the GPU kernel evaluates the input bucket-table cost values associated to each element in the domain of $x_i$, by incrementing the row index $r_j$, $|D_{x_i}| -1$ times, and chooses the minimum cost value (lines~4--5).
At last, it saves to the associated output row the best cost found (line~6).
Note that each thread reads $|D_{x_i}|$ adjacent values of the vector $\chi_{i_k}$, and writes one value in the same vector. Thus, this algorithm \textbf{(1)} perfectly fits the SIMT paradigm,
\textbf{(2)} minimizes the accesses to the global memory as it encourages a coalesced data access pattern, and
\textbf{(3)} uses a relatively small amount of global memory, as it recycles the memory area allocated for the input bucket-table, to output the cost values for the output bucket-table.
The ordering $\prec_T$ adopted by the bucket-tables makes this procedure effective,
by forcing the variables to be eliminated to be always listed as last.
Additionally, we note that reordering the bucket-tables scope may be necessary exclusively when constructing the bucket-table associated to the constraints in $\setf{C}$. Indeed, the bucket-tables constructed by the algorithm preserve this ordering over their scope, since all the problem variables are processed according to the same ordering relation $\prec_T$, guaranteeing that the variables being eliminated are those with lower priority with respect to $\prec_T$. Therefore, no reordering will be required in the bucket functions during the process.
\smallskip
Finally, to reduce the memory transfer time, in addition to the technique described in the previous section, we unrolled
the for-loop in lines~7--9 of Algorithm 3.
Doing so allows us to process all the bucket-tables within a mini-bucket in a single GPU kernel and to copy them to the device using a single transaction.
\section{Theoretical Analysis}
\label{sec:theo}
We report below a theoretical analysis on the runtime and memory complexity of our GpuBE($z$) algorithms.
For the distributed case, we report results on the network load and messages size
complexity provided by the proposed algorithms.
The \emph{network load} and \emph{messages size} are defined, respectively, as the total number of messages exchanged by the agents and as the size of the largest message exchanged by the agents during problem resolution.
Since our algorithms rely on an inference-based procedure, the agent's complexity (i.e., the maximal number of operations performed by the agents while solving the problem) is equivalent to the size of the largest message exchanged. In turn, the latter corresponds to the memory complexity of the algorithm.
We use \emph{GpuBE}($w^*$) and \emph{GpuDBE}($w^*$) to denote our GPU versions of BE and DPOP, respectively,
and \emph{GpuBE}($z$) and \emph{GpuDBE}($z$) to denote our GPU versions of MBE and ADPOP, respectively, with mini-bucket size $z$.
\begin{theorem}
For a problem $P$, given an ordering $\prec_T$ on the \rev{primal graph} $G_P$,
the (mini-)bucket tables (resp.~\mbox{\it UTIL}\ messages) constructed by GpuBE($z$) (resp.~the GpuDBE($z$) agents) are identical to those constructed by (M)BE (resp.~the (A)DPOP agents), for $z \leq w^*$.
\end{theorem}
\begin{proof}
The proof follows from the observation that GpuBE($z$) and (M)BE are executed on the same induced graph $G_P^*$.
Thus, the problem variables are processed in the same order by both versions of the algorithms---lines 1 and 6 of Algorithm~\ref{alg:be} (1 and 8 of Algorithm~\ref{alg:mbe}) for (M)BE, and lines 2 and 11 of Algorithm~\ref{alg:gpu-dbe} for GpuBE($z$).
Analogously, in GpuDBE($z$) and (A)DPOP, agents operate on the same pseudo-tree ordering.
For the centralized case,
during the \emph{Variable Elimination Phase},
the bucket construction and mini-bucket partitioning operations of GpuBE($z$) (lines 3--4 of Algorithm~\ref{alg:gpu-dbe}) are identical to those of MBE (lines 2--3 of Algorithm~\ref{alg:mbe}).
For each mini-bucket $B_{i_j}$ in MBE, the operations to create the bucket function
$\hat{f}_{i_k}$ are identical in both algorithms:
the effect of invoking the $\textnormal{Gpu{::}Aggregate}( T_{i_k}, T_j )$ routine, in GpuBE, for each bucket-table $T_{i_k}$, corresponding to the bucket function $f_{i_k}$ (lines 7--9 of Algorithm~\ref{alg:gpu-dbe}), is analogous to the
aggregation operations performed in MBE:
$F=\sum_{f_j \in B_{i_k}} f_j $ (line 5 of Algorithm~\ref{alg:mbe}, in parenthesis),
and the effect of the $\textnormal{Gpu{::}Eliminate}( T_{i_k}, x_i )$ routine, which projects the variable $x_i$ onto the scope of $T_{i_j}$, produces the bucket function $\hat{f}_{i_k}$, which in turn correspond to the elimination operation performed by MBE:
$\pi_{-x_i}( F )$ (line 5 of Algorithm~\ref{alg:mbe}).
For the distributed cases,
both ADPOP and GpuDBE($z$) agents perform the same operations described above---during the \emph{\mbox{\it UTIL}\ Propagation Phase}---and populate the \mbox{\it UTIL}\ messages they send to their parent.
The equivalence between the Variable Elimination and \emph{\mbox{\it UTIL}\ Propagation Phases} of BE and DPOP, with the respective phases in GpuBE($w^*$) and GpuDBE($w^*$), respectively, follows from the process described above differing exclusively in that partitioning $B_i$ produces a single bucket with the same functions as those listed in $B_i$.
The operations performed during the \emph{Variable Assignment Phases} for (M)BE and GpuBE($z$) (lines 5-7, Algorithm~\ref{alg:be}, for BE , lines 8--9, Algorithm~\ref{alg:mbe}, for MBE, and lines 11--12, Algorithm~\ref{alg:gpu-dbe}, for GpuBE($z$)) are identical. Additionally, the variables are processed in the same order in both algorithms. Thus, the solution assignment for the problem variables returned by (M)BE and GpuBE are identical.
Similarly, for the distributed case, (A)DPOP and GpuDBE($z$) agents perform the same \mbox{\it VALUE}\ Propagation phase.\hfill$\Box$
\end{proof}
\begin{corollary}
\rev{
For a given $z \leq w^*$, the time and memory (message size) requirements of Gpu(D)BE($z$) are, in the worst case, in O($d^{z+1}$), and O($d^{z}$), respectively,
where $d = \max_{x_i \in \setf{X}} D_{x_i}$.}
\end{corollary}
\begin{proof}
This result follows from the equivalence of the \emph{Variable Elimination Phases} of (M)BE and GpuBE($z$), and of the
\emph{\mbox{\it UTIL}\ Propagation Phases} of (A)DPOP and GpuDBE($z$).
During these phases, the construction of the (mini)-buckets requires to save, in the worst case, all possible combinations for the value assignments of the bucket-function with bounded arity $z$. Thus, they require $O(d^z)$ space.
Similarly, for the distributed case, due to the equivalence of (A)DPOP and GpuDBE($z$), the largest message exchanged by the agents has size $O(d^z)$.
Additionally, the total amount of operations (or, equivalently, bucket-tables rows) that can be processed in parallel during the GPU-based Constraint Aggregation and GPU-based Variable Elimination steps,
is bounded by a constant value which depends on the GPU card characteristic. Thus, the time complexity of GpuDBE($z$) is in O$(d^{z+1}$). \hfill$\Box$
\end{proof}
\begin{corollary}
The network load required for GpuDBE($z$) is equivalent to the network load required by (A)DPOP.
\end{corollary}
\begin{proof}
This result follow from the equivalence of DPOP with GpuDBE($w^*$) and ADPOP($z$) with GpuBE($z$) (Theorem 1).
Since (A)DPOP requires each agent to send one \mbox{\it UTIL}\ message to its parent and one \mbox{\it VALUE}\ message to each of its children, there are a total of $n-1$ \mbox{\it UTIL}/\mbox{\it VALUE}\ messages exchanged---one through each tree-edge of the pseudo-tree $T_P$. Thus, the network load required by (A)DPOP and GpuDBE($z$) is in $O(n)$. \hfill$\Box$
\end{proof}
\begin{corollary}
Gpu(D)BE is correct and complete.
\end{corollary}
\begin{proof}
The correctness and completeness of GPU-(D)BE($w^*$) follow from the correctness and completeness of BE \cite{Dechter:99} and DPOP~\cite{petcu:05}, and Theorem 1.\hfill$\Box$
\end{proof}
\section{Experimental Results}
\label{sec:results}
In this section, we evaluate our GPU implementations of BE and MBE (GpuBE) as well as our GPU implementations of DPOP and ADPOP (GpuDBE) and compare them with their CPU counterparts.\footnote{Our source code is available at \url{https://github.com/nandofioretto/GpuBE}, and \url{https://github.com/nandofioretto/GpuDBE}}
Experiments for GpuDBE and (A)DPOP are conducted using a multi-agent DCOP simulator that simulates the concurrent activities of multiple agents, whose actions are activated upon receipt of a message.
All algorithms use the same variable ordering in the centralized case and pseudo-tree in the distributed case.
Performance of the centralized algorithms are evaluated using the algorithms' wallclock runtime, while the performance of distributed algorithms are evaluated using the \emph{simulated runtime} metric~\cite{sultanik:07}. The main focus of the evaluation is on runtime and speedup achieved by the GPU implementations with respect to their CPU counterparts.
Additionally, to compare the quality of the solution bounds reported by the incomplete algorithms, we also report the best solution quality found within the given time limits by \emph{toulbar2} \cite{allouchetoulbar2}, an optimized, exact centralized solver for WCSPs. Toulbar2 is a state-of-the-art solver that uses a depth-first branch-and-bound process to identify a minimum cost assignment and employs the notion of \emph{soft local consistency} to prune the search space using the problem lower bound.
Our experiments are conducted on an \emph{AMD Opteron 6276} with a 2.3GHz CPU and is equipped with a GPU device \emph{GeForce GTX TITAN} with $14$ multiprocessors, 2688 cores, with a clock rate of 837MHz, and 6GB of global memory.
We performed our experiments on both randomly generated instances on different networks topologies and on standard WCSP benchmarks.\footnote{Downloadable from \url{http://costfunction.org/en/benchmark/} and \url{http://graphmod.ics.uci.edu/group/Repository}}
We first analyze the runtimes of the CPU and GPU versions of BE and DPOP on randomly generated instances, where we report the runtimes and lower bounds of the GPU and CPU versions of MBE and ADPOP at varying of the bucket size $z$.
Then, to ensure that the speedups are not due to a specific GPU device configuration, we compare the CPU and GPU speedups achieved on $3$ distinct GPU architectures, characterized by different clock rates, number of SMs, and memory sizes.
Finally, we report the solving time and lower bounds of our GpuBE on an extensive set of WCSP benchmarks to verify the generality of the speedups across different domains.
Each solver has 1-hour timeout of wallclock time in the centralized case and a 1-hour timeout of simulated time in the distributed case. Additionally, they have a memory limit of 32GB to solve each problem instance.
Results are averaged over all instances.
If a solver fails to solve an instance is due to either memory limits (labeled \emph{oom}) or timeout (labeled \emph{oot}).
\subsection{Binary Random Networks}
The instances for each binary network topology are generated as follows:
\bitemize
\item {\bf Random:} We create an $n$-node network, whose density $p_1$ produces $\lfloor n\,(n-1)\,p_1 \rfloor$ edges in total. We do not bound the tree-width, which is based on the underlying graph and randomly generated.
\item {\bf Scale-free:} We create an $n$-node network based on the \emph{Barabasi-Albert model}~\cite{barabasi:99}. Starting from a connected $2$-node network, we repeatedly add a new node, randomly connecting it to two existing nodes. In turn, these two nodes are selected with probabilities that are proportional to the numbers of their connected edges. The total number of edges is $2\,(n-2)+1$.
\item {\bf Grid:} We create an $n$-node network arranged as a rectangular grid, where each internal node is connected to four neighboring nodes, while nodes on the grid perimeter are connected to three neighboring nodes unless they are at the corner of the grid, in which case they are connected to two neighboring nodes.
\end{list}
We generate $50$ instances for each topology, ensuring that the underlying graph is connected. The cost functions are generated using random integer costs in $[0, 100]$, and the constraint tightness (i.e., ratio of entries in the cost table that have a cost of $\infty$) $p_2$ is set to $0.5$ for all experiments.
We set the following as default parameters: For the random and scale-free topology, $n \!=\! 10$, $d \!=\! \max_{D_i \in \setf{D}} \size{D_i} \!=\! 10$, and $p_1 \!=\! 0.3$, and for the grid topology, $\sqrt{n} \!=\! 10$.
\begin{table}
\centering
\resizebox{0.9\textwidth}{!}
\begin{tabular}{|*{4}{c} || *{3}{r}| *{3}{r} |
\hline
\multicolumn{4}{|c||}{Problem} & \multicolumn{3}{| c |}{BE} & \multicolumn{3}{| c |}{DPOP}\\
$n$ & $d$ & $p_1$ & $w^*$ & CPU & GPU & speedup & CPU & GPU & speedup\\% & NC & AC & EDAC \\
\hline\hline
10 & 10 & 0.3 & 2.9 & 0.019 & 0.002 & 10.5 & 0.007 & 0.001 & 7.20\\% & 0.049 & 0.009 & 0.009 \\
11 & 10 & 0.3 & 3.2 & 0.031 & 0.002 & 13.6 & 0.013 & 0.001 & 13.0\\% & 0.115 & 0.011 & 0.010 \\
12 & 10 & 0.3 & 3.6 & 0.069 & 0.003 & 25.7 & 0.028 & 0.001 & 28.5\\% & 0.185 & 0.021 & 0.009 \\
13 & 10 & 0.3 & 4.3 & 0.413 & 0.005 & 79.4 & 0.210 & 0.002 & 116\\% & 0.638 & 0.058 & 0.015 \\
14 & 10 & 0.3 & 4.4 & 0.631 & 0.006 & 98.6 & 0.214 & 0.002 & 134\\% & 1.141 & 0.110 & 0.017 \\
15 & 10 & 0.3 & 5.3 & 4.190 & 0.026 & 158 & 1.609 & 0.009 & 187\\% & 2.188 & 0.181 & 0.025 \\
16 & 10 & 0.3 & 5.8 & 32.29 & 0.189 & 171 & 9.848 & 0.049 & 202\\% & 17.21 & 0.352 & 0.046 \\
17 & 10 & 0.3 & 6.4 & 65.41 & 0.328 & 200 & 28.14 & 0.138 & 204\\% & 11.43 & 0.469 & 0.077 \\
18 & 10 & 0.3 & 7.5 & 206.1 & 0.944 & 218 & 103.0 & 0.483 & 213\\% & 63.76 & 1.505 & 0.153 \\
19 & 10 & 0.3 & 8.0 & 602.1 & 2.541 & 237 & 470.2 & 2.019 & 233\\% & 174.2 & 3.261 & 0.184 \\
20 & 10 & 0.3 & 8.5 & 675.3 & 3.145 & 215 & 508.9 & 2.160 & 236\\% & 284.7 & 9.371 & 0.537 \\
\hline
10 & 5 & 0.3 & 3.0 & 0.001 & 0.002 & 0.56 & 0.001 & 0.001 & 0.80\\% & 0.022 & 0.008 & 0.006 \\
10 & 10 & 0.3 & 2.9 & 0.019 & 0.002 & 10.5 & 0.007 & 0.001 & 7.20\\% & 0.049 & 0.009 & 0.009 \\
10 & 25 & 0.3 & 2.8 & 0.227 & 0.004 & 55.4 & 0.092 & 0.001 & 92.3\\% & 0.198 & 0.021 & 0.014 \\
10 & 50 & 0.3 & 2.9 & 24.81 & 0.095 & 262 & 13.99 & 0.048 & 291\\% & 0.635 & 0.091 & 0.047 \\
10 & 100 & 0.3 & 2.9 & 67.59 & 0.220 & 308 & 35.22 & 0.118 & 299\\% & 1.550 & 0.172 & 0.173 \\
\hline
10 & 10 & 0.2 & 2.0 & 0.001 & 0.001 & 0.62 & 0.001 & 0.001 & 0.94\\% & 0.018 & 0.007 & 0.008 \\
10 & 10 & 0.3 & 2.9 & 0.019 & 0.002 & 10.5 & 0.007 & 0.001 & 7.20\\% & 0.049 & 0.009 & 0.009 \\
10 & 10 & 0.4 & 3.8 & 0.094 & 0.002 & 40.7 & 0.042 & 0.001 & 42.5\\% & 0.149 & 0.017 & 0.008 \\
10 & 10 & 0.5 &4.5 & 0.525 & 0.005 & 105 & 0.234 & 0.002 & 130\\% & 0.444 & 0.042 & 0.019 \\
10 & 10 & 0.6 & 5.4& 3.378 & 0.019 & 176 & 1.941 & 0.011 & 176\\% & 0.597 & 0.068 & 0.024 \\
10 & 10 & 0.7 & 5.9& 14.86 & 0.072 & 205 & 10.00 & 0.053 & 189\\% & 1.003 & 0.116 & 0.043 \\
10 & 10 & 0.8 & 6.7& 56.23 & 0.246 & 228 & 31.29 & 0.147 & 213\\% & 1.513 & 0.203 & 0.093 \\
10 & 10 & 0.9 &7.6 & 72.32 & 0.312 & 232 & 42.47 & 0.201 & 211\\% & 3.264 & 0.313 & 0.195 \\
\hline
\end{tabular}
\caption{Random networks. \label{tab:be_rand}}
\end{table}
\begin{table}
\centering
\resizebox{0.9\textwidth}{!}
\begin{tabular}{|*{3}{c} || *{3}{r}| *{3}{r} |
\hline
\multicolumn{3}{|c||}{Problem} & \multicolumn{3}{| c |}{BE} & \multicolumn{3}{| c |}{DPOP}\\
$n$ & $d$ & $w^*$ & CPU & GPU & speedup & CPU & GPU & speedup \\% & NC & AC & EDAC \\
\hline\hline
10 & 10 & 6.3 & 22.99 & 0.111 & 207 & 13.78 & 0.064 & 215\\% & 1.038 & 0.146 & 0.063 \\
11 & 10 & 6.0 & 25.57 & 0.120 & 212 & 13.21 & 0.057 & 231\\% & 1.451 & 0.148 & 0.065 \\
12 & 10 & 6.0 & 27.96 & 0.132 & 212 & 14.60 & 0.072 & 203\\% & 1.621 & 0.176 & 0.053 \\
13 & 10 & 5.9 & 80.14 & 0.370 & 217 & 36.21 & 0.174 & 208\\% & 4.994 & 0.326 & 0.084 \\
14 & 10 & 6.9 & 78.36 & 0.339 & 231 & 32.50 & 0.145 & 223\\% & 16.08 & 0.675 & 0.166 \\
15 & 10 & 8.2 & 189.4 & 0.887 & 213 & 66.86 & 0.340 & 197\\% & 58.23 & 2.076 & 0.341 \\
16 & 10 & 9.2 & oom & oom & - & oom & oom & - \\% & 132.6 & 4.260 & 0.673 \\
17 & 10 & 9.5 & oom & oom & - & oom & oom & - \\% & 293.2 & 9.223 & 0.811 \\
18 & 10 & 10 & oom & oom & - & oom & oom & - \\% & 1249.1 & 50.17 & 2.768 \\
19 & 10 & 11 & oom & oom & - & oom & oom & - \\% & 2732.3 & 125.6 & 6.015 \\
20 & 10 & 12 & oom & oom & - & oom & oom & - \\% & 3570.4 & 315.5 & 13.62 \\
\hline
10 & 5 & 6.8 & 0.322 & 0.004 & 74.8 & 0.175 & 0.001 & 145\\% & 0.184 & 0.024 & 0.015 \\
10 & 10 & 6.3 & 22.99 & 0.111 & 207 & 13.78 & 0.064 & 215\\% & 1.038 & 0.146 & 0.063 \\
10 & 25 & 6.6 & 242.5 & 0.888 & 273 & 127.9 & 0.593 & 216\\% & 38.37 & 5.249 & 2.274 \\
10 & 50 & 6.4 & oom & oom & - & oom & oom & - \\%& 227.2 & 48.82 & 15.23 \\
10 & 100 & 6.4 & oom & oom & - & oom & oom & - \\%& 1735.5 & 1161.9 & 363.9 \\
\hline
\end{tabular}
\caption{Scale-free networks. \label{tab:be_scalefree}}
\medskip
\resizebox{0.9\textwidth}{!}
\begin{tabular}{|*{3}{c} || *{3}{r}| *{3}{r} |
\hline
\multicolumn{3}{|c||}{Problem} & \multicolumn{3}{| c |}{BE} & \multicolumn{3}{| c |}{D-BE}\\
$\sqrt{n}$ & $d$ & $w^*$ & CPU & GPU & speedup & CPU & GPU & speedup \\% & NC & AC & EDAC \\
\hline\hline
5 & 10 & 3.3 & 0.259 & 0.005 & 44.7 & 0.022 & 0.001 & 21.8\\% & 94.31 & 3.140 & 0.010 \\
6 & 10 & 3.7 & 0.267 & 0.008 & 33.4 & 0.022 & 0.001 & 22.6\\% & 1123.2 & 17.29 & 0.028 \\
7 & 10 & 3.7 & 0.515 & 0.012 & 42.9 & 0.037 & 0.001 & 30.8\\% & oot & 1034.8 & 0.316 \\
8 & 10 & 3.6 & 0.848 & 0.018 & 47.1 & 0.041 & 0.001 & 29.6\\% & oot & 3517.4 & 25.57 \\
9 & 10 & 3.9 & 1.460 & 0.028 & 52.0 & 0.049 & 0.001 & 33.7\\% & oot & 3428.6 & 362.3\\
10 & 10 & 4.0 & 1.881 & 0.035 & 34.4 & 0.054 & 0.002 & 31.7\\% & oot & oot & 1668.6 \\
11 & 10 & 3.7& 1.934 & 0.040 & 48.3 & 0.073 & 0.002 & 38.4\\% & oot & oot & oot \\
12 & 10 & 3.8 & 2.174 & 0.042 & 48.4 & 0.089 & 0.002 & 38.7\\%& oot & oot & oot \\
13 & 10 & 3.9 & 2.430 & 0.045 & 42.2 & 0.102 & 0.003 & 37.8\\% & oot & oot & oot \\
14 & 10 & 4.0 & 2.996 & 0.055 & 54.5 & 0.127 & 0.003 & 39.7\\% & oot & oot & oot \\
15 & 10 & 4.0 & 3.785 & 0.071 & 53.3 & 0.151 & 0.004 & 38.7\\% & oot & oot & oot \\
\hline
10 & 5 & 3.7 & 0.043 & 0.020 & 2.21 & 0.001 & 0.001 & 1.00\\% & oot & 3594 & 4.541 \\
10 & 10 & 4.0 & 1.881 & 0.035 & 34.4 & 0.054 & 0.002 & 31.7\\% & oot & oot & 1668.6 \\
10 & 25 & 3.9 & 97.29 & 0.388 & 251 & 2.930 & 0.011 & 266\\% & oot & oot &oot \\
10 & 50 & 4.0 & oom & oom & - & oom & oom & - \\% & oot & oot & oot\\
10 & 100 & 4.0 & oom & oom & - & oom & oom & - \\% & oot & oot & oot \\
\hline
\end{tabular}
\caption{Grid networks. \label{tab:be_grid}}
\end{table}
Tables \ref{tab:be_rand}--\ref{tab:be_grid} report the runtime, in seconds, for random, scale-free, and grid topologies, respectively, varying the number of variables (resp.~agents) for the centralized (resp.~distributed) algorithms, the size of the variables domains, and the constraint tightness of the \rev{primal graph}.
The first four (three) columns of Table \ref{tab:be_rand}, (\ref{tab:be_scalefree} and \ref{tab:be_grid}) describe the problem setting adopted for each experiment. The induced width $w^*$ is averaged across all instances. All other columns report the average runtime and GPU vs.~CPU speedup in parenthesis.
We make the following observations:
\bitemize
\item
The GPU-based inference-algorithms are consistently faster that their CPU counterparts, with speedups of up to 307x.
Only two exceptions arise for the random networks, where in the small instances with $n=10$, $d=5$, $p_1=0.3$, and $n=10$, $d=10$, $p_1=0.2$, the GPU versions of the algorithms are slower than their CPU counterparts.
\item
The speedup increases with the problem size. In particular, the speedup increases with increasing induced width and with increasing domain size of the problem variables. Both these factors influence the size of the bucket-tables to be processed.\footnote{Recall that BE needs to process bucket-tables whose number of rows is in $O(d^{w^*})$.}
This observation corroborates the effectiveness of the GPU parallelism exploited in the construction of these tables.
\item
As expected, the inference-based algorithms are unable to process instances characterized by large induced widths
or large domain sizes,
as the size of the bucket-tables become intractable with the memory limitations.
This is evident in the scale-free and grid networks, where the solvers run out of memory for instances with $n \geq 16$ and $d \geq 50$, respectively.
\item
The simulated runtimes of the DCOP algorithms are consistently smaller than the wallclock runtimes of the WCSPs ones.
This is due to the fact that agents in different branches of the pseudo-tree can compute their bucket-tables independently from each other.
\item
Finally, the speedup trends of the distributed algorithms are similar to those of the centralized algorithms.
\end{list}
\begin{figure}[!t]
\centering
\includegraphics[width=0.7\textwidth]{Constraints-AggrElim}
\caption{Analysis of the average speedup obtained by the GPU-based constraint aggregation, and GPU-based variable elimination w.r.t.~their CPU-based counterparts in the random, scale-free, and grid network instances. \label{fig:funcSpeed}}
\end{figure}
Next, we analyze the performance of the individual kernels that implement the constraint aggregation and the variable elimination processes described, respectively, in Sections~\ref{sec:aggregation} and~\ref{sec:elimination}.
Figure~\ref{fig:funcSpeed} illustrates the average speedup obtained by the GPU-based constraint aggregation, and the GPU-based variable elimination with respect to their CPU-based counterparts when considering the largest bucket processed in each instance of the random, scale-free, and grid network instances.
The reported average speedup for the constraint aggregation operations range from 363x (in scale free networks) to 613x (in random networks).
The variable elimination operations achieve an even higher speedup, ranging from 830x (for grid networks) to 911x (for scale free networks). This is due to the high locality of data exploited by the GPU-based variable elimination kernel, which encourages coalesced data accesses, and through memory reuse, where we overwrite the input bucket-table of the variable elimination process with the resulting bucket-table from the same process.
\begin{figure}[!t]
\centering
\includegraphics[width=0.47\textwidth]{Constraints-RandomMiniBE}
\includegraphics[width=0.47\textwidth]{Constraints-ScaleFreeMiniBE}\\
{\small (a) \hspace{150pt} (b)}\\
\includegraphics[width=0.47\textwidth]{Constraints-GridMiniBE}\hspace{18pt}
\includegraphics[width=0.44\textwidth]{Constraints-memTransfers}\\
\caption{
(a)~MBE Results on Random Networks: $n = 20, d = 25, p_1 =0.3$;
(b)~MBE results on Scale-free Networks: $n = 20, d = 25$;
(c)~MBE results on Grid Networks: $\sqrt{n} = 10, d = 25$;
(d)~Normalized data allocation and transfer times (blue) vs.~kernel times (white) on different GPUs.
\label{fig:res_mbe1}}
\end{figure}
Next, we compare our centralized and distributed versions of GpuBE with MBE~\cite{Dechter:99} and ADPOP~\cite{petcu:05} at varying of the mini-bucket size $z \in \{2,\ldots, 10\}$, on binary constraint networks with \emph{random}, \emph{scale-free}, and \emph{grid} topologies, using the same settings described in the previous section. The instances for each topology are generated as described above.
Fig.~\ref{fig:res_mbe1}(a--c) illustrate the speedup of the CPU and GPU versions of MBE, respectively, on random networks with $n = 20$, $d = 25$, $p_1 =0.3$, on scale-free networks with $n = 20$, $d = 25$, and on grid networks with $\sqrt{n} = 10, d = 25$.
The intensity of the color illustrates the solution quality of the bound returned (darker color denotes better solution quality).
We make the following observations:
\bitemize
\item
The speedup obtained by the GPU vs.~CPU solvers increases as the size of the mini-buckets increases. This observation is consistent with the previous observation that the speedup increases with increasing induced widths.
\item
The speedup saturates when $z=7$ in all benchmarks, reporting maximal speedups of 235\textsl{x}, 274\textsl{x}, and 156\textsl{x}, for random, scale-free, and grid networks, respectively.
This phenomena occurs when the maximum concurrent number of GPU threads are scheduled and executed simultaneously by all the GPU SMs---i.e., when there is enough work to saturate the GPU maximal occupancy.
\item
As for the previous experiment, the speedup trends of the distributed algorithms are similar to those of the centralized algorithms. The correlation\footnote{We use the \emph{Pearson product-moment correlation} coefficient.} of the CPU vs.~GPU speedup between the centralized and the distributed solutions are 0.93, 0.95, and 0.99, respectively for the grid, random, and scale-free network topologies.
\end{list}
\begin{table}
\centering
\resizebox{\textwidth}{!}
\begin{tabular}{ l l l l }
\hline
& \bemph{TESLA M2075} & \bemph{GeForce GTX Titan} & \bemph{GeForce GTX Titan X} \\[2pt]
\hline
CUDA Capability & 2.0 & 3.5 & 5.2 \\
Global Memory Size & 5375 MB & 6137 MB & 12286 MB \\
Number of SMs & 14 & 14 & 24 \\
Cores per SM & 32 & 192 & 128 \\
GPU Max Clock Rate & 1.15 GHz & 0.88 GHz & 1.08 GHz \\
Memory Clock Rate & 1566 Mhz & 3004 Mhz & 3505 Mhz \\
L2 Cache Size & 786 KB & 1572 KB & 3145 KB \\
Max Number of Threads per SM & 1536 & 2048 & 2048 \\
Concurrent copy and execution & yes & yes & yes \\
\hline
\end{tabular}
}
\caption{GPU device specifics. \label{tab:gpus}}
\end{table}
\begin{table}
\centering
\resizebox{\textwidth}{!}
\begin{tabular}{ c | *{3}{r} | *{3}{r} | *{3}{r}}
\hline
& \multicolumn{3}{c |}{Grid} & \multicolumn{3}{c |}{Random} & \multicolumn{3}{c}{Scale-Free}\\
z & TESLA & Titan & Titan X & TESLA & Titan & Titan X & TESLA & Titan & Titan X\\
\hline
2 & 0.22 & 0.85 & 0.49 & 0.21 & 0.28 & 0.60 & 0.27 & 0.43 & 0.57\\
3 & 1.54 & 3.05 & 2.80 & 1.42 & 1.99 & 3.25 & 1.48 & 2.42 & 1.61\\
4 & 21.9 & 29.1 & 33.0 & 10.7 & 11.6 & 16.1 & 12.2 & 7.46 &12.0\\
5 & 117 & 150 & 232 & 60.6 & 49.8 & 66.4 & 59.3 & 51.3 & 66.8\\
6 & 117 & 143 & 237 & 144 & 145 & 223 & 163 & 159 & 285\\
7 & 117 & 152 & 235 & 198 & 207 & 392 & 208 & 244 & 435\\
8 & 118 & 153 & 241 & 198 & 235 & 645 & 211 & 274 & 627\\
9 & 115 & 156 & 238 & 197 & 234 & 620 & oom & oom & oom\\
10 & 117 & 155 & 235 & 199 & 233 & 628 & oom & oom & oom\\
\hline
\end{tabular}
}
\caption{CPU vs.~GPU speedup on different GPU devices. \label{tab:gpusPerformance}}
\end{table}
\smallskip
Table \ref{tab:gpusPerformance} illustrates a comparison of the speedups obtained with three different GPU hardware configurations: \emph{TESLA M2075}, \emph{GeForce GTX Titan} and \emph{GeForce GTX Titan X}, whose specifics are summarized in Table~\ref{tab:gpus}.\footnote{In all other experiments we used the GeForce GTX Titan, as this is the best, most affordable card at our disposal.}
Among the three GPU devices, the TESLA M2075 achieve the lowest maximal speedups, which range from 117.8\textsl{x} to 213,7\textsl{x}.
Additionally, the speedup saturates when $z=5$ for grid networks and $z=7$ for random and scale-free networks.
This is due to the fact that this card can schedule the smallest number of cores per each SMs (32). Since each core can run concurrently a wrap (32 threads), its maximal level of concurrency is $14 \times 32 \times 32 = 14,336$ threads (and is thus the maximum number of parallel aggregation operations).
In contrast, the speedup obtained by our GpuBE is the highest on the GeForce GTX Titan X---obtaining a maximal speedup of 646.9\textsl{x}---and saturates when $z=8$ in all networks.
The maximum number of threads that can run concurrently on this card is $24 \times 128 \times 32 = 98,304$.
The speedups obtained by our solver on the GeForce GTX Titan, used in the rest for the experiments in this paper, are larger than those obtained on the TESLA but smaller than those obtained on the GeForce GTX Titan X. This card can run up to $86,016$ threads. In addition to the number of threads than can run concurrently, the GPU clock rate and L2 cache size play a substantial role in the GPU performance.
Finally, Fig.~\ref{fig:res_mbe1}(d) illustrates the time spent by the GPU devices while executing the kernel functions (in white) in contrast to the time used for memory transfers and allocations (in blue),
at varying mini-bucket size $z=\{4,6,8\}$.
These times are averaged among all instances for the three network topologies examined and are normalized with the respect to the wallclock runtime.
The results show that the time spent by the device in performing actual computations increases, with the respect to the memory transfer time, as the mini-bucket size increase. Allocations and memory transfers on the Titan device are slower than on the TESLA and the GTX Titan X. Finally, these times account for the 36\% to 55\%, 18\% to 34\%, and 8\% to 18\% of the total time, respectively for the mini-bucket sizes $4, 6$, and $8$.
\subsection{WCSPs Benchmarks}
We now report the evaluation of our GpuBE on the following standard WCSPs benchmarks:
\bitemize
\item \emph{Celar}: Radio link frequency assignment problems.
\item \emph{Coloring}: Graph coloring instances cast into minimum coloring instances.
\item \emph{Iscas89}: WCSPs derived from digital circuits.
\item \emph{Pedigree}: Instances from the genetic linkage analysis domain that is associated with the task of haplotyping.
\item \emph{Spot}: Instances of the daily photograph scheduling problem of Earth observation satellites.
\end{list}
Tables \ref{tab:celar}--\ref{tab:pedigree},
tabulate the results for the above benchmarks.
In each table and for each instance, we report, in order, the instance name---as appearing in the original benchmark---the number of variables $n$ of the problem, the maximum size of their domains $d$, the number of constraints $c$, the graph density $p_1$, and the induced width $w^*$ of the underlying \rev{primal graph}.
In each table, the top row shows the runtimes in seconds of GpuBE(z) at varying bucket size $z$ and GpuBE.
The bottom row shows the returned solutions' qualities, where for GpuBE(z), we report the lower bound it returned.
When GpuBE failed to report a solution (due to memory limits), we report the solution quality found by toulbar2 (shown in parenthesis) or a dash symbol, if toulbar2 did not terminate within the time limit.
The speedup of GpuBE(z) and GpuBE w.r.t. their CPU counterparts are shown in parentheses.
For each instance, we vary the bucket size $z$ from $2$ to $20$, and report the minimum bucket size $z_{\min}$, which is the largest constraint arity of the instance, the maximum bucket size $z_{\max} = \min\{w^z, 20\}$, where
$w^z$ is defined as the maximal bucket size that can be processed within the hardware memory limits, and the intermediate bucket sizes $z_2 = z_{\min} + \frac{1}{3} (z_{\max} - z_{\min})$ and $z_3 = z_{\min} + \frac{2}{3} (z_{\max} - z_{\min})$.
Consistent with our previous observations, the algorithms' speedups and solution qualities increase as the bucket size increases.
Additionally, for several large problems instances (e.g., \textit{scen06-24reduc}---\textit{scen06reduc} in the Celar benchmark), our GPU implementation of MBE can report good lower bounds quickly (within a few seconds), whereas solving the entire problem with the most competitive soft consistency technique in toulbar2 requires from 6 to 48 minutes.
For other large instances (e.g., in the Spot benchmark), we observe that toulbar2 ran out of time for the majority of the instances, while our GpuBE(z) can quickly find lower bounds, which could be used in a AND-and-OR search type as proposed by Marinescu and Dechter~\cite{marinescu:09}.
\iffalse
\fi
\begin{table}
\begin{center}
\resizebox{0.9\textwidth}{!}
\begin{tabular}{| l | p{6pt} p{2pt} p{12pt} p{6pt} p{6pt} || *{5}{r} |
\hline
\prob{Problem}{$n$}{$d$}{$c$}{$p_1$}{$w^*$} & \multicolumn{4}{|c}{GpuBE($z$)} & GpuBE \\% & \multicolumn{3}{|c|}{toulbar2} \\
\prob{}{}{}{}{}{} & $z_{\min}$ & $z_2$ & $z_3$ & $z_{\max}$ & $w^*$\\% & NC & AC & EDAC \\
\hline\hline
\prob{ CELAR6-SUB0 }{16}{44}{207}{0.47}{7} & 0.116 (1.26x) & 0.116 (13.5x) & 0.31 (182x) & 0.31 (182x) & oom \\% & 15.63 & 0.28 & 0.06 \\
& \multicolumn{5}{|c||}{} & 0 & 13 & 13 & 13 & (159) \\% & & & 159 \\
\hline
\prob{ CELAR7-SUB1-20 }{14}{20}{300}{0.98}{9} & 0.111 (3.37x) & 0.129 (42.8x) & 0.451 (215x) & 5.1 (313x) & oom \\% & 6.22 & 0.74 & 0.3 \\
& \multicolumn{5}{|c||}{} & 20102 & 40931 & 71023 & 81433 & (132538) \\% & & & 132538 \\
\hline
\prob{ CELAR6-SUB1-24 }{14}{24}{300}{0.82}{9} & 0.188 (0.59x) & 0.122 (4.81x) & 0.161 (71.8x) & 0.837 (279x) & oom \\% & 166.8 & 5.39 & 2.55 \\
& \multicolumn{5}{|c||}{} & 0 & 280 & 598 & 772 & (2656) \\% & & & 2656 \\
\hline
\prob{ CELAR7-SUB0 }{16}{44}{188 }{0.66}{9} & 0.173 (0.63x) & 0.157 (10.1x) & 0.317 (187x) & 0.317 (187x) & oom \\% & 25.06 & 0.14 & 0.03 \\
& \multicolumn{5}{|c||}{} & 0 & 104 & 10001 & 10001 & (10310)\\% & & & 10310 \\
\hline
\prob{ CELAR6-SUB1 }{14}{44}{300 }{0.82}{9} & 0.201 (1.25x) & 0.317 (10.6x) & 0.723 (171x) & 0.723 (171x) & oom \\% & 2726.3 & 13.74 & 10 \\
& \multicolumn{5}{|c||}{} & 0 & 308 & 626 & 626 & (2669) \\%& & & 2669 \\
\hline
\prob{ CELAR7-SUB1 }{14}{44}{300}{0.82}{9} & 0.295 (1.03x) & 0.316 (10.5x) & 0.593 (210x) & 0.593 (210x) & oom \\% & 465.6 & 10.73 & 2.46 \\
& \multicolumn{5}{|c||}{} & 0 & 21523 & 51123 & 51123 & (142640)\\% & & & 142640 \\
\hline
\prob{ CELAR6-SUB2 }{16}{44}{353}{0.77}{10} & 0.254 (1.3x) & 0.328 (12.6x) & 0.82 (188x) & 0.82 (188x) & oom \\% & oot & 26.31 & 11.38 \\
& \multicolumn{5}{|c||}{} & 0 & 231 & 387 & 387 & (2746)\\% & & & 2746 \\
\hline
\prob{ CELAR7-SUB2 }{16}{44}{353}{0.77}{10} & 0.407 (0.54x) & 0.254 (16.4x) & 0.95 (164x) & 0.95 (164x) & oom \\%& oot & 103.1 & 20.9 \\
& \multicolumn{5}{|c||}{} & 0 & 20420 & 40931 & 40931 & (173252) \\% & & & 173252 \\
\hline
\prob{ CELAR6-SUB3 }{18}{44}{421}{0.71}{10} & 0.375 (1.08x) & 0.304 (16.1x) & 0.975 (188x) & 0.975 (188x) & oom \\% & oot & 101.1 & 31.58 \\
& \multicolumn{5}{|c||}{} & 0 & 265 & 452 & 452 & (3079)\\% & & & 3079 \\
\hline
\prob{ CELAR7-SUB3 }{18}{44}{421}{0.71}{10} & 0.449 (0.97x) & 0.46 (10.8x) & 1.002 (178x) & 1.002 (178x) & oom \\% & oot & 827.1 & 88.94 \\
& \multicolumn{5}{|c||}{} & 0 & 20615 & 51434 & 51434 & (203460) \\% & & & 203460 \\
\hline
\prob{ scen06-30reduc }{81}{14}{399}{0.11}{10} & 0.071 (2.65x) & 0.068 (10.3x) & 0.325 (155x) & 1.738 (278x) & oom \\% & 2825.4 & 22.21 & 2.02 \\
& \multicolumn{5}{|c||}{} & 285 & 690 & 975 & 1447 & (2080)\\% & & & 2080 \\
\hline
\prob{ scen06-30 }{99}{14}{1178}{0.09}{10} & 0.19 (1.5x) & 0.189 (14.2x) & 1.313 (238x) & 13.3 (345x) & oom \\%& oot & 1200.6 & 13.77 \\
& \multicolumn{5}{|c||}{} & 450 & 411 & 1201 & 1100 & (2080) \\% & & & 2080 \\
\hline
\prob{ CELAR6-SUB4-20 }{22}{20}{477}{0.82}{11} & 0.197 (4.09x) & 0.232 (49.6x) & 1.021 (225x) & 11.29 (344x) & oom \\% & 95.21 & 4.13 & 1.53 \\
& \multicolumn{5}{|c||}{} & 494 & 598 & 732 & 1359 & (2716) \\% & & & 2716 \\
\hline
\prob{ CELAR7-SUB4-22 }{22}{22}{473}{0.67}{11} & 0.221 (0.69x) & 0.211 (3.54x) & 0.158 (82.4x) & 0.922 (286x) & oom \\% & 786.1 & 87.02 & 14.21 \\
& \multicolumn{5}{|c||}{} & 0 & 40104 & 60214 & 31530 & (202342) \\% & & & 202342 \\
\hline
\prob{ CELAR6-SUB4reduc }{20}{44}{149}{0.77}{11} & 0.106 (2.01x) & 0.213 (10.9x) & 0.357 (241x) & 0.357 (241x) & oom \\%& oot & 14.21 & 14.21 \\
& \multicolumn{5}{|c||}{} & 0 & 44 & 283 & 283 & (202342) \\% & & & 202342 \\
\hline
\prob{ CELAR6-SUB4 }{22}{44}{477 }{0.65}{1} & 0.387 (0.73x) & 0.343 (17.4x) & 1.013 (229x) & 1.013 (229x) & oom \\% & oot & 316 & 79.59 \\
& \multicolumn{5}{|c||}{} & 0 & 170 & 405 &405 & (3230) \\% & & & 3230 \\
\hline
\prob{ CELAR7-SUB4 }{22}{44} {477}{0.65}{1} & 0.347 (0.82x) & 0.344 (18.1x) & 1.24 (188x) & 1.24 (188x) & oom \\% & oot & 2857.9 & 419.8 \\
& \multicolumn{5}{|c||}{} & 0 & 30118 & 31442 & 31442 & (242443) \\% & & & 242443 \\
\hline
\prob{ scen06-24reduc }{81}{20}{403}{0.12}{12} & 0.099 (4.76x) & 0.101 (57.9x) & 0.375 (217x) & 4.001 (303x) & oom \\% & oot & oot & 294.2 \\
& \multicolumn{5}{|c||}{} & 278 & 599 & 634 & 1411 & (2857)\\% & & & 2857 \\
\hline
\prob{ scen06-22reduc }{81}{22}{404 }{0.12}{12} & 0.164 (0.68x) & 0.091 (5.87x) & 0.122 (67.1x) & 0.52 (243x) & oom \\% & oot & oot & 370.2 \\
& \multicolumn{5}{|c||}{} & 0 & 453 & 717 & 793 & (3159)\\% & & & 3159 \\
\hline
\prob{ scen06-20reduc }{82}{24}{409}{0.12}{12} & 0.203 (0.68x) & 0.095 (7.53x) & 0.142 (86.1x) & 0.838 (277x) & oom \\% & oot & oot & 611.1 \\
& \multicolumn{5}{|c||}{} & 0 & 447 & 717 & 794 & (3163)\\% & & & 3163 \\
\hline
\prob{ scen06-18reduc }{82}{26}{409}{0.12}{12} & 0.221 (0.76x) & 0.194 (4.7x) & 0.303 (56x) & 1.189 (292x) & oom \\%& oot & oot & 1462.5 \\
& \multicolumn{5}{|c||}{} & 0 & 458 & 718 & 796 & (3263)\\% & & & 3263 \\
\hline
\prob{ scen06-24 }{99}{20}{1203}{0.10}{12} & 0.25 (0.52x) & 0.236 (3.97x) & 0.278 (47.1x) & 0.867 (233x) & oom \\% & oot & oot & 386.6 \\
& \multicolumn{5}{|c||}{} & 0 & 437 & 319 & 900 & \\% & & & 2857 \\
\hline
\prob{ scen06-16reduc }{82}{28}{409}{0.12}{12} & 0.22 (0.45x) & 0.113 (10.4x) & 0.235 (101x) & 1.695 (304x) & oom \\% & oot & oot & 1273.6 \\
& \multicolumn{5}{|c||}{} & 0 & 458 & 717 & 812 & \\% & & & 3277 \\
\hline
\prob{ scen06-22 }{99}{22}{1210}{0.10}{12} & 0.271 (0.58x) & 0.26 (4.93x) & 0.358 (56.3x) & 1.415 (256x) & oom \\% & oot & oot & 532.6 \\
& \multicolumn{5}{|c||}{} & 0 & 437 & 403 & 803 & \\% & & & 3159 \\
\hline
\prob{ scen06-20 }{100}{24}{1215}{0.10}{12} & 0.306 (1.2x) & 0.263 (6.37x) & 0.371 (78.2x) & 1.979 (291x) & oom \\% & oot & oot & 1015.4 \\
& \multicolumn{5}{|c||}{} & 0 & 437 & 352 & 804 & \\% & & & 3163 \\
\hline
\prob{ scen06-18 }{100}{26}{1221}{0.10}{12} & 0.352 (0.83x) & 0.299 (7.54x) & 0.457 (94.4x) & 2.995 (303x) & oom \\% & oot & oot & 2429.8 \\
& \multicolumn{5}{|c||}{} & 0 & 437 & 327 & 813 & \\% & & & 3263 \\
\hline
\prob{ scen06-16 }{100}{28}{1222}{0.1}{12} & 0.36 (1.34x) & 0.389 (7.12x) & 0.537 (122x) & 4.382 (317x) & oom \\%& oot & oot & 2840.5 \\
& \multicolumn{5}{|c||}{} & 0 & 437 & 328 & 813 & \\% & & & 3277 \\
\hline
\prob{ scen06reduc }{82}{44}{409}{0.12}{14} & 0.343 (0.68x) & 0.306 (15.1x) & 0.787 (204x) & 0.787 (204x) & oom \\% & oot & oot & 2840.5 \\
& \multicolumn{5}{|c||}{} & 0 & 137 & 318 & 318 & \\% & & & 3277 \\
\hline
\hline
\end{tabular}
{\small
\caption{Celar Benchmark: Runtime (in seconds) of GpuBE, at varying of the bucket size $z$ and GpuBE($w^*$) (top),
and solution quality (bottom). The speedup of GpuBE($z$) and GpuBE($w^*$) w.r.t. their CPU counterparts are shown in parenthesis.}
\label{tab:celar}}
\end{center}
\end{table}
\begin{table}
\begin{center}
\resizebox{0.9\textwidth}{!}
\begin{tabular}{| l | p{6pt} p{2pt} p{12pt} p{6pt} p{6pt} || *{5}{r} |
\hline
\prob{Problem}{$n$}{$d$}{$c$}{$p_1$}{$w^*$} & \multicolumn{4}{|c}{GpuBE($z$)} & GpuBE \\% & \multicolumn{3}{|c|}{toulbar2} \\
\prob{}{}{}{}{}{} & $z_{\min}$ & $z_2$ & $z_3$ & $z_{\max}$ & $w^*$ \\%& NC & AC & EDAC \\
\hline\hline
\prob{ GEOM40-2 }{40}{2}{78}{0.12}{5} & 0.004 (0.25x) & 0.004 (0.25x) & 0.004 (0.25x) & 0.004 (0.25x) & 0.004 (0.25x)\\% & 6.64 & 6.39 & 0.36 \\
& \multicolumn{5}{|c||}{} & 22 & 22 & 22 & 22 & 22 \\% & & & 22 \\
\hline
\prob{ GEOM40-3 }{40}{3}{78}{0.12}{5} & 0.009 (0.44x) & 0.009 (0.44x) & 0.009 (0.44x) & 0.009 (0.44x) & 0.009 (0.44x)\\% & 39.82 & 43.83 & 2.23 \\
& \multicolumn{5}{|c||}{} & 7 & 7 & 7 & 7 & 7 \\% & & & 7 \\
\hline
\prob{ GEOM40-4 }{40}{4}{78}{0.12}{5} & 0.004 (4.25x) & 0.004 (4.25x) & 0.004 (4.25x) & 0.004 (4.25x) & 0.004 (4.25x) \\% & 198.2 & 198.6 & 4.48 \\
& \multicolumn{5}{|c||}{} & 3 & 3 & 3 & 3 & 3 \\% & & & 3 \\
\hline
\prob{ GEOM40-5 }{40}{5}{78}{0.12}{5} & 0.005 (10.8x) & 0.005 (10.8x) & 0.005 (10.8x) & 0.005 (10.8x) & 0.005 (10.8x)\\% & 0.38 & 0.31 & 0.26 \\
& \multicolumn{5}{|c||}{} & 1 & 1 & 1 & 1 & 1 \\% & & & 1 \\
\hline
\prob{ GEOM40-6 }{40}{6}{78}{0.12}{5} & 0.011 (5.73x) & 0.011 (5.73x) & 0.011 (5.73x) & 0.011 (5.73x) & 0.011 (5.73x) \\% & 0.001 & 0.001 & 0.01 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & 0 \\% & & & 0 \\
\hline
\prob{ GEOM30a-3 }{30}{3}{81}{0.2}{6} & 0.014 (0.21x) & 0.024 (0.5x) & 0.011 (1.73x) & 0.024 (0.79x) & 0.003 (2.67x)\\% & 12.25 & 12.12 & 1.94 \\
& \multicolumn{5}{|c||}{} & 0 & 10 & 11 & 11 & 11 \\% & & & 11 \\
\hline
\prob{ GEOM30a-4 }{30}{4}{81}{0.2}{6} & 0.028 (0.14x) & 0.024 (2.25x) & 0.012 (10.3x) & 0.012 (10.2x) & 0.003 (14.7x) \\% & 2.84 & 2.82 & 0.65 \\
& \multicolumn{5}{|c||}{} & 0 & 4 & 4 & 4 & 4 \\%& & & 4 \\
\hline
\prob{ GEOM30a-5 }{30}{5}{81}{0.2}{6} & 0.029 (0.21x) & 0.012 (16x) & 0.012 (21.3x) & 0.012 (20.2x) & 0.004 (21.8x)\\% & 0.02 & 0.02 & 0.05 \\
& \multicolumn{5}{|c||}{} & 0 & 1 & 1 & 1 & 1 \\% & & & 1 \\
\hline
\prob{ GEOM30a-6 }{30}{6}{81}{0.2}{6} & 0.03 (0.17x) & 0.013 (35.3x) & 0.028 (31.6x) & 0.015 (68.8x) & 0.006 (50.3x) \\% & 0.01 & 0.01 & 0.01 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & 0 \\% & & & 0 \\
\hline
\prob{ queen5-5-3 }{25}{3}{160}{0.87}{18} & 0.031 (0.19x) & 0.034 (48.2x) & 0.471 (203x) & 2.899 (250x) & oom \\% & 60.09 & 59.44 & 18.55 \\
& \multicolumn{5}{|c||}{} & 6 & 18 & 23 & 25 & (29) \\% & & & 29 \\
\hline
\prob{ queen5-5-4 }{25}{4}{160}{0.87}{18} & 0.031 (1x) & 0.038 (53.3x) & 0.158 (169x) & 1.355 (238x) & oom \\% & 76.07 & 80.85 & 26.3 \\
& \multicolumn{5}{|c||}{} & 0 & 1 & 4 & 5 & (12) \\% & & & 12 \\
\hline
\prob{ queen5-5-5 }{25}{5}{160}{0.87}{18} & 0.031 (2.19x) & 0.121 (97x) & 1.031 (247x) & 4.659 (267x) & oom \\%& 0.05 & 0.02 & 0.03 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & (0) \\% & & & 0 \\
\hline
\prob{ myciel5g-3 }{47}{3}{236}{0.44}{20} & 0.033 (2.58x) & 0.069 (27.2x) & 1.999 (201x) & 11.97 (308x) & oom \\% & 15.14 & 15.31 & 0.81 \\
& \multicolumn{5}{|c||}{} & 0 & 3 & 9 & 12 & (16) \\% & & & 16 \\
\hline
\prob{ myciel5g-4 }{47}{4}{236}{0.44}{20} & 0.031 (1.03x) & 0.045 (52.6x) & 0.23 (143x) & 7.852 (293x) & oom \\% & 76.86 & 82.87 & 7.93 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & (4) \\% & & & 4 \\
\hline
\prob{ myciel5g-5 }{47}{5}{236}{0.44}{20} & 0.072 (0.92x) & 0.051 (62.7x) & 0.41 (160x) & 6.513 (278x) & oom \\%& 45.73 & 47.47 & 48.15 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & (1) \\% & & & 1 \\
\hline
\prob{ myciel5g-6 }{47}{6}{236}{0.44}{20} & 0.035 (3.57x) & 0.069 (5.39x) & 0.123 (91x) & 1.636 (220x) & oom \\% & 0.01 & 0.001 & 0.001 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & (0) \\% & & & 0 \\
\hline
\prob{ DSJC125.1.4 }{125}{4}{736}{0.72}{72} & 0.241 (0.62x) & 0.168 (19.4x) & 0.477 (96.5x) & 3.551 (197x) & oom \\%& oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & -- \\% & & & \\
\hline
\prob{ DSJC125.1.5 }{125}{5}{736}{0.72}{72} & 0.285 (1.1x) & 0.17 (4.43x) & 0.393 (38.8x) & 1.801 (191x) & oom \\% & 743.1 & 828.4 & 151.3 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & (0) \\% & & & 0 \\
\hline
\prob{ le450-5a-2 }{450}{2}{5714}{0.81}{344} & 1.725 (0.18x) & 1.611 (1.46x) & 2.934 (40.7x) & 10.21 (67.3x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 618 & 734 & 833 & 878 & -- \\% & & & \\
\hline
\prob{ le450-5a-3 }{450}{3}{5714}{0.81}{344} & 1.728 (0.19x) & 1.847 (2.39x) & 2.042 (17.6x) & 7.207 (44.4x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 42 & 55 & 57 & 58 & --\\% & & & \\
\hline
\prob{ le450-5a-4 }{450}{4}{5714}{0.81}{344} & 1.422 (0.44x) & 1.549 (1.29x) & 2.088 (13.7x) & 6.517 (68.6x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 4 & 1 & -- \\% & & & \\
\hline
\prob{ le450-5a-5 }{450}{5}{5714}{0.81}{344} & 1.67 (0.98x) & 1.844 (3.7x) & 3.505 (39.3x) & 10.23 (65.1x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & -- \\%& & & \\
\hline
\end{tabular}
{\small
\caption{Coloring Benchmark: Runtime (in seconds) of GpuBE, at varying of the bucket size $z$ and GpuBE($w^*$) (top),
and solution quality (bottom). The speedup of GpuBE($z$) and GpuBE($w^*$) w.r.t. their CPU counterparts are shown in parenthesis.}
\label{tab:coloring}}
\end{center}
\end{table}
\begin{table}
\begin{center}
\resizebox{0.9\textwidth}{!}
\begin{tabular}{| l | p{6pt} p{2pt} p{12pt} p{6pt} p{6pt} || *{5}{r} |
\hline
\prob{Problem}{$n$}{$d$}{$c$}{$p_1$}{$w^*$} & \multicolumn{4}{|c}{GpuBE($z$)} & GpuBE \\% & \multicolumn{3}{|c|}{toulbar2} \\
\prob{}{}{}{}{}{} & $z_{\min}$ & $z_2$ & $z_3$ & $z_{\max}$ & $w^*$ \\% & NC & AC & EDAC \\
\hline\hline
\prob{ s386 }{172}{2}{172}{0.04}{19} & 0.054 (0.28x) & 0.051 (2.55x) & 0.053 (16.1x) & 0.185 (71.9x) & 0.129 (82.8x) \\% & 11.43 & 0.25 & 0.01 \\
& \multicolumn{5}{|c||}{} & 29 & 29 & 29 & 29 & 29 \\% & & & 29 \\
\hline
\prob{ s1423 }{748}{2}{748}{{0.06}}{38} & 0.184 (0.12x) & 0.182 (1.08x) & 0.189 (6.05x) & 0.546 (57.8x) & oom \\%& oot & oot & 0.47 \\
& \multicolumn{5}{|c||}{} & 231 & 231 & 231 & 231 & (231)\\% & & & 231 \\
\hline
\prob{ c499 }{499}{2}{499}{0.01}{42} & 0.133 (0.89x) & 0.131 (1.08x) & 0.141 (9.72x) & 0.391 (78.1x) & oom \\% & oot & oot & 1041.5 \\
& \multicolumn{5}{|c||}{} & 111 & 111 & 111 & 111 & (111)\\% & & & 111 \\
\hline
\prob{ c432 }{432}{2}{432}{0.01}{54} & 0.225 (0.33x) & 0.237 (0.72x) & 0.270 (7.93x) & 0.622 (69.6x) & oom \\%& oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 101 & 101 & 101 & 101 & --\\% & & & \\
\hline
\prob{ s1494 }{661}{2}{661}{0.01}{57} & 0.238 (0.19x) & 0.222 (1.14x) & 0.249 (17.1x) & 1.285 (101x) & oom \\%& oot & 0.08 & 0.28 \\
& \multicolumn{5}{|c||}{} & 32 & 32 & 32 & 32 & (32)\\% & & & 32 \\
\hline
\prob{ s1488 }{667}{2}{667}{0.01}{62} & 0.232 (0.13x) & 0.219 (1.07x) & 0.244 (16.7x) & 1.268 (98.5x) & oom \\%& oot & 0.07 & 0.04 \\
& \multicolumn{5}{|c||}{} & 32 & 32 & 32 & 32 & (32)\\% & & & 32 \\
\hline
\prob{ c880 }{880}{2}{880}{0.04}{68} & 0.245 (0.11x) & 0.241 (0.69x) & 0.295 (8.86x) & 1.1 (91.7x) & oom \\%& oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 162 & 162 & 162 & 162 & --\\% & & & \\
\hline
\prob{ s1196 }{561}{2}{561}{0.01}{92} & 0.19 (0.27x) & 0.185 (0.93x) & 0.305 (10.6x) & 1.049 (101x) & oom \\% & oot & 3.73 & 0.02 \\
& \multicolumn{5}{|c||}{} & 95 & 95 & 95 & 95 & (95)\\% & & & 95 \\
\hline
\prob{ s953 }{440}{2}{440}{0.01}{93} & 0.234 (0.13x) & 0.139 (1.62x) & 0.262 (11.2x) & 0.781 (98.8x) & oom \\% & oot & 1.09 & 0.02 \\
& \multicolumn{5}{|c||}{} & 124 & 124 & 124 & 124 & (124)\\% & & & 124 \\
\hline
\prob{ s1238 }{540}{2}{540}{0.01}{95} & 0.195 (0.28x) & 0.184 (1.04x) & 0.21 (16.8x) & 0.964 (94.6x) & oom \\% & oot & 3.14 & 0.04 \\
& \multicolumn{5}{|c||}{} & 95 & 95 & 95 & 95 & (95)\\% & & & 95 \\
\hline
\end{tabular}
{\small
\caption{Iscas-89 Benchmark: Runtime (in seconds) of GpuBE, at varying of the bucket size $z$ and GpuBE($w^*$) (top),
and solution quality (bottom). The speedup of GpuBE($z$) and GpuBE($w^*$) w.r.t. their CPU counterparts are shown in parenthesis.}
\label{tab:iscas}}
\end{center}
\end{table}
\begin{table}
\begin{center}
\resizebox{0.9\textwidth}{!}
\begin{tabular}{| l | p{6pt} p{2pt} p{12pt} p{6pt} p{8pt} || *{5}{r} |
\hline
\prob{Problem}{$n$}{$d$}{$c$}{$p_1$}{$w^*$} & \multicolumn{4}{|c}{GpuBE($z$)} & GpuBE \\% & \multicolumn{3}{|c|}{toulbar2} \\
\prob{}{}{}{}{}{} & $z_{\min}$ & $z_2$ & $z_3$ & $z_{\max}$ & $w^*$ \\% & NC & AC & EDAC \\
\hline\hline
\prob{ eye }{36}{21}{53}{0.09}{2} & 0.048 (3.83x) & 0.045 (3.67x) & 0.045 (2.2x) & 0.045 (4.13x) & 0.006 (12.7x) \\%& 0.1 & 0.01 & 0.04 \\
& \multicolumn{5}{|c||}{} & 1 & 1 & 1 & 1 & 1 \\%& & & 1 \\
\hline
\prob{ wijsmanguo }{49}{36}{68}{0.06}{3} & 0.278 (2.15x) & 0.279 (2.18x) & 0.278 (2.21x) & 0.278 (2.15x) & 0.012 (18.8x) \\% & oot & 0.84 & 0.22 \\
& \multicolumn{5}{|c||}{} & 1 & 1 & 1 & 1 & 1 \\%& & & 1 \\
\hline
\prob{ cancer }{49}{36}{68}{0.06}{3} & 0.304 (2.34x) & 0.28 (2.91x) & 0.291 (2.44x) & 0.278 (2.55x) & 0.012 (18.8x) \\% & oot & 0.8 & 0.19 \\
& \multicolumn{5}{|c||}{} & 1 & 1 & 1 & 1 & 1 \\%& & & 1 \\
\hline
\prob{ sobel }{7}{6}{8}{0.61}{3} & 0.002 (0.5x) & 0.002 (1x) & 0.001 (2x) & 0.001 (2x) & 0.001 (1x) \\% & 0.001 & 0.001 & 0.001 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & 0 \\%& & & 0 \\
\hline
\prob{ connell }{12}{6}{15}{0.39}{3} & 0.004 (0.25x) & 0.004 (1.5x) & 0.002 (2.5x) & 0.001 (6x) & 0.001 (4x) \\%& 0.01 & 0.01 & 0.001 \\
& \multicolumn{5}{|c||}{} & 0 & 1 & 1 & 1 & 1 \\%& & & 1 \\
\hline
\prob{ pedck60-L2 }{60}{10}{106}{0.08}{5} & 0.017 (1.71x) & 0.052 (78x) & 0.052 (77x) & 0.051 (82.6x) & 0.02 (144x) \\%& 0.01 & 0.02 & 0.01 \\
& \multicolumn{5}{|c||}{} & 2 & 2 & 2 & 2 & 2 \\%& & & 2 \\
\hline
\prob{ pedck60-L1 }{60}{10}{108}{0.08}{5} & 0.016 (1.56x) & 0.033 (125x) & 0.044 (94.1x) & 0.033 (124x) & 0.02 (137x) \\% & 0.05 & 0.02 & 0.02 \\
& \multicolumn{5}{|c||}{} & 2 & 2 & 2 & 2 & 2 \\%& & & 2 \\
\hline
\prob{ pedck60-L12 }{60}{10}{108}{0.08}{5} & 0.023 (1.26x) & 0.033 (124x) & 0.033 (127x) & 0.033 (126x) & 0.02 (143x) \\% & 0.4 & 0.18 & 0.02 \\
& \multicolumn{5}{|c||}{} & 6 & 6 & 6 & 6 & 6 \\%& & & 6 \\
\hline
\prob{ saudiarabia }{37}{15}{43}{0.16}{5} & 0.015 (4.13x) & 0.267 (231x) & 0.286 (211x) & 0.282 (210x) & 0.187 (212x) \\% & 0.04 & 0.01 & 0.02 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & 0 \\%& & & 0 \\
\hline
\prob{ parkinson }{37}{15}{43}{0.16}{5} & 0.015 (3.4x) & 0.292 (206x) & 0.265 (223x) & 0.274 (223x) & 0.186 (237x) \\%& 0.03 & 0.01 & 0.01 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & 0 \\%& & & 0 \\
\hline
\prob{ pedck350l3 }{350}{10}{578}{0.03}{24} & 0.092 (2.13x) & 0.094 (7.23x) & 0.109 (35.7x) & 0.192 (112x) & oom \\%& 0.03 & 0.04 & 0.06 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & (0)\\% & & & 0 \\
\hline
\prob{ pedck350l2 }{350}{10}{578}{0.03}{24} & 0.091 (1.02x) & 0.131 (4.21x) & 0.109 (32.9x) & 0.235 (81.1x) & oom \\% & 0.03 & 0.08 & 0.06 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 1 & 1 & (1)\\% & & & 1 \\
\hline
\prob{ pedck350 }{350}{10}{580}{0.03}{26} & 0.099 (1.04x) & 0.139 (3.89x) & 0.117 (27.9x) & 0.252 (114x) & oom \\%& 0.19 & 0.18 & 0.03 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 2 & 2 & (2)\\% & & & 2 \\
\hline
\prob{ sheep4r-4-3 }{2662}{10}{5021}{0.00}{38} & 1.16 (0.98x) & 1.173 (5.83x) & 1.551 (29.9x) & 3.139 (110x) & oom \\% & oot & oot & 0.34 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 1 & (1)\\% & & & 1 \\
\hline
\prob{ sheep4r-4-2 }{2172}{10}{4026}{0.00}{46} & 0.874 (0.96x) & 1.3 (4.11x) & 1.145 (29.9x) & 2.339 (110x) & oom \\% & oot & oot & 0.22 \\
& \multicolumn{5}{|c||}{} & 1 & 0 & 0 & 0 & (1)\\% & & & 1 \\
\hline
\prob{ sheep4r-4-1 }{641}{10}{1196}{0.05}{63} & 0.273 (0.96x) & 0.291 (5.92x) & 0.36 (35.7x) & 0.782 (120x) & oom \\% & oot & 0.25 & 0.03 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & (0)\\% & & & 0 \\
\hline
\prob{ sheep4r-4-0 }{1541}{10}{2941}{0.03}{108} & 0.76 (0.95x) & 0.786 (5.93x) & 1.066 (31.3x) & 1.066 (31.3x) & oom \\% & oot & 13.66 & 0.11 \\
& \multicolumn{5}{|c||}{} & 0 & 0 & 0 & 0 & (0)\\% & & & 0 \\
\hline
\prob{ pedck1000 }{928}{6}{1736}{0.09}{126} & 0.368 (0.34x) & 0.393 (1.32x) & 0.399 (4.46x) & 0.447 (20.1x) & oom \\%& oot & oot & 1.39 \\
& \multicolumn{5}{|c||}{} & 15 & 15 & 12 & 16 & (19)\\% & & & 19 \\
\hline
\end{tabular}
{\small
\caption{Pedigree Benchmark: Runtime (in seconds) of GpuBE, at varying of the bucket size $z$ and GpuBE($w^*$) (top),
and solution quality (bottom). The speedup of GpuBE($z$) and GpuBE($w^*$) w.r.t. their CPU counterparts are shown in parenthesis.}
\label{tab:pedigree}}
\end{center}
\end{table}
\begin{table}
\begin{center}
\resizebox{0.9\textwidth}{!}
\begin{tabular}{| l | p{6pt} p{2pt} p{12pt} p{6pt} p{8pt} || *{5}{r} |
\hline
\prob{Problem}{$n$}{$d$}{$c$}{$p_1$}{$w^*$} & \multicolumn{4}{|c}{GpuBE($z$)} & GpuBE \\% & \multicolumn{3}{|c|}{toulbar2} \\
\prob{}{}{}{}{}{} & $z_{\min}$ & $z_2$ & $z_3$ & $z_{\max}$ & $w^*$ \\% & NC & AC & EDAC \\
\hline\hline
\prob{ 8 }{8}{4}{15}{0.28}{2} & 0.003 (0.33x) & 0.001 (1.0x) & 0.002 (1.5x) & 0.003 (1.33x) & 0.003 (2.0x)\\% & 0.01 & 0.01 & 0.01 \\
& \multicolumn{5}{|c||}{} & 2 & 2 & 2 & 2 & 2 \\% & & & 2 \\
\hline
\prob{ 1502 }{209}{4}{411}{0.01}{5} & 0.082 (0.32x) & 0.039 (1.18x) & 0.08 (0.57x) & 0.079 (0.61x) & 0.029 (0.66x)\\% & oot & oot & 0.02 \\
& \multicolumn{5}{|c||}{} & 28042 & 28042 & 28042 & 28042 & 28042 \\% & & & 28042 \\
\hline
\prob{ 54 }{67}{4}{271}{0.15}{11} & 0.037 (0.92x) & 0.036 (1.86x) & 0.036 (7.31x) & 0.045 (53.7x) & 0.027 (46.1x) \\% & 9.01 & 10.24 & 0.48 \\
& \multicolumn{5}{|c||}{} & 31 & 32 & 35 & 37 & 37 \\% & & & 37 \\
\hline
\prob{ 503 }{143}{4}{635}{0.07}{12} & 0.088 (0.74x) & 0.085 (2.88x) & 0.088 (18.3x) & 0.203 (31.9x) & 0.044 (59.9x) \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 7093 & 9106 & 11111 & 11113 & 11113 \\% & & & \\
\hline
\prob{ 29 }{82}{4}{462}{0.17}{14} & 0.085 (0.11x) & 0.07 (3.71x) & 0.147 (97.7x) & 7.281 (236x) & 3.327 (141x) \\% & 31.69 & 30.96 & 0.7 \\
& \multicolumn{5}{|c||}{} & 7035 & 8048 & 8055 & 8059 & 8059 \\% & & & 8059 \\
\hline
\prob{ 404 }{100}{4}{710}{0.16}{19} & 0.112 (0.6x) & 0.207 (5.95x) & 0.188 (113x) & 0.701 (224x) & 0.301 (124x) \\% & oot & oot & 210.4 \\
& \multicolumn{5}{|c||}{} & 76 & 106 & 112 & 114 & 114 \\% & & & 114 \\
\hline
\prob{ 42b }{190}{4}{1140}{0.09}{19} & 0.455 (0.13x) & 0.358 (0.65x) & 0.276 (43.6x) & 12.98 (239x) & oom \\%& oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 58049 & 95050 & 135050 & 155050 & --\\% & & & \\
\hline
\prob{ 505b }{240}{4}{1716}{0.06}{23} & 0.392 (0.12x) & 0.484 (0.61x) & 0.361 (48.6x) & 7.531 (238x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 6106 & 11159 & 15204 & 19244 & --\\% & & & \\
\hline
\prob{ 1504 }{605}{4}{4187}{0.03}{25} & 1.047 (0.18x) & 0.884 (2.41x) & 1.221 (55.3x) & 12.09 (263x) & oom \\%& oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 80104 & 107175 & 131227 & 141251 & --\\% & & & \\
\hline
\prob{ 408b }{200}{4}{1843}{0.10}{29} & 0.552 (0.14x) & 0.309 (2.77x) & 0.862 (132x) & 5.774 (218x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 2104 & 5162 & 6206 & 6215 & --\\% & & & \\
\hline
\prob{ 42 }{190}{4}{1394}{0.11}{30} & 0.295 (0.11x) & 0.251 (7.53x) & 0.59 (56.9x) & 9.419 (241x) & oom \\%& oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 60049 & 97050 & 105050 & 133050 & --\\% & & & \\
\hline
\prob{ 505 }{240}{4}{2242}{0.11}{30} & 0.61 (0.09x) & 0.547 (1.05x) & 0.712 (83.3x) & 8.757 (220x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 5103 & 7149 & 10178 & 15211 & --\\% & & & \\
\hline
\prob{ 408 }{200}{4}{2232}{0.17}{40} & 0.637 (0.12x) & 0.506 (2.14x) & 0.816 (96x) & 10.2 (265x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 2100 & 3123 & 4169 & 5185 & --\\% & & & \\
\hline
\prob{ 5 }{309}{4}{5621}{0.19}{44} & 1.45 (0.14x) & 1.18 (1.09x) & 1.879 (37x) & 6.573 (158x) & oom \\%& oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 42 & 53 & 86 & 106 & --\\% & & & \\
\hline
\prob{ 412 }{300}{4}{4348}{0.16}{48} & 1.366 (0.08x) & 1.097 (2.05x) & 1.228 (41.3x) & 7.048 (183x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 2106 & 2131 & 8176 & 11258 & --\\% & & & \\
\hline
\prob{ 507 }{311}{4}{5732}{0.18}{68} & 1.788 (0.13x) & 1.379 (1.59x) & 1.744 (38.2x) & 5.343 (140x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 6114 & 5140 & 7194 & 10226 & --\\% & & & \\
\hline
\prob{ 1506 }{940}{4}{15240}{0.05}{77} & 4.82 (0.13x) & 3.805 (0.82x) & 3.929 (8.52x) & 16.93 (128x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 50122 & 68152 & 86172 & 88235 & --\\% & & & \\
\hline
\prob{ 28 }{230}{4}{5226}{0.42}{87} & 1.397 (0.12x) & 1.173 (1.89x) & 1.946 (50x) & 8.584 (226x) & oom \\%& oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 43075 & 52105 & 76105 & 88105 & --\\% & & & \\
\hline
\prob{ 509 }{348}{4}{8624}{0.22}{92} & 2.82 (0.16x) & 2.204 (5.56x) & 5.664 (75.6x) & 10.61 (278x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 3105 & 3217 & 5191 & 6157 & --\\% & & & \\
\hline
\prob{ 414 }{364}{4}{10108}{0.24}{104} & 3.61 (0.16x) & 2.673 (1.95x) & 3.091 (16.7x) & 7.608 (80.7x) & oom \\%& oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 3113 & 3120 & 4195 & 6141 & --\\% & & & \\
\hline
\prob{ 1401 }{488}{4}{10963}{0.17}{105} & 3.283 (0.16x) & 3.047 (1.62x) & 3.543 (13x) & 13.65 (157x) & oom \\%& oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 64057 & 61066 & 66071 & 87071 & --\\% & & & \\
\hline
\prob{ 1403 }{665}{4}{13616}{0.11}{105} & 4.52 (0.14x) & 3.829 (1.63x) & 4.246 (15x) & 10.11 (69.6x) & oom \\% & oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 58099 & 71129 & 74118 & 82120 & --\\% & & & \\
\hline
\prob{ 1405 }{855}{4}{18258}{0.09}{105} & 5.708 (0.14x) & 4.791 (2.08x) & 6.787 (17.3x) & 19.98 (74.5x) & oom \\%& oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 58099 & 71129 & 74118 & 84177 & --\\% & & & \\
\hline
\prob{ 1407 }{1057}{4}{21786}{0.07}{105} & 7.018 (0.15x) & 6.283 (2.21x) & 8.931 (19x) & 18.15 (53.9x) & oom \\%& oot & oot & oot \\
& \multicolumn{5}{|c||}{} & 58127 & 74180 & 75164 & 84202 & --\\% & & & \\
\hline
\end{tabular}
\small{
\caption{Spot Benchmark: Runtime (in seconds) of GpuBE, at varying of the bucket size $z$ and GpuBE($w^*$) (top),
and solution quality (bottom). The speedup of GpuBE($z$) and GpuBE($w^*$) w.r.t. their CPU counterparts are shown in parenthesis.}
\label{tab:spot}}
\end{center}
\end{table}
\section{Related Work}
\label{sec:relatedwork}
The use of GPUs to solve difficult combinatorial problems has been explored by several proposals in different areas of constraint optimization.
For instance, Meyer \emph{et al.}~\cite{lalami:11} proposed a multi-GPU implementation of the \emph{simplex tableau} algorithm that relies on a vertical problem decomposition to reduce communication between GPUs.
In constraint programming, Arbelaez and Codognet~\cite{arbelaez:14} proposed a GPU-based version of the \emph{Adaptive Search} algorithm, which explores several \emph{large neighborhoods} in parallel, resulting in a speedup factor of $17$. Campeotto \emph{et al.}~\cite{campeotto:14} proposed a GPU-based framework that exploits both parallel propagation and parallel exploration of several large neighborhoods using local search techniques, leading to a speedup factor of up to $38$.
The combination of GPUs with dynamic programming has also been explored to solve different combinatorial optimization problems. For instance, Boyer \emph{et al.}~\cite{boyer:12} proposed the use of GPUs to compute the classical DP recursion step for the knapsack problem, which led to a speedup factor of $26$.
Paw{\l}owski \emph{et al.}~\cite{pawlowski:14} presented a DP-based solution for the \emph{coalition structure formation problem} on GPUs, reporting up to two orders of magnitude of speedup.
In a recent work, Bistaffa \emph{et al.}~\cite{bistaffa:16} study the parallelization of an inference-based algorithm to solve COPs using GPUs, albeit exclusively in the centralized case.
Silberstein \emph{et al.}~\cite{Silberstein:08} study a GPU-based kernel for the sum-product operations that arise in \emph{marginalize a product of functions} (MPF) problems. The authors report an average speedup factor of $15$ for random benchmarks and Bayesian networks and higher average speedups (up to two orders of magnitude) for log domains due to the difference in performance of the \emph{log2f} and \emph{exp2f} functions on the CPU and GPU.
In the distributed constraint optimization context, GPU parallelism has been applied to speed up several DCOP solving techniques. Fioretto \emph{et al.}~\cite{fioretto:16} proposed a multi-variable agent decomposition strategy to solve \emph{general} DCOPs with complex local subproblems, which makes use of GPUs to implement a search-based and a sampling-based algorithm to speed up the agents' local subproblems resolution.
Le \emph{et al.}~\cite{le:16} studied a GPU accelerated algorithm in the context of stochastic DCOPs---DCOPs where the values of the cost tables are stochastic. The authors used SIMT-style parallelism on a DP-based approach, which resulted in a speedup of up to two orders of magnitude.
Recently, a combination of GPUs with \emph{Markov Chain Monte Carlo} (MCMC) sampling algorithms has been proposed in the context of solving DCOPs~\cite{fioretto:16c}, where the authors adopted GPUs to accelerate the computation of the normalization constants used in the MCMC sampling process as well as to compute several samples in parallel, resulting in a speedup of up to one order of magnitude.
Finally, a parallelization of the AND/OR Branch-and-Bound search with mini-bucket heuristic has been presented in \cite{lars:17}.
Differently from other proposals, our approach aims at using GPUs to exploit SIMT-style parallelism from DP-based methods to solve general, exact and approximated, WCSPs and DCOPs.
\section{Conclusions and Discussions}
\label{sec:conclusions}
Inference-based algorithms are powerful tools for solving discrete optimization problems. However, their applicability is limited by their high time and space requirements. Motivated by the increasing availability of GPUs, in this paper, we proposed a scheme to speed up the resolution of inference-based methods for centralized and distributed constraint optimization by exploiting SIMT-style parallelism.
We introduced an exact algorithm and an approximated algorithm that are inspired by BE and MBE for WCSPs \rev{and tasks over belief networks (e.g., MPE)}, and by DPOP and ADPOP for DCOPs. These procedures make use of multiple threads in the GPU cards to parallelize the aggregation and elimination procedures, which are responsible for the high complexity in the inference-based approaches.
Additionally, we detailed the design of the data structures adopted to process cost functions with GPUs, and of the mapping adopted to associate GPU threads to cost functions' entries, which allows us to efficiently exploit the data parallelism (SIMT) supported by GPUs.
Finally, we reported an extensive experimental evaluation of our inference-based GPU implementations on both centralized and distributed benchmarks.
We showed that the use of GPUs provides significant advantages in terms of runtime and scalability, achieving speedups of up to two order of magnitude, showing a considerable reduction in runtime (up to 345 times faster) with respect to the serialized version, and that the speedups increase with the induced width of the problem and with the size of the domain of the problem's variables.
The proposed results are significant---the wide availability of GPUs provides access to parallel computing solutions that can be used to improve efficiency of WCSPs and DCOP solvers.
Furthermore, GPUs are renowned for their complex architectures (multiple memory levels with very different size and speed characteristics; relatively slow cores), which often create challenges to the effective exploitation of parallelism from irregular applications. The strong experimental results indicate that the proposed algorithms are well-suited to GPU architectures.
These results hint that our approach could be exploited in the the context of a dynamic search that makes use of the mini-bucket elimination method as an heuristic to infer bounds on the solution quality, potentially allowing dynamic variable orderings.
Indeed, the main drawback of this type of search (and various look-ahead methods) is that, since the heuristic is (re)computed in various nodes during search process, the time invested in the heuristic computation may not be cost-effective.
Leveraging the use of GPUs to infer bounds faster during the dynamic search,
may therefore produce dramatic speedup to the whole search process.
While this paper describes the applicability of our approach to (M)BE and (A)DPOP, analogous techniques can be derived and applied to other inference-based approaches to solve discrete optimization problems (e.g.,~to implement the logic of inference-based propagators) and optimization on probabilistic graphical models---(e.g.,~in finding maximum probability explanation (MPE) in belief networks).
\rev{
Additionally, our work can be extended to solve \emph{sum-product} problems, also known as \emph{weighted counting}, \emph{partition function}, or \emph{probability of evidence}. Due to the difficulty of these problems the value of accelerated versions of the bucket elimination algorithms is especially important. We plan to work in this direction as future work.
}
We also envision this technology could open the door to efficiently enforcing higher forms of consistencies than domain consistency (e.g.,~\emph{path consistency}~\cite{montanari:74}, \emph{adaptive consistency}~\cite{dechter:88}, or the more recently proposed \emph{branch consistency} for DCOPs \cite{fioretto:14b}), especially when the cost functions need to be represented explicitly.
\begin{acknowledgements}
This research is partially supported by the National Science Foundation under grants 1345232, 1550662, 1458595, and 1401639. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the sponsoring organizations, agencies, or the U.S. government.
\end{acknowledgements}
\bibliographystyle{spmpsci}
|
\section{Introduction}
\label{sec:introduction}
Answer set programming (ASP) \cite{iclp:GelfondL88,coll:MarekT99,%
cacm:BrewkaET11,book:GebserKKS12} is a well-established logic programming
paradigm based on the stable model semantics of logic programs. Its main
advantage is an intuitive, declarative language, and the fact that, generally,
each answer set of a given logic program describes a valid answer to the
original question. Moreover, ASP solvers---see e.g.\
\cite{ai:GebserKS12,lpnmr:AlvianoDFLR13,lpnmr:ElkabaniPS05,%
datalog:AlvianoFLPPT10}---have made huge strides in efficiency.
A logic program usually consists of a set of logical implications by which new
facts can be inferred from existing ones, and a set of facts that represent the
concrete input instance. Logic programming in general, and ASP in particular,
have also gained popularity because of their intuitive, declarative syntax. The
following example illustrates this:
\begin{example}\label{ex:intro}
The following rule naturally expresses the fact that two people are relatives
of the same generation up to second cousin if they share a great-grandparent.
%
\vspace{-.5ex}
\small
\begin{verbatim}
uptosecondcousin(X, Y) :-
parent(X, PX), parent(PX, GPX),
parent(GPX, GGP), parent(GPY, GGP),
parent(PY, GPY), parent(Y, PY), X != Y.
\end{verbatim}
\normalsize\vspace{-6ex}\qed
\end{example}
Rules written in an intuitive fashion, like the one in the above example, are
usually larger than strictly necessary. Unfortunately, the use of large rules
causes problems for current ASP solvers since the input program is grounded
first (i.e.\ all the variables in each rule are replaced by all possible, valid
combinations of constants). This grounding step generally requires exponential
time for rules of arbitrary size. In practice, the grounding time can thus
become prohibitively large. Also, the ASP solver is usually quicker in
evaluating the program if the grounding size remains small.
In order to increase solving performance, we could therefore split the rule in
Example~\ref{ex:intro} up into several smaller ones by hand, keeping track of
grandparents and great-grandparents in separate predicates, and then writing a
smaller version of the second cousin rule. While this is comparatively easy to
do for this example, this can become very tedious if the rules become even more
complex and larger, maybe also involving negation or arithmetic expressions.
However, since current ASP grounders and solvers become increasingly slower with
larger rules, and noting the fact that ASP programs often need expert
hand-tuning to perform well in practice, this represents a significant entry
barrier and contradicts the fact that logic programs should be fully
declarative: in a perfect world, the concrete formulation should not have an
impact on the runtime. In addition, to minimize solver runtime in general, it is
therefore one of our goals to enable logic programs to be written in an
intuitive, fully declarative way without having to think about various technical
encoding optimizations.
To this end, in this paper we propose the \verb!lpopt! tool that automatically
optimizes and rewrites large logic programming rules into multiple smaller ones
in order to improve solving performance. This tool, based on an idea proposed
for very simple ASP programs in \cite{iclp:MorakW12}, uses the concept of tree
decompositions of rules to split them into smaller chunks. Intuitively, via a
tree decomposition joins in the body of a rule are arranged into a tree-like
form. Joins that belong together are then split off into a separate rule, only
keeping the join result in a temporary atom. We then extend the algorithm to
handle the entire standardized ASP language \cite{web:aspcore}, and also
introduce new optimizations for complex language constructs such as weak
constraints, arithmetic expressions, and aggregates.
The main contributions of this paper are therefore as follows:
\vspace{-1ex}
\begin{itemize}
\item we extend, on a theoretical basis, the \algo{lpopt}\ algorithm proposed
in \cite{iclp:MorakW12} to the full syntax of the ASP language according to
the ASP-Core-2 language specification~\cite{web:aspcore};
\item we establish how to treat complex constructs like aggregates, and
propose an adaptation of the decomposition approach so that it can split up
large aggregate expressions into multiple smaller rules and expressions,
further reducing the grounding size;
\item we implement the \algo{lpopt}\ algorithm in C++, yielding the
\verb!lpopt! tool for automated logic program optimization, and give an
overview of how this tool is used in practice; and
\item we perform an experimental evaluation of the tool on the encodings and
instances used in the fifth Answer Set Programming Competition which show
the benefit of our approach, even for encodings already heavily
hand-optimized by ASP experts.
\end{itemize}
\section{Preliminaries}
\label{sec:preliminaries}
\paragraph*{General Definitions.} We define two pairwise disjoint countably
infinite sets of symbols: a set $\mathbf{C}$ of \emph{constants} and a set $\mathbf{V}$
of \emph{variables}. Different constants represent different values
(\emph{unique name assumption}). By $\variables{X}$ we denote sequences (or, with
slight notational abuse, sets) of variables $\variable{X}_1, \ldots, \variable{X}_k$ with $k
\geqslant 0$. For brevity, let $[n] = \{1,\ldots,n\}$, for any integer $n
\geqslant 1$.
A (\emph{relational}) \emph{schema} $\schema{S}$ is a (finite) set of
\emph{relational symbols} (or \emph{predicates}). We write $\relation{p}/n$ for
the fact that $\relation{p}$ is an $n$-ary predicate. A \emph{term} is a
constant or variable. An \emph{atomic formula} $\atom{a}$ over $\schema{S}$ (called
\emph{$\schema{S}$-atom}) has the form $\fullatom{p}{\terms{t}}$, where $\relation{p}
\in \schema{S}$ and $\terms{t}$ is a sequence of terms. An \emph{$\schema{S}$-literal} is
either an $\schema{S}$-atom (i.e.\ a positive literal), or an $\schema{S}$-atom preceded
by the negation symbol ``$\neg$'' (i.e.\ a negative literal). For a literal
$\ell$, we write $\domof{\ell}$ for the set of its terms, and $\varof{\ell}$ for
its variables. This notation naturally extends to sets of literals. For brevity,
we will treat conjunctions of literals as sets. For a domain $C \subseteq
\mathbf{C}$, a (\emph{total} or \emph{two-valued}) \emph{$\schema{S}$-interpretation} $I$
is a set of $\schema{S}$-atoms containing only constants from $C$ such that, for
every $\schema{S}$-atom $\fullatom{p}{\constants{a}} \in I$,
$\fullatom{p}{\constants{a}}$ is true, and otherwise false. When obvious from
the context, we will omit the schema-prefix.
A \emph{substitution} from a set of literals $L$ to a set of literals $L'$ is a
mapping $s: \mathbf{C} \cup \mathbf{V} \to \mathbf{C} \cup \mathbf{V}$ that is defined on
$\domof{L}$, is the identity on $\mathbf{C}$, and $\fullatom{p}{\term{t_1}, \ldots,
\term{t_n}} \in L$ (resp.\ $\neg \fullatom{p}{\term{t_1}, \ldots, \term{t_n}}
\in L$) implies $\fullatom{p}{s(\term{t_1}), \ldots, s(\term{t_n})} \in L'$
(resp., $\neg \fullatom{p}{s(\term{t_1}), \ldots, s(\term{t_n})} \in L'$).
\paragraph*{Answer Set Programming (ASP).} A \emph{logic programming rule} is a
universally quantified reverse first-order implication of the form
$$\mathcal{H}(\variables{X}, \variables{Y}) \gets \mathcal{B}^+(\variables{X}, \variables{Y}, \variables{Z}, \variables{W}) \wedge
\mathcal{B}^-(\variables{X}, \variables{Z}),$$ where $\mathcal{H}$ (the \emph{head}), resp.\ $\mathcal{B}^+$
(the \emph{positive body}), is a disjunction, resp. conjunction, of atoms, and
$\mathcal{B}^-$ (the \emph{negative body}) is a conjunction of negative literals, each
over terms from $\mathbf{C} \cup \mathbf{V}$. For a rule $\pi$, let $\head{\pi}$,
$\pbody{\pi}$, and $\nbody{\pi}$ denote the set of atoms occurring in the head,
the positive, and the negative body, respectively. Let $\body{\pi} = \pbody{\pi}
\cup \nbody{\pi}$. A rule $\pi$ where $\head{\pi} = \emptyset$ is called a
\emph{constraint}. Substitutions naturally extend to rules. We focus on
\emph{safe} rules where every variable in the rule occurs in the positive body.
A rule is called $\emph{ground}$ if all its terms are constants. The grounding
of a rule $\pi$ w.r.t.\ a domain $C \subseteq \mathbf{C}$ is the set of rules
$\groundc{C}{\pi} = \{ s(\pi) \mid s \text{ is a substitution, mapping }
\varof{\pi} \text{ to elements from } C \}$.
A \emph{logic program} $\Pi$ is a finite set of logic programming rules. The
schema of a program $\Pi$, denoted $\schof{\Pi}$, is the set of predicates
appearing in $\Pi$. The \emph{active domain} of $\Pi$, denoted $\adomof{\Pi}$,
with $\adomof{\Pi} \subset \mathbf{C}$, is the set of constants appearing in $\Pi$.
A program $\Pi$ is ground if all its rules are ground. The \emph{grounding of a
program $\Pi$} is the ground program $\ground{\Pi} = \bigcup_{\pi \in \Pi}
\groundc{\adomof{\Pi}}{\pi}$. The \emph{(Gelfond-Lifschitz) reduct} of a ground
program $\Pi$ w.r.t.\ an interpretation $I$ is the ground program $\Pi^I = \{
\head{\pi} \gets \pbody{\pi} \mid \pi \in \Pi, \nbody{\pi} \cap I = \emptyset
\}$.
A $\schof{\Pi}$-interpretation $I$ is a \emph{(classical) model} of a ground
program $\Pi$, denoted $I \models \Pi$ if, for every ground rule $\pi \in \Pi$,
it holds that $I \cap \pbody{\pi} = \emptyset$ or $I \cap (\head{\pi} \cup
\nbody{\pi}) \neq \emptyset$, that is, $I$ satisfies $\pi$. $I$ is a
\emph{stable model} (or \emph{answer set}) of $\Pi$, denoted $I \models_s \Pi$
if, in addition, there is no $J \subset I$ such that $J \models \Pi^I$, that is,
$I$ is subset-minimal w.r.t.\ the reduct $\Pi^I$. The set of answer sets of
$\Pi$, denoted $\smods{\Pi}$, are defined as $\smods{\Pi} = \{ I \mid I \text{
is a } \schof{\Pi}\text{-interpretation, and } I \models_s \Pi \}$. For a
non-ground program $\Pi$, we define $\smods{\Pi} = \smods{\ground{\Pi}}$. When
referring to the fact that a logic program is intended to be interpreted under
the answer set semantics, we often refer to it as an \emph{ASP program}.
\paragraph*{Tree Decompositions.} A \emph{tree decomposition} of a graph $G =
(V,E)$ is a pair $\mathcal{T} = (T, \chi)$, where $T$ is a rooted tree and $\chi$ is a
labelling function over nodes $t$ of $T$, with $\chi(t) \subseteq V$ called the
\emph{bag of $t$}, such that the following holds: (i) for each $v \in V$, there
exists a node $t$ in $T$, such that $v \in \chi(t)$; (ii) for each $\{v,w\} \in
E$, there exists a node $t$ in $T$, such that $\{v, w\} \subseteq \chi(t)$; and
(iii) for all nodes $r$, $s$, and $t$ in $T$, such that $s$ lies on the path
from $r$ to $t$, we have $\chi(r) \cap \chi(t) \subseteq \chi(s)$. The
\emph{width} of a tree decomposition is defined as the cardinality of its
largest bag minus one. The \emph{treewidth} of a graph $G$, denoted by $\tw{G}$,
is the minimum width over all tree decompositions of $G$. To decide whether a
graph has treewidth at most $k$ is \ensuremath{\textsc{NP}}-complete \cite{siamjadm:ArnborgCP87}. For
an arbitrary but fixed $k$ however, this problem can be solved (and a tree
decomposition constructed) in linear time \cite{siamcomp:Bodlaender96}.
Given a non-ground logic programming rule $\pi$, we let its \emph{Gaifman graph}
$G_\pi = (\varof{\pi}, E)$ such that there is an edge $(X, Y)$ in $E$ iff
variables $\variable{X}$ and $\variable{Y}$ occur together in the head or in a body atom of
$\pi$. We refer to a tree decomposition of $G_\pi$ as a \emph{tree decomposition
of rule $\pi$}. The treewidth of rule $\pi$ is the treewidth of $G_\pi$.
\section{Rule Decomposition}
\label{sec:algorithms}
This section lays out the theoretical foundations for our rule decomposition
approach. First, we recall the algorithm from \cite{iclp:MorakW12}, and then
describe how it can be extended to handle three of the main extensions of the
ASP language, namely arithmetic expressions, aggregates, and weak constraints
(i.e.\ optimization statements), as defined in the ASP-Core language standard
\cite{web:aspcore}.
As demonstrated in Example~\ref{ex:intro}, rules that are intuitive to write
and read are not necessarily the most efficient ones to evaluate in practice.
ASP solvers generally struggle with rules that contain many variables since
they rely on a grounder-solver approach: first, the grounding of a logic
program is computed by a grounder. As per the definition in
Section~\ref{sec:preliminaries}, the size of the grounding can, in the worst
case, be exponential in the number of variables. For large rules, the grounding
step can already take a prohibitively large amount of time. However, the solver
is also adversely affected by this blowup. In practice, this leads to long
runtimes and sometimes the inability of the ASP system to solve a given
instance. This also contributes to the fact that, while the syntax of ASP is
fully declarative, writing efficient encodings still takes expert knowledge.
It is therefore desirable to have a way to automatically rewrite such large
rules into a more efficient representation. One way to do this is the rule
decomposition approach, first proposed in \cite{iclp:MorakW12}, which we will
briefly recall next.
\subsection{Decomposition of Simple Rules}
Generally speaking, the approach in \cite{iclp:MorakW12} computes the tree
decomposition of a rule, and then splits the rule up into multiple, smaller
rules according to this decomposition. While in the worst case this
decomposition may not change the rule at all, in practice it is often the case
that large rules can be split up very well. For instance, the large rule in
Example~\ref{ex:intro} will be amenable for such a decomposition.
Let us briefly recall the algorithm from \cite{iclp:MorakW12} which we will
refer to as the \algo{lpopt}\ algorithm. For a given rule $\pi$, the algorithm
works as follows:
\begin{enumerate}
\item\label{decomp:step1} Compute a tree decomposition $\mathcal{T} = (T, \chi)$ of
$\pi$ with minimal width where all variables occurring in the head of $\pi$
are contained in its root node bag.
\item\label{decomp:step2} For each node $n$, let $\mathit{temp}_n$ be a fresh
predicate, and the same for each variable $\variable{X}$ in $\pi$ and predicate
$\mathit{dom}_X$. Let $\variables{Y}_n = \chi(n) \cap \chi(p_n)$, where $p_n$ is
the parent node of $n$. For the root node $\mathit{root}$, let
$\mathit{temp}_\mathit{root}$ be the entire head of $\pi$, and, accordingly,
$\variables{Y}_\mathit{root} = \varof{\head{\pi}}$. Now, for a node $n$, generate
the following rule:
%
\begin{center}
\begin{tabular}{l l l l l}
$\mathit{temp}_n(\variables{Y}_n)$ & $\leftarrow$ & \,\,\; $\{ \atom{a} \in
\body{\pi}$ & $\mid$ & $\varof{\atom{a}} \subseteq \chi(n) \}$\\
%
& & $\cup \, \{ \mathit{dom}_X(X)$ & $\mid$ & $\atom{a} \in \nbody{\pi}, X \in
\varof{\atom{a}}, \varof{\atom{a}} \subseteq \chi(n),$\\
%
& & & & $\not\exists \atom{b} \in \pbody{\pi}: \varof{\atom{b}}
\subseteq \chi(n), \variable{X} \in \varof{\atom{b}} \}$\\
%
& & $\cup \, \{ \mathit{temp}_m(\variables{Y}_m)$ & $\mid$ & $m \text{ is a child of }
n\}.$
\end{tabular}
\end{center}
\item\label{decomp:step3} For each $\variable{X} \in \varof{\nbody{\pi}}$, for which
a domain predicate $\relation{dom}$ is needed to guarantee safety of a rule
generated above, pick an atom $\atom{a} \in \pbody{\pi}$, such that $\variable{X}
\in \varof{\atom{a}}$ and generate a rule $$\fullatom{dom_\variable{X}}{\variable{X}}
\leftarrow \atom{a}.$$
\end{enumerate}
Step~\ref{decomp:step3} is needed because splitting up a rule may make it
unsafe. In order to remedy this, a domain predicate is generated for each unsafe
variable that arises due to the rule splitting in step~\ref{decomp:step2}. The
following example illustrates how the algorithm works.
\begin{example}
Given the rule $$\pi = \fullatom{h}{\variable{X}, \variable{W}} \gets \fullatom{e}{\variable{X},
\variable{Y}}, \fullatom{e}{\variable{Y}, \variable{Z}}, \neg \fullatom{e}{\variable{Z}, \variable{W}},
\fullatom{e}{\variable{W}, \variable{X}},$$ a tree decomposition of $\pi$ could look as
follows (note that we write in each bag of the tree decomposition not just the
variables as per definition but also all literals of rule $\pi$ over these
variables which is a more intuitive notation):
%
\begin{center}
\begin{tikzpicture}
\tikzstyle{every path}=[
very thick,draw=green!50!black!50]
\tikzstyle{every node}=[rectangle,
very thick,draw=green!50!black!50,
top color=white,bottom color=green!50!black!20]
\tikzstyle{level 1}=[level distance=10mm, sibling distance=30mm]
\tikzstyle{level 2}=[sibling distance=30mm]
\tikzstyle{level 3}=[sibling distance=30mm]
%
\node (root) {$\fullatom{h}{\variable{X}, \variable{W}}, \fullatom{e}{\variable{X}, \variable{Y}},
\fullatom{e}{\variable{W}, \variable{X}}$}
child {%
node {$\fullatom{e}{\variable{Y}, \variable{Z}} ,\neg \fullatom{e}{\variable{Z}, \variable{W}}$}
};
\end{tikzpicture}
\end{center}
%
Applying the \algo{lpopt}\ algorithm to $\pi$ with the tree decomposition
above yields the following set of rules $\algo{lpopt}(\pi)$:
$$\fullatom{dom_\variable{W}}{\variable{W}} \gets \fullatom{e}{\variable{W}, \variable{X}},$$
$$\fullatom{temp}{\variable{Y}, \variable{W}} \gets \fullatom{e}{\variable{Y}, \variable{Z}}, \neg
\fullatom{e}{\variable{Z}, \variable{W}}, \fullatom{dom_\variable{W}}{\variable{W}}, \text{ and}$$
$$\fullatom{h}{\variable{X}, \variable{W}} \gets \fullatom{e}{\variable{X}, \variable{Y}},
\fullatom{e}{\variable{W}, \variable{X}}, \fullatom{temp}{\variable{Y}, \variable{W}},$$ where
$\relation{temp}$ is a fresh predicate not appearing anywhere else. \qed
\end{example}
Let $\Pi$ be a logic program. When the above algorithm is applied to all rules
in $\Pi$, resulting in a logic program $\algo{lpopt}(\Pi)$ as stated in
\cite{iclp:MorakW12}, the answer sets of $\Pi$ are preserved in the following
way: when all temporary atoms are removed, each answer set of
$\algo{lpopt}(\Pi)$ coincides with exactly one answer set from the original
program $\Pi$. Furthermore, the size of the grounding now no longer depends on
the rule size. In fact, it now only depends on the rule treewidth as the
following result states:
\begin{theorem}[\cite{iclp:MorakW12}]\label{thm:groundingsize}
The size of $\ground{\algo{lpopt}(\Pi)}$ is bounded by $O(2^k \cdot n)$, where
$n$ is the size of $\Pi$, and $k$ is the maximal treewidth of the rules in
$\Pi$.
\end{theorem}
The above theorem implies that the size of the grounding of a program $\Pi$,
after optimization via the \algo{lpopt}\ algorithm, is no longer exponential in
the size of $\Pi$, but only in the treewidth of its rules. As
\cite{iclp:MorakW12} demonstrates, this decomposition approach already has a
significant impact on the size of the grounding in practical instances.
However, the ASP language standard \cite{web:aspcore} extends the ASP language
with other useful constructs that the \algo{lpopt}\ algorithm proposed in
\cite{iclp:MorakW12} cannot handle. These include arithmetic expressions,
aggregates, and weak constraints. Looking at concrete, practical instances of
ASP programs, e.g.\ the encodings used in recent ASP competitions
\cite{ai:CalimeriGMR16}, a large majority use such constructs. In the following,
we will therefore extend the \algo{lpopt}\ algorithm to be able to treat them in
a similar way.
\subsection{Treating Arithmetic Expressions}
Arithmetic expressions are atoms of the form $\variable{X} = \varphi(\variables{Y})$, that is,
an equality with one variable (or constant number) $\variable{X}$ on the left-hand
side, and an expression $\varphi$ on the right-hand side, where $\varphi$ is any
mathematical expression built using the variables from $\variables{Y}$, constant
numbers, and the arithmetic connectives ``+,'' ``-,'' ``*,'' and ``/.'' In
addition to the positive and negative body, a rule $\pi$ may also contain a set
of such arithmetic expressions describing a relationship between variables with
the obvious meaning.
Clearly, in order to adapt the rule decomposition approach to this it is easy to
extend the definition of the graph representation of $\pi$ to simply contain a
clique between all variables occurring together in an arithmetic expression.
The \algo{lpopt}\ algorithm then works as described above up to
step~\ref{decomp:step2}. However, a problem may arise when, in
step~\ref{decomp:step3} of the \algo{lpopt}\ algorithm, a domain predicate
$\fullatom{dom_\variable{X}}{\variable{X}}$ is to be generated. Consider the following
example:
\begin{example}\label{ex:arithmetics}
Let $\pi$ be the rule $\fullatom{a}{\variable{X}} \gets \neg \fullatom{b}{\variable{X},
\variable{Y}}, \fullatom{c}{\variable{Y}}, \fullatom{d}{\variable{Z}}, \variable{X} = \variable{Z} + \variable{Z}$. A
simple decomposition according to the \algo{lpopt}\ algorithm may lead to the
following rules: $$\fullatom{temp}{\variable{X}} \gets \neg \fullatom{b}{\variable{X},
\variable{Y}}, \fullatom{c}{\variable{Y}}, \fullatom{dom_\variable{X}}{\variable{X}}, \text{ and}$$
$$\fullatom{a}{\variable{X}} \gets \fullatom{d}{\variable{Z}}, \variable{X} = \variable{Z} + \variable{Z},
\fullatom{temp}{\variable{X}}.$$ It remains to define the domain predicate
$\relation{dom_\variable{X}}$. According to the original definition of \algo{lpopt},
we would get $$\fullatom{dom_\variable{X}}{\variable{X}} \gets \variable{X} = \variable{Z} + \variable{Z}$$ which
is unsafe. \qed
\end{example}
The conditions for safety of rules with arithmetic expressions are defined in
the ASP language specification \cite{web:aspcore}. As
Example~\ref{ex:arithmetics} shows, in order for such expressions to work with
the \algo{lpopt}\ algorithm a more general approach to defining the domain
predicates is needed in step~\ref{decomp:step3}. In fact, instead of choosing a
single atom from the rule body to generate the domain predicate, in general a
set of atoms and arithmetic expressions must be chosen. It is easy to see that
if a rule $\pi$ is safe then, for each variable $\variable{X} \in \body{\pi}$, there
is a set $A$ of (positive) atoms and arithmetic expressions in the body of
$\pi$ that makes that variable safe. In step~\ref{decomp:step3} of the
\algo{lpopt}\ algorithm, for a variable $\variable{X}$ we now choose such a set
$A_\variable{X}$ of body elements in a greedy fashion as follows: let $\variables{S} =
\{ \variable{X} \}$ the set of variables that we need to make safe. For each variable
$\variable{S} \in \variables{S}$, pick a (positive) atom from $\body{\pi}$ that
makes $\variable{S}$ safe, add it to $A_\variable{X}$, and remove $\variable{S}$ from
$\variables{S}$. If no such atom exists in the body of $\pi$, greedily add the
smallest arithmetic expression $\variable{S} = \varphi(\variables{Y})$ in $\body{\pi}$
to $A_\variable{X}$ and let $\variables{S} = \variables{S} \setminus \{ \variable{S} \}
\cup \variables{Y}$. Repeat this process until $\variables{S}$ is empty. Since $\pi$
itself is safe and finite in size, the above procedure necessarily terminates.
Finally, generate the rule $\fullatom{dom_\variable{X}}{\variable{X}} \gets A_\variable{X}$. It is
easy to see that this rule is safe and describes the possible domain of variable
$\variable{X}$ as required. Note also that this rule can not be split up futher as
removing any single element of the rule would make it unsafe.
\begin{example}\label{ex:arithmetics:continued}
A correct domain predicate for Example~\ref{ex:arithmetics} would be defined
as follows: $$\fullatom{dom_\variable{X}}{\variable{X}} \gets \variable{X} = \variable{Z} + \variable{Z},
\fullatom{d}{\variable{Z}}.$$ This ensures the proper safety of all rules generated
by the \algo{lpopt}\ algorithm. \qed
\end{example}
Note that the rule generated in Example~\ref{ex:arithmetics:continued} repeats
most of the atoms that the second rule generated in Example~\ref{ex:arithmetics}
already contains. It is not immediately obvious how such situations can be
remedied in general. Investigating this issue is part of ongoing work.
\subsection{Treating Weak Constraints}
As defined in \cite{web:aspcore}, a weak constraint $\pi [\term{k}:\terms{t}]$
is a constraint $\pi$ annotated with a term $\term{k}$ representing a weight
and a sequence of terms $\terms{t}$ occurring in $\pi$. The intended meaning is
that each answer set $I$ is annotated by a total weight $w(I)$, which is the sum
over all $\term{k}$ for each tuple of constants $\constants{c}$ that realize
$\terms{t}$ in $I$ and satisfy the body of $\pi$. Such a weak constraint can
easily be decomposed by replacing $\pi [\term{k}:\terms{t}]$ with the rule $\pi'
= \fullatom{temp}{\term{k}, \terms{t}} \gets \body{\pi}$, where
$\relation{temp}$ is a fresh predicate, and the weak constraint $\bot \gets
\fullatom{temp}{\term{k}, \terms{t}} [\term{k}:\terms{t}]$. Finally, the
\algo{lpopt}\ algorithm is then applied to rule $\pi'$. This allows our rule
decomposition approach also to be applied in an optimization context (i.e.\
where the task for the solver is to find optimal answer sets w.r.t.\ their
weight).
\subsection{Treating Aggregate Expressions}
An aggregate expression, as defined in \cite{web:aspcore}, is an expression of
the form $$\term{t} \preccurlyeq \#\mathit{agg}\{ \terms{t} : \phi(\variables{X}) \},$$
where $\term{t}$ is a term; $\preccurlyeq \, \in \{ <, \leqslant, =, \neq,
\geqslant, > \}$ is a builtin relation; $\mathit{agg}$ is one of $\mathit{sum}$,
$\mathit{count}$, $\mathit{max}$, and $\mathit{min}$; $\terms{t} = \langle
\term{t}_1, \ldots, \term{t}_n \rangle$ is a sequence of terms; and $\phi(\variables{X})$
is a set of literals, arithmetic expressions, and aggregate expressions, called
the \emph{aggregate body}. Aggregates may appear in rule bodies, or recursively
inside other aggregates, with the following semantic meaning: Given an
interpretation $I$, for each valid substitution $s$ such that $s(\phi(\variables{X}))
\subseteq I$, take the tuple of constants $s(\terms{t})$. Let us denote this set
with $T$. Now, execute the aggregate function on $T$ as follows: for
$\#\mathit{count}$, calculate $|T|$; for $\#\mathit{sum}$, calculate
$\Sigma_{\terms{t} \in T} \term{t}_1$, where $\term{t}_1$ is the first term in
$\terms{t}$; for $\#\mathit{max}$ and $\#\mathit{min}$, take the maximum and
minimum term appearing in the first position of each tuple in $T$, respectively.
Finally, an aggregate expression is true if the relation $\preccurlyeq$ between
term $\term{t}$ and the result of the aggregate function is fulfilled.
Extending the \algo{lpopt} algorithm to aggregate expressions is again
straightforward: The rule graph $G_\pi = (V, E)$ of a rule $\pi$ containing
aggregate expressions is defined as follows: Let $V$ be the set of variables
occurring in $\pi$ outside of aggregate expressions. Let $E$ be as before and,
in addition, add, for each aggregate expression $\atom{e}$, a clique between all
variables $\varof{\atom{e}} \cap V$ to $E$. Intuitively, the rule graph should
contain, for each aggregate expression, a clique between all variables that
appear in the aggregate and somewhere else in the rule. Variables appearing
only in aggregates are in a sense ``local'' and are therefore not of interest
when decomposing the rule.
While the above transformation is straightforward, we can, however, go one step
further and also decompose the inside elements of an aggregate expression. To
this end, let $\term{t} \preccurlyeq \#\mathit{agg}\{ \terms{t} : \phi(\variables{X},
\variables{Y}) \}$ be an aggregate expression occurring in some rule $\pi$, where
$\variables{X}$ are variables that occur either in $\terms{t}$ or somewhere else in
$\pi$, and $\variables{Y}$ are variables occurring inside the aggregate only. Replace
the aggregate expression with $\term{t} \preccurlyeq \#\mathit{agg}\{ \terms{t} :
\psi(\variables{X}, \variables{Z}), \fullatom{temp}{\terms{t}, \variables{Z}} \}$, and furthermore,
generate a rule $\fullatom{temp}{\terms{t}, \variables{Z}} \gets
\overline{\psi}(\variables{Y}), \overline{\psi}_\mathit{dom}(\variables{Y})$, for some fresh
predicate $\relation{temp}$. Here, $\psi$ contains all those atoms from $\phi$
that contain a variable from $\variables{X}$, and $\overline{\psi}$ contains the rest.
$\overline{\psi}_\mathit{dom}$ contains domain predicates generated like in
step~\ref{decomp:step3} of the \algo{lpopt}\ algorithm, as needed to make the
temporary rule safe. The temporary rule can then be decomposed via \algo{lpopt}.
This is best illustrated by an example:
\begin{example}\label{ex:aggregates}
Let $\pi$ be the following logic programming rule, saying that a vertex is
``good'' if it has at least two neighbours that, themselves, have a red
neighbour: $$\fullatom{good}{\variable{X}} \gets \fullatom{vertex}{\variable{X}}, 2
\leqslant \#\mathit{count}\{\variable{Y} : \fullatom{edge}{\variable{X}, \variable{Y}},
\fullatom{edge}{\variable{Y}, \variable{Z}}, \fullatom{red}{\variable{Z}} \}.$$ According to the
above approach, the rule can now be split up as follows. Firstly, the
aggregate is replaced: $$\fullatom{good}{\variable{X}} \gets
\fullatom{vertex}{\variable{X}}, 2 \leqslant \#\mathit{count}\{\variable{Y} :
\fullatom{edge}{\variable{X}, \variable{Y}}, \fullatom{temp}{\variable{Y}} \},$$ and furthermore, a
temporary rule is created as follows: $$\fullatom{temp}{\variable{Y}} \gets
\fullatom{edge}{\variable{Y}, \variable{Z}}, \fullatom{red}{\variable{Z}}.$$ The latter rule is now
amenable for decomposition via the \algo{lpopt} algorithm. \qed
\end{example}
Note that the above approach allows us to decompose, to a degree, even the
insides of an aggregate, which, for large aggregate bodies, can lead to a
further significant reduction in the grounding size.
\subsection{Correctness}
The correctness of the above extensions to the original algorithm follows by the
same arguments that prove the correctness of the original algorithm proposed in
\cite{iclp:MorakW12}, and trivially from the construction for arithmetic
expressions and safety. For the latter, note that for domain predicates of a
variable $\variable{X}$ we explicitly select a set of atoms that make the variable
safe, and that such a set always exists, since the original rule is safe. For
the former two (namely weak constraints and aggregate expressions), the only
thing that needs to be examined is the first step: replacing (part of) the body
with a temporary predicate. But correctness of this is easy to see. Instead of
performing all joins within the weak constraint or aggregate, we perform the
join in a new, separate rule and project only relevant variables into a
temporary predicate. The weak constraint or aggregate then only needs to
consider this temporary predicate since, by construction, all other variables
not projected into the temporary predicate do not play a role w.r.t.\
optimization or aggregation. Finally, the original algorithm from
\cite{iclp:MorakW12} extended to handle arithmetic expressions, for which
correctness has already been established, is then applied to this new, separate
rule.
\subsection{Further Language Extensions}
The ASP-Core language specification \cite{web:aspcore}, as well as the
\emph{gringo} grounder\footnote{\url{http://potassco.sourceforge.net}}, allow
further constructs like variable pooling, aggregates with multiple bodies, or
with upper and lower bounds in the same expression, in addition to various
extensions that amount to syntactic sugar. These constructs make the above
explanations unnecessarily more tedious. However, from a theoretical point of
view, all of these additional constructs can be normalized to one of the forms
discussed in the previous subsections. Furthermore, as we shall see in the next
section, we have implemented the \algo{lpopt}\ algorithm to directly treat all
standard ASP language constructs and certain other additions, like variable
pooling. More details about this general approach, and the exact, but more
tedious, algorithm details, can be found in \cite{thesis:Bichler15}.
\section{Implementation}\label{sec:implementation}
A full implementation of the algorithm and its extensions described in
Section~\ref{sec:algorithms} is now available in the form of the \verb!lpopt!
tool, available with relevant documentation and examples at
\url{http://dbai.tuwien.ac.at/proj/lpopt}. The following gives a quick outline
of how to use the tool.
\verb!lpopt! accepts as its input any form of ASP program that follows the ASP
input language specification laid out in \cite{web:aspcore}. The output of the
program in its default configuration is a decomposed program that also follows
this specification. In addition, the tool guarantees that no language construct
is introduced in the output that was not previously present in the input (cf.\
Section~\ref{sec:algorithms}). Therefore, for example, a program without
aggregates will not contain any aggregates as a result of rule decomposition.
The following is a description of the parameters of the tool:
\begin{footnotesize}
\begin{verbatim}
Usage: lpopt [-idbt] [-s seed] [-f file] [-h alg] [-l file]
-d dumb: do not perform optimization
-b print verbose and benchmark information
-t perform only tree decomposition step
-i ignore head variables when decomposing
-h alg decomposition algorithm, one of {mcs, mf, miw (def)}
-s seed initialize random number generator with seed.
-f file the file to read from (default is stdin)
-l file output infos (treewidth) to file
\end{verbatim}
\end{footnotesize}
\noindent
In what follows, we will briefly describe the most important features of the
tool.
\paragraph{Tree Decomposition Heuristics.} As stated in
Section~\ref{sec:preliminaries}, computing an optimal tree decomposition w.r.t.\
width is an \ensuremath{\textsc{NP}}-hard problem. We thus make use of several heuristic algorithms,
namely the \emph{maximum cardinality search (mcs)}, \emph{minimum fill (mf)},
and \emph{minimum induced width (miw)} approaches described in
\cite{iandc:BodlaenderK10}, that yield tree decompositions that provide good
upper bounds on the treewidth (i.e.\ on an optimal decomposition). It turns out
that in practice, since rules in ASP programs are usually not overly large,
these heuristics come close to, and often even yield, an optimal tree
decomposition for rules. The heuristic algorithm to use for decomposition can be
selected using the \verb!-h! command line parameter. Since these heuristic
approaches rely to some degree on randomization, a seed for the pseudo-random
number generator can be passed along with the \verb!-s! command line parameter.
\paragraph{Measuring the Treewidth of Rules.} Theorem~\ref{thm:groundingsize}
allows us to calculate an upper bound on the size of the grounding of the input
program. In order to do this, the maximal treewidth of any rule in an ASP
program must be known. The \verb!-l! switch of the \verb!lpopt! tool allows
this to be calculated. It forces the tool to perform tree decompositions on all
rules inside an input ASP program, simply outputting the maximal treewidth (or,
more accurately, an upper bound; see above) over all of them into the given
file, and then exiting. Clearly, when a single ASP rule is given as input,
this switch will output a treewidth upper bound of that single rule.
\subsection*{Recommended Usage}
Assuming that a file \verb!enc.lp! contains the encoding of a problem as an ASP
program and that a file \verb!instance.db! contains a set of ground facts
representing a problem instance, the recommended usage of the tool is as
follows:
\begin{verbatim}
cat enc.lp instance.db | lpopt | grounder | solver
\end{verbatim}
In the above command, \verb!grounder! and \verb!solver! are programs for
grounding and for solving, respectively. One established solver that we will use
in the next section for our experimental evaluation is \emph{clasp}
\cite{ai:GebserKS12}. If \emph{clasp} is used as a solver together with the
\verb!lpopt! tool, we generally recommend the use of the \verb!--sat-prepro!
flag, which often speeds up the solving process substantially for decomposed
rules generated by \verb!lpopt! (by considering the fact that the truth values
of all temporary atoms generated by \verb!lpopt! are determined exactly by the
rule body, and need never be guessed).
\section{Experimental Evaluation}
\label{sec:evaluation}
We have tested our \verb!lpopt! tool and benchmarked the performance of
grounding and solving of programs preprocessed with \verb!lpopt! against
non-preprocessed ones. All benchmarks were made on the instance sets of the
fifth answer set programming competition 2014
\footnote{\url{https://www.mat.unical.it/aspcomp2014/}}, which, for most problem
classes, provides two encodings, one from 2013, and one from 2014. The
benchmarks have been run on a 3.5GHz AMD Opteron Processor 6308 with 192 GB of
RAM to its disposal. We used the potassco software
suite\footnote{\url{http://potassco.sourceforge.net}}, namely \emph{gringo}
verison 4.5.3 as the grounder and \emph{clasp} version 3.1.3 as the solver. A
timeout of 300 seconds was set for solving, and 1000 seconds for grounding.
Furthermore, as suggested in the previous section, \emph{clasp} was called with
the \verb!--sat-prepro! flag enabled. In this paper, we will survey the most
important results.
\paragraph*{Remark.} One central aim of our tool is to improve solving
performance for hand-written encodings by non-experts of ASP. In the spirit of a
truly declarative language, it shouldn't matter \emph{how} an encoding is
written as long as it is correct (i.e.\ w.r.t.\ runtime, there should not be a
difference between ``good'' and ``bad'' encodings). In this respect, the ASP
competition does not offer an optimal benchmark set since all encodings are
extensively hand-tuned by ASP experts. However, as to the best of our knowledge
there is no better-suited comprehensive benchmark set available, we will show
that even for these extensively hand-tuned ASP competition encodings our tool
can still find decompositions that decrease grounding size and improve solving
performance. However, there are also encodings that are so perfectly hand-tuned
that only trivial optimizations are possible with the current version of
\verb!lpopt!.
\paragraph*{Results.} Let us first note that the runtime of \verb!lpopt! itself,
for all encodings in the benchmark set, was always less than what can be
accurately measured on a computer system today. Applying our rule decomposition
algorithm thus comes virtually for free for hand-written encodings. Out of the
49 encodings provided by the ASP competition, \verb!lpopt! was able to
syntactically rewrite 41 which indicates that, as mentioned above, even
extensively hand-tuned programs can be further decomposed in an automated
manner. The remaining eight encodings contained rules that were so small that no
further decomposition was possible (i.e.\ their Gaifman graph was a clique of
usually 3-4 nodes) and thus the output of \verb!lpopt! was the original,
unmodified encoding in these cases. In 27 of the 41 encodings rewritten by
\verb!lpopt!, the decompositions were trivial and had no significant impact on
the solving performance. This is due to the fact that only rules that were
already very small (and thus did not contribute much to the grounding size in
the first place) could be decomposed. In five cases out of the 41 rewritten
encodings, we noticed a decrease in solving performance (see the paragraph on
limitations of \verb!lpopt! below for an explanation) and in the remaining seven
cases, the \verb!lpopt! rewriting was able to speed up the solving process with
substantial improvements in three of these seven. Two of those were the stable
marriage problem encoding of 2013, and the permutation pattern matching encoding
of 2014 which we will take a closer look at below. Full benchmark results for
the entire dataset can be found in \cite{thesis:Bichler15}.
\begin{figure}[t]
\begin{multicols}{2}
\includegraphics[width=\linewidth]{figures/sm-grounding-time.png}
\vspace{-1.7ex}
\begin{center} (a) \end{center}
\includegraphics[width=\linewidth]{figures/sm-solving-time.png}
\vspace{-1.7ex}
\begin{center} (b) \end{center}
\end{multicols}
\vspace{-3ex}
\caption{Benchmark results for the stable marriage 2013 instances. The
horizontal axis represents the individual test instances, sorted by runtime
without rule decomposition.}
\vspace{-3ex}
\label{fig:stablemarriage}
\end{figure}
As can be seen in Figure~\ref{fig:stablemarriage}, both grounding and solving
time decrease dramatically. Notice that the grounding time is, in general,
directly correlated with the size of the respective grounding. With \verb!lpopt!
preprocessing, the grounding size decreases dramatically by a factor of up to
65. The grounder is thirty times faster when using preprocessing, and the solver
about three times. This is because of the following constraint in the encoding
that can be decomposed very well:
\small
\begin{verbatim}
:- match(M,W1), manAssignsScore(M,W,Smw), W1!=W,
manAssignsScore(M,W1,Smw1), Smw>Smw1, match(M1,W),
womanAssignsScore(W,M,Swm), womanAssignsScore(W,M1,Swm1),
Swm>=Swm1.
\end{verbatim}
\normalsize
The constraint rule above is quite intuitive to read: There cannot be a man $M$
and a woman $W$, such that they would both be better off if they were matched
together, instead of being matched as they are (that is, to $W1$ and $M1$,
respectively). It encodes, precisely and straightforwardly, the condition of a
stable marriage. The 2014 encoding splits this rule up, making the encoding much
harder to understand. However, with \verb!lpopt! preprocessing, the grounding
and solving performance matches that of the hand-tuned 2014 encoding. This again
illustrates that the \algo{lpopt}\ algorithm allows for efficient processing of
rules written by non-experts that are not explicitly hand-tuned.
\begin{figure}
\begin{multicols}{2}
\includegraphics[width=\linewidth]{figures/ppm-grounding-time.png}
\vspace{-1.7ex}
\begin{center} (a) \end{center}
\includegraphics[width=\linewidth]{figures/ppm-solving-time.png}
\vspace{-1.7ex}
\begin{center} (b) \end{center}
\end{multicols}
\vspace{-3ex}
\caption{Benchmark results for permutation pattern matching 2014. The
horizontal axis represents the individual test instances, sorted by runtime
without rule decomposition.}
\vspace{-3ex}
\label{fig:patternmatching}
\end{figure}
A second example of \verb!lpopt!'s capabilities is the permutation pattern
matching problem illustrated in Figure~\ref{fig:patternmatching}. The grounding
time of the largest instance is $980$ seconds without preprocessing and $17$
seconds with preprocessing. This instance was also impossible to solve within
the timeout window of $300$ seconds without \verb!lpopt! preprocessing, but
finishing within $88$ seconds when \verb!lpopt! was run first.
\paragraph*{Other Use Cases.} \verb!lpopt! has also been employed in other works
that illustrate its performance benefits. In particular, several solvers for
other formalisms rely on a rewriting to ASP in order to solve the original
problem. Such rewritings can easily lead to the generation of large rules that
current ASP solving systems are generally unable to handle. For example, in
\cite{thesis:Heissenberger16} ASP rewritings for several problems from the
abstract argumentation domain, proposed in \cite{ecai:BrewkaW14}, are
implemented. In Section~4.6 of the thesis, the performance benefits of
\verb!lpopt! are clearly demonstated for these rewritings. Interestingly, these
rewritings also make heavy use of aggregates which goes to show that
\verb!lpopt! also handles these constructs well. Another example is
\cite{iclp:BichlerMW16}, where multiple rewritings for \SIGMA{P}{2}\ and
\SIGMA{P}{3}-hard problems are proposed and then benchmarked, again showcasing
that without \verb!lpopt! these rewritings could not be solved by current ASP
solvers in all but the most simple cases.
\paragraph*{Limitations.} However, we also want to point out some limitations of
the \algo{lpopt}\ algorithm. When a domain predicate is used by the algorithm,
the selection of atoms that generate this domain predicate is at the moment
essentially random, since the greedy selection depends on the order of the atoms
appearing in the rule. This approach, as discussed in
Section~\ref{sec:algorithms}, may thus not pick an optimal set of atoms.
However, it depends on this selection how many ground rules this domain
predicate rule will generate when passed to the grounder. Therefore, it may at
the moment be the case that the increased grounding size caused by the domain
predicate rules may destroy any benefit caused by splitting up the main rule.
This is precisely what caused the increase in solving time for the five
encodings out of 49 that \verb!lpopt! was able to rewrite but where solving
performance deteriorated. Clearly, this begs the question of what the best
strategy is to select atoms to generate domain predicates. This is part of
ongoing work.
\section{Conclusions}
\label{sec:conclusions}
In this paper, we present an algorithm, based on a prototype from
\cite{iclp:MorakW12}, that allows the decomposition of large logic programming
rules into smaller ones that current state-of-the-art answer set programming
solvers are better equipped to handle. Our implementation handles the entire
ASP-Core-2 language \cite{web:aspcore}. Benchmark results show that in practice,
even for extensively hand-tuned ASP programs, our rule decomposition algorithm
can improve solving performance significantly. Future work will include
implementing this approach directly into state-of-the-art grounders like the
\emph{gringo} grounder used in our benchmarks, as well as further refining the
algorithm w.r.t.\ selection of domain predicate atoms, as discussed at the end
of Section~\ref{sec:evaluation}.
|
\section{Introduction and Statement of the Results}
\label{sec1}
Let $L\left(q,\alpha ,\beta \right)$ denote the Sturm--Liouville problem
\begin{equation}\label{eq1.1}
- y'' + q\left(x\right)y = \mu y, \; x \in \left(0, \pi \right), \; \mu \in \mathbb{C},
\end{equation}
\begin{equation}\label{eq1.2}
y\left(0 \right)\cos \alpha + y'\left( 0 \right)\sin \alpha = 0, \; \alpha \in \left(0, \pi\right],
\end{equation}
\begin{equation}\label{eq1.3}
y\left(\pi\right)\cos \beta + y'\left(\pi \right)\sin \beta = 0, \; \beta \in \left[0, \pi\right),
\end{equation}
where $q$ is a real-valued, summable function on $\left[0, \pi\right]$ (we write $q \in L_\mathbb{R}^1\left[0, \pi \right]$). By $L\left(q, \alpha, \beta \right)$ we also denote the self-adjoint operator, generated by the problem \eqref{eq1.1}--\eqref{eq1.3} in Hilbert space $L^2\left[0, \pi\right]$ (see \cite{Naimark:1969,Levitan-Sargsyan:1970}). It is well-known that the spectra of $L\left(q, \alpha, \beta \right)$ is discrete and consists of real, simple eigenvalues (see \cite{Naimark:1969,Levitan-Sargsyan:1970,Marchenko:1977}), which we denote by $\mu_n\left(q, \alpha, \beta \right),$ $n=0,1,2\dots,$ emphasizing the dependence of $\mu_n$ on $q,$ $\alpha$ and $\beta.$ For $\mu_n$ the following asymptotic formula have been proven in \cite{Harutyunyan:2008}:
\begin{equation}\label{eq1.4}
\mu_n \left(q, \alpha, \beta \right) = \left[n+\delta_n \left(\alpha, \beta \right)\right]^2+\cfrac{1}{\pi} \displaystyle\int_{0}^{\pi}q\left(t\right)dt+r_n\left(q, \alpha, \beta \right),
\end{equation}
where $\delta_n$ is the solution of the equation (for $n \geq 2$)
\begin{multline}\label{eq1.5}
{\delta_n}(\alpha, \beta)=\cfrac{1}{\pi}\arccos \cfrac{\cos \alpha}{\sqrt {\left(n+{\delta_n}(\alpha, \beta)\right)^2\sin^2 \alpha+\cos^2 \alpha}}- \\
-\cfrac{1}{\pi}\arccos \cfrac{\cos \beta}{\sqrt {\left(n+{\delta_n}(\alpha, \beta)\right)^2\sin^2 \beta+\cos^2 \beta}},
\end{multline}
and $r_n \left(q, \alpha, \beta \right)=o\left(1\right),$ when $n \to \infty,$ uniformly in $\alpha, \beta \in \left[0, \pi \right]$ and $q$ from any bounded subset of $L_\mathbb{R}^1 \left[0, \pi \right]$ (we will write $q \in {BL}_\mathbb{R}^1 \left[0, \pi \right]$). It follows from \eqref{eq1.5} (for details see \cite{Harutyunyan:2008}), that
$$\delta_n \left(\alpha, \beta \right)=\cfrac{\cot \beta-\cot \alpha}{\pi n}+O\left( 1/n^2 \right), \; \alpha, \beta \in \left(0, \pi \right), \eqno (1.5\mbox{a})$$
$$\delta_n \left(\pi, \beta \right)=\cfrac{1}{2}+\cfrac{\cot \beta}{\pi \left(n+\frac{1}{2}\right)}+O\left( 1/n^2 \right)=\cfrac{1}{2}+O\left( 1/n \right), \; \beta \in \left(0, \pi \right), \eqno (1.5\mbox{b})$$
$$\delta_n \left(\alpha, 0 \right)=\cfrac{1}{2}-\cfrac{\cot \alpha}{\pi \left(n+\frac{1}{2}\right)}+O\left( 1/n^2 \right)=\cfrac{1}{2}+O\left( 1/n \right), \; \alpha \in \left(0, \pi \right), \eqno (1.5\mbox{c})$$
$$\delta_n \left(\pi, 0 \right)=1. \; \eqno (1.5\mbox{d})$$
Let $y=\varphi \left(x, \mu, \alpha, q\right)$ and $y=\psi \left(x, \mu, \beta, q\right)$ be the solutions of \eqref{eq1.1} with initial values
$$\varphi \left(0, \mu, \alpha, q \right)=\sin \alpha, \; \varphi' \left(0, \mu, \alpha, q \right)=-\cos \alpha,$$
$$\psi \left(\pi, \mu, \beta, q \right)=\sin \beta, \; \psi' \left(\pi, \mu, \beta, q \right)=-\cos \beta.$$
The eigenvalues $\mu_n$ of $L\left(q, \alpha, \beta \right)$ are the solutions of the equation
\begin{multline*}
\Phi \left(\mu \right)=\varphi \left(\pi, \mu, \alpha, q \right) \cos \beta + \varphi' \left(\pi, \mu, \alpha, q \right) \sin \beta=\\
=-\left[\psi \left(0, \mu, \beta, q \right) \sin \alpha + \psi' \left(0, \mu, \beta, q \right) \cos \alpha \right]=0.
\end{multline*}
It is easy to see that for arbitrary $n=0,1,2,\dots,$ \, $\varphi_n \left( x \right):=\varphi \left(x, \mu_n \left(q, \alpha, \beta \right), \alpha, q \right)$ and
$\psi_n \left( x \right):=\psi \left( x, \mu_n \left(q, \alpha, \beta \right), \beta, q \right)$ are eigenfunctions, corresponding to the eigenvalue $\mu_n \left( q, \alpha, \beta \right).$ The squares of the $L^2$-norm of these eigenfunctions:
$$a_n \left( q, \alpha, \beta \right):=\int_0^{\pi}{\left| \varphi_n \left( x \right) \right|^2 dx}, \;
b_n \left( q, \alpha, \beta \right):=\int_0^{\pi}{\left| \psi_n \left( x \right) \right|^2 dx}$$
are called the norming constants.
The main results of this paper are the following theorems:
\begin{theorem}\label{thm1.1}
For norming constants $a_n$ and $b_n$ the following asymptotic formulae hold (when $n \to \infty$):
\begin{multline}\label{eq1.6}
a_n \left( q, \alpha, \beta \right) =\cfrac{\pi}{2} \left[ 1+ \cfrac {2 \, \ae_n \left( q, \alpha, \beta \right)}{\pi \left[n + \delta \left( \alpha, \beta \right)\right]} + r_n \right] \sin^2 \alpha + \\
+\cfrac{\pi}{2 \left[n + \delta_n(\alpha, \beta)\right]^2} \left[ 1+ \cfrac {2 \, \ae_n \left( q, \alpha, \beta \right)}{\pi \left[n + \delta \left( \alpha, \beta \right)\right]} + \tilde{r}_n \right]\cos^2 \alpha,
\end{multline}
\begin{multline*}
b_n \left( q, \alpha, \beta \right) =\cfrac{\pi}{2} \left[ 1+ \cfrac {2 \, \ae_n \left( q, \alpha, \beta \right)}{\pi \left[n + \delta \left( \alpha, \beta \right)\right]} + p_n \right] \sin^2 \beta + \\
+\cfrac{\pi}{2 \left[n + \delta_n(\alpha, \beta)\right]^2} \left[ 1+ \cfrac {2 \, \ae_n \left( q, \alpha, \beta \right)}{\pi \left[n + \delta \left( \alpha, \beta \right)\right]} + \tilde{p}_n \right]\cos^2 \beta,
\end{multline*}
where
\begin{equation}\label{eq1.7}
\ae_n = \ae_n \left( q, \alpha, \beta \right) = -\cfrac {1}{2} \displaystyle \int_{0}^{\pi} \left( \pi - t \right) q\left(t\right) \sin 2 \left[ n + \delta_n\left(\alpha, \beta\right)\right]t dt,
\end{equation}
$r_n = r_n \left( q, \alpha, \beta \right) = O \left( \cfrac {1} {n^2}\right)$ and $\tilde{r}_n=\tilde{r}_n \left( q, \alpha, \beta \right) = O \left( \cfrac {1} {n^2}\right)$ (the same estimate is true for $p_n$ and $\tilde{p}_n$), when $n \to \infty,$ uniformly in $\alpha, \beta \in \left[0, \pi\right]$ and $q \in {BL}^1_\mathbb{R}\left[0, \pi\right].$
\end{theorem}
\begin{theorem}\label{thm1.2}
For both $\alpha, \beta \in (0, \pi)$ and $\alpha = \pi, \; \beta = 0$ cases the function $k,$ defined as the series
\begin{equation*}
k(x) = \displaystyle \sum_{n=2}^{\infty} \cfrac {\ae_n}{n + \delta_n \left( \alpha, \beta \right)} \cos \left[n + \delta_n \left( \alpha, \beta \right)\right]x
\end{equation*}
is absolutely continuous function on arbitrary segment $\left[a,b\right] \subset \left(0, 2 \pi\right),$ i.e. $k \in AC \left(0, 2 \pi\right).$
\end{theorem}
The dependence of norming constants on $\alpha$ and $\beta$ (as far as we know) hasn't been investigated before. The dependence of spectral data (by spectral data here we understand the set of eigenvalues and the set of norming constants) on $\alpha$ and $\beta$ has been usually studied (see \cite{Naimark:1969,Levitan-Sargsyan:1970,Marchenko:1977,Isaacson-Trubowitz:1983,Isaacson-McKean-Trubowitz:1984,
Dahlberg-Trubowitz:1984,Poschel-Trubowitz:1987,Yurko:2007}) in the following sense: the boundary conditions are separated into four cases:
\begin{itemize}
\item[1)]
$\sin \alpha \neq 0, \; \sin \beta \neq 0,$ \; i.e. $\alpha, \beta \in \left( 0, \pi \right);$
\item[2)]
$\sin \alpha = 0, \; \sin \beta \neq 0,$ \; i.e. $\alpha = \pi, \; \beta \in \left( 0, \pi \right);$
\item[3)]
$\sin \alpha \neq 0, \; \sin \beta = 0,$ \; i.e. $\alpha \in \left( 0, \pi \right), \; \beta = 0;$
\item[4)]
$\sin \alpha = 0, \; \sin \beta = 0,$ \; i.e. $\alpha = \pi, \; \beta = 0,$
\end{itemize}
and results are formulated separately for each case. For eigenvalues, formula \eqref{eq1.4} generalizes and unites four different formulae that were known before in four mentioned cases (see \cite{Harutyunyan:2008}).
So far, for norming constants the following is known.
In the case $ \sin \alpha \neq 0$ it is known that for smooth $q$
\begin{equation}\label{eq1.8}
\cfrac{a_n \left( q, \alpha, \beta \right)}{\sin^2 \alpha}=\cfrac{\pi}{2}+O \left(\cfrac{1}{n^2} \right).
\end{equation}
For absolutely continuous $q$ (we will write $q \in AC \left[ 0, \pi \right]$) the proof of \eqref{eq1.8} can be found in \cite{Levitan-Sargsyan:1970}. Let us note, that if $q \in AC\left[0, \pi\right],$ then $\ae_n = O \left(\cfrac{1}{n}\right),$ and it is easy to see, that in this case \eqref{eq1.6} takes the form \eqref{eq1.8}. In \cite{Zhikov:1967}, under the condition $q \left( x \right) = \cfrac{dF\left(x\right)}{dx}$ (almost everywhere), and $\sin\alpha \neq 0,$ where $F$ is a function of bounded variation (we will write $F \in BV\left[0, \pi\right]$), the author asserts that
\begin{equation}\label{eq1.9}
\cfrac{a_n \left( q, \alpha, \beta \right)}{\sin^2 \alpha} = \cfrac{\pi}{2}+\alpha_n,
\end{equation}
where the sequence $\left\{\alpha_n\right\}_{n = 0}^\infty$ is characterized by the condition that the function $f\left( x \right):=\displaystyle \sum\limits_{n = 0}^\infty \alpha _n \cos nx$ has a bounded variation on $\left[0, \pi \right],$ i.e. $f \in BV \left[ 0, \pi \right].$ Our result is similar to this, but there are some differences, in particular, we assert that $k \in AC \left(0, 2\pi\right).$
In \cite{Yurko:2007}, for $q \in L_\mathbb{R}^2 \left[0, \pi \right],$ it was proved that
\begin{equation}\label{eq1.10}
\cfrac{a_n \left( q, \alpha, \beta \right)}{\sin^2 \alpha} = \cfrac{\pi}{2} + \cfrac{\kappa_n}{n},
\end{equation}
where $\left\{\kappa_n \right\}_{n=0}^\infty \in l^2$ (i.e. $\displaystyle \sum\limits_{n = 0}^\infty \left| \kappa _n \right|^2 < \infty$), and $\kappa_n = \ae_n + O \left( \cfrac{1}{n} \right)$ (see \eqref{eq1.7}).
It is also important to note that norming constants $a_n\left(q, \alpha, \beta \right)$ are analytic functions on $\alpha$ and $\beta$. It easily follows from formulae \eqref{eq3.1}, \eqref{eq3.2} and \eqref{eq3.4} below and from the result in \cite{Harutyunyan:2008}, which states that $\lambda_n \left( q, \alpha, \beta \right)$ $\left(\lambda_n^2 \left( q, \alpha, \beta \right)=\mu_n \left( q, \alpha, \beta \right)\right)$ depend analytically on $\alpha$ and $\beta.$
In the case $\sin \alpha =0,$ $\sin \beta \neq 0$ it is known that for smooth $q$ (for $q \in AC \left[ 0, \pi \right]$ the proof of \eqref{eq1.11} can be found in \cite{Levitan-Sargsyan:1970})
\begin{equation}\label{eq1.11}
a_n \left( q, \pi, \beta \right) = \cfrac{\pi}{2 \left(n+1/2\right)^2} \left[1+O\left(\cfrac{1}{n^2}\right)\right].
\end{equation}
Since $\delta_{n} \left( \pi, \beta \right)=\cfrac{1}{2}+O \left( \cfrac{1}{n} \right)$ (see (1.5b)), then it is easy to see that \eqref{eq1.11} follows from \eqref{eq1.6}. Besides, we see that \eqref{eq1.6} smoothly turns into \eqref{eq1.11} when $\alpha \to \pi$ (for $q \in AC \left[0, \pi\right]$).
In the case $\sin \alpha =0$, $\sin \beta =0$ the following result can be found in \cite{Levitan-Sargsyan:1970} for $q \in AC \left[ 0, \pi \right]$:
\begin{equation*}
a_n \left(q, \pi, 0\right)=\cfrac{\pi}{2 n^{2}} \left[1+O\left(\cfrac{1}{n^2}\right)\right].
\end{equation*}
We think that it is more correct to write this result in the form (note that $\delta_n \left(\pi, 0 \right)=1$)
\begin{equation}\label{eq1.12}
a_n \left(q, \pi, 0\right)=\cfrac{\pi}{2 \left( n+1 \right)^2} \left[1+O\left(\cfrac{1}{n^2}\right)\right]
\end{equation}
to keep the beginning of the enumeration of eigenvalues and norming constants starting from $0,$ but not from $1,$ as in \cite{Levitan-Sargsyan:1970}.
Our proofs of the theorems are based on the detailed study of the dependence of eigenfunctions $\varphi_{n}$ and $\psi_{n}$ on parameters $\alpha$ and $\beta.$ We will present it in the sections \ref{sec3} and \ref{sec4}. But first we need to prove some properties of the solutions of the equation \eqref{eq1.1}.
\section{Asymptotics of the solutions}
\label{sec2}
Let $q\in L_\mathbb{C}^1 \left[ 0, \pi \right],$ i.e. $q$ is a complex-valued, summable function on $\left[0, \pi\right],$ and let us denote by $y_i \left( x, \lambda \right),$ $i=1,2,3,4,$ the solutions of the equation
\begin{equation}\label{eq2.1}
-y''+q \left( x \right)y=\lambda^2 y,
\end{equation}
satisfying the initial conditions
\begin{equation}\label{eq2.2}
\begin{aligned}
y_1 \left(0, \lambda \right)=1, \; y_2 \left(0, \lambda \right)=0, \; y_3 \left(\pi, \lambda \right)=1, \; y_4 \left(\pi, \lambda \right)=0, \\
y'_1 \left(0, \lambda \right)=0, \; y'_2 \left(0, \lambda \right)=1, \; y'_3 \left(\pi, \lambda \right)=0, \; y'_4 \left(\pi, \lambda \right)=1.
\end{aligned}
\end{equation}
Let us recall that by a solution of \eqref{eq2.1} (which is the same as \eqref{eq1.1}) we understand the function $y,$ such that $y, \, y' \in AC \left[0, \pi \right]$ and which satisfies \eqref{eq2.1} almost everywhere (see \cite{Naimark:1969}).
The solutions $y_1$ and $y_2$ (as well as the second pair $y_3$ and $y_4$) form a fundamental system of solutions of \eqref{eq1.1}, i.e. any solution $y$ of \eqref{eq1.1} can be represented in the form:
\begin{equation}\label{eq2.3}
y \left(x \right)=y \left(0 \right)y_1 \left(x, \lambda \right)+y' \left(0 \right)y_2\left(x, \lambda \right)=y \left(\pi \right)y_3 \left(x, \lambda \right)+y' \left(\pi \right)y_4 \left(x, \lambda \right).
\end{equation}
The existence and uniqueness of the solutions $y_i,$ $i=1,2,3,4$ (under the condition $q \in L_\mathbb{C}^1 \left[0, \pi\right]$) were investigated in \cite{Naimark:1969,Marchenko:1952,Chudov:1949,Atkinson:1964,Harutyunyan-Hovsepyan:2005}. The following lemma in some sense extends the results of the mentioned papers related to asymptotics (when $\left| \lambda \right| \to \infty$) of the solutions $y_i,$ $i=1,2,3,4.$
\begin{lemma}\label{lem2.1}
Let $q \in L_\mathbb{C}^1 \left[0, \pi\right].$ Then for the solutions $y_i,$ $i=1,2,3,4,$ the following representations hold (when $\left| \lambda \right| \geq 1$):
\begin{equation}\label{eq2.4}
y_1 \left(x, \lambda \right)=\cos \lambda x+\cfrac{1}{2\lambda} \, a \left(x, \lambda \right),
\end{equation}
\begin{equation}\label{eq2.5}
y_2 \left(x, \lambda \right)=\cfrac{\sin \lambda x}{\lambda}-\cfrac{1}{2 \lambda^2} \, b \left(x, \lambda \right),
\end{equation}
\begin{equation}\label{eq2.6}
y_3 \left(x, \lambda \right)=\cos \lambda \left(\pi -x \right)+\cfrac{1}{2\lambda} \, c \left(x, \lambda \right),
\end{equation}
\begin{equation}
y_4 \left(x, \lambda \right)=\cfrac{\sin \lambda \left(\pi -x \right)}{\lambda}-\cfrac{1}{2 \lambda^2} \, d\left(x, \lambda \right),
\end{equation}
where $a$, $b$, $c$, $d$ are twice differentiable with respect to $x$ and entire functions with respect to $\lambda,$ and have the form
\begin{equation}\label{eq2.8}
a\left(x, \lambda\right)= \sin \lambda x \displaystyle \int \limits^x_0 q\left(t\right) dt + \displaystyle \int \limits^x_0 q\left(t\right) \sin \lambda \left(x-2t\right) dt + R_1\left(x, \lambda, q\right),
\end{equation}
\begin{equation}\label{eq2.9}
b\left(x, \lambda\right)= \cos \lambda x \displaystyle \int \limits^x_0 q\left(t\right) dt - \displaystyle \int \limits^x_0 q\left(t\right) \cos \lambda \left(x-2t\right) dt + R_2\left(x, \lambda, q\right),
\end{equation}
\begin{equation}\label{eq2.10}
c\left(x, \lambda\right)= \sin \lambda \left(\pi -x\right) \displaystyle \int \limits^\pi_x q\left(t\right) dt + \displaystyle \int\limits^\pi_x q\left(t\right) \sin\lambda \left(2t-\pi-x\right) dt + R_3\left(x,\lambda, q\right),
\end{equation}
\begin{equation}\label{eq2.11}
d\left(x, \lambda\right)= \cos \lambda \left(\pi -x\right) \displaystyle \int \limits^\pi_x q\left(t\right) dt + \displaystyle \int \limits^\pi_x q\left(t\right) \cos \lambda \left(\pi+x-2t\right) dt + R_4\left(x,\lambda, q\right),
\end{equation}
and $R_i,$ $i=1,2,3,4,$ satisfy the estimates (when $\left| \lambda \right| \geq 1$)
\begin{equation}\label{eq2.12}
R_1 \left(x, \lambda, q\right), \; R_2 \left(x, \lambda, q\right) = O \left( \cfrac{e^{\left| Im \lambda \right| x}}{\left| \lambda \right|} \right),
\end{equation}
\begin{equation}\label{eq2.13}
R_3 \left(x, \lambda, q\right), \; R_4 \left(x, \lambda, q\right) = O \left( \cfrac{e^{\left| Im \lambda \right| (\pi -x)}}{\left| \lambda \right|} \right),
\end{equation}
uniformly with respect to $q \in BL_{\mathbb{C}}^1 \left[0, \pi \right].$
\end{lemma}
\begin{proof} In \cite{Harutyunyan-Hovsepyan:2005} the authors have proved that $y_2 \left(x, \lambda \right)$ can be obtained as a sum of series
$$y_2 \left(x, \lambda, q \right)=\displaystyle \sum_{k=0}^{\infty}{S_k \left(x, \lambda, q \right)},$$
which converge to $y_2 \left(x, \lambda, q \right)$ uniformly on bounded subsets of the set $\left[0, \pi \right] \times \mathbb{C} \times L_\mathbb{C}^1 \left[0, \pi\right],$ and where $S_0 \left(x, \lambda, q \right)=\cfrac{\sin \lambda x}{\lambda},$
$$S_k \left(x, \lambda, q \right)=\displaystyle \int_0^x \cfrac{\sin \lambda \left(x-t \right)}{\lambda} \, q\left(t\right) S_{k-1} \left(t, \lambda, q \right)dt, \; k=1,2,\dots.$$
For $S_k$ we have the estimate (when $\left| \lambda \right| \geq 1$):
\begin{equation}\label{eq2.14}
\left| S_k \left(x, \lambda, q \right) \right| \leq \cfrac{e^{\left| Im \lambda \right| x}}{{\left| \lambda \right|}^{k+1}} \; \cfrac{\sigma_0^k \left(x\right)}{k!}, \; k=0,1,2,\dots,
\end{equation}
where $\sigma_0 \left( x \right) \equiv \displaystyle \int_0^x \left| q \left(t \right) \right|dt$ (see \cite{Harutyunyan-Hovsepyan:2005}). To prove \eqref{eq2.5}, \eqref{eq2.9} and the estimate \eqref{eq2.12}, we write $S_1$ in the form
\begin{multline*}
S_1 \left(x, \lambda, q \right)=\displaystyle \int_0^x \cfrac{\sin \lambda \left(x-t \right)}{\lambda} \; \cfrac{\sin \lambda t}{\lambda} \, q \left(t \right)dt=\\
=\cfrac{1}{2 \lambda^2} \displaystyle \int_0^x \left[\cos \lambda \left(x-2t \right)-\cos \lambda x \right] q \left(t \right)dt=\\
=-\cfrac{\cos \lambda x}{2 \lambda^2} \displaystyle \int_0^x q \left(t \right)dt+\cfrac{1}{2 \lambda^2} \displaystyle \int_0^x{\cos \lambda \left(x-2t \right)} q \left(t \right)dt,
\end{multline*}
and note that
\begin{equation*}
S'_k \left(x, \lambda, q \right)=\displaystyle \int_0^x {\cos \lambda \left(x-t \right)} q \left(t \right) S_{k-1} \left(t, \lambda, q \right)dt, \; k=1,2,\dots.
\end{equation*}
This implies that $S'_k \in AC \left[0, \pi \right].$ By writing $y_2 \left(x, \lambda, q \right)=S_0+S_1+\displaystyle \sum_{k=2}^\infty S_k \left(x, \lambda, q\right),$ we obtain
\begin{equation*}
y_2 \left(x, \lambda, q \right)=\cfrac{\sin \lambda x}{\lambda}-\cfrac{1}{2 \lambda^2} \, b\left(x, \lambda \right),
\end{equation*}
where $-\cfrac{1}{2 \lambda^2} \, b \left(x, \lambda \right)=\displaystyle\sum_{k=1}^\infty S_k \left(x, \lambda, q\right)$ and therefore $b \left(x, \lambda \right)$ has the form \eqref{eq2.9}, where $R_2 \left(x, \lambda\right)= -2 \lambda^2 \displaystyle\sum_{k=2}^\infty S_k \left(x, \lambda\right).$ Now, from the estimate \eqref{eq2.14}, we obtain that
\begin{multline*}
\sum_{k=2}^\infty \left|S_k \left(x, \lambda, q \right) \right| \leq \sum_{k=2}^\infty \cfrac{e^{\left| Im \lambda \right|x}}{\left| \lambda \right|^{k+1}} \; \cfrac{\sigma_0^k \left(x \right)}{k!}=\cfrac{e^{\left| Im \lambda \right|x} \, \sigma_0^2 \left(x \right)}{\left| \lambda \right|^3} \sum_{k=2}^{\infty} \cfrac{\sigma_0^{k-2} \left(x \right)}{\left| \lambda \right|^{k-2} k!}< \\
<\cfrac{e^{\left| Im \lambda \right|x} \, \sigma_0^2 \left(x \right)}{\left| \lambda \right|^3} \sum_{k=2}^{\infty} \cfrac{\sigma_0^{k-2} \left(x \right)}{\left| \lambda \right|^{k-2} \left(k-2 \right)!}=\cfrac{e^{\left| Im \lambda \right|x} \, \sigma_0^2 \left(x\right)}{\left|\lambda \right|^3} \sum_{n=0}^{\infty} \cfrac{\sigma_0^n \left(x \right)}{\left| \lambda \right|^n n!}= \\
=\cfrac{e^{\left| Im \lambda \right|x} \, \sigma_0^2 \left(x \right)}{\left| \lambda \right|^3} \; e^{\tfrac{\sigma_0 \left(x \right)}{\left| \lambda \right|}}=\cfrac{\sigma_0^2 \left(x \right)}{\left| \lambda \right|^3} \; e^{\left| Im \lambda \right|x+\tfrac{\sigma_0 \left(x \right)}{\left|\lambda \right|}}.
\end{multline*}
This implies \eqref{eq2.12} for $\left| \lambda \right| \geq 1$. Since $y'_2 \in AC \left[ 0, \pi \right]$ and $S_0 \left( x, \lambda \right)=\cfrac{\sin \lambda x}{\lambda},$ then we obtain that $-\cfrac{1}{2 \lambda^2} \,
b \left( x, \lambda \right)= y_2 - S_0$ is also a twice differentiable function (more precisely $b' \in AC \left[ 0, \pi \right]$). Assertions for $y_1,$ $y_3,$ $y_4$ can be proven similarly.
\end{proof}
\section{The proof of the Theorem \ref{thm1.1}}
\label{sec3}
According to \eqref{eq2.3}, the solution $\varphi \left(x, \mu, \alpha, q \right),$ which we will denote by $\varphi \left(x, \lambda^2, \alpha \right)$ for brevity, has the form
\begin{equation}\label{eq3.1}
\varphi \left(x, \lambda^2, \alpha \right)=y_1 \left(x, \lambda \right) \sin \alpha - y_2 \left(x, \lambda \right) \cos \alpha,
\end{equation}
and according to \eqref{eq2.4} and \eqref{eq2.5} we arrive at:
\begin{equation*}
\varphi \left(x, \lambda^2, \alpha \right)=\left[\cos \lambda x+\cfrac{1}{2\lambda} \, a \left(x, \lambda \right) \right] \sin \alpha - \left[\cfrac{\sin \lambda x}{\lambda}-\cfrac{1}{2 \lambda^2} \, b \left(x, \lambda \right) \right] \cos \alpha.
\end{equation*}
Taking the squares of both sides of the last equality, we obtain:
\begin{multline}\label{eq3.2}
\varphi^2 \left(x, \lambda^2, \alpha \right)=\cos^2 \lambda x \sin^2 \alpha + \cfrac{1}{\lambda} \left[a\left(x,\lambda \right) \cos \lambda x + \cfrac{a^2 \left(x, \lambda \right)}{4\lambda}\right] \sin^2 \alpha - \\
-\cfrac{2}{\lambda} \left[\cos \lambda x \sin \lambda x-\cfrac{b\left(x, \lambda \right) \cos \lambda x} {2\lambda}+\cfrac{a\left(x, \lambda \right) \sin \lambda x}{2\lambda}-\cfrac{a\left(x, \lambda \right) b\left(x, \lambda \right)}{4 \lambda^2} \right] \times \\
\times \sin \alpha \cos \alpha +\cfrac{\sin^2 \lambda x}{\lambda^2} \, \cos^2 \alpha +\left[\cfrac{b^2 \left(x, \lambda \right)}{4 \lambda^4}-\cfrac{b\left(x, \lambda \right) \sin \lambda x}{\lambda^3}\right] \cos^2 \alpha. \end{multline}
Recalling the formulae $\cos^2 \lambda x = \cfrac{1}{2}\left(1+\cos 2 \lambda x \right)$ and $\sin^2 \lambda x = \cfrac{1}{2} \left(1-\cos 2 \lambda x \right),$ from \eqref{eq3.2}, we obtain:
\begin{multline}\label{eq3.3}
\displaystyle \int_0^{\pi} \varphi^2 \left(x, \lambda^2, \alpha \right)dx = \cfrac{\pi}{2} \sin^2 \alpha +\cfrac{\sin 2\lambda \pi}{4\lambda} \, \sin^2 \alpha + \\
+\cfrac{1}{\lambda}\left(\int_0^{\pi} a\left(x, \lambda \right) \cos \lambda xdx + \cfrac{1}{4\lambda} \int_0^{\pi} a^2 \left(x, \lambda \right)dx \right) \sin^2 \alpha-\cfrac{\sin^2 \lambda \pi}{\lambda^2} \sin \alpha \cos \alpha + \\
+\cfrac{1}{\lambda^2} \left(\displaystyle\int_0^{\pi} b\left(x, \lambda \right) \cos \lambda xdx - \displaystyle\int_0^{\pi} a\left(x, \lambda \right) \sin \lambda xdx \right) \sin \alpha \cos \alpha + \\
+\cfrac{\sin \alpha \cos \alpha}{2\lambda^3} \int_0^{\pi} a\left(x, \lambda \right) b\left(x, \lambda \right) dx + \cfrac{\pi}{2 \lambda^2} \, \cos^2 \alpha - \cfrac{\sin 2\lambda \pi}{4 \lambda^3} \, \cos^2 \alpha - \\
-\cfrac{1}{\lambda^3}\left(\displaystyle\int_0^{\pi} b\left(x, \lambda \right) \sin \lambda xdx - \cfrac{1}{4\lambda}\displaystyle\int_0^{\pi} b^2 \left(x, \lambda \right)dx \right) \cos^2 \alpha.
\end{multline}
We are going to receive the asymptotic formula \eqref{eq1.6} by the substitution $\lambda=\lambda_n \left(q, \alpha, \beta \right)=\sqrt{\mu_n \left(q, \alpha, \beta \right)}$ in \eqref{eq3.3}. To this aim, we estimate each term of the right-hand side of \eqref{eq3.3} for $\lambda ={\lambda}_n$. It can be easily deduced from \eqref{eq1.4} that for $\lambda_n = \sqrt{\mu_n}$ we have the following asymptotic formula:
\begin{equation}\label{eq3.4}
\lambda_{n} \left( q, \alpha, \beta \right) = n+ \delta_{n} \left( \alpha, \beta \right)+ \cfrac{\left[ q \right]}{2{\left( n+\delta_{n} \left( \alpha, \beta \right) \right)}}+l_{n},
\end{equation}
where $ \left[ q \right] := \cfrac{1}{\pi} \displaystyle\int_0^{\pi} q \left(t\right)dt,$ $l_n=l_n \left(q, \alpha, \beta \right)=o \left(\cfrac{1}{n} \right)$ uniformly with respect to $\alpha, \beta \in \left[0, \pi \right]$ and $q \in {BL}_\mathbb{R}^1 \left[0,\pi \right]$ (see \cite{Harutyunyan:2016}).
It follows from (1.5a)--(1.5d) that $\sin 2\pi \delta_n \left(\alpha, \beta \right)=O\left(\cfrac{1}{n}\right)$ and
\begin{equation}\label{eq3.5}
\sin 2\pi \lambda_n \left(q, \alpha, \beta\right) = O\left(\cfrac{1}{n} \right), \; \cos 2\pi \lambda_n \left( q, \alpha, \beta \right) = 1 - O\left(\cfrac{1}{n^2} \right),
\end{equation}
for all $\left(\alpha, \beta\right) \in \left(0, \pi \right] \times \left[0, \pi \right).$
Thus, the second term
\begin{equation}\label{eq3.6}
\cfrac{\sin 2\pi \lambda_n}{\lambda_n} \, \sin^2 \alpha = O\left( \cfrac{1}{n^2} \right) \sin^2 \alpha.
\end{equation}
Important is the third term: $\cfrac{1}{\lambda_n} \displaystyle \int_0^{\pi} a\left(x, \lambda_n \right) \cos \lambda_n x dx.$ According to \eqref{eq2.8} and \eqref{eq2.12} we have
\begin{equation}\label{eq3.7}
a \left(x, \lambda_n\right) = A \left(x, \lambda_n\right) + O \left( \cfrac {1}{\lambda_n}\right),
\end{equation}
where
\begin{equation}\label{eq3.8}
A \left(x, \lambda_n\right) = \displaystyle \int_0^x q\left(t\right)dt \sin \lambda_n x + \int_0^x q\left(t\right)\sin \lambda_n \left(x - 2t\right) dt.
\end{equation}
After multiplying both sides by $\cos \lambda_n x,$ integrating over $\left[0,\pi\right]$ and changing the order of integration we get
\begin{multline}\label{eq3.9}
\displaystyle \int_0^{\pi} A\left(x, \lambda_n\right) \cos \lambda_n x dx =
\cfrac{\sin^2 \lambda_n \pi}{\lambda_n} \displaystyle \int_0^{\pi} q\left(t\right) \cos^2 \lambda_n t dt - \\
- \cfrac{\sin 2 \lambda_n \pi}{4 \lambda_n} \displaystyle \int_0^{\pi} q\left(t\right) \sin 2 \lambda_n t dt - \cfrac{1}{2} \displaystyle \int_0^{\pi} \left(\pi - t\right) q\left(t\right) \sin 2 \lambda_n tdt.
\end{multline}
Taking into account the formulae \eqref{eq3.5} and denoting (see \eqref{eq1.7})
\begin{equation*}
\tilde{\ae}_n \equiv \tilde{\ae}_n \left( q, \alpha, \beta \right) := -\cfrac {1}{2} \displaystyle \int_0^{\pi} \left( \pi - t \right) q\left(t\right) \sin 2 \lambda_n tdt,
\end{equation*}
we can rewrite \eqref{eq3.9} in the form
\begin{equation}\label{eq3.10}
\displaystyle \int_{0}^{\pi} A(x, \lambda_n) \cos \lambda_n x dx = \tilde {\ae}_n + O \left( \cfrac{1}{n} \right).
\end{equation}
Since $\sin 2 \lambda_n t = \sin 2 \left(n+\delta_n + O \left( \cfrac{1}{n} \right)\right)t = \sin 2 \left(n+\delta_n \right)t + O \left( \cfrac{1}{n} \right)$ holds uniformly with respect to $t \in \left[0, \pi \right],$ then $\tilde{\ae}_n = \ae_n + O \left( \cfrac{1}{n} \right),$ and therefore the third term of \eqref{eq3.3} has the form
\begin{equation*}
\cfrac{1}{\lambda_n} \displaystyle \int_{0}^{\pi} a(x, \lambda_n) \cos \lambda_n x dx=\cfrac{\ae_n}{n + \delta_n (\alpha, \beta)} + O \left( \cfrac{1}{n^2} \right).
\end{equation*}
Now, let us focus on the remained terms of the equality \eqref{eq3.3} for $\lambda=\lambda_n.$ The terms from the fourth to the eighth have the coefficient $\cfrac{1}{\lambda_n^{\gamma}},$ where $\gamma \geq 2,$ and therefore they have the order $O \left( \cfrac{1}{n^2} \right).$ Concerning the last four terms of \eqref{eq3.3}, we observe that both $\cfrac{\sin 2 \pi \lambda_n}{4 \lambda_n^3} \cos^2 \alpha$ and $\cfrac{1}{\lambda_n^4} \displaystyle \int_{0}^{\pi} b^2 \left(x, \lambda_n\right) dx \cos^2 \alpha$ have the same order $O \left( \cfrac{1}{\lambda_n^4} \right) \cos^2 \alpha.$ An important term is $\cfrac{1}{\lambda_n^3} \displaystyle \int_{0}^{\pi} b\left(x, \lambda_n\right) \sin \lambda_n x dx.$ According to \eqref{eq2.9} and \eqref{eq2.12} we can write $b \left(x, \lambda_n\right)$ in the form
\begin{equation}\label{eq3.11}
b \left(x, \lambda_n \right) = B \left(x, \lambda_n \right) + O \left( \cfrac{1}{\lambda_n} \right),
\end{equation}
where
\begin{equation}\label{eq3.12}
B\left(x, \lambda_n\right) = \displaystyle \int_{0}^{x} q\left(t\right) dt \cos \lambda_n x - \displaystyle \int_{0}^{x} q\left(t\right) \cos \lambda_n \left(x-2t\right) dt.
\end{equation}
A simple computation yields:
\begin{equation*}
B\left(x, \lambda_n \right) \sin \lambda_n x = \displaystyle \int_{0}^{x} q\left(t\right) dt \sin 2 \lambda_n x - \displaystyle \int_{0}^{x} q\left(t\right) \sin 2 \lambda_n t dt - A\left(x, \lambda_n \right) \cos \lambda_n x.
\end{equation*}
After integrating the latter equality from $0$ to $\pi,$ changing the order of integration and taking into consideration \eqref{eq3.9} we get:
\begin{multline*}
\displaystyle \int_{0}^{\pi} B\left(x,\lambda_n\right) \sin \lambda_n x dx = - \cfrac{\cos^2 \lambda_n \pi}{\lambda_n} \displaystyle \int_{0}^{\pi} q\left(t\right) \sin^2 \lambda_n t dt + \\
+\cfrac{\sin 2 \lambda_n \pi}{4 \lambda_n} \displaystyle \int_{0}^{\pi} q\left(t\right) \sin 2 \lambda_n t dt
- \cfrac{1}{2} \displaystyle \int_{0}^{\pi} \left(\pi - t\right) q\left(t\right) \sin 2 \lambda_n t dt = \\
= O \left( \cfrac{1}{\lambda_n} \right) + O \left( \cfrac{1}{\lambda_n^2} \right) + \tilde{\ae}_n=\ae_n+O\left(\cfrac{1}{n}\right),
\end{multline*}
and therefore the eleventh term of the equality \eqref{eq3.3} for $\lambda=\lambda_n$ has the form
\begin{equation*}
\cfrac{1}{\lambda_n^3}\displaystyle\int_0^{\pi} b\left(x, \lambda_n \right) \sin \lambda_n xdx=\cfrac{1}{\lambda_n^2} \left(\cfrac{\ae_n}{n+\delta_n \left(\alpha,\beta\right)}+O\left(\cfrac{1}{n^2}\right)\right).
\end{equation*}
Let us remark that from \eqref{eq3.4} we have $\cfrac{1}{\lambda_n} - \cfrac{1}{n+\delta_n} = O \left( \cfrac{1}{n^3} \right).$ Thus,
\begin{multline}\label{eq3.13}
a_n \left(q, \alpha, \beta \right) = \cfrac{\pi}{2} \left[ 1+ \cfrac{2 \, \ae_n}{\pi \left[n+\delta_n \left(\alpha,\beta\right)\right]} + O \left( \cfrac{1}{n^2} \right) \right] \sin^2 \alpha + O \left( \cfrac{1}{n^2} \right) \sin \alpha \cos \alpha+\\
+\cfrac{\pi}{2 \left[n+\delta_n \left(\alpha,\beta\right)\right]^2} \left[ 1+ \cfrac{2 \, \ae_n}{\pi \left[n+\delta_n \left(\alpha,\beta\right)\right]} + O \left( \cfrac{1}{n^2} \right) \right] \cos^2 \alpha.
\end{multline}
If $\sin \alpha \neq 0,$ then $O \left( \cfrac{1}{n^2} \right) \sin \alpha \cos \alpha$ can be included into the term $O \left( \cfrac{1}{n^2} \right) \sin^2 \alpha,$ and if $\sin \alpha = 0,$ then these terms are absent. Finally, we can write \eqref{eq3.13} in the form \eqref{eq1.6}. For $b_n$ everything can be done similarly. Theorem \ref{thm1.1} is proved.
\section{The proof of the Theorem \ref{thm1.2}}
\label{sec4}
In the sequel the following notations will be used:
\begin{equation}\label{eq4.1}
\tilde{q} \left(t \right):= \left(\pi-t\right)q\left(t\right) \;\; \mbox{and} \;\; \sigma\left(x\right) := \displaystyle \int_{0}^{x} \tilde{q} \left(t \right) dt = \displaystyle \int_{0}^{x} \left(\pi - t \right) q \left(t \right) dt.
\end{equation}
Now, we have
\begin{multline}\label{eq4.2}
\cfrac{\ae_n}{n+\delta_n\left(\alpha, \beta\right)} = -\cfrac{1}{2 \left[n+\delta_n \left(\alpha, \beta\right)\right]} \displaystyle \int_{0}^{\pi} \tilde {q} \left(t\right) \sin 2 \left( n+\delta_n \right)t
dt = \\
=-\cfrac{1}{2 \left(n+\delta_n\right)} \displaystyle \int_{0}^{\pi} \sin 2 \left(n+\delta_n\right) t d \sigma \left(t\right) = -\cfrac{\sigma \left(\pi\right) \sin 2 \pi \delta_n}{2 \left(n+\delta_n \right)} + \\
+\displaystyle \int_{0}^{\pi} \sigma\left(t\right) \cos 2 \left(n+\delta_n \right)t dt.
\end{multline}
It was observed in \eqref{eq3.5} that $\sin 2 \pi \delta_n = O \left( \cfrac{1}{n} \right)$. If we denote by $\tilde{\sigma} \left(x\right) := \sigma\left(\cfrac{x}{2}\right)$ and $c_n := \cfrac {\sin 2 \pi \delta_n}{2 \left(n+\delta_n\right)}=O\left(\cfrac{1}{n^2}\right),$ then we can rewrite $k\left(x \right)$ in the form
\begin{equation}\label{eq4.3}
k\left(x\right) = k_1 \left(x\right) + k_2 \left(x\right),
\end{equation}
where
\begin{equation}\label{eq4.4}
k_1 \left(x\right) = -\sigma\left(\pi\right) \displaystyle \sum_{n=2}^{\infty} c_n \cos \left[n+\delta_n \left(\alpha, \beta \right) \right]x,
\end{equation}
\begin{equation}\label{eq4.5}
k_2 \left(x\right) = \displaystyle \sum_{n=2}^{\infty} \displaystyle \int_{0}^{2 \pi} \tilde{\sigma}\left(t\right) \cos \left[ n+\delta_n \left(\alpha, \beta \right) \right]t dt \cos \left[n+\delta_n \left(\alpha, \beta \right) \right]x.
\end{equation}
Since $c_n=O\left(\cfrac{1}{n^2}\right),$ then the series in \eqref{eq4.4} converges absolutely and uniformly on $\left[0,2\pi \right],$ and $k_1 \in AC \left[0,2\pi \right].$
Next, we consider two cases:
{\bf Case I:} If $\alpha, \beta \in (0, \pi),$ then by (1.5a) we have
\begin{equation*}
\delta_n(\alpha, \beta)=\cfrac{\cot \beta - \cot \alpha}{\pi n} + O \left( \cfrac{1}{n^2} \right)= \cfrac{d}{n} + O \left( \cfrac{1}{n^2}\right)=O \left( \cfrac{1}{n}\right),
\end{equation*}
where $d=\cfrac{\cot \beta - \cot \alpha}{\pi}.$
Recalling the Maclaurin expansions of the functions $\sin x$ and $\cos x$ around the point $x=0,$ we obtain
\begin{equation}\label{eq4.6}
\cos \left[n+\delta_n \left(\alpha, \beta\right)\right]x= \cos nx - d \cdot x \, \cfrac{\sin nx}{n}+ e_n \left(x\right),
\end{equation}
where $e_n \left(x\right),$ as all the other entries of \eqref{eq4.6}, is a smooth function $\left(e_n \in C^{\infty}\right)$ and
\begin{equation}\label{eq4.7}
e_n \left(x\right)=O\left(\cfrac{1}{n^2}\right)
\end{equation}
uniformly on $x \in \left[0,2\pi \right].$ Therefore $k_2$ can be written in the form
\begin{equation*}
k_2 \left(x\right) = l_1 \left(x\right)+l_2 \left(x\right)+l_3 \left(x\right),
\end{equation*}
where
\begin{multline}\label{eq4.8}
l_1 \left(x\right)=-d \cdot x \displaystyle \sum_{n=2}^{\infty} \cfrac{1}{n} \int_{0}^{2 \pi} \tilde{\sigma}\left(t\right) \cos nt dt \sin nx - \\
-d \cdot \displaystyle \sum_{n=2}^{\infty} \cfrac{1}{n} \int_{0}^{2 \pi} t \tilde{\sigma}\left(t\right) \sin nt dt \cos nx + d^2 \cdot x \displaystyle \sum_{n=2}^{\infty} \cfrac{1}{n^2} \int_{0}^{2 \pi} t \tilde{\sigma}\left(t\right) \sin nt dt \sin nx + \\
+\displaystyle \sum_{n=2}^{\infty} \int_{0}^{2 \pi} e_n \left(t \right) \tilde{\sigma}\left(t\right) dt \cos nx -d \cdot x \displaystyle \sum_{n=2}^{\infty} \cfrac{1}{n} \int_{0}^{2 \pi} e_n \left(t \right) \tilde{\sigma}\left(t\right) dt \sin nx,
\end{multline}
\begin{multline}\label{eq4.9}
l_2 \left(x\right)=\displaystyle \sum_{n=2}^{\infty} e_n \left(x \right) \int_{0}^{2 \pi} \tilde{\sigma}\left(t\right) \cos nt dt - \\
-d \cdot \displaystyle \sum_{n=2}^{\infty} \cfrac{e_n \left(x \right)}{n} \int_{0}^{2 \pi} t \tilde{\sigma}\left(t\right) \sin nt dt + \displaystyle \sum_{n=2}^{\infty} e_n \left(x \right) \int_{0}^{2 \pi} e_n \left(t \right) \tilde{\sigma}\left(t\right) \sin nt dt,
\end{multline}
\begin{equation}\label{eq4.10}
l_3 \left(x\right)=\displaystyle \sum_{n=2}^{\infty}\displaystyle \int_{0}^{2 \pi} \tilde{\sigma}\left(t\right) \cos nt dt \cos nx.
\end{equation}
Since $\tilde \sigma \in AC [0, 2 \pi],$ then Fourier coefficients are
\begin{equation}\label{eq4.11}
\displaystyle \int_{0}^{2 \pi} \tilde{\sigma}\left(t\right) \cos nt dt = O \left( \cfrac{1}{n} \right), \;
\displaystyle \int_{0}^{2 \pi} t\tilde{\sigma}\left(t\right) \sin nt dt = O \left( \cfrac{1}{n} \right).
\end{equation}
Also we note that
\begin{equation}\label{eq4.12}
\displaystyle \int_{0}^{2 \pi} e_n \left(t \right) \tilde{\sigma}\left(t\right) dt=O\left(\cfrac{1}{n^2}\right).
\end{equation}
Therefore the trigonometric series in \eqref{eq4.8} converges absolutely and uniformly on $\left[0,2\pi \right],$ and $l_1 \in AC \left(0,2\pi \right).$
It follows from \eqref{eq4.7}, \eqref{eq4.11} and \eqref{eq4.12} that the terms of the series in \eqref{eq4.9} have the order $O\left(\cfrac{1}{n^3}\right),$ and therefore $l_2 \in AC \left[0, 2\pi \right].$
About $l_3 \left(x\right)$ we can say the following:
Since $\tilde{\sigma} \in AC \left[0, 2 \pi\right],$ then the Fourier series of $\tilde{\sigma}$
\begin{equation*}
\tilde{\sigma}\left(x\right) =
\cfrac{a_0 \left(\tilde{\sigma}\right)}{2} + \displaystyle \sum_{n=1}^{\infty}\left( a_n \left(\tilde{\sigma}\right) \cos nx + b_n \left(\tilde{\sigma}\right) \sin x\right),
\end{equation*}
where $a_n \left(\tilde{\sigma}\right)=\cfrac{1}{\pi} \displaystyle \int_{0}^{2\pi} \tilde{\sigma} \left(t\right) \cos ntdt,$ $b_n \left(\tilde{\sigma}\right)=\cfrac{1}{\pi} \displaystyle \int_{0}^{2\pi} \tilde{\sigma}\left(t\right) \sin nt dt,$
converges to $\tilde{\sigma}\left(x\right)$ in every point of $\left[0, 2 \pi\right]$ and this series is a function from $AC\left[0, 2 \pi\right].$
The same is true for $\sigma^{\star} (x) = \tilde{\sigma}\left(2 \pi - x\right):$
\begin{equation*}
\tilde{\sigma}\left(2 \pi - x\right) = \cfrac{a_0 \left(\sigma^{\star}\right)}{2} + \displaystyle \sum_{n=1}^{\infty}\left( a_n \left(\sigma^{\star}\right) \cos nx + b_n \left(\sigma^{\star}\right) \sin x \right).
\end{equation*}
But it is easy to see, that $a_n \left(\sigma^{\star}\right) = a_n \left(\tilde{\sigma}\right)$ and $b_n \left(\sigma^{\star}\right) = -b_n \left(\tilde{\sigma}\right).$ So
\begin{equation*}
\cfrac{1}{2} \left( \tilde{\sigma}\left(x\right) + \tilde{\sigma}\left(2\pi - x\right) \right) = \cfrac{a_0 \left(\tilde{\sigma}\right)}{2} + \displaystyle \sum_{n=1}^{\infty} a_n \left(\tilde{\sigma}\right) \cos nx,
\end{equation*}
i.e. this is "the even part" of Fourier series of $\tilde{\sigma} \left(x\right),$ and is absolutely continuous on $\left[0, 2\pi\right].$
Thus, for the case $\alpha, \beta \in \left(0, \pi\right)$ Theorem \ref{thm1.2} is proved.
{\bf Case II:} If $\alpha = \pi, \beta = 0,$ then $\delta_n \left(\pi, 0 \right) = 1,$ and the function $k_2 \left(\cdot\right)$ takes the form $k_2 \left(x\right) = \displaystyle \sum_{n=3}^{\infty} \int_{0}^{2 \pi} \tilde{\sigma} \left(t\right) \cos nt dt \cos nx$ and again it is "the even part" of Fourier series (without the zeroth, the first and the second terms) of an absolutely continuous function. Theorem \ref{thm1.2} is proved.
\subsection*{Acknowledgment}
T.N.~Harutyunyan was supported by State Committee of Science MES RA, in frame of the research project No. 15T--1A392.
|
\section{Introduction}
Let $(X_1,Y_1),...,(X_n,Y_n)$ be an independent and identically distributed sample of the bivariate random vector $(X,Y)$ with joint cumulative distribution function $H$ and marginal distribution functions $F$ and $G$. Let $F_n$ and $G_n$ be the empirical marginal cumulative distribution functions and let $K(\cdot,\cdot)$ represent a multiplicative kernel integral $K(x,y)=K(x)K(y)$. Then, the transformation kernel estimator of copulas suggested in \cite{r6} is defined as follows :
\begin{equation}\label{ee1}
\hat{C}_{n,h}^{(T)}(u,v)=\frac{1}{n}\sum_{i=1}^{n}K\left(\frac{\phi^{-1}(u)-\phi^{-1}(\hat{U}_i)}{h} \right)K\left(\frac{\phi^{-1}(v)-\phi^{-1}(\hat{V}_i)}{h} \right),
\end{equation}
where $\phi$ is a given distribution function and $\hat{U}_i=\frac{n}{n+1}F_n(X_i)$, $\hat{V}_i=\frac{n}{n+1}G_n(Y_i)$. Taking $\phi$ equal to the standard Gaussian distribution leads to the Probit transformation proposed by Marron and Ruppert \cite{r5} (1994). Furthermore, this estimator presents a great advantage as it does not depend on the marginal distributions.
We shall consider in this paper a general bandwidth $h$ that may depend either on the sample data or/and the location $(u,v)$. It is found in the literature that, for practical use, the most interesting choice of the bandwidth for kernel distribution function estimation is the data-driven method (see, e.g., Altman and L\'eger \cite{r0} (1995)).
\section{Main results and Simulation}\label{ssec2}
\subsection{Results}
Here, we state our theoretical results in three theorems. The first theorem gives the uniform in bandwidth rate of convergence for the maximal deviation of the estimator \eqref{ee1} from its expectation. The second theorem handles the bias, while the third theorem provides optimal asymptotic simultaneous confidence bands for the copula function $C(u,v)$ defined as,
$$ C(u,v)=\mathbb{P}(U\leq u,V\leq v)\qquad\text{for all}\; 0\leq u,v\leq 1,$$
where $U$ and $V$ are $(0,1)-$uniform random variables.
\begin{theorem}\label{tt1}
Suppose that the copula function $C(u,v)$ has bounded first-order partial derivatives on $(0, 1)^2$ and the function $\phi$ admits a bounded derivative $\phi'$. Then, for any sequence of positive constants $(b_n)_{n\geq 1}$ satisfying $0<b_n<1, b_n\rightarrow 0$, $b_n\geq (\log n)^{-1}$, and for some $c>0$, we have almost surely
\begin{equation}\label{ee2}
\limsup_{n\rightarrow\infty}\left\{R_n\sup_{\frac{c\log n}{n}\le h\le b_n}\sup_{(u,v)\in(0,1)^2}\left|\hat{C}_{n,h}^{(T)}(u,v)-\mathbb{E}\hat{C}_{n,h}^{(T)}(u,v)\right|\right\}=A(c),
\end{equation}
where $A(c)$ is a positive constant less than or equal to 3, and $R_n =\left(\frac{n}{2\log\log n}\right)^{1/2}$.
\end{theorem}
\begin{theorem} \label{tt2}
Suppose that the copula function $C(u,v)$ has bounded second-order partial derivatives on $(0, 1)^2$ and the function $\phi$ admits a bounded derivative $\phi'$. Then, for any sequence of positive constants $(b_n)_{n\geq 1}$ satisfying $0<b_n<1$,\\ $\sqrt{n}b_n^2/\sqrt{\log\log n}=o(1),$ and for some $c>0$, we have almost surely
\begin{equation} \label{biais}
R_n\sup_{\frac{c\log n}{n}\le h\le b_n}\sup_{(u,v)\in(0,1)^2}\vert \mathbb{E}\hat{C}_{n,h}^{(T)}(u,v) - C(u,v)\vert \rightarrow 0,\,\, n\rightarrow\infty.
\end{equation}
\end{theorem}
\begin{remark}{\rm
We can infer from Theorem \ref{tt1} that for any data-driven bandwidth $\hat{h}_n$ such that
\begin{equation} \label{ee4'}
\mathbb{P}(\frac{c\log n}{n}\le \hat{h}_n\le b_n)\rightarrow 1,\; n\rightarrow\infty,
\end{equation}
we have
\begin{equation} \label{ee5}
\sup_{(u,v)\in(0,1)^2}\frac{R_n}{A(c)}\left|\hat{C}_{n,\hat{h}_n}^{(T)}(u,v)-\mathbb{E}\hat{C}_{n,\hat{h}_n}^{(T)}(u,v)\right|\stackrel{\mathbb{P}}{\longrightarrow}1,\ \ \ n\rightarrow \infty,
\end{equation}
where $\stackrel{\mathbb{P}}{\longrightarrow} $ stands for convergence in probability. To make use of \eqref{ee5} for providing confidence bands, we must ensure that the bias of the estimator may be neglected, in the sense that,
\begin{equation}\label{ee6}
\sup_{(u,v)\in(0,1)^2}\frac{R_n}{A(c)}\left|\mathbb{E}\hat{C}_{n,\hat{h}_n}^{(T)}(u,v)-C(u,v)\right|\stackrel{\mathbb{P}}{\longrightarrow}0,\ \ \ n\rightarrow \infty.
\end{equation}
}
\end{remark}
\begin{theorem}\label{tt3}
Suppose that assumptions of Theorem \ref{tt1} and Theorem \ref{tt2} hold and condition \eqref{ee4'} is fulfilled.
Then, for every $\epsilon \in (0,1)$, as $n\rightarrow\infty$, one has
\begin{multline}\label{lleq1}
\mathbb{P}\bigg(C(u,v)\in \left[\hat{C}_{n,\hat{h}_n}^{(T)}(u,v)-E_{n,\epsilon}(u,v),\hat{C}_{n,\hat{h}_n}^{(T)}(u,v)+E_{n,\epsilon}(u,v)\right],\forall\:0\leq u,v\leq 1 \bigg)\longrightarrow 1,
\end{multline}
\begin{multline}\label{lleq2}
\mathbb{P}\bigg(C(u,v)\in \left[\hat{C}_{n,\hat{h}_n}^{(T)}(u,v)-\Delta_{n,\epsilon}(u,v),\hat{C}_{n,\hat{h}_n}^{(T)}(u,v)+\Delta_{n,\epsilon}(u,v)\right], \forall\:0\leq u,v\leq 1 \bigg)\longrightarrow 0,
\end{multline}
where $E_{n,\epsilon}(u,v)=(1+\epsilon)\frac{A(c)}{R_n},\;\Delta_{n,\epsilon}(u,v)=(1-\epsilon)\frac{A(c)}{R_n}$.
\end{theorem}
The proofs of our theorems are direct adaptations of those of the results on the Local linear estimator established in \cite{r1}. So we omit the proofs and suggest the interested reader to follow the lines from page 2084 to 2089 in this paper. Instead, we will focus in the next subsection on simulation studies to compare the performance of our results with the confidence bands obtained from the classical asymptotic normality approach.
\begin{remark} {\rm
Whenever (\ref{lleq1}) and (\ref{lleq2}) hold jointly for every $\epsilon > 0$, we will say that the intervals
\begin{equation}\label{iinterv}
\left[\alpha_n(u,v),\beta_n(u,v)\right]= \left[\hat{C}_{n,\hat{h}_n}^{(T)}(u,v)-\frac{A(c)}{R_n}\:,\:\hat{C}_{n,\hat{h}_n}^{(T)}(u,v)+\frac{A(c)}{R_n}\right]
\end{equation}
provide asymptotic simultaneous confidence bands for the copula function $C(u,v),\ 0\leq u,v\leq 1$. These bands are optimal as their asymptotic confidence level tends to 100\%. Therefore we can write, with a probability near to 1, that for all $(u,v)\in [0,1]^2,$ as $n\rightarrow\infty$,
\begin{equation*}
C(u,v)\in\left[\alpha_n(u,v),\beta_n(u,v)\right].
\end{equation*}
}
\end{remark}
\subsection{Simulation study}
Here, we make some simulation experiments to show the finite sample performance of our confidence bands. To this end, we choose Frank copula family which has bounded second-order partial derivatives. To compute the estimator $\hat{C}_{n,\hat{h}}^{(T)}$, we employ the conditional sampling method and generate a random sample of $n$ pairs of data from Frank copula, say $C_{\theta}$, of parameter $\theta\in \mathbb{R}$, defined as
\begin{equation}
C_{\theta}(u,v)= -\frac{1}{\theta}\log\left[ 1+ \frac{(e^{-\theta u}-1)(e^{-\theta v}-1)}{(e^{-\theta}-1)} \right].
\end{equation}
To control the behavior of the data-driven bandwidth $\hat{h}_n$, we take it close to a sequence of constants $h_n=1/\log(n)$. While the kernel $K(\cdot)$ is taken as the integral of the Epanechnikov kernel density function
$k(t)=0.75(1-t^2)\mathbb{I}(|t|\leq 1)$. We then compute the lower and upper bounds of the confidence bands established in \eqref{iinterv} by substracting and adding respectively the term $A(c)/R_n$, where $A(c)= 1/2$.\\
To measure the performance of our bands, we first compute the true curve $C_{\theta}$, for specific values of $\theta$. Then, we consider $B=1000$ replications of the experiment and determine the frequency with which the bands cover the true curve $C_{\theta}(u,v)$, at all values $0\leq u,v\leq 1$. This approximates the coverage probability of our confidence bands which is reported in Table \ref{tab1}, for different values of $\theta$ and sample size $n=50,100,500$.\\
\begin{table}[htbp]
\begin{center}
$$\begin{array}{ccccc}
\hline
& \text{sample size}\, n & \theta=-2 & \theta=1 & \theta=10\\
\hline
\multirow{3}*{The proposed method}
& 50 & 0.95 & 0.96 & 0.93\\
& 100 & 0.96 & 0.97 &0.94\\
& 500& 0.98 & 0.99 &0.99\\
\hline
\multirow{3}*{The Normal approximation method}
& 50 & 0.55 & 0.54 & 0.54\\
& 100 & 0.63 & 0.62 & 0.56\\
& 500& 0.60 & 0.66 & 0.62\\
\hline
\end{array}
$$
\end{center}
\caption{Coverage probabilities of the asymptotic confidence bands based on our method and the Normal approximation method.}
\label{tab1}
\end{table}
We can observe that the coverage probability is increasing with $n$ and is satisfactory even for enough small sample sizes, as $n=50$, and is close to 1, for sample sizes reaching 500.\\
We now compare our proposal to the $100(1-\alpha)\%$ confidence bands based on the asymptotic normality of the estimator. Indeed, Omelka et \textit{al.} \cite{r6} (2009) have established, under some regularity assumptions which are fulfilled here by taking $\phi$ equal to the Probit transformation, the weak convergence of the normalized process $\sqrt{n}[\hat{C}_{n,\hat{h}_n}^{(T)}(\cdot,\cdot) - C(\cdot,\cdot)]$ to a Gaussian limit process with explicit covariance function. This implies that, for any fixed $0\leq u,v\leq 1$, $\sqrt{n}[\hat{C}_{n,\hat{h}_n}^{(T)}(u,v) - C(u,v)]$ converges in distribution to Gaussian random variable with asymptotic variance,
\begin{eqnarray*}
\sigma^2(u,v)= C(u,v)[1-C(u,v)-2\{(1-u)C_u(u,v)-(1-v)C_v(u,v)+C_u(u,v)C_v(u,v)\}]&\\
+\; u(1-u)C^2_u(u,v)+v(1-v)C^2_v(u,v)-2uv C_u(u,v)C_v(u,v),&
\end{eqnarray*}
where $C_u(u,v)$ and $C_v(u,v)$ are the first-order partial derivatives of the copula $C(u,v)$.
Taking $C=C_{\theta}$ representing the Frank copula, we get the explicit expression of $\sigma^2(u,v)$ :
$$ \sigma^2(u,v)=\alpha(u,v)+\beta(u,v),$$
where
\begin{eqnarray*}
\alpha(u,v)& = & C_{\theta}(u,v)\left[ 1-C_{\theta}(u,v)-\frac{2(1-u)e^{-\theta u}(e^{-\theta v}-1)+2(1-v)e^{-\theta v}(e^{-\theta u}-1)}{e^{-\theta(u+ v)}-e^{-\theta u}-e^{-\theta v}+e^{-\theta}} \right]\\
& & + C_{\theta}(u,v)\left[\frac{ 2e^{-\theta(u+ v)}(e^{-\theta u}-1)(e^{-\theta v}-1)}{(e^{-\theta(u+ v)}-e^{-\theta u}-e^{-\theta v}+e^{-\theta})^2} \right]
\end{eqnarray*}
and
\begin{eqnarray*}
\beta(u,v) &= &\frac{u(1-u)e^{-2\theta u}(e^{-\theta v}-1)^2+v(1-v)e^{-2\theta v}(e^{-\theta u}-1)^2}{(e^{-\theta(u+ v)}-e^{-\theta u}-e^{-\theta v}+e^{-\theta})^2}\\
& & + \frac{2uv e^{-\theta(u+ v)}(e^{-\theta u}-1)(e^{-\theta v}-1)}{(e^{-\theta(u+ v)}-e^{-\theta u}-e^{-\theta v}+e^{-\theta})^2}.
\end{eqnarray*}
To compare our method to the normal asymptotic approximation method, we gives the coverage probabilities of the $99\%$ confidence bands for the true Frank copula $C_{\theta}$ in the last three lines in Table \ref{tab1}. From this table, it is clear that our proposed asymptotic confidence bands are more accurate than those obtained from the normal asymptotic approximation.\\
\begin{center}
The authors would like to thank the World Bank center for excellence - CEA-MITIC - based at the Universit\'e Gaston Berger for financial support of this work.
\end{center}
\addcontentsline{toc}{section}{References}
|
\section{Conclusions}
We propose Spatially Regularized Discriminative Correlation Filters (SRDCF) to address the limitations of the standard DCF. The introduced spatial regularization component enables the correlation filter to be learned on larger image regions, leading to a more discriminative appearance model.
By exploiting the sparsity of the regularization operation in the Fourier domain, we derive an efficient optimization strategy for learning the filter. The proposed learning procedure employs the Gauss-Seidel method to solve for the filter in the Fourier domain. We perform comprehensive experiments on four benchmark datasets. Our SRDCF outperforms existing trackers on all four datasets.
\noindent\textbf{Acknowledgments}:
This work has been supported by SSF (CUAS) and VR (VIDI, EMC${}^2$, ELLIIT, and CADICS).
\section{Experiments}
Here, we present a comprehensive evaluation of the proposed method. Result are reported on four benchmark datasets: OTB-2013, OTB-2015, ALOV++ and VOT2014.
\subsection{Details and Parameters}
\label{sec:details}
The weight function $w$ is constructed by starting from a quadratic function $w(m,n) = \mu + \eta (m/P)^2 + \eta (n/Q)^2$ with the minimum located at the sample center. Here $P \times Q$ denotes the target size, while $\mu$ and $\eta$ are parameters. The minimum value of w is set to $\mu = 0.1$ and the impact of the regularizer is set to $\eta = 3$. In practice, only a few DFT coefficients in the resulting function have a significant magnitude. We simply remove all DFT coefficients smaller than a threshold to ensure a sparse spectrum $\hat{w}$, containing about 10 non-zero coefficients. Figure~\ref{fig:periodic} visualizes the resulting weight function $w$ used in the optimization.
Similar to recent DCF based trackers \cite{DanelljanBMVC14,HenriquesPAMI15,Li2014}, we also employ HOG features, using a cell size of $4 \times 4$ pixels. Samples are represented by a square $M \times N$ grid of cells (\ie $M = N$), such that the corresponding image area is proportional to the area of the target bounding box. We set the image region area of the samples to $4^2$ times the target area and set the initial scale to ensure a maximum sample size of $M=50$ cells. Samples are multiplied by a Hann window \cite{MOSSE2010}.
We set the label function $y_t$ to a sampled Gaussian with a standard deviation proportional to the target size \cite{DanelljanBMVC14,Henriques12d}. The learning rate is set to $\gamma = 0.025$ and we use $N_\text{GS} = 4$ Gauss-Seidel iterations. All parameters remain fixed for all videos and datasets. Our Matlab implementation\footnote{Available at \url{http://www.cvl.isy.liu.se/research/objrec/visualtracking/regvistrack/index.html}.} runs at 5 frames per second on a standard desktop computer.
\subsection{Baseline Comparison}
Here, we evaluate the impact of the proposed spatial regularization component and compare it with the standard DCF formulation. First, we investigate the consequence of simply replacing the proposed regularizer with the standard DCF regularization in \eqref{eq:mdcf_cost}, without altering any parameters. This corresponds to using uniform regularization weights $w(m,n) = \sqrt{\lambda}$, in our framework. We set $\lambda = 0.01$ following \cite{DanelljanBMVC14,DanelljanCVPR14,Henriques12d}. For a fair comparison, we also evaluate both our and the standard regularization using a smaller sample size relative to the target, by setting the size as in \cite{DanelljanBMVC14,DanelljanCVPR14,Henriques12d}.
Table~\ref{tab:baseline_reg} shows the mean overlap precision (OP) for the four methods on the OTB-2013 dataset. The OP is computed as the fraction of frames in the sequence where the intersection-over-union overlap with the ground truth exceeds a threshold of $0.5$ (PASCAL criterion).
The standard DCF benefits from using smaller samples to avoid corrupting the positive training samples with background information. On the other hand, the proposed spatial regularization enables an expansion of the image region used for training the filter, without corrupting the target model. This leads to a more discriminative model, resulting in a gain of $7.0 \%$ in mean OP compared to the standard DCF formulation.
Additionally, we compare our method with Correlation Filters with Limited Boundaries (CFLB) \cite{GaloogahiCVPR2015}. For a fair comparison, we use the same settings as in \cite{GaloogahiCVPR2015} for our approach: single grayscale channel, no scale estimation, no sub-grid detection and the same sample size. On the OTB-2013, the CFLB achieves a mean OP of $48.6 \%$. Whereas the mentioned baseline version of our tracker obtains a mean OP of $54.3 \%$, outperforming \cite{GaloogahiCVPR2015} by $5.7 \%$.
\begin{table}[!t]
\centering
\resizebox{0.48\textwidth}{!}{\input{tables/baseline_regularization}}\vspace{1mm}
\caption{
A comparison of tracking performance on OTB-2013 when using the standard regularization \eqref{eq:mdcf_cost} and the proposed spatial regularization \eqref{eq:ours_cost_spatial}, in our tracking framework. The comparison is performed both with a conventional sample size (used in existing DCF based trackers) and our expanded sample size.
}
\label{tab:baseline_reg}
\end{table}
\vspace{1mm}
\subsection{OTB-2013 Dataset}
\begin{table}
\centering
\resizebox{0.48\textwidth}{!}{\input{tables/OTB_comparison}}\vspace{1mm}
\caption{A comparison with state-of-the-art trackers on the OTB-2013 and OTB-2015 datasets using mean overlap precision (in percent). The best two results for each dataset are shown in red and blue fonts respectively. Our SRDCF achieves a gain of $8.0 \%$ and $8.2 \%$ on OTB-2013 and OTB-2015 respectively compared to the second best tracker on each dataset.
}\vspace{-2mm}
\label{tab:OTB_comparison}
\end{table}
We provide a comparison of our tracker with 24 state-of-the-art methods from the literature: MIL \cite{Babenko09b}, IVT \cite{Ross08d}, CT \cite{Zhang12c}, TLD \cite{Mikolajczyk10d}, DFT \cite{Laura12d}, EDFT \cite{Felsberg13c}, ASLA \cite{Jia12d}, L1APG \cite{Bao12d}, CSK \cite{Henriques12d}, SCM \cite{Zhong12g}, LOT \cite{Oron12b}, CPF \cite{Perez02b}, CXT \cite{Dinh11}, Frag \cite{Adam6c}, Struck \cite{Torr11b}, LSHT \cite{Shengfeng13b}, LSST \cite{Wang13d}, ACT \cite{DanelljanCVPR14}, KCF \cite{HenriquesPAMI15}, CFLB \cite{GaloogahiCVPR2015}, DSST \cite{DanelljanBMVC14}, SAMF \cite{Li2014}, TGPR \cite{TGPR2014} and MEEM \cite{MEEM2014}.
\subsubsection{State-of-the-art Comparison}
Table~\ref{tab:OTB_comparison} shows a comparison with state-of-the-art methods on the OTB-2013 dataset, using mean overlap precision (OP) over all 50 videos. Only the results for the top 10 trackers are reported.
The MEEM tracker, based on an online SVM, provides the second best results with a mean OP of $70.1\%$. The best result on this dataset is obtained by our tracker with a mean OP of $78.1\%$, leading to a significant gain of $8.0 \%$ compared to MEEM.
Figure~\ref{fig:sota_ope} shows the success plot over all the 50 videos in OTB-2013. The success plot shows the mean overlap precision (OP), plotted over the range of intersection-over-union thresholds. The trackers are ranked using the \emph{area under the curve} (AUC), displayed in the legend.
Among previous DCF based trackers, DSST and SAMF provides the best performance, with an AUC score of $56.0\%$ and $57.7\%$. Our approach obtains an AUC score of $63.3\%$ and significantly outperforms the best existing tracker (SAMF) by $5.6\%$.
\begin{figure}[!t]
\centering\vspace{-3.5mm}
\newcommand{0.49\textwidth}{0.24\textwidth}
\subfloat[OTB-2013\label{fig:sota_ope}]{\includegraphics[width = 0.49\textwidth]{figures/sota_OPE/quality_plot_overlap_OPE_AUC}}
\subfloat[OTB-2015\label{fig:sota_ope_100}]{\includegraphics[width = 0.49\textwidth]{figures/sota_OPE_100/quality_plot_overlap_OPE_AUC}}
\caption{Success plots showing a comparison with state-of-the-art methods on OTB-2013 (a) and OTB-2015 (b). For clarity, only the top 10 trackers are displayed. Our SRDCF achieves a gain of $5.6 \%$ and $5.7 \%$ on OTB-2013 and OTB-2015 respectively, compared to the second best methods.}\vspace{-2mm}
\label{fig:OTB_ope}
\end{figure}
\begin{figure}[!t]
\centering
\newcommand{0.49\textwidth}{0.24\textwidth}
\includegraphics[width = 0.49\textwidth]{figures/sota_SRE/quality_plot_overlap_SRE_AUC}%
\includegraphics[width = 0.49\textwidth]{figures/sota_TRE/quality_plot_overlap_TRE_AUC}\vspace{-1mm}
\caption{Comparison with respect to robustness to initialization on OTB-2013. We show success plots for both the spatial (SRE) and temporal (TRE) robustness. Our approach clearly demonstrates robustness in both scenarios.}\vspace{-2mm}
\label{fig:sota_sretre}
\end{figure}
\subsubsection{Robustness to Initialization}
Visual tracking methods are known to be sensitive to initialization. We evaluate the robustness of our tracker by following the protocol proposed in \cite{Wu13g}. Two different types of initialization criteria, namely: temporal robustness (TRE) and spatial robustness (SRE), are evaluated. The SRE corresponds to tracker initialization at different positions close to the ground-truth in the first frame. The procedure is repeated with 12 different initializations for each video in the dataset. The TRE criteria evaluates the tracker by initializations at 20 different frames, with the ground-truth.
\begin{figure}[!t]
\centering
\newcommand{0.49\textwidth}{0.12\textwidth}
\includegraphics*[trim = 130 50 70 0,width = 0.49\textwidth]{figures/quality/Soccer0003}%
\includegraphics*[trim = 130 50 70 0,width = 0.49\textwidth]{figures/quality/Soccer0059}%
\includegraphics*[trim = 70 50 130 0,width = 0.49\textwidth]{figures/quality/Soccer0165}%
\includegraphics*[trim = 70 50 130 0,width = 0.49\textwidth]{figures/quality/Soccer0261}
\includegraphics*[trim = 50 50 5 180,width = 0.49\textwidth]{figures/quality/Human60113}%
\includegraphics*[trim = 50 0 5 230,width = 0.49\textwidth]{figures/quality/Human60345}%
\includegraphics*[trim = 50 0 5 230,width = 0.49\textwidth]{figures/quality/Human60360}%
\includegraphics*[trim = 50 0 5 230,width = 0.49\textwidth]{figures/quality/Human60370}
\includegraphics*[trim = 50 180 230 30,width = 0.49\textwidth]{figures/quality/Tiger20012}%
\includegraphics*[trim = 100 180 180 30,width = 0.49\textwidth]{figures/quality/Tiger20213}%
\includegraphics*[trim = 100 180 180 30,width = 0.49\textwidth]{figures/quality/Tiger20322}%
\includegraphics*[trim = 50 180 230 30,width = 0.49\textwidth]{figures/quality/Tiger20358}
\includegraphics*[trim = 2 2 2 5,width = 0.40\textwidth]{figures/quality/qualitative_legend}\vspace{-1mm}
\caption{Qualitative comparison of our approach with state-of-the-art trackers on the \emph{Soccer}, \emph{Human6} and \emph{Tiger2} videos. Our approach provides consistent results in challenging scenarios, such as occlusions, fast motion, background clutter and target rotations.}\vspace{-2mm}
\label{fig:qualitative}
\end{figure}
Figure~\ref{fig:sota_sretre} shows the success plots for TRE and SRE on the OTB-2013 dataset with 50 videos. We include all the top 7 trackers in figure~\ref{fig:sota_ope} for this experiment. Among the existing methods, SAMF and MEEM provide the best results. Our SRDCF achieves a consistent gain in performance over these trackers on both robustness evaluations.
\begin{figure*}[!t]
\centering
\newcommand{0.49\textwidth}{0.25\textwidth}
\newcommand{figures/sota_OPE}{figures/sota_OPE}
\newcommand{OPE}{OPE}
\includegraphics[width=0.49\textwidth]{figures/sota_OPE /out-of-plane_rotation_overlap_OPE _AUC.pdf}%
\includegraphics[width=0.49\textwidth]{figures/sota_OPE /scale_variations_overlap_OPE _AUC.pdf}%
\includegraphics[width=0.49\textwidth]{figures/sota_OPE /blur_overlap_OPE _AUC.pdf}%
\includegraphics[width=0.49\textwidth]{figures/sota_OPE /occlusions_overlap_OPE _AUC.pdf}\vspace{-0.5mm}
\caption{Attribute-based analysis of our approach on the OTB-2013 dataset with 50 videos. Success plots are shown for four attributes. Each plot title includes the number of videos associated with the respective attribute. Only the top 10 trackers for each attribute are displayed for clarity. Our approach demonstrates superior performance compared to existing trackers in these scenarios.}\vspace{-3mm}
\label{fig:attribute}
\end{figure*}
\subsubsection{Attribute Based Comparison}
We perform an attribute based analysis of our approach on the OTB-2013 dataset. All the 50 videos in OTB-2013 are annotated with 11 different attributes, namely: illumination variation, scale variation, occlusion, deformation, motion blur, fast motion, in-plane rotation, out-of-plane rotation, out-of-view, background clutter and low resolution. Our approach outperforms existing trackers on 10 attributes.
Figure~\ref{fig:attribute} shows example success plots of four different attributes. Only the top 10 trackers in each plot are displayed for clarity. In case of out-of-plane rotations, (MEEM) achieves an AUC score of $57.2\%$. Our tracker provides a gain of $3.3\%$ compared to MEEM. Among the existing methods, the two DCF based trackers DSST and SAMF provide the best results in case of scale variation. Both these trackers are designed to handle scale variations. Our approach achieves a significant gain of $4.1\%$ over DSST. Note that the standard DCF trackers struggle in the cases of motion blur and fast motion due to the restricted search area. This is caused by the induced boundary effects in the detection samples of the standard DCF trackers. Our approach significantly improves the performance compared to the standard DCF based trackers in these cases.
Figure~\ref{fig:qualitative} shows a qualitative comparison of our approach with existing methods on challenging example videos.
Despite no explicit occlusion handling component, our tracker performs favorably in cases with occlusion.
\subsection{OTB-2015 Dataset}
We provide a comparison of our approach on the recently introduced OTB-2015. The dataset extends OTB-2013 and contains 100 videos. Table~\ref{tab:OTB_comparison} shows the comparison with the top 10 methods, using mean overlap precision (OP) over all 100 videos. Among the existing methods, SAMF and MEEM provide the best results with mean OP of $64.7\%$ and $63.4\%$ respectively. Our tracker outperforms the best existing tracker by $8.2\%$ in mean OP.
Figure~\ref{fig:sota_ope_100} shows the success plot over all the 100 videos. Among the standard DCF trackers, SAMF provides the best results with an AUC score of $54.8\%$. The MEEM tracker achieves an AUC score of $53.8\%$. Our tracker obtains an AUC score of $60.5\%$, outperforming SAMF by $5.7\%$.
\begin{figure}[!t]
\centering
\includegraphics[width = 0.47\textwidth]{figures/alov300_results_test.pdf}\vspace{0mm}
\caption{Survival curves comparing our approach with 24 trackers on ALOV++. The mean F-scores for the top 10 trackers are shown in the legend. Our approach achieves the best overall results.}\vspace{-3mm}
\label{fig:alov}
\end{figure}
\subsection{ALOV++ Dataset}
We also perform experiments on the ALOV++ dataset \cite{ALOV}, containing 314 videos with 89364 frames in total. The evaluation protocol employs survival curves based on F-score, where a higher F-score indicates better performance.
The survival curve is constructed by plotting the sorted F-scores of all 314 videos.
We refer to \cite{ALOV} for details.
Our approach is compared with the 19 trackers evaluated in \cite{ALOV}. We also add the top 5 methods from our OTB comparison. Figure~\ref{fig:alov} shows the survival curves and the average F-scores of the trackers. MEEM obtains a mean F-score of $0.708$. Our approach obtains the best overall performance compared to 24 trackers with a mean F-score of $0.787$.
\subsection{VOT2014 Dataset}
Finally, we present results on VOT2014 \cite{VOT2014}. Our approach is compared with the 38 participating trackers in the challenge. We also add MEEM in the comparison. In VOT2014, the trackers are evaluated both in terms of accuracy and robustness. The accuracy score is based on the overlap with ground truth, while the robustness is determined by the failure rate.
The trackers are restarted at each failure. The final rank is based on the accuracy and robustness in each video. We refer to \cite{VOT2014} for details.
Table~\ref{tab:VOT_comparison} shows the final ranking scores over all the videos in VOT2014. Among the existing methods, the DSST approach provides the best results. Our tracker achieves the top final rank of $8.26$, outperforming DSST and SAMF.
\begin{table}[!t]
\centering
\resizebox{0.47\textwidth}{!}{\input{tables/VOT_comparison}}\vspace{1mm}
\caption{Results for the top 3 trackers on VOT2014. The mean overlap and failure rate is reported in the first two columns. The accuracy rank, robustness rank and the combined final rank are shown in the remaining columns. Our tracker obtains the best performance on this dataset.}\vspace{-3mm}
\label{tab:VOT_comparison}
\end{table}
\section{Introduction}
Visual tracking is a classical computer vision problem with many applications. In generic tracking the task is to estimate the trajectory of a target in an image sequence, given only its initial location. This problem is especially challenging. The tracker must generalize the target appearance from a very limited set of training samples to achieve robustness against, \eg occlusions, fast motion and deformations. Here, we investigate the key problem of learning a robust appearance model under these conditions.
\begin{figure}[!t]
\newcommand{0.49\textwidth}{0.23\textwidth}
\newcommand{30}{30}
\centering\vspace{-1mm}
\subfloat[Original image.]{\includegraphics*[trim=0 0 0 0, width=0.49\textwidth]{figures/soccer_nonperiodic.pdf}}\hspace{0.1mm}
\subfloat[Periodicity in correlation filters.]{\includegraphics*[trim= 0 0 0 0, width=0.49\textwidth]{figures/soccer_periodic.pdf}}\vspace{1mm}
\caption{Example image (a) and the underlying periodic assumption (b) employed in the standard DCF methods. The periodic assumption (b) leads to a limited set of negative training samples, that fails to capture the true image content (a). As a consequence, an inaccurate tracking model is learned.}
\vspace{-4mm}
\label{fig:periodic}
\end{figure}
Recently, Discriminative Correlation Filter (DCF) based approaches \cite{MOSSE2010,DanelljanBMVC14,DanelljanCVPR14,Henriques12d,HenriquesPAMI15,Li2014} have successfully been applied to the tracking problem \cite{VOT2014}. These methods learn a correlation filter from a set of training samples. The correlation filter is trained to perform a circular sliding window operation on the training samples. This corresponds to assuming a periodic extension of these samples (see figure~\ref{fig:periodic}). The periodic assumption enables efficient training and detection by utilizing the Fast Fourier Transform (FFT).
As discussed above, the computational efficiency of the standard DCF originates from the periodic assumption at both training and detection. However, this underlying assumption produces unwanted boundary effects. This leads to an inaccurate representation of the image content, since the training patches contain periodic repetitions. The induced boundary effects mainly limit the standard DCF formulation in two important aspects. Firstly, inaccurate negative training patches reduce the discriminative power of the learned model. Secondly, the detection scores are only accurate near the center of the region, while the remaining scores are heavily influenced by the periodic repetitions of the detection sample. This leads to a very restricted target search region at the detection step.
The aforementioned limitations of the standard DCF formulation hamper the tracking performance in several ways. (a) The DCF based trackers struggle in cases with fast target motion due to the restricted search region. (b) The lack of negative training patches leads to over-fitting of the learned model, significantly affecting the performance in cases with \eg target deformations. (c) The mentioned limitations in training and detection also reduce the potential of the tracker to re-detect the target after an occlusion. (d) A naive expansion of the image area used for training the correlation filter corresponds to using a larger periodicity (see figure~\ref{fig:periodic}). Such an expansion results in an inclusion of substantial amount of background information within the positive training samples. These corrupted training samples severely degrade the discriminative power of the model, leading to inferior tracking results. In this work, we tackle these inherent problems by re-visiting the standard DCF formulation.
\subsection{Contributions}
In this paper, we propose Spatially Regularized Discriminative Correlation Filters (SRDCF) for tracking. We introduce a spatial regularization component within the DCF formulation, to address the problems induced by the periodic assumption. The proposed regularization weights penalize the correlation filter coefficients during learning. The spatial weights are based on the a priori information about the spatial extent of the filter. Due to the spatial regularization, the correlation filter can be learned on larger image regions. This enables a larger set of negative patches to be included in the training, leading to a more discriminative model.
Due to the online nature of the tracking problem, a computationally efficient learning scheme is crucial. Therefore, we introduce a suitable optimization strategy for the proposed SRDCF. The online capability is achieved by exploiting the sparsity of the spatial regularization function in the Fourier domain. We propose to apply the iterative Gauss-Seidel method to solve the resulting normal equations.
Additionally, we introduce a strategy to maximize the detection scores with sub-grid precision.
We perform comprehensive experiments on four benchmark datasets: OTB-2013 \cite{Wu13g} with 50 videos, ALOV++ \cite{ALOV} with 314 videos, VOT2014 \cite{VOT2014} with 25 videos and OTB-2015 \cite{OTB2015} with 100 videos. Compared to the best existing trackers, our approach obtains an absolute gain of $8.0 \%$ and $8.2 \%$ on OTB-2013 and OTB-2015 respectively, in mean overlap precision. Our method also achieves the best overall results on ALOV++ and VOT2014. Additionally, our tracker won the OpenCV State of the Art Vision Challenge in tracking \cite{OpenCVchallenge} (there termed DCFSIR).
\section{Discriminative Correlation Filters}
Discriminative correlation filters (DCF) is a supervised technique for learning a linear classifier or a linear regressor. The main difference from other techniques, such as support vector machines \cite{VapnikSVM}, is that the DCF formulation exploits the properties of circular correlation for efficient training and detection. In recent years, the DCF based approaches have been successfully applied for tracking. Bolme \etal \cite{MOSSE2010} first introduced the MOSSE tracker, using only grayscale samples to train the filter.
Recent work \cite{DanelljanSCIA2015,DanelljanBMVC14,DanelljanCVPR14,HenriquesPAMI15,Li2014} have shown a notable improvement by learning multi-channel filters on multi-dimensional features, such as HOG \cite{Dalal05} or Color-Names \cite{Weijer09a}. However, to become computationally viable, these approaches rely on harsh approximations of the standard DCF formulation, leading to sub-optimal learning. Other work have investigated offline learning of multi-channel DCF:s for object detection \cite{galoogahiICCV13,henriquesICCV13} and recognition \cite{BoddetiCVPR13}. But these methods are too computationally costly for online tracking applications.
The circular correlation within the DCF formulation has two major advantages. Firstly, the DCF is able to make extensive use of limited training data by implicitly including all shifted versions of the given samples. Secondly, the computational effort for training and detection is significantly reduced by performing the necessary computations in the Fourier domain and using the Fast Fourier Transform (FFT). These two advantages make DCF:s especially suitable for tracking, where training data is scarce and computational efficiency is crucial for real-time applications.
By employing a circular correlation, the standard DCF formulation relies on a periodic assumption of the training and detection samples. However, this assumption produces unwanted boundary effects, leading to an inaccurate description of the image. These inaccurate training patches severely hamper the learning of a discriminative tracking model. Surprisingly, this problem has been largely ignored by the tracking community. Galoogahi \etal \cite{GaloogahiCVPR2015} investigate the boundary effect problem for single-channel DCF:s. Their approach solve a constrained optimization problem, using the Alternating Direction Method of Multipliers (ADMM), to ensure a correct filter size. This however requires a transition between the spatial and Fourier domain in each ADMM iteration, leading to an increased computational complexity. Different to \cite{GaloogahiCVPR2015}, we propose a spatial regularization component in the objective. By exploiting the sparsity of our regularizer, we efficiently optimize the filter directly in the Fourier domain. Contrary to \cite{GaloogahiCVPR2015}, we target the problem of multi-dimensional features, such as HOG, crucial for the overall tracking performance \cite{DanelljanCVPR14,HenriquesPAMI15}.
\subsection{Standard DCF Training and Detection}
In the DCF formulation, the aim is to learn a multi-channel convolution\footnote{We use convolution for mathematical convenience, though correlation can equivalently be used.} filter $f$ from a set of training examples $\{(x_k, y_k)\}_{k=1}^t$. Each training sample $x_k$ consists of a $d$-dimensional feature map extracted from an image region. All samples are assumed to have the same spatial size $M \times N$. At each spatial location $(m,n) \in \Omega \defeq \{0,\ldots,M-1\} \times \{0,\ldots,N-1\}$ we thus have a $d$-dimensional feature vector $x_k(m,n) \in \reals^d$. We denote feature layer $l \in \{1,\ldots,d\}$ of $x_k$ by $x_k^l$. The desired output $y_k$ is a scalar valued function over the domain $\Omega$, which includes a label for each location in the sample $x_k$.
The desired filter $f$ consists of one $M \times N$ convolution filter $f^l$ per feature layer. The convolution response of the filter $f$ on a $M \times N$ sample $x$ is given by
\begin{equation}
\label{eq:mdcf_conv}
S_f(x) = \sum_{l=1}^d x^l \conv f^l .
\end{equation}
Here, $\conv$ denotes circular convolution. The filter is obtained by minimizing the $L^2$-error between the responses $S_f(x_k)$ on the training samples $x_k$, and the labels $y_k$,
\begin{equation}
\label{eq:mdcf_cost}
\varepsilon_t(f) = \sum_{k=1}^t \alpha_k \big\| S_f(x_k) - y_k \big\|^2 + \lambda \sum_{l=1}^d \big\| f^l \big\|^2 .
\end{equation}
Here, the weights $\alpha_k \geq 0$ determine the impact of each training sample and $\lambda \geq 0$ is the weight of the regularization term. Eq.~\ref{eq:mdcf_cost} is a linear least squares problem. Using Parseval's formula, it can be transformed to the Fourier domain, where the resulting normal equations have a block diagonal structure. The Discrete Fourier Transformed (DFT) filters $\hat{f}^l = \ft\{f^l\}$ can then be obtained by solving $MN$ number of $d \times d$ linear equation systems \cite{galoogahiICCV13}.
For efficiency reasons, the learned DCF is typically applied in a sliding-window-like manner by evaluating the classification scores on all cyclic shifts of a test sample. Let $z$ denote the $M \times N$ feature map extracted from an image region. The classification scores $S_f(z)$ at all locations in this image region can be computed using the convolution property of the DFT,
\begin{equation}
\label{eq:mdcf_detection}
S_f(z) = \ftinv \Bigg\{ \sum_{l=1}^d \hat{z}^l \pmult \hat{f}^l \Bigg\} .
\end{equation}
Here, $\pmult$ denotes point-wise multiplication, the hat denotes the DFT of a function and $\ftinv$ denotes the inverse DFT. The FFT hence allows the detection scores to be computed in $\ordo(dMN \log MN)$ complexity instead of $\ordo(dM^2N^2)$.
Note that the operation $S_f(x)$ in \eqref{eq:mdcf_conv} corresponds to applying the linear classifier $f$, in a sliding window fashion, to the periodic extension of the sample $x$ (see figure~\ref{fig:periodic}). This introduces unwanted periodic boundary effects in the training \eqref{eq:mdcf_cost} and detection \eqref{eq:mdcf_detection} steps.
\section{Spatially Regularized Correlation Filters}
\label{sec:method}
We propose to use a spatial regularization component in the standard DCF formulation. The resulting optimization problem is solved in the Fourier domain, by exploiting the sparse nature of the proposed regularization.
\begin{figure}[!t]
\newcommand{0.49\textwidth}{0.46\textwidth}
\centering
\includegraphics[width=0.49\textwidth,height=5cm]{figures/weights_plot.png}\vspace{-0.5mm}
\caption{Visualization of the spatial regularization weights $w$ employed in the learning of our SRDCF, and the corresponding image region used for training. Filter coefficients residing in the background region are penalized by assigning higher weights in $w$. This significantly mitigates the emphasis on background information in the learned classifier.}
\vspace{-1mm}
\label{fig:weights}
\end{figure}
\subsection{Spatial Regularization}
To alleviate the problems induced by the circular convolution in \eqref{eq:mdcf_conv}, we replace the regularization term in \eqref{eq:mdcf_cost} with a more general Tikhonov regularization. We introduce a spatial weight function $w: \Omega \rightarrow \reals$ used to penalize the magnitude of the filter coefficients in the learning.
The regularization weights $w$ determine the importance of the filter coefficients $f^l$, depending on their spatial locations. Coefficients in $f^l$ residing outside the target region are suppressed by assigning higher weights in $w$ and vice versa.
The resulting optimization problem is expressed as,
\begin{equation}
\label{eq:ours_cost_spatial}
\varepsilon(f) = \sum_{k=1}^t \alpha_k \big\| S_f(x_k) - y_k \big\|^2 + \sum_{l=1}^d \big\| w \pmult f^l \big\|^2 .
\end{equation}
The regularization weights $w$ in \eqref{eq:ours_cost_spatial} are visualized in figure~\ref{fig:weights}. Visual features close to the target edge are often less reliable than those close to the target center, due to \eg target rotations and occlusions. We therefore let the regularization weights change smoothly from the target region to the background. This also increases the sparsity of $w$ in the Fourier domain.
Note that \eqref{eq:ours_cost_spatial} simplifies to the standard DCF \eqref{eq:mdcf_cost} for uniform weights $w(m,n) = \sqrt{\lambda}$.
\begin{figure*}[!t]
\newcommand{0.49\textwidth}{0.49\textwidth}
\centering
\subfloat[Standard DCF.]{\includegraphics[width=0.49\textwidth]{figures/filter_standard.png}}\hspace{1mm}
\subfloat[Our SRDCF.]{\includegraphics[width=0.49\textwidth]{figures/filter_ours.png}}\vspace{0.0mm}
\caption{Visualization of the filter coefficients learned using the standard DCF (a) and our approach (b). The surface plots show the filter values $f^l$ and the corresponding image region used for training. In the standard DCF, high values are assigned to the background region. The larger influence of background information at the detection stage deteriorates tracking performance. In our approach, the regularization weights penalizes filter values corresponding to features in the background. This increases the discriminative power of the learned model, by emphasizing the appearance information within the target region (green box).}
\vspace{-4mm}
\label{fig:filters}
\end{figure*}
By applying Parseval's theorem to \eqref{eq:ours_cost_spatial}, the filter $f$ can equivalently be obtained by minimizing the resulting loss function \eqref{eq:ours_cost_dft} over the DFT coefficients $\hat{f}$,
\begin{equation}
\label{eq:ours_cost_dft}
\check{\varepsilon}(\hat{f}) = \sum_{k=1}^t \alpha_k \Bigg\| \sum_{l=1}^d \hat{x}_k^l \pmult \hat{f}^l - \hat{y}_k \Bigg\|^2 + \sum_{l=1}^d \Bigg\| \frac{\hat{w}}{MN} \conv \hat{f}^l \Bigg\|^2 .
\end{equation}
The second term in \eqref{eq:ours_cost_dft} follows from the convolution property of the inverse DFT.
A vectorization of \eqref{eq:ours_cost_dft} gives,
\begin{equation}
\label{eq:ours_cost_vec1}
\check{\varepsilon}(\hat{f}) = \sum_{k=1}^t \alpha_k \Bigg\| \sum_{l=1}^d \diagm(\vecn{\hat{x}}_k^l) \vecn{\hat{f}}^l - \vecn{\hat{y}}_k \Bigg\|^2 \hspace{-1.0mm} + \sum_{l=1}^d \Bigg\| \frac{\circm(\vecn{\hat{w}})}{MN} \vecn{\hat{f}}^l \Bigg\|^2 \hspace{-1.0mm}.
\end{equation}
Here, bold letters denote a vectorization of the corresponding scalar valued functions and $\diagm(\vecn{v})$ denotes the diagonal matrix with the elements of the vector $\vecn{v}$ in its diagonal. The $MN \times MN$ matrix $\circm(\vecn{\hat{w}})$ represents circular 2D-convolution with the function $\hat{w}$, \ie $\circm(\vecn{\hat{w}}) \vecn{\hat{f}}^l = \vecnf(\hat{w} \conv \hat{f}^l)$. Each row in $\circm(\vecn{\hat{w}})$ thus contains a cyclic permutation of $\vecn{\hat{w}}$.
The DFT of a real-valued function is known to be Hermitian symmetric. Therefore, minimizing \eqref{eq:ours_cost_spatial} over the set of real-valued filters $f^l$, corresponds to minimizing \eqref{eq:ours_cost_dft} over the set of Hermitian symmetric DFT coefficients $\hat{f}^l$. We reformulate \eqref{eq:ours_cost_vec1} to an equivalent real-valued optimization problem, to ensure faster convergence by preserving the Hermitian symmetry. Let $\rho : \Omega \rightarrow \Omega$ be the point-reflection $\rho(m,n) = (-m \mod M,-n \mod N)$. The domain $\Omega$ can be partitioned into $\Omega_0$, $\Omega_+$ and $\Omega_-$, where $\Omega_0 = \rho(\Omega_0)$ and $\Omega_- = \rho(\Omega_+)$. Thus, $\Omega_0$ denote the part of the spectrum with no corresponding reflected frequency, and $\Omega_-$ contains the reflected frequencies in $\Omega_+$. We define,
\begin{equation}
\label{eq:real_trans}
\tilde{f}^l(m,n) =
\begin{cases}
\hat{f}^l(m,n) , & (m,n) \in \Omega_0 \\
\frac{\hat{f}^l(m,n) + \hat{f}^l(\rho(m,n))}{\sqrt{2}} , & (m,n) \in \Omega_+ \\
\frac{\hat{f}^l(m,n) - \hat{f}^l(\rho(m,n))}{i\sqrt{2}} , & (m,n) \in \Omega_-
\end{cases}
\end{equation}
such that $\tilde{f}^l$ is real-valued by the Hermitian symmetry of $\hat{f}^l$. Here, $i$ denotes the imaginary unit. Eq.~\ref{eq:real_trans} can be expressed by a unitary $MN \times MN$ matrix $B$ such that $\vectil{f}^l = B\vecft{f}^l$. By \eqref{eq:real_trans}, $B$ contains at most two non-zero entries in each row.
The reformulated variables from \eqref{eq:ours_cost_vec1} are defined as $\vectil{y}_k = B \vecft{y}_k$, $\redftmat{D}_k^l = B \diagm(\vecn{\hat{x}}_k^l) B\ctp$ and $\redftmat{C} = \frac{1}{MN} B \circm(\vecn{\hat{w}}) B\ctp$,
where $\ctp$ denotes the conjugate transpose of a matrix. Since $B$ is unitary, \eqref{eq:ours_cost_vec1} can equivalently be expressed as,
\begin{equation}
\label{eq:ours_cost_vec2}
\tilde{\varepsilon}(\vectil{f}^1\ldots\vectil{f}^d) = \sum_{k=1}^t \alpha_k \Bigg\| \sum_{l=1}^d \redftmat{D}_k^l \vectil{f}^l - \vectil{y}_k \Bigg\|^2 \hspace{-2mm} + \sum_{l=1}^d \left\| \redftmat{C} \vectil{f}^l \right\|^2 \hspace{-2mm}.
\end{equation}
All variables in \eqref{eq:ours_cost_vec2} are real-valued. The loss function \eqref{eq:ours_cost_vec2} is then simplified by defining the fully vectorized real-valued filter as the concatenation $\vectil{f} = \big((\vectil{f}^1)\tp \cdots (\vectil{f}^d)\tp \big)\tp$,
\begin{equation}
\label{eq:ours_cost_vec3}
\tilde{\varepsilon}(\vectil{f}) = \sum_{k=1}^t \alpha_k \Big\| \redftmat{D}_k \vectil{f} - \vectil{y}_k \Big\|^2 + \Big\| \redftmat{W} \vectil{f} \Big\|^2 .
\end{equation}
Here we have defined the concatenation $\redftmat{D}_k = (\redftmat{D}_k^1 \cdots \redftmat{D}_k^d)$ and $\redftmat{W}$ to be the $dMN \times dMN$ block diagonal matrix with each diagonal block being equal to $\redftmat{C}$. Finally, \eqref{eq:ours_cost_vec3} is minimized by solving the normal equations $\redftmat{A}_t \vectil{f} = \vectil{b}_t$, where%
\begin{subequations}
\label{eq:normal_equations_real}
\begin{align}
\label{eq:normal_A_real}
\redftmat{A}_t &= \sum_{k=1}^t \alpha_k \redftmat{D}_k\tp \redftmat{D}_k + \redftmat{W}\tp \redftmat{W} \\
\tilde{\mathbf{b}}_t &= \sum_{k=1}^t \alpha_k \redftmat{D}_k\tp \tilde{\vecn{y}}_k .
\end{align}
\end{subequations}
Here, \eqref{eq:normal_equations_real} defines a real $dMN \times dMN$ linear system of equations. The fraction of non-zero elements in $\redftmat{A}_t$ is smaller than $\frac{2d + K^2}{dMN}$, where $K$ is the number of non-zero Fourier coefficients in $\hat{w}$. Thus, $\redftmat{A}_t$ is sparse if $w$ has a sparse spectrum. The DFT coefficients for the filters are obtained by solving the system \eqref{eq:normal_equations_real} and applying $\vecft{f}^l = B\ctp \vectil{f}^l$.
Figure~\ref{fig:filters} visualizes the filter learned by optimizing the standard DCF loss \eqref{eq:mdcf_cost} and the proposed formulation \eqref{eq:ours_cost_spatial}, using the spatial regularization weights $w$ in figure~\ref{fig:weights}. In the standard DCF, large values are spatially distributed over the whole filter. By penalizing filter coefficients corresponding to background, our approach learns a classifier that emphasizes visual information within the target region.
A direct application of a sparse solver to the normal equations $\redftmat{A}_t \vectil{f} = \vectil{b}_t$ is computationally very demanding (even when the standard regularization $\redftmat{W}\tp \redftmat{W} = \lambda I$ is used and the number of features is small ($d > 2$)). Next, we propose an efficient optimization scheme to solve the normal equations for online learning scenarios, such as tracking.
\subsection{Optimization}
For the standard DCF formulation \eqref{eq:mdcf_cost} the normal equations have a block diagonal structure \cite{galoogahiICCV13}. However, this block structure is not attainable in our case due to the structure of the regularization matrix $\redftmat{W}\tp \redftmat{W}$ in \eqref{eq:normal_A_real}. We propose an iterative approach, based on the Gauss-Seidel, for efficient online computation of the filter coefficients.
The Gauss-Seidel method decomposes the matrix $\redftmat{A}_t$ into a lower triangular part $\redftmat{L}_t$ and a strictly upper triangular part $\redftmat{U}_t$ such that $\redftmat{A}_t = \redftmat{L}_t + \redftmat{U}_t$. The algorithm then proceeds by solving the following triangular system for $\vectil{f}^{(j)}$ in each iteration $j = 1, 2, \ldots$,
\begin{equation}
\label{eq:gauss_seidel}
\redftmat{L}_t \vectil{f}^{(j)} = \tilde{\mathbf{b}}_t - \redftmat{U}_t \vectil{f}^{(j-1)} .
\end{equation}
This lower triangular equation system is solved efficiently using forward substitution and by exploiting the sparsity of $\redftmat{L}_t$ and $\redftmat{U}_t$. The Gauss-Seidel recursion \eqref{eq:gauss_seidel} converges to the solution of $\redftmat{A}_t \vectil{f} = \tilde{\mathbf{b}}$ whenever the matrix $\redftmat{A}_t$ is symmetric and positive definite. The construction of the weights $w$ (see section~\ref{sec:details}) ensures that both conditions are satisfied.
\section{Our Tracking Framework}
Here, we describe our tracking framework, based on the Spatially Regularized Discriminative Correlation Filters (SRDCF) proposed in section~\ref{sec:method}.
\subsection{Training}
At the training stage, the model is updated by first extracting a new training sample $x_t$ centered at the target location. Here, $t$ denotes the current frame number. We then update $\redftmat{A}_t$ and $\tilde{\mathbf{b}}_t$ in \eqref{eq:normal_equations_real} with a learning rate $\gamma \geq 0$,%
\begin{subequations}
\label{eq:normal_equations_real_update}
\begin{align}
\label{eq:normal_A_real_update}
\redftmat{A}_t &= (1 - \gamma) \redftmat{A}_{t-1} + \gamma \left( \redftmat{D}_t\tp \redftmat{D}_t + \redftmat{W}\tp \redftmat{W} \right) \\
\tilde{\mathbf{b}}_t &= (1 - \gamma) \tilde{\mathbf{b}}_{t-1} + \gamma \redftmat{D}_t\tp \tilde{\vecn{y}}_t .
\end{align}
\end{subequations}
This corresponds to using exponentially decaying weights $\alpha_k$ in the loss function \eqref{eq:ours_cost_spatial}. In the first frame, we set $\redftmat{A}_1 = \redftmat{D}_1\tp \redftmat{D}_1 + \redftmat{W}\tp \redftmat{W}$ and $\tilde{\mathbf{b}}_1 = \redftmat{D}_1\tp \tilde{\vecn{y}}_1$. Note that the regularization matrix $\redftmat{W}\tp \redftmat{W}$ can be precomputed once for the entire sequence. The update strategy \eqref{eq:normal_equations_real_update} ensures memory efficiency, since it does not require storage of all samples $x_k$.
After the model update \eqref{eq:normal_equations_real_update}, we perform a fixed number $N_\text{GS}$ of Gauss-Seidel iterations \eqref{eq:gauss_seidel} per frame to compute the new filter coefficients.
For the initial iteration $\vectil{f}_t^{(0)}$ in frame $t$, we use the filter computed in the previous frame, \ie $\vectil{f}_t^{(0)} = \vectil{f}_{t-1}^{(N_\text{GS})}$. In the first frame, the initial estimate $\vectil{f}_1^{(0)}$ is obtained by solving the $MN \times MN$ linear system,
\begin{equation}
\label{eq:initial_solution}
\left(\sum_{p=1}^d (\redftmat{D}_1^p)\tp \redftmat{D}_1^p + d \redftmat{C}\tp \redftmat{C} \right) \vectil{f}_1^{l,(0)} = (\redftmat{D}_1^l)\tp \tilde{\vecn{y}}_1
\end{equation}
for $l = 1,\ldots,d$.
This provides a starting point for the Gauss-Seidel optimization in the first frame. The systems in \eqref{eq:initial_solution} share the same sparse coefficients and can be solved efficiently with a direct sparse solver.
\subsection{Detection}
At the detection stage, the location of the target in a new frame $t$ is estimated by applying the filter $\hat{f}_{t-1}$ that has been updated in the previous frame. Similar to \cite{Li2014}, we apply the filter at multiple resolutions to estimate changes in the target size. The samples $\{z_r\}_{r \in \left\{\left\lfloor\frac{1-S}{2}\right\rfloor, \ldots, \left\lfloor\frac{S-1}{2}\right\rfloor\right\}}$ are extracted centered at the previous target location and at the scales $a^r$ relative to the current target scale. Here, $S$ denotes the number of scales and $a$ is the scale increment factor. The sample $z_r$ is constructed by resizing the image according to $a^r$ before the feature computation.
\noindent \textbf{Fast Sub-grid Detection}:
Generally, the training and detection samples $x_k$ and $z_k$ are constructed using a grid strategy with a stride greater than one pixel. This leads to only computing the detection scores \eqref{eq:mdcf_detection} on a coarser grid. We employ an interpolation approach that allows computation of pixel-dense detection scores. The detection scores \eqref{eq:mdcf_detection} are efficiently interpolated with trigonometric polynomials by utilizing the computed DFT coefficients. Let $\hat{s} \defeq \ft\{S_f(z)\} = \sum_{l=1}^d \hat{z}^l \pmult \hat{f}^l$ be the DFT of the detection scores $S_f(z)$ evaluated at the sample $z$. The detection scores $s(u,v)$ at the continuous locations $(u,v) \in [0,M) \times [0,N)$ in $z$ are interpolated as,
\begin{equation}
\label{eq:score_interp}
s(u,v) = \frac{1}{MN} \sum_{m=0}^{M-1} \sum_{n=0}^{N-1} \hat{s}(m,n) e^{i2\pi \left(\frac{m}{M}u + \frac{n}{N}v \right)} .
\end{equation}
Here, $i$ denotes the imaginary unit. We aim to find the sub-grid location that corresponds to the maximum score: $(u^*,v^*) = \argmax_{(u,v) \in [0,M) \times [0,N)} s(u,v)$.
The scores $s$ are first evaluated at all grid locations $s(m,n)$ using \eqref{eq:mdcf_detection}. The location of the maximal score $(u^{(0)},v^{(0)}) \in \Omega$ is used as the initial estimate. We then iteratively maximize \eqref{eq:score_interp} using Newton's method, by starting at the location $(u^{(0)},v^{(0)})$. The gradient and Hessian in each iteration are computed by analytically differentiating \eqref{eq:score_interp}. We found that only a few iterations is sufficient for convergence.
We apply the sub-grid interpolation strategy to maximize the classification scores $s_r$ computed at the sample $z_r$. The procedure is applied for each scale level independently. The scale level with the highest maximal detection score is then used to update target location and scale.
Excluding the feature extraction, the total computational complexity of our tracker sums up to $\mathcal{O}(dSMN\log{MN}+SMNN_\text{Ne}+(d+K^2)dMNN_\text{GS})$. Here, $N_\text{Ne}$ denotes the number of iterations in the sub-grid detection. In our case, the expression is dominated by the last term, which originates from the filter optimization.
\section{Related Work}
The MOSSE tracker \cite{MOSSE2010} and the kernelized extension \cite{Henriques12d} rely on using only grayscale image samples for learning the filter. Later works \cite{DanelljanCVPR14,HenriquesPAMI15,DanelljanBMVC14,Li2014} have gained a notable improvement by learning multi-channel filters on multi-dimensional features, such as HOG \cite{Dalal05} or Color-Names \cite{Weijer09a}. However, to become computationally viable, these approaches rely on harsh approximations of the original multi-channel correlation filter formulation. Typically, a filter is computed for each sample independently and then averaged over time. This may lead to suboptimal learning since all samples are not considered simultaneously in the training. In this work we therefore aim at learning multi-channel correlation filters for the online tracking problem by directly minimizing the associated loss function.
|
\section{Introduction}
Traversable wormholes have long been a source of fascination as a
method of long distance transportation \cite{Morris1988a}. However,
such configurations require matter that violates the null energy condition,
which is believed to apply in physically reasonable classical theories.
In quantum field theory, the null energy condition is false, but in
many situations there are other no-go theorems that rule out traversable wormholes.
In this work we find that adding certain interactions that couple
the two boundaries of eternal AdS-Schwarzschild results in a quantum
matter stress tensor with negative average null energy, rendering
the wormhole traversable after gravitational backreaction. The coupling we turn on has the effect of modifying the boundary conditions of a scalar field in the bulk, which changes the metric at 1-loop order.
Violation of the averaged null energy condition (ANEC) is a prerequisite
for all traversable wormholes \cite{Morris1988,Visser1996,Visser2003,Hochberg1998}.
It states that there must be infinite null geodesics passing through the wormhole, with tangent vector $k^{\mu}$ and affine parameter $\lambda$,
along
which
\begin{equation}
\int^{+\infty}_{-\infty} T_{\mu\nu}k^{\mu}k^{\nu}d\lambda<0.
\end{equation}
The physical picture is that by Raychaudhuri's equation for null geodesic
congruence, light rays will defocus only when ANEC is violated. In
that case, the light rays that focus in one end
of the wormhole can defocus when going out the other end.
There are reasonable arguments that the ANEC is always obeyed along infinite achronal geodesics \cite{Graham:2007va, Kontou:2012ve, Kontou:2015yha, Kelly:2014mra, Wall:2009wi}.\footnote{A set of points is achronal if no two of the points can be connected by a timelike curve; otherwise it is chronal.} This is sufficient to rule out traversable wormholes joining two otherwise disconnected regions of spacetime \cite{Graham:2007va}. Furthermore, the generalized second law (GSL) of causal horizons also rules out traversable wormholes connecting two disconnected (asymptotically flat or AdS) regions, due to the fact that the future horizon of a lightray crossing through the wormhole has divergent area at very early times, which contradicts the increase of generalized entropy along the future horizon \cite{Wall2013}.
For small semiclassical perturbations to a stationary causal horizon, both the GSL and the ANEC follow from lightfront quantization methods that are valid for free or superrenormalizable field theories \cite{Wall2012}. (There is also evidence that these results extend to more general field theories \cite{Faulkner:2016mzt,Hofman:2008ar,Hofman:2016awc,Koeller:2015qmn,Bunting:2015sfa}).
In our configuration, signals from early times on the horizon can intersect it again at late times, by passing through the directly coupled boundaries. The causal structure of the manifold is modified as a result, changing the commutation relations along null geodesics through the wormhole and making them no longer achronal. For the same reason, a causal horizon extending through the wormhole intersects itself, removing the piece with divergent area. Hence the above impossibility results do not apply. The negative energy matter in our configuration is similar to the Casimir effect, since the interaction between the boundaries implies that the radial direction is effectively a compact circle.
Another problematic aspect of traversable wormholes is that
they have the potential to lead to causal inconsistencies. For example,
by applying a boost to one end of a wormhole one could attempt to
create a configuration with closed time-like curves \cite{Morris1988}. The direct interaction
of the boundaries that we require implies that no such paradoxes may arise (for a more detailed discussion, see section 4).
The traversable wormhole we find is the first such solution that has been shown to be embeddable in a UV complete theory of gravity. A phenomenological model of a static BTZ wormhole that becomes traversable as a result of nonperturbative effects in a $1/N$ expansion was proposed in \cite{Solodukhin:2005qy}, however it was not shown that the metric obeys any field equations. A traversable wormhole solution of five dimensional Einstein-Gauss-Bonnet gravity was found in \cite{Bhawal:1992sz, Thibeault:2005ha, Arias:2010xg}, however that low energy effective theory appears to lack a UV completion \cite{Camanho:2014apa}. Another example was found \cite{Barcelo:1999hq} in a theory with a conformally coupled scalar, in a regime in which the effective Newton's constant becomes negative. This suggests that this solution also cannot arise in a UV complete model. The important fact that the boundary CFT dual of a traversable wormhole must involve interactions between the two CFTs was noted in \cite{Solodukhin:2005qy, Arias:2010xg}.
The eternal black hole with two asymptotically AdS regions is the
simplest setting to investigate these questions \cite{Maldacena2003}.
We will deform the system by turning on a relevant double trace deformation \cite{Aharony2005}
\begin{equation}
\delta S=\int dt\ d^{d-1}x\ h(t,x)\mathcal{O}_{R}(t,x)\mathcal{O}_{L}(-t,x), \label{eq:1.2}
\end{equation}
where ${\cal O}$ is a scalar operator of dimension less than $d/2$, dual to a scalar field $\varphi$.
This connects the boundaries with the same time orientation, since the $t$
coordinate runs in opposite directions in two wedges (see
Fig. \ref{fig:Penrose-diagram-of}). The small deformation $h(t,x)$ has support only after some turn-on time $t_{0}$. By the AdS/CFT
correspondence, we can be certain that this relevant deformation corresponds to a consistent
configuration in quantum gravity.
The eternal black hole has a Killing symmetry which is time-like outside
the horizon. Null rays along the horizon $V=0$ pass through the
bifurcation surface of the Killing vector, and asymptote to $t\rightarrow-\infty$
on the left boundary and $t\rightarrow+\infty$ on the right boundary
(see Fig. \ref{fig:Kruskal}). Denote the affine parameter along this ray as
$U$. In the linearized analysis around this solution, the throat
will become marginally traversable if $\int dU\ T_{UU}<0$, where the integral is over the whole $U$ coordinate.
\begin{figure}[h]
\begin{centering}
\subfloat[\label{fig:Penrose-diagram-of}]{\protect\begin{centering}
\protect\includegraphics[width=4cm]{Penrose_diagram}\protect
\par\end{centering}
}$\qquad$\subfloat[\label{fig:Kruskal}]{\protect\begin{centering}
\protect\includegraphics[width=4.5cm]{Kruskal_coord}\protect
\par\end{centering}
}
\par\end{centering}
\protect\caption{(a) is the Penrose diagram and (b) shows the Kruskal coordinates of the eternal black
hole}
\end{figure}
It is instructive to see explicitly in this case that a small spherically symmetric perturbation of the stress tensor $T_{\mu\nu}\sim O(\epsilon)$ results in a traversable wormhole exactly when the ANEC is violated, by solving the linearized Einstein equation for $h_{\mu\nu}=\delta g_{\mu\nu}\sim O(\epsilon)$.
Using Kruskal coordinates for the background metric, we find that at $V=0$,
\begin{equation}
\f{(d-2)}4\left[\left((d-3)r_{h}^{-2}+(d-1)\ell^{-2}\right)\left(h_{UU}+\partial_{U}(Uh_{UU})\right)-2r_{h}^{-2}\partial_{U}^{2}h_{\phi\phi}\right]= 8\pi G_N\, T_{UU}\label{eq:1.4}
\end{equation}
where $\phi$ is the
azimuthal angle, $r_{h}$ is the horizon
radius of the black hole and the cosmological constant is $\Lambda=-\f{(d-2)(d-1)}{2l^{2}}<0$.
Since the deformation of the Hamiltonian is small,
after the scrambling time, the fields ought to approach a stationary state with respect to an asymptotic Killing symmetry $U \partial U$. Hence $T_{UU}$ must decay faster than $U^{-2}$, as does
each term in LHS of (\ref{eq:1.4}) after imposing a suitable gauge at past and future infinity. Therefore, if we integrate (\ref{eq:1.4}) over $U$ the total derivative terms drop out and we obtain
\begin{equation}
8\pi G_N \int dU T_{UU}=\f{(d-2)}4\left((d-3)r_{h}^{-2}+(d-1)\ell^{-2}\right)\int dU h_{UU}
\end{equation}
Linearized diffeomorphisms around the stationary black hole background act on $h_{\mu \nu}$, but when the AdS asymptotic conditions are imposed the quantity $\int^{+\infty}_{-\infty} d U \ h_{U U}$ is gauge invariant.
Note that the null ray originating on the past horizon is given in coordinates by
\begin{equation}
V(U)=-(2g_{UV}(0))^{-1}\int_{-\infty}^{U}dU h_{UU}
\end{equation}
after including the perturbation to linear order,
where $g_{UV}(0)<0$ is the $UV$ component of the original metric on the $V=0$
slice. If the ANEC is violated, $V(+\infty)<0$, and a light
ray from left boundary will hit the right boundary after
finite time.
Note that if there existed any state in which the wormhole was traversable in
the system defined by the decoupled Hamiltonian, $H_{L}+H_{R}$, then
it would contradict the AdS/CFT duality. This is because in the decoupled
system, no operator on the left can influence the right, which implies
that no signal can be transmitted between the boundaries through the
bulk.
At the linearized level, if one modifies the state as $|\textrm{tfd}\rangle\rightarrow e^{i\epsilon A}|\textrm{tfd}\rangle$
for small $\epsilon$, the average null energy becomes $\langle\int dUT_{UU}\rangle=i\epsilon\langle[\int dUT_{UU},A]\rangle$.
If this were non-vanishing for any operator $A$, then by adjusting
the sign of $\epsilon$, the throat could be made traversable. It
is easy to check that the expectation value of this commutator indeed vanishes.
In fact, $\bra{\text{tfd}}$ is invariant under $H_{R}-H_{L}$, which
corresponds to the bulk Killing symmetry $i\partial_{t}$ (note the
directions are opposite in left and right wedges). On the horizon $V=0$,
one can show $\partial_{t}=U\partial_{U}$ in Kruskal coordinates, which
is just a dilation of the $U$ direction. Note that under the
$U\rightarrow\lambda U$ scaling, $T_{UU}\rightarrow\lambda^{-2}T_{UU}$ and $dU\rightarrow\lambda dU$,
which implies $[H_{R}-H_{L},\int dUT_{UU}]=-i\int dUT_{UU}$. Therefore
\begin{equation}
(H_{R}-H_{L})\int dUT_{UU}|\textrm{tfd}\rangle=[H_{R}-H_{L},\int dUT_{UU}]|\textrm{tfd}\rangle+\int dUT_{UU}(H_{R}-H_{L})|\textrm{tfd}\rangle=-i\int dUT_{UU}|\textrm{tfd}\rangle.
\end{equation}
This implies that $\int dUT_{UU}\bra{\text{tfd}}$ is either an eigenvector
of $H_{R}-H_{L}$ with eigenvalue $-i$, or identically zero. Since
$H_{R}-H_{L}$ is a Hermitian operator, whose eigenvalues must be real,
it follows that $\int dUT_{UU}|\textrm{tfd}\rangle=0$. In
other words, $T_{UU}$ in the modified state along $U>0$ will exactly
cancel that along $U<0$. Beyond the linearized level, one can show that the backreaction
always causes the throat to lengthen \cite{Maldacena2013, Shenker2013}, so that
it cannot be traversed in any state of the decoupled system, as expected.
We will consider a deformation of the Hamiltonian that turns on at
some time $t_{0}$ in (\ref{eq:1.2}).\footnote{We do not consider the case of a time-independent interaction, in order to prevent the quantum state from becoming non-regular on the past horizon.} At the linearized level, this
has the same effect as changing the state to the future of $t_{0}$.
Now there is no reason for the above cancellation to occur since $T_{UU}$
along $U<0$ is unchanged. Therefore, one expects that generically
by an appropriate choice of sign one will render the Einstein-Rosen bridge traversable, as long as the deformation couples the two boundaries.\footnote{A deformation of only $H_R$ has the same effect on the ANE as a change in the state, by bulk causality, since the past causal cone of the deformation does not intersect the $V=0$ null sheet. This again agrees with the fact that when the boundaries are decoupled, no traversable wormhole can exist.}
The simplest option in the large $N$ limit is a double trace deformation. This has the effect of modifying the boundary conditions for the dual scalar field, such that some amplitude of a wave hitting one boundary will be transmitted to the opposite one. This does not change the eternal black hole solution classically, but results in a quantum correction to the matter stress tensor.
In order to be sure that the configuration is an allowed one, we choose the deformation to be relevant. Then it will be a renormalizable deformation of the CFT, and the dual geometry will not be modified by backreaction in an uncontrolled way at the AdS boundaries. Also, heuristically, the effect of such a deformation coupling the two CFT's should be strong in the IR, which suggests that it renders the deep interior traversable.
Recall that the conformal weight of a scalar operator $\mathcal{O}_{i}$
is given by $\Delta=\f d2\pm\sqrt{(\f d2)^{2}+M^{2}}$, where $M$ is the mass of the bulk field, and the plus
or minus sign depends on the choice of asymptotic boundary conditions.
In the case $M^{2}>0$, only the plus sign leads to normalizable
modes. However, unitarity in AdS space \cite{Breitenlohner1982} allows
a slightly tachyonic bulk field: $M^{2}>-(\f d2)^{2}$, in which
modes of both signs are normalizable and we are free to choose either
one. To have a relevant deformation, we start with the alternative boundary condition, associated with the minus sign.
A brief overview of this paper is as follows. In section 2, we calculate the bulk two-point function with the modified Hamiltonian at linear order in $h$. In section 3, we use the point-splitting method to calculate $T_{UU}$ on the $V=0$ slice. Numerical result shows that $T_{UU}$ is rendered negative by our boundary interaction. We find an analytic expression for $\int dU T_{UU}$, which is negative for all $0<\Delta<1$. In section 4 we calculate the energy and entropy of the resulting CFT state, and describe their holographic bulk duals. In section 5, we discuss the properties of this traversable wormhole and propose a quantum teleportation interpretation in the ER=EPR context. The appendix is a detailed calculation of $\int dU T_{UU}$.
Throughout we use units where $c = \hbar = 1$.
\section{Modified bulk two-point function}
For simplicity, we consider the eternal BTZ black hole \cite{Banados1992,Banados1993}
(for a review, see \cite{Carlip1995}), whose metric is
\begin{equation}
ds^{2}=-\f{r^{2}-r_{h}^{2}}{\ell^{2}}dt^{2}+\f{\ell^{2}}{r^{2}-r_{h}^{2}}dr^{2}+r^{2}d\phi^{2}\label{eq:2.1-1}
\end{equation}
The inverse temperature of the BTZ black hole is determined by its horizon
radius $r_{h}$ as $\beta=2\pi\ell^{2}/r_{h}$. Here and below we set
AdS length $\ell$ to 1. Without any deformation of the Hamiltonian, the
bulk free field two-point function in the BTZ background with $r^{-\Delta}$
fall-off was first derived by the mode sum method in \cite{Ichinose1995}.
In right wedge, it is
\begin{equation}
\avg{\varphi_{R}(x)\varphi_{R}(x')}_{0}=\frac{1}{2^{3-\Delta}\pi}(G_{+}+G_{-})(G_{+}^{-1}+G_{-}^{-1})^{1-2\Delta}\label{eq:2.1}
\end{equation}
where
\begin{equation}
G_{\pm}\equiv\left(\frac{rr'}{r_{h}^{2}}\cosh r_{h}\Delta\phi\pm1-\frac{(r^{2}-r_{h}^{2})^{1/2}(r'^{2}-r_{h}^{2})^{1/2}}{r_{h}^{2}}\cosh r_{h}\Delta t\right)^{-1/2}.
\end{equation}
The bulk field operator $\varphi_{R}(x)$ in the eternal black hole background
can be understood as a non-local CFT operator \cite{Papadodimas:2012aq}. In particular, $\varphi_{R}(x)$
can be expanded in terms of the right boundary dual operator as
\begin{equation}
\varphi_{R}(t,r,\phi)=\int_{\omega>0}d\omega\,dm\left(f_{\omega m}(r)e^{-i\omega t+im\phi}\mathcal{O}_{\omega m}+f_{\omega m}^{*}(r)e^{i\omega t-im\phi}\mathcal{O}_{\omega m}^{\dagger}\right)
\end{equation}
where $f_{\omega m}(r)e^{-i\omega t+im\phi}$ are bulk positive frequency normalizable
modes approaching $r^{-\Delta}$ when $r\rightarrow\infty$ and $\mathcal{O}_{\omega m}$ is
the boundary annihilation operator defined by
\begin{equation}
\mathcal{O}(t,\phi)=\int d\omega\,dm\left(e^{-i\omega t+im\phi}\mathcal{O}_{\omega m}+e^{i\omega t-im\phi}\mathcal{O}_{\omega m}^{\dagger}\right).
\end{equation}
Therefore, the bulk to boundary correlation function is given by
\begin{align}
K_{\Delta}(r,t,\phi) & \equiv\avg{\varphi_{R}(t,r,\phi)\mathcal{O}(0,0)}=\lim_{r'\rightarrow\infty}r'^{\Delta}\avg{\varphi_{R}(t,r,\phi)\varphi_{R}(0,r',0)}_{0}\nonumber \\
& =\f{r_{h}^{\Delta}}{2^{\Delta+1}\pi}\left(-\f{(r^{2}-r_{h}^{2})^{1/2}}{r_{h}}\cosh r_{h}t+\f r{r_{h}}\cosh r_{h}\phi\right)^{-\Delta},\label{eq:2.5}
\end{align}
where we used translation symmetry in $t$ and $\phi$
to move $(t',\phi')$ to the boundary origin. This expression is real
only when $(r,t,\phi)$ is space-like separated from the boundary origin.
For time-like separation, general analytic properties of Wightman functions imply that
one should change $t$ to $t-i\epsilon$, which assigns a phase of $e^{-i\pi\Delta}$
when $t>0$ and of $e^{i\pi\Delta}$ when $t<0$.
Now we consider the time dependent modified Hamiltonian of (\ref{eq:1.2}):
\begin{equation}
\delta H(t)=-\int d\phi\,h(t,\phi)\mathcal{O}_{R}(t,\phi)\mathcal{O}_{L}(-t,\phi),\label{eq:2.6}
\end{equation}
where
$h(t,\phi)=0$ when $t<t_{0}$. Using evolution operator $U(t,t_{0})=\mathcal{T} e^{-i\int_{t_{0}}^{t}dt\delta H(t)}$ in interaction
picture, the bulk two-point function is
\begin{equation}
\avg{\varphi^H_{R}(t,r,\phi)\varphi^H_{R}(t',r',\phi')}=\avg{U^{-1}(t,t_{0})\varphi^I_{R}(t,r,\phi)U(t,t_{0})U^{-1}(t',t_{0})\varphi^I_{R}(t',r,\phi)U(t',t_{0})}\label{eq:2.7}
\end{equation}
where superscripts $H$ and $I$ represent Heisenberg and interaction picture respectively. To leading order in $h$, (\ref{eq:2.7})
is (suppressing $r$ and $\phi$ coordinates and omitting $I$)
\begin{align}
G_{h}\equiv & -i\int_{t_{0}}^{t}dt_{1}h(t_{1})\avg{[\mathcal{O}_{L}(-t_{1})\mathcal{O}_{R}(t_{1}),\varphi_{R}(t)]\varphi_{R}(t')}-i\int_{t_{0}}^{t'}dt_{1}h(t_{1})\avg{\varphi_{R}(t)[\mathcal{O}_{L}(-t_{1})\mathcal{O}_{R}(t_{1}),\varphi_{R}(t')]}\nonumber \\
\simeq & -i\int_{t_{0}}^{t}dt_{1}h(t_{1})\avg{\varphi_{R}(t')\mathcal{O}_{L}(-t_{1})}\avg{[\mathcal{O}_{R}(t_{1}),\varphi_{R}(t)]}+(t\leftrightarrow t')\nonumber \\
= & i\int_{t_{0}}^{t}dt_{1}h(t_{1})K_{\Delta}(t'+t_{1}-i\beta/2)\left[K_{\Delta}(t-t_{1}-i\epsilon)-K_{\Delta}(t-t_{1}+i\epsilon)\right]+(t\leftrightarrow t')\nonumber \\
= & 2\sin\pi\Delta\int dt_{1}h(t_{1})K_{\Delta}(t'+t_{1}-i\beta/2)K_{\Delta}^{r}(t-t_{1})+(t\leftrightarrow t')\label{eq:2.8}
\end{align}
where in the second line we used large $N$ factorization and causality,
in that $\mathcal{O}_{L}$ commutes with any $\varphi_{R}$, in the third line
we used the KMS condition \cite{Haag1967}
\begin{equation}
\avg{\mathcal{O}_{R}(t)\mathcal{O}_{L}(t')}_{tfd}=\avg{\mathcal{O}_{R}(t)\mathcal{O}_{R}(t'+i\beta/2)}_{tfd}
\end{equation}
and in the last line $K_{\Delta}^{r}$ is the retarded correlation function
\begin{equation}
K_{\Delta}^{r}(t,r,\phi)=|K_{\Delta}(t,r,\phi)|\ \theta(t)\ \theta\left(\f{(r^{2}-r_{h}^{2})^{1/2}}{r_{h}}\cosh r_{h}t-\f r{r_{h}}\cosh r_{h}\phi\right)\label{eq:2.10}
\end{equation}
One can also derive (\ref{eq:2.8}) using the bulk mode sum method with
modified boundary conditions. This approach would allow one to compute the stress tensor for finite $h$, not just perturbatively. According to Lorentzian AdS/CFT, the
double trace deformation \cite{Witten2001,Berkooz2002}, from the point
of view of the right wedge, is equivalent to a source term $h(t,\phi)\mathcal{O}_{L}(-t,\phi)$,
for $\mathcal{O}_{R}(t)$,
activating the initially frozen fall-off component of the bulk field.
The same applies to the left
wedge. Therefore the asymptotic behavior of a global bulk mode
$\varphi$ living in the entire eternal black hole should satisfy
\begin{align}
\varphi(r\rightarrow\infty_{R}) & \rightarrow\alpha_{R}(t,\phi)r^{-\Delta}+\beta_{R}(t,\phi)r^{-2+\Delta},\ \beta_{L}(t,\phi)=h(-t,\phi)\alpha_{R}(-t,\phi)\label{eq:2.11}\\
\varphi(r\rightarrow\infty_{L}) & \rightarrow\alpha_{L}(t,\phi)r^{-\Delta}+\beta_{L}(t,\phi)r^{-2+\Delta},\ \beta_{R}(t,\phi)=h(t,\phi)\alpha_{L}(-t,\phi)\label{eq:2.12}
\end{align}
where the subscript $1$ is for right wedge and $2$ is for left wedge.
The thermofield double state of the eternal black hole is the vacuum state
in the Kruskal patch \cite{Israel1976}. This is analogous to the relation
between the Minkowski vacuum and the Rindler thermofield double state \cite{Unruh1976}.
Choosing the appropriate global bulk modes $H_{\omega m}^{(\pm)}$\footnote{This step is very tricky because at order $h$, the $r^{-\Delta}$
component is not constrained by the deformation. The only requirement
is that the modified two point function must be regular on
horizon. We were able to find a choice to reproduce (\ref{eq:2.8})
up to normalization. } and applying the method of \cite{Israel1976}, we can construct $\varphi$ as
\begin{equation}
\varphi(x)=\int_{\omega>0}d\omega\,dm(H_{\omega m}^{(+)}(x)b_{\omega m}^{(+)}+H_{\omega m}^{(-)}(x)b_{\omega m}^{(-)\dagger}+h.c.)
\end{equation}
where $b_{\omega m}^{(\pm)}$ are annihilation operators used to define the
vacuum. We find the two-point function in this vacuum is the same
as (\ref{eq:2.8}) up to normalization. Since the calculation is quite
involved, we do not include it in this paper.
\section{1-loop stress tensor}
The stress tensor is given by variation of action with respect to
$g^{\mu\nu}$,
\begin{equation}
T_{\mu\nu}=\partial_{\mu}\varphi\partial_{\nu}\varphi-\frac{1}{2}g_{\mu\nu}g^{\rho\sigma}\partial_{\rho}\varphi\partial_{\sigma}\varphi-\frac{1}{2}g_{\mu\nu}M^{2}\varphi^{2}
\end{equation}
The 1-loop expectation value can be calculated by point splitting,
\begin{align}
\langle T_{\mu\nu}\rangle & =\lim_{x\rightarrow x'}\partial_{\mu}\partial'_{\nu}G(x,x')-\frac{1}{2}g_{\mu\nu}g^{\rho\sigma}\partial_{\rho}\partial'_{\sigma}G(x,x')-\frac{1}{2}g_{\mu\nu}M^{2}G(x,x')\label{eq:3.2}
\end{align}
where $G(x,x')$ is 2-point function. In this formula, one must renormalize the stress tensor by subtracting the coincident point singularities from the 2-point function, which are given by the Hadamard conditions \cite{parker2009quantum}.
Since these are determined by the short distance dynamics, this subtraction is unchanged when we modify the boundary conditions, and it has no effect on the order $h$ correction that we are interested in.
At leading order, as we reviewed in the Introduction, $\int dUT_{UU}$ is zero on the horizon $V=0$. Indeed, the leading order two point function
in the BTZ black hole is (\ref{eq:2.1}) where $\phi$ has periodicity $2\pi$
and all $\Delta\phi+2\pi n$ images are summed. The only coincident point pole
comes from the $n=0$ component. Summing over the other $n$ components, one finds that in Kruskal coordinates the leading order stress
tensor $T_{UU}\sim O(V^{2})$ in the $V\ra0$ limit, so that $T_{UU}=0$
along the horizon.
The subleading 2-point function is given by (\ref{eq:2.8}). Note
that $h(t,\phi)$ is dimensionful and its dimension is $2-2\Delta$ because
in (\ref{eq:2.6}) $\mathcal{O}$ has dimension $\Delta$\footnote{Here we implicitly defined the unit length angular coordinate $x\equiv\phi\ell$.
Taking the limit $r\rightarrow\infty$ in BTZ metric (\ref{eq:2.1-1}), the boundary
metric is flat $ds_{b}^{2}=-dt^{2}+dx^{2}$.}. Moreover, since $h(t,\phi)$ is a boundary CFT smearing function, it
should not depend on any bulk length scale (e.g. $r_{h}$ and $\ell$)
explicitly but only on the inverse temperature $\beta$. Let us assume that
$h(t,\phi)$ is uniform over $\phi$:
\begin{equation}
h(t,\phi)=\begin{cases}
h(2\pi/\beta)^{2-2\Delta} & t\geq t_{0}\\
0 & t<t_{0} \label{eq:3.3}
\end{cases}
\end{equation}
where $h$ is a dimensionless constant. In Kruskal coordinates
\begin{equation}
e^{2r_{h}t}=-\frac{U}{V},\;\frac{r}{r_{h}}=\frac{1-UV}{1+UV}
\end{equation}
the change in the 2-point function is
\begin{align}
G_{h}= & C_{0}\left(\f{2\pi}{\beta}\right)^{2\Delta-2}r_{h}\int\f{dU_{1}}{U_{1}}d\phi_{1}h(U_{1},\phi_{1})\left(\f{1+UV}{U/U_{1}-VU_{1}-(1-UV)\cosh r_{h}(\phi-\phi_{1})}\right)^{\Delta}\nonumber \\
& \times\left(\f{1+U'V'}{U'U_{1}-V'/U_{1}+(1-U'V')\cosh r_{h}(\phi'-\phi_{1})}\right)^{\Delta}+(U,V,\phi\leftrightarrow U'V',\phi')
\end{align}
where $C_{0}=\f{r_{h}^{2\Delta-2}\sin\Delta\pi}{2(2^{\Delta}\pi)^{2}}\left(\f{2\pi}{\beta}\right)^{2-2\Delta}$
and we transformed the integral over $t_{1}$ to Kruskal coordinates in which the boundary is $U_{1}V_{1}=-1$. Note that this result applies to both the black hole and black brane cases because the integration
of $\phi_{1}$ over $0$ to $2\pi$ and summing over $n$ with modification
$\phi_{1}\rightarrow\phi+2\pi n$ is equivalent to the integration of $\phi_{1}$
over the whole real axis. Since we only focus on $T_{UU}$ component on the
horizon $V=0$ and the derivative on $U$ and $U'$ in (\ref{eq:3.2})
has nothing to do with the value of $V$ and $V'$, we can take both
points to the horizon first, namely $V=V'=0$. Similarly, we can take
$\phi=\phi'$ first for simplicity. Since $h(t_{1},\phi_{1})$ is
uniform in $\phi_{1}$, $\partial_{\phi}$ is still a Killing vector of
the system and therefore $G_{h}$ should not depends on $\phi$. Defining
$y=\cosh r_{h}(\phi_{1}-\phi)$, on horizon we have
\begin{equation}
G_{h}=hC_{0}\int_{U_{0}}^{U}\f{dU_{1}}{U_{1}}\int_{1}^{\f U{U_{1}}}\f{2dy}{\sqrt{y^{2}-1}}\left(\f{U_{1}}{U-U_{1}y}\right)^{\Delta}\left(\f 1{U'U_{1}+y}\right)^{\Delta}+(U\leftrightarrow U')\equiv F(U,U')+F(U',U)\label{eq:3.6}
\end{equation}
where $U_{0}=e^{r_{h}t_{0}}$. The integral range of (\ref{eq:3.6})
is given by the step function in (\ref{eq:2.10}), which ensures that $U-U_{1}y\geq0$.
Note that the integral in (\ref{eq:3.6}) is dimensionless. Since
$G_{h}$ has dimension 1 ($\varphi_{R}$ has dimension $\f 12$ in 3-dimension
spacetime), if we restore $\ell$ in (\ref{eq:3.6}), we find the
total length scale dependence of $G_{h}$ is $\ell^{-1}$.
Note that $g_{UU}=0$ in the original BTZ geometry. By (\ref{eq:3.2}),
$T_{UU}$ on horizon is
\begin{equation}
T_{UU}=\lim_{U'\rightarrow U}\partial_{U}\partial_{U'}(F(U,U')+F(U',U))=2\lim_{U'\rightarrow U}\partial_{U}\partial_{U'}F(U,U')
\end{equation}
where we should note the dimension of $T_{UU}$ is the same as $G_{h}$
because $U$ is dimensionless. Since the integration ranges are only
functions of $U$, we can take the $U'$ derivative before evaluating
the integral
\begin{equation}
T_{UU}=-4h\Delta C_{0}\lim_{U'\rightarrow U}\partial_{U}\int_{U_{0}}^{U}dU_{1}\int_{1}^{\f U{U_{1}}}\f{dy}{\sqrt{y^{2}-1}}\f{U_{1}^{\Delta}}{(U-U_{1}y)^{\Delta}(U'U_{1}+y)^{\Delta+1}}\label{eq:3.8}
\end{equation}
Defining a new variable $x=\f{y-1}{U/U_{1}-1}$ and
integrating over $x$ we get
\begin{align}
T_{UU}= & -\f{4h\Delta C_{0}\Gamma(\f 12)\Gamma(1-\Delta)}{\sqrt{2}\Gamma(\f 32-\Delta)}\lim_{U'\rightarrow U}\partial_{U}\int_{U_{0}}^{U}dU_{1}\f{F_{1}(\f 12;\f 12,\Delta+1;\f 32-\Delta;\f{U_{1}-U}{2U_{1}},\f{U_{1}-U}{U_{1}(1+U'U_{1})})}{U_{1}^{-\Delta+1/2}(U-U_{1})^{\Delta-1/2}(1+U'U_{1})^{\Delta+1}}\label{eq:3.9}
\end{align}
where we used the integral representation of Appell hypergeometric
function. The integral over $U_{1}$ is finite as long as $\Delta-1/2<1$,
namely $\Delta<3/2$, because in the integrated region, the only potentially
divergent point is around $U_{1}\rightarrow U$ from below since $F_{1}$
is a complete function when $\Delta<3/2$. In particular, when $U_{1}\sim U$,
$F_{1}\ra1$, which implies $\Delta<3/2$ is the sufficient and necessary
condition for integrability. Defining a new variable $z=\f{U_{1}-U_{0}}{U-U_{0}}$,
the domain of integration in (\ref{eq:3.9}) becomes 0 to 1 and therefore
we can exchange the order of $\partial_{U}$ and $\int dz$. After differentiating
w.r.t. $U$, and restoring the variable $U_{1}$, we get
\begin{align}
T_{UU}= & -\f{2h\Delta C_{0}\Gamma(\f 12)\Gamma(1-\Delta)}{\Gamma(\f 32-\Delta)}\int_{U_{0}}^{U}\f{dU_{1}U_{1}^{2\Delta}(f_{1}+f_{2}+f_{3})}{(U-U_{0})(U-U_{1})^{\Delta-1/2}(1+U_{1}^{2})^{\Delta+1}U^{\Delta+1}(U+U_{1})^{1/2}}\label{eq:3.10}
\end{align}
where
\begin{align}
f_{1} & =\f{-2\Delta(UU_{1}^{2}+U_{0})+3UU_{0}U_{1}+U_{0}+2U_{1}}{1+UU_{1}}F_{1}(1-\Delta,\f 12,1+\Delta,\f 32-\Delta,u,v)\\
f_{2} & =\f{2(1+\Delta)(U-U_{1})(U_{0}+2UU_{0}U_{1}-UU_{1}^{2})}{(2\Delta-3)U(1+U_{1}^{2})(1+UU_{1})}F_{1}(1-\Delta,\f 12,2+\Delta,\f 52-\Delta,u,v)\\
f_{3} & =\f{U_{0}(U-U_{1})}{(2\Delta-3)(U+U_{1})}F_{1}(1-\Delta,\f 32,1+\Delta,\f 52-\Delta,u,v)\\
u & =\f{U-U_{1}}{U+U_{1}},\qquad v=\f{U-U_{1}}{U(1+U_{1}^{2})}
\end{align}
Performing the final integral numerically, we plot the result in
Fig. \ref{fig:3.1a}.
\begin{figure}
\begin{centering}
\subfloat[\label{fig:3.1a}]{\protect\begin{centering}
\protect\includegraphics[width=8cm]{TUU1}\protect
\par\end{centering}
}\subfloat[\label{fig:3.1b}]{\protect\begin{centering}
\protect\includegraphics[width=8cm]{TUU2}\protect
\par\end{centering}
}
\par\end{centering}
\protect\caption{(a) shows the null energy along the horizon when the interaction is turned on at $U=U_{0}=1$ and never
shut off, with our choice for the sign of the coupling $h$; (b) shows the case where it is turned on at $U=U_{0}=1$ and turned
off at $U=U_{f}=2$. In both cases, $h=1$. We see clearly in both
(a) and (b) that $T_{UU}$ becomes negative after turn-on; in (b) $T_{UU}$
becomes positive after turn-off. Blue is for $\protect\Delta=0.1$; yellow
is for $\protect\Delta=0.2$; green is for $\protect\Delta=0.4$; pink is
for $\protect\Delta=0.6$; purple is for $\protect\Delta=0.8$}
\end{figure}
In the figure, we see that the null energy is negative after we turn on the insertion
at $U_{0}=1$ if we take positive $h$. Physically,
this means the light-like ray $V=0$ becomes time-like after $U_{0}$
and a spaceship that enters early enough may escape the black hole!
One may note that when $\Delta<1/2$, $T_{UU}$ is finite but when $\Delta>1/2$,
$T_{UU}$ is singular near insertion time $U_{0}$. However, this
singularity is not essential because it is integrable, as we will see
later when we calculate $\int dUT_{UU}$ along the horizon $V=0$. Indeed,
the classical solution of Einstein equations for a shockwave insertion
in the bulk in Kruskal coordinates contains a delta function, which
is also an integrable singularity \cite{Shenker2013}. One might also worry that the derivative of $g_{UU}$ and the Riemann curvature are singular
at the turn-on and turn-off times,
although $T_{UU}$ and $\int dUT_{UU}$ are not. However, this is simply due
to the fact that we turned the insertion on and off as a step function.
If this process were taken to be smooth enough, there would be no singularity.
To see the late time behavior, we can use the $z$ variable to rewrite
(\ref{eq:3.10}) in the large $U$ limit. In this limit, $f_{1}$ dominates among all $f_{i}$'s
in (\ref{eq:3.10}). Using the identity $F_{1}(a;b,b';c;z,0)={}_{2}F_{1}(a,b;c;z)$
we obtain
\begin{equation}
T_{UU}\rightarrow\f{4h\Delta^{2}C_{0}\Gamma(\f 12)\Gamma(1-\Delta)}{\Gamma(\f 32-\Delta)U^{2\Delta+2}}\int_{0}^{1}\f{dz\,z^{2\Delta+1}{}_{2}F_{1}(1-\Delta,\f 12,\f 32-\Delta,\f{1-z}{1+z})}{((z+\epsilon)^2+\epsilon)^{\Delta+1}(1-z)^{\Delta-1/2}(1+z)^{1/2}}\ra0_{+}
\end{equation}
where $\epsilon$ is a small number of order $U^{-1}$ and which implies that $T_{UU}$ becomes positive and decays to zero at
late times.
If we turn off the interaction at some finite time $U_{f}$, when $U>U_f$, we can safely pass $\partial_{U}$ into the $U_{1}$ integral,
which leads to
\begin{align}
T_{UU} & =-\f{4h\Delta C_{0}\Gamma(\f 12)\Gamma(1-\Delta)}{\Gamma(\f 12-\Delta)}\int_{U_{0}}^{U_{f}}dU_{1}\f{U_{1}^{2\Delta+1}F_{1}(-\Delta;\f 12,\Delta+1;\f 12-\Delta;\f{U-U_{1}}{U+U_{1}},\f{U-U_{1}}{U(1+U_{1}^{2})})}{(U-U_{1})^{\Delta+1/2}(U+U_{1})^{1/2}U^{\Delta+1}(1+U_{1}^{2})^{\Delta+1}}\label{eq:3.16}
\end{align}
In deriving (\ref{eq:3.16}), we used a property of the derivative of the
Appell hypergeometric function and equation (7a) in \cite{Schlosser2013}.
The numerical result is plotted in Fig. \ref{fig:3.1b}.
In this figure, we see that after turning off the interaction, $T_{UU}$
has a jump and becomes positive at late times. In particular, when
$\Delta>1/2$, $T_{UU}$ becomes divergent again right after $U_{f}$.
Fortunately, it is again an integrable divergence which should not
cause any physical problem. By the identity \cite{Prudnikov1992}:
\begin{equation}
F_{1}(a;b,b';c;x,y)=\sum_{m\geq0}\f{(a)_{m}(b)_{m}}{m!(c)_{m}}x^{m}{}_{2}F_{1}(a+m,b';c+m;y)\label{eq:3.17-1}
\end{equation}
the late time behavior can be analyzed by taking the $U\rightarrow\infty$ limit
in (\ref{eq:3.16}):
\begin{align*}
T_{UU} & \sim\f{4h\Delta^{2}C_{0}}{U^{2\Delta+2}}\log U\log\f{U_{f}}{U_{0}}\ra0_{+}
\end{align*}
Again, we find $T_{UU}$ becomes positive after some time and decays
to zero. In both late time analyses, $T_{UU}$ decays like
$U^{-2\Delta-2}$, which validates the assumption that $Uh_{UU}$ and
$\partial_{U}h_{\phi\phi}$ vanish when $U\rightarrow\infty$ in (\ref{eq:1.4}).
\begin{figure}
\begin{centering}
\includegraphics[width=10cm]{intTUU}
\par\end{centering}
\protect\caption{$\int dUT_{UU}$ as a function of $\protect\Delta$; blue is for $U_{0}=1$;
yellow is for $U_{0}=2$; green is for $U_{0}=1$ and $U_{f}=2$\label{fig:intTUU}}
\end{figure}
In the above discussion, we see that at some finite time $T_{UU}$ becomes
positive whether or not we turn off the insertion, which might appear dangerous for the fate of the worm hole. The crucial diagnostic is
the sign of the integral of $T_{UU}$ over the
whole $V=0$ slice. This is what determines whether a light ray on horizon
eventually reaches the boundary at spatial infinity.
It looks horrible to integrate
$U$ in (\ref{eq:3.10}) from $U_{0}$ to infinity. Interestingly
and surprisingly, by some tricks, we can get a closed form for
it (see Appendix \ref{sec:intTUU}):
\begin{equation}
\int_{U_{0}}^{\infty}dUT_{UU}=-\f{h\Gamma(2\Delta+1)^{2}}{2^{4\Delta}(2\Delta+1)\Gamma(\Delta)^{2}\Gamma(\Delta+1)^{2}\ell}\f{_{2}F_{1}(\f 12+\Delta,\f 12-\Delta;\f 32+\Delta;\f 1{1+U_{0}^{2}})}{(1+U_{0}^{2})^{\Delta+1/2}}\label{eq:3.17}
\end{equation}
If we turn off the interaction at $U_{f}$, the integral is just the
difference between $\int_{U_{0}}^{\infty}dUT_{UU}$ and $\int_{U_{f}}^{\infty}dUT_{UU}$.
We plot the result as a function of $\Delta$ in Fig. \ref{fig:intTUU}.
In this figure, we see that for all $\Delta$ values from $0$
to $1$, the integral of $T_{UU}$ is always negative, which demonstrates the existence of a traversable wormhole.
Furthermore, the earlier
we turn on the insertion, the larger the effect is. In particular,
even if $T_{UU}$ becomes positive in late times, the wormhole still
exists since the integral of $T_{UU}$ remains negative.
Note that $\Delta=0$ is a special case where $\int dUT_{UU}=0$. Indeed,
the only $\Delta=0$ operator in CFT is the identity and of course adding
the product of identity operators to Hamiltonian has no effect on the
system.
\section{Holographic Energy and Entropy}\label{HEE}
In this section we will consider the implications of a traversable wormhole for the holographic entanglement entropy conjecture, which in this context relates the entanglement entropy between the two boundary CFT's to the area/entropy of certain extremal surfaces in the bulk theory \cite{Ryu:2006bv,Hubeny:2007xt,Barrella:2013wja,Faulkner:2013ana,EngelhardtWall}.
As a preliminary, we discuss the change of energy of the CFT state. Long after the interaction is shut off, the system returns to thermal equilibrium. Thus the final horizon area can be determined from the energy of the system, measured on the left or the right. It is straightforward to check that, in our state, the energy decreases at linear order in $h$ with the sign choice that rendered the wormhole traversable:
After deforming the Hamiltonian $(t>t_{0})$, the state in Schr\"{o}dinger
picture is
\begin{equation}
\bra{\Psi(t)}=e^{-iH_{0}(t-t_{0})}U(t,t_{0})\bra{\text{tdf}}.
\end{equation}
Expanding $U(t,t_{0})$ to leading order in $h(t)$ given by (\ref{eq:3.3}), we find that the change in the energy on the right is
\begin{align}
\delta E_R &
i\int_{t_{0}}^{t}dt_{1}h(t_{1})\ket{\text{tdf}}[\delta H(t_{1}),H_{R}]\bra{\text{tdf}}\nonumber \\
& =\int_{t_{0}}^{t}dt_{1}d\phi h(t_{1})\ket{\text{tdf}}\partial_{t}\mathcal{O}_{R}(t_{1},\phi)\mathcal{O}_{L}(-t_{1},\phi)\bra{\text{tdf}}\nonumber \\
& =\f{hr_h^2}{2^{\Delta+1}\ell^3}\sum_{n}\left(\f 1{(\cosh2r_{h}t+\cosh2\pi r_{h}n)^{\Delta}}-\f 1{(\cosh2r_{h}t_{0}+\cosh2\pi r_{h}n)^{\Delta}}\right)\label{eq:12}
\end{align}
where in the second line we used the Heisenberg equation and in last line the boundary two-point function is obtained by taking limit $r\rightarrow\infty$
in (\ref{eq:2.1}) where $\phi$ has period $2\pi$, and all of its
images are summed over in the global BTZ black hole.\footnote{We consider global AdS here so that the total energy is finite.} If the interaction shuts off at $t_{f}$, the energy obviously becomes constant for $t>t_f$, and
$t$ in (\ref{eq:12}) is replaced by $t_{f}$.
Therefore, the effect of the interaction with $h>0$ is to reduce the energy. Note that if there are any UV divergences in the energy they cannot appear at linear order in $h$, since the interaction involves just one field in each CFT.
At least at first order in $h$, the entropy of entanglement $S_{EE}$ between the left and right boundaries should also be well-defined (and time dependent) even during the period of time when the interaction is turned on, if one thinks of the state as evolving by the deformed Hamiltonian in the original tensor product Hilbert space.
By the first law of entanglement, at linear order in $h$, the change in $S_{EE}$ is equal to $\beta \delta H_R$, thus it also decreases until the turn-off time $t_f$ after which it remains constant (as it must under decoupled unitary evolution on the left and right).
The change in $S_{EE}$ is ${\cal O}(1)$ in a $1/N$ expansion. At this order, in the bulk interpretation $S_{EE}$ has two parts, namely the small gravitational correction to the area $A/4G$ of the extremal surface, and the entanglement entropy of bulk fields $S_\text{bulk}$ on the spacelike slice from the extremal surface to the boundary slice at time $t$ \cite{Faulkner:2013ana,Barrella:2013wja}. More generally, it was proposed in \cite{EngelhardtWall} that, at general orders in $1/N$, one should consider the entropy outside the \emph{quantum extremal} surface, obtained by extremizing the total generalized entropy $S_\mathrm{gen} = A/4G + S_\text{bulk}$. When calculating the ${\cal O}(1)$ piece of the entropy, these two prescriptions agree on the value of the entropy but \cite{EngelhardtWall,Dong:2016eik} argued that the location of the quantum extremal surface is also physically important, because it provides a natural boundary for how much of the bulk can be reconstructed from the CFT state on a single boundary. One useful constraint on the location quantum extremal surface is the GSL, which states that $S_\mathrm{gen}$ is nondecreasing on any future horizon.
On a Cauchy slice prior to the time when the interaction is turned on, the geometry and bulk quantum state are that of the Hartle-Hawking state. Thus the quantum (and classical) extremal surface is located at the bifurcation surface of the original black hole ($E_1$ of Fig. \ref{fig:4.1}). On the other hand, after the interaction is over, the bulk quantum state of the fields changes and thus the quantum extremal surface must move. By left-right symmetry of the spacetime (together with the fact that the joint state of the entire system is pure so that $S_{EE}$ is the same on both sides) it can it can only move along the vertical axis of symmetry of the spacetime. Also, the GSL implies that the new location must be on or behind the causal horizon \cite{EngelhardtWall}, because otherwise it lies on a future horizon whose $S_\mathrm{gen}$ is generically increasing.
In fact, at first order in $h$, the GSL implies that the quantum extremal surface must lie exactly at the point $E_2$ in Fig. \ref{fig:4.1}, where the two future horizons intersect. For since the GSL is true in every state \cite{Wall2012}, and saturated for the Hartle-Hawking state, it must also be saturated for any first order perturbation to the Hartle-Hawking state \cite{Wall:2009wm}. But if $S_\mathrm{gen}$ is stationary along two linearly independent normal directions of $E_2$, then it must be a quantum extremal surface. Since the geometry near the bifurcation surface is unaffected by the perturbation, at order $h$ the area of the quantum extremal surface is unchanged from the original state. On the other hand, $S_\text{bulk}$ has nonlocal aspects. Therefore, the decrease of $S_{EE}$ at first order must be entirely due to a corresponding decrease in $S_\text{bulk}$ evaluated at the bifurcation surface $E_1$. Any effects arising from differences between $E_1$ and $E_2$ are suppressed by additional powers of $h$.
At second order in $h$, the GSL should no longer be saturated on the future horizon. Hence $S_\text{bulk}$ is increasing with time at $E_2$, and the quantum extremal surface will instead be located slightly above the point $E_2$.
We have not followed the evolution of the quantum extremal surface at intermediate times, but it seems that it must gradually move upwards from $E_1$ to its final location above $E_2$. After the interaction is over the boundary evolution is unitary, and hence neither $S_{EE}$ nor the quantum extremal surface changes.
\cite{EngelhardtWall} argued that the quantum extremal surface should always be spacelike to its corresponding CFT region. In a sense this continues to be true, since $E_1$ is spacelike to all the boundary points prior to turning on the interaction, while $E_2$ is spacelike to all the points after the interaction is turned off. But neither one is spacelike to the entire boundary for all time. For example, a unitary operator applied to the right boundary at sufficiently early times might affect the value of $S_\mathrm{gen}(E_2)$, and hence the right CFT entropy after the interaction. But that does not contradict any of the properties of the right CFT, since it does not have unitary time evolution (independent of the left CFT) during the period of the interaction.
Note that, if we assume that our holographic entropy prescription is correct when the CFT's are not coupled, it must necessarily also be correct when the CFT's are coupled. Before the interaction is turned on, we can simply consider the Hartle-Hawking spacetime as if there were no interaction. Similarly, after the interaction is over, we can consider a new spacetime which is dual to extrapolating the final state backwards in time, without any interaction. Neither of these spacetimes corresponds to a traversable wormhole, but they can be used for purposes of calculating $S_{EE}$ before or after the interaction is turned on. It is only when these two spacetimes are patched together, that they are seen to be a traversable wormhole geometry.
\begin{figure}[ht]
\begin{centering}
\includegraphics[width=8cm]{EE_plot}
\par\end{centering}
\protect\caption{The throat size is $\Delta V\sim h$. The red thick interval on the boundary is the duration of the deformation beginning at $t_0$ and ending at $t_f$. The metric in the light yellow region is unchanged and only that of the white region will have a nonzero backreaction correction. The orange thick curve is the future event horizon and the grey thick curve is the past event horizon.
$E_1$ is the original bifurcation surface. $E_2$ is the location where the right and left future horizons cross. The magenta curve is a null ray that passes through wormhole, deviating to right boundary.
\label{fig:4.1}}
\end{figure}
\section{Discussion}
We have demonstrated that the Einstein-Rosen bridge of a BTZ black hole becomes
slightly traversable after the addition of a two-boundary coupling. (We expect that a similar effect also occurs in $D > 3$ bulk spacetime dimensions, although it is harder to calculate the exact form of the stress-tensor.)
From (\ref{eq:3.17}), we see that the integral $\int dUT_{UU}$, giving the
deviation of null rays from the horizon, is proportional to $h$, which implies that the wormhole opens up by an amount (in units where $\hbar = 1$)
\begin{equation}
\Delta V \sim \frac{h G_N}{R^{D-2}}
\end{equation}
where $\Delta V$ is the difference of $V$ coordinate between the future horizon and the first lightray which can get through the wormhole (see Fig. \ref{fig:4.1}), and we assume that the black hole radius $r_h$, the AdS length $l$, and the amount of time $\Delta t$ the interaction is turned on for are all of the same characteristic length scale $R$.
The wormhole is only open for a small proper time in the interior region. This is quite different from the usual static wormhole solutions which do not have event horizons (e.g. \cite{Morris1988a}). Nevertheless, radial lightrays originating on the boundary at arbitrarily early times will cross through the portal to the other side; in this sense the wormhole is open at arbitrarily early boundary times on either side.
A (test) astronaut
from one boundary can only go through the wormhole before it closes,
and she reaches the other boundary long after the boundary-boundary interaction is turned on. One should note that since the coupling
we add breaks the Killing symmetry $H_{L}-H_{R}$, there is no
way to boost her back to a time before she entered the worm hole. Thus the way we glue the two boundaries fixes the relative time coordinate
between them, excluding the possibility of having closed time-like curves \cite{Morris1988}. Note that the traversable throat size depends
on the strength of the coupling and a signal transmitted through the
wormhole is only received at the other end after a very long time delay
if the gravitational effects of the coupling are small. Furthermore, the thermofield double state that we require is an extremely fine tuned state, so it would be very difficult to prepare such a configuration in which the astronaut could enter at early times from the left.
We have not yet considered the backreaction on the geometry coming from an actual (non-test) astronaut traversing the wormhole throat. An object travelling at light speed from left to right contributes to $T_{VV}$ but not to $T_{UU}$, so at the level of linearized gravity it will prevent objects from traversing in the \emph{other direction} (i.e. from right to left) but it will have no tendency to close the wormhole in the same direction that it is travelling. This suggests that the objects can still traverse the wormhole even after taking into account their own gravitational back-reaction.\footnote{Presumably there is \emph{some} limit on how much information can get through, since the black hole on the other side cannot radiate more energy than its initial mass, but determining the precise limit would require going beyond the linearized regime. There might also be an interesting limit on the total amount of \emph{information} which can get through the wormhole, coming from the Bousso bound \cite{Bousso:1999xy,Flanagan:1999jp} or its quantum generalization \cite{Bousso:2015mna,Strominger:2003br}.}
Another question concerns the interaction of the astronaut with the negative energy pulse of radiation travelling in the other direction. In the frame of reference defined by Kruskal coordinates, a quantum traversing the wormhole must be blueshifted up to a frequency $1/{\Delta V}$, while the pulse coming in the other direction has a frequency of order $1/R$. Here we are assuming that the interaction is turned on for about one light-crossing time $R$, and that there is no other time scale of relevance in the problem. Although an incoming pulse with negative total energy is not allowed in classical scattering problems, we will nevertheless attempt to build intuition by comparing the situation to a normal field theory scattering problem. The center-of-mass energy scale of the collision is given by
\begin{equation}
\sqrt{s} \sim \sqrt{\frac{R^{D-4}}{h G_N} }.
\end{equation}
Since $G \sim L_\text{planck}^{D-2}$, the center-of-mass energy is below the Planck scale in $D = 3$ (i.e. a BTZ black hole with any extra dimensions compactified at the Planck scale) but not when $D > 3$. However, even in higher dimensions we do not expect that full quantum gravity effects will be important. We nevertheless expect that it is legitimate to use the eikonal approximation, in which one solves for the propagation of each particle on the background field generated by the other particle. This corresponds to resumming ladder Feynman diagrams, whose amplitude scales with various powers of
\begin{equation}
\frac{G_N s}{b^{D-4}}\sim h^{-1},
\end{equation}
where $b$ is the impact parameter, and we have used the fact that $b \sim R$ (except for small tails of the wavefunction). Non-eikonal Feynman diagrams should be suppressed by additional powers of $G_N$ relative to eikonal diagrams with the same $s$ dependence \cite{Amati:1993tb}. Therefore we can consistently consider scenarios in which only the eikonal scattering is relevant, in which our calculation of the geometry shows that the wormhole is traversable.
It is interesting to consider what would happen if the two black holes were in the same component of space, rather than in different asymptotic regions. If the black holes were in a suitably entangled state, they should be connected by an Einstein-Rosen bridge \cite{Maldacena2013}, with the QFT state near the horizon close to the Hartle-Hawking state. The direct boundary interaction could then be produced by propagation through the ambient spacetime---this would be the same as the interaction we studied, except with a time delay. A similar calculation would then lead to a traversable wormhole. The negative ANE could be understood as coming from the Casimir effect associated to the cycle in space going from one black hole to the other in the ambient space and then threading the wormhole. Of course, the effect would be enhanced if the signals sent between the black holes were directed and amplified (otherwise the Casimir energy would be extremely tiny if the black holes were far apart). No causal paradoxes would arise because the traversability depends on backreaction due to the existence of a casual path between the black holes in the ambient spacetime.
Since any infinite null geodesic which makes it through a wormhole must be chronal (as discussed in the Introduction), such wormholes do not enable one to travel faster than light over long distances through space. Hence traversable wormholes are like getting a bank loan: you can only get one if you are rich enough not to need it.
The traversable wormhole we found has an interesting interpretation in the context of ER=EPR \cite{Maldacena2013}.
Maldacena and Susskind conjectured that any pair of entangled
quantum systems are connected by an Einstein-Rosen bridge (the non-traversable
wormhole).
The crucial difference in our work is that
we allow interaction between the entangled systems, which is
assumed to be negligible in ER=EPR.
What we have shown is that
in this case the Einstein-Rosen bridge can open to become a traversable wormhole.
Our example thus provides a way to operationally verify a salient feature of ER=EPR that observers from opposite sides of an entangled pair of systems may meet in the connected interior. Since in \cite{Maldacena2013} any such meeting is trapped behind the horizon, it is not obvious how its occurrence could be confirmed by exterior or CFT measurements. What we found is that if, after the observers jump into their respective black holes, a boundary-boundary coupling is activated, then the Einstein-Rosen can be rendered traversable, and the meeting inside may be seen from the boundary. This seems to suggest that the ER=EPR wormhole connection was physically ``real''. But since all measurements in the CFT description are governed by the rules of linear quantum mechanics, it seems like any explicit operational verification of the existence of the wormhole would also correspond to a linear quantum measurement. It might be interesting to check the compatibility of these ideas with the linearity of measurements made behind the horizon, discussed in \cite{marolf2012eternal}.
What is the quantum information theory interpretation of such a traversable wormhole? A curious feature of the transmission of a qubit, $Q$, through the wormhole is that it appears to be sent ``via the entanglement'', rather than directly by the inter-boundary coupling. (Note that the traversable portion of the wormhole is close to the bifurcation point, which describes the subspaces of the left and right Hilbert spaces that are the most entangled in the thermofield double state.) There are several ways to see that the quantum information of $Q$ is not simply being sent directly through the boundaries. First, the commutator of $Q$ (for example when it is first injected into the interior from the left boundary) with the interaction Hamiltonian is extremely small near the thermofield double state. Furthermore, at the time the interaction is activated, $Q$ is in fact spacelike separated from the boundary in the bulk picture, so in the bulk approximation $Q$ and ${\cal O}$ are independent quantum variables. From the CFT perspective, this is because $Q$ has thermalized into the left system before the ${\cal O}_L {\cal O}_R$ interaction is turned on, so no quantum information about $Q$ appears to be accessible to the operator ${\cal O}$. Of course, the boundary coupling is nevertheless crucial for the existence of the traversable wormhole.
This situation is somewhat analogous to what occurs in quantum teleportation. Entanglement alone cannot be used to transmit information, and no qubit, $Q$, from the left can traverse the bridge to the right if the left and right systems are dynamically decoupled. However, if additional classical information is sent from the left to the right, a qubit can be transmitted - this is referred to as quantum teleportation.
Suppose Alice and Bob share a maximally entangled pair of qubits, $A$ and $B$. Alice can then transmit the qubit $Q$ to Bob by sending only the classical output of a measurement on the $Q$-$A$ system. Depending on which of the 4 possible results are obtained, Bob will perform a given unitary operation on the qubit $B$, which is guaranteed to turn it into the state $Q$.
Unlike the usual description of quantum teleportation, in our example it is essential that the channel between the left CFT, $A$, and the right CFT, $B$, is a quantum one. For example, if one projected onto eigenstates of the operator ${\cal O}_L$, then the configuration would simply look like a particular quantum state (the projection of $|\textrm{tfd}\rangle$) evolving under the decoupled Hamiltonians together with an action by a purely right unitary, which can never lead to a traversable wormhole. This makes sense, because in the standard description of quantum teleportation, the measurement performed by Alice is a projection onto an eigenstate, which instantly results in the pattern of $Q$ being contained in the system $B$. This would not be described by a physical motion through the wormhole in the bulk. Teleportation in this sense has been discussed in the dual gravity language by \cite{susskind2016er, numasawa2016epr, marolf2012eternal}.
However, in the exact, fully quantum description of the quantum teleportation protocol, there is a particular dynamical process given by the unitary evolution $V = \sum_i P^{Q A}_i U^B_i$ that governs the transmission of the ``classical'' information and the subsequent appropriate transformation of a qubit in the $B$ quantum system. Here $P^{Q A}_i$ are a complete mutually exclusive set of projectors on the $Q$-$A$ system that describe Alice's measurement, and $U^B_i$ is the unitary transformation performed by Bob given the data $i$. The classical information transmitted from Alice to Bob was encoded by the index $i$.
Treating $V$ as a time dependent interaction Hamiltonian can result in negative ANE along the horizon if the original entanglement between $A$ and $B$ was well described by a large Einstein-Rosen bridge, which will render the wormhole traversable. This is a description in which the time scales and processes of decoherence and measurement by Alice are resolved, and treated as physical dynamical evolution. In such a ``microscopic'' description of quantum teleportation, the qubit $Q$ must physically evolve from the left to the right. Of course in the limit that Alice's measurement is essentially instantaneous and classical, the traversable window will be very small (and not well described by a semiclassical spacetime) - just enough to let the single qubit $Q$ pass through.
Therefore, we propose that the gravitational dual description of quantum teleportation understood as a dynamical process is that the qubit passes through the ER=EPR wormhole of the entangled pair, $A$ and $B$, which has been rendered traversable by the required interaction.
Another possible interpretation of our result is to relate it to the recovery of information described in \cite{PreskillHayden}. Assuming that black hole evaporation is unitary, it is in principle possible to eventually recover a qubit which falls into a black hole, from a quantum computation acting on the Hawking radiation. Assuming that you have access to an auxiliary system maximally entangled with the black hole, and that the black hole is an efficient scrambler of information, it turns out that you only need a small (order unity) additional quantity of Hawking radiation to reconstruct the qubit. In our system, the qubit may be identified with the system that falls into the black hole from the left and gets scrambled, the auxiliary entangled system is the CFT on the right, and the boundary interaction somehow triggers the appropriate quantum computation to make the qubit reappear again, after a time of order the scrambling time $R \ln (R / L_\text{planck})$.\footnote{We thank Juan Maldacena for suggesting this interpretation.}
\section*{Acknowledgements}
We thank Ofer Aharony, Daniel Harlow, Juan Maldacena, Sudipta Sarkar, Douglas Stanford and Andy Strominger for helpful and stimulating discussions. DLJ and PG were supported in part by NSFCAREER grant PHY-1352084 and by a Sloan Fellowship. AW was supported by the Institute for Advanced Study, by the Martin A. and Helen Chooljian Membership Fund, and NSF grant PHY-1314311.
|
\section{Introduction}
Cosmic strings are linear defects that could be formed at a symmetry breaking phase transition in the early universe (for a review and references see \cite{Book}). Strings predicted in most grand unified models respond to external electromagnetic fields as thin superconducting wires \cite{Witten}. As they move through cosmic magnetic fields, such strings develop electric currents. Oscillating loops of superconducting string emit short bursts of highly beamed electromagnetic radiation \cite{VV87,Spergel87,Babul87} and high-energy particles \cite{Barr87,BOSV,Tanmay}.
The bursts originate at short string segments, called cusps, where the string briefly develops a large Lorentz factor. Even though particle emission from cusps has been extensively studied in the literature, it appears that some important aspects of this process have not been adequately described. The ejected particles and their decay products were treated individually, disregarding their interaction with one another. And the electromagnetic fields produced by the string were treated classically, disregarding the possibility of pair production. Here, we point out that in a cusp event a large energy is injected in a very small volume. This energy thermalizes, resulting in a relativistic fireball. The fireball then expands and cools, until it becomes optically thin, at which point the particles freely disperse. The fireball is boosted by the large Lorentz factor $\gamma$ of the cusp, so it appears to an observer as a narrow jet with an opening angle $\sim \gamma^{-1}$. This is similar to earlier treatments, but the particle composition of the jet is now very different. The predicted observational effects of superconducting strings should therefore be reconsidered.
We begin in the next section with a brief review of cosmic string properties and evolution. Then we describe the cusp formation and emission from cusps (\S\ref{CoL}), thermalization of the cusp emission (\S\ref{FfC}), the fireball expansion (\S\ref{FE}), and finally the resulting observables (\S\ref{Obs}).
\section{String properties and evolution}
Numerical simulations of string evolution indicate that strings evolve in a self-similar manner. A Hubble-size volume at any time $t$ contains a few long strings stretching across the volume and a large number of closed loops of length $l\ll t$ (for an up to date review and references, see \cite{BPOS}). The loops oscillate periodically and lose their energy, mostly by gravitational radiation. For a loop of length $l$, the oscillation period is $l/2$ and the lifetime is
\begin{equation}
\tau_l \sim l/k_g G\mu.
\end{equation}
Here, $k_g\sim 50$ is a numerical coefficient, $G$ is Newton's constant, and $\mu$ is the string mass per unit length, which is related to the symmetry breaking energy scale $\eta$ as
\begin{equation}
\mu \sim \eta^2.
\end{equation}
Gravitational waves emitted by loops over the cosmic history add up to a stochastic gravitational wave background with a wide spectrum of frequencies. Requiring that the predicted amplitude of this background is not in conflict with the millisecond pulsar observations, one can impose an upper bound on the string mass parameter $\mu$: $G\mu \lesssim 3\times 10^{-9}$ \cite{BPOS}. The corresponding bound on $\eta$ is
\begin{equation}
\eta_{14}\lesssim 5,
\end{equation}
where $\eta_{14} = \eta/10^{14} GeV$.
The most interesting for our discussion here are the most numerous loops surviving at the preset time $t_0$. These loops were formed in the radiation era and have lifetimes $\tau_l \sim t_0$. Their length is
\begin{equation}
l\sim 50 G\mu t_0 \sim 10^{20}{\rm cm}~ \eta_{14}^2
\end{equation}
and their number density is
\begin{equation}
n\sim 10^{-5} (G\mu)^{-3/2} t_0^{-3}.
\end{equation}
There are
\begin{equation}
N \sim 10^{10}\eta_{14}^{-3}
\end{equation}
such loops within the present horizon.
\section{Cusps on Loops}\label{CoL}
At the time of their formation, cosmic string loops consist of nearly straight segments separated by sharp kinks \cite{BPO16}. Such loops do not develop cusps. However, the kinks are gradually smoothed by the gravitational back-reaction and all but disappear towards the end of the loop's life. The most numerous loops, decaying at the present epoch, are expected to be very smooth and to
develop cusps during each oscillation period \cite{BPO16}. Cusp events occur on such loops every
\begin{equation}
t_c \sim 3\times 10^{9}{\rm s}~ \eta_{14}^2.
\end{equation}
A string loop of length $l$ oscillating in a magnetic field $B$ acquires an electric current
\begin{equation}
J_0 \sim 0.1 e^2 Bl.
\end{equation}
The current is strongly enhanced at cusp events, resulting in powerful bursts of EM radiation. In the vicinity of a cusp, a string segment of initial length $l/\gamma$ shrinks by a factor of $\gamma$, so its length becomes
\begin{equation}
l_\gamma \sim \gamma ^{-2}l,
\end{equation}
and develops a Lorentz factor $\gamma$. In the local rest frame of the string, the current is enhanced by the same factor, and an EM pulse of energy
\begin{equation} \label{Ec}
E_c \sim 2\times 10^{37}{\rm erg}~ B_{-9}^2\eta_{14}^6
\end{equation}
is released into a region of characteristic size $l_\gamma$ during the time $\sim l_\gamma$.\footnote{Eq.~(\ref{Ec}) follows from the angular distribution of the EM energy in the lab system derived in \cite{VV87,Spergel87}. The energy emitted within angle $\theta$ of the string velocity at the cusp is ${E}_\theta \sim 10 J_0^2 l /\theta$. It comes from a segment with $\gamma\sim 1/\theta$. Transforming to the frame of the segment, we have $E_c\sim 10 J_0^2 l$, independent of $\gamma$, which gives Eq.~(\ref{Ec}).}
The Lorentz factor $\gamma$ increases towards the central point of the cusp, terminating at
\begin{equation}\label{Gac}
\gamma _c \sim 2\times 10^{10}~ B_{-9}^{-1}\eta_{14}^{-1}.
\end{equation}
This limiting value of $\gamma$ is due to the inertia of the charge carriers on the string \cite{BPOV,BPO}.
Charged particles moving in opposite directions along the string can scatter off one another and escape from the string \cite{Barr87}. This process can be very efficient in the cusp region, where the particle density is very high. However, currents flowing in opposite directions are not expected to be balanced, so ejection of particles does not change the order of magnitude of the current, and the electromagnetic burst continues to develop with the Lorentz factor growing up to $\gamma_c$.
The ejected particles scatter off one another and pair-create in the
strong electromagnetic field near the cusp.\footnote{Interactions of
the ejected particles with the EM field of the string have been
discussed in Ref.~\cite{Rubinstein}.} They quickly thermalize and form a relativistic fireball. This alternative mechanism of fireball formation is sensitive to the string microphysics. For some parameter values the energy carried away by the ejected particles can be comparable to the electromagnetic energy \cite{BOSV}, but it can also be much smaller. Here we shall focus on fireballs of EM origin.
The proper energy Eq.(\ref{Ec}) is negligibly small for an object at a cosmological distance, but the Lorentz boost Eq.(\ref{Gac}) is very large. In the laboratory frame (the CMB frame): (i) the energy of the EM pulse is Lorentz boosted, (ii) its length scale is Lorentz contracted, (iii) the pulse is beamed into an opening angle $\sim \gamma ^{-1}$. Altogether, this gives an equivalent isotropic luminosity
\begin{equation} \label{iso}
L_{\rm iso} \sim 4\times 10^{89}\frac{\rm erg}{\rm s}~ B_{-9}^{-4}\eta_{14}^{-2}\left( \frac{\gamma}{\gamma _c} \right)^6
\end{equation}
-- unheard of in astrophysics. However, as the beaming angle is small, the probability to get illuminated by the most energetic part of cusp emission, with $\gamma \sim \gamma _c$, is negligibly small.
Cusp events might become observable in two ways. Emission into larger beaming angles does occur at smaller $\gamma$. Also, emission initially beamed into a small solid angle can spread out by propagation through the intergalactic medium (IGM). Both of these possibilities are considered in the following sections. Here we will show that astrophysically interesting fluences and fluxes can arise in both scenarios, so that further investigation might be worthwhile.
First consider cusp events that illuminate a given observer once per time $t=t_y$ years. This requires beaming angles $\theta \sim \gamma ^{-1}$ given by
\begin{equation}
t \sim \frac {t_c}{N}\theta ^{-2}.
\end{equation}
Knowing $\gamma$, we estimate the fluence (for an event at a cosmological distance $r\sim 3$ Gpc), $f\sim \gamma ^3 E_c/r^2$, or
\begin{equation} \label{flue}
f \sim 2\times 10^{-7}\frac{\rm erg}{\rm cm ^2}~ B_{-9}^{2}\eta_{14}^{-3/2}t_y^{3/2}.
\end{equation}
Another possibility is that a fair fraction of the cusp energy, initially narrowly beamed, spreads out without great losses, where "losses" mean photons of energy $\lesssim 1$GeV which would be unobservable on the diffuse gamma-ray background. Then the figure of merit is the time averaged flux, $F\sim \frac{N\gamma _cE_c}{t_cr^2}$, giving
\begin{equation} \label{flux}
F \sim 10^{-8}\frac{\rm erg}{\rm cm ^2\cdot s}~ B_{-9}.
\end{equation}
Intergalactic magnetic field is likely inhomogeneous, with different magnetic field values $\sim B$ occupying different volumetric fractions of the universe $f_B$. Then the fluence Eq.(\ref{flue}) is multiplied by $f_B^{3/2}$, the flux Eq.(\ref{flux}) is multiplied by $f_B$, and the results are summed over all $B$ values.
Throughout this paper we assume that cosmic string loops lose their energy mostly by gravitational radiation. This holds for $\eta_{14}\gtrsim 10^{-4}f_B B_{-9}$, while for smaller values of $\eta$ the EM radiation is the dominant energy loss mechanism. In the latter case, the lifetime of the loops is independent of their size and all loops formed in the radiation era decay before the present epoch \cite{BOSV}.
We have only considered the cusp events at redshifts $z\lesssim 1$. The analysis can be extended to include higher redshifts \cite{BHV,BOSV}. One finds that the wait time for cusp events at $z\gtrsim 1$ with a given energy fluence $f = 10^{-8} f_{-8} ~erg/cm^2$ is
\begin{equation}
t_y \sim 10^{-2} \eta_{14} f_{-8}^{2/3}f_B^{-1}(z)B_{-9}^{-4/3}(z)(1+z)^{7/4}(\sqrt{1+z}-1)^{-2/3}.
\end{equation}
This shows that for a uniform primordial magnetic field with $f_B(z) = 1$ and $B(z)\propto (1+z)^2$, the contribution of high redshift events could be significant. On the other hand, if the intergalactic fields originated during and after the galaxy formation epoch, most of the detectable cusp events are likely to occur at $z\sim 1$.
We see that, unlike $L_{\rm iso}$, both the fluence $f$ and the flux $F$ have reasonable values. They are large enough to be detectable, but not too large to be immediately ruled out.
\section{Fireballs from Cusps}\label{FfC}
Observability of fluences and fluxes Eqs.(\ref{flue},\ref{flux}) depends on which particle (photon or neutrino) and at what energy reaches Earth. We must then describe the evolution of the emitted EM pulse. Here we show that for large enough $\gamma$, the EM pulse thermalizes, producing a fireball of initial temperature $T_f\gtrsim 0.1$ MeV. This allows to predict what gets injected into the ISM and then also what reaches Earth.
The idea that large energy, quickly released into a small volume, thermalizes is nothing new in astrophysics. In the context of GRBs (gamma ray bursts) it has been discussed by \cite{Pac,Goo}, and has observational confirmations \cite{Gui}.
Assuming the thermalization does occur, we can estimate the fireball temperature from $E_c\sim l_\gamma ^3T_f^4$, giving
\begin{equation} \label{Tf}
T_f \sim 1{\rm GeV}~ B_{-9}^{-1}\eta_{14}^{-3/2}\left( \frac{\gamma}{\gamma _c} \right)^{3/2}.
\end{equation}
The cusp event really produces an inhomogeneous fireball which is hotter and moves faster (with larger Lorentz factors) on the inside than on the outside. But for the sake of simplicity, we will just speak of separate fireballs with different characteristic Lorentz factors $\gamma$.
Obviously, a necessary condition of thermalization is that the fireball be optically thick. Similar to \cite{Pac,Goo}, this requires $T_f\gtrsim 0.1$MeV, giving
\begin{equation}
\frac{\gamma}{\gamma _c}\gtrsim 2\times 10^{-3}~ B_{-9}^{2/3}\eta_{14}.
\end{equation}
One can argue that $T_f\gtrsim 0.1$MeV is also a sufficient condition of thermalization. This is because in the near zone of the EM pulse (at distances $\lesssim l_\gamma$ in the fireball rest frame) the electric and magnetic fields generated by the string, $E\sim B\sim T_f^2$ are generally not perpendicular to each other, with both field invariants of the same order. Then, for super-Schwinger fields, that is for $T_f\gtrsim e^{-1/2}m_e\sim 2$ MeV, copious pair production does make the fireball optically thick.
At somewhat lower effective temperatures $T_f$, Schwinger effect does not operate, but the tree-level QED processes are still expected to produce an avalanche of pairs, similar to what happens in pulsars \cite{Rud}: an accelerated electron emits gamma rays which produce pairs on the background field, pairs are then accelerated, etc... Quasi-thermalization is already seen in Crab-like pulsars, even though the equivalent temperature, as measured at the light cylinder, is only about 1keV.
A pulsar-like avalanche requires that curvature radiation photons be energetic enough to pair produce on the large-scale EM field. Consider then, closely following \cite{Rud}, a seed electron in a generic EM field $E\sim B\sim T_f^2$ with characteristic length scale $r\sim l_\gamma$. The electron quickly synchrotron cools, but as there is generically an electric field component along the magnetic field, the electron will keep moving relativistically along a curved magnetic field line, with curvature radius $\sim r$. The terminal Lorentz factor of the electron, $\gamma _e$, is determined by the balance between the electric field acceleration and the curvature radiation damping:
\begin{equation}
eE\sim \frac{e^2\gamma_e ^4}{r^2}.
\end{equation}
The electron, as it moves with Lorentz factor $\gamma_e$, emits curvature photons of energy
\begin{equation}
\epsilon \sim \frac{\gamma_e^3}{r}.
\end{equation}
These photons pair produce on the background field $B$, provided
\begin{equation}
\epsilon B\gtrsim 0.1\frac{m_e^3}{e},
\end{equation}
giving the avalanche criterion
\begin{equation}
T_f\gtrsim 1{\rm keV}~B_{-9}^{-2/17}\eta_{14}^{-6/17},
\end{equation}
which is always satisfied by a large margin if we use the proposed thermalization requirement $T_f\gtrsim 0.1$MeV. We note also that particle ejected from the string can provide an additional mechanism for thermalizing fireballs with sub-Schwinger fields.
For effective temperatures somewhat smaller than 0.1 MeV, thermalization does not occur, as the plasma, if thermalized, would have become optically thin. Still a very rough, but physically well motivated, estimate of the emission is possible in some cases. This happens when the string parameter $\eta$, the ISM field $B$, and the fireball parameter $\gamma$ are such that the resulting EM field amplitude and length scale emulate the corresponding pulsar parameters as measured at the light cylinder. Assuming, as appears likely, that pulsar emission is mostly produced at the light cylinder, one can then expect that string loop cusps will produce similar pulsar-like emission. We will not pursue this anymore in this paper, limiting ourselves only to the "cleaner" cases with $T_f\gtrsim 0.1$MeV.
For even smaller $\gamma$ parameters, that is for larger beaming angles, a particle avalanche is not expected to develop. A solid angle $\sim \gamma ^{-2}$ is then illuminated by a pure-EM pulse. The EM pulse has very high energy density; it sweeps up the ISM plasma, and some conversion of the pulse energy into high-energy photons (synchrotron and inverse Compton) must occur. It has even been suggested that the lower-$\gamma$ EM pulses from strings are responsible for GRBs \cite{BHV}. We will not consider the low-$\gamma$ cases here because we do not have a full picture of the EM pulse interaction with the ISM. (Exactly for this reason, \cite{BHV} do not predict the photon energy in their GRB-from-strings scenario.)
\section{The Fireball Evolution}\label{FE}
In what follows we assume that fireballs with effective temperatures $T_f\gtrsim 0.1$MeV do thermalize at about the temperature $T_f$. A distant observer will then see photons of characteristic energy $\gamma T_f$ (and neutrinos with nearly the same characteristic temperature, if $T_f$ is high enough to produce thermalized neutrinos). Here the $\gamma$ factor simply accounts for the fireball frame motion with respect to the observer.
Why $T_f$ gives the photon energy in the fireball rest frame, and why only the massless particles are released has been explained in \cite{Pac,Goo}. As the fireball expands, the plasma temperature, of course, drops. At the same time, the plasma is radially accelerated to high Lorentz factors, and in exactly the same proportion as the temperature drop, so that the characteristic energy of a massless particle remains unchanged. The fireball becomes a spherical shell with a growing radius, but with the shell thickness close to the initial fireball radius. The photon distribution function evolves as if photons were not constantly scattered, produced and annihilated, although the photon-electron interactions truly terminate only when the temperature drops to about $0.1$MeV. At this temperature, the pair density is negligibly small, explaining why the fireball energy goes almost solely into photons (and possibly neutrinos), with only a negligible fraction given to leftover electrons/positrons and protons/antiprotons.
We have described fireball expansion into vacuum, while it actually expands into an ISM wind, blowing with Lorentz factor $\gamma$ in the fireball frame. One can check, however, that the swept-up protons are strongly energetically subdominant at the photon release temperature (negligible baryon loading, in the GRB terminology). This is intuitively clear, because the fireball proper size is initially $\lesssim 1$cm.
\section{Observables} \label{Obs}
Three observationally promising scenarios will be considered in turn:
(i) The entire energy is released into the ISM as photons of energy $\sim \gamma T_f$. Some of these photons reach Earth.
(ii) An order unity fraction of the fireball energy is converted into neutrinos of energy $\sim \gamma T_f$. These neutrinos freely propagate to Earth.
(iii) Almost all of the initial $\sim \gamma T_f$ photons are converted into cascade photons, producing an observable diffuse background.
\subsection{TeV Bursts}
By Eq.(\ref{Tf}), the fireball thermalizes (meaning, by our assumption, that the fireball has effective temperatures $T_f\gtrsim 0.1$MeV) provided
\begin{equation}
\gamma \gtrsim \gamma _B\equiv 4\times 10^7~B_{-9}^{-1/3}.
\end{equation}
This will inject photons of energy
\begin{equation}\label{ep}
\epsilon \sim 4{\rm TeV}~B_{-9}^{-1/3}\left( \frac{\gamma}{\gamma _B} \right)^{5/2}.
\end{equation}
For photon energy $\epsilon\sim $few TeV, the photon mean free path (in the infrared background) can be crudely approximated \cite{Dom} by
\begin{equation}
\lambda \sim 1{\rm Gpc}~\left( \frac{\epsilon}{1 {\rm TeV}} \right)^{-1}.
\end{equation}
The characteristic wait time for a burst of such photons is
\begin{equation}
t\sim \frac{\gamma ^2t_c}{N}\left( \frac{3{\rm Gpc}}{\lambda}\right)^3,
\end{equation}
or
\begin{equation}\label{tp}
t\sim 3\times 10^{10}{\rm y}~B_{-9}^{-5/3}\eta_{14}^{5}\left( \frac{\gamma}{\gamma _B} \right)^{19/2}.
\end{equation}
The fluence (here defined as the number of photons per unit area) is
\begin{equation}
f\sim \frac{\gamma ^3E_c}{\epsilon \lambda ^2},
\end{equation}
or
\begin{equation}\label{fp}
f\sim 5\times 10^{9}\frac{1}{{\rm m}^2}~B_{-9}^{2/3}\eta_{14}^{6}\left( \frac{\gamma}{\gamma _B} \right)^{11/2}.
\end{equation}
To illustrate, take $B\sim 1$nG and $\eta \sim 10^{12}$GeV. Consider only the barely thermalized fireballs, with $\gamma \sim \gamma _B$. Then the photon energy, Eq.(\ref{ep}), the wait time, Eq.(\ref{tp}), and the fluence, Eq.(\ref{fp}), are
\begin{equation}
\epsilon \sim 4{\rm TeV},~~~~t\sim 3{\rm y},~~~~f\sim 0.005\frac{1}{{\rm m}^2}.
\end{equation}
Effective area of an all-sky TeV detector HAWC reaches $10^5{\rm m}^2$ at these energies. One can then get a (never seen before) many-photon event every few years.
A slight modification of this scenario is also worth mentioning. Individual sources, like the giant lobes of the nearby radio galaxy Cen A, can become visible. A sufficiently high event rate is achievable for small enough symmetry breaking scale $\eta$. To illustrate, consider a constantly monitored object of size $r$ ($\sim 1$ Mpc) at a distance $d$ ($\sim 3$ Mpc) with magnetic field $B$ ($\sim 1\mu$G). If $\eta _{14}<<1$, there are many loops within a sphere of radius $r$, and an estimate along the same lines as above gives the event wait time, photon energy, and fluence
\begin{equation}
\epsilon \sim 400{\rm GeV},~~~~t\sim 5{\rm y}~\eta_{-11}^5,~~~~f\sim 0.3\frac{1}{{\rm m}^2}~\eta_{-11}^6.
\end{equation}
Effective area of HAWC at about 300 GeV is about $10^3{\rm m}^2$, but Cen A is too far south for HAWC. Effective area of HESS at about 300 GeV is about $10^5{\rm m}^2$, but HESS has a small field of view and observed Cen A for only about 100 hours.
\subsection{Ultra-High-Energy Neutrino Bursts}
Neutrinos are produced and thermalized if the fireball is optically thick to electron-positron annihilation into neutrinos,
\begin{equation}
\sigma nr\sim (G_F^2T_f^2)T_f^3l_\gamma\gtrsim 1,
\end{equation}
or
\begin{equation}
\gamma \gtrsim \gamma _\nu\equiv 5\times 10^{9}~B_{-9}^{-5/11}\eta_{14}^{-4/11}, ~~~B_{-9}\lesssim 10~\eta_{14}^{-7/6},
\end{equation}
where the last inequality ensures that $\gamma _\nu$ does not exceed the maximal possible $\gamma =\gamma _c$.
Bursts of neutrinos of energy
\begin{equation}
\epsilon_\nu \sim 6\times 10^{17}{\rm eV}~B_{-9}^{-7/11}\eta_{14}^{-10/11}\left( \frac{\gamma}{\gamma _\nu} \right)^{5/2},
\end{equation}
are observed with the wait time
\begin{equation}
t_\nu \sim 3\times 10^{11}{\rm y}~B_{-9}^{-10/11}\eta_{14}^{47/11}\left( \frac{\gamma}{\gamma _\nu} \right)^{2},
\end{equation}
and fluences
\begin{equation}
f_\nu\sim 3\times 10^{14}\frac{1}{{\rm km}^2}~B_{-9}^{14/11}\eta_{14}^{64/11}\left( \frac{\gamma}{\gamma _\nu} \right)^{1/2}.
\end{equation}
To illustrate, take $B\sim 1$nG and $\eta \sim 4\times 10^{11}$GeV. Consider only the barely $\nu$-thermalized fireballs, with $\gamma \sim \gamma _\nu$. Then the photon energy, the wait time, and the fluence are
\begin{equation}
\epsilon \sim 10^{20}{\rm eV},~~~~t\sim 15{\rm y},~~~~f\sim 3\frac{1}{{\rm km}^2},
\end{equation}
giving a multi-neutrino event in a detector with $\sim {\rm km}^2$ effective area.
\subsection{Diffuse gamma-ray background}
Most energy (for a CMB-frame observer) is coming from the most compact part of the fireball, with $\gamma \sim \gamma _c$. These fireballs produce very narrowly beamed bursts of photons with characteristic energy
\begin{equation}\label{ep0}
\epsilon_0 \sim 2\times 10^{19}{\rm eV}~B_{-9}^{-2}\eta_{14}^{-5/2}.
\end{equation}
In about 10 Mpc, these ultra-high-energy photons produce an electron-positron pair on the CMB (or other radio background) \cite{Set}.
The electron and the positron have the same characteristic energy $\sim \epsilon_0$. These particles inverse-Compton scatter, pair produce, and emit synchrotron radiation. The synchrotron dominates roughly for
\begin{equation}
B_{-9}\gtrsim \left( \frac{\epsilon_0}{10^{19} {\rm eV}} \right)^{-1/2},
\end{equation}
which gives, using Eq.(\ref{ep0}), a $B$-independent criterion for the synchrotron domination
\begin{equation}
\eta _{14}\lesssim 1.
\end{equation}
If the synchrotron losses are the dominant ones, most initial energy gets converted into second generation photons, with characteristic energy
\begin{equation}
\epsilon \sim 4{\rm GeV}~B_{-9}^{-3}\eta_{14}^{-5}.
\end{equation}
For $\epsilon \lesssim 100$GeV, the second generation photons freely propagate to Earth. What we see at Earth is not a burst, but a diffuse background. This is because synchrotron-emitting electrons and positrons were also deflected by the magnetic field. This spreads the cascade in angle and leads to large time delays. The diffuse gamma-ray flux is
\begin{equation}
F \sim 10^{-8}\frac{\rm erg}{\rm cm ^2\cdot s}~ B_{-9}.
\end{equation}
To illustrate, take $B\sim 1$nG and $\eta \sim 10^{14}$GeV. Then the photon energy and flux are
\begin{equation}
\epsilon \sim 4{\rm GeV},~~~F \sim 10^{-8}\frac{\rm erg}{\rm cm ^2\cdot s}.
\end{equation}
This is in very strong conflict with observations. Not only is the actual background at 4GeV smaller ($\sim 3\times 10^{-10}\frac{\rm erg}{\rm cm ^2\cdot s\cdot sr}$, \cite{Fer}), almost all of the observed background seems to have a mundane (astrophysical) explanation \cite{Hoo}.
\section{Conclusions}
We have argued that the energy discharge from cusps of superconducting strings occurs in the form of thermalized relativistic fireballs. The large Lorentz factor $\gamma$ of the cusp boosts the energies of thermal particles and collimates them in a narrow beam with opening angle $\gamma^{-1}$. As the fireball expands and cools, practically all massive particle-antiparticle pairs annihilate, so only photons and neutrinos can reach the observer.
The predicted neutrino and photon fluxes are not sensitive to string microphysics and depend only on the string energy scale $\eta$ and on the magnitude $B$ and the volume filling factor $f_B$ of the cosmic magnetic fields. Since the magnitude and the distribution of intergalactic magnetic fields are highly uncertain, we could not derive any firm bounds on the string parameter $\eta$. However, with plausible assumptions about $B$ and $f_B$, we showed that strings with $\eta \sim 10^{10}-10^{15} ~GeV$ can produce observable effects with unique signatures. These include TeV photon bursts (coincident multiple hits of the detector area by TeV photons), UHE neutrino bursts, and TeV photon bursts from individual astrophysical sources.\footnote{The possibility of such effects was first indicated in Ref.\cite{BOSV}, but their predicted rate and fluxes are different in the fireball scenario.} Both neutrino and photon bursts from cusps can be accompanied by potentially detectable bursts of gravitational radiation \cite{BHV}. The diffuse gamma-ray background from strings does not have unique signatures but provides stringent combined constraints on the symmetry breaking scale and the ISM magnetic field.
\section*{Acknowledgements}
A.V. is grateful to Venya Berezinsky and to Ken Olum for useful discussions and to Gregory Gabadadze for his hospitality at New York University, where this work was initiated. This work was supported in part by the National Science Foundation under grant PHY-1518742 (A.V.).
|
\section{Introduction}
Phase transitions are an interesting trait of physical systems. In particular, when dealing with quantum many-body set-ups it is curious that a well defined microscopic description can lead to non-trivial and singular behaviours in the thermodynamic limit. The study of the equilibrium properties of such systems is well established in, for example, exactly solvable one-dimensional spin chains~\cite{Sachdev}. With the help of several tools from quantum information, notably quantum correlations, the implications of the presence of quantum phase transitions (QPTs) have been explored~\cite{CorrelationsQPT}.
Beyond the static properties, there is a growing interest in studying the dynamics of many-body quantum systems. This is further catalysed by the increasing interest in understanding the thermodynamic properties of genuinely quantum systems~\cite{GooldJPA}. Despite the significantly more involved nature of studying dynamics, remarkable progress has been made in elucidating the behaviour of important quantities, e.g. (irreversible) work, entropy production, and residual energy, when a many-body system is evolved through its critical point. The most frequent evolution considered is that of a ``sudden quench" of the order parameter~\cite{EchoQuenches,SilvaPRL,DornerPRL,LorenzoPRX,GabrielePRB,DasPRB} (however finite time protocols have also been addressed~\cite{SaroPRB}). Focusing on such a sudden change allows us to capture the salient features of the ensuing non-equilibrium dynamics, while leaving the study of more qualitative differences to a more involved temporal analysis. Typically, the sudden change to the Hamiltonian kicks the system out of equilibrium and can lead to interesting consequences. Notably, for the Ising model the dynamics of a sudden quench have been explored and it has been shown that the irreversible entropy production provides signals of the presence of the equilibrium QPT~\cite{SilvaPRL}, and also can be used to explain emergent phenomena such as the vanishing gap between ground and first excited energy levels in the thermodynamic limit~\cite{DornerPRL}. Recently, the irreversible work was shown to faithfully capture the critical features even for so-called impurity QPTs~\cite{GabrielePRB}.
In this work we add to this endeavour by studying the Lipkin-Meshkiv-Glick model~\cite{LMG65}. The model has attracted substantial interest as it serves as the paradigmatic example of an infinite range interacting system. It can be solved in the thermodynamic limit and exhibits a complex phase diagram~\cite{VidalPRLPRE,CastanosPRB}. We will be interested in exploring how clear signatures of the equilibrium QPT is manifest in the dynamics when the model is quenched through its critical point. We remark that the evolution of this model through its QPT have been studied previously in Ref.~\cite{DasPRB} wherein the equal time-order parameter correlation function was examined and Ref.~\cite{SaroPRB} where the adiabatic dynamics were explored. Our study is set apart from these as it seeks to establish a rigorous link between the thermodynamic quantities such as work and free energy, with the presence of the known QPT. By exploiting the time-dependent fidelity we show that the dynamics are manifestly different when the quench is restricted to a particular phase compared to when the system is quenched through the critical point. More interestingly, we explicitly show that while the average work performed on the system due to the quench is blind to the QPT, the free energy, and therefore the irreversible work, appears acutely sensitive, showing a markedly different rates of change. Finally, we use the spectral function to further understand the fundamental excitations governing the dynamics of the system.
The remainder of the paper is organised as follows. In Sec.~\ref{model} we present the model and introduce the quantities that will be of interest to our analysis. In Sec.~\ref{results} we explore these quantities for quenching the system across the critical point and show that the time-dependent fidelity exhibits interesting features that reveal the critical nature of the system as well as showing that the irreversible work neatly reveals the QPT. Sec.~\ref{spectral} assess the spectral function for various quenches. Finally, Sec.~\ref{conclusions} we present our conclusions and some discussions on our results.
\section{The Model and Figures of Merit}
\label{model}
We consider the ferromagnetic spin-1/2 Lipkin-Meshkov-Glick (LMG) model in a transverse field,
\begin{equation}
\mathcal{H} = -\frac{1}{N}\left( \sum_{i<j} \sigma_x^i \otimes \sigma_x^j+\gamma \sigma_y^i \otimes \sigma_y^j \right) -
h \sum_{i} \sigma_z^i,
\end{equation}
with $\sigma_{x,y,z}$ the Pauli spin-operators, $h$ the magnetic field strength, and $\gamma$ the anisotropy parameter (which we set to zero for simplicity in our simulations, however we remark that qualitatively similar results can be obtained for any $0\leq\gamma<1$, cfr. the appendix). By considering the collective spin operators $S_\alpha=\sum_i \sigma_{\alpha}^i/2$ with $\alpha=\{x,y,z\}$, up to a constant energy shift, the model can be written as
\begin{equation}
\label{collspinLMG}
\mathcal{H} =-\frac{2}{N}\left( S_x^2 + \gamma S_y^2 \right) - 2h S_z.
\end{equation}
In the following we work in the basis of maximum angular momentum (which is a constant of motion) and using the eigenstates of $S_z$ we can diagonalize Eq.~\eqref{collspinLMG} to find the complete spectrum (see e.g. Refs~\cite{VidalPRLPRE,CastanosPRB} for further details). In Fig.~\ref{fig1} we show the energy difference between the 5 lowest excited states and the ground state against $h$. We see when $h>1$ each energy level is distinct. As $h$ is decreased the gap between the first excited state and the ground state closes, and similarly the energy gap between subsequent pairs of excited states also closes. However, an important remark, only when $N\to \infty$ does the gap vanish and all eigenstates become doubly degenerate. Hence, for any finite size there is a small difference between the ground and first excited states~\cite{SaroPRB}.
\begin{figure}[t]
\includegraphics[width=0.85\columnwidth]{figures/GroundStateEnergy2ndDerivative.pdf}
\caption{(Color online) Energy difference between the ground state and the first 5 excited states for $N=400$ plotted against magnetic field strength $h$. {\it Inset:} Second derivative of the ground state energy per site for $N=100$ (red) to $700$ (magenta) in steps of 100.}
\label{fig1}
\end{figure}
The inset in Fig.~\ref{fig1} shows the second derivative with respect to $h$ of the ground state energy (per site) for system sizes ranging from $N=100$ to $N=700$. We see the emergence of a discontinuity appearing, thus signalling the known second order QPT at $h=1$~\cite{VidalPRLPRE}. We are interested in studying the dynamics when the ground state of one phase is evolved using the propagator of another. In what follows we will assume the system is initialized in the ground state of Eq.~\eqref{collspinLMG} corresponding to $h=h_i$. At time $t=0$, we quench the field strength $h_i \to h_f$ and we evolve the initial state according to the new Hamiltonian $\mathcal{H}_f$, so that
\begin{equation}
\ket{\psi(t)} = e^{-i \mathcal{H}_f t} \ket{\psi(0)}.
\end{equation}
Using this we can readily evaluate the time dependent overlap
\begin{equation}
\mathcal{O} = \bra{\psi(0)} \psi(t) \big>.
\end{equation}
This quantity will be central to our analysis as it allows us to access several important quantities that indicate that signatures of the equilibrium QPT are clearly manifest in the system's evolution.
A particularly important quantity will be the time-dependent fidelity (TDF)
\begin{equation}
\mathcal{L} = \vert \mathcal{O} \vert^2,
\end{equation}
which quantifies how different the evolved state is compared to the initial one (we remark this quantity is sometimes referred to as the Loschmidt echo in the literature). The TDF has already proven to be a useful tool in studying critical dynamics~\cite{EchoQuenches}. Additionally, we can determine the average work due to the sudden quench~\cite{LorenzoPRX,CampbellMarchFogarty}
\begin{equation}
\big< W \big> = \sum_j (E_j^f -E_0^i) \big\vert \big<\psi_0^i \vert\psi_j^f\big> \big\vert^2
\end{equation}
where $E_j^f$ and $\vert\psi_j^f \big>$ are the $j$-th eigenenergy and eigenstate of the post-quench Hamiltonian, and $E_0^i$ and $\vert \psi_0^i\big>$ are ground state energy and ground state for the initial Hamiltonian. The sudden nature of the quench drives the system out of equilibrium, and thus introduces a degree of irreversibility of the process. We can quantitatively define the irreversible work as~\cite{GabrielePRB,DeffnerPRL,CampbellMarchFogarty}
\begin{equation}
\big< W_\text{irr} \big> = \big< W \big> - \Delta F
\end{equation}
where $\Delta F$ is the free energy difference. By considering closed dynamics, and since we assume our system begins in the ground state of the initial Hamiltonian, $\Delta F$ is simply given by the difference between the post- and pre-quench ground state energies, i.e. $\Delta F = E_0^f - E_0^i$.
\section{Quench Dynamics Across a Quantum Critical Point}
\label{results}
\subsection{From the Paramagnetic to the Ferromagnetic Phase}
We begin analysing the case of quenching from the paramagnetic phase, setting $h_i=1.5$. In this regime there is a significant energy difference between the ground state and the first excited state, cfr Fig.~\ref{fig1}. In Fig.~\ref{fig2} we show the TDF for $N=400$ and the thick black curve corresponds to $h_f=1$. The dashed curves above this are for quenches that evolve the initial state taking a value of $h_f$ that is still in the paramagnetic phase. We clearly see the regular oscillatory behaviour persists even when quenching close to the critical point. When we evolve the state using $h_f<1$ (lower dotted curves) we see the dynamics loses the clean periodic behaviour, and dynamically the TDF no longer reaches unity. Additionally there is a significant decrease in the values of $\mathcal{L}$, even in some cases reaching exactly 0 indicating that the evolved state is orthogonal to the initial state.
\begin{figure}[t]
\includegraphics[width=0.85\columnwidth]{figures/LargeToSmall_LE_N400.pdf}
\caption{(Color online) Time-dependent fidelity for a quench from $h_i=1.5$ to $h_f \in [0.6,1.4]$ with $N=400$. The dashed curves are for $h_f>1$ with $h_f=1.4$ (topmost, blue) and $h_f=1.2$ (purple). The lowest two dotted curves are for $h_f<1$ with $h_f=0.8$ (red) and $h_f=0.6$ (bottom-most orange). The solid black curve is for $h_f=1$.}
\label{fig2}
\end{figure}
\begin{figure}[t]
\includegraphics[width=0.95\columnwidth]{figures/LargeToSmall_LEmin.pdf}
\caption{(Color online) {\it Main Panel:} minimum value of TDF achieved for $t\in(0,10)$ quenching the field from $h_i=1.5$ to $h_f$. Each curve from left to right corresponds to an increasing size of $N=100$ (dotted, red), 200 (dash-dotted, orange), 300 (dashed, blue) and 400 (solid, black). {\it Upper left inset:} Zoomed in cross-section of main panel. {\it Lower right inset:} Finite size scaling for the first value of field at which the $\mathcal{L}_\text{min}=0$.}
\label{fig3}
\end{figure}
The precise value at which the TDF reaches zero is examined in Fig.~\ref{fig3}. We determine the minimum value of TDF observed, $\mathcal{L}_\text{min}$, within the same time window for Fig.~\ref{fig2} against $h_f$. Clearly, when the quench is small, e.g. $h_f\in(1.2,1.4)$ the minimum value of TDF is still quite large. As the strength of the quench is increased we find this minimum value decreases. Interestingly, $\mathcal{L}_\text{min}=0$ only when $h_f<1$. We see an oscillatory behaviour appearing, however the amplitude of the oscillations is decreasing as the system size is increased. Furthermore, the first value of $h_f$ where $\mathcal{L}_\text{min}=0$, denoted $h_0$, shifts closer to 1 as we increase the system size. Through a finite size scaling with a quadratic fit, the lower right inset shows that this accurately determines the critical point. Such a result is remarkable as it clearly indicates that the equilibrium QPT can be witnessed by the occurrence of dynamical orthogonality.
A final peculiarity appears in studying $\mathcal{L}_\text{min}$, for $N>200$ a kink appears close to the critical point as shown in the upper left inset of Fig.~\ref{fig3}. However, as the system size is increased, this feature appears to move further from the critical value.
\begin{figure}[t]
{\bf (a)}\\
\includegraphics[width=0.8\columnwidth]{figures/LargeToSmall_Work.pdf} \\
{\bf (b)}\\
\includegraphics[width=0.8\columnwidth]{figures/LargeToSmall_DeltaF.pdf}\\
{\bf (c)} \\
\includegraphics[width=0.8\columnwidth]{figures/LargeToSmall_IrrWork.pdf}
\caption{(Color online) {\bf (a)} Average Work {\bf (b)} Free energy {\bf (c)} Irreversible work for a quench from $h_i=1.5$ to $h_f$. The insets of panels {\bf (b)} and {\bf (c)} are the first derivative of the functions. In all panels $N=400$.}
\label{fig4}
\end{figure}
From Fig.~\ref{fig2} it is clear that for small quenches the system dynamically comes close to the initial state, however for larger quenches this is no longer the case, indicating that a degree of irreversibility has been introduced into the system~\cite{BatalhaoPRL}. In Fig.~\ref{fig4} {\bf (a)} we examine the average work done, $\big< W \big>$, against $h_f$. We find $\big< W \big>$ is linearly dependent on the value of $h_f$, the larger the quench $h_i \to h_f$, the more work is done. In panels {\bf (b)} and {\bf (c)} we show the free energy and average irreversible work, respectively. Again, the free energy increases as we increase the size of the quench. However in the inset we examine its rate of change, and we see for $h_f>1$ this rate is linear, and there is a sudden change near $h_f\sim1$. While it is not surprising that the free energy exhibits a non-trivial behavior as we go through the critical point in light of the fact that is defined in terms of the ground state energy, the fact that both $\big< W \big>$ and $\Delta F$ are of the same order of magnitude leads to a trade off between the two quantities. This has interesting consequences for the irreversibility of the process captured by $\big<W_\text{irr} \big>$. Panel {\bf (c)} shows that when the quench is small and confined to the same phase as the initial state, $\big< W_\text{irr} \big> = 0$, indicating that the process is fully reversible, as confirmed by the behavior of the TDF which achieves values of unity during the dynamics. For large quenches, when the system is evolved according to a Hamiltonian in the ferromagnetic phase the average irreversible work becomes non-zero, and the degree of irreversibility grows as the magnitude of the quench increases.
\begin{figure}[t]
\includegraphics[width=0.8\columnwidth]{figures/SmallToLarge_LE_N400.pdf}
\caption{(Color online) Time-dependent fidelity for a quench from $h_i=0.5$ to $h_f \in [0.6,1.4]$ with $N=400$. The dashed curves are for $h_f<1$ with $h_f=0.6$ (blue) and $h_f=0.8$ (purple). The dotted curves are for $h_f>1$ with $h_f=1.2$ (red) and $h_f=1.4$ (left-most, orange). The solid black curve is for $h_f=1$.}
\label{fig5}
\end{figure}
\subsection{From the Ferromagnetic to the Paramagnetic Phase}
We next consider the complementary case of beginning in the ferromagnetic phase, setting $h_i=0.5$, and quenching to increasingly larger values of $h_f$. This scenario is markedly different to that of the previous section as the ground state is now {\it nearly} degenerate, i.e. there is a exponentially (in $N$) vanishing gap between the ground state the the first excited state, cfr. Fig.~\ref{fig1}. Therefore, by quenching the field strength and kicking the system out-of-equilibrium, it quickly becomes excited and occupies higher order states. In Fig.~\ref{fig5} we see this effect clearly, contrary to the previous section, we see even for moderately small quenches ($h_f\gtrsim0.6$) the TDF reaches zero, indicating the evolved state is orthogonal. We further remark, the larger the system the smaller the quench required to achieve dynamical orthogonality, again this is a consequence of the fact that the energy gap between the ground and first excited states decreases with increasing $N$. Thus we cannot use the presence of orthogonality to witness signatures of the QPT in the dynamics. However, there is a clear qualitative difference appearing when the quench is near to or above the critical point. For quenches to $h_f\gtrsim 1.0$ the TDF evolves into fully orthogonal states for a period, before exhibiting short time revivals. The height of these peaks are steadily decreasing in the considered time window, and the width of the revivals broadens. This indicates a sizeable increase in the irreversibility of the process when quenches into the paramagnetic phase are considered.
We confirm this behavior in Fig.~\ref{fig6}. Panel {\bf (a)} shows that the work is a linear function of the magnitude of the quench. However, the free energy and the irreversibility show the same qualitative behavior as shown in the previous section. Focussing on the average irreversible work, due to the vanishingly small gap between the ground and first excited states, we see even small quenches are accompanied by a degree of irreversibility, the rate of which grows as the magnitude of the quench is increased. However, beyond the critical point we find the irreversibility grows linearly with the size of the quench.
\begin{figure}[t]
{\bf (a)}\\
\includegraphics[width=0.8\columnwidth]{figures/SmallToLarge_Work.pdf}\\
{\bf (b)}\\
\includegraphics[width=0.8\columnwidth]{figures/SmallToLarge_DeltaF.pdf}\\
{\bf (c)}\\
\includegraphics[width=0.8\columnwidth]{figures/SmallToLarge_IrrWork.pdf}
\caption{(Color online) {\bf (a)} Average Work {\bf (b)} Free energy {\bf (c)} Irreversible work for a quench from $h_i=0.5$ to $h_f$. The insets of panels {\bf (b)} and {\bf (c)} are the first derivative of the functions. In all panels $N=400$.}
\label{fig6}
\end{figure}
\section{Analysis Based on the Spectral function}
\label{spectral}
\begin{figure*}[t]
{\bf (a)} \hskip1\columnwidth {\bf (b)}\\
\includegraphics[width=1.0\columnwidth]{figures/spectral_function_large_to_small.pdf}~
\includegraphics[width=1.0\columnwidth]{figures/spectral_function_small_to_large.pdf}
\caption{(Color online) Spectral function, Eq.~\eqref{specEq}, for $N=50$. {\bf (a)} Starting in the paramagnetic phase with $h_i=1.5$ we examine the spectral function for several decreasing values of $h_f$. {\bf (b)} Spectral function for several quenches to $h_f$ starting from the ferromagnetic phase with $h_i=0.5$.}
\label{fig7}
\end{figure*}
Finally assess the behaviour of the spectral function (SF)
\begin{equation}
\label{specEq}
A(\omega) = 2 \Re \int e^{i \omega t} \mathcal{O} dt.
\end{equation}
This gives insight into the fundamental excitations that are governing the evolution, and therefore serves as an informative tool in understanding the the dynamics of the system~\cite{Mossy}. In Fig.~\ref{fig7} we show the SFs for several quenches, both when the quench remains in the same initial phase, and when it is across the critical point. We restrict ourselves to $N=50$ for simplicity, although qualitatively similar results hold for larger systems.
In panel {\bf (a)} we assume the system begins in the paramagnetic phase fixing $h_i=1.5$. For no quench, i.e. $h_f=1.5$, the SF is a single peak exactly at the ground state energy. For small quenches staying within the paramagnetic phase we see the dynamics continues to be dictated only by the ground state, and this helps understand why the process is fully reversible. As we approach the critical point a second peak appears, corresponding to the second excited state of the final Hamiltonian. When $h_f \sim 1$ this second peak becomes more prominent. Quenching into the ferromagnetic phase, we see significantly more levels enter into the dynamics of the system. Interestingly although the the ground state still contributes to the dynamics, higher excited states play a significantly more dominant role and this results in dynamical orthogonality. We remark, the model naturally has two distinct subspaces. Since our initial state is in the even excitation subspace, only the even states play a role in the dynamics.
In Fig.~\ref{fig7} {\bf (b)} we show the complementary analysis starting from the ferromagnetic phase $h_i=0.5$ and quenching to larger values of the field. Again for reference, we see when no quench is performed, the SF is a single peak exactly at the ground state energy. However, now even for small quenches, $h_f=0.6$, due to the significantly more dense energy spectrum in this phase, more (even excitation) levels play a role in dictating the dynamics of the system, and therefore the system almost immediately witnesses dynamic orthogonality. This in turns allows us to understand the significantly larger irreversibility of the process when quenching from the ferromagnetic into the paramagnetic phase. However, similar to the previous case, when the quench remains in the same phase, i.e. $h_f \leq 1$ in this case, the ground state is still dominant. Quenching (near) to the critical point we see the SF spreads. For $h_f$ deep in the paramagnetic phase the SF is very spread, and again since we have quenched through the QPT, we see higher excited states play the most dominant role in the dynamics.
\section{Discussions and Conclusions}
\label{conclusions}
We have examined the dynamics arising by quenching the parameters of the many-body interacting Lipkin-Meshkov-Glick (LMG) model. Starting from the ground state in a particular phase, and using the time dependent fidelity, we have shown that manifestly different dynamics occur when the quench is restricted to the same initial phase compared to a quench through the critical point. By employing tools from quantum thermodynamics we have shown that the average work maintains a linear relationship with the magnitude of the quench, regardless if it is through the quantum phase transition (QPT) or not. In contrast, the free energy and irreversible work are acutely sensitive to this difference. We find that quenching through the QPT leads to significant increases in the degree of irreversibility. This result can also help in understanding why controlling such many-body systems is so difficult through their QPTs~\cite{AdolfoPRL,CampbellPRL,SaberiPRA}, as it is this irreversibility that needs to be controlled. Starting from the paramagnetic phase, where there is a sizeable energy gap between the ground and first excited states, we have shown the occurrence of dynamical orthogonality serves as a remarkable witness of criticality in the model. Furthermore, by examining the spectral function we have shown that when the quench is through the critical point, the fundamental excitations that govern the dynamics are no longer dictated primarily by the ground state, but in fact higher excited states play the most prominent role. It is important to remark that our analysis is restricted to zero temperature and a natural question arises regarding the situation for finite temperatures. In the case of a quench from the paramagnetic to the ferromagnetic phase, the presence of the energy gap means that for reasonably small temperatures (i.e. temperatures that fail to provide enough thermal energy to excite the first excited state) the results remain largely unaffected. Conversely, when quenching from the ferromagnetic phase, due to the vanishingly energy gap even small temperatures lead to the first excited state becoming populated and thus can significantly change the dynamics. Our results highlight the interesting role the static properties of a many-body system can play in its dynamics. Indeed, such a role has recently been explored in Refs.~\cite{Heyl1,Heyl2,Heyl3,Heyl4,Heyl5} where, for the Ising model and also long-range interacting spin models (including the LMG model), so-called ``{\it dynamical quantum phase transitions}'' have been characterized. Our results add further evidence that equilibrium QPTs are clearly manifest in non-equilibrium processes. Finally we remark that we expect similar features to appear for other 1-dimensional spin systems such as the Ising model, however we leave this for a future study.
\section*{Appendix - Finite Anisotropy}
Here we examine a finite value for the anisotropy parameter $\gamma$ showing that qualitatively the results in the main text are unaffected. We choose $\gamma=0.5$ and restrict to the case of quenching from the paramagnetic to the ferromagnetic phase for brevity. In Fig.~\ref{fig_appendix} we show the TDF [panel {\bf (a)}] and its corresponding dynamical minimum [panel {\bf (b)}], which are complementary to the results shown in Figs.~\ref{fig2} and \ref{fig3} of the main text. For finite $\gamma$ the TDF exhibits the same change in behavior when the quench is through the critical point. However, by changing $\gamma$ we are altering the energy of the system, and therefore this will be evidenced by a change in the frequency of the TDF. We clearly see this effect in panel {\bf (a)} as the time at which the first minimum is achieved is larger than in the $\gamma=0$ case. When the quench is restricted to a single phase the clean periodic behavior is maintained, while for values of $h<1$ this feature is lost and we find that the system can become dynamically orthogonal. Interestingly the `kink' in $\mathcal{L}_\text{min}$ is also still present.
\begin{figure}[h!]
{\bf (a)}\hskip0.35\columnwidth{\bf (b)}\\
\includegraphics[width=0.49\columnwidth]{figures/FigApp_1.pdf}
\includegraphics[width=0.49\columnwidth]{figures/FigApp_2.pdf}
\caption{(Color online) Finite anisotropy, $\gamma=0.5$: {\bf (a)} Time-dependent fidelity for a quench from $h_i=1.5$ to $h_f \in [0.6,1.4]$ with $N=300$. The dashed curves are for $h_f>1$ with $h_f=1.4$ (topmost, blue) and $h_f=1.2$ (purple). The lowest two dotted curves are for $h_f<1$ with $h_f=0.8$ (red) and $h_f=0.6$ (bottom-most orange). The solid black curve is for $h_f=1$. {\bf (b)} {\it Main Panel:} minimum value of TDF achieved for $t\in(0,12)$ quenching the field from $h_i=1.5$ to $h_f$. Each curve from left to right corresponds to an increasing size of $N=100$ (dotted, red), 200 (dash-dotted, orange), and 300 (dashed, blue). {\it Inset:} Zoomed in cross-section of main panel.}
\label{fig_appendix}
\end{figure}
\acknowledgements
I am grateful to Gabriele De Chiara, Thom\'as Fogarty and Mauro Paternostro for useful discussions and exchanges. This work is supported by the EU Collaborative Project TherMiQ (Grant Agreement 618074), the Julian Schwinger Foundation (JSF-14-7-0000), and COST Action MP1209 ``Thermodynamics in the quantum regime".
\section{Introduction}
Phase transitions are an interesting trait of physical systems. In particular, when dealing with quantum many-body set-ups it is curious that a well defined microscopic description can lead to non-trivial and singular behaviours in the thermodynamic limit. The study of the equilibrium properties of such systems is well established in, for example, exactly solvable one-dimensional spin chains~\cite{Sachdev}. With the help of several tools from quantum information, notably quantum correlations, the implications of the presence of quantum phase transitions (QPTs) have been explored~\cite{CorrelationsQPT}.
Beyond the static properties, there is a growing interest in studying the dynamics of many-body quantum systems. This is further catalysed by the increasing interest in understanding the thermodynamic properties of genuinely quantum systems~\cite{GooldJPA}. Despite the significantly more involved nature of studying dynamics, remarkable progress has been made in elucidating the behaviour of important quantities, e.g. (irreversible) work, entropy production, and residual energy, when a many-body system is evolved through its critical point. The most frequent evolution considered is that of a ``sudden quench" of the order parameter~\cite{EchoQuenches,SilvaPRL,DornerPRL,LorenzoPRX,GabrielePRB,DasPRB} (however finite time protocols have also been addressed~\cite{SaroPRB}). Focusing on such a sudden change allows us to capture the salient features of the ensuing non-equilibrium dynamics, while leaving the study of more qualitative differences to a more involved temporal analysis. Typically, the sudden change to the Hamiltonian kicks the system out of equilibrium and can lead to interesting consequences. Notably, for the Ising model the dynamics of a sudden quench have been explored and it has been shown that the irreversible entropy production provides signals of the presence of the equilibrium QPT~\cite{SilvaPRL}, and also can be used to explain emergent phenomena such as the vanishing gap between ground and first excited energy levels in the thermodynamic limit~\cite{DornerPRL}. Recently, the irreversible work was shown to faithfully capture the critical features even for so-called impurity QPTs~\cite{GabrielePRB}.
In this work we add to this endeavour by studying the Lipkin-Meshkiv-Glick model~\cite{LMG65}. The model has attracted substantial interest as it serves as the paradigmatic example of an infinite range interacting system. It can be solved in the thermodynamic limit and exhibits a complex phase diagram~\cite{VidalPRLPRE,CastanosPRB}. We will be interested in exploring how clear signatures of the equilibrium QPT is manifest in the dynamics when the model is quenched through its critical point. We remark that the evolution of this model through its QPT have been studied previously in Ref.~\cite{DasPRB} wherein the equal time-order parameter correlation function was examined and Ref.~\cite{SaroPRB} where the adiabatic dynamics were explored. Our study is set apart from these as it seeks to establish a rigorous link between the thermodynamic quantities such as work and free energy, with the presence of the known QPT. By exploiting the time-dependent fidelity we show that the dynamics are manifestly different when the quench is restricted to a particular phase compared to when the system is quenched through the critical point. More interestingly, we explicitly show that while the average work performed on the system due to the quench is blind to the QPT, the free energy, and therefore the irreversible work, appears acutely sensitive, showing a markedly different rates of change. Finally, we use the spectral function to further understand the fundamental excitations governing the dynamics of the system.
The remainder of the paper is organised as follows. In Sec.~\ref{model} we present the model and introduce the quantities that will be of interest to our analysis. In Sec.~\ref{results} we explore these quantities for quenching the system across the critical point and show that the time-dependent fidelity exhibits interesting features that reveal the critical nature of the system as well as showing that the irreversible work neatly reveals the QPT. Sec.~\ref{spectral} assess the spectral function for various quenches. Finally, Sec.~\ref{conclusions} we present our conclusions and some discussions on our results.
\section{The Model and Figures of Merit}
\label{model}
We consider the ferromagnetic spin-1/2 Lipkin-Meshkov-Glick (LMG) model in a transverse field,
\begin{equation}
\mathcal{H} = -\frac{1}{N}\left( \sum_{i<j} \sigma_x^i \otimes \sigma_x^j+\gamma \sigma_y^i \otimes \sigma_y^j \right) -
h \sum_{i} \sigma_z^i,
\end{equation}
with $\sigma_{x,y,z}$ the Pauli spin-operators, $h$ the magnetic field strength, and $\gamma$ the anisotropy parameter (which we set to zero for simplicity in our simulations, however we remark that qualitatively similar results can be obtained for any $0\leq\gamma<1$, cfr. the appendix). By considering the collective spin operators $S_\alpha=\sum_i \sigma_{\alpha}^i/2$ with $\alpha=\{x,y,z\}$, up to a constant energy shift, the model can be written as
\begin{equation}
\label{collspinLMG}
\mathcal{H} =-\frac{2}{N}\left( S_x^2 + \gamma S_y^2 \right) - 2h S_z.
\end{equation}
In the following we work in the basis of maximum angular momentum (which is a constant of motion) and using the eigenstates of $S_z$ we can diagonalize Eq.~\eqref{collspinLMG} to find the complete spectrum (see e.g. Refs~\cite{VidalPRLPRE,CastanosPRB} for further details). In Fig.~\ref{fig1} we show the energy difference between the 5 lowest excited states and the ground state against $h$. We see when $h>1$ each energy level is distinct. As $h$ is decreased the gap between the first excited state and the ground state closes, and similarly the energy gap between subsequent pairs of excited states also closes. However, an important remark, only when $N\to \infty$ does the gap vanish and all eigenstates become doubly degenerate. Hence, for any finite size there is a small difference between the ground and first excited states~\cite{SaroPRB}.
\begin{figure}[t]
\includegraphics[width=0.85\columnwidth]{figures/GroundStateEnergy2ndDerivative.pdf}
\caption{(Color online) Energy difference between the ground state and the first 5 excited states for $N=400$ plotted against magnetic field strength $h$. {\it Inset:} Second derivative of the ground state energy per site for $N=100$ (red) to $700$ (magenta) in steps of 100.}
\label{fig1}
\end{figure}
The inset in Fig.~\ref{fig1} shows the second derivative with respect to $h$ of the ground state energy (per site) for system sizes ranging from $N=100$ to $N=700$. We see the emergence of a discontinuity appearing, thus signalling the known second order QPT at $h=1$~\cite{VidalPRLPRE}. We are interested in studying the dynamics when the ground state of one phase is evolved using the propagator of another. In what follows we will assume the system is initialized in the ground state of Eq.~\eqref{collspinLMG} corresponding to $h=h_i$. At time $t=0$, we quench the field strength $h_i \to h_f$ and we evolve the initial state according to the new Hamiltonian $\mathcal{H}_f$, so that
\begin{equation}
\ket{\psi(t)} = e^{-i \mathcal{H}_f t} \ket{\psi(0)}.
\end{equation}
Using this we can readily evaluate the time dependent overlap
\begin{equation}
\mathcal{O} = \bra{\psi(0)} \psi(t) \big>.
\end{equation}
This quantity will be central to our analysis as it allows us to access several important quantities that indicate that signatures of the equilibrium QPT are clearly manifest in the system's evolution.
A particularly important quantity will be the time-dependent fidelity (TDF)
\begin{equation}
\mathcal{L} = \vert \mathcal{O} \vert^2,
\end{equation}
which quantifies how different the evolved state is compared to the initial one (we remark this quantity is sometimes referred to as the Loschmidt echo in the literature). The TDF has already proven to be a useful tool in studying critical dynamics~\cite{EchoQuenches}. Additionally, we can determine the average work due to the sudden quench~\cite{LorenzoPRX,CampbellMarchFogarty}
\begin{equation}
\big< W \big> = \sum_j (E_j^f -E_0^i) \big\vert \big<\psi_0^i \vert\psi_j^f\big> \big\vert^2
\end{equation}
where $E_j^f$ and $\vert\psi_j^f \big>$ are the $j$-th eigenenergy and eigenstate of the post-quench Hamiltonian, and $E_0^i$ and $\vert \psi_0^i\big>$ are ground state energy and ground state for the initial Hamiltonian. The sudden nature of the quench drives the system out of equilibrium, and thus introduces a degree of irreversibility of the process. We can quantitatively define the irreversible work as~\cite{GabrielePRB,DeffnerPRL,CampbellMarchFogarty}
\begin{equation}
\big< W_\text{irr} \big> = \big< W \big> - \Delta F
\end{equation}
where $\Delta F$ is the free energy difference. By considering closed dynamics, and since we assume our system begins in the ground state of the initial Hamiltonian, $\Delta F$ is simply given by the difference between the post- and pre-quench ground state energies, i.e. $\Delta F = E_0^f - E_0^i$.
\section{Quench Dynamics Across a Quantum Critical Point}
\label{results}
\subsection{From the Paramagnetic to the Ferromagnetic Phase}
We begin analysing the case of quenching from the paramagnetic phase, setting $h_i=1.5$. In this regime there is a significant energy difference between the ground state and the first excited state, cfr Fig.~\ref{fig1}. In Fig.~\ref{fig2} we show the TDF for $N=400$ and the thick black curve corresponds to $h_f=1$. The dashed curves above this are for quenches that evolve the initial state taking a value of $h_f$ that is still in the paramagnetic phase. We clearly see the regular oscillatory behaviour persists even when quenching close to the critical point. When we evolve the state using $h_f<1$ (lower dotted curves) we see the dynamics loses the clean periodic behaviour, and dynamically the TDF no longer reaches unity. Additionally there is a significant decrease in the values of $\mathcal{L}$, even in some cases reaching exactly 0 indicating that the evolved state is orthogonal to the initial state.
\begin{figure}[t]
\includegraphics[width=0.85\columnwidth]{figures/LargeToSmall_LE_N400.pdf}
\caption{(Color online) Time-dependent fidelity for a quench from $h_i=1.5$ to $h_f \in [0.6,1.4]$ with $N=400$. The dashed curves are for $h_f>1$ with $h_f=1.4$ (topmost, blue) and $h_f=1.2$ (purple). The lowest two dotted curves are for $h_f<1$ with $h_f=0.8$ (red) and $h_f=0.6$ (bottom-most orange). The solid black curve is for $h_f=1$.}
\label{fig2}
\end{figure}
\begin{figure}[t]
\includegraphics[width=0.95\columnwidth]{figures/LargeToSmall_LEmin.pdf}
\caption{(Color online) {\it Main Panel:} minimum value of TDF achieved for $t\in(0,10)$ quenching the field from $h_i=1.5$ to $h_f$. Each curve from left to right corresponds to an increasing size of $N=100$ (dotted, red), 200 (dash-dotted, orange), 300 (dashed, blue) and 400 (solid, black). {\it Upper left inset:} Zoomed in cross-section of main panel. {\it Lower right inset:} Finite size scaling for the first value of field at which the $\mathcal{L}_\text{min}=0$.}
\label{fig3}
\end{figure}
The precise value at which the TDF reaches zero is examined in Fig.~\ref{fig3}. We determine the minimum value of TDF observed, $\mathcal{L}_\text{min}$, within the same time window for Fig.~\ref{fig2} against $h_f$. Clearly, when the quench is small, e.g. $h_f\in(1.2,1.4)$ the minimum value of TDF is still quite large. As the strength of the quench is increased we find this minimum value decreases. Interestingly, $\mathcal{L}_\text{min}=0$ only when $h_f<1$. We see an oscillatory behaviour appearing, however the amplitude of the oscillations is decreasing as the system size is increased. Furthermore, the first value of $h_f$ where $\mathcal{L}_\text{min}=0$, denoted $h_0$, shifts closer to 1 as we increase the system size. Through a finite size scaling with a quadratic fit, the lower right inset shows that this accurately determines the critical point. Such a result is remarkable as it clearly indicates that the equilibrium QPT can be witnessed by the occurrence of dynamical orthogonality.
A final peculiarity appears in studying $\mathcal{L}_\text{min}$, for $N>200$ a kink appears close to the critical point as shown in the upper left inset of Fig.~\ref{fig3}. However, as the system size is increased, this feature appears to move further from the critical value.
\begin{figure}[t]
{\bf (a)}\\
\includegraphics[width=0.8\columnwidth]{figures/LargeToSmall_Work.pdf} \\
{\bf (b)}\\
\includegraphics[width=0.8\columnwidth]{figures/LargeToSmall_DeltaF.pdf}\\
{\bf (c)} \\
\includegraphics[width=0.8\columnwidth]{figures/LargeToSmall_IrrWork.pdf}
\caption{(Color online) {\bf (a)} Average Work {\bf (b)} Free energy {\bf (c)} Irreversible work for a quench from $h_i=1.5$ to $h_f$. The insets of panels {\bf (b)} and {\bf (c)} are the first derivative of the functions. In all panels $N=400$.}
\label{fig4}
\end{figure}
From Fig.~\ref{fig2} it is clear that for small quenches the system dynamically comes close to the initial state, however for larger quenches this is no longer the case, indicating that a degree of irreversibility has been introduced into the system~\cite{BatalhaoPRL}. In Fig.~\ref{fig4} {\bf (a)} we examine the average work done, $\big< W \big>$, against $h_f$. We find $\big< W \big>$ is linearly dependent on the value of $h_f$, the larger the quench $h_i \to h_f$, the more work is done. In panels {\bf (b)} and {\bf (c)} we show the free energy and average irreversible work, respectively. Again, the free energy increases as we increase the size of the quench. However in the inset we examine its rate of change, and we see for $h_f>1$ this rate is linear, and there is a sudden change near $h_f\sim1$. While it is not surprising that the free energy exhibits a non-trivial behavior as we go through the critical point in light of the fact that is defined in terms of the ground state energy, the fact that both $\big< W \big>$ and $\Delta F$ are of the same order of magnitude leads to a trade off between the two quantities. This has interesting consequences for the irreversibility of the process captured by $\big<W_\text{irr} \big>$. Panel {\bf (c)} shows that when the quench is small and confined to the same phase as the initial state, $\big< W_\text{irr} \big> = 0$, indicating that the process is fully reversible, as confirmed by the behavior of the TDF which achieves values of unity during the dynamics. For large quenches, when the system is evolved according to a Hamiltonian in the ferromagnetic phase the average irreversible work becomes non-zero, and the degree of irreversibility grows as the magnitude of the quench increases.
\begin{figure}[t]
\includegraphics[width=0.8\columnwidth]{figures/SmallToLarge_LE_N400.pdf}
\caption{(Color online) Time-dependent fidelity for a quench from $h_i=0.5$ to $h_f \in [0.6,1.4]$ with $N=400$. The dashed curves are for $h_f<1$ with $h_f=0.6$ (blue) and $h_f=0.8$ (purple). The dotted curves are for $h_f>1$ with $h_f=1.2$ (red) and $h_f=1.4$ (left-most, orange). The solid black curve is for $h_f=1$.}
\label{fig5}
\end{figure}
\subsection{From the Ferromagnetic to the Paramagnetic Phase}
We next consider the complementary case of beginning in the ferromagnetic phase, setting $h_i=0.5$, and quenching to increasingly larger values of $h_f$. This scenario is markedly different to that of the previous section as the ground state is now {\it nearly} degenerate, i.e. there is a exponentially (in $N$) vanishing gap between the ground state the the first excited state, cfr. Fig.~\ref{fig1}. Therefore, by quenching the field strength and kicking the system out-of-equilibrium, it quickly becomes excited and occupies higher order states. In Fig.~\ref{fig5} we see this effect clearly, contrary to the previous section, we see even for moderately small quenches ($h_f\gtrsim0.6$) the TDF reaches zero, indicating the evolved state is orthogonal. We further remark, the larger the system the smaller the quench required to achieve dynamical orthogonality, again this is a consequence of the fact that the energy gap between the ground and first excited states decreases with increasing $N$. Thus we cannot use the presence of orthogonality to witness signatures of the QPT in the dynamics. However, there is a clear qualitative difference appearing when the quench is near to or above the critical point. For quenches to $h_f\gtrsim 1.0$ the TDF evolves into fully orthogonal states for a period, before exhibiting short time revivals. The height of these peaks are steadily decreasing in the considered time window, and the width of the revivals broadens. This indicates a sizeable increase in the irreversibility of the process when quenches into the paramagnetic phase are considered.
We confirm this behavior in Fig.~\ref{fig6}. Panel {\bf (a)} shows that the work is a linear function of the magnitude of the quench. However, the free energy and the irreversibility show the same qualitative behavior as shown in the previous section. Focussing on the average irreversible work, due to the vanishingly small gap between the ground and first excited states, we see even small quenches are accompanied by a degree of irreversibility, the rate of which grows as the magnitude of the quench is increased. However, beyond the critical point we find the irreversibility grows linearly with the size of the quench.
\begin{figure}[t]
{\bf (a)}\\
\includegraphics[width=0.8\columnwidth]{figures/SmallToLarge_Work.pdf}\\
{\bf (b)}\\
\includegraphics[width=0.8\columnwidth]{figures/SmallToLarge_DeltaF.pdf}\\
{\bf (c)}\\
\includegraphics[width=0.8\columnwidth]{figures/SmallToLarge_IrrWork.pdf}
\caption{(Color online) {\bf (a)} Average Work {\bf (b)} Free energy {\bf (c)} Irreversible work for a quench from $h_i=0.5$ to $h_f$. The insets of panels {\bf (b)} and {\bf (c)} are the first derivative of the functions. In all panels $N=400$.}
\label{fig6}
\end{figure}
\section{Analysis Based on the Spectral function}
\label{spectral}
\begin{figure*}[t]
{\bf (a)} \hskip1\columnwidth {\bf (b)}\\
\includegraphics[width=1.0\columnwidth]{figures/spectral_function_large_to_small.pdf}~
\includegraphics[width=1.0\columnwidth]{figures/spectral_function_small_to_large.pdf}
\caption{(Color online) Spectral function, Eq.~\eqref{specEq}, for $N=50$. {\bf (a)} Starting in the paramagnetic phase with $h_i=1.5$ we examine the spectral function for several decreasing values of $h_f$. {\bf (b)} Spectral function for several quenches to $h_f$ starting from the ferromagnetic phase with $h_i=0.5$.}
\label{fig7}
\end{figure*}
Finally assess the behaviour of the spectral function (SF)
\begin{equation}
\label{specEq}
A(\omega) = 2 \Re \int e^{i \omega t} \mathcal{O} dt.
\end{equation}
This gives insight into the fundamental excitations that are governing the evolution, and therefore serves as an informative tool in understanding the the dynamics of the system~\cite{Mossy}. In Fig.~\ref{fig7} we show the SFs for several quenches, both when the quench remains in the same initial phase, and when it is across the critical point. We restrict ourselves to $N=50$ for simplicity, although qualitatively similar results hold for larger systems.
In panel {\bf (a)} we assume the system begins in the paramagnetic phase fixing $h_i=1.5$. For no quench, i.e. $h_f=1.5$, the SF is a single peak exactly at the ground state energy. For small quenches staying within the paramagnetic phase we see the dynamics continues to be dictated only by the ground state, and this helps understand why the process is fully reversible. As we approach the critical point a second peak appears, corresponding to the second excited state of the final Hamiltonian. When $h_f \sim 1$ this second peak becomes more prominent. Quenching into the ferromagnetic phase, we see significantly more levels enter into the dynamics of the system. Interestingly although the the ground state still contributes to the dynamics, higher excited states play a significantly more dominant role and this results in dynamical orthogonality. We remark, the model naturally has two distinct subspaces. Since our initial state is in the even excitation subspace, only the even states play a role in the dynamics.
In Fig.~\ref{fig7} {\bf (b)} we show the complementary analysis starting from the ferromagnetic phase $h_i=0.5$ and quenching to larger values of the field. Again for reference, we see when no quench is performed, the SF is a single peak exactly at the ground state energy. However, now even for small quenches, $h_f=0.6$, due to the significantly more dense energy spectrum in this phase, more (even excitation) levels play a role in dictating the dynamics of the system, and therefore the system almost immediately witnesses dynamic orthogonality. This in turns allows us to understand the significantly larger irreversibility of the process when quenching from the ferromagnetic into the paramagnetic phase. However, similar to the previous case, when the quench remains in the same phase, i.e. $h_f \leq 1$ in this case, the ground state is still dominant. Quenching (near) to the critical point we see the SF spreads. For $h_f$ deep in the paramagnetic phase the SF is very spread, and again since we have quenched through the QPT, we see higher excited states play the most dominant role in the dynamics.
\section{Discussions and Conclusions}
\label{conclusions}
We have examined the dynamics arising by quenching the parameters of the many-body interacting Lipkin-Meshkov-Glick (LMG) model. Starting from the ground state in a particular phase, and using the time dependent fidelity, we have shown that manifestly different dynamics occur when the quench is restricted to the same initial phase compared to a quench through the critical point. By employing tools from quantum thermodynamics we have shown that the average work maintains a linear relationship with the magnitude of the quench, regardless if it is through the quantum phase transition (QPT) or not. In contrast, the free energy and irreversible work are acutely sensitive to this difference. We find that quenching through the QPT leads to significant increases in the degree of irreversibility. This result can also help in understanding why controlling such many-body systems is so difficult through their QPTs~\cite{AdolfoPRL,CampbellPRL,SaberiPRA}, as it is this irreversibility that needs to be controlled. Starting from the paramagnetic phase, where there is a sizeable energy gap between the ground and first excited states, we have shown the occurrence of dynamical orthogonality serves as a remarkable witness of criticality in the model. Furthermore, by examining the spectral function we have shown that when the quench is through the critical point, the fundamental excitations that govern the dynamics are no longer dictated primarily by the ground state, but in fact higher excited states play the most prominent role. It is important to remark that our analysis is restricted to zero temperature and a natural question arises regarding the situation for finite temperatures. In the case of a quench from the paramagnetic to the ferromagnetic phase, the presence of the energy gap means that for reasonably small temperatures (i.e. temperatures that fail to provide enough thermal energy to excite the first excited state) the results remain largely unaffected. Conversely, when quenching from the ferromagnetic phase, due to the vanishingly energy gap even small temperatures lead to the first excited state becoming populated and thus can significantly change the dynamics. Our results highlight the interesting role the static properties of a many-body system can play in its dynamics. Indeed, such a role has recently been explored in Refs.~\cite{Heyl1,Heyl2,Heyl3,Heyl4,Heyl5} where, for the Ising model and also long-range interacting spin models (including the LMG model), so-called ``{\it dynamical quantum phase transitions}'' have been characterized. Our results add further evidence that equilibrium QPTs are clearly manifest in non-equilibrium processes. Finally we remark that we expect similar features to appear for other 1-dimensional spin systems such as the Ising model, however we leave this for a future study.
\section*{Appendix - Finite Anisotropy}
Here we examine a finite value for the anisotropy parameter $\gamma$ showing that qualitatively the results in the main text are unaffected. We choose $\gamma=0.5$ and restrict to the case of quenching from the paramagnetic to the ferromagnetic phase for brevity. In Fig.~\ref{fig_appendix} we show the TDF [panel {\bf (a)}] and its corresponding dynamical minimum [panel {\bf (b)}], which are complementary to the results shown in Figs.~\ref{fig2} and \ref{fig3} of the main text. For finite $\gamma$ the TDF exhibits the same change in behavior when the quench is through the critical point. However, by changing $\gamma$ we are altering the energy of the system, and therefore this will be evidenced by a change in the frequency of the TDF. We clearly see this effect in panel {\bf (a)} as the time at which the first minimum is achieved is larger than in the $\gamma=0$ case. When the quench is restricted to a single phase the clean periodic behavior is maintained, while for values of $h<1$ this feature is lost and we find that the system can become dynamically orthogonal. Interestingly the `kink' in $\mathcal{L}_\text{min}$ is also still present.
\begin{figure}[h!]
{\bf (a)}\hskip0.35\columnwidth{\bf (b)}\\
\includegraphics[width=0.49\columnwidth]{figures/FigApp_1.pdf}
\includegraphics[width=0.49\columnwidth]{figures/FigApp_2.pdf}
\caption{(Color online) Finite anisotropy, $\gamma=0.5$: {\bf (a)} Time-dependent fidelity for a quench from $h_i=1.5$ to $h_f \in [0.6,1.4]$ with $N=300$. The dashed curves are for $h_f>1$ with $h_f=1.4$ (topmost, blue) and $h_f=1.2$ (purple). The lowest two dotted curves are for $h_f<1$ with $h_f=0.8$ (red) and $h_f=0.6$ (bottom-most orange). The solid black curve is for $h_f=1$. {\bf (b)} {\it Main Panel:} minimum value of TDF achieved for $t\in(0,12)$ quenching the field from $h_i=1.5$ to $h_f$. Each curve from left to right corresponds to an increasing size of $N=100$ (dotted, red), 200 (dash-dotted, orange), and 300 (dashed, blue). {\it Inset:} Zoomed in cross-section of main panel.}
\label{fig_appendix}
\end{figure}
\acknowledgements
I am grateful to Gabriele De Chiara, Thom\'as Fogarty and Mauro Paternostro for useful discussions and exchanges. This work is supported by the EU Collaborative Project TherMiQ (Grant Agreement 618074), the Julian Schwinger Foundation (JSF-14-7-0000), and COST Action MP1209 ``Thermodynamics in the quantum regime".
|
\section{Introduction}
\label{intro}
Quantum dots and molecules offer unique opportunities to design nanoelectronic devices and to understand fundamental properties of open quantum many-body systems such as quantization, interference,
exchange and nonequilibrium effects.\cite{Kastner2000,Aleiner2002,Fujisawa2006,cuevasscheer2010,Baldea15} Conventional studies
focus on the coupling to noninteracting leads, which is appropriate when the electrons in the leads are Fermi liquids. \cite{datta1997electronic,Andergassen2010}
In contrast, interacting electrons in a quasi-one-dimensional lead can organize themselves into a Luttinger liquid giving rise to a unique set of correlations and properties. \cite{Haldane1981,giamarchi2003quantum} The importance of such correlations on the corresponding transport properties can be demonstrated by Kane-Fischer theory, \cite{Kane1992, Kane1992a} which shows that the conductance through a tunneling barrier vanishes if it is attached to repulsive Luttinger liquid leads, yet becomes perfect for attractive ones. Experimentally, it is challenging to realize Luttinger liquid leads, while the interest in such systems is growing due to possible realizations in quantum wires, \cite{Auslaender2000,Matzdorf2009,Matzdorf2011,Laroche2014} or by the edge states of a fractional quantum Hall system. \cite{Milliken1994,Chang2003,Wurstbauer2013,Pascher2014,Li2015} Thus, it is very interesting to study the (non-equilibrium) properties of such transport setups.
Transport with Luttinger liquid leads has been investigated before. \cite{vonDelft1998,giamarchi2003quantum,Schoenhammer2005,Metzner2012} Two Luttinger liquids separated by a tunneling barrier were first studied by Kane and Fisher,\cite{Kane1992, Kane1992a} and later by others. \cite{Furusaki1993, Matveev1993, Yue1994, Moon1995, Egger1995, Fendley1995, Furusaki1997, Aristov2008, Aristov2009, Einhellinger2012} Fabrizio et al. \cite{Fabrizio1994} and Maurey and Giamarchi \cite{Maurey1995} extended these studies by long-range Coulomb interactions in the leads. noninteracting structures or quantum dots with a single-level that replaces the tunneling barrier of the latter studies were also considered,\cite{Weiss1995, Nazarov2003, Komnik2003, Komnik2003a, Waechter2007, Goldstein2010, Goldstein2010b} including the dual, side-coupled case. \cite{Lerner2008,Goldstein2010a} Here, the position of the energy level is crucial; a level aligned with the chemical potential yields perfect conductance while the conductance is zero otherwise.
Transport with a single level quantum dot, which can be described by an Anderson impurity model has
been investigated by Andergassen et al.,\ \cite{Andergassen2006} where the authors find that screening of the dot spin (Kondo physics) occurs even in the presence of the power-law singularity inherent to a Luttinger liquid.\cite{Haldane1981,Jeckelmann2012}
These considerations were extended to Coulombic dot-lead interactions by Elste et al.,\cite{Elste2011a}
revealing a mechanism for negative differential resistance (NDR) at the Coulomb blockade edge (similar to the NDR mechanisms reported by Matveev and Larkin \cite{Matveev1992} and Dubi \cite{Dubi2013}).
In this work, we study the same model; we consider transport through an Anderson impurity coupled to two Luttinger liquid leads
where all interactions are local and short-ranged. Note that transport through more complex, multi-orbital systems has also been considered,\cite{Furusaki1998,Durganandini2006,Durganandini2006b,Kawaguchi2009,Yang2014} including the effect of electron-phonon interactions.\cite{Maier2010,Yang2010,Skorobagatko2012}
The early studies on the low-energy transport properties of a tunneling barrier connected to Luttinger liquid leads are
based on analytic results. They have been derived either by employing Bethe ansatz solution \cite{Fendley1995} or by
bosonic \cite{Kane1992,Furusaki1993} or fermionic renormalization group.\cite{Matveev1993, Yue1994, Aristov2008, Aristov2009}
Approximate solutions on the more complex setups with double barrier or quantum dot structures have been derived using rate equations,\cite{Furusaki1998,Durganandini2006,Durganandini2006b,Elste2011a} functional renormalization group,\cite{Andergassen2006,Waechter2007} equation of motion techniques,\cite{Skorobagatko2012} full counting statistics \cite{Maier2010} or
nonequilibrium Green's functions.\cite{Kawaguchi2009,Yang2010,Yang2014} Numerically converged or exact results have become available only recently.
This includes a study of the tunneling barrier case by density matrix renormalization group (DMRG) \cite{Einhellinger2012} and
an imaginary-time quantum Monte Carlo (QMC) scheme where a quantum dot is described that is coupled to a single Luttinger liquid lead. \cite{Hattori2014}
Thus, the number of numerically exact schemes that can be used to describe this transport setup is very limited. This is
in contrast to the transport setups with noninteracting leads, where density matrix renormalization group,\cite{Hassanieh2006,daSilva2008,Kirino2008,HeidrichMeisner2009,Kirino2010,Branschadel2010}
numerical renormalization group, \cite{Anders2008,Anders2008b,Schmitt2010,Nghiem2014}
multilayer multiconfiguration time-dependent Hartree theory, \cite{Thoss2013,Balzer2015} iterative \cite{Thorwart2008,Segal2010,Segal2011,Huetzen2012,Weiss2013} and stochastic path-integral
schemes \cite{Werner2006,Schmidt2008,Werner2009,Schiro2010,Gull2010,Cohen2011,Muhlbacher2011,Cohen2013} or
the hierarchical quantum master equation (HQME) method \cite{Jin2008,Zheng2009,Hartle2013,Hartle2015,Cheng2015,Wenderoth2016} have been employed.
In this paper, we extend the HQME framework to describe transport with Luttinger liquid leads.
The method was originally developed to describe finite quantum systems coupled to a bosonic
environment \cite{Tanimura1989,Tanimura1990,Tanimura2006} and was later extended to fermionic
reservoirs. \cite{Welack2006,Jin2008,Hartle2013,Hartle2015}
The method is based on a hybridization expansion that, for noninteracting leads, can be systematically
converged if the temperature of the environment is not too low (for a single-impurity Anderson model, the numerical effort becomes prohibitive below the Kondo temperature). \cite{Hartle2015}
Thus, higher order processes between the dot and the leads, and also correlations
in the dot and the leads, can be accounted for systematically. In parallel, a perturbative analysis to a fixed, given order
is also possible. We present a scheme that allows to obtain converged results with respect to single-particle correlations
in the Luttinger liquid leads, including the effect of the power-law singularity. The effect of two- and $n$-particle correlations,
however, is included only in leading order. Thus, our method represents an important step towards a numerically exact scheme and
allows us to estimate, at least, the role of multi-particle correlations.
In contrast to DMRG or imaginary-time QMC, it also allows us to account for the full nonequilibrium character of this transport problem.
As a standard example, we study the transport properties of a single-level quantum dot
or molecule that can be described by the Anderson impurity model. For this model, we corroborate previous results on NDR at the Coulomb edge, \cite{Elste2011a} assessing the role of cotunneling and beyond-second-order processes and
the role of two-particle correlations in the leads. Processes beyond second order and two-particle correlations turn out to be
negligible in the range of temperatures that we consider ($T\gg T_\text{Kondo}$).
The rest of the paper is organized as follows. In Sec.~\ref{model}, the model that we study is introduced. In Sec.~\ref{method}, we outline the derivation of the HQME with particular emphasis on the differences between the hierarchy construction for noninteracting and interacting leads.
We also compare our new method with the rate equation scheme of Elste et al.. \cite{Elste2011,Elste2011a} We present our results in Sec.~\ref{results}, where we discuss current-voltage characteristics in the steady state for various interaction types and strengths in the leads. The effects of two-particle correlations functions, which become important when interactions in the leads are non-negligible, are also discussed. We present
our conclusions in Sec.~\ref{conclusion}. Technical details are outlined in the Appendices.
\section{Model}
\label{model}
\begin{figure}[tb]
\includegraphics[width=\columnwidth]{model.pdf}
\caption{Graphical scheme of the transport setup that we consider, i.e., a quantum dot coupled to two, semi-infinite Luttinger liquid leads with chemical potentials that differ by the applied bias voltage. }
\label{fig:model}
\end{figure}
The model that we consider (see Fig.~\ref{fig:model}) is described by the Hamiltonian
\begin{equation}
H = H_\text{dot} + \sum_{\alpha = L, R} \left( H_\text{lead}^{\alpha} + H_\text{tun}^{\alpha} \right).
\end{equation}
It includes a quantum dot (located at position $x=0$) that can be described by an Anderson impurity model with a spin-degenerate level $\epsilon$ and an on-dot Coulomb repulsion $U$,
\begin{equation}
H_\text{dot} = \epsilon \sum_{s = \uparrow, \downarrow} n_s +U n_{\downarrow} n_{\uparrow},
\end{equation}
where $n_{s} = d^{\dagger}_s d_s$ is the electron density with spin $s$, and $d^{(\dagger)}_s$ is the dot annihilation (creation) operator with spin $s$.
The dot is coupled to two, semi-infinite Luttinger liquid leads that terminate at $x_{\alpha = L, R}$. We model the coupling between the dot and the leads by a single-particle tunneling Hamiltonian
\begin{equation}
H_\text{tun}^{\alpha} = \sum_{ s} \int_{\alpha} dx \left[ \mathcal{T}_{\alpha} (x) \psi_{\alpha s}^{\dagger} (x) d_{s} +\text{H.c.} \right],
\end{equation}
where $\mathcal{T}_{\alpha } (x)$ is the tunneling amplitude and $\psi_{\alpha s}^{(\dagger)} (x)$ is the electron annihilation (creation) operator in the lead $\alpha$. The integral $\int_{\alpha}$ covers the range from $-\infty$ to $x_L <0$ for the left lead, and from $x_R >0$ to $\infty$ for the right lead. This field-theoretical form of the tunneling Hamiltonian is often used in the description of Luttinger liquid leads. It reduces to a simpler form
\begin{equation}
H_\text{tun}^{\alpha} (t) = \sum_{s} f^{+}_{\alpha s } (t)d_{ s} + \text{H.c.},
\end{equation}
where $f_{\alpha s}^{\sigma}$ is a creation ($\sigma = +$) and annihilation ($\sigma = -$) operator in the interaction picture defined as
\begin{equation}
\begin{split}
f_{\alpha s}^{-(+)} (t) &\equiv e^{i H_\text{lead}^{\alpha} t} \left( \int_{\alpha} dx \mathcal{T}_{\alpha} (x) \psi_{\alpha s}^{(\dagger)} (x) \right) e^{-i H_\text{lead}^{\alpha} t} \\
&\simeq a \bar{\mathcal{T}}_{\alpha} \bar{\psi}_{\alpha s}^{(\dagger)} (t) ,
\end{split}
\end{equation}
if one considers tunneling amplitudes that are non-zero only within a small range $a$ at the edge of each lead. $\bar{\mathcal{T}}_{\alpha}$ and $\bar{\psi}_{\alpha s}^{(\dagger)} (t)$ represent the average tunneling amplitude and field strength over this small range, respectively.
The leads are, for example, Hubbard chains whose low-energy properties can be described by an effective model of a Luttinger liquid. We do not assume a specific form of a microscopic Hamiltonian, because only the effective Luttinger form is used in the following. We set the notation using the bosonized Hamiltonian for an infinite lead \cite{giamarchi2003quantum}
\begin{equation}
H_\text{lead}^{\alpha} = \sum_{\nu = c, s} \frac{1}{2\pi} \int dx \left[ u^{\alpha}_{\nu} K^{\alpha}_{\nu} \left( \nabla \theta^{\alpha}_{\nu}\right)^2 + \frac{u^{\alpha}_{\nu}}{K^{\alpha}_{\nu}} \left( \nabla \phi^{\alpha}_{\nu}\right)^2 \right] ,
\end{equation}
where $\phi_{s,c}$ and $\theta_{s,c}$ are conjugate bosonic variables for spin and charge sectors, $u^{\alpha}_{\nu}$ is the renormalized velocity of sound, and $K^{\alpha}_{\nu}$ are Luttinger parameters. For a Hubbard chain, for example, the Luttinger parameter $K^{\alpha}_c$ is given by the on-site interaction $U_\alpha$ and the bare Fermi velocity $v_{F,\alpha}$ as
\begin{equation}
(K^{\alpha}_c)^2 = \frac{1}{1+U_\alpha/\pi v_{F,\alpha}}.
\end{equation}
We assume a SU(2) invariant system about spins, i.e.\ we set the effective Luttinger constant $K^{\alpha}_s = 1$
and $u^{\alpha}_s$ is equal to $v_{F,\alpha}$. The transformation formula from fermions to bosons is
\begin{equation}
\begin{split}
\psi_{\alpha s } (x,t) &= \sum_{r = \pm} \psi_{\alpha r s} (x,t) \\
&= \sum_{r = \pm} \eta_{\alpha rs} e^{i \mu_{\alpha} t} \frac{e^{ i r k_{F,\alpha} x}}{\sqrt{2\pi a}} e^{-\frac{i}{\sqrt{2}} \left[ \left( r \phi^{\alpha}_{ c} - \theta^{\alpha}_{ c} \right) + s \left( r \phi^{\alpha}_{s} - \theta^{\alpha}_{s} \right) \right]},
\end{split}
\end{equation}
where $r$ is the chirality of the fermion, and $\eta_{\alpha r s}$ is a Klein factor taking care of the anticommutation relations among fermions. $\mu_{\alpha}$ denotes the chemical potential of lead $\alpha$.
An essential ingredient for the HQME is the correlation function of the lead electrons at the dot-lead boundary, which is calculated over the equilibrium density matrix of the leads $\rho_{L+R}$ as $\braket{\cdot}_{L+R} \equiv \mathrm{Tr}_{L+R} \left[ \rho_{L+R} \cdots \right]$. Including the boundary condition for the wave functions $\psi_{L s}^{(\dagger)} (x >x_L) = \psi_{R s}^{(\dagger)} (x < x_R) = 0$, the correlation functions at the dot-lead boundary can be calculated as \cite{Furusaki1998, Elste2011, Elste2011a} (from now on we drop the spatial dependence in the field operators)
\begin{equation}
\begin{split}
C^{\sigma}_{\alpha } (t) &\equiv \left\langle f_{\alpha s}^{\sigma} (t) f_{\alpha s}^{\bar{\sigma}} (0) \right\rangle_{L+R} \\
&= \Gamma_{\alpha} e^{i \sigma \mu_{\alpha} t} \left\{ \frac{i \hbar v_{F,\alpha} \beta_{\alpha}}{\pi a} \sinh \left[ \frac{\pi (t - i \delta)}{\beta_{\alpha} \hbar} \right] \right\}^{-Y_{\alpha}},
\end{split}
\label{psi_corr}
\end{equation}
where $\Gamma_{\alpha} = a |\bar{\mathcal{T}}_{\alpha}|^2/ \pi$ is the system-lead coupling and $\bar{\sigma} \equiv - \sigma$. $\delta$ is a short-time cut-off, which we choose $\delta = 1/W_{\alpha}$ with the band width $W_{\alpha}$ that determines the scale where the fermionic dispersion can be considered linear, $W_{\alpha} \simeq v_{F,\alpha} /a $. $\beta_{\alpha}$ is the inverse temperature of lead $\alpha$. The parameter $Y_{\alpha} = 1/(2K^{\alpha}_c)+1/2$ determines the character and strength of the interactions in the Luttinger liquid in the charge sector. $Y_{\alpha}$ greater (smaller) than 1 represents repulsive (attractive) interactions in the Luttinger liquid. Elste et al.\ \cite{Elste2010, Elste2011a}
showed that a short-range excitonic (density-density) coupling between the dot and the leads effectively renormalizes $Y_{\alpha}$ to
\begin{equation}
Y_{\alpha} = \frac{\left(1 - \frac{K^{\alpha}_c}{u^{\alpha}_c} V_{\alpha} \right)^{2}}{2 K^{\alpha}_c} + \frac{K^{\alpha}_c V_{\bar{\alpha}}^2 }{2 u_c^{\alpha,2}} + \frac{1}{2},
\label{Y_renorm}
\end{equation}
where $V_{\alpha}$ is the excitonic coupling strength. This means that an originally repulsive lead, which is usually the case for quantum wires and carbon nanotubes, can be effectively attractive. Thus in the following we consider both repulsive and attractive cases. The excitonic coupling also renormalizes the energy $\epsilon$ and the interaction strength $U$, but we assume that such effects are already included in these parameters.
We note that the correlation function, Eq.~\eqref{psi_corr}, gives a power-law singularity in the single-particle tunneling density of states at low frequencies and low temperatures,
\begin{equation}
\begin{split}
J_{\alpha}(\omega) &\equiv \int_{-\infty}^{\infty} \frac{dt}{2\pi} e^{-i \omega t} C^{+}_{\alpha}(t) + \text{h.c.} \\
&\propto |\omega - \mu_{\alpha}|^{Y-1} \ \ \text{as} \ \ |\omega - \mu_{\alpha}| \ll W_{\alpha}, T_{\alpha}.
\end{split}
\end{equation}
Numerical examples for the single-particle tunneling density of states are depicted in Fig.~\ref{fig:psd}. While attractive interactions ($Y=0.8$) induces a sharp peak at the chemical potential, $\omega =0$, repulsive interactions ($Y=1.2$) lead to a sharp dip. This dependence of the low energy single-particle tunneling density of states on the interaction parameter $Y_{\alpha}$ crucially affects the transport properties, even at the single-particle level.
\begin{figure}[tb]
\includegraphics[width=0.9 \columnwidth]{psd.pdf}
\caption{Single-particle tunneling density of states, $J(\omega)$, at the edge of a Luttinger liquid lead for attractive ($Y=0.8$) and repulsive interactions ($Y=1.2$) with $T = 200$\,K, and $W = 4$\,eV. $\Gamma$ is the system-lead coupling strength.}
\label{fig:psd}
\end{figure}
\section{Method}
\label{method}
\subsection{Hierarchical equations of motion}
In this section, we briefly outline the derivation of the hierarchical equations of motion for transport with Luttinger liquid leads. The derivation is quite general, and not necessarily restricted to Luttinger liquids, while, to implement the formalism in practice, we make use of the specific properties of Luttinger liquids. Note that our derivation follows closely Refs.\ \onlinecite{Jin2008} and \onlinecite{Hartle2013}.
We first trace out the lead variables from the total density matrix $\rho_T (t)$ of the system (the dot and leads) to obtain an equation of motion for the reduced density matrix for the dot degrees of freedom,
\begin{equation}
\rho (t) \equiv \mathrm{Tr}_{L+R} \left [ \rho_T (t)\right] .
\end{equation}
Formally the time propagation of the reduced density matrix can be written as $\rho (t) = \mathcal{U} (t, t_0) \rho (t_0)$, where $\mathcal{U} (t, t_0)$ can be written in terms of a path integral representation as
\begin{equation}
\mathcal{U} (t, t_0) = \int_{\xi_0}^{\xi} \mathcal{D} \tilde{\xi} \int_{\xi'_0}^{\xi'} \mathcal{D} \tilde{\xi}' e^{ i S[\tilde{\xi}]} \mathcal{F}[\tilde{\xi}, \tilde{\xi}']e^{- i S[\tilde{\xi}']},
\end{equation}
where $\ket{\xi} \equiv \exp \left(\sum_{ s} \xi_{s} d^{\dagger}_{ s} \right) | 0 \rangle$ is a fermionic coherent state, and $S$ is the action of the dot Hamiltonian. $\mathcal{F} $ is the Feynman-Vernon influence functional
\begin{multline}
\mathcal{F} = \Bigg\langle T \exp \left[ -i \int_{t_0}^t d\tau \sum_{\alpha, s} f^{+}_{\alpha s } (\tau) \tilde{\xi}_{s}(\tau) + \text{H.c.} \right] \\ \times \bar{T} \exp \left[ i \int_{t_0}^t d\tau \sum_{\alpha, s} f^{+}_{\alpha s } (\tau) \tilde{\xi}_{s}(\tau) + \text{H.c.} \right] \Bigg\rangle_{L+R} ,
\label{FV}
\end{multline}
with time-ordering $T$ and anti-time-ordering $\bar{T}$. Expanding the exponentials in Eq.~\eqref{FV} and integrating the lead variables, $\mathcal{F}$ contains a series of correlation functions of $f^{(\dagger)}_{\alpha s}$. Let us schematically write this as
\begin{equation}
\mathcal{F} \sim 1 + \int C^{(1)} \tilde{\xi}^2 + \int C^{(2)}\tilde{\xi}^4 + \dots,
\label{FV_exp}
\end{equation}
where $C^{(k)}$ denotes a $k$-particle correlation function (including disconnected diagrams). The formally exact series expansion of $\mathcal{F}$ is given in Appendix \ref{appA}. We can reorganize the expansion by using the influence exponent
\begin{equation}
\Phi \equiv - \log \mathcal{F}.
\label{replica_id}
\end{equation}
The corresponding expansion of $\Phi$, i.e., the cumulant expansion of $\mathcal{F}$ is found to be a series of connected correlation functions,
\begin{equation}
\begin{split}
\Phi &\sim - \int C^{(1)} \tilde{\xi}^2 - \int C^{(2)}\tilde{\xi}^4 + \int [C^{(1)}]^2 \tilde{\xi}^4 + \dots \\
&\sim - \int C^{(1)} \tilde{\xi}^2 - \int \hat{C}^{(2)}\tilde{\xi}^4 + \dots,
\end{split}
\end{equation}
where $\hat{C}^{(2)}$ is the connected two particle correlation functions in the leads. For noninteracting leads, only the first term remains since all the higher order correlation functions can be written as products of $C^{(1)}$ by Wick's theorem, and are summed up to be zero. In the Appendix\ \ref{appB}, we show an explicit formula for $\Phi$ up to $\hat{C}^{(2)}$ for interacting leads.
Now using Eq.~\eqref{replica_id}, we can express $\mathcal{F}$ as the exponential of connected correlation functions.
Neglecting ($k>2$)-particle correlations, the equation of motion of $\mathcal{F}$ can thus be written as
\begin{equation}
\begin{split}
\dot{\mathcal{F}} &\simeq -i \sum_{\sigma \alpha s} \mathcal{A}^{\bar{\sigma}}_{ s} \left( \mathcal{B}^{\sigma}_{\alpha s} + \hat{\mathcal{B}}^{\sigma}_{1, \alpha s} + \hat{\mathcal{B}}^{\sigma}_{2, \alpha s} \right) \mathcal{F} \\
&\equiv -i \sum_{\sigma \alpha s} \mathcal{A}^{\bar{\sigma}}_{ s} \left( \mathcal{F}^{\sigma}_{\alpha s} + \hat{\mathcal{F}}^{\sigma}_{1, \alpha s} + \hat{\mathcal{F}}^{\sigma}_{2, \alpha s} \right),
\end{split}
\label{FV_EOM_0}
\end{equation}
with
\begin{equation}
\begin{split}
&\mathcal{A}^{\sigma}_{s} \equiv \xi_{ s}^{\sigma} + \xi_{ s}^{\prime \sigma}, \\
&\mathcal{B}^{\sigma}_{\alpha s} \equiv -i \int_{t_0}^{t} d\tau \left[ C^{\sigma}_{\alpha } (t, \tau) \tilde{\xi}_{s}^{\sigma} (\tau) - C^{\bar{\sigma} *}_{\alpha } (t, \tau) \tilde{\xi}_{ s}^{\prime \sigma} (\tau) \right] ,\\
&C^{\sigma}_{\alpha } (t-t') \equiv \left\langle f_{\alpha s}^{\sigma} (t) f_{\alpha s}^{\bar{\sigma}} (t') \right\rangle_{L+R}.
\end{split}
\label{ABC}
\end{equation}
The first term is related to the $C^{(1)}$ correlator in Eq.~\eqref{FV_exp}. The second and third terms are associated with the $\hat{C}^{(2)}$ correlator. Explicit expressions of the $\hat{\mathcal{B}}_{1, \alpha s}^{\sigma}$ and $\hat{\mathcal{B}}_{2,\alpha s}^{\sigma}$ operators are given in Appendix \ref{appB}. They include the correlation function $\hat{C}^{(2)}$ and three $\tilde{\xi}^{(\prime)}$ Grassmann numbers convoluted in three time integrals.
In the second line, we define the $1$st tier auxiliary influence functionals (AIF's) $\mathcal{F}^{\sigma}_{\alpha s}$, $\hat{\mathcal{F}}^{\sigma}_{1, \alpha s} $, and $\hat{\mathcal{F}}^{\sigma}_{2, \alpha s} $. They enter the equation of motion of the reduced density matrix by the following set of auxiliary density operators (ADO's)
\begin{equation}
\begin{split}
\rho^{\sigma}_{\alpha s} &= \mathcal{U}^{\sigma}_{\alpha s} (t, t_0) \rho(t_0), \\
\mathcal{U}^{\sigma}_{\alpha s} (t, t_0) &= \int_{\xi_0}^{\xi} \mathcal{D} \tilde{\xi} \int_{\xi'_0}^{\xi'} \mathcal{D} \tilde{\xi}' e^{ i S[\tilde{\xi}]} \mathcal{F}^{\sigma}_{\alpha s}[\tilde{\xi}, \tilde{\xi}']e^{- i S[\tilde{\xi}']},
\end{split}
\end{equation}
with similar relations for $\hat{\rho}^{\sigma}_{1(2), \alpha s}$ and $\hat{\mathcal{F}}^{\sigma}_{1(2), \alpha s}$. The equation of motion of the reduced density matrix in terms of ADO's can thus be written as
\begin{equation}
\dot{\rho}(t) = -i\mathcal{L} \rho (t) - i \sum_{\alpha s \sigma} \left[ d^{\bar{\sigma}}_{s}, \rho^{\sigma}_{\alpha s} (t) + \hat{\rho}^{\sigma}_{1, \alpha s} (t) + \hat{\rho}^{\sigma}_{2, \alpha s} (t) \right] .
\end{equation}
The Louiville term $\mathcal{L} \rho = [H_\text{dot}, \rho ]$ arises from the time derivative of the action $S$ in Eq.~\eqref{FV}.
Taking the time derivative of ADO's opens an infinite hierarchy of equations of motion, including higher order ADO's. In particular, a new type of ADO's emerges from the time derivative of the correlation functions, for example, $C^{\sigma}_{\alpha }$ as
\begin{equation}
\tilde{\mathcal{B}}^{\sigma}_{\alpha s} = -i \int_{t_0}^{t} d\tau \left[ \dot{C}^{\sigma}_{\alpha } (t, \tau) \tilde{\xi}_{s}^{\sigma} (\tau) - \dot{C}^{\bar{\sigma} *}_{\alpha } (t, \tau) \tilde{\xi}_{ s}^{\prime \sigma} (\tau) \right]\neq \mathcal{B}^{\sigma}_{\alpha s}
\end{equation}
In order to obtain a closed set of ADO's, the correlation functions $C_{\alpha }^{\sigma} (t )$ are translated to a sum of exponentials (Meier-Tannor parametrization method),
\begin{equation} \begin{split}
C_{\alpha }^{\sigma} (t ) &\equiv \sum_{l = 0}^{\infty} h_{\alpha l } e^{- \omega_{\alpha l }^{\sigma} t}, \\
h_{\alpha l}&= |\Gamma_{\alpha}| 2^{Y_{\alpha}} \frac{(Y_{\alpha})_l}{l!} \left( \frac{W_{\alpha} \beta_{\alpha}}{\pi} \right)^{-Y_{\alpha}} e^{-i Y_{\alpha} \frac{\pi}{2} } e^{ i (Y_{\alpha} + 2 l )\frac{\pi}{\beta_{\alpha} W_{\alpha}}}, \\
\omega_{\alpha l }^{\sigma} &= (Y_{\alpha} + 2 l) \frac{\pi}{\beta } - i \sigma \mu_{\alpha},
\end{split}
\label{C_approx}
\end{equation}
where $(x)_n \equiv \prod_{k=0}^{n-1} (x+k)$ is the Pochhammer symbol. Correspondingly, $ \mathcal{B}^{\sigma}_{\alpha s}$ is also decomposed into a sum as $ \mathcal{B}^{\sigma}_{\alpha s} = \sum_{l=0}^{\infty} \mathcal{B}^{\sigma}_{\alpha l s}$, and
\begin{equation}
\mathcal{B}^{\sigma}_{\alpha l s} = -i h_{\alpha l } \int_{t_0}^{t} d\tau e^{- \omega_{\alpha l }^{\sigma} (t-\tau)}\left[ \tilde{\xi}_{s}^{\sigma} (\tau) - \tilde{\xi}_{ s}^{\prime \sigma} (\tau) \right] .
\end{equation}
Now the zeroth tier equation becomes
\begin{equation}
\dot{\rho} = -i\mathcal{L} \rho - i \sum_{j} \left[ d^{\bar{\sigma}}_{s}, \rho_j + \hat{\rho}_{1, j} + \hat{\rho}_{2,j} \right],
\label{0th_tier}
\end{equation}
where we introduced abbreviated notations $j = \{ \alpha l s \sigma \}$, and $\bar{j} = \{ \alpha l s \bar{\sigma} \}$. Note that it is in general impossible to decompose $\hat{C}^{(2)}$ as in Eq.~\eqref{C_approx}. Therefore, we use a local time approximation to reduce its complexity to an exponential series. This approximation is motivated by the fact that the $\hat{C}^{(2)} (t_1, t_2, t_3, t_4)$ are exponentially small when any two of the four time variables differ by $|t_i - t_j| \geq 1/W$ (see Ref.~\onlinecite{Jang2002} and Appendix \ref{appC} for details); the approximation is thus better in the wide band limit. We now discuss the equations of motion of the two classes of ADO's, $\rho_j$ and $\hat{\rho}_{1(2),j}$, separately.
We start with the equation of motion of the $\rho_j$ operators,
ignoring the contributions from ($k>1$)-particle correlations (which is an exact procedure for noninteracting leads).
It leads to a hiearchy of general $n$th tier ADOs, $\rho_{\mathbf{j}}$, associated with an AIF $\mathcal{B}_{j_n}\mathcal{B}_{j_{n-1}} \cdots \mathcal{B}_{j_1} \mathcal{F}$, and obey the following equations of motion
\begin{multline}
\dot{\rho}^{(n)}_{\mathbf{j}} = - i \mathcal{L} \rho^{(n)}_{\mathbf{j}} - \left( \sum_{k=1}^n \omega_{j_k} \right)\rho^{(n)}_{\mathbf{j}} \\
- i \sum_{k = 1}^n (-1)^{n-k} \tilde{\mathcal{C}}_{j_k} \rho^{(n-1)}_{\mathbf{j}_k}- i \sum_{j \not \in \mathbf{j}} \mathcal{A}_{\bar{j}} \rho^{(n+1)}_{\mathbf{j}j}
\label{HQME}
\end{multline}
with superoperators $\tilde{\mathcal{C}}$ and $\mathcal{A}$ as
\begin{equation} \begin{split}
\tilde{\mathcal{C}}_{j} \rho^{(n)} &= \sum_{\nu} \left[ h_{\alpha l} d_{s}^{\sigma} \rho^{(n)} - (-1)^{n} h^{ *}_{\alpha l } \rho^{(n)} d_{ s}^{\sigma} \right], \\
\mathcal{A}_{\bar{j}} \rho^{(n)} &= d_{ s}^{\bar{\sigma}} \rho^{(n)} +(-1)^{n} \rho^{(n)} d^{\bar{\sigma}}_{ s} .
\end{split}\end{equation}
The second and third terms originate from the time derivative of exponential time dependence and of the integral limit in $\mathcal{B}_{j_k}$ respectively. In the third term, $\rho^{(n-1)}_{\mathbf{j}_k}$ is an ADO where $\mathcal{B}_{j_k}$ is removed from $\rho^{(n)}_{\mathbf{j}}$. The last term is associated with the time derivative of $\mathcal{F}$, which adds $\mathcal{B}_{j}$ on the left of $\mathcal{B}_{j_n}\mathcal{B}_{j_{n-1}} \cdots \mathcal{B}_{j_1} \mathcal{F}$. The constraint on the summation, $ j \not \in \mathbf{j}$, is due to the anticommutation property of Grassmann numbers. We note that the current is related to the first tier ADO as \cite{Jin2008}
\begin{equation}
I_{\alpha} = i \sum_{l, s} \mathrm{Tr}{\left[ \left( \rho_{\alpha l s}^{\dagger} + \hat{\rho}_{1, \alpha l s}^{\dagger} + \hat{\rho}_{2, \alpha l s}^{\dagger} \right)d_{s} - \text{h.c.} \right] }.
\end{equation}
The contributions from the operators $\hat{\rho}$ always conserves the current, i.e., $I_L = - I_R$, which can be easily seen from the equation of motions of $\hat{\rho}$ considered next.
The corresponding equation of motion is given by
\begin{gather}
\begin{split}
\partial_t \hat{\rho}_{1, \alpha l s}^{\sigma} &\simeq -i\mathcal{L} \hat{\rho}_{1, \alpha l s}^{ \sigma} -\omega_{\alpha \sigma l} \hat{\rho}_{1, \alpha l s}^{\sigma}\\
&-i \Gamma \sum'_{\substack{s_1 s_2 s_{3} \\ \sigma_1 \sigma_2 \sigma_{3} }} c^{\alpha}_{ j j_1 j_2 j_3} \left[ \rho d^{ \bar{\sigma}_{3}}_{s_{3}}d^{\bar{\sigma}_{2}}_{s_{2}} d^{\bar{\sigma}_{1}}_{s_{1}} h_{\alpha l}^* - d^{\bar{\sigma}_1}_{s_1} d^{\bar{\sigma}_2}_{s_2} d^{\bar{\sigma}_3}_{s_3} \rho h_{\alpha l} \right] ,
\end{split} \\
\begin{split}
\partial_t \hat{\rho}_{2, \alpha l s}^{\sigma} &\simeq -i\mathcal{L} \hat{\rho}_{2, \alpha l s}^{ \sigma} -\omega_{\alpha \sigma l} \hat{\rho}_{2, \alpha l s}^{\sigma} \\
&-i \Gamma \sum'_{\substack{s_1 s_2 s_{3} \\ \sigma_1 \sigma_2 \sigma_{3} }} c^{\prime \alpha}_{ j_1 j j_2 j_3} \left[d^{ \bar{\sigma}_2}_{s_{2}} d^{ \bar{\sigma}_3}_{s_{3}} \rho d^{ \bar{\sigma}_{1}}_{s_{1}} - d^{ \bar{\sigma}_1}_{s_{1}}\rho d^{ \bar{\sigma}_3}_{s_{3}} d^{ \bar{\sigma}_{2}}_{s_{2}} \right] \hat{h}_{\alpha l},
\end{split}
\label{2p-order_eom}
\end{gather}
including the time-local approximation on $\hat{C}^{(2)}$ that we mentioned earlier and neglecting higher order terms.
The summation is restricted to the combinations that conserve charges and spins; $\sigma + \sum_{i=1}^3\sigma_i = 0$, and $\sigma s + \sum_{i=1}^3 \sigma_i s_i= 0$. This is because $\hat{C}^{(2)}$ is only nonzero when these conditions are satisfied, as is known from Luttinger liquid theory. \cite{gogolin2004bosonization} The coefficient $c^{(\prime), \alpha}_{j j_1 j_2 j_3 }$ takes one of the values
\begin{equation}
0, \ 2^{1-Y_{\alpha}}-1, \ 2^{Y_{\alpha}}-2, \ 2^{Y_{\alpha}-1}-1, \ 2^{-Y_{\alpha}}[1 - (-1)^{1-Y_{\alpha} }] ,
\label{C2_coeff}
\end{equation}
depending on the indices (see the explicit form in Appendix C). As expected, they vanish in the noninteracting limit $Y_{\alpha} = 1$.
Please note that, in the following, we ignore the effect of higher orders of the $\hat{\rho}$ operators by truncating the hierarchy at this point, that is at the first tier of the $\hat{\rho}$ operators.
This treatment is not, per se, systematic but allows us to estimate the leading order effect. As it turns out to be marginal in the parameter
regime that we consider, we assume that higher order $\hat{\rho}$ contributions are even smaller and, therefore, negligible. In that sense,
our results that we present in Sec.\ \ref{results} can be considered to be converged.
We are thus left with the infinite hierarchy of equations of the standard $\rho$ operators. We truncate it according to the arguments in Ref.~\onlinecite{Hartle2013}, where the importance of a $n (\geq 2)$th tier ADO, $\rho_\mathbf{j}$, is estimated by assigning it
the following amplitude
\begin{equation}
\prod_{k=1}^n \left( \frac{\left| h_{j_k} \right|}{\Re \omega_{j_k} \sum_{p=1}^{k-1} \Re \omega_{j_p}} \right).
\label{amplitude}
\end{equation}
In our calculations, we then consider only operators that have amplitudes larger than a given threshold value $A_\text{th}$. Convergence is achieved by reducing the threshold value systematically. In addition to the Grassmann and hermite natures of AIF's, this preselection greatly reduces the number of ADO's. Thus, we can obtain numerically converged, exact results for the ADO's of $\rho$ type (as was shown, for example, by a direct comparison to quantum Monte Carlo simulations only recently).\cite{Hartle2015} The $\hat{\rho}$ hierarchy is truncated at the first tier, ignoring higher order terms which correspond to $\mathcal{O}(\Gamma^{4})$. Since $n$-particle connected correlation functions scale as $\mathcal{O} (\Gamma^{n})$ and
we typically obtain converged results already at second or third order for high enough temperatures, the truncation of the $\hat{\rho}$ hierarchy goes well along the lines of our convergence criterion.
In the next section, we will show that the contributions from $\hat{C}^{(2)}$ have indeed only a very minor effect on the current-voltage characteristics. Therefore, we believe that essential correlation effects in the leads are included in our scheme. Yet, it is an important open problem how to go beyond the limitations and approximations of our scheme in order to get a systematic improvement of our results.
\subsection{Rate equations}
In the next section, we compare the HQME results with rate equation results. The latter contains only the lowest order of the hybridization expansion outlined above, ignoring, in addition, non-Markovian memory effects. We will see how these two assumptions affect the transport properties. The rate equations are derived for the weak dot-lead tunneling with Born-Markov-secular approximations:
\begin{equation}
\begin{split}
\dot{\rho} &= - \int^{\infty}_0 d\tau \sum_{\alpha s } \Big\{ \\
&\left[d_s(t) d^{\dagger}_s (t-\tau) \rho(t) - d^{\dagger}_s(t-\tau) \rho(t) d_s(t) \right] C^{+}_{\alpha }(\tau) \\
&+ \left[d^{\dagger}_s(t) d_s (t-\tau) \rho(t) - d_s(t-\tau) \rho(t) d^{\dagger}_s(t) \right] C^{-}_{\alpha }(\tau) \\
&+ \text{h.c.} \Big\}.
\end{split}
\end{equation}
Using the diagonal form of the density matrix,
\begin{equation}
\rho = P_0 \ket{0}\bra{0}+ \frac{P_{1}}{2}\left( \ket{\uparrow}\bra{\uparrow} + \ket{\downarrow}\bra{\downarrow} \right) + P_2\ket{\uparrow \downarrow}\bra{\uparrow \downarrow},
\label{population}
\end{equation}
we obtain the rate equations
\begin{equation}
\begin{split}
\dot{P}_0 &= \sum_{\alpha} \left[ -2 P_0 \mathcal{R}^{\alpha}_{01} +P_1 \mathcal{R}^{\alpha}_{10} \right], \\
\dot{P}_1 &= \sum_{\alpha} \left[ -P_1 \left( \mathcal{R}^{\alpha}_{10} + \mathcal{R}^{\alpha}_{12} \right) + 2 P_0 \mathcal{R}^{\alpha}_{01} + 2 P_2 \mathcal{R}^{\alpha}_{21} \right], \\
\dot{P}_2 &= \sum_{\alpha} \left[ -2 P_2 \mathcal{R}^{\alpha}_{21} +P_1 \mathcal{R}^{\alpha}_{12} \right].
\end{split}
\end{equation}
The coefficients $\mathcal{R}^{\alpha}_{pq}$ are given by Laplace transforms of the correlation function $C^{\alpha}_0(\tau) \equiv C^{\sigma}_{\alpha } (\tau) e^{ -i \sigma \mu \tau}$ as
\begin{equation}
\mathcal{R}^{\alpha}_{pq} = 2 \Re \int_{0}^{\infty} d \tau e^{- i \omega_{pq} \tau} C^{\alpha}_0(\tau) ,
\end{equation}
with $\omega_{01} = -\omega_{10} = \epsilon - \mu_{\alpha}$ and $\omega_{12} = -\omega_{21} = \epsilon + U - \mu_{\alpha}$. The current from lead $\alpha$ is
\begin{equation}
\langle I_{\alpha} (t) \rangle = 2 P_0 \mathcal{R}_{01}^{\alpha} + P_1 \mathcal{R}_{12}^{\alpha} - P_1 \mathcal{R}^{\alpha}_{10} - 2 P_2 \mathcal{R}^{\alpha}_{21}.
\end{equation}
For the steady state, we find
\begin{multline}
\langle I_{\alpha} (t) \rangle = \frac{2e}{R} \Big[ \mathcal{R}_{21} \left(\mathcal{R}^{\alpha}_{01}\mathcal{R}^{\bar{\alpha}}_{10}-\mathcal{R}^{\bar{\alpha}}_{01} \mathcal{R}^{\alpha}_{10} \right) \\
+ \mathcal{R}_{01} \left(\mathcal{R}^{\alpha}_{12}\mathcal{R}^{\bar{\alpha}}_{21}-\mathcal{R}^{\bar{\alpha}}_{12} \mathcal{R}^{\alpha}_{21} \right) \Big],
\end{multline}
where $\mathcal{R}_{ij} = \sum_{\alpha} \mathcal{R}_{ij}^{\alpha}$, and
\begin{equation}
R = \mathcal{R}_{10} \mathcal{R}_{21} + 2 \mathcal{R}_{01} \mathcal{R}_{21} + \mathcal{R}_{01} \mathcal{R}_{12} .
\end{equation}
\section{Results}
\label{results}
\subsection{Weak coupling regimes}
\begin{table}[tb!]
\centering
\begin{ruledtabular}
\caption{Parameters used in our simulations. Note that these parameters are borrowed from molecular systems, but
can be easily scaled to describe, for example, semiconductor heterostructeres, etc.}
\begin{tabular}{cccccc}
$W$ & $\Gamma$ & $U$ & $l_\text{max}$ & $A_\text{th}$ \\ \hline
4.0eV & 0.001eV & 0.25eV & $ 2500$ & $5.0\times10^{-10}$
\end{tabular}
\label{parameters}
\end{ruledtabular}
\end{table}
\begin{figure}[tb]
\includegraphics[width=\columnwidth]{IV_1.pdf}
\caption{Current-voltage characteristics for the quantum dot at the charge-symmetric point ($\epsilon=-U/2$), for
various interaction parameters of the Luttinger liquid leads ($Y=0.8/1/1.2$ in the upper/middle/lower panels) and temperatures ($T=100/200$\,K in the left/right panels). Solid (dashed) lines are obtained by HQME (rate equations).}\label{IV_1}
\end{figure}
In this section, we show the current-voltage ($I-V$) characteristics of the model outlined in Sec.~\ref{model}. For simplicity, we consider symmetric leads; $W$, $\beta$, and $Y$ are independent of $\alpha$. The chemical potentials are taken in the standard form $\mu_{L} = -\mu_{R} = V/2$, corresponding to the choice of symmetric leads. We consider two scenarios where the quantum dot is at the charge symmetric point, $\epsilon = - U/2 $, and where it is unpopulated at zero bias, $\epsilon = U/2$. Other parameters of the model are given in Table \ref{parameters}.
In Fig.~\ref{IV_1}, we plot the current-voltage characteristics of a charge-symmetric quantum dot for three cases that are relevant, for example, for carbon nanotubes:\cite{Elste2011} attractive leads ($Y = 0.8$, upper panels), noninteracting leads ($Y=1$, middle panels), and repulsive leads ($Y= 1.2$, lower panels) for $T=100$\,K (left panels) and $T=200$\,K (right panels). In the attractive cases ($Y = 0.8 $), we observe a pronounced NDR above the Coulomb-blockade threshold at $V/2 = |\epsilon| = |\epsilon + U|$. This NDR is closely related to the peak in the single-particle tunneling density of states (see Fig.\ \ref{fig:psd}) and can also appear for noninteracting electrodes but attractive electron-electron interactions at the terminal site, \cite{Dubi2013}
or slightly more general cases. \cite{Matveev1992} It appears quenched in the HQME results as compared to the rate equation results due to the effect of broadening or the hybridization with the leads. We also find that higher order or cotunneling processes which are not included in the rate equation results enhance the low bias conductance. The enhancement is strongest for the attractive case. This behavior can also be understood in terms of the peaked tunneling density of states (cf.\ Fig.\ \ref{fig:psd}), which is larger around the chemical potential
in the attractive case as compared to the noninteracting or repulsive one. When $Y = 1$, the lead is noninteracting, and thus our formalism is numerically exact within the model. We find that due to the finite band width, $W = 4.0$ eV, the current has a small negative slope or NDR at high voltages $V \geq 1$. For repulsive leads ($Y = 1.2$), the curves monotonically increase in the range of voltages that we consider, and the deviation from the rate equations is negligible. This is because the density of states near the Fermi energy depends on the interaction parameter $Y$ with an explicit factor $W^{-Y}$ as seen in Eq.~\eqref{psi_corr}. Thus, effectively, the tunneling density of states is suppressed for repulsive interactions. This means that the perturbation parameter $\Gamma/T$ is also smaller and, thus, the differences between HQME and rate equations are also smaller.
Fig.~\ref{IV_2} shows the current-voltage characteristics for a quantum dot that is unpopulated at zero bias, $\epsilon = U/2$, i.e., far away from the charge symmetric point. A similar situation has been considered, for example, in Ref.~\onlinecite{Elste2011}. Here, in contrast to the charge-symmetric case, we observe two steps at $V/2=|\epsilon|$ and $V/2=|\epsilon+U|$. In the charge-symmetric case, they appear at the same voltages. Both steps are followed by a decrease of the current level in the attractive case $Y=0.8$, similar as in the charge-symmetric case. The effect is again more pronounced in the rate equation solutions due to broadening that is missing in the rate equation results.
Similar to the charge-symmetric case, the effect of cotunneling processes at low bias voltages is more pronounced at lower temperatures (see the upper left and upper right panels). It is generally less pronounced in the charge-non-symmetric case, because the single-particle levels at $\epsilon+U$ are farther away from the chemical potentials, that is the associated probability for virtual excitations is lower as compared to the charge-symmetric case.
\begin{figure}[!tb]
\includegraphics[width=\columnwidth]{IV_2.pdf}
\caption{Current-voltage characteristics for the quantum dot far away from the charge-symmetric point ($\epsilon=U/2$), for
various interaction parameters of the Luttinger liquid leads ($Y=0.8/1/1.2$ in the upper/middle/lower panels) and temperatures ($T=100/200$\,K in the left/right panels). Solid (dashed) lines are obtained by HQME (rate equations).}\label{IV_2}
\end{figure}
Differences between the rate equation and HQME results reveal the effect of higher order processes, including second and beyond-second order processes. Some more insights can be obtained by comparing other, approximate solutions. The role of the Markov approximation, for example, can be assessed by comparing rate equation results with a first-tier truncation of the HQME. Such a comparison is depicted by the blue and green lines in Fig.\ \ref{IV_comp}. The negligible differences between the two curves shows that the Markov approximation is well justified in the parameter regime that we consider. Moreover, comparing a second-tier truncation of the HQME with the full converged result (yet, including the approximations with respect to multi-particle correlations; see Sec.\ \ref{method}) allows us to reveal the effect of beyond-second order processes (cf.\ orange and purple lines in Fig.\ \ref{IV_comp}). They also turn out to be negligible.
Please note that a different behavior can be expected at low temperatures, especially when, for example, Kondo correlations emerge \cite{Hartle2015}.
\begin{figure}[tb]
\includegraphics[width=0.8 \columnwidth]{comparison.pdf}
\caption{Comparison of current-voltage characteristics for the quantum dot at the charge-symmetric point ($\epsilon=-U/2$) for
$Y=0.8$ and $T=100$\,K obtained with rate equations (blue line), full first order (dashed green line), second order (orange line) and the converged HQME results (dashed purple line).}\label{IV_comp}
\end{figure}
\subsection{Strong coupling regimes}
So far, we looked at the cases where interactions of the leads are relatively weak ($|Y-1| \lesssim 20\%$), since our decomposition scheme,
Eq.~\eqref{C_approx}, converges well only for these cases. However, in this regime, the effect of the two-particle correlation functions $\hat{C}^{(2)}$ is quite small since the order of the corresponding ADO's $\hat{\rho}$ is
\begin{equation}
\mathcal{O} (\hat{\rho} )\sim \left( \frac{\Gamma}{T} \right)^2 \left(\frac{T}{W} \right)^{Y} c_{j_1j_2j_3j_4}.
\end{equation}
This estimate is motivated by Eq.~\eqref{2p-order_eom}; in the steady state, the amplitude of the $\hat{\rho}_{1,\alpha s}^{\sigma}$ and $\hat{\rho}_{2,\alpha s}^{\sigma}$ operators
scale with $\Gamma h$ and $\Gamma \hat{h}$ respectively and with $1/\omega_l$, including an additional factor that vanishes as the interaction strength decreases. Thus, in the above weak coupling regimes, we find the effect of $\hat{\rho}$ ignorable. The complete treatment of both $\hat{\rho}$ and regular ADO's at higher tiers is desirable but not possible within our current scheme. That being said, we would like to illustrate the effects of two-particle correlation functions for relatively strong couplings realized by very high temperatures ($Y= 0.5$, $\Gamma = 0.02$ eV, and $T=4000$\,K, other parameters are the same as in the previous subsection), and stop the hierarchy of the regular ADO's at the first tier. This allows us to disentangle the higher order tunneling effect from regular ADO's and from two-particle correlations that enter via the $\hat{\rho}$ operators.
Fig.~\ref{rho_hat} shows the current-voltage characteristics and dot populations with and without the $\hat{\rho}$ contributions. For this extreme case, we find that the effect of two particle correlation functions is still negligible for the current level ($\sim 2 \%$), while the dot populations are slightly more affected ($\sim 5\%$). The current is slightly enhanced by two-particle correlations. The probability of having a single particle on the dot is also increased, while the ones of having no particle or two particles are reduced.
\begin{figure}[tb]
\includegraphics[width=0.8\columnwidth]{rho_hat.pdf}
\caption{The top panel shows the current-voltage characteristics with and without contributions of two-particle correlations $C^{(2)}$, i.e., $\hat{\rho}$ ADO's. The bottom panel shows the dot populations, Eq.~\eqref{population}, with and without contributions of two-particle correlations $C^{(2)}$. }
\label{rho_hat}
\end{figure}
\section{Conclusions}
\label{conclusion}
We presented a generalized hierarchical quantum master equation technique to describe transport through an interacting quantum dot or molecule that is coupled to interacting, semi-infinite, Luttinger liquid leads. Our method paves the way towards a numerically exact description of this transport problem. This, however, is not fully achieved yet. While the results can be converged with respect to single-particle correlations in the leads [where we included contributions up to fifth order, $\mathcal{O}(\Gamma^5)$, in order to obtain fully converged results], multi-particle correlations can be included only in leading order, i.e., $\mathcal{O}(\Gamma^2)$. A systematic improvement of the latter approximation represents the next step. Nevertheless, we were able to present well based arguments and numerical results, which show that these contributions are negligible for the parameters that we considered, in particular the high temperature regime. Thus, assuming that a hybridization expansion converges, our method can give results that should be very close to the exact result.
As a test case, we considered transport through a quantum dot or molecule that can be described by a single-level Anderson impurity model. We corroborated previous results on a mechanism for negative differential resistance that can be associated with the power-law singularity of an attractive Luttinger liquid. In the repulsive case, the power-law singularity has the opposite effect and even overrides the NDR effect originating from a finite band width in the leads. Higher order effects smear the NDR effects slightly due to broadening. In the low bias regime, cotunneling effects enhance the conductivity significantly. While this is well known, we were able to show that beyond-second order processes do not change these effects significantly. This justifies, inter alia, a treatment by second-order perturbation theory. Our statements and findings are restricted to the high temperature regime and the reduced complexity of our model. More complex quantum dot structures, including, for example, also spin-orbit interactions or the coupling to vibrational degrees of freedom, may show a richer behavior. \\
\section{Acknowledgements}
We thank K.\ Sch\"onhammer for helpful discussions and comments.
RH is supported by Deutsche Forschungsgemeinschaft (DFG) under grant No.\ HA 7380/1-1.
JO and LM are supported by the Deutsche Forschungsgemeinschaft through the SFB 925 and the Hamburg Centre for Ultrafast Imaging, and from the Landesexzellenzinitiative Hamburg, supported by JoachimHerzStiftung.
\onecolumngrid
|
\section{Introduction}\label{Intro}
The history of kinetic theory starts with the classical Boltzmann equation, which describes the dynamics of dilute monoatomic gases (cf. \cite{Cercignani:1988:BEI,CercignaniIllnerPulvirenti:1994:TMT,Cercignani:1975:TAB,Glassey:1996:CPK}). In order to extend the Boltzmann equation from classical particles to quantum particles, the Uehling-Uhlenbeck equation was introduced \cite{Nordheim:OTK:1928,UehlingUhlenbeck:TPI:1933}.
However, the Uehling-Uhlenbeck equation is only true for a dilute Bose gas, which is above the Bose-Einstein Condensate (BEC) critical temperature.
After the first observation of the Bose-Einstein condensation of dilute Bose gases in a gas of rubidium $^{87}$Rb atoms \cite{WiemanCornell} and then in a gas of sodium $^{23}$Na atoms \cite{Ketterle}, which lead to the 2001 Nobel Prize in physics of Wieman, Cornell and Ketterle \cite{BarbaraGoss}, there has been a huge amount of research investigating properties of dilute BECs.
After the pioneering work of Kirkpatrick and Dorfman \cite{KD1,KD2} deriving a kinetic model for bosonic gases below the BEC critical temperature, there has been an explosion of research on the kinetic theory for BECs (see \cite{semikoz1997condensation,Matthew:Thesis:2001,semikoz1995kinetics,IG,KD1,KD2,ZarembaNikuniGriffin:1999:DOT,Spohn:2010:KOT,GriffinNikuniZaremba:2009:BCG,Stoof:1999:CVI,ArkerydNouri:AMP:2013,ArkerydNouri:2015:BCI,Allemand:DOF:2009,PomeauBrachetMetensRica,josserand2001nonlinear,josserand2006self,LacazeLallemandPomeau:2001:DFO,QK0,QK1,QK2,QK3,QK4,QK5} and references therein). By using the PY method from quantum field theory, Reichl and Gust discovered a new collision operator (cf. \cite{ReichlGust:2013:TTF,ReichlGust:2013:RRA,ReichlGust:2012:CII}). The model of Reichl and Gust is more precise when describing the dynamics of BECs; it can be used to obtain microscopic expressions for the six hydrodynamic modes of a dilute Bose-Einstein condensate: two transverse (shear) modes and four longitudinal modes corresponding to the first sound (density waves) and second sound (temperature waves). We refer to the series of books \cite{ColdAtoms1,ColdAtoms2,ColdAtoms3,ColdAtoms4} and the review paper \cite{anglin2002bose} for a complete discussion and more references on this rapidly growing research on quantum gases. We also refer to \cite{seiringer2011excitation,rougerie2013quantum,LewinNamSerfatySolovej:2015:BSO,GrillakisMachedon:PEA:2013,KirkpatrickSchlein:2013:ACL,bach2016time,DeckertFrohlichPickl:2016:DSW,mitrouskas2016bogoliubov,boccato2015quantum,ErdosSchleinYau:DOT:2010}, and cited references, for recent works on the rigorous derivation of nonlinear Sch\"odinger and related equations starting from the Quantum N-body problem.
\bigskip\\ While the mathematical theory for classical gases has been significantly developed, with the work of Carleman, DiPerna, Lions, Villani and many other mathematicians (for example, see \cite{Carleman:1933:TEI,DiPernaLions:1989:OCP,Villani:2002:RMT,DesvillettesVillani:2005:OTG,Villani:2002:RMT}), there are still many open research topics for quantum gases, which need to be explored. An important result on the mathematical side is the work of Escobedo and Velazquez \cite{EscobedoVelazquez:2015:FTB}, in which they prove that the Uehling-Ulenbeck equation has a blow-up solution. The aim of this work is to provide the first mathematical study of the convergence to equilibrium of solutions to the discrete quantum Boltzmann model describing the dynamics of thermal cloud by Kirkpatrick-Dorfman and others, and the modified quantum Boltzmann model by Reichl and Gust.\\
\bigskip
{\\\\\bf Quantum Boltzmann model of the thermal cloud.} The quantum Boltzmann equation of the thermal cloud describes the evolution of the density distribution function $f$ of the excited atoms outside of the condensate. The distribution $f(t,p)$ is a function of time $t$ and momenta $p$. If we denote
$$f_1=f(t,p_1), f_2=f(t,p_2), f_3=f(t,p_3), f_4=f(t,p_4),$$
then $f_1$ satisfies
\begin{eqnarray}
\label{QuantumBoltzmann1}
\frac{\partial f_1}{\partial t}&=&C_{12}[f_1]+C_{22}[f_1],
\end{eqnarray}
where
\begin{eqnarray}
C_{22}[f_1]&:=&\int_{\mathbb{R}^{9}}\mathcal K^{22}_{p_1, p_2, p_3,p_4}\delta(p_1+p_2-p_3-p_4)\delta(\mathcal{E}_{p_1}+\mathcal{E}_{p_2}-\mathcal{E}_{p_3}-\mathcal{E}_{p_4})\\\label{QuantumBoltzmann2}
& &\times[(1+f_1)(1+f_2)f_3f_4-f_1f_2(1+f_3)(1+f_4)]dp_2dp_3dp_4,\\\nonumber
C_{12}[f_1]&:=&\int_{\mathbb{R}^{6}}\mathcal K^{12}_{p_1, p_2, p_3}\delta(p_1-p_2-p_3)\delta(\mathcal{E}_{p_1}-\mathcal{E}_{p_2}-\mathcal{E}_{p_3})\\\label{QuantumBoltzmann3}
& &\times[(1+f_1)f_2f_3-f_1(1+f_2)(1+f_3)]dp_2dp_3\\\nonumber
&&-2\int_{\mathbb{R}^{6}}\mathcal K^{12}_{p_1, p_2, p_3}\delta(p_2-p_1-p_3)\delta(\mathcal{E}_{p_2}-\mathcal{E}_{p_1}-\mathcal{E}_{p_3})\\\nonumber
& &\times[(1+f_2)f_1f_3-f_2(1+f_1)(1+f_3)]dp_2dp_3,
\end{eqnarray}
and
$\mathcal K^{22}_{p_1, p_2, p_3,p_4}, \mathcal K^{12}_{p_1, p_2, p_3}\geq 0$ are the collision kernels, which are radially symmetric, and symmetric with respect to the permutation of $p_1$, $p_2$, $p_3,$ and $p_4$:
\begin{align*}
&~~\mathcal K^{22}_{p_1, p_2, p_3,p_4}=\mathcal K^{22}_{|p_1|,|p_2|,|p_3|,|p_4|}=\mathcal K^{22}_{|p_2|,|p_1|,|p_3|,|p_4|}=\mathcal K^{22}_{|p_3|,|p_2|,|p_1|,|p_4|}\\
&~~=\mathcal K^{22}_{|p_4|,|p_2|,|p_3|,|p_1|}=\mathcal K^{22}_{|p_1|,|p_3|,|p_2|,|p_4|}=\mathcal K^{22}_{|p_1|,|p_4|,|p_3|,|p_2|}=\mathcal K^{22}_{|p_1|,|p_2|,|p_4|,|p_3|},
\end{align*}
and
\begin{align*}
&~~\mathcal K^{12}_{p_1, p_2, p_3}=\mathcal K^{12}_{|p_1|,|p_2|,|p_3|}=\mathcal K^{12}_{|p_2|,|p_1|,|p_3|}=\mathcal K^{12}_{|p_3|,|p_2|,|p_1|}=\mathcal K^{12}_{|p_1|,|p_3|,|p_2|},
\end{align*}
where $|p|$ denotes the length of the vector $p$. We suppose that the temperature is very low compared to the Bose-Einstein critical temperature. As a result, the energy $\mathcal{E}_p=\mathcal{E} (p)$ is given by the phonon dispersion law (cf. \cite{ReichlBook}):
\begin{eqnarray}
&&\mathcal{E} (p)=c |p|, \,\,\,c=\sqrt{\frac {gn_c} {m}}. \label{E3a}
\end{eqnarray}
The operator $C_{22}$ is often referred to as the Uehling-Uhlenbeck collision integral, and describes two-body collisions between excited atoms ($2$ atoms $\rightleftharpoons$ $2$ atoms). The collision operator $C_{12}$ describes collisions which involve one condensate atom ($1$ atom $\rightleftharpoons$ $2$ atoms). Let us note that $\mathcal K^{12}_{p_1, p_2, p_3}$ is a function of $n_c$, $|p_1|$, $|p_2|$, $|p_3|$ and it becomes $0$ when $n_c=0$ (cf. \cite{ZarembaNikuniGriffin:1999:DOT}).
\\\\ {\it Nordheim or Uehling-Uhlenbeck equation.}
If there is no condensate, i.e. $n_c=0$, then $C_{12}=0$. As a result, Equation \eqref{QuantumBoltzmann1} can be reduced to the Nordheim or Uehling-Uhlenbeck equation
\begin{equation}\label{Norheim}
\frac{\partial f_1}{\partial t}=C_{22}[f_1].
\end{equation}
~~\\ {\it Quantum Boltzmann equation of phonons and quantum Boltzmann equation of bosons at very low temperature.}
If the temperature of the quasiparticles in the dilute Bose gas is below the Bose-Einstein transition temperature $T_c$, one can assume the dominance of the bosons-condensate interactions $C_{12}$ over the bosons-bosons interactions $C_{22}$ and the quantum Boltzmann equation of bosons at very low temperature reads (see \cite{ArkerydNouri:2012:BCI,EPV}):
\begin{equation}\label{QuantumPhonon}
\frac{\partial f_1}{\partial t}=C_{12}[f_1].
\end{equation}
Interestingly, the quantum {\it phonon} Boltzmann equation \cite{Peierls:1960:QTS,Buot:ORT:1997} coming from the quantum theory of {\it solids} has exactly the same mathematical formulation as \eqref{QuantumPhonon}. {\it Moreover, it is in the context of the study of phonon interactions in anharmonic crystals \cite{Peierls:1993:BRK} that the first derivation of a kinetic model of weak turbulence was obtained}: R. Peierls suggested the theoretical option of considering the anharmonicities as a small perturbation to the perfectly harmonic crystal, which leads to a kinetic model of an interacting phonons in terms of a nonlinear Boltzmann equation. The phonon Boltzmann equation is then used to carry on the actual computation of the thermal conductivity of dielectric crystals. {\it Moreover, let us emphasize that in the Zakharov's weak turbulence theory approach to Bose-Einstein condensates, an equation similar to \eqref{QuantumPhonon} has also been obtained (cf. \cite{ZakharovNazarenko:DOT:2005})}. As an attempt to build a theory for quantum Boltzmann equations, some mathematical results have been obtained in \cite{AlonsoGambaBinh,JinBinh,Binh9,GambaSmithBinh,GermainIonescuTran,ToanBinh,nguyen2017quantum,SofferBinh1,ReichlTran,soffer2016coupling} .
{\\\\\bf The modified quantum Boltzmann model of the thermal cloud.}
As mentioned in the introduction, in \cite{ReichlGust:2013:TTF,ReichlGust:2013:RRA,ReichlGust:2012:CII}, Reichl and Gust discovered a new collision operator, which makes the model \eqref{QuantumBoltzmann1}-\eqref{QuantumBoltzmann3} more complicated:\begin{eqnarray}\label{QuantumBoltzmannLinda}
\frac{\partial f_1}{\partial t}&=&C_{12}[f_1]+C_{22}[f_1]+C_{13}[f_1],\end{eqnarray}
{where}\begin{equation}\label{QuantumBoltzmannLinda13}
\begin{aligned}
C_{13}[f_1]\ =&\int_{\mathbb{R}^{3\times 3}} \mathcal K^{13}_{p_1, p_2, p_3,p_4}\delta(p_1-p_2-p_3-p_4)\delta(\mathcal{E}_{p_1}-\mathcal{E}_{p_2}-\mathcal{E}_{p_3}-\mathcal{E}_{p_4})\\
&\ \times[(1+f_1)f_2f_3f_4-f_1(1+f_2)(1+f_3)(1+f_4)]dp_2dp_3dp_4\\
&\ -3\int_{\mathbb{R}^{3\times 3}}\mathcal K^{13}_{p_1, p_2, p_3,p_4}\delta(p_2-p_1-p_3-p_4)\delta(\mathcal{E}_{p_2}-\mathcal{E}_{p_1}-\mathcal{E}_{p_3}-\mathcal{E}_{p_4})\\
&\ \times[(1+f_2)f_1f_3f_4-f_2(1+f_1)(1+f_3)(1+f_4)]dp_2dp_3dp_4,
\end{aligned}
\end{equation}
and $C_{12}$, $C_{22}$ are defined in \eqref{QuantumBoltzmann2}, \eqref{QuantumBoltzmann3} and $\mathcal K^{13}_{p_1, p_2, p_3,p_4}$ is positive, radially symmetric, and symmetric with respect to the permutation of $p_1$, $p_2$, $p_3$, $p_4$
\begin{align*}
&~~\mathcal K^{13}_{p_1, p_2, p_3,p_4}=\mathcal K^{13}_{|p_1|,|p_2|,|p_3|,|p_4|}=\mathcal K^{13}_{|p_2|,|p_1|,|p_3|,|p_4|}=\mathcal K^{13}_{|p_3|,|p_2|,|p_1|,|p_4|}\\
&~~=\mathcal K^{13}_{|p_4|,|p_2|,|p_3|,|p_1|}=\mathcal K^{13}_{|p_1|,|p_3|,|p_2|,|p_4|}=\mathcal K^{13}_{|p_1|,|p_4|,|p_3|,|p_2|}=\mathcal K^{13}_{|p_1|,|p_2|,|p_4|,|p_3|}.
\end{align*}
\\ Notice that $p=0$ corresponds to the condensate and it turns out that we need to consider the collisions between atoms only in the high-temperature region with large momenta. Therefore, in this paper, we impose the assumption that $\mathcal K^{22}_{p_1, p_2, p_3,p_4}$, $\mathcal K^{13}_{p_1, p_2, p_3,p_4}$ become $0$ if one of the variables $p_1,p_2,p_3,p_4$ is $0$.
{\\\\\bf Reaction networks and a toric dynamical system approach for the relaxation to equilibrium problem.}
The mathematical program about the trend to equilibrium for kinetic equations was initiated by Desvillettes and Villani in \cite{DesvillettesVillani:2001:OTG,DesvillettesVillani:2005:OTG}. The study of the relaxation of BECs to thermodynamic equilibrium has also played very important role in the theory of Bose gases \cite{ReichlGust:2013:TTF,ReichlGust:2013:RRA,ReichlGust:2012:CII,GriffinNikuniZaremba:2009:BCG,ZarembaNikuniGriffin:1999:DOT}. In \cite{Binh9} the authors have proved the existence and uniqueness of solutions of the linearized model of \eqref{QuantumPhonon} satisfying conservation of energy. They have showed that these solutions converge to the corresponding stationary state, at an explicit algebraic rate as time tends to infinity.
\\ In order to study the Peierls model \eqref{QuantumPhonon}, the quantum Boltzmann model of the thermal cloud \eqref{QuantumBoltzmann1}, and the modified quantum Boltzmann model of the thermal cloud \eqref{QuantumBoltzmannLinda}, our main tool is to convert these equations into chemical reaction systems and use an extension of the theory of toric dynamical systems (cf. \cite{MR2561288}).
\bigskip
\\ In general, there is great interest in understanding the qualitative behavior of deterministically modeled chemical reaction systems, including the existence of positive equilibria, stability properties of equilibria, and the non-extinction, or persistence, of species, which are the constituents of these systems \cite{Feinberg:1972:CBI,Feinberg:1995:TEA,Horn:1974:TDO,Anderson:2001:APG,AngeliLeenheerSontag:2011:PRF,MR3199409, MR2604624,MR2734052,MR2561288,desvillettes2016trend}. Toric dynamical systems -- originally called complex-balanced systems (cf. \cite{MR2561288,HornJackson:1972:GMA}) -- are models used to describe an important class of chemical kinetics. The complex-balanced condition was first introduced by Boltzmann \cite{Boltzmann} for modeling collisions in kinetic gas theory. Based on this condition, it was shown by Horn and Jackson \cite{HornJackson:1972:GMA,Horn:1972:NAS,Feinberglecture,Gunawardena} that a complex-balanced system has a unique locally stable equilibrium within each linear invariant subspace. To underline the tight connection to the algebraic study of toric varieties, the name ``toric dynamical system'' was proposed in \cite{MR2561288}. The most important problem in the theory of toric dynamical systems is the Global Attractor Conjecture, which says that the complex balanced equilibrium of a toric dynamical system is a globally attracting point within each linear invariant subspace. This global attractor question is strongly related to the convergence to equilibrium problem in the study of kinetic equations. A proof to the Global Attractor Conjecture for small dimensional systems has been supplied in \cite{CraciunNazarovPantea:2013:PAP}, and a complete proof has been proposed in \cite{Craciun:2015:TDI}.
\bigskip\\
{\it Our goal is to use the tools developed in \cite{CraciunNazarovPantea:2013:PAP,Craciun:2015:TDI} to prove the relaxation to equilibrium of discrete versions of the Peierls model \eqref{QuantumPhonon}, whose collision operator is $C_{12}$. Similarly, we will prove the relaxation to equilibrium of the quantum Boltzmann model of the thermal cloud \eqref{QuantumBoltzmann1}, whose collision operator is $C_{12}+C_{22}$, and modified quantum Boltzmann model of the thermal cloud \eqref{QuantumBoltzmannLinda}, whose collision operator is $C_{12}+C_{22}+C_{13}$.}
\bigskip\\
Notice that our relaxation to equilibrium results also agree with results from physics \cite{ReichlGust:2013:TTF,ReichlGust:2013:RRA,ReichlGust:2012:CII}. On the other hand, it was proved in \cite{EscobedoVelazquez:2015:FTB}, that \eqref{Norheim} has a blow-up solution. However, the initial condition which creates blow-up solutions is concentrated around $p=0$, which does not correspond to the high temperature regime of the thermal cloud under investigation here. Moreover, the discrete models considered here are the usual ones obtained from the derivations of these physical systems: for example, see equations $(A23)$ and $(A27)$ of \cite{ZarembaNikuniGriffin:1999:DOT} for $C_{12}$ and $C_{22}$.
\\
The plan of our paper is the following:
\begin{itemize}
\item In section \ref{Sec:Q12}, we show that the discrete version of equation \eqref{QuantumPhonon} could be rewritten as a chemical reaction network. By using an approach inspired by the theory of toric dynamical system, we prove in Theorem 2.1. that the solution of the discrete version of \eqref{QuantumPhonon} converges to the equilibrium exponentially in time.
\item In section \ref{Sec:Qnm}, we generalize Theorem \ref{TheoremQ12} to collision operators of the forms $C_{13}$ and $C_{22}$. We prove that the solutions of the discrete versions of these equations, associated with the collision operators $C_{13}$ and $C_{22}$ converge to equilibria exponentially in Theorems 3.1 and \ref{TheoremQnm}. In the case of $C_{22}$, we consider a one-dimensional version of the model.
\item In Theorem \ref{TheoremSumQnm} of Section 4, we extend Theorem \ref{TheoremQnm} to the quantum Boltzmann model of the thermal cloud \eqref{QuantumBoltzmann1}, and the modified quantum Boltzmann model of the thermal cloud \eqref{QuantumBoltzmannLinda}.
\end{itemize}
\section{A reaction network approach for the case of $C_{12}$}\label{Sec:Q12}
\subsection{The dynamical system associated to $C_{12}$}
As mentioned in the introduction, the model derived from physics to describe the system that couples BEC-excitations at very low temperature (and phonon interactions in anharmonic crystals) is the discrete version of \eqref{QuantumPhonon} described below.
\\\\ Let $\mathcal{L}_R$ denote the lattice of integer points
$$\mathcal{L}_R=\{p\in\mathbb{Z}^3, |p|<R\}.$$
The discrete version of the quantum Boltzmann equation \eqref{QuantumPhonon} reads
\begin{equation}\label{DiscreteQuantum}\begin{aligned}
\dot{f}_{p_1}=
&\sum_{\substack{p_2,p_3\in\mathcal{L}_R,\\ p_1-p_2-p_3=0, \\ \mathcal{E}(p_1)-\mathcal{E}(p_2)-\mathcal{E}(p_3)=0}}\mathcal{K}^{12}_{p_1,p_2,p_3}\left\{(f_{p_1}+1)f_{p_2}f_{p_3}-f_{p_1}(f_{p_2}+1)(f_{p_3}+1)\right\}\\
&\quad -2\sum_{\substack{p_2,p_3\in\mathcal{L}_R,\\ p_1+p_2-p_3=0,\\ \mathcal{E}(p_1)+\mathcal{E}(p_2)-\mathcal{E}(p_3)=0}}\mathcal{K}^{12}_{p_1,p_2,p_3}\left\{(f_{p_3}+1)f_{p_1}f_{p_2}-f_{p_3}(f_{p_1}+1)(f_{p_2}+1)\right\},~~\end{aligned}
\end{equation}
for all $p_1$ in $\mathcal{L}_R$, where $\mathcal{E}(p)$ is defined in \eqref{E3a}.
\subsection{Decoupling the quantum Boltzmann equation associated to $C_{12}$}
Note that when $p_1=0$, $\mathcal{K}^{12}_{p_1,p_2,p_3}$ is also $0$, and therefore, we get
\begin{eqnarray}\label{DiscreteQuantumIndex0}
\dot{f}_0=0,
\end{eqnarray}
which says that $f_0(t)$ is a constant for all time $t$. Moreover, $f_{p_1}$ does not depend on $f_0$ for all $p_1\ne 0$. Therefore, without loss of generality, we can suppose that $f_0(0)=0$, which leads to $f_0(t)=0$ for all $t$.
\\ Taking into account the fact $\mathcal{E}(p)=c|p|$, note that if $p_1,p_2,p_3\in\mathcal{L}_R$ are different from $0$ and $p_3=p_1+p_2$ and $|p_3|=|p_1|+|p_2|$ (like in the second sum of \eqref{DiscreteQuantum}), then $p_1,p_2,p_3$ must be collinear and on the same side of the origin. Therefore, we infer that there exists a vector $P$ and $k_1$, $k_2$, $k_3>0$, $k_1,k_2,k_3\in\mathbb{Z}$ such that
$$p_1=k_1 P;~~~p_2=k_2 P;~~~p_3=k_3 P,~~~k_1+k_2=k_3.$$
Since $\mathcal{L}_R$ is bounded, it follows that $k_1,k_2,k_3$ belong to a finite set of integer indices $\mathbb{I}=\{1,\dots,I\}$.
Arguing similarly for the first sum in \eqref{DiscreteQuantum}, we deduce that \eqref{DiscreteQuantum} is equivalent with the following system for $k_1\in\mathbb{I}$
\begin{equation}\label{DiscreteQuantum1DVersion2}\begin{aligned}
\dot{f}_{Pk_1}
=&\quad\sum_{\substack{k_2,k_3\in
\mathbb{I},\\ k_1-k_2-k_3=0}}\mathcal{K}^{12}_{Pk_1,Pk_2,Pk_3}\left\{(f_{Pk_1}+1)f_{Pk_2}f_{Pk_3}-f_{Pk_1}(f_{Pk_2}+1)(f_{Pk_3}+1)\right\}\\
&\quad-{2}\sum_{\substack{k_2,k_3\in
\mathbb{I},\\ k_1+k_2-k_3=0}}\mathcal{K}^{12}_{Pk_1,Pk_2,Pk_3}\left\{(f_{Pk_3}+1)f_{Pk_1}f_{Pk_2}-f_{Pk_3}(f_{Pk_1}+1)(f_{Pk_2}+1)\right\}.
\end{aligned}
\end{equation}
Note that the system of equations \eqref{DiscreteQuantum1DVersion2} shows a {\it decoupling} of the system of equations \eqref{DiscreteQuantum} along a ray $\{kP_0\}$ with $k>0$ (see Figure 1). As a consequence, it is sufficient to study the system of equations \eqref{DiscreteQuantum1DVersion2} for a fixed value of $P_0$, instead of the system of equations \eqref{DiscreteQuantum}.
If we denote $f_{k_1P_0}$ by $f_{k_1}$ (with $k_1\in \mathbb{I}$) and $\mathcal{K}^{12}_{k_1P_0,k_2P_0,k_3P_0}$ by $\mathcal{K}^{12}_{k_1,k_2,k_3}$, with an abuse of notation, we obtain the following new system for the ray $\{k_1P_0|k_1>0\}$:
\begin{equation}\label{DiscreteQuantum1D}\begin{aligned}
\dot{f}_{k_1}=&\quad\sum_{\substack{k_2,k_3\in
\mathbb{I},\\ k_1=k_2+k_3}}\mathcal{K}^{12}_{k_1,k_2,k_3}\{(f_{k_1}+1)f_{k_2}f_{k_3}-f_{k_1}(f_{k_2}+1)(f_{k_3}+1)\}
\\
&\quad-2\sum_{\substack{k_2,k_3\in
\mathbb{I},\\ k_1+k_2=k_3}}\mathcal{K}^{12}_{k_1,k_2,k_3}\{(f_{k_3}+1)f_{k_1}f_{k_2}-f_{k_3}(f_{k_1}+1)(f_{k_2}+1)\}, ~\forall k_1\in\mathbb{I}.\end{aligned}
\end{equation}
A simple calculation leads to the following {\it conservation of energy}
\begin{equation}\label{MassConservation1}
\sum_{k=1}^I k\dot{f_k}=0,
\end{equation}
or equivalently
\begin{equation}\label{MassConservation2}
\sum_{k=1}^I k{f_k}=\mbox{const}.
\end{equation}
\begin{figure}
\centering
\includegraphics[width=.9\linewidth]{Decoupling}
\caption{We decouple the system \eqref{DiscreteQuantum} into rays}
\label{Fig2}
\end{figure}
Also by abuse of notation, we denote this discrete version of $C_{12}$ by
\begin{equation}\label{DiscreteC12}
\begin{aligned}
\mathcal{C}_{12}[f_{k_1}]:= &~~\sum_{k_2+k_3=k_1}\mathcal{K}^{12}_{k_1,k_2,k_3}[(f_{k_1}+1)f_{k_2}f_{k_3}-f_{k_1}(f_{k_2}+1)(f_{k_3}+1)]\\
&~~-2\sum_{k_1+k_3=k_2}\mathcal{K}^{12}_{k_2,k_1,k_3}[(f_{k_2}+1)f_{k_1}f_{k_3}-f_{k_2}(f_{k_1}+1)(f_{k_3}+1)].
\end{aligned}
\end{equation}
\subsection{The chemical reaction network associated to $C_{12}$}\label{Sec:ChemC12}
For $x\in\mathbb{R}_{>0}^n$ and $\alpha\in \mathbb{R}_{\geq 0}^n$, we denote by $x^\alpha$ the monomial $\Pi_{i=1}^n x_i^{\alpha_i}$.
\begin{definition} Consider a chemical reaction of the form
$$\ce{\alpha_1 X_1 + \alpha_2 X_2 + ... + \alpha_n X_n ->[\mathcal{K}] \beta_1 X_1 + \beta_2 X_2 + ... + \beta_n X_n},$$
where $\mathcal{K}$ is a positive parameter, called reaction rate constant. Then the mass-action dynamical system generated by this reaction is
\begin{equation}\label{ExampleEquation}
\dot{x}= \mathcal{K} x^\alpha (\beta-\alpha),
\end{equation}
where $\alpha=(\alpha_1,\cdots,\alpha_n)^T$, $\beta=(\beta_1,\cdots,\beta_n)^T$, $\alpha_i,\beta_i>0$ and $x=(x_1,\cdots,x_n)^T$, in which $x_i$ is the concentration of the chemical species $X_i$. For the case of a network that contains several reactions
$$\ce{\alpha_1^j X_1^j + \alpha_2^j X_2^j + ... + \alpha_n^j X_n^j ->[\mathcal{K}_j] \beta_1 X_1^j + \beta_2^j X_2^j + ... + \beta_n^j X_n^j},$$
for $1\le j\le m$, its associated mass-action dynamical system is given by
\begin{equation}\label{DynSys}
\dot{x}=\sum_{j=1}^m \mathcal{K}_jx^{\alpha^j}(\beta^j-\alpha^j).
\end{equation}
\end{definition}
In this section, we will show that the system \eqref{DiscreteQuantum1D} has the form \eqref{DynSys} for a well-chosen set of reactions.
If $y\to y'$ and $y'\to y$ are reactions, we combine them together into a {\it ``reversible'' reaction $y\leftrightarrow y'$.}
We will derive the system \eqref{DiscreteQuantum1D} from the network of chemical reactions of the form:
\begin{eqnarray}\label{BioChemEq01}
&&\ce{X_{k_2} + X_{k_3} <-> X_{k_1}}\\\label{BioChemEq02}
&&\ce{X_{k_2} + X_{k_1} -> 2X_{k_2} + X_{k_3}},
\end{eqnarray}
for all $k_1,k_2,k_3$ in $\mathbb{I}$ such that $k_2+k_3=k_1$. If we denote by $F_k$ the concentration of the species $X_k$, we will show that, {\it for appropriate choices of the reaction rate constants} in \eqref{BioChemEq01} and \eqref{BioChemEq02}, the differential equations satisfied by $F_k$ according the mass-action kinetics are exactly the same as \eqref{DiscreteQuantum1D}.\\
In order to describe the connection between the mass-action system given by reactions of the form \eqref{BioChemEq01}-\eqref{BioChemEq02} and our system \eqref{DiscreteQuantum1D}, we need to consider several cases.
\bigskip
{\it Case 1:} For $k_2+k_3=k_1$, $k_2\ne k_3$, $k_1,k_2,k_3\in\mathbb{I}$, we consider
\begin{eqnarray}\label{BioChemEq1a}
&&\ce{X_{k_2} + X_{k_3} <->[2\mathcal{K}^{12}_{k_1,k_2,k_3}] X_{k_1}}\\\label{BioChemEq1b}
&&\ce{X_{k_2} + X_{k_1} ->[2\mathcal{K}^{12}_{k_1,k_2,k_3}] 2X_{k_2} + X_{k_3}},
\end{eqnarray}
and for the reversible reaction \eqref{BioChemEq1a} the forward and backward rate constants are the same, i.e., we choose the reaction rate constants of the three reactions $X_{k_2}+X_{k_3}\to X_{k_1}$, $X_{k_1}\to X_{k_2}+X_{k_3}$, $X_{k_2}+X_{k_1}\to 2X_{k_2}+X_{k_3}$ to be $2\mathcal{K}^{12}_{k_1,k_2,k_3}$.
For example, consider the reversible reaction \eqref{BioChemEq1a}:
in this reaction, $X_{k_1}$ is created from $X_{k_2}+X_{k_3}$ with the rate $2\mathcal{K}^{12}_{k_1,k_2,k_3}F_{k_2}F_{k_3}$ and $X_{k_1}$ is decomposed into $X_{k_2}+X_{k_3}$ with the rate $-2\mathcal{K}^{12}_{k_1,k_2,k_3}F_{k_1}$. Therefore, the rate of change of the species $X_{k_1}$ due to this reaction is $ 2\mathcal{K}^{12}_{k_1,k_2,k_3}[F_{k_2}F_{k_3} - F_{k_1}]$.
For the irreversible reaction \eqref{BioChemEq1b},
$X_{k_1}$ is lost with the rate $-2\mathcal{K}^{12}_{k_1,k_2,k_3}F_{k_2}F_{k_1}$ to create $2X_{k_2}+X_{k_3}$. Therefore the rate of change of the species $X_{k_1}$ due to this reaction is $ -2\mathcal{K}^{12}_{k_1,k_2,k_3}F_{k_2}F_{k_1}$. By exchanging the roles of $X_{k_2}$ and $X_{k_3}$ in \eqref{BioChemEq1b}, we obtain the rate $ -2\mathcal{K}^{12}_{k_1,k_2,k_3}[F_{k_2}F_{k_1}+F_{k_3}F_{k_1}]$.
Therefore, the total rate of change of $X_{k_1}$ due to the reactions in \eqref{BioChemEq1a}-\eqref{BioChemEq1b} is
\begin{equation}\label{FirstChangeRate}
\begin{aligned}
&2\mathcal{K}^{12}_{k_1,k_2,k_3}[F_{k_2}F_{k_3}- F_{k_1}-F_{k_2}F_{k_1}-F_{k_3}F_{k_1}].
\end{aligned}
\end{equation}
\bigskip
{\it Case 2:} For $2k_2=k_1$, $k_1,k_2\in\mathbb{I}$, we consider
\begin{eqnarray}\label{BioChemEq1a2}
&&\ce{2X_{k_2} <->[\mathcal{K}^{12}_{k_1,k_2,k_3}] X_{k_1}}\\\label{BioChemEq1b2}
&&\ce{X_{k_2} + X_{k_1} ->[2\mathcal{K}^{12}_{k_1,k_2,k_3}] 3X_{k_2} }.
\end{eqnarray}
We choose the reaction rate constant of $2X_{k_2}\to X_{k_1}$ and the reaction rate constant of $X_{k_1}\to 2 X_{k_2}$ to be $\mathcal{K}^{12}_{k_1,k_2,k_3}$. Also, we choose the reaction rate constant of $X_{k_2}+X_{k_1}\to 3X_{k_2}$ to be $2\mathcal{K}^{12}_{k_1,k_2,k_3}$.
Consider the first reaction \eqref{BioChemEq1a2}:
In this reaction, $X_{k_1}$ is created from $2X_{k_2}$ with the rate $\mathcal{K}_{k_1,k_2,k_2}F_{k_2}^2$ and $X_{k_1}$ is decomposed into $2X_{k_2}$ with the rate $-\mathcal{K}^{12}_{k_1,k_2,k_2}F_{k_1}$. The rate of change of the species $X_{k_1}$ is $\mathcal{K}^{12}_{k_1,k_2,k_2}[F_{k_2}^2 -F_{k_1}]$.
For the second reaction \eqref{BioChemEq1b2}:
$X_{k_1}$ is lost with the rate $-2\mathcal{K}^{12}_{k_1,k_2,k_2}F_{k_2}F_{k_1}$ to create $3X_{k_2}$.
As a result, the rate of change of $X_{k_1}$ due to the reactions \eqref{BioChemEq1a2}-\eqref{BioChemEq1b2} is
\begin{equation}\label{FirstChangeRate2}
\begin{aligned}
&\mathcal{K}^{12}_{k_1,k_2,k_3}[F_{k_2}^2- F_{k_1}-2F_{k_2}F_{k_1}].
\end{aligned}
\end{equation}
\bigskip
{\it Case 3:} Next, for $k_2=k_3+k_1$, $k_1\ne k_3$, $k_1,k_2,k_3\in\mathbb{I}$, let us look at the rate of change of $X_{k_1}$ in
\begin{eqnarray}\label{BioChemEq2a}
&&\ce{X_{k_1} + X_{k_3} <->[2\mathcal{K}^{12}_{k_2,k_1,k_3}] X_{k_2}}\\\label{BioChemEq2b}
&&\ce{X_{k_2} + X_{k_1} ->[2\mathcal{K}^{12}_{k_2,k_1,k_3}] 2X_{k_1} + X_{k_3}}\\\label{BioChemEq2c}
&&\ce{X_{k_2} + X_{k_3} ->[2\mathcal{K}^{12}_{k_2,k_1,k_3}] X_{k_1} + 2X_{k_3}},
\end{eqnarray}
For \eqref{BioChemEq2a}, the rate of change of $X_{k_1}$ is $2\mathcal{K}^{12}_{k_2,k_1,k_3}[F_{k_2}-F_{k_1}F_{k_3}]$. For \eqref{BioChemEq2b}, the rate of change of $X_{k_1}$ is $2\mathcal{K}^{12}_{k_2,k_1,k_3}F_{k_1}F_{k_2}.$ By exchanging the roles of $X_1$ and $X_3$, we obtain the rate $2\mathcal{K}^{12}(k_2,k_1,k_3)[F_{k_1}F_{k_2}+F_{k_2}F_{k_3}]$.\\
Therefore, the rate of change of $X_{k_1}$ due to reactions in \eqref{BioChemEq2a}-\eqref{BioChemEq2c} is
\begin{equation}\label{SecondChangeRate}
\begin{aligned}
&-2\mathcal{K}^{12}_{k_2,k_1,k_3}[F_{k_1}F_{k_3}- F_{k_2}-F_{k_2}F_{k_3}-F_{k_1}F_{k_2}].
\end{aligned}
\end{equation}
\bigskip
{\it Case 4:} Now, for $k_2=2k_1$, $k_1,k_2\in\mathbb{I}$, let us look at the rate of change of $X_{k_1}$ in
\begin{eqnarray}\label{BioChemEq2a2}
&&\ce{2X_{k_1} <->[\mathcal{K}^{12}_{k_2,k_1,k_1}] X_{k_2}}\\\label{BioChemEq2b2}
&&\ce{X_{k_2} + X_{k_1} ->[2\mathcal{K}^{12}_{k_2,k_1,k_3}] 3X_{k_1}},
\end{eqnarray}
For \eqref{BioChemEq2a2}, the rate of change of $X_{k_1}$ is $2\mathcal{K}^{12}_{k_2,k_1,k_3}[F_{k_2}-F_{k_1}^2]$.
For \eqref{BioChemEq2b2}, the rate of change of $X_{k_1}$ is $4\mathcal{K}^{12}_{k_2,k_1,k_3}F_{k_1}F_{k_2}.$
Therefore, the rate of change of $X_{k_1}$ due to the reactions \eqref{BioChemEq2a2}-\eqref{BioChemEq2b2} is
\begin{equation}\label{SecondChangeRate2}
\begin{aligned}
&-2\mathcal{K}^{12}_{k_2,k_1,k_3}[F_{k_1}^2- F_{k_2}-2F_{k_1}F_{k_2}].
\end{aligned}
\end{equation}
From \eqref{FirstChangeRate}, \eqref{FirstChangeRate2}, \eqref{SecondChangeRate}, \eqref{SecondChangeRate2},
the total rate of change of $X_{k_1}$ is
\begin{equation}\label{TotalChangeRate}
\begin{aligned}
&~~\sum_{k_2+k_3=k_1, k_2< k_3}2\mathcal{K}^{12}_{k_1,k_2,k_3}[(F_{k_1}+1)F_{k_2}F_{k_3}-F_{k_1}(F_{k_2}+1)(F_{k_3}+1)]\\
&~~+\sum_{2k_2=k_1}\mathcal{K}^{12}_{k_1,k_2,k_2}[(F_{k_1}+1)F_{k_2}F_{k_2}-F_{k_1}(F_{k_2}+1)(F_{k_2}+1)]\\
&~~-\sum_{k_1+k_3=k_2}2\mathcal{K}^{12}_{k_2,k_1,k_3}[(F_{k_2}+1)F_{k_1}F_{k_3}-F_{k_2}(F_{k_1}+1)(F_{k_3}+1)],
\end{aligned}
\end{equation}
which can be written as
\begin{equation}\label{TotalChangeRateEquation}
\begin{aligned}
\dot{F}_{k_1}\ =&~~\sum_{k_2+k_3=k_1}\mathcal{K}^{12}_{k_1,k_2,k_3}[(F_{k_1}+1)F_{k_2}F_{k_3}-F_{k_1}(F_{k_2}+1)(F_{k_3}+1)]\\
&~~-2\sum_{k_1+k_3=k_2}\mathcal{K}^{12}_{k_2,k_1,k_3}[(F_{k_2}+1)F_{k_1}F_{k_3}-F_{k_2}(F_{k_1}+1)(F_{k_3}+1)],
\end{aligned}
\end{equation}
which shows that the system of differential equations satisfied by the concentrations $F_k$ is exactly the same as the system of differential equations \eqref{DiscreteQuantum1D} satisfied by the densities $f_k$.
\subsection{A change of variables}
In this section, we introduce a change of variables that will help us to investigate the dynamics of the system \eqref{TotalChangeRateEquation}.
Define $$G_k=\frac{F_k}{F_k+1},$$
then
$$F_k=\frac{G_k}{1-G_k},$$
and
$$F_{k_3}+F_{k_1}F_{k_3}+F_{k_2}F_{k_3}-F_{k_1}F_{k_2}=\frac{G_{k_3}-G_{k_1}G_{k_2}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})},$$
$$F_{k_1}+F_{k_1}F_{k_2}+F_{k_1}F_{k_3}-F_{k_3}F_{k_2}=\frac{G_{k_1}-G_{k_2}G_{k_3}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})}.$$
Notice that $0<F_k<\infty$ and $0<G_k<1$.\\
The system \eqref{TotalChangeRateEquation} is converted into
\begin{eqnarray}\label{DiscreteQuantumConverted}
\frac{\dot{G}_{k_1}}{(1-G_{k_1})^2}\nonumber
&=& \tilde{\mathcal{C}}_{12}[G](k_1):=2\sum_{k_1+k_2=k_3}\mathcal{K}^{12}_{k_1,k_2,k_3}\frac{G_{k_3}-G_{k_1}G_{k_2}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})}\\
&+&\sum_{k_1=k_2+k_3}\mathcal{K}^{12}_{k_1,k_2,k_3}\frac{-G_{k_1}+G_{k_2}G_{k_3}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})}, \forall k_1\in\mathbb{I}.
\end{eqnarray}
Suppose that $G$ represents the column vector $(G_1,\dots,G_I)^T$. Let us also denote by $X_k$, with an abuse of notation, the vector
$$\left( {\begin{array}{cc}
0\\ \cdots\\ 1\\ \cdots \\ 0 \end{array} } \right),$$
in which the only element that different from $0$ is the $k$-th one.
Also, for $k_1\ne k_2$, we denote
$$K_{X_{k_1}+X_{k_2} \rightarrow X_{k_3}}(G):=2\mathcal{K}^{12}_{k_1,k_2,k_3}\frac{G_{k_1}G_{k_2}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})},$$
$$K_{X_{k_3} \rightarrow X_{k_1}+X_{k_2}}(G):=2\mathcal{K}^{12}_{k_1,k_2,k_3}\frac{G_{k_3}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})},$$
$$\mathcal{K}_{X_{k_1}+X_{k_2}\leftrightarrow X_{k_3}}:=2\mathcal{K}^{12}_{k_1,k_2,k_3}.$$
Otherwise, if $k_1= k_2$ , we denote
$$K_{2X_{k_1}\rightarrow X_{k_3}}(G):=\mathcal{K}^{12}_{k_1,k_1,k_3}\frac{G_{k_1}G_{k_2}}{(1-G_{k_1})^2(1-G_{k_3})},$$
$$K_{X_{k_3} \rightarrow 2X_{k_1}}(G):=\mathcal{K}^{12}_{k_1,k_1,k_3}\frac{G_{k_3}}{(1-G_{k_1})^2(1-G_{k_3})},$$
$$\mathcal{K}_{2X_{k_1}\leftrightarrow X_{k_3}}:=2\mathcal{K}^{12}_{k_1,k_1,k_3}.$$
Using these notations, the system \eqref{DiscreteQuantumConverted} could be rewritten as:
\begin{eqnarray}\label{EqG}
{\dot{G}}
&=&\mbox{diag}\left( {\begin{array}{cc}
{(1-G_1)^2}\\ \cdots \\ {(1-G_I)^2} \end{array} } \right)\times\\\nonumber
& &\times\sum_{k_1+k_2=k_3}\left[K_{X_{k_1}+X_{k_2} \rightarrow X_{k_3}}(G)-K_{X_{k_3} \rightarrow X_{k_1}+X_{k_2} }(G)\right](X_{k_3}-X_{k_1}-X_{k_2} ).
\end{eqnarray}
Equivalently, we can also write
\begin{equation}\label{EqG2}
{\dot{G}}
=\mbox{diag}\left( {\begin{array}{cc}
{(1-G_1)^2}\\ \cdots \\ {(1-G_I)^2} \end{array} } \right)\sum_{y\leftrightarrow y'}\left[K_{y \rightarrow y'}(G)-K_{y' \rightarrow y}(G)\right](y'-y ),
\end{equation}
where $y\leftrightarrow y'$ belongs to the set of reversible reactions
\begin{eqnarray}\label{ChemicalNetworkReversible1a}
X_{k_1}+X_{k_2}&\longleftrightarrow &X_{k_3},
\end{eqnarray}
with $k_1+k_2=k_3$.
\subsection{Convergence to equilibrium}
\begin{theorem}\label{TheoremQ12} For any positive initial condition, the solution $$f(t)=(f_p(t))_{p\in\mathcal{L}_R}$$ of the discrete quantum Boltzmann equation \eqref{DiscreteQuantum} converges to an equilibrium state $f^*=(f_p^*)_{p\in\mathcal{L}_R}$. For each ray $\{kP_0\}_{k\geq 1}$ there exists a positive constant $\rho({P_0})$ such that if $p=kP_0$ then $$f^*_p=\frac{1}{e^{k\rho({P_0})}-1}.$$ Moreover, the solution $f(t)$ of \eqref{DiscreteQuantum} converges to $f^*$ exponentially fast in the following sense: there exist positive constants $C_1$, $C_2$ such that $$\max_{p\in\mathcal{L}_R}|f_p(t)-f_p^*|<C_1e^{-C_2t}.$$
\end{theorem}
\begin{proof}
By using the decoupling and the change of variables discussed in the previous sections, for each ray $\{kP_0\}_{k\geq 1}$, we can reduce the study of $f$ to $F$, which satisfies \eqref{TotalChangeRateEquation}. From $F$, we can switch to study $G$, which is the solution of \eqref{EqG2}.
\bigskip
{\\\it Step 1: The Lyapunov function.}
We recall that \eqref{DiscreteQuantumConverted} could be rewritten under the form
\begin{equation}\label{EQ}
{\dot{G}}=\mbox{diag}\left( {\begin{array}{cc}
{(1-G_1)^2}\\ \cdots \\ {(1-G_I)^2} \end{array} } \right)\sum_{y \leftrightarrow y'}\left[K_{y \to y'}(G)-K_{y'\to y}(G)\right](y'-y).
\end{equation}
We define the function
\begin{equation}\label{Lyapunov}
L(G)=\sum_{k=1}^I \left(\log (1-G_k)+\frac{G_k \log G_k}{1-G_k}-\frac{\log G_k^*}{1-G_k}\right),
\end{equation}
where $G^*_k=\frac{1}{e^{k\rho}}$, for some $\rho>0$,
and we will show that $L$ is a Lyapunov function for the system \eqref{DiscreteQuantumConverted}.
We have
\begin{equation}
\nabla L=\left( {\begin{array}{cc}
\frac{1}{(1-G_1)^2}\log \frac{G_1}{G^*_1}\\ \cdots \\ \frac{1}{(1-G_I)^2}\log \frac{G_I}{G^*_I} \end{array} } \right),
\end{equation}
which implies that
\begin{eqnarray}\label{eq1}
\mbox{diag}\left( {\begin{array}{cc}
{(1-G_1)^2}\\ \cdots \\ {(1-G_I)^2} \end{array} } \right)\cdot(y'-y)\cdot\nabla L&=&\log \left(\frac{G}{G^*}\right)^{y'-y}\\\nonumber
& = &\log\left(\frac{G}{G^*}\right)^{y'}-\log\left(\frac{G}{G^*}\right)^{y}.
\end{eqnarray}
If we define
$$\mathcal{H}_{y,y'}(G)=\frac{K_{y\to y'}(G)}{\mathcal{K}_{y\leftrightarrow y'}G^y},$$
then $\mathcal{H}_{y,y'}=\mathcal{H}_{y',y}$ for $y$ and $y'$ as in \eqref{ChemicalNetworkReversible1a}. Moreover,
we have
\begin{equation}\label{eq2}
\begin{aligned}
& K_{y\to y'}(G)-K_{y'\to y}(G)=\\
=& \mathcal{K}_{y\leftrightarrow y'}G^y \mathcal{H}_{y,y'}(G)-\mathcal{K}_{y\leftrightarrow y'}G^{y'} \mathcal{H}_{y,y'}(G)\\
=&\mathcal{K}_{y\leftrightarrow y'}\mathcal{H}_{y,y'}(G)[G^y-G^{y'}]\\
=&\mathcal{K}_{y\leftrightarrow y'}(G^*)^y \mathcal{H}_{y,y'}(G)\left[\frac{G^y }{(G^*)^y}
-\frac{G^{y'} }{(G^*)^{y'} }\right],
\end{aligned}
\end{equation}
since $(G^*)^y=(G^*)^{y'}$.\\
Combining \eqref{EQ}, \eqref{eq1} and \eqref{eq2}, we obtain
\begin{equation}
\begin{aligned}
& \ \dot{G}\cdot\nabla L = \\
= & \sum_{y\leftrightarrow y'}\left[\log\left(\frac{G}{G^*}\right)^{y'}-\log\left(\frac{G}{G^*}\right)^{y}\right] \mathcal{K}_{y\leftrightarrow y'}(G^*)^y \mathcal{H}_{y,y'}(G)\left[\frac{G^y }{(G^*)^y}
-\frac{G^{y'} }{(G^*)^{y'} }\right]\\
\leq & \ 0,
\end{aligned}
\end{equation}
since $\mathrm{log}$ is an increasing function.
Also, note that the above inequality is strict unless
\begin{equation}\label{C12equilibrium}
\frac{G^y }{(G^*)^y}
=\frac{G^{y'} }{(G^*)^{y'} },\end{equation}
for all reactions $y \leftrightarrow y'.$
\\Since $(G^*)^y=(G^*)^{y'}$ for all reactions $y \leftrightarrow y',$ this implies $G^{*}_{k_1}\cdot G^{*}_{k_2} = G^*_{k_1+k_2}$ for all $k_1$ and $k_2$ such that $k_1+k_2\le I$. As a consequence $G^*_k=e^{-\rho k}$, for some positive constant $\rho$. Moreover, \eqref{C12equilibrium} implies that at equilibrium $(G)^y=(G)^{y'}$ for all reactions $y \leftrightarrow y',$ which leads to $G_k=e^{-\rho' k}$, for some positive constant $\rho'$.
\\ By the conservation relation
$$\sum_{k=1}^Ik\frac{G_k}{1-G_k}=\sum_{k=1}^Ik\frac{G_k^*}{1-G_k^*},$$
we deduce that
$$\sum_{k=1}^Ik\frac{e^{-\rho k}}{1-e^{-\rho k}}=\sum_{k=1}^Ik\frac{e^{-\rho' k}}{1-e^{-\rho' k}}.$$
By the monotonicity of the function $\rho\to\frac{e^{-\rho k}}{1-e^{-\rho k}}$, we conclude that $\rho=\rho'$, i.e., $G^*$ is the only equilibrium point that satisfies the same conservation relation as the initial condition.
Now, we will prove that there exists exactly one critical point of the Lyapunov function $L$ within each invariant set $$\mathfrak{S}_{c}:=\left\{\sum_{k=1}^Ik\frac{G_k}{1-G_k}=c\right\}.$$
Since
$$\nabla L=\mbox{diag}\left( {\begin{array}{cc}
\frac{1}{(1-G_1)^2}\\ \cdots \\ \frac{1}{(1-G_I)^2} \end{array} } \right)[\log G-\log G^*],$$
the projection of $\nabla L$ on the tangent space to the set $\mathfrak{S}_{c}$ is $0$ if and only if there exists a constant $\varrho$ such that
$$\nabla L=\varrho \cdot \nabla \left(\sum_{k=1}^Ik\frac{G_k}{1-G_k}\right),$$
which is equivalent with $$\mbox{diag}\left( {\begin{array}{cc}
\frac{1}{(1-G_1)^2}\\ \cdots \\ \frac{1}{(1-G_I)^2} \end{array} } \right)[\log G-\log G^*]=\varrho\left( {\begin{array}{cc}
\frac{1}{(1-G_1)^2}\\ \cdots \\ \frac{1}{(1-G_I)^2} \end{array} } \right).$$
A direct consequence of the above is the following system of identities
\begin{eqnarray*}
\log G_1 - \log G_1^* &=& \varrho,\\
\log G_2 - \log G_2^* &=& 2\varrho,\\
&\cdots&\\
\log G_I - \log G_I^* &=& I\varrho,
\end{eqnarray*}
yielding
$$\frac{G_k}{G_k^*}=e^{k\varrho}, \ \ \ \forall k\in \{1,\cdots,I\}.$$
Moreover, since $G_k$ and $G_k^*$ satisfy the same conservation law then it follows that $G=G^*$. This implies that $G^*$ is the only critical point of $L$ on the invariant set $\mathfrak{S}_c$.
\bigskip
{\\\it Step 2: Differential inclusions and persistence.}
Now let us observe that \eqref{DiscreteQuantum1D} could be regarded as a $\mathcal{K}$-variable mass-action system for the reversible network \eqref{ChemicalNetworkReversible1a}. For this we write
$$F_{k''}+F_kF_{k''}+F_{k'}F_{k''}=(1+F_k+F_{k'})F_{k''},$$
and note that $1+F_k+F_{k'}$ is bounded below by $1$ and above by $1+2C$, where
$$C=\sum_{k=1}^I kF_k.$$
Therefore, the results of \cite{Craciun:2015:TDI} about persistence of $\mathcal{K}$-variable reversible mass-action systems can be applied and we conclude that the system is persistent. Alternatively, we can also use the Petri net argument of \cite{AngeliDeLeenheerSontag:PRF:2011}, to prove that the system is persistent, as follows. Note that $F_k$ is the density function of the species $X_k$. It is straightforward that each {\it siphon} is $\{X_1, X_2,\cdots, X_I\}$, which contains the support of the {\it $P$-semiflow} (see \cite{AngeliDeLeenheerSontag:PRF:2011} for the definition of siphons and P-semiflows) given by
$$\sum_{k=1}^IkF_k=\mbox{constant}.$$
As a result, the Petri net theory developed in \cite{AngeliDeLeenheerSontag:PRF:2011} can be applied and it follows that the system is persistent.
Therefore, by using the existence of the globally defined strict Lyapunov function $L$, and the LaSalle invariance principle, it follows that all trajectories converge to the unique positive equilibrium $G^*$ that we discussed in~Step~1.
\bigskip
{\\\\\it Step 3: Exponential rate of convergence.}
Define
\begin{eqnarray}\label{DefineR}\nonumber
&&\mathcal{R}(G)=\\
&=&\mbox{diag}\left( {\begin{array}{cc}
{(1-G_1)^2}\\ \cdots \\ {(1-G_I)^2} \end{array} } \right)\sum_{y \leftrightarrow y'}\left[K_{y \to y'}(G)-K_{y'\to y}(G)\right](y'-y)\\\nonumber
&=&\mbox{diag}\left( {\begin{array}{cc}
{(1-G_1)^2}\\ \cdots \\ {(1-G_I)^2} \end{array} } \right)\sum_{y \leftrightarrow y'}[\mathcal{K}_{y\leftrightarrow y'}G^y -\mathcal{K}_{y\leftrightarrow y'}G^{y'}] \mathcal{H}_{y,y'}(G)(y'-y),
\end{eqnarray}
and define
\begin{eqnarray}\label{DefineRR}\nonumber
\mathcal{S}(G)
&=&\sum_{y \leftrightarrow y'}[\mathcal{K}_{y\leftrightarrow y'}G^y -\mathcal{K}_{y\leftrightarrow y'}G^{y'}]\mathcal{H}_{y,y'}(G)(y'-y).
\end{eqnarray}
\\ Following \cite{CraciunFeinberg:2005:MEI}, we compute the Jacobian of $\mathcal{S}$ at the equilibrium point $G^*$, applied to an arbitrary vector $\delta\ne 0$ that belongs to the span of the vectors $y'-y$
\begin{eqnarray}\label{JacR}
\mbox{Jac}(\mathcal{S}(G^*))\delta =\sum_{y \leftrightarrow y'} \mathcal{K}_{y\leftrightarrow y'}(G^*)^y ((y-y')*\delta)\mathcal{H}_{y,y'}(G^*)(y-y'),
\end{eqnarray}
in which the inner product $*$ is defined as
$$y*\delta=\sum_{1}^I \frac{y_k \delta_k}{G_k}.$$
Therefore
\begin{eqnarray}\label{JacR1}
&& [\mbox{Jac}(\mathcal{S}(G^*))\delta]*\delta =\\\nonumber
& =& \sum_{y \leftrightarrow y'} \mathcal{K}_{y\leftrightarrow y'}(G^*)^y\mathcal{H}_{y,y'}(G^*)[(y-y')*\delta][(y'-y)*\delta]<0.
\end{eqnarray}
Now, we compute the Jacobian of $\mathcal{R}$ at the equilibrium point $G^*$,
\begin{eqnarray}\nonumber
& &\mbox{Jac}(\mathcal{R}(G^*))\\\nonumber
&=&\mbox{diag}\begin{bmatrix}
\partial_{G_1} {(1-G^*_1)^2}\mathcal{S}(G^*)_1\\\cdots \\ \partial_{G_I} {(1-G^*_I)^2}\mathcal{S}(G^*)_I\end{bmatrix}+\mbox{diag}\begin{bmatrix}
{(1-G^*_1)^2}\\\cdots \\ {(1-G^*_I)^2}\end{bmatrix} \mbox{Jac}(\mathcal{S}(G^*))\\\nonumber
&=&\mbox{diag}\begin{bmatrix}
{(1-G^*_1)^2} \\\cdots \\ {(1-G^*_I)^2} \end{bmatrix} \mbox{Jac}(\mathcal{S}(G^*)),
\end{eqnarray}
where the second equality is due to the fact that since $G^*$ is an equilibrium we have that $\mathcal{S}(G^*)=0$.\\
Since $$D:=\mbox{diag}\begin{bmatrix}
{(1-G^*_1)^2} \\\cdots \\ {(1-G^*_I)^2}\end{bmatrix} $$
is a diagonal matrix and $ A:=\mbox{Jac}(\mathcal{S}(G^*))$ is negative definite, then $D^{1/2}AD^{1/2}$ is also negative definite with respect to this inner product. Since
$$\mbox{det}(DA-\lambda Id)=\mbox{det}(D^{1/2}AD^{1/2}-\lambda Id),~~\forall \lambda\in\mathbb{R},$$
it follows that $D^{1/2}AD^{1/2}$ and $DA$ have the same eigenvectors, so $DA$ is negative definite. In other words, $\mbox{Jac}(\mathcal{R}(G^*))$ is negative definite.
The exponential rate of convergence
$$\max\{|G_1(t)-G_1^*|,\cdots, |G_I(t)-G_I^*|\}\leq C_1 e^{-C_2 t}.
$$
then follows from the fact that the Jacobian above is negative definite. This leads to the conclusion of the theorem.
\end{proof}
\begin{remark}
The Lyapunov function \eqref{Lyapunov} in the variable $F$ reads
\begin{equation}\label{LyapunovF}
L(F)=\sum_{k=1}^I [F_k\log F_k-(1+F_k)\log (1+F_k)+(\log (F_{k}^*+1)-\log F_k^*)(F_{k}+1)],
\end{equation}
and it is a strictly convex function.
\end{remark}
\begin{remark}
If the intersection between the ray $\{kP_0\}_{k\geq 1}$ and $\mathcal{L}_R$ contains a single point, then the solution $f(t)$ of \eqref{DiscreteQuantum} has $f_{P_0}\equiv 0$, so $f_{P_0}\equiv$ constant.
\end{remark}
\section{A reaction network approach for the case of $C_{13}$ and $C_{22}$}\label{Sec:Qnm}
\subsection{The dynamical system associated to $C_{13}$}
As we discussed in the Introduction, we are also interested in the dynamics given by the discrete model of the collision operator $C_{13}$, described in \eqref{QuantumBoltzmannLinda13}.
\\\\ Let $\mathcal{L}_R$ denote the lattice of integer points
$$\mathcal{L}_R=\{p\in\mathbb{Z}^3~~~~|~~~~|p|<R\}.$$
The discretized quantum Boltzmann equation for $C_{13}$ reads
\begin{equation}\label{DiscreteQuantuma}\begin{aligned}
&\dot{f}_{p_1}=C_{13}^D[f_{p_1}]:=\\
:=
&\quad\sum_{\substack{p_2,p_3,p_4\in\mathcal{L}_R,\\ p_1=p_2+p_3+p_4,\\ \mathcal{E}(p_1)=\mathcal{E}(p_{2})+\mathcal{E}(p_{3})+\mathcal{E}(p_{4})}}\mathcal{K}^{13}_{p_1,p_2,p_3,p_4}\{(f_{p_1}+1)f_{p_2}f_{p_3}f_{p_4}-(f_{p_2}+1)(f_{p_3}+1)(f_{p_4}+1)f_{p_1}\}\\
&\quad -3\sum_{\substack{p_2,p_3,p_4\in\mathcal{L}_R,\\ p_2=p_1+p_3+p_4,\\ \mathcal{E}(p_2)=\mathcal{E}(p_{1})+\mathcal{E}(p_{3})+\mathcal{E}(p_{4})}}\mathcal{K}^{13}_{p_2,p_1,p_3,p_4}\left\{(f_{p_2}+1)f_{p_1}f_{p_3}f_{p_4}-(f_{p_1}+1)(f_{p_3}+1)(f_{p_4}+1)f_{p_2}\right\},~~
\end{aligned}\end{equation}
for all $p_1$ in $\mathcal{L}_R$, where $\mathcal{E}(p)$ is defined in \eqref{E3a}.
Similar to the $C_{12}$ case, when $p=0$, $\mathcal{K}^{13}_{p_1,p_2,p_3,p_4}=0$, and we obtain
\begin{eqnarray*}\label{DiscreteQuantumIndex0}
\dot{f}_0=0,
\end{eqnarray*}
which means $f_0(t)$ is a constant for all time $t$, and we can assume $f_0(t)=0$ for all $t$.
\\ Since in the first sum of \eqref{DiscreteQuantuma}, we consider $(p_1,p_2,p_3,p_4)$ satisfying
\begin{equation}\label{EnergyC13}
p_1 \ = \ p_2 \ + \ p_3\ +\ p_4, \ \ \ \ \mathcal{E}(p_1)\ = \ \mathcal{E}(p_2)\ +\ \mathcal{E}(p_3) \ + \mathcal{E}(p_4),
\end{equation}
we infer that there exists a vector $P$ and $k_1$, $k_2$, $k_3, k_4\geq0$, $k_1, k_2, k_3, k_4\in\mathbb{Z}$ such that
$$p_1=k_1 P;~~~p_2=k_2 P;~~~p_3=k_3 P;~~~p_4=k_4 P;~~~k_1=k_2+k_3+k_4.$$ Using the same arguments as the case of $C_{12}$, we can deduce that Equation \eqref{DiscreteQuantuma} for $C_{13}$ is equivalent with the following family of decoupled systems for $k_1\in\mathbb{I}=\{1,2,\dots,I\}$ where $P$ is the closest point to the origin among the lattice points on its ray:
\begin{equation}\label{DiscreteC13Pre}\begin{aligned}
&\dot{f}_{k_1P}=\\
=&\quad\sum_{\substack{k_2,k_3,k_4\in
\mathbb{I},\\ k_1=k_2+k_3+k_4}}\mathcal{K}^{13}_{k_1P,k_2P,k_3P,k_4P}\{(f_{k_1P}+1)f_{k_2P}f_{k_3P}f_{k_4P}\\
& -f_{k_1P}(f_{k_2P}+1)(f_{k_3P}+1)(f_{k_4P}+1)\}\\
&\quad-{3}\sum_{\substack{k_2,k_3,k_4\in
\mathbb{I},\\ k_2=k_1+k_3+k_4}}\mathcal{K}^{13}_{k_2P,k_1P,k_3P,k_4P}\{(f_{k_2P}+1)f_{k_1P}f_{k_3P}f_{k_4P}\\
& -f_{k_2P}(f_{k_1P}+1)(f_{k_3P}+1)(f_{k_4P}+1)\}.\end{aligned}
\end{equation}
Denoting $f_{kP}$ by $F_k$ (with $k\in \mathbb{I}$) and $\mathcal{K}^{12}_{k_1P,k_2P,k_3P,k_4P}$ by $\mathcal{K}^{12}_{k_1,k_2,k_3,k_4}$, we obtain
\begin{equation}\label{DiscreteC13}\begin{aligned}
\dot{F}_{k_1}
=&\ \mathcal{C}_{13}[F](k_1)=\quad\sum_{k_1=k_2+k_3+k_4}\mathcal{K}^{13}_{k_1,k_2,k_3,k_4}\{(F_{k_1}+1)F_{k_2}F_{k_3}F_{k_4}-\\
& -F_{k_1}(F_{k_2}+1)(F_{k_3}+1)(F_{k_4}+1)\}
\\
&\quad-3\sum_{k_1+k_2+k_3=k_4}\mathcal{K}^{13}_{k_1,k_2,k_3,k_4}\{(F_{k_4}+1)F_{k_1}F_{k_2}F_{k_3}-\\
&-F_{k_4}(F_{k_1}+1)(F_{k_2}+1)(F_{k_3}+1)\}, ~\forall k_1\in\mathbb{I}.\end{aligned}
\end{equation}
In order to ensure that all the variables $F_k$ are coupled with each other, let us assume that $I\geq 4$. We have the following conservation of energy for $C_{13}$
\begin{equation}\label{ConservationE1}
\sum_{k=1}^I k\dot{F_k}=0,
\end{equation}
or equivalently
\begin{equation}\label{ConservationE2}
\sum_{k=1}^I k{F_k}=\mbox{const}.
\end{equation}
Similar to the case of $C_{12}$, we define $$G_k=\frac{F_k}{F_k+1},$$
and then we have
$$F_k=\frac{G_k}{1-G_k}.$$
Note that, similar to the previous section, $0<F_k<\infty$ and $0<G_k<1$.\\
The system \eqref{DiscreteC13} can be now written
\begin{equation}\label{DiscreteQuantumaConverteda}\begin{aligned}
&\frac{\dot{G}_{k_1}}{(1-G_{k_1})^2}=\overline{\mathcal{C}_{13}}[G]:=\\
:=&\quad\mathcal{K}^{13}_{k_1,k_2,k_3,k_4}\sum_{\substack{k_1=k_{2}+k_3+k_4,\\
|k_1|=|k_2|+|k_3|+|k_4|
}}\frac{G_{k_2}G_{k_3}G_{k_4}-G_{k_1}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})(1-G_{k_4})}
\\
&-3\mathcal{K}^{13}_{k_2,k_1,k_3,k_4}\sum_{\substack{k_2=k_{1}+k_3+k_4,\\
|k_2|=|k_1|+|k_3|+|k_4|
}}\frac{G_{k_1}G_{k_3}G_{k_4}-G_{k_2}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})(1-G_{k_4})}.\end{aligned}
\end{equation}
This system can also be rewritten as
\begin{equation}\begin{aligned}
{\dot{G}}
=&\mbox{diag}\left( {\begin{array}{cc}
{(1-G_1)^2}\\ \cdots \\ {(1-G_I)^2} \end{array} } \right)\times\\
&\times\sum_{\substack{k_1=k_{2}+k_3+k_4,\\
|k_1|=|k_2|+|k_3|+|k_4|
}}\big[K_{X_{k_{2}}+X_{k_{3}}+ X_{k_{4}}\rightarrow X_{k_{1}}}(G)-K_{X_{k_{1}}\rightarrow X_{k_{2}}+X_{k_{3}}+X_{k_{4}}}(G)\big](X_{k_{1}}-X_{k_{2}}-X_{k_{3}}-X_{k_{4}}).\end{aligned}
\end{equation}
where $X_k$ is, with an abuse of notation, as mentioned earlier, the vector
$$\left( {\begin{array}{cc}
0\\ \cdots\\ 1\\ \cdots \\ 0 \end{array} } \right),$$
in which the only element that is $1$ is the $k$-th one,
and
$$K_{X_{k_2}+X_{k_3}+ X_{k_4}\rightarrow X_{k_{1}} }(G):=\mathcal{K}^{13}_{k_1,k_2,k_3,k_4}\frac{G_{k_1}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})(1-G_{k_4})},$$
$$K_{X_{k_{1}}\rightarrow X_{k_2}+X_{k_3}+ X_{k_4}}(G):=\mathcal{K}^{13}_{k_1,k_2,k_3,k_4}\frac{G_{k_2}G_{k_3}G_{k_4}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})(1-G_{k_4})}.$$
We can also write
\begin{eqnarray}\nonumber
{\dot{G}}
&=&\mbox{diag}\left( {\begin{array}{cc}
{(1-G_1)^2}\\ \cdots \\ {(1-G_I)^2} \end{array} } \right)\sum_{y\leftrightarrow y'}\left[K_{y \rightarrow y'}(G)-K_{y' \rightarrow y}(G)\right](y'-y ),
\end{eqnarray}
where $y\leftrightarrow y'$ rang over the reversible reactions shown above.
\begin{theorem}\label{TheoremQnm} For any initial condition, the solution $$f(t)=(f_p(t))_{p\in\mathcal{L}_R}$$ of the quantum Boltzmann equation \eqref{DiscreteQuantuma} converges to an equilibrium state $f^*=(f_p^*)_{p\in\mathcal{L}_R}$. For each ray $\{kP_0\}_{k\geq 1}$ that intersects $\mathcal{L}_R$ in at least $4$ points there exists a constant $\rho_{P_0}$ such that if $p=kP_0$ then $$f^*_p=\frac{1}{e^{k\rho_{P_0}}-1}.$$ Moreover, the solution $f(t)$ of \eqref{DiscreteQuantuma} converges to $f^*$ exponentially fast in the following sense: there exists positive constants $C_1$, $C_2$ such that $$\max_{p\in\mathcal{L}_R}|f_p(t)-f_p^*|<C_1e^{-C_2t}.$$
\end{theorem}
\begin{proof}
The proof of Theorem \ref{TheoremQnm} then follows exactly from the same Lyapunov function
\eqref{Lyapunov} and arguments as in Theorem \ref{TheoremQ12}.
\end{proof}
\subsection{The dynamical system associated to $C_{22}$}
Let us consider a discretized version of the quantum Boltzmann model associated to the collision operator given by $C_{22}$:
\\\\ Let $\mathcal{L}_R$ denote the lattice of integer points
$$\mathcal{L}_R=\{p~~|~~|p|\in\mathbb{Z}^3, |p|<R\}.$$
The discretized quantum Boltzmann equation associated to $C_{22}$ reads $\forall p_1\in \mathcal{L}_R$
\begin{equation}\label{DiscreteQuantumb}\begin{aligned}
&\dot{f}_{p_1}=C_{22}^D[f_{p_1}]:=\\
:=
&\quad\sum_{\substack{p_2,p_3,p_4\in\mathcal{L}_R,\\ p_1+p_2=p_3+p_4,\\ \mathcal{E}(p_1)+\mathcal{E}(p_{2})=\mathcal{E}(p_{3})+\mathcal{E}(p_{4})}}\mathcal{K}^{13}_{p_1,p_2,p_3,p_4}\{(f_{p_1}+1)(f_{p_2}+1)f_{p_3}f_{p_4}-f_{p_1}f_{p_2}(f_{p_3}+1)(f_{p_4}+1)\},~~
\end{aligned}\end{equation}
where $\mathcal{E}(p)$ is defined in \eqref{E3a}.
Similar to the $C_{12}$ case, when $p=0$, $\mathcal{K}^{22}_{p_1,p_2,p_3,p_4}=0$, and we obtain
\begin{eqnarray*}\label{DiscreteQuantumIndex0}
\dot{f}_0=0,
\end{eqnarray*}
which means $f_0(t)$ is a constant for all time $t$. As a consequence, we can suppose that $f_0(0)=0$, which implies $f_0(t)=0$ for all $t$.
\\ In \eqref{DiscreteQuantumb}, the sums for $C_{22}$ are taken over $(p_1,p_2,p_3,p_4)$ satisfying
\begin{equation}\label{EnergyC22}
p_1 \ + \ p_2 \ = \ p_3\ +\ p_4, \mbox{ and } \ \ \ \ \mathcal{E}(p_1)\ + \ \mathcal{E}(p_2)\ = \ \mathcal{E}(p_3) \ + \mathcal{E}(p_4).
\end{equation}
In this case, unlike in the case of $C_{12}$ and $C_{13}$, we {\it cannot} infer from \eqref{EnergyC22} that there exists a vector $P$ and $k_1$, $k_2$, $k_3, k_4\geq0$, $k_1, k_2, k_3, k_4\in\mathbb{Z}$ such that
$$p_1=k_1 P;~~~p_2=k_2 P;~~~p_3=k_3 P;~~~p_4=k_4 P,~~~k_1+k_2=k_3+k_4.$$
However, let us consider the following simplified version of \eqref{DiscreteQuantumb} for $C_{22}$
\begin{equation}\label{DiscreteC22}\begin{aligned}
\dot{F}_{k_1}
=&\ \mathcal{C}_{22}[F](k_1):=\quad\sum_{\substack{k_1+k_2=k_3+k_4\\k_2,k_3,k_4\in\mathbb{I}}}\mathcal{K}^{13}_{k_1,k_2,k_3,k_4}\{(F_{k_1}+1)(F_{k_2}+1)F_{k_3}F_{k_4}-\\
& -F_{k_1}F_{k_2}(F_{k_3}+1)(F_{k_4}+1)\}, ~\forall k_1\in\mathbb{I}.\end{aligned}
\end{equation}
Recall that $\mathbb{I}=\{1,\cdots,I\}$. We also suppose that $I\geq3$.
We have the following conservation of energy
\begin{equation}\label{ConservationE1b}
\sum_{k=1}^I k\dot{F_k}=0,
\end{equation}
or equivalently
\begin{equation}\label{ConservationE2b}
\sum_{k=1}^I k{F_k}=\mbox{const}.
\end{equation}
For $C_{22}$, the following ``conservation of mass'' also holds
\begin{equation}\label{ConservationM1}
\sum_{k=1}^I \dot{F_k}=0,
\end{equation}
or equivalently
\begin{equation}\label{ConservationM2}
\sum_{k=1}^I {F_k}=\mbox{const}.
\end{equation}
Similar to the case of $C_{12}$, define $$G_k=\frac{F_k}{F_k+1},$$
then
$$F_k=\frac{G_k}{1-G_k},$$
and
the system \eqref{DiscreteQuantumb} can be now written
\begin{equation}\label{DiscreteQuantumaConvertedb}\begin{aligned}
&\frac{\dot{G}_{k_1}}{(1-G_{k_1})^2}=\overline{\mathcal{C}_{22}}[G]:=\\
:=&\quad\mathcal{K}^{13}_{k_1,k_2,k_3,k_4}\sum_{\substack{k_1+k_{2}=k_3+k_4,\\
|k_1|+|k_2|=|k_3|+|k_4|
}}\frac{G_{k_3}G_{k_4}-G_{k_1}G_{k_2}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})(1-G_{k_4})}
.\end{aligned}
\end{equation}
This system can be rewritten as
\begin{equation}\begin{aligned}
{\dot{G}}
=&\mbox{diag}\left( {\begin{array}{cc}
{(1-G_1)^2}\\ \cdots \\ {(1-G_I)^2} \end{array} } \right)\times\\
&\times\sum_{\substack{k_1+k_{2}=k_3+k_4,\\
|k_1|+|k_2|=|k_3|+|k_4|
}}\big[K_{X_{k_{3}}+ X_{k_{4}}\rightarrow X_{k_{2}}+X_{k_{1}}}(G)\\
&-K_{X_{k_{2}}+X_{k_{1}}\rightarrow X_{k_{3}}+X_{k_{4}}}(G)\big](X_{k_{1}}+X_{k_{2}}-X_{k_{3}}-X_{k_{4}}).\end{aligned}
\end{equation}
where $X_k$ is, with an abuse of notation, the vector
$$\left( {\begin{array}{cc}
0\\ \cdots\\ 1\\ \cdots \\ 0 \end{array} } \right),$$
in which the only element that is $1$ is the $k$-th one,
and
$$K_{X_{k_3}+ X_{k_4}\rightarrow X_{k_2}+X_{k_{1}} }(G)=\mathcal{K}^{13}_{k_1,k_2,k_3,k_4}\frac{G_{k_1}G_{k_2}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})(1-G_{k_4})},$$
$$K_{X_{k_2}+X_{k_{1}}\rightarrow X_{k_3}+ X_{k_4}}(G)=\mathcal{K}^{13}_{k_1,k_2,k_3,k_4}\frac{G_{k_3}G_{k_4}}{(1-G_{k_1})(1-G_{k_2})(1-G_{k_3})(1-G_{k_4})}.$$
We can also write
\begin{eqnarray}\nonumber
{\dot{G}}
&=&\mbox{diag}\left( {\begin{array}{cc}
{(1-G_1)^2}\\ \cdots \\ {(1-G_I)^2} \end{array} } \right)\sum_{y\leftrightarrow y'}\left[K_{y \rightarrow y'}(G)-K_{y' \rightarrow y}(G)\right](y'-y ),
\end{eqnarray}
where $y\leftrightarrow y'$ range over the reversible reactions shown above.
\begin{theorem}\label{TheoremQnm}
For any initial condition, the solution $$F(t)=(F_k(t))_{k\in\mathbb{I}}$$ of the quantum Boltzmann equation \eqref{DiscreteC22} converges to an equilibrium state $F^*=(F_k^*)_{k\in\mathbb{I}}$, where $$F^*_k=\frac{1}{e^{\rho_2(k-1)-\rho_1(k-2)}-1}.$$ Moreover, the solution $F(t)$ of \eqref{DiscreteC22} converges to $F^*$ exponentially fast in the following sense: there exists positive constants $C_1$, $C_2$ such that $$\max_{k\in\mathbb{I}}|F_k(t)-F_k^*|<C_1e^{-C_2t}.$$
\end{theorem}
\begin{proof}
We notice that \begin{equation}\label{Cmnequilibrium}
\frac{G^y }{(G^*)^y}
=\frac{G^{y'} }{(G^*)^{y'} },\end{equation}
holds true for all reactions $y \leftrightarrow y',$ if and only if $G^y=G^{y'}$
since $(G^*)^y=(G^*)^{y'}$ for all reactions $y \leftrightarrow y'.$ In the case of $C_{22}$, we obtain the relation $G^{*}_{k_1}\cdot G^{*}_{k_2} = G^{*}_{k_3}\cdot G^*_{k_4}$ for all $k_1$, $k_2$, $k_3$, $k_4$ such that $k_1+k_2=k_3+k_4\le I$. From the relation $G_{k_1}\cdot G_{k_2} = G_{k_3}\cdot G_{k_4}$ and the fact that $k+(k-2)=2(k-1)$, the following identity holds true
$$G_k(G_1)^{k-2}=(G_2)^{k-1}.$$
We then obtain $G_k=(G_2)^{k-1}/(G_1)^{k-2}$. Identity \eqref{Cmnequilibrium} then implies that $(G)^y=(G)^{y'}$ for all reactions $y \leftrightarrow y',$ which leads to $G_k=e^{\rho' k}$.
\\ The conservation relation
$$\sum_{k=1}^n\frac{G_k}{1-G_k}=\sum_{k=1}^n\frac{G_k^*}{1-G_k^*},$$
implies that that $\rho=\rho'$, then $G_k=G_k^*$.
We can still use the Petri net argument of \cite{AngeliDeLeenheerSontag:PRF:2011} or the result in \cite{Craciun:2015:TDI}, to prove that the system is persistent. For example, to use the method from \cite{AngeliDeLeenheerSontag:PRF:2011}, we note that we have two {\it siphons} $\{X_1, X_2,\cdots, X_I\}$, $\{X_2,\cdots, X_I\}$. However, we also have the conservations of mass and energy
$$\sum_{k=1}F_k=\mbox{constant},$$
$$\sum_{k=1}kF_k=\mbox{constant},$$
that leads to the {\it $P$-semiflow}
$$\sum_{k=2}(k-1)F_k=\mbox{constant}.$$
Therefore, similar to the case of $C_{12}$, it follows that the system is persistent, and we can use the same Lyapunov function as in the proof of Theorem \ref{TheoremQ12} to obtain the desired convergence result.
\end{proof}
\begin{remark}
If $I<3$ then $F_{k}^*\equiv 0$. If $I=3$ then $F_{2}^*\equiv 0$ and $F^*_{1}=\frac{1}{e^\rho-1}$, $F^*_{3}=\frac{1}{e^{3\rho}-1}$ for some $\rho=\rho(P_0)$.
\end{remark}
\section{A reaction network approach for the sum of $C_{12}, C_{22}, C_{13}$}\label{Sec:SumQnm}
Let us consider the following equations
\begin{equation}\label{DiscreteC12C22}
\dot{F}_{k_1}=\mathcal{C}_{12}[F](k_1) \ + \ \mathcal{C}_{22}[F](k_1),
\end{equation}
and
\begin{equation}\label{DiscreteC12C22C13}
\dot{F}_{k_1}=\mathcal{C}_{12}[F](k_1) \ + \ \mathcal{C}_{22}[F](k_1)\ + \ \mathcal{C}_{13}[F](k_1),
\end{equation}
where $\mathcal{C}_{12}$, $\mathcal{C}_{22}$, $\mathcal{C}_{13}$ are the operators defined in \eqref{DiscreteC12}, \eqref{DiscreteC13}, \eqref{DiscreteC22}.
The following theorem then follows by exactly the same argument as in Theorem \ref{TheoremQnm}
\begin{theorem}\label{TheoremSumQnm} For any initial condition, the solution $$F(t)=(F_k(t))_{k\in\mathbb{I}}$$ of the quantum Boltzmann equation \eqref{DiscreteC12C22} or \eqref{DiscreteC12C22C13} converges to an equilibrium state $F^*=(F_k^*)_{k\in\mathbb{I}}$, where $F^*_k=\frac{1}{e^{\rho k}-1}$ for some constant $\rho$. Moreover, the solution $F(t)$ of \eqref{DiscreteC13} converges to $F^*$ exponentially fast in the following sense: there exists positive constants $C_1$, $C_2$ such that $$\max_{k\in\mathbb{I}}|F_k(t)-F_k^*|<C_1e^{-C_2t}.$$
\end{theorem}
~~ \\{\bf Acknowledgements.} G. Craciun was supported by NSF grant DMS-1412643.
M.-B Tran was supported by NSF Grant RNMS (Ki-Net) 1107444, ERC Advanced Grant DYCON. M.-B Tran would like to thank Professor Linda Reichl and Professor Robert Dorfman for fruiful discussions on the topic. The research was carried on partially while M.-B. Tran was visiting University of Texas at Austin. He would like to thanks the institution for the hospitality. \bibliographystyle{plain} |
\section{Introduction}
In 2015, LHCb collaboration reported that
two candidates of the new exotic baryons, $P_c$(4380) and $P_c$(4450),
had been observed in the $\Lambda_b^0\rightarrow \mbox{$J\!/\!\psi$} p K^-$ decay.
The mass of the higher peak, $P_c$(4450),
is 4449.8$\pm$1.7$\pm$2.5 MeV with a width of 39$\pm$5$\pm$19 MeV,
while the lower and broader peak, $P_c$(4380),
has a mass of 4380$\pm$8$\pm$29 MeV and a width of 205$\pm$18$\pm$86 MeV.
The most favorable set of the spin parity for the lower and the higher peaks is
$J^P = (\half3^-,\half5^+)$,
though $(\half3^+,\half5^-)$ or $(\half5^+,\half3^-)$ are also acceptable \cite{Aaij:2015tga}.
Also, because the $\mbox{$J\!/\!\psi$} p$ contribution is necessary to describe the decay data \cite{Aaij:2016phn},
it is almost certain
the peaks have the $c\overline{{c}}$ pair and are considered
as the isospin-$\half1$ $uudc\overline{{c}}$ pentaquarks.
Beside of the predicting work \cite{Wu:2010jy},
the LHCb observation of $P_c$(4380) and $P_c$(4450)
has evoked many theoretical studies: the hadronic molecule
with the meson exchange interaction, the chiral unitary approach with the hidden local gauge symmetry,
the QCD sum rule, the chiral quark model,
the diquark/triquark model as well as that of the kinematical effects \cite{Chen:2016qju}.
At present, the theoretical and experimental knowledge
is
not enough and one cannot draw a definite picture of these peaks.
Here, we concentrate our attention on the short range part of
the hidden-charm pentaquark structure, which is governed by the quark and gluon dynamics.
For this purpose, we employ the quark cluster model,
which successfully explained the short range part of the baryon-baryon interaction
\cite{Oka:2000wj} and
the structure of the light flavored pentaquark $\Lambda(1405)$ \cite{Takeuchi:2007tv}.
It is also shown that
the baryon-baryon interaction derived from the lattice QCD
is found to be similar to that of the quark cluster model
\cite{Sasaki:2015ab}.
Since we are interested in the short range region,
we have investigated the $S$-wave five-quark systems
as a first step.
They correspond to the negative-parity pentaquarks.
In order to discuss the positive-parity pentaquark state,
which has also been observed by the LHCb experiments,
one has to investigate the $P$-wave five-quark systems,
which is beyond the scope of the present paper.
\begin{table}[t]
\renewcommand\arraystretch{1.4}
\tabcolsep=1mm
\caption{The classification of the isospin-$\half1$ negative parity $qqqc\overline{{c}}$ states.
The $uud$ spin ($s_q$), color ($c$),
CMI of the five quark systems at the heavy quark limit
($\langle {{\cal O}_\text{CMI}} \rangle_{5q}^{(HQ)}$), the $c\overline{{c}}$ spin ($s_c$), the total spin of $uudc\overline{{c}}$ ($J$),
the lowest $S$-wave threshold (T) and the CMI contribution to the threshold energy ($\langle {{\cal O}_\text{CMI}} \rangle_\text{T}^{(HQ)}$) are listed.}
\begin{center}
\begin{tabular}{ccccccccccccccccccccccc}\hline
& $s_q$ & $c$ & $\langle {{\cal O}_\text{CMI}} \rangle_{5q}^{(HQ)}$ & $s_c$ & $J$ & T & $\langle {{\cal O}_\text{CMI}} \rangle_\text{T}^{(HQ)}$
\\\hline
{[$q^3$1\half1]}\ &$\half1$ & {\bf 1} &$-8$& 0 & $\half1$ &~~~$N\mbox{$\eta_c$}$~~~ & $-8$
\\
&& && 1 & $\half1$ &$N\mbox{$J\!/\!\psi$}$&$-8$
\\
&& && 1 & $\half3$ &$N\mbox{$J\!/\!\psi$}$&$-8$
\\\hline
{[$q^3$8\half1]}\ & $\half1$ & {\bf 8} &$-2$& 0 & $\half1$ & \mbox{$\Lambda_c$}\overline{{D}}{}&$-8$
\\
&& && 1 & $\half1$ & \mbox{$\Lambda_c$}\overline{{D}}{}&$-8$
\\
& &&& 1 & $\half3$ & \mbox{$\Lambda_c$}\mbox{$\overline{D}{}^*$}& ${-8}$
\\\hline
{[$q^3$8\half3]}\ & $\half3$ & {\bf 8} &\phantom{$-$}2& 0 & $\half3$ & \mbox{$\Sigma_c^*$}\overline{{D}}{}& ${8\over 3}$
\\
&& && 1 & $\half1$ & \mbox{$\Sigma_c$}\overline{{D}}{}&
${8\over 3}$
\\
&& && 1 & $\half3$ & \mbox{$\Sigma_c^*$}\overline{{D}}{}& ${8\over 3}$
\\
& &&& 1 & $\half5$ & \mbox{$\Sigma_c^*$}\mbox{$\overline{D}{}^*$}& ${8\over 3}$
\\\hline
\end{tabular}
\end{center}
\label{tbl:hq}
\end{table}%
Let us first discuss possible configurations of $uud$ quarks in the $uudc\overline{{c}}$ systems.
Since the whole system is the color-singlet and
the $c\overline{{c}}$ pair
is color-singlet or octet,
the remaining three light quarks are also color-singlet or color-octet.
So, when the orbital configuration is totally symmetric,
the $uud$ configuration in the $uudc\overline{{c}}$ systems is totally symmetric ({\bf 56}-plet)
or mixed symmetric ({\bf 70}-plet) in the flavor-spin
SU$_{f\sigma}$(6)
space. They are classified as:
\begin{align}
{\bf 56}_{f\sigma}
&=
{\bf 8}_f \times {\bf 2}_\sigma
+ {\bf 10}_f\times {\bf 4}_\sigma
\\
{\bf 70}_{f\sigma}
&=
{\bf 1}_f \times {\bf 2}_\sigma
+ {\bf 8}_f \times {\bf 2}_\sigma
+ {\bf 8}_f \times {\bf 4}_\sigma
+ {\bf 10}_f\times {\bf 2}_\sigma\ .
\end{align}
Here the numbers are the dimension of the corresponding representations.
The color-singlet $uud$ systems
correspond to the usual {\bf 56}-plet baryons.
The color-octet {\bf 70}-plet systems
can be decomposed into
the flavor-singlet spin-$\half1$ (${\bf 1}_f\times {\bf 2}_\sigma$),
the flavor-octet spin-$\half1$ (${\bf 8}_f\times {\bf 2}_\sigma$),
the flavor-octet spin-$\half3$ states (${\bf 8}_f\times {\bf 4}_\sigma$),
and the flavor-decuplet spin-$\half1$ (${\bf 10}_f\times {\bf 2}_\sigma$).
Since the present work concerns systems of the isospin $\half1$ and the strangeness zero,
namely, flavor-octet systems,
the configurations of the three light quarks correspond to
one of the following three:
(a) color-singlet spin-$\half1$,
(b) color-octet spin-$\half1$, and
(c) color-octet spin-$\half3$,
each of them we denote by {[$q^3$1\half1]}, {[$q^3$8\half1]}, and {[$q^3$8\half3]}\ in the following, respectively.
Since the spin of the $c\overline{{c}}$ pair is either 0 or 1,
the total spin of the $uudc\overline{{c}}$ systems is
either $\half1$ (5-fold), $\half3$ (4-fold), or $\half5$ (1-fold). (See Table \ref{tbl:hq}.)
In the short range part of the two-hadron interaction, the color-magnetic interaction (CMI)
plays an important role.
We evaluate the color flavor spin part of CMI,
\begin{align}
{{\cal O}_\text{CMI}}&=-\sum_{ij}{m_u^2\over m_im_j}{\lambda_i\cdot\lambda_j}{\sigma_i\cdot\sigma_j} \ ,
\label{eq:Ocmi}
\end{align}
by the quark wave function.
In eq.\ (\ref{eq:Ocmi}),
$m_i$, $\lambda_i$, and $\sigma_i$ are
the constituent mass, the Gell-Mann matrix in the color space,
and the Pauli spin matrix for the $i$th (anti)quark, respectively.
Because of the factor ${m_u^2\over m_im_j}$, only the operators between the light quarks give non-zero
contribution at the
heavy quark limit.
So, CMI estimated
by the above three-light-quark configurations
actually correspond to the estimates of the whole $uudc\overline{{c}}$
at that limit, $\langle {{\cal O}_\text{CMI}}\rangle_{5q}^{(HQ)}$,
which is listed in Table \ref{tbl:hq}.
There, we also show the lowest $S$-wave baryon-meson threshold for each state
together with the contribution of
${{\cal O}_\text{CMI}}$ to that threshold at the heavy quark limit:
\begin{align}
\langle {{\cal O}_\text{CMI}}\rangle_\text{T}^{(HQ)}&=\langle {{\cal O}_\text{CMI}}\rangle_B^{(HQ)}+\langle {{\cal O}_\text{CMI}}\rangle_M^{(HQ)}\ ,
\end{align}
where $\langle {{\cal O}_\text{CMI}}\rangle_B^{(HQ)}$ and $\langle {{\cal O}_\text{CMI}}\rangle_M^{(HQ)}$ are ${{\cal O}_\text{CMI}}$ evaluated
by the baryon and the meson wave functions
at the heavy quark limit, respectively.
Suppose we estimate the baryon-meson potential arising from CMI by
\begin{align}
V_\text{cmi}^{\text{eff}(HQ)} &\sim
\langle {{\cal O}_\text{CMI}}\rangle_{5q}^{(HQ)} -\langle {{\cal O}_\text{CMI}}\rangle_\text{T}^{(HQ)},
\label{eq:Veff}
\end{align}
then only those which include the {[$q^3$8\half3]}\ configuration
is attractive
though its energy is actually the highest.
As in Table \ref{tbl:hq},
there are four states which gain such an attraction
in the isospin-\half1 $uudc\overline{{c}}$ systems.
Since $uudc\overline{{c}}$ is color-singlet as a whole,
the system of the color-octet $uud$ with color-octet $c\overline{{c}}$
can be observed as \mbox{$\Lambda_c$}\overline{{D}}{}${}^{(*)}$ or \mbox{$\Sigma_c$}${}^{(*)}$\overline{{D}}{}${}^{(*)}$ baryon meson states,
where each of the hadrons is color-singlet.
The attraction within the color-octet spin-\half3 $uud$ is observed
as the attraction between the $\Sigma_c^{(*)}$ baryon and the $\overline{{D}}{}^{(*)}$ meson.
In this sense, when this interaction causes pentaquark states,
one may call them `hidden color-octet $uud$ baryons.'
As we will show in this letter,
the above behavior remains visible even after we employ the realistic quark masses and perform
dynamical calculations.
We have found that there is a bound state
in the $J=\half5$ system,
a resonance and a cusp in the $\half1$ system, and a resonance in $\half3$ system;
which exactly correspond to those of the {[$q^3$8\half3]}\ configuration
as listed in Table \ref{tbl:hq}.
Since the expectation values of CMI by the $uud$ configurations
can be calibrated by the observed hadron masses,
and since they do not depend on the heavy quark mass,
the above four structures are robust to change of the parameters.
Though we use a rather complicated model Hamiltonian in the following
in order to produce the threshold energies correctly,
the results do not depend much on the model details;
the situation is the same, for example,
when the system goes to the bottom sector.
Thus, we would like to argue that the negative parity peak of the LHCb pentaquark
may consist of (some of) these structures caused by the color-octet $uud$ configuration.
\section{Method}
We employ the coupled-channel quark cluster model to investigate the
$uudc\overline{{c}}$ $I(J^P)=\half1(\half1^-,\half3^-,\half5^-)$ systems.
This model becomes a $(0s)^5$ quark model in the baryon-meson short range region.
In the long range region,
this model becomes
essentially
a baryon-meson model.
There, the interaction between the baryon and the meson
arises from the quark degrees of freedom and from the interaction between quarks.
The model Hamiltonian,
$H_q$, consists of the central spin-independent term, $H_c$,
and the color spin term, $V_\text{CMI}$.
The $H_c$ consists of
the kinetic term, $K$, the confinement term, $V_\text{conf}$, and the
color Coulomb term, $V_\text{Coul}$:
\begin{align}
H_q&=H_c+V_\text{CMI}
\\
H_c&= K+V_\text{conf}+V_\text{Coul}\ .
\end{align}
Both of $V_\text{Coul}$ and $V_\text{CMI}$ come from the effective one-gluon exchange interaction
between quarks.
Each of the terms is taken to be slightly different from the
conventional quark model \cite{Godfrey:1985xj,Capstick:1986bm}.
It is because we use a single Gaussian
for the orbital wave function of each of the $q\overline{{q}}$ and $q^3$ hadrons
in order to make it feasible to solve the five-quark systems.
The kinetic term is taken as nonrelativistic:
\begin{align}
K &=\sum m_i+{1\over 2m_i}\Big({\xbld{p}}_i-{m_i\over M_G}{\xbld{P}}_G\Big)^2\ ,
\end{align}
where ${\xbld{p}}_i$ is the momentum of the $i$th (anti)quark, and
$M_G$ and ${\xbld{P}}_G$ are the total mass and momentum of the five-quark system.
The linear confinement term is
\begin{align}
V_\text{conf}&=\sum_{i<j} {\lambda_i\cdot\lambda_j}\,( -a_c r_{ij} + c_1+{c_2^2\over \mu_{ij}}+c_{q\overline{{q}}} )\ .
\end{align}
The value of the confinement strength, $a_c$, is taken from the
Lattice QCD calculation \cite{Kawanai:2011xb},
whose value corresponds 1.12 GeV$^2$ for the $q\overline{{q}}$ systems.
The $r_{ij}$ and $\mu_{ij}$ are the relative distance and the reduced masses of the $i$th and the $j$th quarks, respectively.
In the above equation, the $c_1$, $c_2$, and $c_{q\overline{{q}}}$
are the constant parameters.
The $c_1$ and $c_2$ express the constant mass shift
which is expanded up to the $\mu_{ij}^{-1}$ term.
The parameter $c_{q\overline{{q}}}$ is nonzero only when this operates on quark-antiquark pairs.
We use these constants as
free parameters so that the model produces the observed masses of the
relevant single hadrons, which have a quite wide energy range.
The constant mass shift itself appears
when the potential is constructed from the lattice QCD calculation,
though the values are different \cite{Kawanai:2011xb}.
The color-Coulomb term is written as
\begin{align}
V_\text{Coul}&=\sum_{i<j}{{\lambda_i\cdot\lambda_j}\over 4}{\alpha_s(r_{ij})\over r_{ij}}
\\
\alpha_s(r) &= \sum_{k=1}^3 \alpha_k \,\text{erf}[\gamma_k r] \ .
\end{align}
The strong coupling constant, $\alpha_s$,
is assumed to depend on the relative distance of the interacting quarks,
following the manner of refs.\ \cite{Godfrey:1985xj,Capstick:1986bm}.
In this work, $\alpha_s$ at small $Q^2$
is refitted so that it corresponds to the running coupling constant for $Q^2>3$ GeV
in the momentum space,
whereas it goes to 0.8 at $Q^2=0$:
($4\gamma_k^2$ in GeV$^2$, \ $\alpha_k$)=(1.5, 0.45), (10, 0.15), (1000, 0.20).
The CMI term is
\begin{align}
V_\text{CMI} &= -\sum_{i<j}{{\lambda_i\cdot\lambda_j}\over 4}\alpha_s^{ss}{}_{ij}{2\pi\over 3m_im_j}{\sigma_i\cdot\sigma_j}\delta^3({\xbld{r}}_{ij})
\\
\alpha_s^{ss}{}_{ij}&=\begin{cases}\displaystyle \alpha_{s1}^{ss}+ \alpha_{s2}^{ss}{m_u\over \mu_{ij}}& \text{for a $qq$ pair}\\
\alpha_{s3}^{ss}& \text{for a $q\overline{{q}}$ pair.}
\end{cases}
\end{align}
We use $\alpha_{s1}^{ss}$ and $\alpha_{s2}^{ss}$ as parameters to
fit the contribution of the $uu$ pairs in the baryons such as
$2m_{\Sigma_Q^*}+m_{\Sigma_Q}-3m_{\Lambda_Q}$ and that of the $uQ$ pairs such as
$m_{\Sigma_Q^*}-m_{\Sigma_Q}$.
The value of $\alpha_{s3}^{ss}$ is determined by taking an average of the \mbox{$\overline{D}{}^*$}-\overline{{D}}{}\ and $D_s^*$-$D_s$ mass differences.
Again these mass dependent coupling constants have to be introduced
so that the model gives the correct hyperfine splitting of the hadron masses.
In this way, we calibrate the size of the CMI,
which is the origin of the attraction focused in this work,
from the observables.
The parameters are summarized in Table \ref{tbl:param}.
\begin{table}
\caption{Model parameters. The quark masses, $m_Q$, and the constant parameters, $c_i$, are
in MeV.}
\begin{center}
\renewcommand\arraystretch{1.2}
\begin{tabular}{cccccccccccccc}\hline
$m_u(=m_d)$ &$m_c$ &$c_1$ &$c_2$&$c_{q\overline{{q}}}$\\ \hline
300 &1741.5&86.4& 113.9 &$-$5.65\\ \hline
$a_c$(MeV/fm)
&$\alpha_{s1}^{ss}$&$\alpha_{s2}^{ss}$&$\alpha_{s3}^{ss}$ \\ \hline
196.9 &$-$1.0967 & 0.4756 &0.5668\\ \hline
\end{tabular}
\end{center}
\label{tbl:param}
\end{table}%
The color flavor spin part of the $q^3$ or $q\overline{{q}}$ wave functions is
taken as a conventional way \cite{Stancu:1996,pdg}.
The orbital wave function of the mesons, $\phi_M$, and that of the baryons, $\phi_B$,
are written by Gaussian with a size parameter $b$, $\phi({\xbld{r}},b)$:
\begin{align}
\phi_M({\xbld{r}}_M) &= \phi({\xbld{r}}_{12},{x_{0}\over \sqrt{\mu_{12}}})
\\
\phi_B({\xbld{r}}_B)
&=\phi({\xbld{r}}_{12},{x_{0}\over \sqrt{\mu_{12}}}) \phi({\xbld{r}}_{12-3},{x_{0}\over \sqrt{\mu_{12-3}}})
\ ,
\end{align}
where the reduced masses, $\mu_{12}$ and $\mu_{12-3}$,
correspond to the Jacobi coordinates, ${\xbld{r}}_{12}$ and ${\xbld{r}}_{12-3}$.
We assume that the size parameter of the orbital motion can be
approximated by $b = x_0/\sqrt{m}$ and
minimize the central part of the Hamiltonian, $H_c$, against
$x_0$ for each flavor set:
$u\overline{{c}}$, $c\overline{{c}}$, $uud$, $udc$.
For the baryons, this means that the ratio of the size parameters is kept to a certain mass ratio;
{\it e.g.}, $b_{uc}/b_{ud}$
in \mbox{$\Lambda_c$}\ or \mbox{$\Sigma_c$}\ is equal to $\sqrt{\mu_{ud}/\mu_{uc}}$.
\begin{table}
\caption{Single hadrons masses obtained by the present model. The isospin-averaged masses are taken from \cite{pdg}.}
\begin{center}
\renewcommand\arraystretch{1.2}
\begin{tabular}{ccccc}\hline
Baryon &
N &
$\Lambda_c$ &
$\Sigma_c$ &
$\Sigma_c^*$ \\\hline
Mass&
~922.3&
2291.8&
2453.6&
2516.0\\
Obs.\ &
~938.9&
2286.5&
2453.5&
2518.1\\ \hline
Meson &
$\eta_c$ &
\mbox{$J\!/\!\psi$}\ &
\overline{{D}}{}\ &
\mbox{$\overline{D}{}^*$}\ \\\hline
Mass&
2981.3&
3100.9&
1863.4&
2004.9\\
Obs.\ &
2983.6&
3096.9&
1867.2&
2008.6\\
\hline
\end{tabular}
\label{tbl:cmibm}
\end{center}
\end{table}
The masses of the relevant hadrons obtained by the models are summarized in table \ref{tbl:cmibm}
together with the observed masses.
The obtained $x_0$'s are listed in Table \ref{tbl:x0}
with the corresponding size parameters.
It is found that $x_0$ does not vary much while
the difference between $b_{uu}$, $b_{uc}$ and $b_{cc}$ is large.
In order to investigate systems with more than one charm quark
one needs to take into account the flavor dependence of the orbital motion.
\begin{table}
\caption{The size parameter $b_{ij}$ (fm) and the parameter $x_0$ (fm$^{1/2}$) obtained by minimizing the central part of the Hamiltonian, $H_c$, for each of the systems. }
\begin{center}
\begin{tabular}{lcccccccccccccc}\hline
system & $x_0$&$b_{uu}$&$b_{uc}$&$b_{cc}$\\\hline
$uud$ & 0.60 & 0.68 & & \\
$uuc$ & 0.62 & 0.71 & 0.54 & \\
$ucc$ & 0.65 & & 0.57 & 0.31 \\
$u\overline{{c}}$ & 0.56 & & 0.49 & \\
$c\overline{{c}}$ & 0.61 & & & 0.29 \\ \hline
\end{tabular}
\end{center}
\label{tbl:x0}
\end{table}%
We employ the resonating group method (RGM) to solve the five-quark systems.
The wave function, $\Psi$,
can be expanded by the locally peaked Gaussians for each of
the baryon-meson channel $\nu$ as \cite{Oka:2000wj,Takeuchi:2007tv}
\begin{align}
\Psi &= \sum_{\nu,i} c^\nu_i
{\cal A}_q \big\{\psi_B^\nu({\xbld{r}}_B)\psi_M^\nu({\xbld{r}}_M)\chi({\xbld{R}},{\xbld{S}}_i)\big\}
\label{eq:16}
\\
\chi({\xbld{R}},{\xbld{S}}_i)&=i_0({1\over b^2}{\xbld{R}}\cdot{\xbld{S}}_i)\exp[-{1\over 2b^2}(R^2+S_i^2)]
\ ,
\label{eq:17}
\end{align}
where ${\cal A}_q$ stands for the quark antisymmetrization, which operates on the four quarks,
and $i_\ell(z)$ is the modified spherical Bessel function.
As for the scattering state, the wave function of the relative motion
is connected smoothly to the spherical Hankel functions in the long range region.
By integrating out the internal wave function of the hadrons,
the RGM equation can be obtained from the equation of motion for the quarks
$(H_q-E) \Psi = 0$, as
\begin{align}
\sum_{\nu'j} (H_{ij}^{\nu\nu'}-EN_{ij}^{\nu\nu'}) c^{\nu'}_j=0
\end{align}
with the Hamiltonian and normalization kernels
\begin{align}
\left\{
\begin{matrix}
H_{ij}^{\nu\nu'}\\N_{ij}^{\nu\nu'}
\end{matrix}
\right\}
&=
\int{\rm d} {\xbld{r}} {\rm d}{\xbld{r}}'\;
\psi_B^{\nu\dag}({\xbld{r}}_B)\psi_M^{\nu\dag}({\xbld{r}}_M)\chi^\dag({\xbld{R}},{\xbld{S}}_i)
\nonumber\\
&\times
\left\{
\begin{matrix}
H_q \\1
\end{matrix}
\right\}
{\cal A}_q
\big\{\psi_B^{\nu'}({\xbld{r}}'_B)\psi_M^{\nu'}({\xbld{r}}'_M)\chi({\xbld{R}}',{\xbld{S}}_j)\big\}
\ ,
\label{eq:RGMk}
\end{align}
where d${\xbld{r}}$ stands for the integration over all the Jacobi coordinates of the five-quark system.
For the detail of the calculation, see, for example, appendix B of ref.\ \cite{Takeuchi:2007tv}.
We choose the parameters $b$ and ${\xbld{S}}_i$ in eqs.\ (\ref{eq:16}) and
(\ref{eq:17})
so that the calculating results are stable against changing the parameters.
Thus, after fitting single hadron masses, the model is essentially parameter-free.
Here we define a three-body operator, ${\cal P}$,
to extract the $uud$ color-$c$, spin-$s$, orbital $(0s)^3$ configuration
from the resonance as well as from the bound states
in order to evaluate its size.
It is defined as
\begin{align}
{\cal P}{}_{cs} &= |uud;cs(0s)^3\rangle \langle uud;cs(0s)^3|
\\
{\cal P}&=\sum_{c,s}{\cal P}_{cs}
\ .
\end{align}
We use the same value as that of $b_{uu}$ in $\Sigma_c$ for the size parameter of the $(0s)^3$ component.
\section{Results and Discussions}
\subsection{$uudc\overline{{c}}$ $I(J^P)$=$\half1(\half5^-)$ state}
Suppose the orbital part of the $uudc\overline{{c}}$ system is
totally symmetric, such as the system in the orbital $(0s)^5$ configuration,
the color-spin-flavor part of the $uudc$ quarks should be totally antisymmetric.
For the $S$-wave $uudc\overline{{c}}$ $I(J^P)$=$\half1(\half5^-)$ channel,
this state is
\mbox{$\Sigma_c^*$}\mbox{$\overline{D}{}^*$}\
antisymmetrized over the quarks.
The color flavor spin part of its
normalization, $\langle \mbox{$\Sigma_c^*$}\mbox{$\overline{D}{}^*$} |{\cal A}_q|\mbox{$\Sigma_c^*$}\mbox{$\overline{D}{}^*$}\rangle$, is ${4\over 3}$.
For a channel which has a repulsion from Pauli blocking over the quarks,
this normalization becomes smaller than 1.
When the normalization is larger than 1, like this case,
an attraction between the two hadrons arises in the short range region.
Moreover, the $c\overline{{c}}$ pair in this state is spin-1 and the three light quarks are
color-octet spin-\half3,
{[$q^3$8\half3]}.
The attraction is expected to come also from CMI.
We investigate the \mbox{$\Sigma_c^*$}\mbox{$\overline{D}{}^*$}\ $\half1(\half5^-)$ channel
by the quark cluster model.
A very shallow bound state is found
1.0 MeV below the threshold.
The probability to find the $(0s)^3$ $uud$
in this bound state, $\langle {\cal P}\rangle$,
is found to be 0.21.
In this channel,
the $uud$ quarks in the $(0s)^3$ configuration are all color-octet spin-\half3.
Suppose there is no
Pauli exclusion principle applied between the $ud$ quarks in \mbox{$\Sigma_c$}\ and the $u$ quark in \mbox{$\overline{D}{}^*$},
the ratio of the color-singlet to the color-octet probabilities should be 1 to 8.
The color-singlet configuration vanishes by introducing the quark antisymmetrization,
and this small change of the configuration in size, $\sim 0.21\times$1/9,
induces the bound state.
\subsection{$uudc\overline{{c}}$ $I(J^P)$=$\half1(\half3^-)$ states}
The $uudc\overline{{c}}\ \half1(\half3^-)$ states consist of five baryon-meson channels:
$N\mbox{$J\!/\!\psi$}$, \mbox{$\Lambda_c$}\mbox{$\overline{D}{}^*$}, \mbox{$\Sigma_c^*$}\overline{{D}}{}, \mbox{$\Sigma_c$}\mbox{$\overline{D}{}^*$}, and \mbox{$\Sigma_c^*$}\mbox{$\overline{D}{}^*$}.
When the baryon and the meson come close to the overlapping region,
these channels are not orthogonal to each other any more.
There are
four $uud$ $\half1(\half3^-)$
states, {[$q^3$1\half1]}, {[$q^3$8\half1]}, and {[$q^3$8\half3]}\ (2-fold),
whereas five baryon-meson channels exist.
Thus one forbidden baryon-meson state, whose norm is zero, appears
when the system is
totally symmetric in the orbital space.
All the diagonal elements of the color flavor spin part of the normalization, however,
are close to or larger than 1;
the baryon-meson single channels do not gain
the large repulsion from the quark Pauli-blocking.
We have performed the quark cluster model calculation
where all the relevant five baryon-meson channels are coupled
for the $S$-wave $uudc\overline{{c}}$ $\half1(\half3^-)$ channel.
Before that, however,
we first discuss the results of the three-channel calculation of the \mbox{$\Sigma_c^*$}\overline{{D}}{},
\mbox{$\Sigma_c$}\mbox{$\overline{D}{}^*$}, and \mbox{$\Sigma_c^*$}\mbox{$\overline{D}{}^*$}\
in order to see the effects of the {[$q^3$8\half3]}\ configuration.
It is found that
a bound state appears 0.08 MeV below the \mbox{$\Sigma_c^*$}\overline{{D}}{}\ threshold.
Also, a very sharp resonance with the width less than 0.1 MeV
is found in the \mbox{$\Sigma_c^*$}\overline{{D}}{}\ channel at the energy
0.8 MeV below the \mbox{$\Sigma_c$}\mbox{$\overline{D}{}^*$}\ threshold.
The probability of the $uud$ to form one of the $(0s)^3$ configurations, $\langle{\cal P}\rangle$,
in the bound state
is found to be 0.08.
The proportion of the factor to find each {[$q^3$1\half1]}, {[$q^3$8\half1]}, and {[$q^3$8\half3]}\ configurations,
$\langle {\cal P}_{cs}\rangle/\langle{\cal P}\rangle$
is
listed in Table \ref{tbl:prob0s} with an identification of C$'$,
together with those of the resonance (B$'$).
Those proportions are very similar to the
antisymmetrized \mbox{$\Sigma_c^*$}\overline{{D}}{}\ (0.05 0.19 0.76) and \mbox{$\Sigma_c$}\mbox{$\overline{D}{}^*$} (0.01 0.06 0.93).
Thus, the bound state C$'$ is considered
essentially as the antisymmetrized \mbox{$\Sigma_c^*$}\overline{{D}}{}, and
the resonance B$'$
is considered as the antisymmetrized
\mbox{$\Sigma_c$}\mbox{$\overline{D}{}^*$}.
\begin{table}[tb]
\caption{The bound state, resonances and cusp obtained by the present model.
The four structures are identified by the letter A-D in the text. The identification with a dash (B$'$-D$'$) is used for the result of the \mbox{$\Sigma_c$}${}^{(*)}$\overline{{D}}{}${}^{(*)}$ three-channel calculation.
The energies, $E$, are shown in MeV.
The proportions of the factors to find each of {[$q^3$1\half1]}, {[$q^3$8\half1]}, and {[$q^3$8\half3]},
$\langle {\cal P}_{cs}\rangle/\langle{\cal P}\rangle$, are listed under the entry [$q^3cs$].
}
\begin{center}
\tabcolsep=0.9mm
\renewcommand\arraystretch{1.2}
\begin{tabular}{lllcrrrrrr} \hline
id.&\multicolumn{2}{l}{initial channel $(J^P)$}& $E$ & {[$q^3$1\half1]} & {[$q^3$8\half1]} & {[$q^3$8\half3]} \\\hline
A&\mbox{$\Sigma_c^*$}\mbox{$\overline{D}{}^*$} $(\half5^-)$ & bound state & 4519.9 & 0.00&0.00&1.00\\
B&$N$\mbox{$J\!/\!\psi$}\ $(\half3^-)$ & cusp & 4458.0 & 0.21&0.02&0.77\\
C&$N$\mbox{$J\!/\!\psi$}\ $(\half3^-)$ & resonance & 4379.3 & 0.24&0.16&0.60\\
D&$N$\mbox{$J\!/\!\psi$}\ $(\half1^-)$ & resonance & 4316.5 & 0.75&0.13&0.12\\\hline
B$'$&\mbox{$\Sigma_c^*$}\overline{{D}}{}\ $(\half3^-)$ & resonance & 4457.8 & 0.02&0.08&0.90\\
C$'$&\mbox{$\Sigma_c^*$}\overline{{D}}{}\ $(\half3^-)$ & bound state & 4379.3 & 0.05&0.21&0.74\\
D$'$&\mbox{$\Sigma_c$}\overline{{D}}{}\ $(\half1^-)$ & bound state & 4317.0 & 0.05&0.22&0.73\\\hline
\end{tabular}
\end{center}
\label{tbl:prob0s}
\end{table}%
\begin{figure}[btp]
\begin{center}
\includegraphics[width=9cm]{ps_23all_v2.eps}
\caption{The diagonal scattering phase shifts of the $S$-wave $uudc\overline{{c}}$ $I(J^P)$=$\half1(\half3^-)$ system.
The solid line is that of the $N\mbox{$J\!/\!\psi$}$ channel,
the long-dot-dashed line is for the \mbox{$\Lambda_c$}\mbox{$\overline{D}{}^*$} channel,
and the dot-dashed, double-dot-dashed, and triple-dot-dashed lines are
for the \mbox{$\Sigma_c^*$}\overline{{D}}{}, \mbox{$\Sigma_c$}\mbox{$\overline{D}{}^*$}, and \mbox{$\Sigma_c^*$}\mbox{$\overline{D}{}^*$} channels,
respectively. (Color online.)}
\label{fig:ps23}
\end{center}
\end{figure}
In figure \ref{fig:ps23},
we show phase shifts of the $S$-wave $uudc\overline{{c}}$ $\half1(\half3^-)$
by the calculation where all five baryon-meson channels are coupled.
The plotted phase shifts are the diagonal ones;
namely, the initial and the final channels are taken to be the same.
The bound state and the resonance found in the above three-channel calculation
now become a sharp resonance and a cusp in the \mbox{$\Lambda_c$}\mbox{$\overline{D}{}^*$}\ channel, respectively.
The energy of the resonance or the cusp does not move much when the $N\mbox{$J\!/\!\psi$}$ and \mbox{$\Lambda_c$}\mbox{$\overline{D}{}^*$}\ channels are introduced.
Their energies are summarized in Table \ref{tbl:prob0s}.
Among the $\langle {\cal P}_{cs}\rangle/\langle{\cal P}\rangle$ evaluated by the scattering wave function with the initial $N$\mbox{$J\!/\!\psi$}\ channel,
the factor to find {[$q^3$8\half3]}\ is the largest
in both of the resonance and in the cusp.
In table \ref{tbl:prob0s}, we list their proportions at the resonance or the cusp energy
(with identifications of B and C).
In the five-channel calculation, the proportion of {[$q^3$1\half1]}\ becomes larger than that of the three-channel calculation, but still the {[$q^3$8\half3]}\ component is the largest.
The listed energy of each of the structures is that where the $\langle {\cal P}\rangle$ becomes local maximum.
All the resonance and cusp energies read from the phase shifts differ by less than 1 MeV from
the listed ones except for the resonance C,
where the phase shift increases up to above $\pi/2$ at by 4 MeV above the listed energy.
\subsection{$uudc\overline{{c}}$ $I(J^P)$=$\half1(\half1^-)$ states}
The $uudc\overline{{c}}\ \half1(\half1^-)$ states consist of seven baryon-meson channels:
$N\mbox{$\eta_c$}$, $N\mbox{$J\!/\!\psi$}$, \mbox{$\Lambda_c$}\overline{{D}}{}, \mbox{$\Lambda_c$}\mbox{$\overline{D}{}^*$}, \mbox{$\Sigma_c$}\overline{{D}}{}, \mbox{$\Sigma_c$}\mbox{$\overline{D}{}^*$},
and \mbox{$\Sigma_c^*$}\mbox{$\overline{D}{}^*$},
whereas there are five $(0s)^5$ states.
So, there are two forbidden states when the system is
totally symmetric in the orbital space.
Also in this case, all the diagonal elements
of the normalization are close to 1;
no baryon meson
state is affected strongly by the quark Pauli-blocking.
Again, we first discuss the results of the three-channel quark cluster model calculation:
\mbox{$\Sigma_c$}\overline{{D}}{},
\mbox{$\Sigma_c$}\mbox{$\overline{D}{}^*$}, and
\mbox{$\Sigma_c^*$}\mbox{$\overline{D}{}^*$}.
There is one bound state with the binding energy 0.13 MeV,
but no resonance is found.
As is seen from the entry with an identification D$'$ in Table \ref{tbl:prob0s}, major component of this bound state is {[$q^3$8\half3]}.
In the antisymmetrized \mbox{$\Sigma_c$}\overline{{D}}{}\ state, the proportion is (0.05 0.19 0.76).
Thus this bound state is essentially an antisymmetrized \mbox{$\Sigma_c$}\overline{{D}}{}.
In figure \ref{fig:ps21},
we show the diagonal phase shifts
of the 7-channel calculation
of this system.
There is a sharp resonance in the \mbox{$\Lambda_c$}\mbox{$\overline{D}{}^*$}\ channel,
which corresponds to the bound state of the three-channel calculation.
In Table \ref{tbl:prob0s}, we list $\langle {\cal P}_{cs}\rangle/\langle{\cal P}\rangle$
at the resonance energy under an identification D.
The factor to find the {[$q^3$8\half3]}\ configuration becomes
small compared to that of the three-channel calculation;
the {[$q^3$1\half1]}\ configuration becomes the largest at the resonance.
As we will show later,
the coupling to the $N\mbox{$J\!/\!\psi$}$ channel to the \mbox{$\Lambda_c$}\overline{{D}}{}\ and \mbox{$\Sigma_c$}\overline{{D}}{}\ channels is
stronger in this
$\half1(\half1^-)$ case.
The existence of the attraction in the \mbox{$\Sigma_c$}\overline{{D}}{}\ or {[$q^3$8\half3]}, however, is important to create a resonance.
\begin{figure}[bt]
\begin{center}
\includegraphics[width=9cm]{ps_21all_v2.eps}
\caption{The diagonal scattering phase shifts of the $S$-wave $uudc\overline{{c}}$ $I(J^P)$=$\half1(\half1^-)$ system.
The solid and the dotted lines are those of the $N\mbox{$J\!/\!\psi$}$ and the $N\eta_c$ channels,
the long-dashed and the long-dot-dashed lines are for
\mbox{$\Lambda_c$}\overline{{D}}{}\ and \mbox{$\Lambda_c$}\mbox{$\overline{D}{}^*$} channels,
and the dashed, double-dot-dashed, and triple-dot-dashed lines are
for the \mbox{$\Sigma_c$}\overline{{D}}{}, \mbox{$\Sigma_c$}\mbox{$\overline{D}{}^*$}, and \mbox{$\Sigma_c^*$}\mbox{$\overline{D}{}^*$} channels,
respectively. (Color online.)
}
\label{fig:ps21}
\end{center}
\end{figure}
\subsection{$uudc\overline{{c}}$ pentaquarks}
As summarized in Table \ref{tbl:prob0s},
in the present model,
one bound state is found in $\half1(\half5^-)$,
one resonance and a cusp in $\half1(\half3^-)$, and
one resonance in $\half1(\half1^-)$.
As seen in Table \ref{tbl:hq},
these structures exactly correspond to the
number of {[$q^3$8\half3]}\ configurations in the negative parity baryon meson channels.
The energy of the color-octet $uud$ configuration itself
is higher than that of the color-singlet $uud$ configurations.
CMI works as an attractive force in the {[$q^3$8\half3]}\ configuration
just because its energy is lower than the that of the
relevant baryon-meson threshold, $\Sigma^{(*)}\overline{{D}}{}{}^{(*)}$.
So, when the model space is enlarged from the $\Sigma^{(*)}\overline{{D}}{}{}^{(*)}$
to all the relevant baryon-meson systems, for example, D$'$ to D,
the proportion of the {[$q^3$8\half3]}\ configuration becomes smaller,
from 0.73 to 0.12.
The resonance, however, is still there.
Situation is similar for the resonances B and C,
though the reduction of the proportion
by introducing the lower channels is less extreme.
The {[$q^3$8\half3]}\ configuration plays an important role even though the
proportion becomes small.
These structures are considered to be an appearance of the hidden color-octet $uud$ baryon.
These resonances and cusp in $(\half1^-,\half3^-)$ exist in the energy range of the $P_c(4380)$ peak,
$E\pm \Gamma/2 =$ 4278--4483 MeV.
All the structures are very close to the baryon meson thresholds,
and each of the resonances has a width of a few MeV,
which is far smaller than the widths of the observed peaks.
Including the light meson exchange effects
between the $Y_c$ baryon and the \overline{{D}}{}\ meson may enlarge the width,
which is an interesting topic and will be investigated in future works.
We argue that these resonances and cusp may combine to form the broad peak of $P_c(4380)$.
Or, if the parity of $P_c(4450)$ is found to be negative in future experiments,
the cusp at 4458 MeV may correspond to that peak.
The bound state in $\half5^-$, whose energy is higher than both of the observed pentaquark peaks, does not couple to the $S$-wave $N\mbox{$J\!/\!\psi$}$ channel;
the higher partial wave mode is necessary to
see this bound state from the $N\mbox{$J\!/\!\psi$}$ channel.
There are arguments that
peaks which correspond to the pentaquarks should appear at
$\pi N\rightarrow \mbox{$J\!/\!\psi$} N$ \cite{Kim:2016cxr}
or
$\gamma N\rightarrow \mbox{$J\!/\!\psi$} N$ \cite{Wang:2015jsa}
reactions
if
the coupling between $N\mbox{$J\!/\!\psi$}$ and the pentaquarks
is large.
The diagonal elasticities
of the present calculation, $\eta$, which is
the absolute value of the scattering matrix of the same initial and final channels,
are plotted in Figure \ref{fig:eta}.
In $\half1(\half3^-)$, $\eta$ of the $N\mbox{$J\!/\!\psi$}$ channel
goes down only
to 0.81 at the resonance C
or to 0.70 at the cusp B.
The elasticity of the $\half1(\half1^-)$ goes down to 0.65 at the resonance D.
In both of the channels, the mixing between $N\mbox{$J\!/\!\psi$}$ and the other channels is rather small
except for the resonance energies,
because the quark overlap is small and
because the rearrangement of the charm quark
is necessary.
The average
inelasticity of $N\mbox{$J\!/\!\psi$}$, $1-\eta$,
over the energy range of the broad $P_c(4380)$ width
is about 0.1--0.2.
It should be checked whether this value
is consistent with the above \mbox{$J\!/\!\psi$}\ production experiments.
\begin{figure}[t]
\begin{center}
\includegraphics[width=8cm]{eta_23all.eps}
\includegraphics[width=8cm]{eta_21all.eps}
\caption{The diagonal elasticity of the $S$-wave $uudc\overline{{c}}$ $I(J^P)$=$\half1(\half3^-)$ (figure (a)),
and $\half1(\half1^-)$ systems (figure (b)).
For the notation of figures (a) and (b),
see fig.\ \ref{fig:ps23} and
fig.\ \ref{fig:ps21}, respectively.
}
\label{fig:eta}
\end{center}
\end{figure}
The estimate of the cross section $\sigma_{N\mbox{$J\!/\!\psi$}}$
from the photo production experiment ($\gamma N\rightarrow \mbox{$J\!/\!\psi$} N$)
is $3.5 \pm 0.8 \pm 0.5$ mb \cite{Anderson:1976hi}.
It corresponds to the scattering length $|a_{N\mbox{$J\!/\!\psi$}}| = 0.17^{+0.02}_{-0.03}$
fm by using $\sigma=4\pi a^2$.
This scattering length has been
calculated by many theories.
The QCD sum rule gives
$a_{N\mbox{$J\!/\!\psi$}} = -0.10\pm 0.02$ fm
\cite{Hayashigaki:1998ey}.
The effects of the QCD van der Waals force
was estimated by \cite{Brodsky:1989jd,Brodsky:1997gh},
$a_{N\mbox{$J\!/\!\psi$}}=-0.24$ fm.
Two quenched Lattice QCD calculations
were reported:
the spin averaged scattering length was obtained as
$-0.71\pm 0.48$ fm or $-0.39\pm 0.14$ fm
\cite{Yokokawa:2006td},
or about
$-0.35$ fm (read from figure)\cite{Kawanai:2010ru}.
In the present calculation, the
$\mbox{$J\!/\!\psi$} N$ scattering length is
$-0.077$ fm for $J$=\half3 and
$-0.103$ fm for $J$=\half1.
The spin averaged value is $a_{N\mbox{$J\!/\!\psi$}} =-0.085$ fm.
Since the quark interaction does not produce
a direct interaction between
$N$ and the $c\overline{{c}}$ mesons,
this attraction here solely comes from
the channel coupling between the $N\mbox{$J\!/\!\psi$}$ and the \mbox{$\Lambda_c$}\overline{{D}}{}$^{(*)}$ and \mbox{$\Sigma_c$}$^{(*)}$\overline{{D}}{}$^{(*)}$
channels.
Our results
suggest that
roughly half of the observed attraction between the $N$ and \mbox{$J\!/\!\psi$}\ comes from the channel coupling.
In this work, we employ the quark interaction arising from the gluons.
It is because
we would like to investigate the features of the color-octet $uud$ baryons,
which newly appears in the hidden-charm pentaquarks.
As for the \mbox{$\Lambda_c$}\overline{{D}}{}\ or \mbox{$\Sigma_c$}\overline{{D}}{}\ baryon meson channel, however,
it is necessary to include the pion-exchange force in the long range region.
As is reported in ref.\ \cite{Yamaguchi:2016ote},
the meson-exchange models may give many bound states and resonances
though the results seem to depend strongly on the cutoff value.
It is very interesting to see whether the energies of the currently obtained resonances move
or their widths become broader by introducing the meson-exchange in our model.
Moreover, in order to discuss the $P_c(4450)$, which has an opposite parity to the $P_c(4380)$,
simultaneously,
the $P$-wave baryon-meson relative motion and the positive parity mesons should be introduced.
Since the lowest orbital excitation
is considered to be the $S$-wave $uud$ configuration with the $P$-wave $c\overline{{c}}$ pair,
the color-octet $uud$ configuration may again play an important role there.
We take both of them as future problems.
It is also interesting to investigate the hidden-charm pentaquarks with the strangeness,
where the flavor-octet $uds$ may play a similar role to the present work.
As for the isospin-\half3 $uuuc\overline{{c}}$ systems, the three light quarks are either
color-octet spin-\half1
or color-singlet spin-\half3.
Because CMI of these light quark configurations
contributes repulsively,
and because the Pauli-blocking gives strong repulsion there,
all the five-quark systems
seem to dissolve by the baryon-meson couplings,
which we will also discuss elsewhere.
The isospin-\half1,
or the flavor-octet hidden-$Q\overline{{Q}}$ baryons are a special place to look into for the color-octet $uud$ baryons.
When we introduce the $b\overline{{b}}$ pairs instead of $c\overline{{c}}$ to the five-quark systems,
the situation will become more manifest.
The system is closer to the heavy quark limit,
and the thresholds of the
$\Sigma_b^{(*)}B^{(*)}$ are closer to each other.
The four structures we find in the $uudc\overline{{c}}$ states are also found to exist.
The bound state in $J^P=\half5^-$ has a binding energy of more than 10 MeV.
We discuss the system with the $b\overline{{b}}$ pairs as well as those with
the $s\overline{s}$ pairs elsewhere.
\section{Conclusions}
The $I(J^P)=\half1(\half1^-)$, $\half1(\half3^-)$, and $\half1(\half5^-)$ $uudc\overline{{c}}$ systems
are investigated by the quark cluster model.
There is no strong repulsion due to the quark Pauli blocking
in the relevant baryon meson systems.
It is shown that the color-octet isospin-\half1 spin-\half3 $uud$ configuration
gains an attraction from the color magnetic interaction.
The $uudc\overline{{c}}$ states with this configuration
cause structures around the $\mbox{$\Sigma_c$}{}^{(*)}\overline{{D}}{}{}^{(*)}$
thresholds.
We have found
one bound state in $\half1(\half5^-)$,
one resonance and a cusp in $\half1(\half3^-)$, and
one resonance in $\half1(\half1^-)$
in the negative parity channels.
We argue that one of these structures may give $P_c(4450)$ if its parity is found to be negative, or
these resonances and cusp may combine to form the broad peak of $P_c(4380)$.
\bigskip
We would like to thank Professors M.\ Oka and
A.\ Hosaka for
useful discussions.
|
\section{Introduction}
\label{sec:intro}
Star formation is a slow process on galactic size and time scales,
with a mere $\sim 2\%$ of the gas mass turning into stars in the disk dynamical time
\citep{kennicutt89, kennicutt98}. Stars
in the Milky Way and nearby galaxies form in giant molecular clouds (GMCs),
with the mass in newly formed stars proportional to the mass in host GMCs \citep[see e.g.,][]{mooney88,scoville89}.
The bulk of the molecular gas resides in the most massive
GMCs \citep[e.g.,][]{solomon87} and, as implied by the results of \citet{mooney88}, most
star formation occurs in the most massive GMCs \citep[see also e.g.,][]{murray11}.
There is considerable disagreement regarding the rate of star formation and the
star formation efficiencies on scales of GMCs and smaller.
Star formation efficiency $\epsilon$ is defined as the ratio of the mass in
protostars to the total mass in a given star-forming region:
\be \label{eqn:efficiency}
\epsilon \equiv \frac{M_\star}{M_g+M_\star},
\ee
where the star-forming region may be a GMC or a smaller sub-region of a GMC.
The star formation rate per free fall time $\epsilon_{\rm ff}$ is
\be \label{eqn:sfe_ff}
\epsilon_{\rm ff}\equiv \epsilon \frac{\tau_{\rm ff}}{\tau_\star},
\ee
where $\tau_*$ is the lifetime of the (proto-)stellar object in question, and
$\tau_{\rm ff}\equiv\sqrt{3\pi/32G\rho}$ is the free fall time
of the star-forming region (a GMC or its sub-region),
which is assumed to have a mean density $\rho$.
\citet{mooney88} and \citet{scoville89} showed that there is
a wide (maximum to minimum of approximately $2.5$ dex)
spread in the efficiencies of star formation in GMCs, a measurement based on
the ratio of far infrared luminosity $L_{\rm FIR}$ to CO luminosity $L_{\rm CO}$.
More recent estimates also employ
counts of protostars in nearby Milky Way molecular clouds
\citep[e.g.][]{evans09, heiderman10, lada10}, or measurements of
the free-free emission associated with massive stars \citep{murray11}
to find a similarly large spread.
\citet{heiderman10} and \citet{lada10} (as well as \citealt{evans09} to a
smaller degree) find broad distributions
of $\epsilon_{\rm ff}$, which range both well below and
well above $\epsilon_{\rm ff}=0.02$, by factors of $\sim 20$ or
more in either direction; they also note that Galactic clouds
with high surface densities may have
higher-than-expected $\epsilon_{\rm ff}$ compared to
their extragalactic counterparts.
\citet{kdm12} argue that the broad distribution
in $\epsilon_{\rm ff}$ of Galactic molecular clouds can be
explained by variations in volumetric
densities (the density determining the free-fall time) among
clouds. They argue that the star formation rate
on all scales is $\epsilon_{\rm ff}\approx 0.02 M_{\rm GMC}/\tau_{\rm ff}$,
with only a factor of 3 scatter above and below the mean value,
once the variations in $\tau_{\rm ff}$ are taken into account
\citep[see also e.g.,][]{krumholz07}.
However, using a sample of clouds from c2d and Gould Belt {\it Spitzer} legacy programs,
\citet{evans14} find a large scatter in $\epsilon_{\rm ff}$
even after taking into account the variations in volumetric densities.
Similarly, \citet{heyer16} study dense clumps in
ATLASGAL survey and find $\epsilon_{\rm ff} \sim 0.001$--0.01 (see their Figure 10c).
The rate of small scale star formation, and the dispersion of its distribution,
is important for galaxy formation. If the small scale star formation rate is steady
and hence low, the mass in live stars will be steady, and small.\footnote{We
refer to stars less than 4 Myrs old as ``young" or ``live" stars. Clusters
containing more than $\sim5000 M_\odot$--$10000 M_\odot$,
which sample the initial mass function (IMF)
fairly well, have ionizing
photon rates $Q\,(\s^{-1})$ that are dominated by massive stars with $M_\star\gtrsim30M_\odot$. Such
stars have lifetimes of order 4 Myrs.}
It follows that the kinetic and thermal feedback from stellar winds,
radiation, and supernovae will be steady and low.
If the small scale star formation is sporadic, with pronounced peaks and long-lived lows,
the feedback will be both temporally and spatially concentrated. Because stars form
preferentially in dense gas, the resultant feedback will be deposited in regions of
dense gas, where it has the potential to move the most material, but also where cooling is
most rapid. Observations of starburst galaxies show that massive star clusters are
prominent sources of galaxy scale winds \citep[e.g.,][]{schwartz06}. These winds
are believed to be crucial for determining both the global star formation rate and the
total stellar mass by regulating the amount of gas in the disk \citep[e.g.,][]{oppenheimer10,hopkins11,hopkins14},
as well as the distribution of metals in the intergalactic medium \citep[e.g.,][]{oppenheimer10}.
For a fixed global star formation rate, sporadic small scale star formation will tend
to produce more massive clusters than will steady small scale star formation, with
important consequences for wind properties.
Star formation is promoted by gravity and by convergent fluid flows, and
suppressed by a number of physical effects including thermal gas pressure,
turbulent kinetic pressure, magnetic fields, and stellar feedback --- i.e. stellar winds,
radiation pressure, protostellar jets, and, at late times, supernovae.
Two leading candidates (at the time of writing) for the suppression
of star formation rates are stellar feedback and turbulent pressure support.
On galactic scales, the rate of star formation is believed to
be regulated by stellar feedback, which can keep the gas disk in
a state of marginal stability \citep[e.g.,][]{TQM05, Ostriker_Shetty11}.
The turbulent pressure scenario
is well motivated: the large linewidths seen in massive star forming regions
\citep[e.g.,][]{Caselli_Myers95,Plume97}
show that the kinetic energy density greatly exceeds the thermal pressure on scales larger
than $\sim 0.01\pc$, and is comparable to the gravitational potential
energy density.\footnote{The two pictures --- stellar feedback
and turbulence --- are not necessarily
in conflict, since a source of energy is needed to power the turbulence
seen in GMCs, and stellar feedback may provide this energy.}
A number of authors have suggested that these turbulent motions
support GMCs and hence slow the rate of star
formation. The extreme version of the argument says that turbulence maintains GMCs
in hydrostatic equilibrium, preventing large scale collapse \citep[e.g.,][]{Myers_Fuller_92,mclaughlin97,mckee03}.
If GMCs are in hydrostatic
equilibrium (which implies that the clouds live for at least a few free fall times)
then $\epsilon_{\rm ff}= 0.02$ implies that the mass in live stars is roughly the
same in most GMCs of a given mass. However, recent numerical and semi-analytic studies show
that $\epsilon_{\rm ff}$ increases roughly linearly with time
\citep[e.g.,][]{lee15,murray15,dmurray15}.
If this is true, then most $10^6M_\odot$ GMCs will have
very few live stars, while a small subset of $10^6M_\odot$ GMCs will host very
massive clusters of live stars; the distribution of $\epsilon_{\rm ff}$
will be very broad.
Measuring the width of the distribution in $\epsilon_{\rm ff}$ and $\epsilon$
therefore provides an important diagnostic for testing the idea that
$\epsilon_{\rm ff}$ is independent of time.
We will show that the observed scatter in $\epsilon_{\rm ff}$ of
the Milky Way GMCs is significantly larger than what is predicted by models of
constant star formation rate per free fall time~\citep[e.g.,][]{krumholz05,padoan12}.
To estimate either $\epsilon$ or $\epsilon_{\rm ff}$ we must estimate
the mass in young stars. We use the free-free flux to do so.
Our sample of star-forming Milky Way GMCs is built by cross-correlating
a new all-sky cloud catalog from M-A., Miville-Desch\^enes et al.
(2016, in preparation; MML16 from hereon) with star forming complexes
(SFCs) from \citet{lee12}.
This paper is organized as follows: in Section \ref{sec:sfgmc}, we match
SFCs with GMCs in (l,b,v)-space (galactic longitude, galactic latitude, radial velocity);
in Section \ref{sec:flux}, we describe how we convert free-free flux into stellar mass;
in Sections \ref{sec:sfe} and \ref{sec:sfr_ff}, we present our
analysis of $\epsilon$ and $\epsilon_{\rm ff}$, respectively;
in Section \ref{sec:cause}, we compare
various models of star formation rate to observations; in Section \ref{sec:sigma},
we present the surface density star formation rate profile across the Galactic
plane; we summarize and discuss our results in Section \ref{sec:disc}.
\section{Star Forming Giant Molecular Clouds}
\label{sec:sfgmc}
In this section, we describe how we cross-correlate the SFCs
in \citet{lee12} with the GMC catalog of MML16.
\subsection{Star Forming Complexes}
\label{ssec:sfc}
We present a brief description of the SFC catalog here \citep[see][for more detail]{lee12}.
\citet{lee12}, following the approach of \citet{murray10} and \citet{rahman10}, identified
280 SFCs in and near the Galactic plane.
Ionizing photons from young clusters can travel
tens or even hundreds of parsecs through the interstellar medium before being absorbed.
It follows, and it is observed, that free-free emission regions can be much larger
($\sim 1$--2$^\circ$ in diameter)
than the star clusters that power them.
In addition, star clusters tend to form in associations (just like
stars tend to form in clusters),
so that the free-free emission seen
in a given direction may be powered by more than one star cluster.
Star forming complexes represent systems of clusters that are able to carve out
bubbles of size $\sim$10--100 parsecs in their host GMCs.
We assign a free-free flux $f_{\nu}$ to each SFC by dividing up the
free-free flux of their host {\it WMAP}
sources (the 1--2$^o$ wide regions of free-free emission) in
proportion to the 8$\mu$m flux of each SFC, as seen by {\it Spitzer}.
This is motivated by the linear correlation
between free-free and 8$\mu$m flux seen on small (parsec) scales.
\citet{lee12} calculated the distance to each SFC by fitting the galactic longitude
and the central local standard at rest velocity $v_{\rm lsr}$ to the \citet[hereafter C85]{clemens85}
rotation curve. For SFCs that reside inside the solar circle ($R_\odot\leq8.5\kpc$),
they use the radial velocities of the absorption lines along the line of sight
to disambiguate between near (absorption line velocity is less than that of
the SFC) and far (absorption line velocity can be as large as the tangent point
velocity) distances.
\begin{figure}
\includegraphics[width=0.5\textwidth]{sfc_dist_comp_brand}
\caption{\label{fig:dist_comp}Comparison of the kinematic distances to SFCs
inferred from using the \citet{clemens85} and \citet{brand93} rotation curves.
Black circles represent near
distances, while red crosses represent far distances. The blue dashed line
shows the relationship that would be found if the two curves gave the
same distances. For the four outliers, where the \citet{clemens85} curve predicts
near-zero distances, the \citet{brand93} curve predicts a unique large distance.
The \citet{clemens85} curve yields a galactocentric radius that is slightly inside the solar circle,
while the \citet{brand93} rotation curve finds a galactocentric radius that is slightly outside the solar circle.}
\end{figure}
\subsection{Rotation Curve and Kinematic Distance}
\label{ssec:rotcurve}
The rotation curve calculated by C85 is known to introduce
significant errors in the distance to the objects in the outer galaxy due to
Perseus arm streaming motions~\citep{fich89}.
We therefore update the \citet{lee12} distances using the rotation curve of
\citet[][hereafter BB93]{brand93} at $R_0 = 8.5 \kpc$ and $\Theta_0 = 220 \kms$.
Figure \ref{fig:dist_comp} shows that the distances given by the two rotation curves are
generally in good agreement.
The four outliers with C85 `near' distances that are close to zero have C85
solutions that are slightly inside (by $\sim 0.03\kpc$) the solar circle,
while the BB93 solutions for these objects are slightly outside
(by $\sim 0.12\kpc$) the solar circle.
We reject solutions with galactocentric radius greater than 16 kpc, i.e., we are using a prior
that there is no significant massive (O star) star formation at such large radii.
The typical error in distance is $\sim$35\%, the dominant sources of error being
the streaming velocity of ionized gas, the expansion velocities of bubbles carved out
by SFCs, and the motions of spiral arms~\citep{lee12}.
\subsection{Giant Molecular Clouds}
\label{ssec:gmccat}
We use the new all-sky catlaog of Galactic molecular clouds
presented by MML16, where we describe in detail how
clouds are identified. MML16 present a number of correlations between
cloud properties. We provide a brief summary here.
Clouds are identified as coherent molecular structures from
the $^{12}$CO~(J1-0) survey of \citet{dame01}.
This survey has a modest angular resolution (7.5\,arcmin) compared to
other observations, e.g., \citet{jackson2006}, but it has the advantage of providing
a uniform data set covering the entire Galactic plane in longitude.
We limit our study to $-5^\circ < b < 5^\circ$.
The identification of clouds from position-position-velocity (or [$l$, $b$, $v$])
cubes is challenging, as the interstellar medium (ISM) has a fractal structure,
and because unrelated clouds along a given line of sight can have similar
projected velocities.
Given this difficulty, it is striking that
studies using different data sets and different structure
identification techniques find
relatively consistent
scaling relations between various cloud properties,
such as size, velocity dispersion,
mass and surface density \citep[see the review by][]{hennebelle2012a}.
MML16 find scaling relations similar to those reported in the earlier catalogs.
MML16 employ a combination of Gaussian decomposition
and hierarchical clustering analysis to identify clouds.
First, every spectra of the the entire CO position-position-velocity (PPV) cube
is decomposed into a sum of Gaussian components.
Next, they build a cube of integrated emission $W_{\rm CO}$
where each grid cell ($l$, $b$, $v$) is assigned
$W_{\rm CO}$ integrated over Gaussian components
whose central velocities equal $v$.
This cube is much more sparse than the original
brightness cube as the integrated emission of each
Gaussian component is concentrated in a single cell
($l$, $b$, $v$) and not spread out in velocity. This
new cube facilitates the identification of coherent
structures down to the noise level of the data. To do
so, they used a classical threshold descent. First,
they identified islands in PPV space as neighboring
cells with $W_{\rm CO}$ higher than some high threshold value.
The threshold is progressively lowered down to 0.5 ${\rm K\,\kms}$.
At each step of this descent, new cells are revealed.
They are either attached to previously identified structures
or classified as new structures. This watershed method is
similar in spirit to {\em clumpfind} \citep{williams1994}
or to {\em dendrogrammes} \citep{rosolowsky2008}.
A total of 8107 coherent structures / clouds are
identified over the whole Milky Way disk,
recovering 91\% of the total CO emission.
As for the SFCs in this paper, the distance to each
GMC is estimated assuming that its average
velocity follows the Galactic rotation curve of \citet{brand93}. In the
inner Galaxy, there is an ambiguity as two distances along the line of
sight (dubbed near and far) have the same observed velocity.
MML16 relied on a statistical method that
has been used by many other studies
\citep[e.g.,][]{dame1986,solomon87,grabelsky88,garca2014}.
The original idea was to choose the distance that provides a cloud
physical size, $R$, that
matches more closely the size-linewidth
$\sigma_v \propto R^\alpha$ relation, established
using clouds not subject to the distance ambiguity.
The size-linewidth relation appears to vary with the cloud
column density $\Sigma$\,\citep[see e.g.,][]{heyer09}.
To account for this dependence on $\Sigma$,
MML16 select the distance that
best matches the relation $\sigma \propto (R\Sigma)^{0.42}$.
The cloud distances chosen in this
manner agree well with the distances of associated SFCs,
as we show in Section \ref{ssec:correl}.
The cloud catalog provides the position in ($l$, $b$, $v$), the size,
the velocity dispersion, and the distance to each cloud.
It also provides the gas mass estimated as
\begin{equation}
\label{eq:CO mass}
M_{\rm g} = W_{\rm CO}^{\rm tot} \, X_{\rm CO} \, 2 \mu m_{\rm H} \, D^2 \tan(\delta)^2
\end{equation}
where $W_{\rm CO}^{\rm tot}$ is the total CO emission of all the
Gaussian components associated to the cloud,
$X_{\rm CO} = 2 \times 10^{20}$\,cm$^{-2}$\,K$^{-1}$\,km$^{-1}$\,s \citep{bolatto2013},
$\mu=1.36$ takes into account the contribution fron Helium,
$m_{\rm H}$ is the mass
of Hydrogen, $D$ is the distance to the cloud and $\delta = 0.125^\circ$
is the angular size of the pixel.
We also provide the {\it WMAP} free-free and
{\it IRAS} 100$\mu$m flux of clouds,
measured by integrating the emission over pixels associated with each cloud.
\subsection{Cross-Correlating Star Forming Complexes with Giant Molecular Clouds}
\label{ssec:correl}
We identify matches between SFCs and GMCs if they meet the following criteria:
\begin{equation}
\sqrt{(\delta l^2 + \delta b^2 + \delta v^2)} \leq 1
\label{eq:match}
\end{equation}
where
\begin{align}
\delta l &= (l_{GMC} - l_{SFC})/\rm{max}(\sigma_{l,GMC}, R_{SFC}/D,0.5^o) \nonumber \\
\delta b &= (b_{GMC} - b_{SFC})/\rm{max}(\sigma_{b,GMC}, R_{SFC}/D,0.5^o) \nonumber \\
\delta v &= (v_{GMC} - v_{SFC})/\rm{max}(\sigma_{v,GMC}, \sigma_{v,SFC}, 7\kms).
\label{eq:lbv}
\end{align}
By $R_{SFC}/D$ we mean the angular size of the SFC measured in degrees rather than radians;
$\sigma_{v,{\rm GMC}}$ is the RMS velocity dispersion of a GMC while $\sigma_{v,{\rm SFC}}$
is the half-spread velocity of SFCs (i.e., $(v_{\rm max}-v_{\rm min})/2$ along the
bubble walls; see \citealt{lee12} for more detail).
Out of 280 SFCs, 256 have measured median velocities $v_{\rm SFC}$.
Since the host {\it WMAP} sources
are $\sim 2^o$ in size, we only allow SFC-GMC matches to the clouds that are
large enough to encompass more than 10 pixels but
smaller than $2^o$ in their longest axis; larger objects
are likely not isolated, self-gravitating clouds.\footnote{We note that
there are a couple of SFCs with mean radii that are close to
or exceed 2$^o$, such as SFC Nos.~110 and 111. These SFCs are associated with
the Cygnus region within a particularly large {\it WMAP} free-free source
(it was originally identified as three separate free-free sources by \citet{murray10}
but \citet{lee12} merged them after a visual inspection).}
The selection criteria on the GMC size limits the cloud count to 5469.
In summary, we perform a cross-correlation between 256 SFCs and 5469 GMCs.
Using the criteria given by equations \ref{eq:match} and \ref{eq:lbv},
we find that approximately half of SFCs are matched with multiple GMCs.
Visual inspection reveals that these clouds---which are often smaller in size than SFCs---trace the
outer rim of bubbles blown by their SFC counterpart when we overlay them on
{\it GLIMPSE} 8$\mu$m images.
It is likely that the clouds originate from a single massive cloud
that was disrupted by stellar winds and radiation pressure from its
massive star clusters.
These SFC-GMC ``complexes'' are often found inside the solar circle,
so that even though multiple objects (SFCs or GMCs) coincide
in the $(l,b,v)$-space, they can be at vastly different distances.
We reject any GMC that is not within $0.3\,d_{\rm SFC}$ from
the centroid of the host SFC, where $d_{\rm SFC}$ is the
heliocentric distance of the SFC.
For SFCs without a distance measurement, we use the distance
of the GMC that is most closely matched in the ($l,b,v$)-space.
Lastly, we make a visual inspection for the luminous SFCs
(those with expected $M_\star \geq 10^4 M_\odot$)
to ensure that we have made a sensible cross-correlation.
Only $\sim$10\% of the initial SFC-GMC complexes are mismatched.
Most of the mismatches stem from conflicting distance ambiguity
resolution between the SFC catalog and the GMC catalog.
The fact that most of the luminous SFCs are matched
with GMCs suggests a good agreement in distance ambiguity resolution
between the SFC sample (resolution by absorption line velocities)
and the GMC sample
(resolution by a fit to a size-linewidth-column-density relation),
enhancing our confidence in the use of the latter technique.
We identify a known cluster from \citet{morales13} in the
direction of SFC-GMC complexes to reassign SFC Nos. 27, 28, 93
to near distances and
GMC Nos. 136, 171, 388, 398, 446, 523, 625, 812, 1312,
1656, 2733, 2734, 2816, 3422
to far distances.\footnote{SFC Nos. 27 and 28 are associated with the W31
cluster while SFC No. 93 is associated with [BDS2003] 135~\citep{bica03}.}
Distances to SFC Nos. 36 and 252 are changed to
far distances after correcting for a bug in a code
used to produce the catalog in \citet{lee12}.
We also manually match SFC No. 202 with GMC No. 2420,
SFC No. 65 with GMC No. 2071 (associated with the W40 cluster~\citep{mallick13}),
and SFC No. 251 with GMC No. 482 (associated with RCW 120~\citep{anderson10}).
Four SFCs (Nos. 31, 35, 245, 248) had no GMC match because their
distances were truncated to 12 kpc (because of their likely
association with $\sim$3\kpc ring; see \citealt{lee12});
we recalculated their distances.
We present the results of our cross-correlation in Table \ref{table1}.
The final result is 191 unique SFCs matched to 389 unique GMCs,
recovering 93.5$\%$ of the total SFC free-free luminosity, 83.8$\%$ of
the total SFC free-free flux, and 9\% of total GMC gas mass. All of
the top 24 most luminous SFCs are matched to at least one GMC. For
the rest of the paper, we identify these 389 GMCs matched to 191 SFCs
as 191 SFC-GMC complexes.
Each complex inherits the sum of the CO fluxes and gas masses of the
matched GMCs, as well as the mass-weighted mean Galactic coordinates $l,b$,
velocity $v$, heliocentric distance $d$, and galactocentric radius $R_{\rm gal}$.
Following MML16, we calculate the angular size and surface area of each SFC-GMC
complex by solving for eigenvalues of the moment of inertia matrix:
\begin{equation}
\label{eq:inertia}
\phi = \left[ \begin{array}{cc}
\sigma_l^2 & \sigma^2_{lb}\\
\sigma^2_{lb} & \sigma_b^2
\end{array} \right]
\end{equation}
where
\begin{gather}
\begin{aligned}
\sigma_l^2 &= \frac{\sum_i M_i \sum^4_{c=1} (l_{c,i} - l_{\rm SG})^2}{4\sum_i M_i} \\
\sigma_b^2 &= \frac{\sum_i M_i \sum^4_{c=1} (b_{c,i} - b_{\rm SG})^2}{4\sum_i M_i} \\
\sigma^2_{lb} &= \frac{\sum_i M_i \sum^4_{c=1} (l_{c,i} - l_{\rm SG}) (b_{c,i}-b_{\rm SG})}{4\sum_i M_i}.
\end{aligned}
\end{gather}
Here, $i$ denotes each constituent MML16 cloud, $M_i$ is the individual
cloud mass, and $(l_{\rm SG}, b_{\rm SG})$ are the mass-weighted
mean Galactic coordinates of the host SFC-GMC complex.
The quantities $(l_{c,i}, b_{c,i})$ are
the Galactic coordinates of the 2 semi-major and 2 semi-minor vertices
($c = 1$--4) of each cloud, whose semi-major axis we define as
3$\sigma_l$ and semi-minor axis as 3$\sigma_b$.
Using the two eigenvalues $R_{\rm max}$ and $R_{\rm min}$ of $\phi$,
we define the angular radius of each SFC-GMC complex as
$R_{\rm ang} \equiv (R_{\rm max} R_{\rm min}^2)^{1/3}$.
We define the velocity dispersion of the SFC-GMC complex as
\begin{equation}
\label{eq:sigV_sg}
\sigma^2_{\rm v, SG} = \frac{\sum_i M_i (\delta v_i^2 + \delta_+ v_i^2 + \delta_- v_i^2)}{\sum_i M_i}
\end{equation}
where $\delta v_i = v_{\rm cent,i} - v_{\rm SG}$,
$\delta_+ v_i = v_{\rm cent,i} + \sigma_{v,i} - v_{\rm SG}$,
and $\delta_- v_i = v_{\rm cent,i} - \sigma_{v,i} - v_{\rm SG}$
with $v_{\rm cent,i}$ the central velocity
of each cloud $i$, $v_{\rm SG}$ the mass-weighted velocity
of the host SFC-GMC complex, and $\sigma_{v,i}$ the velocity dispersion
of each cloud $i$.
We have verified that the SFC-GMC complexes follow
the $\sigma_v$-$M_g$ and $\alpha_{\rm vir}$-$M_g$
(where $\alpha_{\rm vir} \equiv 5 \sigma_v^2 R_g / G M_g$
is the cloud virial parameter in which $R_g = d \tan(R_{\rm ang})$)
relations reported in MML16.
The properties of SFC-GMC complexes are presented
in Table \ref{table2}.
\input{gmc-sfc-match}
\input{sfc-gmc-complex}
\section{Calculating Stellar Mass from Free-Free Flux}
\label{sec:flux}
We aim to measure the spread in the distribution of clouds' star
formation efficiencies $\epsilon$ (equation \ref{eqn:efficiency})
and star formation rates $\epsilon_{\rm ff}$ (equation \ref{eqn:sfe_ff}).
We will compare the observed scatter with what is
expected from turbulence-regulated star formation \citep[][KM05 from hereon]{krumholz05}.
The mass of stars associated with each cloud is evaluated from
the {\it WMAP} free-free fluxes.
We provide a brief summary of photometry here; readers interested in
more details are referred to \citet{lee12}.
For a given {\it WMAP} free-free source
(the large $\sim$1--2$^\circ$ wide regions identified by
their peak free-free flux in {\it WMAP}; see \citealt{murray10}
for more detail), we first perform aperture photometry to
compute the total free-free flux.
The total flux is divided into constituent
SFCs (and by extension constituent SFC-GMC complexes),
proportional to the relative SFC 8$\mu$m fluxes computed
from {\it Spitzer} GLIMPSE and MSX images.
The gas mass of the cloud is calculated from the CO flux $W_{\rm CO}$.
A measure of $\epsilon$ can then be probed by the flux ratio
free-free $f_\nu^{\rm br}$ over $W_{\rm CO}$:
\begin{equation}
\label{eqn:epsilon br}
\epsilon_{\rm br}\equiv {a\,f_\nu^{\rm br}/ W_{\rm CO}\over 1 + a\,f_\nu^{\rm br}/ W_{\rm CO}}
\end{equation}
where the subscript br stands for Bremsstrahlung.
To understand the constant $a$, we review
how $f_{\nu}$'s are
converted to the mass in live stars $M_\star$
and how $f_{\rm CO}$'s are
converted to the gas mass $M_g$.
A cloud with $f_\nu^{\rm br}$ located at a distance $D$ has a luminosity
$L_\nu^{\rm br} = 4\pi D^2 f_\nu^{\rm br}$.
Powering this luminosity requires a streaming rate of ionizing photons of
${\cal Q} = 1.34\times 10^{26}\,(L_{\nu}/\erg\,\s^{-1}\,{\rm Hz}^{-1})~\rm{s}^{-1}$.
The ionizing luminosity is converted to $M_\star$ using
the ratio of ${\cal Q}$ to $M_\star$ averaged over
the modified Muench initial mass function (IMF) from \citet{murray10}:
\be
{\left<m_{*}\right>\over \left<q\right>} = 1.6\times10^{-47} \rm{s}^{-1} \rm{M_{\odot}}.
\label{eqn:mq}
\ee
The live stellar mass in the cloud is then
$M_\star = 1.37{\cal Q}(\left<m_\star\right>/\left<q\right>)$
where the numerical factor 1.37 accounts for the absorption of ionizing photons by dust
(which compete with hydrogen atoms as a sink of ionizing photons)
following \citet{mckee_williams97}.
The conversion between $W_{\rm CO}$ and the gas mass $M_g$ is given
by equation (\ref{eq:CO mass}).
We can now define the constants $a$ and $b$:
\begin{align}
\label{eqn:conv_a}
a &= \frac{M_\star/f_\nu^{\rm br}}{M_g/W_{\rm CO}} \nonumber \\
&= \frac{4\pi\times 10^{-23} \times ({\cal Q}/L_\nu)\times 1.37 \times (\left<m_\star\right>/\left<q\right>)}{X_{\rm CO}\,2\mu m_{\rm H}\,\tan(\delta)^2},
\end{align}
where $10^{-23}$ is the conversion factor from jansky to cgs units,
$X_{\rm CO} = 2\times 10^{20}\,{\rm cm}^{-2}\,{\rm K}^{-1}\,{\rm s}$,
$\mu = 1.36$ to take into account helium, and $\delta = 0.125^\circ$ is
the pixel scale. Note the distance $D$ does not appear in the expression
for $\epsilon$. Any error in the distance measurement will therefore not affect
the scatter in the star formation efficiency.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{Myso_Mflux}
\caption{Comparison between the stellar mass of known
young clusters as probed by counting young stellar objects (YSOs)
and the stellar mass in the corresponding SFCs
as probed by free-free flux. The dashed line delineates $M_{\star,\rm flux} = M_{\star,\rm YSO}$.
We adopt YSO counts for the Orion nebula cluster (ONC) from
\citet{DaRio12}, W5 from \citet{koenig08}, Quintuplet and Arches
from \citet{PZ10} and references therein, Trumpler
14 \& 16 from \citet{PZ10} and \citet{wolk11},
respectively, and W49A from \citet{homeier05}.}
\label{fig2}
\end{figure}
Assuming all star clusters to follow a universal IMF, our
computed stellar mass is reliable for massive
clusters ($M_\star \geq 10^4 M_\odot$; \citealt{krumholz15})
that sample their IMFs well.
There may also be variations in the IMF across
different clusters \citep[see, e.g.,][]{dib14}.
Both the poor sampling of and the variations
in the IMF introduce
a scatter in the inferred stellar mass
and by extension star formation efficiencies.
We quantify the scatter by computing
$\sigma$ in the ratio between the stellar mass
reported by surveys of young stellar objects (YSOs)
and the stellar mass we compute from the free-free
fluxes for known clusters (see Figure \ref{fig2}).
We find $\sigma=0.22$ dex.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{sfe_v_Mtot}
\caption{Left: the star formation efficiency $\epsilon_{\rm br}$
(the subscript br stands for Bremsstrahlung)
of 191 SFC-GMC complexes plotted as a function of
total (gas plus stellar) mass.
The correlation between $\epsilon$ and $M_{\rm tot}$
is statistically significant;
we draw with the red dashed line the least-square fit
correlation.
Right: histograms of $\epsilon$ calculated with
free-free flux. The middle dashed line
represents the median $\epsilon$ while
the dashed lines to the left and right illustrate $\sigma$.
The mean $\epsilon\sim 0.03$ is in rough agreement with that
of nearby clouds \citep[e.g.,][]{evans09,gutermuth11}.
The full width of the distribution spans $\sim$3 orders of magnitude.
Since the p-values of Kendall's $\tau$ test is smaller than 0.05,
we remove the annotated least-square fit between $\epsilon$ and $M_{\rm tot}$
before estimating the $\sigma$.}
\label{fig5}
\end{figure*}
\section{Star Formation Efficiency}
\label{sec:sfe}
We use the measured scatter $\sigma_{\log\epsilon}$ as one metric to
test various models of star formation rate. Using the efficiency of
star formation appears to be particularly advantageous since the
measured value of $\epsilon$ is independent of distance, so that
errors in the distance determination to either GMCs or star clusters
do not introduce any scatter in the distribution of
$\epsilon$. However, the comparison of $\sigma_{\log\epsilon}$ to the
prediction of the constant star formation rate per free-fall time
theory do require the use of a distance estimate, somewhat reducing
the attractiveness of this particular metric for our purposes.
Figure \ref{fig5} shows the star formation efficiency $\epsilon$ as a function
of the mass of the host GMCs (left panels) and the histogram of $\epsilon$ (right panels).
We recover the result of \citet{mooney88} that the ratio of FIR to CO flux has
a very broad distribution. Specifically, we find that $\epsilon$
ranges over three orders of magnitude.
After correcting for the correlation between $\epsilon$ and the total mass,
we calculate the scatter $\sigma_{\log\epsilon}$ as defined in Section \ref{sec:flux}:
about 68.3\% of the data fall within 0.79 dex.
There are a number of sources of scatter in $\epsilon$ that stem from
measurement uncertainties.
Ionizing photons can travel surprising long distance so if the aperture chosen to
measure the free-free flux is too small, $\epsilon_{\rm br}$ will be underestimated.
If there are nearby ($\sim 100\pc$) sources of ionizing radiation and the aperture is
overly large, $\epsilon_{\rm br}$ will be overestimated.
Miss-identifications---regions which are not physically associated are nevertheless
cross-correlated---will also introduce artificial scatter into the value of $\epsilon$.
The aperture for free-free measurements are limited by the
low resolution of the {\it WMAP} free-free map. We have chosen to assign
free-free fluxes of SFCs
according to their relative 8$\mu$m fluxes. Varying the aperture size of SFCs,
\citet{lee12} quote an 88\% measurement error in 8$\mu$m, which we adopt for free-free fluxes
of all GMCs. Propagating these errors in flux measurement, each cloud has an error
of $\delta \log\epsilon = 0.38$ dex.
For the entire sample of 191 star-forming clouds, measurement errors
contribute $\Delta\sigma_{\log\epsilon_{\delta \rm flux}}=0.03$ dex.
The conversion factor between free-free luminosities and stellar masses
depends on the IMF. In Section \ref{sec:flux}, we estimated the
errors introduced by the poor sampling of the IMF and the physical
variations in the IMF as 0.22 dex.
Adding the two sources of error in quadrature, we find
\be \label{eq: match err br}
\Delta \sigma_{\log\epsilon}=0.22\,{\rm dex}.
\ee
Combining all our error estimates, we arrive at our final error estimate for the
width of the distribution of star formation efficiency:
\be \label{eq: match br}
\sigma_{\log\epsilon_{\rm br}}=0.79\pm 0.22\,{\rm dex}
\ee
This is the first of two main observational results in this paper: the distribution
of star formation efficiencies $\log \epsilon$ is very broad.
\section{STAR FORMATION RATE PER FREE FALL TIME $\epsilon_{\rm ff}$ }
\label{sec:sfr_ff}
The quantity $\epsilon_{\rm ff}$ is an estimate of the star
formation rate, normalized to the free-fall time of the
star-forming region (GMCs in our case, see equation \ref{eqn:sfe_ff}).
We will measure the scatter in the distribution of $\epsilon_{\rm ff}$
and compare it against what is expected from various models of star formation
(the main test case being turbulence-regulated star formation proposed by
\citealt[][KM05]{krumholz05} and \citealt[][HC11]{hennebelle11}).
Using the definition of $\epsilon_{\rm br}$
from equations \ref{eqn:epsilon br},
equation \ref{eqn:sfe_ff} can be rewritten as
\begin{equation}
\epsilon_{\rm ff}^{\rm br} = \epsilon_{\rm br} \frac{\tau_{\rm ff}}{\left<t_{\rm ms, q}\right>},
\label{eq:eff_br}
\end{equation}
where $\left<t_{\rm ms,q}\right>\approx 3.9\Myrs$ is the
Q-weighted main-sequence lifetime of O stars.
We calculate the free-fall time $\tau_{\rm ff}$ using the ellipsoidal volume found
by assuming that the length along the projected radial direction to be equal to the
shorter of the two axes in the plane of the sky.
Unlike the measurement of $\epsilon$, the measurement of
$\epsilon_{\rm ff}$ depends on the distance to the source, via the
free-fall time $\tau_{\rm ff}$ factor in equation (\ref{eq:eff_br}).
Thus we expect that the scatter
in $\epsilon$ is smaller than that of $\epsilon_{\rm ff}$, and this is what we find.
The scatter in the GMC free-fall time is
$\sigma_{\log\tau_{\rm ff}}\approx0.27$ dex. This is much smaller than
the scatter in $\epsilon_{\rm br}$, $\sigma_{\log\epsilon_{\rm br}}=0.79$ dex.
Part of the scatter in $\tau_{\rm ff}$ is due to uncertainties in our
distance measurements, but most of it is intrinsic.
More precisely, $\tau_{\rm ff} \propto f_{\rm CO}^{-1/2} D^{1/2}$.
The errors in the measurement of distances are typically $\sim$35\% \citep{lee12}.
Propagating errors in both the flux and distance measurements,
we find errors of $\delta\log\tau_{\rm ff}=0.22$ dex for each cloud.
Combining this with the flux measurement errors in $\epsilon$,
$\delta\log\epsilon_{\rm ff}=0.44$ dex.
For the entire sample of 191 SFC-GMCs, the measurement errors are
$\Delta\sigma_{\log\epsilon_{\rm ff}}=0.03$ dex.
Combining the measurement error with the error introduced by
the poor sampling and the variations in the IMF $\sigma=$0.22 dex,
\be
\sigma_{\log\epsilon_{\rm ff}^{\rm br}}= 0.91\pm 0.22\,{\rm dex}
\ee
This is our second major observational result: the distribution of
star formation rate per free fall time $\epsilon_{\rm ff}$ is very
broad.
We can compare this directly to the prediction of KM05 and HC11, using the
properties of the GMCs we used to measure $\epsilon_{\rm ff}$
(see Section \ref{sec:cause}). We find
$\sigma_{\log\epsilon_{\rm, ff, KM05}}= 0.24$ dex
and $\sigma_{\log\epsilon_{\rm, ff, HC11}}=0.13$ or 0.12 dex,
depending on the choice of parameters (to be elaborated in Section \ref{sec:cause}).
The model-inferred scatter in $\epsilon_{\rm ff}$ are
3, 3.5, and 3.6 standard deviations from the observed value of $0.91\pm 0.22$ dex
(see Figure \ref{fig4hc}).
The fact that the measured scatter in $\epsilon_{\rm ff}$ is
significantly larger than the prediction of either KM05 or HC11 model, which
account for variations in the Mach number and virial parameter of the
host GMC (but ignores any
{\it explicit}
time dependence in the rate of star
formation) is a strong evidence that the model is incomplete.
We discuss in more detail in the next section.
We summarize the result of our calculations of $\epsilon$ and $\epsilon_{\rm ff}$
in Table \ref{table3}.
\input{sfc-gmc-analy}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{sfrff_v_Mtot_HC11}
\caption{Top left: the star formation rate per free-fall
time $\epsilon_{\rm ff}$ of 191 SFC-GMC complexes
as a function of the total (gas + star) mass.
We see a correlation between $\epsilon_{\rm ff}$ and
the total mass that is marginally statistically significant;
the least-square fit correlation is drawn with the red dashed line.
Middle left: $\epsilon_{\rm ff}$ predicted
by equation 30 of \citet{krumholz05} (equation \ref{eq:krumholz05}
in this paper). We see a statistically significant
correlation between $\epsilon_{\rm ff, KM05}$ and the total mass
(also drawn with the red dashed line).
Bottom left: $\epsilon_{\rm ff}$ predicted by the
multi-freefall theory of star formation
from \citet[][labeled as HC11]{hennebelle11},
their equation 8 (our equation \ref{eq:sfrffHC11_4};
see also \citealt{federrath12}).
Following HC11, we compute
the model-inferred $\epsilon_{\rm ff,{\rm HC11}}$ using
two different criteria on the critical gas density over which
star formation occurs: the local Jeans length is equal
to the thickness of a shocked layer
for a given Mach number \citep[][PN]{padoan11};
and the local Jeans length is equal to
some prescribed fraction $y_{\rm cut}$ of the cloud size
\citep[][$y_{\rm cut}$]{hennebelle11};
we adopt their recommended $y_{\rm cut}=0.1$.
We see a strong positive correlation between
$\epsilon_{\rm ff,{\rm HC11}}$ and $M_{\rm tot}$
in contrast to marginal negative correlation
observed in the data (upper left panel).
The least-square fit correlations are drawn
in blue for PN and in magenta for $y_{\rm cut}$.
We normalized $\epsilon_{\rm ff, HC11}$ such that
its median matches that of the observed.
We plot the median $\epsilon_{\rm ff}$ in the
upper right panel with the errorbar indicating
$\sigma$. The observed scatter in
$\epsilon_{\rm ff} = 0.91$ dex is significantly larger
than what any static model predicts:
0.24 dex (KM05), 0.13 dex (HC11, PN), and
0.12 dex (HC11, $y_{\rm cut}$).
All $\sigma$'s are measured after correcting
$\epsilon_{\rm ff}$ for its correlation
with the total mass (annotated in the left panels).
Bottom right panel shows that no $y_{\rm cut}$
within a reasonable range can reconcile
HC11 model with the enormous scatter in $\epsilon_{\rm ff}$
observed in the data (shown as the dashed line).
}
\label{fig4hc}
\end{figure*}
\section{IMPLICATIONS OF THE LARGE SCATTER IN THE EFFICIENCY AND THE RATE OF STAR FORMATION}
\label{sec:cause}
We discuss a number of ideas regarding the physical processes that
regulate the rate of star formation in light of our observational
results.
\subsection{Turbulence regulated star formation}
\label{ssec:turbsf}
KM05 present a semi-analytic model of $\epsilon_{\rm ff}$
based on the idea of turbulence-regulated star formation.
They posit that within a turbulent cloud whose density distribution
is well-characterized by a log-normal distribution---with its width governed
by the Mach number---only the regions of densities larger than some critical value
will collapse to form stars. This critical value depends on both
the virial parameter and the Mach number. Clouds
with larger $\alpha_{\rm vir}$ and ${\cal M}$ are harder to collapse so
the critical $\rho$ for such clouds will be larger, leading to a
smaller star formation rate per free-fall time $\epsilon_{\rm ff}$.
Assuming that the log-normal density
distribution established by turbulence is maintained over the cloud lifetime
(but see e.g., \citealt{vazquez-semadeni08}, \citealt{cho11}, \citealt{kritsuk11},
\citealt{collins12}, \citealt{federrath13}
for the evidence of dynamic density distributions in star-forming regions),
KM05 find
\be \label{eq:krumholz05}
\epsilon_{{\rm ff},0}=0.014
\left({\alpha_{\rm vir}\over 1.3 }\right)^{-0.68}
\left({{\cal M} \over 100}\right)^{-0.32},
\ee
(their equation 30). We append a subscript 0 to $\epsilon_{\rm ff}$
to emphasize that their model {\em assumes} that the star formation rate
is time-independent.
We see in Figure \ref{fig4hc} that
the KM05 model expects a scatter in
$\epsilon_{\rm ff}$, $\sigma=0.24$ dex,
significantly smaller than the observed
$\sigma=0.91$ dex.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{kdm12_sfrff_surfdens}
\caption{Star formation rate per unit area vs.~the cloud gas surface
density divided by the free-fall time, analogous to Figure 3 of
\citet{kdm12}. The red dashed line delineates the volumetric star formation law
proposed by \citet{kdm12} assuming a constant $\epsilon_{\rm ff}=0.01$.
We draw two different constant $\epsilon_{\rm ff}$ (0.1 and 0.001)
in red dot-dashed lines for reference.
We find a significant correlation between
$\Sigma_\star$ and $\Sigma/\tau_{\rm ff}$. However, the scatter about
the volumetric star formation law
$\dot\Sigma_\star \propto \Sigma/\tau_{\rm ff}$ is large:
0.86 dex.}
\label{fig8}
\end{figure}
The KM05 model has been criticized for
characterizing an entire cloud
and its collapsing substructure by
a single free-fall timescale.
Star-forming substructures have
varying densities and therefore
different free-fall times
(see e.g., \citealt{hennebelle11},
\citealt{federrath12}, and references therein).
\citet{hennebelle11} present an analytic model of $\epsilon_{\rm ff}$,
which takes into account not only the different free-fall time
for each collapsing structure but also the
recycling of turbulent flow over a dynamical time
\citep[see also][]{federrath12,hennebelle13}:
\begin{equation}
\epsilon_{\rm ff} \propto e^{3\sigma_{\rho}^2/8}\left[1+{\rm erf}\left(\frac{\sigma_{\rho}^2-\ln x_{\rm crit}}{2^{1/2}\sigma_{\rho}}\right)\right]
\label{eq:sfrffHC11_1}
\end{equation}
where $\sigma_{\rho}$ is the spread in the density distribution, set by
the cloud Mach number:
\begin{equation}
\sigma_{\rho}^2 = \ln(1 + b^2{\cal M}^2)
\label{eq:sfrffHC11_2}
\end{equation}
($b = 0.25$ for purely solenoidal driving and $b=1.0$
for purely compressive driving; we take $b=0.5$)
and $x_{\rm crit} \equiv \rho_{\rm crit}/\rho_0$ is
the critical density over which substructure begins to collapse
normalized by the average bulk density of the cloud $\rho_0$.
We experiment with two criteria for collapsing:
if the local Jeans length becomes comparable
to the local shock width \citep[][PN criterion]{padoan11}:
\begin{equation}
x_{\rm crit} \simeq 0.067 \theta^{-2} \alpha_{\rm vir} {\cal M}^2
\label{eq:sfrffHC11_3}
\end{equation}
with $\theta = 0.35$ and
if the local Jeans length becomes comparable
to some prescribed fraction $y_{\rm cut}$ of the cloud size
\citep[][$y_{\rm cut}$ criterion]{hennebelle11}:
\begin{equation}
x_{\rm crit} = \frac{1}{5 y_{\rm cut}} \frac{\alpha_{\rm vir}}{{\cal M}^2}(1 + {\cal M}y_{\rm cut}^p).
\label{eq:sfrffHC11_4}
\end{equation}
Here, $p$ is the power-law scale from
the size-linewidth relationship $\sigma \propto R^p$; we
adopt p = 0.5.
The normalization of equation \ref{eq:sfrffHC11_1} represents
the efficiency at which a clump gas converts to a star, and
we adjust it so that the median model-inferred
$\epsilon_{\rm ff}$ matches that of the observed.
The match warrants an unusually small
normalization: 0.009 for the PN criterion and
0.002 for the $y_{\rm cut}$ criterion,
compared to the usual 0.02--0.05.
These multi-freefall models of star formation show stronger dependence
of $\epsilon_{\rm ff}$ on the Mach number compared
to the KM05 model. It has been argued
that the observed scatter in $\epsilon_{\rm ff}$ of Milky Way clouds
can be explained if the clouds' Mach number or the scale of
turbulence driving vary by $\gtrsim2$ orders of magnitude
\citep[e.g.,][]{federrath13b,chabrier14}.
Figure \ref{fig4hc} demonstrates that the velocity dispersion
or the size of the clouds simply do not vary enough to reconcile
even the multi-freefall model
with the observed scatter in $\epsilon_{\rm ff}$
of the MML16 clouds.
No reasonable choice
of $y_{\rm cut}$ can
reproduce the large $\sigma=0.91$ dex.
Furthermore, HC11 models predict
a strong positive correlation between $\epsilon_{\rm ff}$
and the total mass in contrast to a weak negative correlation
seen in the data.
Three facts emerge from our analysis that
hints at the controlling parameters
of star formation other than $\alpha_{\rm vir}$, ${\cal M}$,
and the scale of turbulence driving.
First, the measured scatter is significantly larger
than any theoretically expected scatter.
Second, there is little systematic offset between the observed
and the KM05 distribution of $\epsilon_{\rm ff}$.
Third, an unusually small gas-to-core efficiency is required
to match the median $\epsilon_{\rm ff}$ expected from the multi-freefall
models to that observed.
All three facts suggest
the distribution of Milky Way cloud ${\cal M}$ and size
is far too narrow to explain the observed scatter in $\epsilon_{\rm ff}$
with what the multi-freefall models predict.
Both the larger scatter and the lack of systematic offset between the observed
and theoretical distribution of $\epsilon_{\rm ff}$
is also evident in Figure \ref{fig8}.
The figure shows the star formation rate per unit area $\dot\Sigma_\star$
plotted against the ratio between the gas surface density
and the free-fall time, $\Sigma/\tau_{\rm ff}$, a comparison advocated
by \citet{kdm12}.
Like \citet{kdm12}, we find a statistically significant correlation
between $\dot\Sigma_\star$ and $\Sigma/\tau_{\rm ff}$. A least-square
fit produces $\dot\Sigma_\star \propto (\Sigma/\tau_{\rm
ff})^{0.3}$, compared to the linear relationship (the volumetric star
formation law) reported by \citet{kdm12}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{tan00_sfrff_surfdens}
\caption{Star formation rate per unit area $\dot\Sigma_\star$
vs.~the cloud surface density divided by the collision time
$\sim \tau_{\rm orb}Q$.
The orbital time $\tau_{\rm orb} = 2\pi R_{\rm gal} / 220\kms$
where $R_{\rm gal}$ is the galactocentric radius and
we assume a flat rotation curve $v(r) = 220\kms$.
The red dashed line illustrates equation \ref{eq:sf_col}
with $\epsilon_{\rm col} = 0.1$,
assuming a flat rotation curve ($\beta = 0$).
We see a strong correlation between
$\dot\Sigma_\star$ and $Q^{-1}\Sigma/\tau_{\rm orb}$
but the scatter about the model is still large: 0.92 dex.
}
\label{fig9}
\end{figure}
We interpret the difference in the power-law index as the result of
the large scatter in the observed value of $\epsilon_{\rm ff}$.
As Figure \ref{fig8} attests,
the data points span more than an order of magnitude both
above and below $\epsilon_{\rm ff}=0.01$, well beyond the expected
uncertainty $\sim$3 predicted by \citep{kdm12}.
We find a dispersion
about the volumetric law of 0.86 dex.
Similarly large scatters in star
formation rates are observed in local molecular gas from Gould's belt
\citep{evans14} and in ATLASGAL clumps \citep{heyer16}.
The sample of clouds assembled by \citet{evans14}
falls squarely within the range $\epsilon_{\rm ff}=0.1$ and
$\epsilon_{\rm ff}=0.001$. A few ATLASGAL clumps analyzed by \citet{heyer16}
fall below $\epsilon_{\rm ff}=0.001$ but no
clumps lie beyond $\epsilon_{\rm ff}=0.1$.
Our GMC sample is
more complete than the clump samples in either the Gould's belt or
ATLASGAL clumps, which may explain part of the difference in the
dispersions found in the different surveys.
\subsection{Collision-induced Star Formation}
\label{ssec:collision}
An alternate form of star formation law envisions
cloud-cloud collision to regulate the star formation rate
on approximately orbital timescales
\citep[see e.g.,][and references therein]{tan00, suwannajak14}:
\begin{equation}
\dot\Sigma_\star \simeq \epsilon_{\rm col} Q^{-1}(1-0.7\beta)\Sigma / \tau_{\rm orb}
\label{eq:sf_col}
\end{equation}
where $\epsilon_{\rm orb}=0.1$ is the rate at which gas is converted to
stars for each collision, $Q \sim \sqrt{\alpha_{\rm vir}}$ the Toomre parameter,
$\tau_{\rm orb}$ the orbital time, and
$\beta$ is the logarithmic derivative of the velocity profile.
Assuming a flat rotation curve ($\beta=0$), we
show in Figure \ref{fig9}
that our star-forming clouds scatter about the
collision-induced star formation
law with a dispersion of $\sigma=$0.94 dex.
As in Figure \ref{fig8}, the full min-mid or mid-max scatter
is more than an order of magnitude; it is unlikely that
the shear-velocity dependent term $(1-0.7\beta)$
vary by more than a factor of order unity.
As with the volumetric star formation law, the large scatter about the
collision-induced star formation law suggests there are extra
parameters that control the rate at which gas is converted into stars.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{sfrff_tratio}
\caption{Star formation rate per free-fall time $\epsilon_{\rm ff}$
plotted against the ratio between the free-fall and the dynamical time,
$\tau_{\rm ff}/\tau_{\rm dyn}$. The star formation law in
equation \ref{eq:eff_padoan} is drawn as red dashed lines.
The time ratio (equivalently, the square-root of the virial parameter)
may be an important parameter related to the
maximum possible $\epsilon_{\rm ff}$, but by itself
it does not explain the large scatter in $\epsilon_{\rm ff}$.}
\label{fig10}
\end{figure}
\subsection{SFR Parametrized by Virial Parameter}
\label{ssec:sfr-vir}
\citet{padoan12} studied magneto-hydrodynamic simulations of
turbulent, star-forming gas, finding that $\epsilon_{\rm ff}$ depends
exponentially on the ratio between the free-fall and the dynamical
time (which is equivalent to the square-root of the virial parameter):
\begin{equation}
\epsilon_{\rm ff} \simeq \epsilon_{w}e^{\left(-1.6~\tau_{\rm ff}/\tau_{\rm dyn}\right)},
\label{eq:eff_padoan}
\end{equation}
where $\epsilon_{w}$ accounts for the mass loss due to
proto-stellar winds and outflows,
and $\tau_{\rm dyn}$ is the dynamical time in their simulations.
Figure \ref{fig10} shows the star formation rate per free-fall
time for star-forming clouds, plotted
as a function of $\tau_{\rm ff}/\tau_{\rm dyn}$
where $\tau_{\rm dyn} = R_g / \sigma_{\rm GMC}$ is
the dynamical time of the host cloud.
We do see a hint of a decrease in the upper envelope of the distribution of
$\epsilon_{\rm ff}$ with increasing $\tau_{\rm ff}/\tau_{\rm dyn}$.
However, many GMCs are more efficient at producing stars than what
equation \ref{eq:eff_padoan} would suggest. More strikingly,
the $\epsilon_{\rm ff}$ values
of GMCs span $\sim$3 orders of magnitude {\em below} the
prediction by \citet{padoan12}.
\citet{padoan12} noted that $\epsilon_{\rm ff}$
depended on the Alfv\'enic Mach number ${\cal M}_a$, with lower Alfv\'enic Mach number
(stronger magnetic field for a given strength of turbulence)
yielding smaller $\epsilon_{\rm ff}$, but only for ${\cal M}_a\gtrsim 5$;
below this value, $\epsilon_{\rm ff}$ increases again (see their Figure 3).
Their simple fitting formula (our equation \ref{eq:eff_padoan})
corresponds to the minimum star formation rates in their models
with ${\cal M}_a\approx 5$. Can the variation in the magnetic field strength
explain the large scatter in $\epsilon_{\rm ff}$?
Variations in ${\cal M}_a$ may explain those points that lie above
the red dashed line (corresponding to equation \ref{eq:eff_padoan}) in
Figure \ref{fig10}, but the extreme low values seen
below the red dashed line cannot be explained by variations
in the magnetic field strength.
The Alfv\'enic Mach number can be estimated
as ${\cal M}_a \sim \alpha_{\rm vir} (M_g/M_\Phi)$
where $\alpha_{\rm vir}$ is the virial parameter and
$M_\Phi = 0.12\pi B (R_g/2)^2 / \sqrt{G}$ is the magnetic critical mass
where $B$ is the magnetic field strength. SFC-GMC clouds typically have
volumetric number densities of $n_{\rm H} \sim\,$10--200$\,\,{\rm g\,\,cm^{-3}}$
so their $B\sim$1--10$\mu$G \citep{crutcher12}; we find that these clouds
have $M_g/M_\phi \sim\,$10--100.
The median $\alpha_{\rm vir}\sim$0.76 with
a scatter of 0.32 dex. We estimate ${\cal M}_a\sim\,$4--200.
The simulations of \citet{padoan12} show that an order of magnitude change in
${\cal M}_a$ results in a factor of $\sim$3 change in $\epsilon_{\rm ff}$.
Taking the variations in ${\cal M}_a$ of the star-forming clouds into account,
we expect an upward scatter from the reference value
given by equation \ref{eq:eff_padoan} by factors of $\sim$10.
\subsection{Constant vs.~Time-dependent star formation rate per free fall time}
\label{ssec:time_dependent}
All four models described in
Section \ref{sec:cause}---single-freefall turbulence-regulated,
multi-freefall turbulence-regulated,
collision-induced,
and parametrization by virial parameter---assume a
star formation rate that has no explicit dependence on time.
These models predict that any observed
scatter in $\epsilon_{\rm ff}$ should arise from variations
in the internal properties of clouds (e.g., virial parameter, Mach
number, Alfv\'enic Mach number, or
free-fall time) or large scale motions.
We have demonstrated above that the distribution in the
observed $\epsilon_{\rm ff}$ of Milky Way GMCs
is far too broad to be explained by any of these models
(equations \ref{eq:krumholz05}, \ref{eq:sfrffHC11_4}, \ref{eq:sf_col}, and \ref{eq:eff_padoan}).
A number of authors have suggested that the star formation rate on
scales of GMCs and smaller increases systematically with time
(see e.g., \citealt{palla99}, \citealt{gutermuth11}
and \citealt{murray11} for observational evidence
or see e.g., \citealt{lee15}, \citealt{murray15},
and \citealt{dmurray15} for theoretical studies).
Sampling clouds at different evolutionary stages
with time-varying star formation rate
may give rise to the broad distributions
in $\epsilon$ and $\epsilon_{\rm ff}$.
In this section, we allow $\epsilon_{\rm ff}$ to be
time-dependent and compare the expected width in the distribution of
$\epsilon_{\rm ff}$ to that observed.
We note that the explicit time-dependence of $\epsilon_{\rm ff}$
arises from the time-varying turbulent
structure (i.e., the size-linewidth relation itself
changes with time due to the interplay between
turbulence and gravity;
see \citealt{goldbaum11} and \citealt{murray15}).
This explicit time-dependence should not be confused
with the implicit time-dependence portrayed by \citet{hennebelle11}
whose model accounts for the re-assembly of turbulent structure
that is {\it static} (i.e., the turbulent flows follow the same
size-linewidth relation each time they are recycled).
We proceed to test the effect of explicit
time evolution of $\epsilon_{\rm ff}$ on the distribution of $\epsilon_{\rm ff}$.
First, we generalize the \citet{krumholz05} model to allow
for time-variable $\epsilon_{\rm ff}$,
\be
{dM_*\over dt}=\epsilon_{{\rm ff},0}\left({t\over\tau_{\rm ff}}\right)^\delta {M_g\over \tau_{\rm ff}},
\ee
where $t=0$ corresponds to the time at which the first star forms, and
$\delta=0$ corresponds to the constant $\epsilon_{\rm ff}$ model.
Next, we use this prescription for $\epsilon_{\rm ff}$
in the model of \citet{feldmann11}
which describes the concomitant evolution of the
stellar mass and the gas mass of a star-forming cloud:
\begin{eqnarray}\label{eq:feldmann}
{dM_g\over dt}&=&-\epsilon_{{\rm ff},0}\left({t\over \tau_{\rm ff}}\right)^\delta
{M_g\over\tau_{\rm ff}}-\alpha M_*+\gamma,
\\
{dM_*\over dt}&=& \epsilon_{{\rm ff},0}\label{eq:feldmann2}
\left({t\over \tau_{\rm ff}}\right)^\delta
{M_g\over\tau_{\rm ff}}.
\end{eqnarray}
The quantity $\alpha$ parametrizes the rate at which stellar
feedback disrupts the host GMC,
while $\gamma$ is the (possibly time dependent) rate of
gas accretion onto the GMC.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{sfrff_FG_br}
\caption{Left: the evolution of the GMC gas mass
$M_g$ (black line), stellar mass $M_\star$ (cyan line),
and live stellar mass $M_{\star,\rm live}$ (red line; those with effective
lifetime of $\left<t_{\rm ms,q}\right>\sim 4\Myrs$),
together with the star formation rate per free-fall time
$\eta_{\rm ff} \equiv [M_{\star,\rm live} / (M_g + M_{\star,\rm live})] (\tau_{\rm ff} / \left<t_{\rm ms,q}\right>)$
(magenta dashed line) for three different
values of $\delta =$ 0, 1, 2 in the model described by equations
\ref{eq:feldmann} and \ref{eq:feldmann2}.
We use the observed median free-fall time $\tau_{\rm ff} \simeq 6.7\Myrs$.
Right: the predicted distributions of
$\eta_{\rm ff}$ (magenta histograms) together with
the observed distributions (gray histograms).
We assume no gas accretion (results for constant
accretion models are similar).
A quadratic growth $\epsilon_{\rm ff} \propto t^2$ (bottom panels) reproduces
the observed scatter in $\epsilon_{\rm ff}$ the best.}
\label{fig11}
\end{figure*}
Equations \ref{eq:feldmann} and \ref{eq:feldmann2}
are coupled ordinary differential equations (ODEs)
whose solutions are $M_g(t)$ and $M_\star(t)$.
Massive stars that contribute most to the ionizing
radiation typically live for $\left<t_{\rm ms,q}\right>\simeq 3.9\Myrs$
so we define the mass in these ``live'' stars as
$M_{\rm \star, live} (t) = M_\star(t) - M_\star(t - \left<t_{\rm ms,q}\right>)(t)$.
We can then write the star formation efficiency as
$\epsilon(t) \equiv M_{\rm \star, live}(t) / (M_g(t) + M_{\rm \star, live}(t))$
and the star formation rate per free fall time as
$\eta_{\rm ff}(t) \equiv \epsilon(t) \tau_{\rm ff} / \left<t_{\rm ms,q}\right>$.
Note that $\eta_{\rm ff}$ is analogous to the
observed $\epsilon_{\rm ff}$, not to be confused with
$\epsilon_{\rm ff,0} (t / \tau_{\rm ff})^\delta$.
We illustrate the predicted evolution and the distribution of
$\eta_{\rm ff}$ of this model in Figure \ref{fig11} for three different
$\delta=0,1,2$ assuming no accretion ($\gamma=0$).
We set $\tau_{\rm ff}$ to the median GMC free-fall
time $\tau_{\rm ff} \sim 6.7\Myrs$
and set $\alpha=3.5$ so that the cloud disperses in $\sim 20\Myrs$
e.g., \citealt{williams97} and \citealt{murray11}.
In Section \ref{ssec:lifetime} below we obtain a mean
cloud lifetime using MML16 clouds of 21--24$\Myrs$.
To isolate the effect of $\delta$, we assume all
the host GMCs have the same $\alpha_{\rm vir}$ and ${\cal M}$, or more
precisely, the same intrinsic value of $\epsilon_{\rm ff,0}$. We
account for variations in these parameters after integrating the ODEs
by adding the variations in quadrature.
The distribution of $\eta_{\rm ff}$ that we calculate from the model with
$\delta=0$ (the constant $\epsilon_{\rm ff}$ model) features a very
sharp peak. Because the lifetime of GMCs ($\sim20\Myrs$) is
substantially longer than the lifetime of live stars ($\sim4\Myrs$),
the distribution in $\eta_{\rm ff}$ is dominated by the period in which
both the live stellar mass and the host GMC mass are roughly constant
with $\epsilon \sim \epsilon_{\rm ff, 0} (4 \Myrs / \tau_{\rm ff})$
and therefore a constant $\eta_{\rm ff} \sim \epsilon_{\rm ff, 0}$.
The small scatter (0.16 dex) arises partly from the initial rise in
the live stellar mass, and partly from the final rapid decline in GMC
mass around $\sim20\Myrs$, when the stellar feedback disrupts the host
GMC.
To make a fair comparison to the observed scatter in
$\epsilon_{\rm ff}$, we add in quadrature the scatter
from the initial and final transients, 0.16 dex, and the dispersion in
$\epsilon_{\rm ff,0}=0.24$ dex, to find a total predicted dispersion
of $\sigma_{\log\eta_{\rm ff}}=0.29$.
Compared to the observed $\sigma_{\rm br} = 0.91 \pm 0.22$ dex,
the scatter predicted by the models of
time-independent $\epsilon_{\rm ff}$
is too small by 2.8 standard deviations.
Allowing for a monotonic increase in $\epsilon_{\rm ff}$ with time significantly
broadens the distribution in $\epsilon$ and $\epsilon_{\rm ff}$
(see second and third rows in Figure \ref{fig11}).
Not only does the mass in live stars monotonically increase over most of the cloud
lifetime, it also increases more gradually from time zero.
When $\delta=1$ (as predicted by \citealt{lee15} and \citealt{murray15}),
we find $\sigma=0.54$. Adding this in quadrature to the scatter in
GMC properties, we find a total dispersion of
$\sigma_{\log\eta_{\rm ff}}=0.59$
(1.4-sigma away from $\sigma_{\rm br}$).
Similar calculations show a total dispersion of
$\sigma_{\log\eta_{\rm ff}}=0.93$ for $\delta=2.0$
(within one standard deviation from $\sigma_{\rm br}$).
Models assuming constant star formation rate per free-fall time
predict too narrow a distribution in $\epsilon_{\rm ff}$.
Using the simple model of \citet{feldmann11},
$\epsilon_{\rm ff} \propto t^2$---equivalently, the stellar mass
increases as $M_*(t)\sim t^3$---fits the data better.
Our result demonstrates that a time-varying
star formation rate is in better agreement
with the data compared to time-independent models.
Determining the exact form
of $\epsilon_{\rm ff}$ that best fits the observation
will require a careful parameter study which
is beyond the scope of this paper.
We note that all predicted distributions of
$\eta_{\rm ff}$
(whether the star formation rate is assumed to be time-dependent or constant)
feature a deficit towards the high end and an excess towards the low end
compared to the observations.
The simple model of
\citet{feldmann11} assumes that feedback from stars
begins to destroy the host GMC as soon as the stars form. If the feedback
takes the form of gas pressure in HII regions, or of radiation pressure, the host GMC
will not be significantly affected until either pressure (or their sum) overcomes the
pressure associated with the self-gravity of the GMC. Altering the model to account for
this threshold effect will extend the distribution of $\epsilon$ to lower values. We leave
this and similar model building efforts to future work.
\subsection{Cloud Lifetimes}
\label{ssec:lifetime}
We estimate the average cloud lifetime using the star-forming clouds
(191 SFC-GMC complexes) and non-star-forming clouds that are
massive enough---and gravitationally bound---to birth stars.
We show that clouds indeed live substantially longer than $\sim4\Myrs$. This
separation of time scales is crucial if we are to use the width of the
distribution of $\epsilon$ (or $\epsilon_{\rm ff}$) to distinguish between constant
star formation rate models and models which allow for variations of
the star formation rate with time.
Clouds that harbor at least one SFC are likely near the end of their lives:
their SFCs have already carved out bubbles of size 10--100 pc.
The lifetime of clouds can then be estimated by multiplying
the effective lifetime of live stars by the ratio of the
total number of potentially star-forming clouds to the
number of clouds that have SFCs.
We define potentially star-forming clouds as those
that are both massive enough to birth
Orion Nebula Cluster (i.e., $M_{\rm gas} \geq M_{*} \sim 1000 M_\odot$)
and gravitationally bound ($\alpha_{\rm vir} < 3.3$; see Appendix of MML16);
there are 1014 such clouds.
The average cloud lifetime is then $((1014+191)/191) \times 4 \Myrs\sim 24\Myrs$.
The lifetime reduces to 21$\Myrs$ if we place the cloud lower mass limit
at $10^4 M_\odot$ instead.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{sigSFR_gas_prof}
\caption{Top: star formation rate per unit
area $\dot\Sigma_\star$ as a function of galactocentric radius.
Middle: gas surface density profile $\Sigma_{\rm gas}$.
Bottom: the ratio between the two profiles $\dot\Sigma_\star/\Sigma_{\rm gas}$.
Star formation is observed to follow the gas: they all show similar scale radius.
Both profiles show a break at $\sim4\kpc$, consistent
with the presence of a molecular ring.}
\label{fig12}
\end{figure}
\section{THE MILKY WAY STAR FORMATION RATE SURFACE DENSITY PROFILE}
\label{sec:sigma}
\citet{kennicutt12}, in their Figure 7, show that the star formation rate per unit area
$\dot\Sigma_\star$ falls off more slowly than the surface density of molecular gas
$\Sigma_{\rm gas}$. By contrast, in the nearby galaxy NGC 6946,
$\dot\Sigma_\star$ is observed to follow $\Sigma_{\rm gas}$ closely.
The information on star formation rate used by \citet{kennicutt12}
date back to \citet{guesten82}.
The catalog of SFCs and MML16 clouds used in this paper should
provide the most up-to-date and the most
complete estimate of star formation rate and the molecular gas mass.
Using the sample of star-forming clouds, we illustrate in
Figure \ref{fig12} how even in the
Milky Way, $\dot\Sigma_\star$ tracks well $\Sigma_{\rm gas}$
as a function of galactocentric radius.
In particular, both the star formation rate and the gas density profiles
are well-fitted with an exponential profile of scale length
$\sim$2.0. The trend observed in $\dot\Sigma_\star$ is anti-correlated
with the profile of $\alpha_{\rm vir}$ shown
in Figure 14 of MML16: where $\alpha_{\rm vir}$ rises, $\dot\Sigma_\star$
dips (inside 4 $\kpc$ and outside 8 $\kpc$) and where $\alpha_{\rm vir}$
dips, $\dot\Sigma_\star$ rises (between 4 and 8 $\kpc$). This anti-correlation
is expected since stars should form in gravitationally-bound regions.
\section{DISCUSSION AND SUMMARY}
\label{sec:disc}
We have demonstrated in this paper the importance of the scatter
in the distribution of $\epsilon_{\rm ff}$ and
$\epsilon$. Any successful model of cloud-scale star formation should
reproduce not only the mean star formation rate but also
the large scatter about the median.
The spread in $\epsilon$ has long been known to be enormous:
\citet{mooney88}, \citet{scoville89}, and \citet{mead90} find a
minimum-maximum range of three orders of magnitude in $\epsilon$ as
probed by FIR luminosity in a small sample of nearby molecular clouds.
More recently, \citet{lada10} and \citet{heiderman10} count young stellar
objects in a sample of nearby clouds, and show that the values
of $\epsilon_{\rm ff}$ range up to a factor of ten larger than
the nominal value of $0.02$. \citet{evans14} report approximately
an order of magnitude spread in $\epsilon_{\rm ff}$ both above
and below 0.01 in nearby molecular clouds from the c2d and Gould's belt survey.
Using dense clumps from ATLASGAL, \citet{heyer16} find $\epsilon_{\rm ff}$
can be as low as $\sim$0.001.
We combined the Milky Way all-sky catalog of GMCs by
M-A., Miville-Desch\^enes et al. (2016) and
an all-sky catalog of SFCs by \citet{lee12} to build a large collection
of star-forming clouds (see Section \ref{sec:sfgmc}).
Our MML16 catalog of GMCs contains
5469 valid clouds, which is about an order of magnitude more clouds
than all previously published catalogs.
Using the statistical power afforded by this catalog, we showed that
both $\epsilon$ and $\epsilon_{\rm ff}$ of star-forming GMCs
in the Milky Way have large dispersions:
$\sigma_{\log\epsilon,}=0.79\pm 0.22$ dex,
and $\sigma_{\log\epsilon_{\rm ff}}=0.91\pm0.22$ dex,
(see Sections \ref{sec:sfe} and \ref{sec:sfr_ff}).
These results confirm and extend earlier observations.
The error in the scatter is dominated by
the statistical variations in the IMF
(see Section \ref{sec:flux}).
Variations in internal cloud properties or their large scale motions
cannot account for the large scatter in
$\epsilon_{\rm ff}$.
We rule out the constant star-formation rate model of \citet{krumholz05}
since cloud-to-cloud variations in the virial parameter $\alpha_{\rm vir}$ or
the Mach number ${\cal M}$ do not produce a large enough scatter
in $\epsilon_{\rm ff}$ (see Section \ref{ssec:turbsf}).
Even the improved model of turbulence-regulated star
formation by \citet{hennebelle11}---that has a strong
dependence on the Mach number---cannot account for the
observed scatter.
Similarly, we find at least an order of magnitude scatter in star formation rate
about the collision-induced star formation law proposed by \citet{tan00};
the rate at which gas is converted to star on a cloud-to-cloud
collisional timescale cannot be a constant (see Section \ref{ssec:collision}).
We see some evidence of decreasing
$\epsilon_{\rm ff}$ with larger $\tau_{\rm ff}/\tau_{\rm dyn}$
as expected by \citet{padoan12}
but the large scatter for a given time ratio cannot be explained by
varying magnetic field intensities (see Section \ref{ssec:sfr-vir}).
One way to produce a large dispersion in both
$\epsilon$ and $\epsilon_{\rm ff}$ is to arrange for a
time-variable rate of star formation in a GMC with fixed gas mass, density,
and velocity dispersion (see Section \ref{ssec:time_dependent}).
Generalizing the turbulence-regulated star formation model of \citet{krumholz05}
to allow time-variable $\epsilon_{\rm ff}$ in the prescription of \citet{feldmann11},
we find that $\epsilon_{\rm ff} \propto t^2$ is most
consistent with our data.
We conclude that star formation is dynamic on the
GMC-scale.\footnote{The time-varying $\epsilon_{\rm ff}$ may arise from
the evolution of star-forming clumps embedded in each GMC.}
Our study concerns the properties of star formation in clouds
that have already formed or in the process of forming stars.
By construction, we have limited
our analysis to 191 SFC-GMC complexes out of 5469.
We found 1014 clouds unmatched to SFCs that have the potential
to form star clusters (see Section \ref{ssec:lifetime}).
We surmise that most (5469 - 1014 - 191 = 4264; 78\% by number) of the GMCs
are gravitationally unbound
and will never form massive star clusters.
It is possible that some locally
dense clumps within an unbound GMC form
stars \citep[see e.g.,][]{dobbs11}. These clumps likely
give birth to lower mass star clusters
that our star formation tracers are not sensitive to.
Throughout our analysis, we used free-free emission
as a proxy for stellar mass. A more direct measurement would be to count
young stellar objects (YSOs). For known massive clusters,
we find a generally good agreement in the measured stellar mass
between free-free emission and YSO counts (see Figure \ref{fig2}).
A more complete comparison may be possible
using all-sky YSO catalogs \citep[e.g.,][]{marton16}. This is an important
and natural avenue for future improvement.
\subsection{Comments on Stellar Feedback}
\label{ssec:feedback}
If the rate of star formation in the most massive
($10^6M_\odot$ and higher) GMCs does accelerate,
the process of star formation must halt before $\epsilon\gtrsim 0.1$,
roughly the largest value we see in such clouds.
Stellar feedback---in the form of stellar winds, radiation pressure, protostellar jets,
and supernovae---from massive star clusters can disrupt the natal cloud and
inhibit future star formation.
The idea that stellar feedback destroys GMCs was proposed long ago by \citet{larson81}.
Galactic-scale numerical simulations (e.g., \citealt{hopkins11} and \citealt{fg13})
find that the effects of stellar feedback are necessary for regulating star
formation rates to the Kennicutt-Schmidt value. Turbulence in the interstellar medium
alone cannot slow the rate of star formation.
The existence of large expanding bubbles---a few to $\sim$100 parsecs wide---associated with
regions bright in free-free emission is evidence for the effects of stellar feedback
from young clusters (referred to as SFCs by \citealt{rahman10} and \citealt{lee12}).
These massive young clusters, which power
most of the free-free emission in the Milky Way, are identified by clear cavities in
the 8$\mu$m emission. The clusters are found in massive GMCs, but the
star clusters are not enshrouded in molecular gas---they are instead
in regions of ionized gas which are enshrouded in molecular gas.
In some cases, the massive stars may still be accreting gas from their natal protostellar
disks but the disks are not accreting gas from the host GMC.
We have examined our sample to see if the feedback from stars
is affecting their host GMCs globally.
There was no clear correlation between the variations in $\alpha_{\rm vir}$
and the ratio of disruptive forces (gas and radiation pressure) to the
binding force (dynamical pressure) of the host GMC. We also found no clear
correlation between the size of the bubbles blown by SFCs in the host clouds
and $\alpha_{\rm vir}$.
It may be that the effects of radiation and gas pressure will only become evident
later in the evolutionary history of the GMC, or it may be that other forms
of feedback (e.g., supernovae) are responsible for disrupting the
clouds. Given that GMCs in the Milky Way are found near spiral arms, combined with our finding
that the a large fraction of $10^6M_\odot$ GMCs are most likely gravitationally
bound ($\alpha_{\rm vir} < 1$), it seems unlikely that they simply disperse on their own.
\acknowledgments
The creation of the MML16 catalog of GMCs is made possible by
the all-sky CO data provided by T. Dame.
We thank the anonymous referee for their constructive feedback
that helped to improve this paper significantly.
We thank L. Blitz, G. Chabrier, E. Chiang, C-A. Faucher-Gigu\'{e}re, C. Federrath,
R. Feldmann, N. Gnedin, P. Hopkins, M. Krumholz, P-S. Li, C. Matzner, C. Mckee,
E. Ostriker, E. Quataert, S. Stahler, J. Tan,
and E. Vazquez-Semadeni for valuable discussions.
We also thank H. Isaacson, R. Trainor,
and J. Wang for their advice in statistical analyses.
EJL is supported in part by NSERC of Canada under PGS D3 and the Berkeley fellowship.
NWM is supported in part by NSERC of Canada. This research was undertaken, in part,
thanks to funding from the Canada Research Chairs program.
|
\section{Introduction} \label{sec:Intro}
A binary differential equation on the plane, or a BDE, is an implicit quadratic differential equation
\begin{equation}\label{BDE}
a(x,y) dy^2 + 2b(x,y)dx dy + c(x,y)dx^2 =0,
\end{equation}
where the coefficients $a, b, c$ are real functions which we assume to be smooth on an open set $U \subseteq {\mathbb{R}}^2$. The function $\delta: U \rightarrow \mathbb{R}, $ $\delta(x,y)= (b^2 - ac) (x,y)$, is the {\it discriminant function} and its zero set
\[ \Delta=\{(x,y) \in U: (b^2 - ac) (x,y)=0 \} \]
is the {\it discriminant set } of the BDE.
The investigation of occurrence of symmetries is converted in purely algebraic terms, so there is no loss of generality in assuming that $U$ is the whole plane, which we do from now on for simplicity.
At points where $\delta>0$, (\ref{BDE}) defines a pair of transversal directions, and by the {\it configuration} associated with this equation we mean the distribution of all solution curves tangent to these directions. The geometry of a BDE configuration is a subject of great interest, with important applications in differential geometry, as the equations of lines of curvatures, characteristic curves and asymptotic curves of smooth surfaces, and in control theory (see \cite{tari} and references therein). Conditions for local stability of positive binary differential equations ($\delta >0$) and their classification are given in \cite{guinez2} and \cite{guinez1}, with a description of the topological patterns that bifurcate in one-parameter families of these equations. Singular points of a class of positive binary equations associated with a smooth surface are also studied in\cite{guinez2} and \cite{sotoguitierrez}, the coefficients of BDE being given in terms of the coefficients of the first and second fundamental forms of the surface.
Determining models of configurations associated with BDEs has been also addressed in many works; for example, in \cite{brucetari, brucetari2, brucetari3, dara, gutierrezguinez, tari2, tari3} the classification of BDEs is performed up to topological, formal, analytic or smooth equivalences. We refer to \cite{tari} for a survey on these topics.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.5]{picture.png}
\caption{Configurations of symmetric BDEs. In (a) and (b) the symmetry group is $\mathbf{Z}_2\times \mathbf{Z}_2$ and in (c) the symmetry group is $\mathbf{D}_6$. }
\label{Figura 1}
\end{figure}
This work is motivated by the recognition of symmetries in most normal forms that appear in the works mentioned above. In the linear case for example, namely when the coefficients of (\ref{BDE}) are linear functions, presence of at least one nontrivial symmetry (namely minus identity) is a necessary condition. As a consequence, our study shows which symmetry groups are attained or never attained in the nonlinear cases. The purpose of the present paper is to introduce the systematic study of symmetries in binary differential equations. \cite{ManTemp1} is a continuation of this study: we investigate symmetries of a BDE in connection with an associated pair of equivariant vector fields (see Remark~\ref{rmk: 1-forms}); also, quadratic forms with homogeneous coefficients are studied through an analysis of the number of invariant lines that appear in the configuration space imposed by their group of symmetries; in addition, we analyse possible symmetry groups of BDEs with Morse type discriminant introduced in \cite{brucetari}.
More recently we have driven attention to some questions relating symmetries in differential forms to pairs of foliations of special classes of surfaces; for example, for a given equivariant BDE, we ask whether this can be realized as an equation of lines of curvatures or of asymptotic lines of a surface immersed in ${\mathbb{R}}^n$ for some $n$.
To remark on evidences of occurrence of symmetries in configurations associated with BDEs, let us consider the configurations that appear in \cite{brucefidal}, which are generic topological structures of principal direction fields at umbilic points of surfaces on Euclidean spaces. The normal forms are given as triples $(a, b,c)$ for $c=\pm a$ in (\ref{BDE}) and their configurations are reproduced in Fig.~\ref{Figura 1}, the so-called (a) lemon, (b) monstar and (c) star. The solution curves determine two foliations of the plane distinguished by colour, and the black point is the discriminant set.
The pictures clearly suggest an invariance of the three configurations under reflection with respect to the $x$-axis. There is another invariance with respect to the $y$-axis, which is given by this operation followed by a change of colour. As a consequence, the composition of these two elements (minus identity) must be a symmetry which interchanges colour. In fact, we should recognize {\it a priori} minus identity in the set of symmetries of all these cases by their linearity, as mentioned above. The third picture has also six rotational symmetries, three of which are colour-preserving and the other three are colour-interchanging. In fact, the full symmetry group of pictures (a) and (b) is ${\bf Z}_2 \times {\bf Z}_2$, generated by the reflections across the axes, and the full symmetry group of picture (c) is the dihedral group $\mathbf{D}_6$, generated by a reflection and a rotation of order six. As these examples illustrate, the group action must be defined taking colour changes into consideration at the region on the plane where (\ref{BDE}) defines a bivalued direction field. As we shall see, the action of a symmetry group $\Gamma$ of a binary differential equation must be defined on the tangent bundle through its representation on the plane combined with a group homomorphism $\eta : \Gamma \to {\bf Z}_2 = \{\pm 1\}$ (Definition~\ref{def:symmetric BDE}). This action is translated into an action of $\Gamma$ on the space of symmetric matrices under conjugacy. This is then used to obtain the general forms of equivariant BDEs under all compact subgroups of ${\bf O}(2)$ through invariant theory.
The paper is organized as follows: in Section~\ref{sec: symmetry action} we introduce the notion of symmetries in a BDE, namely when the equation is invariant under the linear action of a subgroup $\Gamma \leq {\bf O}(2)$. We formalize the concept using group representation theory on the tangent bundle on which the associated quadratic 1-form is defined. One of our two main results is Theorem~\ref{thm: homomorphisms formula}, which establishes a formula that reveals the effect of a symmetry in the configuration geometry in simple algebraic terms. In Section \ref{sec: inv theory} we generalize the results in \cite{patricia} for $\Gamma$-equivariant mappings with distinct actions on the source and target. The results allow the computation of general forms of equivariant mappings using an algebraic algorithm (Algorithm~\ref{algoritmo das apl geta equivariantes}). In Section~\ref{sec: general forms} we implement these results to calculate the general forms of equivariant BDEs under any compact subgroup of ${\bf O}(2)$, which is our second main result, summarized in Table~1; in this section we also present a number of examples.
\section{Symmetry groups of binary differential equations} \label{sec: symmetry action}
In this section we formalize the concept of a symmetric binary differential equation under the linear action of a compact Lie subgroup $\Gamma$ of $\mathbf{O}(2)$ .
Let ${\cal F}({\mathbb{R}}^2)$ denote the set of real $C^\infty$ quadratic differential forms $\omega$ on $M= {\mathbb{R}}^2$, $\omega : TM \to {\mathbb{R}} $ defined on the tangent bundle
$TM = \sqcup_{u \in M} T_u M$ given by
\eq \label{eq:1-form} \omega = a(x,y) dy^2 + 2b(x,y) dx dy + c(x,y) dx^2,
\feq
with $a,b,c$ $C^\infty$ functions on ${\mathbb{R}}^2$.
Let $\Gamma$ be a compact Lie group acting linearly on $\mathbb{R}^2$. This induces an action of $\Gamma$ on the tangent bundle,
\[ \begin{array}{ccc}
\Gamma \times \sqcup_{u \in M} T_u M & \to & \sqcup_{u \in M} T_u M \\
(\gamma, u, v) &\mapsto & \gamma \cdot (u,v) \ = \ \bigl(\gamma u, (d \gamma\bigr)_u (v) ).
\end{array} \]
Since the action is linear, $\gamma \cdot (u,v) =(\gamma u, \gamma v )$.
Let $\eta:~\Gamma~\rightarrow~\mathbf{Z}_2~=~\{\pm 1\}$ be a one-dimensional representation of $\Gamma$, acting on the target of $\omega$. We then give the following:
\begin{definition} \label{def:equivariant}
An element $\omega \in {\cal F}({\mathbb{R}}^2)$ is $\Gamma_\eta$-equivariant if, for all $\gamma \in \Gamma$,
\begin{equation}
\label{eq:equivariant form}\omega(\gamma \cdot (x,y,dx,dy)) \ = \ \eta(\gamma) \omega(x,y,dx,dy).
\end{equation}
\end{definition}
The BDE (\ref{BDE}) is the equation of zeros of (\ref{eq:1-form}), so we define
\begin{definition} \label{def:symmetric BDE}
For $\Gamma$ a compact Lie group acting linearly on ${\mathbb{R}}^2$ and $\eta:~\Gamma~\rightarrow~\mathbf{Z}_2$ a group homomorphism, the binary differential equation $(\ref{BDE})$ is $\Gamma$-invariant, or $\Gamma$ is the symmetry group of $(\ref{BDE})$, if $\omega$ in $(\ref{eq:1-form})$ is $\Gamma_\eta$-equivariant.
\end{definition}
We notice that the group of symmetries of a BDE generally admits, by its nature, an order-2 normal subgroup, which is equivalent to saying that the group homomorphism $\eta$ in Definition \ref{def:symmetric BDE} is nontrivial. Hence, in the present context, symmetries are established through the group together with this homomorphism, and hence this action is denoted below by $\Gamma_{\eta}$.
\begin{remark} \label{rmk: inclusion1}
Let $\Sigma(\Delta) \leq {\bf O}(2)$ denote the group of symmetries of $\Delta$, namely, the subgroup of elements of ${\bf O}(2)$ that leave $\Delta$ setwise invariant. Then
\eq \Gamma_{\eta} \leq \Sigma(\Delta).\label{eq: inclusion}\feq
In other words, symmetries of a BDE are at most the symmetries of the discriminant set. This can be of practical use when detecting the symmetry group of the equation if we know the shape of $\Delta$.
\end{remark}
Solutions of (\ref{BDE}) are nonoriented curves, associated with direction fields. At the region on the plane where the discriminant $\delta$ is positive, these form a pair of transverse foliations ${\mathscr F}_1$ and ${\mathscr F}_2$. The set of symmetries of each foliation, that is, the set of elements that leave them setwise invariant is a subgroup of $\Gamma$. The structure of this subgroup is discussed below. This pair is, in turn, associated with two oriented foliations given as integral curves of the vector fields
\eq \label{eq:vector fields}
X_i(x,y) \ = \ (a(x,y), b(x,y) +(-1)^i \sqrt{\delta(x,y)}), \ \ i=1,2.
\feq
Consider for example the equation
\[ \omega(x,y) = y dy^2 + 2x dx dy - y dx^2=0, \]
for which the associated vector fields are
\[ X_i(x,y) \ = \ (y, x +(-1)^i \sqrt{x^2 + y^2}), \ \ i=1,2. \]
The configuration of this equation is given in Fig.~\ref{Figura 1}(a). Both $X_1$ and $X_2$ are reversible-equivariant vector fields under the action of the group $\mathbf{Z}_2$ generated by the reflection $\kappa_x$ with respect to the $x$-axis,
\[ X_i(\kappa_x (x,y)) \ = \ - \kappa_x X_i(x,y), \\ i =1,2. \]
As the picture suggests, this reflection is in fact a symmetry of the BDE. Now, the combination of the two foliations adds symmetries to the whole picture, leading to a configuration which is also symmetric with respect to the reflection on the $y$-axis. By the nature of this additional symmetry, this element should invert foliations. In fact, we prove that $\mathbf{Z}_2 \times \mathbf{Z}_2$ is the symmetry group of the BDE. However, it is more subtle to realize how each element should act on the form $\omega$, in the sense that it is not obvious whether $\eta(\gamma) =1$ or $-1$ for each $\gamma$ in the group. As we shall see in Theorem~\ref{thm: homomorphisms formula}, this depends not only whether $\gamma$ preserves or interchanges foliations, but also whether it preserves or inverts orientation on the plane. \\
Before we state the result, we introduce another homomorphism: Consider the open region in ${\mathbb{R}} ^2$
\[ \Omega = \{ (x,y) \in {\mathbb{R}} ^2 \ : \ \delta(x,y) > 0\}, \]
and consider the restriction of the action of $\Gamma_\eta$ on $\Omega$. This is well-defined since the discriminant set $\Delta$ is $\Gamma$-invariant and splits the plane into two invariant regions, where $\delta$ is positive or negative. For BDEs (\ref{BDE}) for which $\Omega$ is non-empty, we introduce the homomorphism $\lambda : \Gamma \to \mathbf{Z}_2 = \{\pm1\}$,
\begin{equation} \label{eq:lambda}
\lambda (\gamma) = \left\{ \begin{array}{cl}
1, & \gamma({\mathscr F}_i) = {\mathscr F}_i \\
-1, & \gamma({\mathscr F}_i) = {\mathscr F}_j, \ j \neq i, \end{array} \right.
\end{equation}
$i,j \in \{1,2\}$. It follows directly from this definition that the subgroup of symmetries of each foliation ${\mathscr F}_i$, $i=1,2$, is
$\Sigma({\mathscr F}_i) = \ker \lambda$.
\begin{theorem} \label{thm: homomorphisms formula}
Let \ $\eta, \ \lambda : \Gamma \to \mathbf{Z}_2$ be the two group homomorphisms of Definition~$\ref{def:symmetric BDE}$ and of $(\ref{eq:lambda})$. Then, for all $\gamma \in \Gamma$,
\[ \lambda (\gamma) = \det(\gamma) \ \eta(\gamma). \]
\end{theorem}
\begin{proof}
At $(x,y) \in \Omega$ consider the pair of tangent vectors given in (\ref{eq:vector fields}),
$$ X_i(x,y)=(a(x,y), -b(x,y) + (-1)^i\sqrt{\delta(x,y)}), \ i=1,2.$$
From the definition of the action $\Gamma$ on $TM$, the pair of tangent vectors to the two solution curves at $\gamma (x,y)$ is given by $\gamma X_i(x,y)$, $i=1,2$. On the other hand, from the equivariance of $\omega$ under $\Gamma_{\eta}$, the vectors
$$X_i(\gamma(x,y))=(\eta a(x,y), -\eta b(x,y) + (-1)^i \sqrt{\delta(x,y)}), \ i=1,2,$$
are also tangent vectors to the two solution curves at $\gamma (x,y)$. By symmetry it follows that these two pairs must be parallel, i.e., there exists a nonzero $\alpha$ such that, for $i \neq j \in \{1,2\},$
\eq \label{igualdades}
\gamma X_i (x,y)= \left \{ \begin{array}{ll}
\alpha X_i(\gamma(x,y)), & \hbox{if} \ \lambda(\gamma))=1 \\
\alpha X_j(\gamma(x,y)), & \hbox{if} \ \lambda(\gamma))=-1.
\end{array} \right.
\feq
Also, by the orthogonality of the action, we have $\alpha = \pm1$.
Now, consider the two matrices $M_1$ and $M_2$ whose columns are the vectors $\gamma X_1, \gamma X_2$ and $X_1\gamma, X_2\gamma$ both calculated at $(x,y)$, respectively, that is,
$$M_1=\left(
\begin{array}{cc}
\gamma X_1 & \gamma X_2 \\
\end{array}
\right),
\ M_2=\left(
\begin{array}{cc}
\eta(\gamma)a & \eta(\gamma)a\\
-\eta(\gamma)b- \sqrt{\delta}& -\eta(\gamma)b+ \sqrt{\delta} \\
\end{array}
\right)
$$
From (\ref{igualdades}) it follows that
\[ \det(M_1)= \alpha^2 \lambda(\gamma) \det(M_2) = \lambda(\gamma) \det(M_2). \]
Finally, $\det(M_1) = \det(\gamma) \ 2a\sqrt{\delta} $, and $\det(M_2)= \eta(\gamma) 2 a \sqrt{\delta}$. Hence,
\[ \det(\gamma) = \eta(\gamma) \lambda(\gamma), \]
which implies the result since these are all group homomorphisms $\Gamma \to \mathbf{Z}_2=\{\pm 1\} $.
\end{proof}
\begin{remark}
In practice, Theorem $\ref{thm: homomorphisms formula}$ adds information to the inclusion $(\ref{eq: inclusion})$ when detecting the symmetry group $\Gamma_{\eta}$ of a BDE. In fact, it provides the construction of the homomorphism $\eta$ by the geometrical investigation of whether each element $\gamma \in \Gamma_\eta$ preserves the foliations ($\lambda(\gamma) = 1$) or interchanges the foliations ($\lambda(\gamma) = -1$). To illustrate, consider the pictures in Fig.~1. In $(a)$ and $(b)$, foliations are interchanged by $\kappa_y$, whereas they are preserved by $\kappa_x$, and now we use $\det(\kappa_x) = \det(\kappa_y) = -1$ to conclude by Theorem $\ref{thm: homomorphisms formula}$ that $\eta(\kappa_x) = - \eta(\kappa_y) = -1$. These are the generators of the symmetry group $\mathbf{Z}_2 \times \mathbf{Z}_2$, and so the homomorphism $\eta$ is well-established for these examples. In $(c)$ foliations are interchanged by $\kappa_y$ and rotation of $\pi/3$; since these are orientation reserving and orientation preserving respectively, it follows that $\eta$ assumes $1$ and $-1$, respectively. These are the generators of the symmetry group
${\bf D}_6$, and so the homomorphism $\eta$ is well-established.
\end{remark}
The following property is direct from Theorem \ref{thm: homomorphisms formula}:
\begin{corollary} \label{cor: homo formula}
$\ker \lambda \cap \ker \eta $ is a cyclic group of $\Gamma$.
\end{corollary}
A quadratic differential form (\ref{eq:1-form}) is associated to the matrix-valued mapping $B: \mathbb{R}^2 \rightarrow Sym_2,$
\eq \label{forma quadratica} B(x,y)=\left(
\begin{array}{cc}
c(x,y) & b(x,y) \\
b(x,y) & a(x,y) \\
\end{array}
\right), \feq
where $Sym_2$ denotes the set of symmetric matrices of order 2. Then (\ref{eq:1-form}) can be written as
\[ \omega \ = \ \left(
\begin{array}{c}
dx\\
dy\\
\end{array}
\right)^t \ B(x,y) \ \left(
\begin{array}{c}
dx \\
dy \\
\end{array}
\right), \]
where superscript $t$ denotes transposition. From (\ref{eq:equivariant form}), it follows that symmetries of (\ref{BDE}) are given by the following equivariance condition of $B$, with the action on the target given by the homomorphism $\eta$ and conjugacy:
\begin{equation}
B\bigl(\gamma (x,y)\bigr) = \eta(\gamma) \gamma B(x,y) \gamma^t, \ \forall \ \gamma \in \Gamma. \label{relacao}
\end{equation}
\begin{remark} \label{rmk: 1-forms}
Although Definition $\ref{def:equivariant}$ is given here for quadratic 1-forms, the equivariance condition $(\ref{eq:equivariant form})$ can be given for general 1-forms. For the linear case, it follows that
\[ \alpha(x,y, dx, dy) = A(x,y) dy + B(x,y) dx \]
is $\Gamma_\eta$-equivariant if, and only if, the associated planar mapping $F = (A, B)$ is
$\Gamma_{\eta}$-equivariant in the sense of Definition $\ref{def:inv and equiv}$ below. For an investigation of pairs of symmetric linear 1-forms and their connection with symmetries of an associated BDE we refer to \mbox{\cite{ManTemp1}}.
\end{remark}
From now on we shall use the matricial notation (\ref{relacao}) to investigate symmetries in BDEs. For a given $\Gamma_\eta$, the aim is to find the general form of symmetric matrices (\ref{forma quadratica}) satisfying (\ref{relacao}). This is an algebraic problem which is treated in Section \ref{sec: general forms}. \\
\section{\label{sec: inv theory}Invariant theory for the group $\Gamma_{\eta}$}
The aim of this section is to obtain generalizations of some results of \cite{patricia} to be applied to a systematic study of symmetries in BDEs. The idea is to use group representation theory in the study of symmetries in the space of quadratic forms. In Subsection~\ref{subsec:generalization} we register the generalization of results of \cite{patricia} to $\Gamma_{\eta}$-equivariant mappings $V \to W$ (possibly distinct source and target), whose proofs adapt straightforwardly from \cite{patricia}.
In Subsection~\ref{subseq: equiv quadratic forms} we restrict the results of the preceding subsection to equivariant mappings ${\mathbb{R}}^n \to Sym_n$, where $Sym_n$ denotes the space of symmetric matrices of order $n$. These are then used in Section \ref{sec: general forms} in the study of symmetries in binary differential equations.
\subsection{$\Gamma_{\eta}$-equivariant mappings} \label{subsec:generalization}
We start with a brief summary about algebraic invariant theory of compact Lie groups. Throughout we consider a compact Lie group $\Gamma$ acting linearly on a real vector space $V$ of finite dimension $n$. There exists a $\Gamma$-invariant inner product on $V$ under which the representation $(\rho, V)$ associated with the given action of $\Gamma$ is orthogonal, {\it i.e.} for $\gamma \in \Gamma$, $\rho(\gamma) \in \mathbf{O}(n)$, the group of orthogonal matrices of order $n$ (\cite[XII, Proposition 1.3]{golubitsky}). Hence, Lie groups in this paper are the closed subgroups of $\mathbf{O}(n)$.
A real function $f:V \rightarrow \mathbb{R}$ is {\it $\Gamma$-invariant} if
$$\label{fc invariant} f(\rho(\gamma)x)= f(x), \ \ \forall \ \gamma \in \Gamma, \ \forall \ x \in V. $$
The set $\mathcal{P}(\Gamma)$ of $\Gamma$-invariant polynomials is a ring over $\mathbb{R}$. A finite set $\{u_1, ... u_s\}$ of $\Gamma$-invariants generating this ring is called a { \it Hilbert basis} for $\mathcal{P}(\Gamma).$ The existence of a Hilbert basis was proved by Weyl in $1946$, and Schwarz proved in $1975$ that the same set generates the ring of $C^{\infty}$ $\Gamma$-invariant germs (see \cite{golubitsky}).
For $(\rho, V)$ and $(\nu, W)$ representations of $\Gamma,$ a mapping $g: V \rightarrow W$ is {\it $\Gamma$-equivariant} if
$$ \label{apl equivariante} g(\rho(\gamma))=\nu(\gamma)g(x), \ \ \forall \ \gamma \in \Gamma, \ \forall x \in V. $$
The set $\overrightarrow{\mathcal{P}}(\Gamma)$ of $\Gamma$-equivariant mappings $V \to W$ with polynomial entries is a module over $\mathcal{P}(\Gamma)$. Po\'enaru in 1976 proved that $\overrightarrow{\mathcal{P}}(\Gamma)$ is finitely generated over the ring $\mathcal{P}(\Gamma)$ and that the same set generates the module of equivariant $C^{\infty}$~germs over the ring of invariant germs.
We shall also consider a one-dimensional representation of $\Gamma$,
\eq \label{homom eta} \eta: \Gamma \rightarrow \mathbf{Z}_2=\{\pm 1\}. \feq
This is a group homomorphism with $\Gamma_{+} = \ker \eta$ a normal subgroup of $\Gamma$ of index $2$ if $\eta$ is nontrivial. From that, we also have the so-called $\eta$-{\it dual representation} of $(\nu, W)$, denoted by $\nu_\eta$, defined by the product
$$ \label{dual} \gamma \mapsto \nu_\eta(\gamma) = \eta(\gamma) \nu(\gamma).$$
\begin{definition} \label{def:inv and equiv}
For $\eta$ as in $(\ref{homom eta})$ and $(\rho, V)$ and $(\nu, W)$ representations of $\Gamma$, a function $f: V \rightarrow \mathbb{R}$ is $\Gamma_{\eta}$-invariant if
$$\label{fc eta invariant} f(\rho(\gamma)x)=\eta(\gamma)f(x), \ \ \forall \ \gamma \in \Gamma, \ \forall \ x \in V. $$
A mapping $g:V \rightarrow W$ is $\Gamma_{\eta}$-equivariant if
$$\label{ap eta eqvariant} g(\rho(\gamma)x)=\eta(\gamma)\nu(\gamma)g(x), \ \ \forall \ \gamma \in \Gamma, \ \forall \ x \in V. $$
\end{definition}
We shall denote
$\overrightarrow{\mathcal{P}}^{\eta}(\Gamma)$ and $\mathcal{P}^{\eta}(\Gamma)$ the sets of $\Gamma_{\eta}$-equivariant polynomial mappings and $\Gamma_{\eta}$-invariant polynomial functions, respectively. A $\Gamma_{\eta}$-equivariant is an equivariant $(\rho, V) \to (\nu_{\eta}, W)$, and a $\Gamma_{\eta}$-invariant is an equivariant $(\rho, V) \to (\eta, \mathbb{R})$, so the finitude of generators for each as $\mathcal{P}(\Gamma)$-modules follows by Po\'enaru's theorem. It also follows from this theorem that it is no loss of generality to restrict to modules of polynomials when finding generators of $C^\infty$ mappings. In Section~\ref{sec: general forms} we shall see how configurations change when the symmetry group of binary differential equations changes to its dual representation. \\
A connection is established in \cite{patricia} between the invariant theory for $\Gamma$ and $\Gamma_{+}$. This is done through an algebraic algorithm to compute generators of $\overrightarrow{\mathcal{P}}^{\eta}(\Gamma)$ from the knowledge of generators of $\overrightarrow{\mathcal{P}}(\Gamma_{+})$, when source and target spaces are the same. In Proposition~\ref{prop: generalization} and Algorithm~\ref{alg} we generalize this, with a similar algorithm to compute generators of $\Gamma_{\eta}$-equivariants with possibly distinct source and target.
We follow the notation used in \cite{patricia} to introduce the Reynolds operators $R:~\mathcal{P}(\Gamma_{+}) ~\rightarrow~\mathcal{P}(\Gamma_{+})$ and $\overrightarrow{R} : \overrightarrow{\mathcal{P}}(\Gamma_{+}) \rightarrow \overrightarrow{\mathcal{P}}(\Gamma_{+})$,
$$\label{reynods r e s} R(f)(x)=1/2(f(x) + f(\rho(\delta) x)) \ \ \mbox{e} \ \ \overrightarrow{R}(g)(x) =1/2(g(x) + \nu(\delta)^{-1}g(\rho(\delta) x)), $$
and, the $\eta$-Reynolds operators on $\mathcal{P}(\Gamma_{+})$ and on $\overrightarrow{\mathcal{P}}(\Gamma_{+})$, $S: \mathcal{P}(\Gamma_{+}) \rightarrow \mathcal{P}(\Gamma_{+}) $ and $\overrightarrow{S} : \overrightarrow{\mathcal{P}}(\Gamma_{+}) \rightarrow \overrightarrow{\mathcal{P}}(\Gamma_{+})$,
$$ {S}(f)(x)=1/2(f(x) - f( \rho( \delta ) x)) \ \ \mbox{e} \ \ \overrightarrow{S} (g)(x)= 1/2 (g(x) - \nu(\delta)^{-1}g(\rho(\delta) x)), $$
for an arbitrary fixed $\delta \in \Gamma \setminus \Gamma_{+}$.
Let us denote by $I_{\mathcal{P}(\Gamma_{+})}$ and $I_{ \small \overrightarrow{\mathcal{P}}(\Gamma_{+})}$ the identity maps on $\mathcal{P}(\Gamma_{+})$ and on $\overrightarrow{\mathcal{P}}(\Gamma_{+})$, respectively.
\begin{proposition} \label{prop: generalization} The operators above satisfy the following:
\begin{description}
\item[(a)] They are homomorphisms of $\mathcal{P}(\Gamma)$-modules and
$$ R + S = I_{\mathcal{P}(\Gamma_{+})} \ \ \ \mbox{and} \ \ \ \overrightarrow{R} + \overrightarrow{S}= I_{ \small \overrightarrow{\mathcal{P}}(\Gamma_{+})}.$$
\item[(b)] They are idempotent projections and the following direct sum decompositions of $\mathcal{P}(\Gamma)$-modules hold:
\eq \label{decomposition} \mathcal{P}(\Gamma_{+}) = \mathcal{P}(\Gamma) \oplus \mathcal{P}^{\eta}(\Gamma) \ \ \ \mbox{and} \ \ \ \overrightarrow{\mathcal{P}}(\Gamma_{+}) = \overrightarrow{\mathcal{P}}(\Gamma) \oplus \overrightarrow{\mathcal{P}}^{\eta}(\Gamma). \feq
\end{description}
\end{proposition}
\begin{proof}
Analogous to Propositions 2.3 and 2.4 in \cite{patricia}.
\end{proof}
The algorithm is based on the decompositions (\ref{decomposition}) and on the projection operators $S$ and $\overrightarrow{S}$ applied to a given Hilbert basis of $\mathcal{P}(\Gamma_{+})$ and a set of generators of $\overrightarrow{\mathcal{P}}(\Gamma_{+})$. The procedure is:
\begin{algthm} \label{alg}
\label{algoritmo das apl geta equivariantes}
Let $\Gamma$ be a closed subgroup of $\mathbf{O}(n)$ and $\eta: \Gamma \rightarrow \mathbf{Z}_2$ a homomorphism with $\ker \eta =\Gamma_{+}$, $\{ u_1,...,u_s\} $ a Hilbert basis of $\mathcal{P}(\Gamma_{+})$ and $\{H_0,...,H_r\} $ a generator set of $\overrightarrow{\mathcal{P}}(\Gamma_{+})$ as a $\mathcal{P}(\Gamma_{+})$-module;
\begin{description}
\item[1] Fix $\delta \in \Gamma \setminus \Gamma_{+}$ arbitrary;
\item[2] For $i \in \{1, ..., s\}$, do $\tilde{u}_i=S(u_i), \tilde{u_0}:=1;$
\item[3] For $i \in \{1, ..., s\}$ and $j \in \{0, ..., r\}$, do $H_{ij}=\tilde{u}_i H_j;$
\item[4] For $i \in \{1, ..., s\}$ and $j \in \{0, ..., r\}$, do $\tilde{H}_{ij}=\overrightarrow{S}(H_{ij})$.
\end{description}
{\bf Result:} $\{\tilde{H}_{ij}: \ 0 \leq i \leq s, \: 0 \leq j \leq r\}$ is a generator set of $\overrightarrow{\mathcal{P}}^{\eta}(\Gamma)$ as a $\mathcal{P}(\Gamma)$-module.
\end{algthm}
As proved in \cite{patricia}, step 2 above provides a generator set of the $\mathcal{P}(\Gamma)$-module $\mathcal{P}^{\eta}(\Gamma)$ (these are the anti-invariants in that paper). What we also remark at this point is that replacing $\overrightarrow{S}$ by the projection operator $\overrightarrow{R}$ in step 4 we obtain, as expected, a direct way to compute a set of generators for the equivariants under the whole group $\Gamma$ from the knowledge of equivariants under the subgroup $\Gamma_{+}$. This is formalized below:
\begin{proposition} \label{prop rflexa} Let $\Gamma$ be a compact Lie group acting on $V$ and on $W$ and $\{H_{ij}=\tilde{u}_i H_j, 0 \leq i \leq s , 0\leq j \leq r\}$ a generator set of $\overrightarrow{\mathcal{P}}(\Gamma_{+})$ as a $\mathcal{P}(\Gamma)$-module given by step $3$ in Algorithm~$\ref{algoritmo das apl geta equivariantes}$). Then
$$\{\overrightarrow{R}(H_{i,j}),0 \leq i \leq s , 0\leq j \leq r \}$$
generates $\overrightarrow{\mathcal{P}}(\Gamma)$ as a $\mathcal{P}(\Gamma)$-module.
\end{proposition}
\begin{proof}
Let $g \in \overrightarrow{\mathcal{P}}(\Gamma) \subset \overrightarrow{\mathcal{P}}(\Gamma_{+})$. Then
$g=\sum_{i,j}^{s,r} p_{ij}H_{ij}, \ \ \ p_{ij} \in \mathcal{P}(\Gamma),$ $0 \leq i \leq s$ e $0 \leq j \leq r$.
Since $\overrightarrow{R}$ is a $\mathcal{P}(\Gamma)$-homomorphism and $\overrightarrow{R}(g)=g$, then $$g=\overrightarrow{R}(g)=\overrightarrow{R}\left(\sum_{i,j}^{s,r} p_{ij}H_{ij}\right)= \sum_{i,j}^{s,r} p_{ij} \overrightarrow{R}(H_{ij}). $$
\end{proof}
\subsection{$\Gamma_{\eta}$-equivariant quadratic forms} \label{subseq: equiv quadratic forms}
Let $\Gamma$ be a closed subgroup of $\mathbf{O}(n)$ acting linearly on $\mathbb{R}^n$. In this subsection we just rewrite the Reynolds operators of Subsection~\ref{subsec:generalization} applied to the module of $\Gamma_{\eta}$-equivariant quadratic forms of ordem $n$. In the present context, the representation on the target is defined from the representation on the source by conjugacy.
We consider the action of $\Gamma$ on $Sym_n$, the space of symmetric matrices of order $n$, given by conjugacy,
$$\label{acao Symn} (\gamma, B) \mapsto \gamma B \gamma^t, \ \gamma \in \Gamma, \ B \in Sym_n . $$
\begin{definition}
A quadratic form of order $n$ is a matrix-valued mapping $B:\mathbb{R}^n \rightarrow Sym_n$.
\end{definition}
Following the notation of Subsection~\ref{subsec:generalization}, $\overrightarrow{\mathcal{P}}(\Gamma)$ is the module of the $\Gamma$-equivariants
$$B (\gamma x)=\gamma B(x) \gamma^t, \ \ \forall \gamma \in \Gamma, \ x \in \mathbb{R}^n\label{gamma equiv}, $$
and for $\eta: \Gamma \rightarrow \mathbf{Z}_2=\{\pm 1\}$ a nontrivial group homomorphism, $\overrightarrow{\mathcal{P}}^{\eta}(\Gamma)$ is the module of the $\Gamma_{\eta}$-equivariants
$$ B(\gamma x)=\eta(\gamma) \gamma B(x) \gamma^t, \ \ \forall \gamma \in \Gamma, \ x \in \mathbb{R}^n. \label{geta equiv} $$
The Reynolds operators
$\overrightarrow{R}, \overrightarrow{S} :\overrightarrow{\mathcal{P}}(\Gamma_{+}) \rightarrow \overrightarrow{\mathcal{P}}(\Gamma_{+}),$ are now
$$\label{reynods r e s} \overrightarrow{R} (g)(x) =1/2(g(x) + \delta^{-1}g(\delta x) \delta), \ \
\overrightarrow{S}(g)(x)= 1/2 (g(x) - \delta^{-1}g(\delta x)\delta), $$
for an arbitrary fixed $\delta \in \Gamma \setminus \Gamma_{+}$.
\section{\label{sec: general forms} General forms of symmetric BDEs}
The aim of this section is to present the algebraic forms of BDEs symmetric under the compact subgroups of ${\bf O}(2)$ with its standard action on the plane. These are derived from generator sets of the modules of equivariant quadratic forms on the plane
\eq \label{quadratic form} B(x,y)=\left(
\begin{array}{cc}
c(x,y) & b(x,y) \\
b(x,y) & a(x,y) \\
\end{array}
\right). \feq
These modules are $\overrightarrow{\mathcal{P}}(\Gamma)$ or $\overrightarrow{\mathcal{P}}^{\eta}(\Gamma)$ of Subsection~\ref{subseq: equiv quadratic forms} for $n=2$.
If the group homomorphism $\eta : \Gamma \to \mathbf{Z}_2$ is nontrivial we apply Algorithm~\ref{algoritmo das apl geta equivariantes}. For clarification of exposition, for each $\Gamma$ and for each possible $\eta$ we shall denote $\Gamma_{\eta}$ by $\Gamma[\ker \eta]$ when $\eta$ is nontrivial, and by $\Gamma$ otherwise. We also denote $\mathcal{P}^{\eta}(\Gamma)$ by $\mathcal{P}[\Gamma, \ker \eta]$ and $\overrightarrow{\mathcal{P}}^{\eta}(\Gamma)$ by $\overrightarrow{\mathcal{P}}[\Gamma, \ker \eta]$.
For the computations below we shall use the action of (subgroups of) $\mathbf{O}(2)$ on $\mathbb{R}^2\simeq \C$ with the usual semi-direct product of $\mathbf{SO}(2)$ and $\mathbf{Z}_2(\kappa)$, using complex coordinates,
$$\label{acao padrao} \theta \cdot z=e^{i \theta}z,\ \ \hbox{and} \ \ \kappa z= \bar{z}, \ \ \theta \in [0, 2\pi], \ z \in \C.$$
In Subsections~\ref{subsec: znzn} and \ref{subsec: other zn} we derive the general forms of symmetric BDEs under the cyclic group $\mathbf{Z}_n$, $n \geq 3$, and in Subsection~\ref{subsec:z2} the general forms under ${\mathbf{Z}}_2$, for all possible homomorphisms $\eta$. For the other compact subgroups of ${\bf O}(2)$ the computation is similar and shall be omitted. In Subsection \ref{subsection:table} all general forms are given in Table~\ref{tabela forma gerais}.
\subsection{\label{subsec: znzn}$\mathbf{Z}_n$-equivariant quadratic forms}
Here we consider the cyclic group $\mathbf{Z}_n$, $n \geq 3$, with $\eta$ trivial. We compute generators of $\overrightarrow{\mathcal{P}}(\mathbf{Z}_n)$ by computing generators of $\mathcal{M}(\mathbf{Z}_n)$, the module of $\mathbf{Z}_n$-equivariant matrix-valued mappings $ \mathbb{R}^2 \to M_2(\mathbb{R}^2)$, and projecting onto the space of mappings $\mathbb{R}^2 \to Sym_2$ . In complex coordinates we write any element of $\mathcal{M}(\mathbf{Z}_n)$ as
\eq z \mapsto \alpha(z) w + \beta(z) \bar{w}, \ \forall w \in {\mathbb{R}^2}, \label{matrizcomplexa} \feq
for functions $\alpha=\alpha_1 +i \alpha_2$ and $\beta=\beta_1 + i\beta_2$, with $\alpha_j, \beta_j, j=1,2$, real functions. Associating it with the real matrix
$$
M = \left(
\begin{array}{cc}
\alpha_1 + \beta_1 & \beta_2 - \alpha_2 \\
\alpha_2 + \alpha_2 & \alpha_1-\beta_1 \\
\end{array}
\right),
$$
the desired quadratic forms are obtained by the projection
\begin{equation} \label{matrizsimetrica}
M \mapsto B = 1/2 (M + M^t),
\end{equation}
after imposing the $\mathbf{Z}_n$-symmetry condition.
Write ($\ref{matrizcomplexa}$) as
$$ M(z)w=\sum\alpha_{jk}z^j\bar{z}^kw + \sum \beta_{jk}z^j\bar{z}^k\bar{w}, \ \ \ \ \alpha_{jk}, \beta_{jk} \in \C.$$
The equivariance with respect to $\theta \in \mathbf{Z}_n$ gives
\eq \label{1} M(z)w=\sum\alpha_{jk}e^{i\theta(j-k)}z^j\bar{z}^kw + \sum \beta_{jk} e^{i \theta(j-k-2)} z^j\bar{z}^k\bar{w}. \feq
So $\alpha_{jk}=\alpha_{jk}e^{i\theta(j-k)}$ and $\beta_{jk}=\beta_{jk}e^{i\theta(j-k-2)}$ for $\theta={2k\pi}/{n}, \ k =1, \ldots, n$, and so
\eq \label{equiv 1}\alpha_{jk}=0 \ \mbox{ou} \ j\equiv k\mbox{(mod n)} \ \ \ \mbox{and} \ \ \ \beta_{jk}=0 \ \mbox{ou} \ \ j\equiv k+2\mbox{(mod n)}.\feq
A Hilbert basis for $\mathcal{P}(\mathbf{Z}_n)$ is given in \cite{symmetryinchaos},
$$\{ u_1 = z\bar{z}, u_2 =z^n + \bar{z}^n, u_3 =i(z^n -\bar{z}^n)\}.$$
Factor out $z\bar{z}$ in (\ref{1}) and use (\ref{equiv 1}) to get
\begin{eqnarray*}
M(z)w &=& \sum_{j\geq k} \alpha_{jk}(z\bar{z})^k z^{j-k}w + \sum_{j< k} \alpha_{j k} (z\bar{z})^j \bar{z}^{k-j}w + \sum_{j\geq k} \beta_{jk}(z\bar{z})^k z^{j-k}\bar{w} + \\ & + & \sum_{j < k} \beta_{jk} (z\bar{z})^j \bar{z}^{j-k}\bar{w} \ = \ \sum c^1_{kl_1}(z\bar{z})^k z^{l_1 n}w + \sum c^2_{jl_2} (z\bar{z})^j \bar{z}^{l_2 n}w + \\ &+& \sum c^3_{k l_3}(z\bar{z})^k z^{l_3n+2}\bar{w} + \sum c^4_{jl_4} (z\bar{z})^j \bar{z}^{l_4 n -2}\bar{w},
\end{eqnarray*}
where, $c^t \in \C$, $l_t \in \mathbb{N}$, $t=1, ...,4$, \ $l_1, l_2, l_3 \geq0$ and $l_4\geq1$. We now use the identities
\[ \begin{array}{lll}
z^{ln} &=& (z^n+\bar{z}^n)z^{(l-1)n} - (z \bar{z})^n z^{(l-2)n}\\
\bar{z}^{ln}&= &(z^n+\bar{z}^n)\bar{z}^{(l-1)n} - (z \bar{z})^n \bar{z}^{(l-2)n}\\
z^{ln +2}&= &(z^n+\bar{z}^n)z^{(l-1)n +2} - (z \bar{z})^n z^{(l-2)n+2} \\
\bar{z}^{ln-2}&= &(z^n+\bar{z}^n)\bar{z}^{(l-1)n-2} - (z \bar{z})^n \bar{z}^{(l-2)n-2} \\
\bar{z}^n&= & (z^n + \bar{z}^n)-z^n \\
z^{n+2}&= & (z^n+\bar{z}^n)z^2 - (z \bar{z})^2 \bar{z}^{n-2} \\
\bar{z}^{(l+1)n-2}&= & (z^{ln}+\bar{z}^{ln})\bar{z}^{n-2} - (z \bar{z})^{n-2}z^{(l-1)n+2}
\end{array} \]
to conclude that a set of generators of $\mathcal{M}_2(\mathbf{Z}_n)$ over $\mathcal{P}(\mathbf{Z}_n)$ is given by the elements
$$ M_1(z)w=w, \ M_2(z)w=iw, \ M_3(z)w=z^2\bar{w}, \ M_4(z)w=iz^2\bar{w}, \ M_5(z)w=\bar{z}^{n-2}\bar{w},$$ $$M_6(z)w=i\bar{z}^{n-2}\bar{w}, \ M_7(z)w=z^n w, \ M_8(z)w=iz^n w.$$
We now apply the projection (\ref{matrizsimetrica}) to the elements above to find generators of $\overrightarrow{\mathcal{P}}(\mathbf{Z}_n)$,
$$B_{1}(z)w=w, \ B_3(z)w=z^2\bar{w}, \ B_4(z)w=iz^2\bar{w}, \ B_5(z)w=\bar{z}^{n-2}\bar{w}, \ B_6(z)w= i\bar{z}^{n-2}\bar{w}.$$
\subsection{$\mathbf{Z}_n[\mathbf{Z}_{n/2}]$-equivariant quadratic forms, for $n \geq 4$ even }
\label{subsec: other zn}
In this case, $\ker \eta = \mathbf{Z}_{n/2}$. From the preceding subsection we extract
$$H_0(z)w=w, H_1(z)w=z^2\bar{w}, H_2(z)w=iz^2\bar{w}, H_3(z)w=\bar{z}^{n/2-2}\bar{w}, H_4(z)w=i\bar{z}^{n/2-2}\bar{w}$$
as generators of $\overrightarrow{\mathcal{P}}(\mathbf{Z}_{n/2})$ over the ring $\mathcal{P}(\mathbf{Z}_{n/2})$ whose Hilbert basis is
$$\{u_1(z)=z \bar{z}, u_2(z)=z^{n/2} + \bar{z}^{n/2}, u_3(z)=i(z^{n/2} - \bar{z}^{n/2}) \}, $$
We now apply Algorithm~\ref{algoritmo das apl geta equivariantes}:
\begin{enumerate}
\item Fix $\delta=e^{2 \pi i/n} \in \mathbf{Z}_n \setminus \mathbf{Z}_{n/2}$;
\item Generators of $\mathcal{P}[{\mathbf{Z}_n, \mathbf{Z}_{n/2}}]$ over $\mathcal{P}(\mathbf{Z}_n)$:
$$\tilde{u}_1(z)=S(u_1)(z)=\frac{1}{2}(z\bar{z}-(e^{{2\pi i}/n}z)(e^{-{2 \pi i}/n}\bar{z}))=0.$$
$$\tilde{u}_2(z)=S(u_2)(z)=\frac{1}{2}(z^{n/2} + \bar{z}^{n/2} - (e^{\pi i}z^{n/2} + e^{-\pi i}\bar{z}^{n/2})=z^{n/2} + \bar{z}^{n/2}.$$
$$\tilde{u}_3(z)=S(u_3)(z)=\frac{1}{2}(i(z^{n/2} - \bar{z}^{n/2}) - i((e^{\pi i}z^{n/2} - e^{-2\pi i}\bar{z}^{n/2})= i(z^{n/2} - \bar{z}^{n/2}).$$
\item Generators of $\overrightarrow{\mathcal{P}}[\mathbf{Z}_n, \mathbf{Z}_{n/2}]$ over $\mathcal{P}(\mathbf{Z}_n/2)$: set $\tilde{u}_0(z)=1$,
$$H_{0j}(z)w=\tilde{u}_0(z)H_j(z)w=H_j(z)w, \ j=0, ...,4;$$
$$H_{1j}(z)w=\tilde{u}_1(z)H_j(z)w=0, \ j=0, ..., 4;$$
$$H_{20}(z)w=\tilde{u}_2(z)H_0(z)w=(z^{n/2} + \bar{z}^{n/2})w;$$
$$H_{21}(z)w=\tilde{u}_2(z)H_1(z)w=(z^{n/2 +2} + (z\bar{z})^2 \bar{z}^{n/2-2})\bar{w};$$
$$H_{22}(z)w=\tilde{u}_2(z)H_2(z)w=i (z^{n/2 +2} + (z\bar{z})^2 \bar{z}^{n/2-2})\bar{w};$$
$$H_{23}(z)w=\tilde{u}_2(z)H_3(z)w=(\bar{z}^{n/2 -2} + (z\bar{z})^{n/2-2} z^2)\bar{w};$$
$$H_{24}(z)w=\tilde{u}_2(z)H_4(z)w=i(\bar{z}^{n/2 -2} + (z\bar{z})^{n/2-2} z^2)\bar{w};$$
$$H_{30}(z)w=\tilde{u}_3(z)H_0(z)w=i(z^{n/2}- \bar{z}^{n/2})w;$$
$$H_{31}(z)w=\tilde{u}_3(z)H_1(z)w=i(z^{n/2 +2} -(z\bar{z})^2 \bar{z}^{n/2-2})\bar{w};$$
$$H_{32}(z)w=\tilde{u}_3(z)H_2(z)w=(-z^{n/2 +2} +(z\bar{z})^2 \bar{z}^{n/2-2})\bar{w};$$
$$H_{33}(z)w=\tilde{u}_3(z)H_3(z)w=i (-\bar{z}^{n-2} + (z\bar{z})^{n/2-2}z^2)\bar{w};$$
$$H_{34}(z)w=\tilde{u}_3(z)H_4(z)w=(\bar{z}^{n-2} - (z\bar{z})^{n/2-2}z^2)\bar{w},$$
which, as an intermediate step, we simplify to the reduced list
$$ H_{00}(z)w=w, H_{01}(z)w=z^2\bar{w}, H_{02}(z)w=iz^2\bar{w}, H_{03}(z)w=\bar{z}^{n/2-2}\bar{w},$$
$$ H_{04}(z)w=i\bar{z}^{n/2-2}\bar{w}, H_{20}(z)w=(z^{n/2} - \bar{z}^{n/2}) w, H_{21}(z)w=z^{n/2+2}\bar{w}, H_{22}(z)w=iz^{n/2+2}\bar{w},$$
$$ H_{23}(z)w=\bar{z}^{n-2}\bar{w}, H_{24}(z)w=i\bar{z}^{n-2}\bar{w},
H_{30}(z)w=i(z^{n/2} -\bar{z}^{n/2})w. $$
\item Generators of $\overrightarrow{\mathcal{P}}[\mathbf{Z}_n, \mathbf{Z}_{n/2}]$ over $\mathcal{P}(\mathbf{Z}_n)$:
$$\tilde{H}_{00}(z)w=\tilde{H}_{01}(z)w=\tilde{H}_{02}(z)w=0;$$
$$\tilde{H}_{03}(z)w=\bar{z}^{n/2-2}\bar{w};$$
$$\tilde{H}_{04}(z)w=i\bar{z}^{n/2-2}\bar{w};$$
$$\tilde{H}_{20}(z)w=(z^{n/2} + \bar{z}^{n/2}) w;$$
$$\tilde{H}_{21}(z)w=z^{n/2+2}\bar{w};$$
$$\tilde{H}_{22}(z)w=iz^{n/2+2}\bar{w};$$
$$\tilde{H}_{23}(z)w=\tilde{H}_{24}(z)w=0;$$
$$\tilde{H}_{20}(z)w=i(z^{n/2} - \bar{z}^{n/2}) w.$$
\end{enumerate}
Therefore, $\overrightarrow{\mathcal{P}}[\mathbf{Z}_n, \mathbf{Z}_{n/2}]$ is the $\mathcal{P}(\mathbf{Z}_n)$-module generated by $$\tilde{B}_1(z)w=\bar{z}^{n/2-2}\bar{w}, \ \tilde{B}_2(z)w=i \bar{z}^{n/2 -2}\bar{w}, \ \tilde{B}_3(z)w=(z^{n/2} + \bar{z}^{n/2})w,$$
$$ \tilde{B}_4(z)w=z^{n/2+2}\bar{w}, \ \tilde{B}_5(z)w=iz^{n/2+2}\bar{w}, \ \tilde{B}_6(z)w=i(z^{n/2} - \bar{z}^{n/2}) w.$$
\subsection{${\mathbf{Z}}_2$-equivariant quadratic forms} \label{subsec:z2}
Let ${\mathbf{Z}}_2$ be the group generated by the reflection $\kappa_x$ on the $x$-axis. First we consider $\eta : {\mathbf{Z}}_2(\kappa_x) \to \mathbf{Z}_2$ trivial. Imposing the ${\mathbf{Z}}_2$-equivariance to (\ref{quadratic form}) gives
$$ \label{eq: z2 equi} \left(
\begin{array}{cc}
c(x,-y) & b(x,-y) \\
b(x,-y) & a(x,-y) \\
\end{array}
\right)=\left(
\begin{array}{cc}
c(x,y) & -b(x,y) \\
-b(x,y) & a(x,y) \\
\end{array}
\right).
$$
This is to say that $a$ and $c$ are ${\mathbf{Z}}_2[ {\bf 1}]$ -equivariant and $b$ is ${\mathbf{Z}}_2$-anti-invariant. Therefore, the generators of $\overrightarrow{\mathcal{P}}({\mathbf{Z}}_2)$ under $\mathcal{P}({\mathbf{Z}}_2)$ are
$$(x,y) \mapsto \left(
\begin{array}{cc}
1 & 0 \\
0 & 0 \\
\end{array} \right), \ (x,y) \mapsto \left(
\begin{array}{cc}
0 & 0 \\
0 & 1 \\
\end{array}\right), \ (x,y) \mapsto \left(
\begin{array}{cc}
0 & y \\
y & 0 \\
\end{array} \right).$$
Assume now $\eta$ nontrivial, so $\ker \eta = {\bf 1}$. Imposing the ${\mathbf{Z}}_2[{\bf 1}]$-equivariance to (\ref{quadratic form}) gives
$$ \label{eq1: z2 equi} \left(
\begin{array}{cc}
c(x,-y) & b(x,-y) \\
b(x,-y) & a(x,-y) \\
\end{array}
\right)=\left(
\begin{array}{cc}
-c(x,y) & b(x,y) \\
b(x,y) & -a(x,y) \\
\end{array}
\right).
$$
Hence $b$ is ${\mathbf{Z}}_2$-invariant and the functions $a$ and $c$ are ${\mathbf{Z}_2}[{\bf 1}]$-equivariant. Therefore, the generators for $\overrightarrow{\mathcal{P}}({\mathbf{Z}_2},{\bf 1})$ under $\mathcal{P}({\mathbf{Z}}_2)$ are
$$(x,y) \mapsto \left(
\begin{array}{cc}
0 & 1 \\
1 & 0 \\
\end{array} \right), \ (x,y) \mapsto \left(
\begin{array}{cc}
y & 0 \\
0 & 0 \\
\end{array}\right), \ (x,y) \mapsto \left(
\begin{array}{cc}
0 & 0 \\
0 & y \\
\end{array} \right).$$
\subsection{ \label{subsection:table} Summarizing table and illustrations}
In this subsection we present the general forms of symmetric quadratic differential 1-forms $ \omega = a(x,y)dy^2+2b(x,y)dxdy +c(x,y)dy^2$ under compact subgroups of $\mathbf{O}(2)$.
Table~1 shows each group $\Gamma$ with all possible values of $\ker \eta$, denoted by $\Gamma[\ker \eta]$. Following the previous notation, when $\eta$ is trivial the group is denoted simply by $\Gamma$. Also, $\mathbf{D}_{n}(\kappa_x)$ and $\mathbf{D}_{n}(\kappa_y)$ shall denote the dihedral groups generated by the rotation of angle $2 \pi/n$ and by the reflections with respect to the $x$-axis or $y$-axis, respectively.
In \cite{brucetari} the authors consider BDEs whose discriminant function $\delta=b^2 - ac$ is of Morse type. In this case, the discriminant set is a pair of transversal straight lines by the origin or the origin itself. They prove that these BDEs are topologically equivalent to their linear part. We remark that all the normal forms that they obtain must be equivariant under a finite symmetry group. In fact, it follows from Table~1 that there are no linear BDEs with infinite group of symmetries. As it appears in \cite{brucetari}, the Morse condition is given in terms of the coefficients of the linear part of the smooth functions $a, b$ and $c$. More precisely, if we write $a=a_1x + a_2y + {\bf o}(2)$, $b=b_1x + b_2y + {\bf o}(2)$ and $c=c_1x + c_2y + {\bf o}(2)$, then the condition is
\begin{equation}
\label{morse} (c_2a_1-c_1a_2)^2 -4(b_2a_1-b_1a_2)(c_2b_1-c_1b_2)\neq 0.
\end{equation}
From Table~1, the possible symmetry groups of BDEs whose linear parts satisfy (\ref{morse}) are
\begin{equation} \label{eq: class 1}
\mathbf{Z}_3, \ \mathbf{Z}_6[\mathbf{Z}_3], \ \mathbf{D}_3, \mathbf{D}_3[\mathbf{Z}_3], \ \mathbf{D}_6[\mathbf{D}_3]
\end{equation}
or
\begin{equation} \label{eq: class 2}
\mathbf{Z}_2, \ \mathbf{Z}_2[{\bf 1}], \ \mathbf{Z}_2\times \mathbf{Z}_2[\mathbf{Z}_2(\kappa_x)].
\end{equation}
Recall from Remark \ref{rmk: inclusion1} that the set of all symmetries of a BDE is at most the symmetry group $\Sigma(\Delta)$ of the discriminant set. Hence, for the Morse cases it follows that if $\Delta$ is the origin, then the possible nontrivial symmetry groups are the ones in (\ref{eq: class 1}), whereas the groups listed in (\ref{eq: class 2}) are the possible groups when the discriminant set is a pair of transversal straight lines. We also point out that the finiteness of the symmetry group also holds for equations with constant coefficients. A classification of these two types of BDEs is done in \cite{ManTemp1}, including an analysis of the corresponding group of symmetries of the equation with possible number of invariant lines in the associated configuration.
\begin{remark}
The symmetry group of the configuration shown in Fig.1$(c)$ is $\mathbf{D}_6[\mathbf{D}_3(\kappa_y)]$, whose quadratic form $(y, x, -y)$
appears in Table $1$ by interchanging the variables $x$ and $y$ and taking $p_1 \equiv 1$ and $p_2 \equiv p_3 \equiv 0$ in the general form for the group $\mathbf{D}_6[\mathbf{D}_3(\kappa_x)].$ Similarly, the symmetry group of the configurations in Fig.~\ref{Figura 1}$(a)$ and $(b)$ is $\mathbf{Z}_2 \times \mathbf{Z}_2[\mathbf{Z}_2(\kappa_y)]$, whose quadratic forms appear from the data for $\mathbf{Z}_2\times \mathbf{Z}_2[\mathbf{Z}_2(\kappa_x)]$ in Table~1 by interchanging $x$ and $y$ and taking $p_1\equiv p_2 \equiv1, p_3\equiv -1$, and $p_1\equiv1, p_2 \equiv \frac{1}{4}, p_3 \equiv -1$, respectively.
\end{remark}
\begin{table}[!h] \label{table}
\begin{center}
\begin{tabular}{|c|c|l|}
\hline
$\Gamma[\ker \eta]$ & $\ker \lambda$ & General form \\ \hline
& & $a=p_1 + (y^2-x^2)p_2 + 2xyp_3;$ \\
$\mathbf{SO}(2)$ & $\mathbf{SO}(2)$ & $b= 2xyp_2 + (x^2-y^2)p_3;$ \\
& & $c= p_1 + (x^2-y^2)p_2 - 2xyp_3,$ \\
& & $ p_i \in \mathcal{P}(\mathbf{SO}(2)), i=1,2,3.$ \\ \hline
$\mathbf{O}(2)$ & $\mathbf{SO}(2)$ & $a= p_1 + (y^2-x^2)p_2; \ b=2xyp_2; $ \\ & & $ c=p_1 + (x^2-y^2)p_2, \ p_i \in \mathcal{P}(\mathbf{O}(2)), i=1,2.$ \\ \hline
& & $a= 2xyp;$ \\
$\mathbf{O}(2)[\mathbf{SO}(2)]$ & $\mathbf{O}(2)$ & $b= (x^2-y^2)p;$ \\
& & $ c= -2xyp, \mathcal{P} \in \mathcal{P}(\mathbf{O}(2)). $ \\ \hline
& & $a= p_1 + (y^2-x^2)p_2 + 2xyp_3 -A_1p_4 -A_2p_5;$ \\
$\mathbf{Z}_n, $ & $\mathbf{Z}_n$ & $b= 2xyp_2 + (x^2-y^2)p_3 + A_1p_5 -A_2p_4;$ \\
$n \geq 3$ & & $c= p_1 + (x^2-y^2)p_2 -2xyp_3 +A_1p_4 + A_2p_5, $ \\
& & $ p_i \in \mathcal{P}(\mathbf{Z}_n ), i=1,..., 5.$ \\ \hline
& & $a= -A_3p_1 - A_4p_2 + A_5p_3 -A_7p_4 + A_8p_5 + A_6p_6;$ \\
$\mathbf{Z}_n[\mathbf{Z}_{n/2}], $ & $\mathbf{Z}_{n/2}$ & $b= -A_4p_1 + A_3p_2 + A_8p_4 + A_7p_5; $ \\
$ n \geq 4$ even & & $c= A_3p_1 + A_4p_2 + A_5p_3 +A_7p_4 - A_8p_5 + A_6p_6,$ \\
& & $ p_i \in \mathcal{P}(\mathbf{Z}_n), i=1, ..., 6. $ \\ \hline
& & $a= p_1 + (y^2-x^2)p_2 - A_1p_3;$ \\
$\mathbf{D}_n, $ & $\mathbf{Z}_n$ & $b= 2xyp_2 - A_2p_3; $ \\
$n \geq 3 $ & & $c= p_1 + (x^2-y^2)p_2 + A_1p_3, \ p_i \in \mathcal{P}(\mathbf{D}_n), i=1,2,3.$ \\ \hline
& & $a= 2xyp_1 -A_2p_2 +A_9p_3;$ \\
$\mathbf{D}_n[\mathbf{Z}_n], $ & $\mathbf{D}_n$ & $b= (x^2-y^2)p_1 + A_1p_2; $ \\
$n \geq 3$ & & $c=-2xyp_1 + A_2p_2 +A_9p_3, \ p_i \in \mathcal{P}(\mathbf{D}_n), i=1, 2, 3.$ \\ \hline
& & $a= -A_3p_1 +A_5p_2 -A_7p_3; $ \\
$\mathbf{D}_n[\mathbf{D}_{n/2}(\kappa_x)], $ & $\mathbf{D}_{n/2}(\kappa_y)]$ & $b= -A_4p_1 +A_8p_3; $ \\
$ n \geq 4$ even & & $c= A_3p_1 +A_5p_2 + A_7p_3, \ p_i \in \mathcal{P}(\mathbf{D}_n), i=1, 2, 3.$ \\ \hline
$\mathbf{Z}_2$ & ${\bf 1}$ & $a= p_1; \ b= yp_2; \ c=p_3, \ p_i \in \mathcal{P}(\mathbf{Z}_2), i=1, 2,3.$ \\ \hline
$\mathbf{Z}_2[{\bf 1}]$ & $\mathbf{Z}_2$ & $a= y p_1; \ b= p_2; \ c=yp_3, \ p_i \in \mathcal{P}(\mathbf{Z}_2), i=1, 2, 3.$ \\ \hline
$\mathbf{Z}_2 \times \mathbf{Z}_2$ & $\mathbf{Z}_2(-I)$ & $a=p_1; \ b= xyp_2; \ c=p_3, \ p_i \in \mathcal{P}(\mathbf{Z}_2 \times \mathbf{Z}_2), i=1, 2, 3.$ \\ \hline
$\mathbf{Z}_2 \times \mathbf{Z}_2[\mathbf{Z}_2(-I)]$ & $\mathbf{Z}_2 \times \mathbf{Z}_2 $ & $a=xyp_1; \ b= p_2; \ c=xyp_3, p_i \in \mathcal{P}(\mathbf{Z}_2 \times \mathbf{Z}_2), i=1, 2, 3.$ \\ \hline
$\mathbf{Z}_2 \times \mathbf{Z}_2[\mathbf{Z}_2(\kappa_x)]$ & $\mathbf{Z}_2(\kappa_y)$ & $a= xp_1; \ b= yp_2; \ c=xp_3, p_i \in \mathcal{P}(\mathbf{Z}_2 \times \mathbf{Z}_2), i=1, 2, 3.$ \\ \hline
\multicolumn{3}{|l|}{$A_1= \hbox{Re}(z^{n-2}), \ A_2= \hbox{Im}(z^{n-2}), \ A_3= \hbox{Re}(z^{n/2-2}), \ A_4= \hbox{Im}(z^{n/2-2}), \ A_5= \hbox{Re}(z^{n/2}),$ } \\ \hline
\multicolumn{3}{|l|}{$A_6= \hbox{Im}(z^{n/2}), \ A_7= \hbox{Re}(z^{n/2+2}), \ A_8= \hbox{Im}(z^{n/2+2}), \ A_9= \hbox{Im}(z^n). $} \\ \hline
\end{tabular}
\end{center}
\caption{General forms of equivariant quadratic differential forms on the plane under closed subgroups of {\bf O}$(2)$.}
\label{tabela forma gerais}
\end{table}
We finish this paper with an example of each symmetry type given in Table \ref{tabela forma gerais}. Let us point out that some of these configurations can be realized for example as lines of curvatures or as asymptotic lines of surfaces immersed in $\mathbb{R}^3$ or $\mathbb{R}^4$, whereas some others cannot. This is an interesting issue in differential geometry that we have started to investigate in presence of symmetries. For the context without symmetry we cite \cite{hopf,livroronaldo,gutierrezsoto}.
\begin{figure}[!h]
\centering
\subfigure[$(1 +y^2 -x^2+2xy, x^2-y^2 +2xy, 1+x^2 -y^2 -2xy)$][\label{so2so2}]{
\includegraphics[scale=0.5]{exemplosodsod.png}}
\qquad
\subfigure[$(y^2-x^2, 2xy, x^2-y^2)$][\label{o2o2}]{
\includegraphics[scale=0.5]{exemploodod.png}}
\qquad
\subfigure[$(2xy, x^2-y^2, -2xy )$][\label{o2so2}]{
\includegraphics[scale=0.5]{exemploodsod.png}}
\caption{Configurations with symmetry (a) $\mathbf{SO}(2)$, (b) $\mathbf{O}(2)$ and (c) $\mathbf{O}(2)[\mathbf{SO}(2)]$.}
\label{fig01}
\end{figure}
For $\mathbf{SO}(2)$, we choose $p_1 \equiv p_2 \equiv p_3\equiv1$ in Table~\ref{tabela forma gerais}, so that the differential form is
$$(1 +y^2 -x^2+2xy, x^2-y^2 +2xy, 1+x^2 -y^2 -2xy).$$
The homomorphism $\lambda$ is trivial and the discriminant function is $\mathbf{O}(2)$-invariant given by $$\delta(x,y)=2(x^2 +y^2)^2 -1.$$
This is illustrated in Fig.~\ref{so2so2}.
For $\mathbf{O}(2)$, we choose $p_1\equiv0, p_2\equiv1$, so the differential form is
$$(y^2-x^2, 2xy, x^2-y^2).$$
The homomorphism $\lambda$ is such that $\ker \lambda = \mathbf{SO}(2)$ and the discriminant function is $\mathbf{O}(2)$-invariant given by $$\delta(x,y)=(x^2 +y^2)^2.$$
This is illustrated in Fig.~\ref{o2o2}.
The configuration in Fig.~\ref{o2so2} is $\mathbf{O}(2)[\mathbf{SO}(2)]$-symmetric, whose quadratic form has been chosen by taking $p\equiv1$ in Table~1, that is, $$(2xy, x^2-y^2, -2xy ).$$ The homomorphism $\lambda$ is trivial and the discriminant function is the $\mathbf{O}(2)$-invariant given by $$\delta(x,y)=(x^2 +y^2)^2. $$
\begin{figure}[!h]
\centering
\subfigure[()][\label{znzn}]{
\includegraphics[scale=0.5]{exemploznzn.png}}
\qquad
\subfigure[$()$][\label{znzn2}]{
\includegraphics[scale=0.5]{exemploznznd.png}}
\caption{Configurations with symmetry group given by (a) $\mathbf{Z}_5$ and (b) $\mathbf{Z}_4[\mathbf{Z}_2]$ .}
\label{fig02}
\end{figure}
We now consider $\mathbf{Z}_5$ taking $p_1\equiv p_2 \equiv p_5 \equiv1$ and $p_3\equiv p_4 \equiv p_5 \equiv 0$ in Table~\ref{tabela forma gerais}, so that the differential form is
$$(1 +y^2 -x^2 -3x^2y +y^3, 2xy + x^3 -3xy^2, 1 + x^2 -y^2 +3x^2y -y^3).$$
The homomorphism $\lambda$ is necessarily trivial. The discriminant function is the $\mathbf{Z}_5$-invariant given by
$$\delta(x,y)=(x^2 +y^2)^3 +10x^4y -20x^2y^3+2y^5 +(x^2 +y^2)^2 -1.$$
The picture for this case is shown in Fig. \ref{znzn}. The star shape of the discriminant set is in fact $\mathbf{Z}_5$-symmetric without reflectional symmetries, as it is easily checked by direct calculation.
Fig.~\ref{znzn2} is a $\mathbf{Z}_{4}[\mathbf{Z}_{2}]$ case, considering $p_1\equiv p_2\equiv p_4 \equiv p_5\equiv 1$ and $p_3 \equiv p_6 \equiv0 $ in Table~\ref{tabela forma gerais}, so that the differential form is
$$( -x^4 +6x^2y^2 -y^4+4x^3y -4xy^3, x^4 -6x^2y^2 +y^4+4x^3y -4xy^3, x^4 -6x^2y^2 +y^4-4x^3y +4xy^3).$$
The homomorphism $\lambda $ must be such that $\ker \lambda = \mathbf{Z}_2.$ The discriminant set is just the origin, given as the zero set of (the $\mathbf{O}(2)$-invariant)
$$\delta(x,y)=2(x^2 +y^2)^4.$$
\begin{figure}[!h]
\centering
\subfigure[()][\label{dndn}]{
\includegraphics[scale=0.5]{exemplodndn.png}}
\qquad
\subfigure[$(y^2-x^2, 2xy, x^2-y^2)$][\label{dnzn}]{
\includegraphics[scale=0.5]{exemplodnzn.png}}
\qquad
\subfigure[$(2xy, x^2-y^2, -2xy )$][\label{dndnd}]{
\includegraphics[scale=0.5]{exemplodndnd.png}}
\caption{Configurations with symmetry groups $\mathbf{D}_5$, $\mathbf{D}_6[\mathbf{Z}_6]$ and $\mathbf{D}_6[\mathbf{D}_3(\kappa_x)]$.}
\label{fig03}
\end{figure}
For $\mathbf{D}_5$ in Table~1, we take $ p_1\equiv p_2 \equiv p_3 \equiv1$, so that the differential form is
$$(1 +y^2 -x^2 -x^3 +3xy^2, 2xy -3x^2y +y^3, 1 -y^2 +x^2 +x^3 -3xy^2).$$
In this case, $\ker \lambda= \mathbf{Z}_5 $ and the discriminant function is the $\mathbf{D}_5$-invariant given by
$$\delta(x,y)=(x^2 +y^2)^3 +2x^5 -20x^3y^2+10xy^4 +(x^2 +y^2)^2 -1.$$
This is illustrated in Fig.~\ref{dndn}.
We now consider $\mathbf{D}_6[\mathbf{Z}_6]$ choosing $ p_1\equiv p_2 \equiv 1$ and $p_3 \equiv 0$ in Table~\ref{tabela forma gerais}, so that the form is
$$( 2xy -4x^3y +4xy^3, x^2-y^2 +x^4-6x^2y^2 +y^4, -2xy +4x^3y -4xy^3).$$
In this case $\lambda $ is trivial and the discriminant function is $\mathbf{D}_6$-invariant and given by
$$\delta(x,y)=(x^2 +y^2)^4+2x^6-30x^4y^2+30x^2y^4 +2y^6 + (x^2 +y^2)^2.$$
The picture is given in Fig.~\ref{dnzn}.
We now turn to $\mathbf{D}_6[\mathbf{D}_{3}(\kappa_x)]$ taking $ p_1\equiv1$ and $p_2 \equiv p_3 \equiv 0$ in Table~1, so that the form is
$$( -x, -y, x).$$
In this case, $\ker \lambda = \mathbf{D}_3(\kappa_y)$ and the discriminant set is the origin, given by the zero set of
$$\delta(x,y)=x^2 + y^2.$$
The picture is given in Fig.~\ref{dndnd}.
\begin{figure}[!h]
\centering
\subfigure[$()$][\label{kxkx}]{
\includegraphics[scale=0.5]{exemplozdzd.png}}
\qquad
\subfigure[$(2xy, x^2-y^2, -2xy )$][\label{kxki}]{
\includegraphics[scale=0.5]{exemplozdzi.png}}
\caption{Configurations with symmetry groups $\mathbf{Z}_2$ and $\mathbf{Z}_2[{\bf 1}]$.}
\label{fig04}
\end{figure}
Consider now $\mathbf{Z}_2$ for $p_1 \equiv p_2 \equiv p_3 \equiv 1$ from Table~1, so that the form is
$$( 1, y, 1).$$
We have $\ker \lambda = {\bf 1}$ and the discriminant function is $\mathbf{Z}_2$-invariant and given by
$$\delta(x,y)=y^2-1.$$
See the illustration of this case in Fig.~\ref{kxkx}.
Fig.~\ref{kxki} is a $\mathbf{Z}_2[{\bf 1}]$ case, for which we have chosen $p_1 \equiv p_2 \equiv 1$ and $p_3\equiv-1$ in Table \ref{tabela forma gerais} , so that the form is
$$( y, 1, -y).$$
The homomorphism $\lambda $ is trivial and the discriminant function is $\mathbf{Z}_2$-invariant and given by $$\delta(x,y)=y^2 +1.$$
\begin{figure}[!h]
\centering
\subfigure[()][\label{z2z2z2z2}]{
\includegraphics[scale=0.5]{exemplozdzizdzi.png}}
\qquad
\subfigure[$()$][\label{z2z2zi}]{
\includegraphics[scale=0.5]{exemplozdzizi}}
\qquad
\subfigure[$()$][\label{z2z2z2k}]{
\includegraphics[scale=0.5]{exemplozdzizd.png}}
\caption{ Configurations with symmetry groups $\mathbf{Z}_2 \times \mathbf{Z}_2, \mathbf{Z}_2 \times \mathbf{Z}_2[\mathbf{Z}_2(-I)]$ and $\mathbf{Z}_2 \times \mathbf{Z}_2 [\mathbf{Z}_2(\kappa_x)]$.}
\label{fig05}
\end{figure}
For $\mathbf{Z}_2 \times \mathbf{Z}_2$ in Table \ref{tabela forma gerais}, we take $p_1 \equiv p_2 \equiv 1$ and $p_3\equiv -1$, so that the differential form is $$( -x, -y, x).$$ In this case $\ker \lambda = \mathbf{Z}_2$ and the discriminant function is the $\mathbf{Z}_2 \times \mathbf{Z}_2$-invariant given by $$\delta(x,y)=x^2y^2+1.$$ This is illustrated in Fig. \ref{z2z2z2z2}.
We now consider $\mathbf{Z}_2\times \mathbf{Z}_2[\mathbf{Z}_2(-I)]$ choosing $ p1\equiv p_2 \equiv 1$ and $p_3 \equiv-1$ in Table \ref{tabela forma gerais}, so that the form is $$( xy, 1, -xy).$$ In this case $\lambda $ is trivial and the discriminant function is $\mathbf{Z}_2\times \mathbf{Z}_2$-invariant and given by $$\delta(x,y)=x^2y^2+1.$$ The picture is given in Fig. \ref{z2z2zi}.
Finally, consider $\mathbf{Z}_2\times \mathbf{Z}_2[\mathbf{Z}_2(\kappa_x)]$ taking $p_1 \equiv1, p_2 y^2$ and $p_3 \equiv -1$ in Table \ref{tabela forma gerais}, so that differential form is $$( x, y^3, -x).$$ In this case $\ker \lambda = \mathbf{Z}_2(\kappa_y)$ and the discriminant function is given by $$\delta(x,y)=x^2 +y^6.$$ See the illustration of this case in Fig. \ref{z2z2z2k}. \\
{\bf Acknowledgements}: Research of P. T. was supported by CAPES Grant 8474758/D.
|
\section{Introduction}\
{\bf (a) Subject and Goal.}
The van der Pol equation (in short, vdP, cf. \eqref{vdP}), originally introduced in \citep{VdP} to study stable oscillations in electrical circuits, is very often considered as a starting point of applied nonlinear dynamics. An important feature of equation \eqref{vdP} is that it respects the antipodal symmetry. Different generalizations of vdP, which break this symmetry, have been considered by many authors in connection to a wide range of applied problems (in what follows, we will call these generalizations van der Pol-like equations (in short, vdPl)). Examples of particular importance include the FitzHugh-Nagumo model (see, for example, \citep{FitzHugh}) and the realistic kinetic model of the chlorite-iodide-malonic acid reaction (see, for example, \citep{Boissonade}). To be more concrete about the importance of considering vdPl systems with quadratic terms, we refer to \citep{VdPlike} and the references therein.
In real life models, vdP as well as vdPl serve as nodes in coupled networks.
Symmetries of the coupling have an impact on the symmetries of the actual dynamics.
In this paper, we will consider $N^3$ oscillators arranged in a cubical lattice, where opposite faces are identified in the same way as for a $3$-torus.
In such a configuration, two aspects of the coupling are important:
(i) which nodes impact the dynamics of a given node (which we will call {\it coupling topology}), and
(ii) how a neighboring node impacts a given node (which we will call {\it coupling structure}).
For the coupling topology, we consider the following two cases:
(i) all $6$ neighbors of a given node impact on that node's dynamics (we call such a coupling {\it bi-directional});
(ii) only $3$ neighbors of a given node impact on that node's dynamics (we call such a coupling {\it uni-directional}).
In the case of bi-directional coupling, the system respects a natural action of $\mathbb D_N\times \mathbb D_N\times \mathbb D_N$, while in the case of uni-directional coupling the symmetry generated by $\kappa$ is destroyed, hence the total symmetry group is $\mathbb Z_N \times \mathbb Z_N \times \mathbb Z_N$ (cf. \citep{3d-torus-applications} and references therein). We will distinguish between two linear coupling structures, namely, for a given node, either the $x$-variable of a neighbor or the $y$-variable of a neighbor is coupled to the $x$ variable of the specified node (compare \eqref{nvdP_2d} with \eqref{nvdPl_2d}). We will call these $x$-coupling and $y$-coupling respectively.
The {\it goal} of this paper is three-fold, namely, in the settings introduced above, we will: (i) establish the {\it occurrence} of the Hopf bifurcation, classify symmetric properties of the bifurcating branches and estimate their number; (ii) study {\it stability} of the corresponding periodic solutions, and (iii) investigate the {\it existence} of periodic solutions with prescribed period and symmetry.
\medskip
{\bf (b) Results.} Keeping in mind a wide spectrum of potential applications in natural sciences and engineering, it is worthy to study the above mentioned problems (occurrence, stability and existence) in all possible settings. The usual dilemma of keeping a balance between Scylla of completeness and Charybdis of reasonable size of the manuscript, resulted in our paper as follows:
\medskip
(i) Although, using the methods developed in this paper, the occurrence/multiplicity estimates/symmetry classification of the Hopf bifurcation can be established for any combination of bi-directional/uni-directional vdP/vdPl $x$-coupled/$y$-coupled systems,
we only treat two cases, namely, that of bi-directionally $y$-coupled vdP oscillators and uni-directionally $x$-coupled vdPl oscillators.
(ii) We provide an instability result for bi-directionally $y$-coupled vdP oscillators and stability results for uni-directionally $x$-coupled vdPl oscillators.
(iii) We establish the existence of periodic solutions with prescribed period and symmetry only in the case of bi-directionally $y$-coupled vdP oscillators.
\medskip
{\bf (c) Methods.} The methods used in this paper are based on the results rooted in both equivariant degree theory and (equivariant) singularity theory.
To be more specific:
\medskip
(i) To treat the occurrence/multiplicity estimates/symmetric classification of the Hopf bifurcation, we appeal to the abstract results presented in \citep{AED} (see also \citep{BK-chapter, Sliding}).
(ii) The stability results are obtained in the framework of the theory presented in \citep{Guckenheimer} (see also \citep{GolSteShef}).
(iii) The main ingredient to establish the existence of periodic solutions with prescribed period and symmetry is Theorem \ref{theom-abstract-Hirano-Ryb} which was presented in
\citep{AED} and \citep{BFK} (see also \citep{Hirano-Rybicki}).
\smallskip
For the representation theory background, we refer to \citep{Brocker-tomDieck,Serre}.
\medskip
{\bf (d) Overview.} After the Introduction, the paper is organized as follows. In Section \ref{onetorus_theor}, we formulate main results of the paper. Section \ref{sect:occurrence-result}
is devoted to the proof of the main occurrence/multiplicity/symmetry results (Theorems \ref{thm_vdp} and \ref{thm_vdpl}). The proof (see Subsection \ref{Proof-Theor-occurr}) is based on an abstract Theorem \ref{abs_therom} and equivariant spectral information collected in Subsections \ref{sec_decop}--\ref{iso_sec}. We believe that several algebraic observations related to the computation of maximal twisted orbit types in complexifications of tensor product representations (see Subsection \ref{iso_sec}) may be interesting in their own. Section \ref{sect-stability} contains the proof of the instability result for system \eqref{nvdP_2d} (see Theorem \ref{Dn_instability}) and stability result for system \eqref{nvdPl_2d} (see Theorem \ref{Zn_stab}). The proof follows the standard lines (see \citep{Guckenheimer}, Theorem 3.4.2, and \citep{GolSteShef}), and combines the spectral equavariant data from Subsections \ref{sec_decop}--\ref{iso_sec} with the computations of the first Lyapunov coefficient from Subsection \ref{subsec:Lyapunov-coef}. The proof of the existence result (see Theorem \ref{Existence-Dihedral}) is given in Section
\ref{sec:Existence}
where one can also find an adapted version of the abstract existence result given in \citep{AED}, Theorem 12.7, and \citep{BFK} (cf. Theorem \ref{theom-abstract-Hirano-Ryb}).
We conclude with a short Appendix where several symbols frequently used in this paper to denote some twisted groups are explained (cf. \citep{AED}).
\medskip
{\bf (e) Acknowledgements.} The first two authors acknowledge the support from National Science Foundation through grant DMS-413223. The first author is grateful for the support of the Gelbart Research Institute for mathematical sciences at Bar Ilan University. The third author acknowledges a postdoctoral BITDEFENDER fellowship from
the Institute of Mathematics Simion Stoilow of the Romanian Academy, Contract of Sponsorship No. 262/2016.
\section{Main Results}\label{onetorus_theor}
In this paper, we are interested in networks of identical vdP oscillators \begin{equation}\label{vdP}
\begin{array}{ll}
\dot{x}&=\nu(ax-x^3)-y\\
\dot{y}&=bx
\end{array}
\end{equation}
and vdP-like oscillators
\begin{equation}\label{vdPl}
\begin{array}{ll}
\dot{x}&=-y-x^3-x^2+ax\\
\dot{y}&=bx
\end{array}
\end{equation}
coupled in the symmetric configuration of a three-dimensional torus. To be more precise, we consider $N^3$ oscillators, where $N$ is an {\it odd} number, with both bi-directional coupling
\begin{equation}\label{nvdP_2d}
\left\{
\begin{aligned}
\dot{x}_{(\alpha,\beta,\gamma)}= &
\nu(ax_{(\alpha,\beta,\gamma)}-x_{(\alpha,\beta,\gamma)}^3) -y_{(\alpha,\beta,\gamma)}
\\
&+\delta \left(2y_{(\alpha,\beta,\gamma)}-y_{(\alpha+1,\beta,\gamma)}-y_{(\alpha-1,\beta,\gamma)}\right)\\& +\zeta \left(2y_{(\alpha,\beta,\gamma)}-y_{(\alpha,\beta+1,\gamma)}-y_{(\alpha,\beta-1,\gamma)}\right)
\\
&+\varepsilon \left(2y_{(\alpha,\beta,\gamma)}-y_{(\alpha,\beta,\gamma+1)}y_{(\alpha,\beta,\gamma-1)}\right)
\\
\dot{y}_{(\alpha,\beta,\gamma)}=& \; bx_{(\alpha,\beta,\gamma)}
\end{aligned}
\right.
\end{equation}
and uni-directional coupling
\begin{equation}\label{nvdPl_2d}
\left\{
\begin{aligned}
\dot{x}_{(\alpha,\beta,\gamma)}= &
-y_{(\alpha,\beta,\gamma)}-x_{(\alpha,\beta,\gamma)}^3-x_{(\alpha,\beta,\gamma)}^2+ax_{(\alpha,\beta,\gamma)}
\\
&+\delta \left(x_{(\alpha,\beta,\gamma)}-x_{(\alpha+1,\beta,\gamma)}\right)+\zeta \left(x_{(\alpha,\beta,\gamma)}-x_{(\alpha,\beta+1,\gamma)}\right)
\\
&+\varepsilon \left(x_{(\alpha,\beta,\gamma)}-x_{(\alpha,\beta,\gamma+1)}\right)
\\
\dot{y}_{(\alpha,\beta,\gamma)}=& \; bx_{(\alpha,\beta,\gamma)}
\end{aligned}
\right.
\end{equation}
Here $x,y \in \mathbb R^{N^3}$ and their entries are indexed by the triple $(\alpha,\beta,\gamma)$ where $\alpha,\beta,\gamma \in \{1,\cdots,N\}$, $\delta, \zeta,\varepsilon\in \mathbb R$ and $\nu,b>0$.
\begin{remark}
{\rm To avoid distinctions occurring due to the parity of $N$, we will only consider the case when $N$ is odd.}
\end{remark}
\begin{definition}\label{spa_temp}{\rm
We will say that a periodic function $x:\mathbb R \to U$ with period $T$ has a (spatio-temporal) symmetry $H< G\times S^1$, if for every $(g,e^{i\theta})\in H$ and for every $t$,
$$
g\cdot x(t+\theta T/2\pi)= x(t).
$$
}
\end{definition}
\begin{theorem}\label{thm_vdp}
For each fixed ${\bf t}= (t_1,t_2,t_3)$, where $t_1,t_2,t_3 \in \{1,\cdots,n\}$, put
\begin{equation}\label{Kt_def1}
K_{\bf t}= 1 + 2\delta(1-\cos(2\pi t_1/N))+2\zeta(1-\cos(2\pi t_2/N))+2\varepsilon(1-\cos(2\pi t_3/N)).
\end{equation}
If $K_{\bf t} > 0$, then system \eqref{nvdP_2d} undergoes Hopf bifurcation as $a$ passes $0$.
Furthermore, for each $(H^\varphi) \in \mathcal S({\bf t})$, there exist $\frac{8N^3}{|H_1\times H_2\times H_3|}$ branches of bifurcating non-constant periodic solutions of \eqref{nvdP_2d} with limit frequency $\omega_{\bf t}= \sqrt{bK_{\bf t}}$ and minimal symmetry $(H^\varphi) $ (here $\mathcal S (\bf t)$ is the set of spatio-temporal symmetries associated to {\bf t} which is described in Section \ref{iso_sec}, formula \eqref{desc_iso}).
\end{theorem}
\begin{theorem}\label{thm_vdpl}
For each fixed ${\bf t}= (t_1,t_2,t_3)$, where $t_1,t_2,t_3 \in \{1,\cdots,N\}$, system \eqref{nvdPl_2d} undergoes Hopf bifurcation as $a$ passes
\begin{equation}\label{eq:bifur-points}
a^z_{\bf t}= \delta(1-\cos(2t_1\pi/N))+\zeta(1-\cos(2t_2\pi/N))+\varepsilon(1-\cos(2t_3\pi/N)).
\end{equation}
Furthermore, there exist a branch of bifurcating non-constant periodic solutions of \eqref{nvdPl_2d} with symmetry $(\mathbb Z_n\times \mathbb Z_n \times \mathbb Z_n)^{(t_1, t_2, t_3)}$ and limit period
\begin{align*}
P^1_{\bf t}&=\Big|\frac{4\pi}{H+\sqrt{H^2+4b}}\Big|,
\end{align*}
and a branch with the same symmetry and limit period
\begin{align*}
P^2_{\bf t}&=\Big|\frac{4\pi}{H-\sqrt{H^2+4b}}\Big|,
\end{align*}
where $ H= \delta\sin((2t_1\pi/N))+\zeta\sin(2t_2\pi/N))+\varepsilon(\sin(2t_3\pi/N))$ and the symbol $(\mathbb Z_n \times \mathbb Z_n \times \mathbb Z_n)^{(t_1, t_2, t_3)}$ is described in Appendix.
\end{theorem}
\begin{remark}
{\rm We do not guarantee that $P_{\bf t}^1$ or $P_{\bf t}^2$ are minimal periods. For this reason, it is possible for a single solution to have both $P_{\bf t}^1$ and $P_{\bf t}^2$ as a period, in which case we only guarantee the existence of a single branch. This can occur in the case when $P_{\bf t}^1/P_{\bf t}^2\in\mathbb Q$.}
\end{remark}
\begin{theorem}\label{Dn_instability}
Put $\theta_N:= \frac{(N-1) \pi}{N}.$ Suppose $k_1\delta+k_2\zeta+k_3\varepsilon$ is less than
$$
\frac{1}{2(\cos(\theta_N)-1)}
$$
for some $k_1,k_2,k_3 \in \{0,1\}$. Then, all branches of bifurcating non-constant periodic solutions of \eqref{nvdP_2d} guaranteed by Theorem \ref{thm_vdp} are unstable.
\end{theorem}
\begin{theorem}\label{Zn_stab}
For any fixed $\delta,\zeta$ and $\varepsilon$, the equilibrium of system \eqref{nvdPl_2d} is stable for $a<a^*$ and unstable for $a>a^*$, where $a^*$ is described in
Table~\ref{tabelaBifurca}. Furthermore, the \textbf{\textit {fully synchronized}} branch of system \eqref{nvdPl_2d}, born at $a=0$, is stable if and only if $\delta,\zeta$ and $\varepsilon$ are negative.
\end{theorem}
\begin{table}[h]
\begin{center}
\begin{tabular}{c|c|c|c|c}
\hline
sign$(\delta)$&sign$(\zeta)$& sign$(\epsilon)$& $a_*$ & ${\bf t} = (t_1,t_2,t_3) $ \\
\hline
-&-&-&0&$(0,0,0)$\\[0.25em]
+&-&-&$\delta\left(1-\cos\left(\theta_N\right)\right)$&$\left(\frac{N-1}{2},0,0\right)$\\[0.25em]
-&+&-&$\zeta\left(1-\cos\left(\theta_N\right)\right)$&$\left(0,\frac{N-1}{2},0\right)$\\[0.25em]
-&-&+&$\epsilon\left(1-\cos\left(\theta_N\right)\right)$&$\left(0,0,\frac{N-1}{2}\right)$\\[0.25em]
+&+&-&$(\delta+\zeta)\left(1-\left(\cos\theta_N\right)\right)$&$\left(\frac{N-1}{2},\frac{N-1}{2},0\right)$\\[0.25em]
+&-&+&$(\delta+\epsilon)\left(1-\left(\cos\theta_N\right)\right)$&$\left(\frac{N-1}{2},0,\frac{N-1}{2}\right)$\\[0.25em]
-&+&+&$(\zeta+\epsilon)\left(1-\left(\cos\theta_N\right)\right)$&$\left(0,\frac{N-1}{2},\frac{N-1}{2}\right)$\\[0.25em]
+&+&+&$(\delta+\zeta+\epsilon)\left(1-\left(\cos\theta_N\right)\right)$&$\left(\frac{N-1}{2},\frac{N-1}{2},\frac{N-1}{2}\right)$\\[0.25em]
\hline
\end{tabular}\end{center}
\caption{Details of the Hopf bifurcations from a stable equilibrium(here $\theta_N= \frac{(N-1)\pi}{N}$)}
\label{tabelaBifurca}
\end{table}
\begin{theorem}\label{Existence-Dihedral}
Assume that $K_{\bf t} \neq 0$ for any ${\bf t}=(t_1,t_2,t_3)$ (cf. \eqref{Kt_def1}). Choose $p \not\in\big\{\frac{2\pi(2k-1)}{\sqrt{bd}}\; : \; k \in \mathbb Z, \; d>0,\; d= K_{\bf t} \; \text{for some} \; {\bf t}\big\}$. Then, for each {\bf t} with $K_{\bf t} > (\frac{2\pi}{p})^2$, there exists a value of $\nu$ such that for each $(H^{\varphi}) \in \mathcal S({\bf t})$, system \eqref{nvdP_2d} admits $\frac{8N^3}{|H_1\times H_2\times H_3|}$ $p$-periodic solutions with minimal symmetry $(H^{\varphi})$ (here $\mathcal S (\bf t)$ is the set of symmetries associated to {\bf t} which is described in Section \ref{iso_sec}, formula \eqref{desc_iso}).
\end{theorem}
\section{Equivariant spectral data and first Lyapunov coefficient}
\subsection{Isotypical decomposition of the phase space}\label{sec_decop}
Although \eqref{nvdP_2d} and \eqref{nvdPl_2d} have different symmetry groups, they have the same phase space ($\mathbb R^{N^3}$) as a geometric set.
Since it is usually unambiguous, we will use the same notation for the representations of $G_1:= \mathbb Z_N \times \mathbb Z_N \times \mathbb Z_N$ and $G_2:= \mathbb D_N \times \mathbb D_N \times \mathbb D_N$. Put $V := \mathbb R^{N^3}$ and denote by $W := V \oplus V$ the phase space of systems \eqref{nvdP_2d} and \eqref{nvdPl_2d}.
To describe spatial symmetries of system \eqref{nvdPl_2d}, we will consider $G_1= \mathbb Z_N \times \mathbb Z_N \times \mathbb Z_N$ as a subgroup of $ S^1\times S^1\times S^1$ and define the $G_1$-action on $V$ by specifying how each of its generators acts, namely:
\begin{align*}
((e^{\frac{2\pi i}{N}},1,1)\cdot x)_{(\alpha,\beta,\gamma)}= x_{(\alpha+1,\beta,\gamma)}\\
((1,e^{\frac{2\pi i}{N}},1)\cdot x)_{(\alpha,\beta,\gamma)}= x_{(\alpha,\beta+1,\gamma)}\\
((1,1,e^{\frac{2\pi i}{N}})\cdot x)_{(\alpha,\beta,\gamma)}= x_{(\alpha,\beta,\gamma+1)}.
\end{align*}
Here $+$ is taken modulo $N$. By direct verification, the right-hand side of system \eqref{nvdPl_2d} is $G_1$-equivariant.
To extend this action to a $G_2$-action, we need to specify how the remaining generators act, namely:
\begin{align*}
((\kappa,1,1)\cdot x)_{(\alpha,\beta,\gamma)}= x_{(-\alpha,\beta,\gamma)}\\
((1,\kappa,1)\cdot x)_{(\alpha,\beta,\gamma)}= x_{(\alpha,-\beta,\gamma)}\\
((1,1,\kappa)\cdot x)_{(\alpha,\beta,\gamma)}= x_{(\alpha,\beta,-\gamma)}
\end{align*}
where $-$ is again taken modulo $N$. It is clear that the right-hand side of \eqref{nvdP_2d} is $G_2$-equivariant.
To describe the isotypical decomposition of $V$ as a $G_1$-space, we need to classify all (real) irreducible $G_1$-representations. For each $0\leq t_1,t_2,t_3\leq N-1$, put ${\bf t}:= (t_1,t_2,t_3)$ and denote by $\mathcal V^z_{\bf t}$ an irreducible representation of $G_1$ associated with ${\bf t}$. We have
\begin{equation}\label{Z_triv_rep}
\mathcal V^z_{0,0,0}= \mathbb R
\end{equation}
is the trivial real $G_1$-representation. For $(t_1,t_2,t_3)\neq (0,0,0)$, put
\begin{equation}\label{irr_def}
\mathcal V^z_{\bf t}=\mathbb R^2
\end{equation}
and define the $G_1$-action as follows:
$$
\Big(e^{\frac{2k_1\pi i}{N}},e^{\frac{2k_2\pi i}{N}},e^{\frac{2k_3\pi i}{N}}\Big)\cdot \begin{bmatrix}
x \\ y
\end{bmatrix} = A \begin{bmatrix} x \\ y\end{bmatrix},
$$
where
$$
A:= \begin{bmatrix}
\cos\big((k_1t_1+k_2t_2+k_3t_3)\left(\frac{2 \pi}{N}\right)\big) & \sin\big((k_1t_1+k_2t_2+k_3t_3)\left(\frac{2 \pi}{N}\right)\big)\\
-\sin\big((k_1t_1+k_2t_2+k_3t_3)\left(\frac{2 \pi}{N}\right)\big) & \cos\big((k_1t_1+k_2t_2+k_3t_3)\left(\frac{2 \pi}{N}\right)\big)
\end{bmatrix}
$$
By direct verification, $\mathcal V^z_{\bf t}$ is $G_1$-equivalent to $\mathcal V^z_{-\bf t}$ where $-$ is taken modulo $N$. Hence, there is one one-dimensional trivial representation and $(N^3-1)/2$ non-trivial two-dimensional non-equivalent $G_1$-representations.
For each fixed ${\bf t}$, we define vectors $x^1_{\bf t}$ and $x^2_{\bf t}$ by specifying them component-wisely as follows:
$$
\big(x^1_{\bf t}\big)_{\alpha,\beta,\gamma}= \cos(\alpha t_1+\beta t_2 + \gamma t_3)
$$
$$
\big(x^2_{\bf t}\big)_{\alpha,\beta,\gamma}= \sin(\alpha t_1+\beta t_2 + \gamma t_3).
$$
Define a family of subspaces of $V^z_{\bf t}$ by
$$V^z_{\bf t}= \text{span} (x^1_{\bf t}, x^2_{\bf t}).$$
Notice that $V^z_{\bf t}$ is a $G_1$-irreducible component of $V$ and is $G_1$-equivalent to $\mathcal V^z_{\bf t}$ (cf. \eqref{Z_triv_rep}, \eqref{irr_def}).
\begin{remark}\label{Z_prim_decomp}
{\rm $V$ admits a primary $G_1$-decomposition which includes every $G_1$-irreducible representation.}
\end{remark}
Let us denote by $\mathcal U_0^d$ the trivial one-dimensional real $\mathbb D_N$-representation and by $\mathcal U_t^d$ the natural 2-dimensional real $\mathbb D_N$-representation, where the action is defined by
$$
e^{\frac{2k\pi i}{N}}\cdot \begin{bmatrix}
x\\y\end{bmatrix}= \begin{bmatrix}
\cos\big(kt\left(\frac{2 \pi}{N}\right)\big) & \sin\big(kt\left(\frac{2 \pi}{N}\right)\big)\\
-\sin\big(kt\left(\frac{2 \pi}{N}\right)\big) & \cos\big(kt\left(\frac{2 \pi}{N}\right)\big)
\end{bmatrix} \begin{bmatrix}
x\\y
\end{bmatrix}$$
and
$$
\kappa\cdot \begin{bmatrix}
x\\y
\end{bmatrix} = \begin{bmatrix}
y\\x
\end{bmatrix}
$$
Denote
$$\mathcal V^d_{\bf t}:=
\mathcal U_{t_1}^d \otimes \mathcal U_{t_2}^d \otimes \mathcal U_{t_3}^d. $$
Since $\mathcal U_{t}^d$ is of real type for any $t= 0 ,\cdots, N$, it is easy to see that $\mathcal V_{\bf t}^d$ is a real {\it irreducible} $G_2=(\mathbb D_N \times \mathbb D_N\times \mathbb D_N)$-representation. Furthermore, the dimension of $\mathcal V_{\bf t}^d$ is either $1,2,4$ or $8$ depending on how many non-zero components {\bf t} has.
Put ${\bf t}^\dagger:=(t_1,-t_2,t_3)$, ${\bf t}^\# =(t_1,t_2,-t_3)$ and ${\bf t}^* :=(t_1,t_2,-t_3)$. Put
$$
V_{\bf t}^d := \text{span}(x^1_{\bf t},x^2_{\bf t},x^1_{\bf t^\dagger},x^2_{\bf t^\dagger},x^1_{\bf t^\#},x^2_{\bf t^\#},x^1_{\bf t^*},x^2_{\bf t^*})
$$
By simple but lengthy computations, one can easily show that $V_{\bf t}^d$ is $G_2$-invariant and equivalent to $\mathcal V^d_{\bf t}$.
\begin{remark}\label{prim_decom}
{\rm $V$ admits a primary $G_2$-decomposition, however, unlike its decomposition as a $G_1$-space (cf. Remark \ref{Z_prim_decomp}), the decomposition as a $G_2$-space does not include every $G_2$-irreducible representation.}
\end{remark}
\subsection{Equivariant spectral decomposition}\label{sec_eign}
The linearization of system \eqref{nvdP_2d} at the origin restricted to the isotypical component $V^d_{\bf t}\oplus V^d_{\bf t}$ is given by $A^d_{\bf t}(a):V^d_{\bf t}\oplus V^d_{\bf t} \to V^d_{\bf t}\oplus V^d_{\bf t}$, where
\begin{equation}\label{Dn_blocks}
A^d_{\bf t}(a)= \left(
\begin{array}{cc}
\nu a & -K_{\bf t}\\
b & 0
\end{array}
\right) \otimes \text{Id}_{\bf t},
\end{equation}
\begin{equation}\label{Kt_def}
K_{\bf t}= 1+2\delta(1-\cos(2\pi t_1/N))+2\zeta(1-\cos(2\pi t_2/N))+2\varepsilon(1-\cos(2\pi t_3/N))
\end{equation} and $\text{Id}_{\bf t}$ is the matrix of the identity operator on $V^d_{\bf t}$ (here $\otimes$ stands for the Kroneker product of matrices). It is clear that the eigenvalues of $A^d_{\bf t}(a)$ are given by
\begin{equation}\label{eig_d_syst}
\lambda_{\bf t}^d(a)= \frac{\nu a\pm \sqrt{(\nu a)^2-4bK_{\bf t}}}{2}.
\end{equation}
\bigskip
The linearization of system \eqref{nvdPl_2d} at the origin restricted to the isotypical component $V^z_{\bf t}\oplus V^z_{\bf t}$ is given by $A^z_{\bf t}(a):V^z_{\bf t}\oplus V^z_{\bf t} \to V^z_{\bf t}\oplus V^z_{\bf t}$, where
$$A^z_{\bf t}(a)= \left(
\begin{array}{cccc}
H_{\bf t}(a) & G_{\bf t} & -1 & 0\\
-G_{\bf t} & H_{\bf t}(a)& 0 & -1\\
b & 0 & 0 & 0 \\
0 & b & 0 & 0
\end{array}
\right),$$
$$
H_{\bf t}(a) = a-\delta(1-\cos(2\pi t_1/N))-\zeta(1-\cos(2\pi t_2/N))-\varepsilon(1-\cos(2\pi t_3/N))$$ and
\begin{equation}\label{H_t}
G_{\bf t}=\delta(\sin(2\pi t_1/N))+\zeta(\sin(2\pi t_2/N))+\varepsilon(\sin(2\pi t_3/N)).
\end{equation} To compute the eigenvalues of $A^z_{\bf t}(a)$, we should notice that $A^z_{\bf t}(a) = {}^\mathbb RL_{\bf t}(a)$, where $L_{\bf t}(a)$ is the complex matrix
\begin{equation}
\label{matL_t}
L_{\bf t}(a)= \left(
\begin{array}{cc}
H_{\bf t}(a)-iG_{\bf t} & -1\\
b & 0
\end{array}
\right)
\end{equation}
(here the symbol ${}^\mathbb RL_{\bf t}$ stands for the realification of $L_{\bf t}$). It is well-known that $\sigma ({}^\mathbb RL_{\bf t}(a)) = \sigma (L_{\bf t}(a)) \cup \overline{\sigma (L_{\bf t}(a))}.$ It is easy to see that the eigenvalues of $L_{\bf t}(a)$ are given by
\begin{equation}\label{eig_z_sys}
\lambda^{z}_{\bf t}(a)= \frac{H_{\bf t}(a)-iG_{\bf t} \pm \sqrt{(H_{\bf t}(a)-iG_{\bf t})^2-4b}}{2}
\end{equation}
\begin{remark}{\rm For a generic choice of parameters $\gamma, \zeta, \varepsilon$, all eigenvalues $\lambda^{z}_{\bf t}(a)$ are distinct and in the case $\lambda^{z}_{\bf t}(a)$ is purely imaginary, not in resonance.}
\end{remark}
\subsection{Isotropies in $(V^z_{\bf t})^c$ and $(V^d_{\bf t})^c$}\label{iso_sec}
{\bf Note}:
For an explanation of all symbols used in this section, see the Appendix.
{(a)} Since $(V^z_{\bf t})$ is a real irreducible $G_1$-representation of complex type, $(V^z_{\bf t})^c$ decomposes into the direct sum of two non-equivalent conjugate irreducible $(G_1 \times S^1)$-representations:
$V_1\oplus \overline V_1$. Since $V_1$ is one-dimensional, it has only two orbit types, namely $(G_1\times S^1)$ and $(\mathbb Z_n\times \mathbb Z_n \times \mathbb Z_n)^{(t_1, t_2, t_3)}$.
Similarly, $\overline V_1$ has only $G_1\times S^1$ and $(\mathbb Z_n\times \mathbb Z_n \times \mathbb Z_n)^{(-t_1, -t_2, -t_3)}$.
\bigskip
{(b)} Let us recall the following
\begin{definition}\label{def-max-orbit}
{\rm We will call an orbit type $H < G \times S^1$ maximal in $\mathcal U^c_j$ if for any $\widetilde{H} \neq G \times S^1$ which is an orbit type in $\mathcal U^c_j$, one has $\widetilde{H}<H$.}
\end{definition}
To restrict candidates for maximal isotropies in $(V^d_{\bf t})^c$, we use the following simple observation.
\begin{lemma}\label{lem:factor-product}
Let $\mathcal G_1$ (resp. $\mathcal G_2$) be a finite group and let $U_1$ (resp. $U_2$) be a unitary $\mathcal G_1 \times S^1$-respresentation (resp. $G_2 \times S^1$-respresentation), where
$S^1$ acts on $U_1$ and $U_2$ by complex multiplication.
\medskip
(i) If $H_1^{\varphi_1}$ is a twisted isotropy in $U_1$ and $H_2^{\varphi_2}$ is a twisted isotropy in $U_2$, then $(H_1 \times H_2)^{(\varphi_1,\varphi_2)}$ is an isotropy in $U_1 \otimes U_2$.
\medskip
(ii) If $(H_1 \leftidx{}{_{K_1}}{\times}_{K_2} H_2)^{(\varphi_1,\varphi_2)}$ is an isotropy in $U_1 \otimes U_2$, then for some $v_1 \in U_1$ and $v_2 \in U_2$, one has $\mathcal G_{v_1} \geq K_1^{\varphi_1}$ and
$\mathcal G_{v_2} \geq K_2^{\varphi_2}$ (cf. \eqref{amal_not}).
\end{lemma}
\begin{proof}
(i) Take $v_1 \in (\mathcal U_1)_{H_1^{\varphi_1}}$ and $v_2 \in (\mathcal U_2)_{H_2^{\varphi_2}}$.
Then, for any
$$
g:=(h_1,h_2,\varphi_1(h_1)\varphi_2(h_2)) \in (H_1 \times H_2)^{(\varphi_1,\varphi_2)},
$$
one has $g(v_1 \otimes v_2) = v_1 \otimes v_2$, i.e.,
$\mathcal G_{v_1 \otimes v_2} \geq (H_1 \times H_2)^{(\varphi_1,\varphi_2)}$. On the other hand, if $\mathcal G_{v_1 \otimes v_2} \ni
(h_1,h_2,e^{i \theta})$, then $\Big( e^{i \theta} T_{h_1} \otimes T_{h_1} \Big) v_1 \otimes v_2 = v_1 \otimes v_2$, which implies that for some $\hat{\theta}$ and $\tilde{\theta}$ with
$\theta = \hat{\theta} + \tilde{\theta}$ one has $e^{i\hat{\theta}} T_{h_1} v _1 = v_1$ and $e^{i\tilde{\theta}} T_{h_2} v _2 = v_2$. Hence,
$(h_1, e^{i\hat{\theta}}) \in \mathcal G_{v_1} = H_1^{\varphi_1}$, $(h_1, e^{i\tilde{\theta}}) \in \mathcal G_{v_2} = H_2^{\varphi_2}$ and
$(h_1,h_2,e^{i\theta}) = (h_1,h_2,e^{i\hat{\theta}} \cdot e^{i\tilde{\theta}}) \in (H_1 \times H_2)^{(\varphi_1,\varphi_2)}$
i.e., $\mathcal G_{v_1\otimes v_2}\leq (H_1\times H_2)^{(\varphi_1,\varphi_2)}$.
\medskip
(ii) Take $v \in (U_1 \times U_2)_{(H_1 \times H_2)^{(\varphi_1, \varphi_2)}}$ and decompose it as
$$
v = \sum_{i=1}^n v_i \otimes e_i.
$$
For any $g := (k_1,1_{H_1},\varphi_1(k_1)) \in (H_1 \times H_2)^{(\varphi_1, \varphi_2)}$, one has
$$
g \cdot v = \sum_{i=1}^n e^{\varphi_i(k_i)} T_{h_1} v_i \otimes e_i = \sum_{i=1}^n v_i \otimes e_i,
$$
hence, $v_i \in U_1^{K_1^{\varphi_1}}$
for any $i = 1,...,n$, i.e.,
$(\mathcal G_1)_{v_i} \leq H_1^{\varphi_1}$. A similar argument shows that $(\mathcal G_2)_{v_i} \leq H_2^{\varphi_2}$.
\end{proof}
A direct consequence of Lemma \ref{lem:factor-product} is the following:
\begin{corollary}\label{for:two observations}
Under the notations of Lemma \ref{lem:factor-product}, assume that $H_1^{\varphi_1}$ and $H_2^{\varphi_2}$ are maximal isotropies in $U_1$ and $U_2$, respectively, and $N_{\mathcal G_i}(H_i)= H_i$.
Then, $(H_1 \times H_2)^{(\varphi_1,\varphi_2)}$
is a maximal isotropy in the tensor product.
\end{corollary}
\begin{proof}
From Lemma \ref{lem:factor-product}(i) it immediately follows that $(H_1 \times H_2)^{(\varphi_1,\varphi_2)}$ is an isotropy. Assume for contradiction that it is submaximal. Then
\begin{equation}\label{inclusion2}
(H_1 \times H_2)^{(\varphi_1,\varphi_2)} \lneq (\widetilde H_1 \leftidx{}{_{K_1}}{\times}_{K_2} \widetilde H_2)^{(\widetilde\varphi_1,\widetilde\varphi_2)},
\end{equation} where
\begin{equation}\label{inclusion!}
\widetilde H_i \vartriangleright K_i > H_i
\end{equation} and $\widetilde{\varphi_i}$ is an extension of $\varphi_i$ to $\widetilde H_i$.
If $K_i\gneq H_i$ then by Lemma \ref{lem:factor-product}(ii), $K_i^{\varphi_i}$ is contained in an isotropy, which contradicts maximality of $H_i^{\varphi_i}$.
Since, by assumption, $N_{\mathcal G_i}(H_i) = H_i$ it follows from \eqref{inclusion!} that $\widetilde H_i = H_i$ which contradicts \eqref{inclusion2}.
\end{proof}
\medskip
Returning to the particular situation, where $\mathcal G_i = \mathbb D_N$ and $U_i = (\mathcal U^d_{t_i})^c$, we have the following:
\begin{lemma}\label{YTM2}
If $H_1^{\varphi_1}, H_2^{\varphi_2}$ and $H_3^{\varphi_3}$ are maximal isotropies in $(\mathcal U_{{t}_1}^d)^c$, $(\mathcal U_{{t}_2}^d)^c$ and $(\mathcal U_{{t}_3}^d)^c$ respectively, then $(H_1\times H_2 \times H_3)^{\varphi_1\varphi_2\varphi_3}$ is a maximal isotropy in $(\mathcal{V}_{\bf t}^d)^c$.
\end{lemma}
\begin{proof}
Let us begin by observing that the maximal orbit types in $(\mathcal U_{{t}_i}^d)^c$ are either $\mathbb D_N\times \{1\}$ in the case when $t_i=0$, or $(\mathbb Z_N^{t_i}),(\mathbb D_1^+)$ and $(\mathbb D_1^-)$ if $t_i \neq 0$. By assumption, $N$ is {\it odd}, therefore $N(\mathbb D_1)=\mathbb D_1$. This means that Lemma \ref{lem:factor-product} and Corollary \ref{for:two observations} exclude the possibility that $(H_1\times H_2 \times H_3)^{\varphi_1\varphi_2\varphi_3}$ is not a maximal isotropy except in the case when $(H_1\times H_2 \times H_3)^{\varphi_1\varphi_2\varphi_3}= (\mathbb Z_N\times \mathbb Z_N \times \mathbb Z_N)^{t_1t_2t_3}$. However, it follows from Lemma \ref{lem:factor-product} that the only candidate for a twisted subgroup of $\mathbb D_N\times \mathbb D_N\times \mathbb D_N\times S^1$, which is an isotropy in $(\mathcal{V}_{\bf t}^d)^c$ and contains $(\mathbb Z_N\times \mathbb Z_N \times \mathbb Z_N)^{t_1t_2t_3}$, is of the form $\mathcal H^\varphi$, where
$$
\mathcal H = \{(g_1,g_2,g_3)\in \mathbb D_N\times \mathbb D_N \times \mathbb D_N : \psi(g_1)=\psi(g_2)=\psi(g_3)\}
$$
(here $\psi: \mathbb D_N \to \mathbb Z_2$ is the homomorphism with kernel $\mathbb Z_N$). On the other hand, it can be easily seen that any vector $x\in (\mathcal{V}_{\bf t}^d)^c$ which is fixed by $(\mathbb Z_N\times \mathbb Z_N \times \mathbb Z_N)^{t_1t_2t_3}$ cannot be fixed by the element $(\kappa,\kappa,\kappa,e^{i\theta})$ for any $e^{i\theta}\in S^1$, hence $(\mathbb Z_N\times \mathbb Z_N \times \mathbb Z_N)^{t_1t_2t_3}$ is also maximal.
\end{proof}
\medskip
Being motivated by Lemma \ref{YTM2}, we introduce the following notations:
$$S(t_i)= \begin{cases}
\{(\mathbb D_N\times\{1\})\} &\text{if} \; t_i=0\\
\{ (\mathbb Z_N^{t_i}) ,\; (\mathbb D_1^+), \; (\mathbb D_1^-)\} & \text{if} \; t_i \neq 0
\end{cases}
$$
and
\begin{equation}\label{desc_iso}
\mathcal S({\bf t})= \{(H_1\times H_2\times H_3)^{\varphi_1\varphi_2\varphi_3}: (H_i^{\varphi_i})\in S(t_i)\}
\end{equation}
\subsection{Stability analysis of system \eqref{vdPl}}\label{subsec:Lyapunov-coef}
In this subsection, we analyze stability of Hopf branches of periodic solutions to system
\eqref{vdPl}. This information will be used later
for the stability analysis of the fully synchronized Hopf branches of periodic solutions to system \eqref{nvdPl_2d}.
\begin{lemma}\label{teoorigin}
For the parameter $a$ crossing zero, system \eqref{vdPl} undergoes a
{\bf supercritical} Hopf bifurcation.
\end{lemma}
\begin{proof}
By inspection, the eigenvalues of the linearization of \eqref{vdPl} at the origin have the form $\lambda = {a \pm \sqrt{a^2 - 4b} \over 2}$,
in particular, the Hopf bifurcation takes place when $a$ crosses zero.
To analyze the character of the bifurcation, take the linear
change of coordinates:
\begin{align}\label{changecoord}
\left\{
\begin{array}{l}
\tilde{x}=\displaystyle{\frac{a}{\sqrt{4b-a^2}}x - \frac{2}{\sqrt{4b-a^2}}}y\\
\\
\tilde{y}=x
\end{array}
\right. \\
\hspace{0.01cm}\Rightarrow
\left\{
\begin{array}{l}
\dot{\tilde{x}}=\displaystyle{\frac{a}{2}\tilde{x}-\frac{\sqrt{4b-a^2}}{2}} \tilde{y} \; \displaystyle{-\frac{a\tilde{y}^2-a\tilde{y}^3}{\sqrt{4b-a^2}}}
\\
\\
\dot{\tilde{y}}=\displaystyle{\frac{\sqrt{4b-a^2}}{2}}\tilde{x}+\frac{a}{2}\tilde{y} \; \displaystyle{-\tilde{y}^2-\tilde{y}^3}.
\end{array}
\right.
\end{align}
By direct computation, near $a = 0$, one has:
\begin{equation}\label{Lyap-direction}
\frac{\partial} {\partial a } \Big(\text{Re}( \lambda)\Big) =\frac{1}{2} > 0 \;\; \text {and}\;\; l_1 =-{3 \over 8} < 0,
\end{equation}
where $l_1$ stands for the first Lyapunov coefficient. Combining \eqref{Lyap-direction} with \citep{Guckenheimer}, Theorem 3.4.2, completes the proof.
\end{proof}
\section{Occurrence of Hopf bifurcations}\label{sect:occurrence-result}
\subsection{Abstract result}
Let $G$ be a finite group and $U$ be an orthogonal $G$-representation which admits a $G$-isotypical decomposition
\begin{equation}\label{gen_iso_decom}
U= U_0\oplus \cdots U_k,
\end{equation}
where $U_j$ is modeled on the irreducible representation $\mathcal U_j$. We will denote by $\mathcal U_j^c$ the complexification of $U_j$ which is a $G\times S^1$-representation.
\noindent Suppose $f:\mathbb R \oplus U \to U$ is a $C^1$-smooth function and consider the system
\begin{equation}\label{abst_sys}
\dot x(t)= f(\alpha,x).
\end{equation}
\begin{definition}\label{iso_cen}
We will say that $(\alpha_0,0)$ is an isolated center with limit frequency $\beta_0$ of \eqref{abst_sys} if:
\smallskip
\noindent(i) $(\alpha_0,0)$ is a center of \eqref{abst_sys} with limit frequency $\beta_0$, that is $D_xf(\alpha_0,0)$ admits $i\beta_0$ as a purely imaginary eigenvalue;
\noindent(ii) $(\alpha_0,0)$ is the only center in a neighborhood of $(\alpha_0,0)$ in $\mathbb R \oplus U $.
\end{definition}
We are now in a position to formulate the abstract occurrence result which we will apply to the system considered in this paper.
\begin{theorem}[cf. \citep{AED}]\label{abs_therom}
Suppose $f$ in system \eqref{abst_sys} satisfies the following conditions:
\begin{itemize}
\item[(P1)] $f$ is a $C^1$-smooth equivariant map (we assume $G$ acts trivially on $\mathbb R$);
\item[(P2)] $f(\alpha,0)=0$ for all $\alpha \in \mathbb R$;
\item[(P3)] $(\alpha_0,0)$ is an isolated center for \eqref{abst_sys} (cf. Definition \ref{iso_cen});
\item[(P4)] $\det D_xf(\alpha_0,0) \neq 0$;
\item[(P5)] $D_xf(\alpha_0,0)_{|U_j}$ decreases stability as $\alpha$ passes $\alpha_0$, while the stability of $D_xf(\alpha_0,0)_{|U_k}$ does not increase for any $U_k$ (cf \eqref{gen_iso_decom}).
\end{itemize}
Then, for every maximal orbit type $(H^\varphi)$ in $\mathcal U^c_j$, there exist $|(G\times S^1)/H^\varphi|_{S^1}$ branches of non-constant periodic solutions of \eqref{abst_sys} bifurcating from the origin with (spatio-temporal) symmetry $(H^\varphi)$ and limit period $2\pi/\beta_0$ (cf. Definition \ref{spa_temp}). Here $|(G\times S^1)/H^\varphi|_{S^1}$ is the number of $S^1$-orbits in the space $(G\times S^1)/H$.
\end{theorem}
\begin{remark}
{\rm Using the concept of isotypical crossing number one can relax condition (P5) (see, for example, \citep{AED,BK-chapter,Sliding}).}
\end{remark}
\subsection{Proofs of Theorems \ref{thm_vdp} and \ref{thm_vdpl}}\label{Proof-Theor-occurr}
To detect the occurrence of the equivariant Hopf bifurcation in systems \eqref{nvdP_2d} and \eqref{nvdPl_2d} and to classify symmetric properties of the resulting branches, we will combine the equivariant spectral data collected in Subsections \ref{sec_decop}--\ref{iso_sec} with Theorem \ref{abs_therom}.
\medskip
\noindent{\it (a) Proof of Theorem \ref{thm_vdp}}:
We begin by observing that conditions (P1) and (P2) are obvious. It follows immediately from \eqref{Dn_blocks}--\eqref{eig_d_syst} and $b>0$ that system \eqref{nvdP_2d} can only have a center $(a,0)$ when $a=0$.
Also, since $b,K_{\bf t}>0$, formula \eqref{Dn_blocks} implies (P4). Finally,
all eigenvalues cross the imaginary axis in the same direction (see \eqref{eig_d_syst}), meaning that (P5) is also satisfied. Combining this with the description of isotropies in Section \ref{iso_sec}(i) completes the proof of Theorem \ref{thm_vdp}.
\medskip
\noindent{\it (b) Proof of Theorem \ref{thm_vdpl}}: Observe that (P1) and (P2) are obvious. Plugging $\lambda = i\omega$ into the characteristic equation of matrix $L_{\bf t}(a)$ (cf. \eqref{matL_t}) and seperating real and imaginary parts yields that all centers $(a^z_{\bf t},0)$ of system \eqref{nvdPl_2d} are given by \eqref{eq:bifur-points}.
From this (P3) follows immediately, while (P4) is provided by $b>0$. Finally, differentiating \eqref{eig_z_sys} with respect to $a$ at $a = a^z_{\bf t}$ shows that all eigenvalues cross the imaginary axis in the same direction. Combining this with the description of isotropies in Section \ref{iso_sec}(ii) (in particular, formula \eqref{desc_iso}) completes the proof of Theorem \ref{thm_vdpl}.
\section{Stability of bifurcating branches}\label{sect-stability}
\subsection{Proof of Theorem \ref{Dn_instability}}
Recall that if the equilibrium is unstable then any bifurcating branch of non-constant periodic solutions will also be unstable. Since the eigenvalues of the linearization of system
\eqref{nvdP_2d} at the origin are given by $\sqrt{-bK_{\bf t}}$, it is easy to see that if $K_{\bf t}$ is negative for some {\bf t}, then the theorem is proved. Take $k_1,k_2,k_3$ provided by the hypothesis of Theorem \ref{Dn_instability}. Clearly, $K_{\bf t} < 0 $ for ${\bf t} = \big(k_1\big(\frac{n-1}{2}\big), k_2\big(\frac{n-1}{2}\big), k_3\big(\frac{n-1}{2}\big)\big)$ which completes the proof.
\medskip
\subsection{Proof of Theorem \ref{Zn_stab}} It is clear from formulas \eqref{H_t} and \eqref{eig_z_sys} that the sign of $\text{Re} (\lambda_{\bf t}^z(a))$ is given by the sign of $H_{\bf t}(a)$. Observe that the
largest value of $H_{\bf t}(a)$ is achieved by the value of ${\bf t}$ specified in the Table \ref{tabelaBifurca} and changes its sign at $a = a^{*}$. Therefore, for any $\bf{t}$ and $a < a^*$, one has
$\text{Re} (\lambda_{\bf t}^z(a)) < 0$. Therefore, for $a < a^*$, the equilibrium is stable while for $a > a^*$, the equilibrium is unstable.
In the case when $\delta, \zeta, \varepsilon < 0$, only one pair of eigenvalues crosses the imaginary axis at $a^*=0$. Observe that the central manifold coincides with the central space. To complete the proof we combine Lemma \ref{teoorigin} with the fact that all other eigenvalues have negative real part (cf. \citep{Guckenheimer}, Theorem 3.4.2).
\section{Existence of periodic solutions with prescribed period and symmetry
} \label{sec:Existence}
\subsection{Abstract result.}
To prove Theorem \ref{Existence-Dihedral}, we will use a slight modification of the main result from \citep{AED}, Chapter 12 (cf. Theorem 12.7). To begin with, we need the following
\begin{definition}\label{def:lambda-max-types}
{\rm Let $G$ be a finite group and let $V$ be a real orthogonal $G$-representation. Assume $A : V \to V$ is an equivariant linear operator and $\lambda \in \sigma(A^c)$.
Take the eigenspace $E(\lambda) \subset V^c$
and denote by $\mathcal O(\lambda)$ the set of all maximal $G \times S^1$-orbit types occurring in $E(\lambda)$.
}
\end{definition}
Let $V :=\mathbb R^n$ be an orthogonal permutational $G$-representation and
consider the system
\begin{equation}\label{syst:VDP-Hirano_Ryb}
\begin{cases}
\dot{x}_i = \nu\Big(ax_i - {x^3_i \over 3}\Big) - \sum_{j=1}^n C_{ij}y_j\\
\dot{y}_i = b x_i
\end{cases} \quad (i = 1,\ldots, n).
\end{equation}
\begin{theorem}\label{theom-abstract-Hirano-Ryb} Assume $C$ is a non-singular $G$-equivariant symmetric matrix.
Then, for each real positive $p \not\in \{\frac{2\pi (2k-1)}{\sqrt{\mu}} \; : \; \mu \in \sigma^+(b C), \; k \in \mathbb N \}$ and each $\mu \in \sigma(bC)$ satisfying $0 < (\frac{2\pi}{p})^2 < \mu$, there exists $\nu > 0$ such that for every $(H^\varphi) \in \mathcal O(\mu)$, system \eqref{syst:VDP-Hirano_Ryb} admits $\frac{|G|}{|H|}$ $p$-periodic solutions with minimal symmetry $(H^varphi)$.
\end{theorem}
\subsection {Proof of Theorem \ref{Existence-Dihedral}}
The linearization of system \eqref{syst:VDP-Hirano_Ryb} at the origin has the form:
$$
\mathfrak A=
\begin{bmatrix}
\nu a \text{Id} & -C\\
b \text{Id} & 0
\end{bmatrix}
$$
where Id stands for the $n\times n$-identity matrix. Since in the case of system \eqref{nvdP_2d}, the isotypical decomposition of $V$ contains only irreducible representations of real type, then $\mathfrak A_{|V_i}$ is of the form
$$
\begin{bmatrix}
\nu a & -\lambda\\
b & 0
\end{bmatrix} \otimes \text{Id}_{|V_i}
$$
where $\lambda$ is an eigenvalue of $C_{|V_i}$. It follows from Remark \ref{prim_decom} and formula \eqref{Dn_blocks} that
$$
\sigma(C)=\{K_{\bf_t}: 0<t_i<(N-1)/2\}.$$
Also note that $\mathcal O(K_{\bf t})= \mathcal S({\bf t})$. Combining this with Theorem \ref{theom-abstract-Hirano-Ryb} completes the proof of Theorem \ref{Existence-Dihedral}.
\section{Appendix}
If $W$ is a $G$-representation, then for any function $x:S^1 \to W$ the spatio-temporal symmetry of $x$ is a group $\mathfrak H < G \times S^1$ such that $g\cdot x(t-\theta)=x(t)$ for any $t\in \mathbb R/2\pi\mathbb Z \simeq S^1$ and any $(g,e^{i\theta}) \in \mathfrak H$. If $x$ is non-constant, then its spatio-temporal symmetry group will have the structure of a graph of a homomorphism from some subgroup $H < G$ to $S^1$. In our general discussion, we used the following notation:
$$
H^\varphi := \{(h,\varphi(h)\;:\;h\in H)\}.
$$
We call this a {\it twisted} symmetry group with twisting homomorpism $\varphi$. If the domain of the twisting homomorphism is a direct product of groups, we can describe the twisting homomorphism by its restrictions to each of the components. Therefore, we use the following notation:
\begin{equation}
\label{prod_twist}
\begin{aligned}
(H_1\times H_2\times H_3)^{\varphi_1\varphi_2\varphi_3} := &\{(h_1,h_2,h_3,\varphi_1(h_1)\varphi_2(h_2)\varphi_3(h_3))\\ &\;:\; (h_1,h_2,h_3)\in H_1\times H_2\times H_3\}
\end{aligned}
\end{equation}
Given two groups $\mathscr G_1$ and $\mathscr G_2$, to describe subgroups of $\mathscr G_1\times \mathscr G_2$,
define the projection homomorphisms:
\begin{align*}
\pi _{1}&:\mathscr G_1\times\mathscr G_2\to \mathscr G_1,\quad\pi_1(g_1,g_2)=g_1;\\
\pi _{2}&:\mathscr G_1\times\mathscr G_2\to \mathscr G_2,\quad\pi_2(g_1,g_2)=g_2.
\end{align*}
The following result, being a reformulation of the well-known
Goursat's Lemma (cf.~\citep{Goursat}),
provides the desired description of subgroups $\mathscr H$ of the
product group $\mathscr G_1\times \mathscr G_2$.
\begin{theorem}\label{th:product-1}
Let $\mathscr H$ be a subgroup of the product group
$\mathscr G_1\times \mathscr G_2$. Put $H:=\pi_1(\mathscr H)$
and $K:=\pi_2(\mathscr H)$. Then, there exist a group $L$ and
two epimorphisms $\varphi :H\rightarrow L$ and $\psi :K\rightarrow L$,
such that
\begin{equation}\label{eq:product-rep}
\mathscr H=\{(h,k)\in H\times K: \varphi(h)=\psi(k)\}.
\end{equation}
\end{theorem}
For the needs of our paper, it is enough to characterize such subgroups up to the kernels of the homomorphisms $\varphi$ and $\psi$. For this reason, we put
\begin{equation}\label{amal_not}
\mathscr H =: H_1 \leftidx{}{_{K_1}}{\times}_{K_2} H_2
\end{equation}
where $K_1$ is the kernel of $\varphi$ and $K_2$ is the kernel of $\psi$.
\medskip
In the particular case of $\mathbb D_N$, we denote by $\mathbb Z_N$ the subgroup generated by $e^{\frac{2\pi i}{N}}$, and by $\mathbb D_1$ the subgroup generated by $\kappa$. Furthermore, we put
\begin{eqnarray*}
\mathbb Z_N^{t_i} := \mathbb Z_N^\varphi &\text{where} & \varphi:e^{k\frac{2\pi i}{N}}\mapsto e^{kt_i\frac{2\pi i}{N}}\\
\mathbb D_1^+ := \mathbb D_1^\varphi &\text{where} & \varphi: \kappa \to 1\\
\mathbb D_1^- := \mathbb D_1^\varphi &\text{where} & \varphi: \kappa \mapsto -1
\end{eqnarray*}
We combine this with the notation for twisted groups given in \eqref{prod_twist} in the obvious way, for example:
$$
(\mathbb Z_N\times \mathbb D_1\times \mathbb Z_N)^{(t_1,-,t_2)}:=(\mathbb Z_N\times \mathbb D_1\times \mathbb Z_N)^{\varphi_1\varphi_2\varphi_3}
$$
where
$\varphi_1:e^{k\frac{2\pi i}{N}}\mapsto e^{kt_1\frac{2\pi i}{N}}$, $\varphi_2:\kappa \mapsto -1$ and $\varphi_3:e^{k\frac{2\pi i}{N}}\mapsto e^{kt_2\frac{2\pi i}{N}}$.
\bibliographystyle{siam}
|
\section{Introduction}
We take the Stone-\v Cech compactification of a discrete semigroup
$(S,\cdot)$ to be the set of ultrafilters on $S$, identifying the points of
$S$ with the principal ultrafilters. Given $A\subseteq S$,
we set $\overline A=\{p\in\beta S:A\in p\}$. The set $\{\overline A:A\subseteq S\}$
is a basis for the open sets and a basis for the closed sets of $\beta S$.
The operation on $S$ extends uniquely to $\beta S$ so that $(\beta S,\cdot)$ is
a right topological semigroup with $S$ contained in its topological center,
meaning that $\rho_p$ is continuous for each $p\in\beta S$ and $\lambda_x$ is continuous
for each $x\in S$, where for $q\in\beta S$, $\rho_p(q)=q\cdot p$ and $\lambda_x(q)=x\cdot q$.
So, for every $p,q\in\beta S$, $pq=\displaystyle \lim_{s\to p}\lim_{t\to q}st$, where $s$ and $t$
denote elements of $S$.
If $A\subseteq S$, $A\in p\cdot q$
if and only if $\{x\in S:x^{-1}A\in q\}\in p$, where $x^{-1}A=\{y\in S:xy\in A\}$. (We are following the
custom of frequently writing $xy$ for $x\cdot y$.)
The algebraic structure of $\beta S$ is interesting in its own right, and has had
substantial applications, especially to that part of combinatorics known as
{\it Ramsey Theory\/}. See the book \cite{HS} for an elementary introduction to the
structure of $\beta S$ and its applications.
We are concerned in this paper with the relationship between the algebraic structure of
$\beta S$ and recurrence in {\it dynamical systems\/}.
\begin{definition}\label{defdyn}
A {\it dynamical system\/} is a pair
$(X,\langle T_s\rangle_{s\in S})$ such that
\begin{itemize}
\item[(1)] $X$ is a compact Hausdorff topological
space (called the {\it phase space\/}
of the system);
\item[(2)] $S$ is a semigroup;
\item[(3)] for each $s\in S$, $T_s$ is a continuous function from $X$
to $X$; and
\item[(4)] for all $s,t\in S$, $T_s\circ T_t=T_{st}$.
\end{itemize}
\end{definition}
Associated with any semigroup $S$ are at least two interesting
dynamical systems, namely $(\beta S,\langle \lambda_s\rangle_{s\in S}\rangle)$,
and $(\xtotwo{S},\langle T_s\rangle_{s\in S})$ where $\xtotwo{S}$ is the
set of all functions from $S$ to $\{0,1\}$ with the product topology and $T_s(x)=x\circ\rho_s$.
(We shall verify that this latter example is a dynamical system shortly.)
It is common to assume that the phase space of a dynamical system
is a metric space, but we make no such assumption.
If $S$ is infinite, then $\beta S$ is not a metric space. Everything we
do here is boring if $S$ is finite so whenever we write
``let $S$ be a semigroup" we shall assume that $S$ is infinite. The
interested reader can amuse herself by determining which of our results
remain valid if that assumption is dropped.
The system $(\beta S,\langle \lambda_s\rangle_{s\in S}\rangle)$ has significant
general properties as can be seen in \cite[Section 19.1]{HS}, but will
not be used much in this paper.
Given a product space $\xtotwo{S}$, recall that the product topology has
a subbasis consisting of sets of the form $\pi_{t}^{-1}[\{a\}]$ for
$t\in S$ and $a\in\{0,1\}$, where, for $x\in\xtotwo{S}$, $\pi_t(x)=x(t)$.
\begin{lemma}\label{stotwodyn} Let $R$ be a semigroup and let $S$ be a subsemigroup
of $R$. Let $Z=\xtotwo{R}$, the set of all functions from $R$ to $\{0,1\}$ with
the product topology. For $x\in Z$ and $s\in S$, define $T_s(x)=x\circ\rho_s$.
Then $(Z,\langle T_s\rangle_{s\in S})$ is a dynamical system.
\end{lemma}
\begin{proof}
It is routine to verify that for $s,t\in S$, $T_s\circ T_t=T_{st}$.
To see that $T_s$ is continuous for each $s\in S$, let $s\in S$ be given.
It suffices to show that the inverse image of each subbasic open set
is open, so let $t\in R$ and $a\in\{0,1\}$ be given.
Then $T_s^{-1}\big[\pi_t^{-1}[\{a\}]\big]=\pi_{ts}^{-1}[\{a\}]$.\end{proof}
Recall that, if $T$ is any discrete space, $p\in\beta T$,
$\langle x_t\rangle_{t\in T}$ is any indexed family in a Hausdorff
topological space $X$, and $y\in X$, then $\displaystyle p{-}\!\lim_{t\in T}x_t=y$
if and only if for every neighborhood $U$ of $y$, $\{t\in T:x_t\in U\}\in p$.
In compact spaces $p$-limits always exist.
\begin{definition}\label{defTp}
Let $(X,\langle T_s\rangle_{s\in S})$ be a dynamical
system and let $p\in\beta S$. Then $T_p:X\to X$ is defined by, for $x\in X$,
$T_p(x)=\displaystyle p{-}\lim_{s\in S}T_s(x)$. So $T_p(x)=\displaystyle \lim_{s\to p}T_s(x)$ where
$s$ denotes an element of $S$. \end{definition}
Using \cite[Theorem 4.5]{HS} one easily sees that for $p,q\in\beta S$,
$T_p\circ T_q=T_{pq}$. However, $(X,\langle T_s\rangle_{s\in \beta S})$ is
not in general a dynamical system, since $T_p$ is not likely to be
continuous when $p\in\beta S\setminus S$. However, for each $x\in X$, the map
$p\mapsto T_p(x):\beta S \to X$ is continuous. To see this, define $f_x(p)=T_p(x)$.
If $U$ is a neighborhood of $f_x(p)$ and $A=\{s\in S:T_s(x)\in U\}$, then
$U\in p$ and $f_x[\,\overline A\,]\subseteq U$. Alternatively, one may note
that $p\mapsto T_p(x)$ is the continuous extension to $\beta S$ of the function
$s\mapsto T_s(x):S\to X$.
As a compact Hausdorff right topological semigroup, $\beta S$ has a number of important algebraic properties,
and we list some of those that we shall use. (Proofs can be found in
\cite[Chapters 1 and 2]{HS}. Assume that $T$ is a compact Hausdorff right topological semigroup.
A non-empty subset $V$ of $T$ is a {\it left ideal\/} if $tV\subseteq V$ for every $t\in T$, a {\it right ideal\/} if
$Vt\subseteq V$ for every $t\in T$, and an {\it ideal\/} if it is both a left and a right ideal.
\begin{itemize}
\item[(1)] $T$ contains an idempotent.
\item[(2)] $T$ has a smallest ideal $K(T)$, which is the union of the minimal left ideals of $T$ and the union of the minimal right ideals of $T$.
\item[(3)] For every $t\in K(T)$, $Tt$ is a minimal left ideal of $T$ and $tT$ is a minimal right ideal of $T$.
\item[(4)] The intersection of any minimal left ideal and any minimal right ideal of $T$ is a group.
\item[(5)] Every left ideal of $T$ contains a minimal left ideal, and every right ideal of $T$ contains a minimal
right ideal.
\item[(5)] Every minimal left ideal of $T$ is compact.
\item[(6)] If $\{t\in T:\lambda_t$ is continuous$\}$ is dense in $T$, then the closure of every ideal
in $T$ is also an ideal.
\end{itemize}
We introduce the main objects of study in this paper now. Given a set
$X$, we let $\pf(X)$ be the set of finite nonempty subsets of $X$.
\begin{definition}\label{defsyn}
Let $S$ be a semigroup and let $A\subseteq S$. We say the set $A$ is
{\it syndetic\/} if and only if there exists $F\in\pf(S)$ such that
$S=\bigcup_{t\in F}t^{-1}A$.\end{definition}
In the semigroup $(\ben,+)$ a set is syndetic if and only if it has
bounded gaps.
\begin{definition}\label{defurec} Let $(X,\langle T_s\rangle{s\in S})$ be
a dynamical system and let $x\in X$.
\begin{itemize}
\item[(a)] The point $x$ is {\it uniformly recurrent\/} if and only
if for every neighborhood $V$ of $x$, $\{s\in S:T_s(x)\in V\}$ is syndetic.
\item[(b)] $U(x)=U_X(x)=\{p\in\beta S:T_p(x)$ is uniformly recurrent$\}$.
\end{itemize}
\end{definition}
In Section 2 of this paper we present well known results about $U(x)$ that are valid in
arbitrary dynamical systems as well as the few simple results that we have in
the dynamical system $(\beta S,\langle \lambda_s\rangle_{s\in S})$.
In Section 3 we present results about the dynamical systems described in
Lemma \ref{stotwodyn}.
In Section 4 we consider the effect of slightly modifying the phase
space in the dynamical systems described in
Lemma \ref{stotwodyn}.
In Section 5 we consider surjectivity of $T_p$ and the set
$NS=NS_X=\{p\in \beta S:T_p:X\to X$ is not surjective$\}$ which
is a right ideal of $\beta S$ whenever it is nonempty.
\section{General results}
We begin with some well known basic facts.
\begin{lemma}\label{uniformrecurrence} Let $(X,\langle T_s\rangle_{s\in S})$ be a dynamical system,
let $L$ be a minimal left ideal of $\beta S$, and let $x\in X$. The following are equivalent:
\begin{itemize}
\item[(a)] $x$ is uniformly recurrent.
\item[(b)] There exists $q\in L$ such that $T_q(x)=x$.
\item[(c)] There exists an idempotent $q\in L$ such that $T_q(x)=x$.
\item[(d)] There exists $y\in X$ and $q\in L$ such that $T_q(y)=x$.
\item[(e)] There exists $q\in K(\beta S)$ such that $T_q(x)=x$.
\item[(f)] There exists $y\in X$ and $q\in K(\beta S)$ such that $T_q(y)=x$.
\end{itemize} \end{lemma}
\begin{proof} The equivalence of (a)-(d) is shown in
\cite[Theorem 19.23]{HS}. Since (c) implies (e), and (e)
implies (f), we shall show (f) implies (c) and this will establish the equivalence of all six
statements. So assume that (f) holds. Let $u$ denote the identity of the group $L\cap q\beta S$.
Since $uq=q$, it follows that $T_u(x)=T_uT_q(y)=T_q(y)=x$.
\end{proof}
\begin{corollary}\label{U(x)ideal} Let $(X,\langle T_s\rangle_{s\in S})$ be a dynamical system
and let $x\in X$.
\begin{itemize}
\item[(1)] If $x$ is uniformly recurrent, $U(x)=\beta S$.
\item[(2)] For each $x\in X$, $U(x)$ is a left ideal of $\beta S$.
\item[(3)] For every $x\in X$, $K(\beta S)\subseteq U(x)$.
\item[(4)] $\bigcap_{x\in X}U(x)$ is a two sided ideal of $\beta S$.
\end{itemize} \end{corollary}
\begin{proof}
(1) Suppose that $x$ is uniformly recurrent. Then $T_u(x)=x$ for some $u\in K(\beta S)$.
Thus for every $v\in \beta S$, $T_v(x)=T_vT_u(x)=T_{vu}(x)$; since $vu\in K(\beta S)$,
by Lemma \ref{uniformrecurrence}(f), $T_v(x)$ is uniformly recurrent.
(2) Let $x\in X$, let $p\in U(x)$, and let $r\in\beta S$. By Lemma \ref{uniformrecurrence}(e),
pick $q\in K(\beta S)$ such that $T_q\big(T_p(x)\big)=T_p(x)$. Then
$T_{rp}(x)=T_r\big(T_q\big(T_p(x)\big)\big)=T_{rqp}(x)$. Now $rqp\in K(\beta S)$,
so by Lemma \ref{uniformrecurrence}(f), $T_{rp}(x)$ is uniformly recurrent.
(3) This is immediate from Lemma \ref{uniformrecurrence}(f).
(4) By (3), $\bigcap_{x\in X}U(x)$ is nonempty, so by (2) $\bigcap_{x\in X}U(x)$ is a left
ideal of $\beta S$, so it is enough to show that $\bigcap_{x\in X}U(x)$ is a
right ideal of $\beta S$. So suppose that $x\in X$,
$p\in\bigcap_{x\in X}U(x)$ and $q\in \beta S$. Since $p\in U(T_q(x))$, $T_{pq}(x)$ is uniformly recurrent
and so $pq\in U(x)$.
\end{proof}
The statements of Lemma \ref{basicfacts} below are modifications of basic well known
facts that are proved in \cite{F}.
(Furstenberg assumes that the phase space is metric, but the
proofs given do not use this assumption.) We shall say that a subspace $Z$ of
$X$ is {\it invariant\/} if $T_s[Z]\subseteq Z$ for every $s\in S$.
Of course, if $Z$ is closed and invariant, then $T_p[Z]\subseteq Z$ for every $p\in \beta S$.
(Let $x\in Z$. Then $T_s(x)\in Z$ for each $s\in S$ so $\displaystyle p{-}\lim_{s\in S}T_s(x)\in Z$.)
\begin{lemma}\label{basicfacts} Let $(X,\langle T_s\rangle_{s\in S})$ be a dynamical system.
Let $L$ be a
minimal left ideal of $\beta S$.
\begin{itemize}
\item[(1)] A subspace $Y$ of $X$ is minimal among all closed and invariant subsets of $X$ if and only if
there is some $x\in X$ such that $Y=\{T_p(x):p\in L\}$.
\item[(2)] Let $Y$ be a subspace of $X$ which is minimal among all closed and invariant subsets of $X$.
Then every element of $Y$ is uniformly recurrent.
\item[(3)] If $x\in X$ is uniformly recurrent and $Y=\{T_p(x):p\in \beta S\}$, then $Y$ is
minimal among all closed and invariant subsets of $X$.
\item[(4)] If $x\in X$ is uniformly recurrent, then $T_p(x)$ is uniformly recurrent for every $p\in \beta S$.
\end{itemize} \end{lemma}
\begin{proof} (1) Suppose that $Y$ is a subspace of $X$ which is minimal
among all closed and invariant subsets of $X$. Pick $x\in Y$ and let $Z=\{T_p(x):p\in L\}$. We
claim that $Z$ is a closed and invariant subspace of $Y$ and is therefore equal to $Y$.
If $p\in L$ and $s\in S$, then $T_s\big(T_p(x)\big)=T_{sp}(x)$ and $sp\in L$, so $Z$ is invariant and
obviously $Z\subseteq Y$. To see that $Z$ is closed, it suffices to show that any net in $Z$ has a cluster
point in $Z$. To this end, let $\langle p_\alpha\rangle_{\alpha\in D}$ be a net in $L$ and
pick a cluster point $p$ in $L$ of $\langle p_\alpha\rangle_{\alpha\in D}$. Then
$T_p(x)$ is a cluster point of $\langle T_{p_\alpha}(x)\rangle_{\alpha\in D}$.
Conversely, let $x\in X$ and let $Y=\{T_p(x):p\in L\}$.
Then $Y$ is invariant and one sees as above that $Y$ is closed.
We shall show that $Y$ is minimal among all closed and invariant subsets of $X$.
To see this, suppose that $Z$ is a subset of $Y$ which is closed and
invariant. We shall show that $Y\subseteq Z$, so let $y\in Y$ be given. Pick $z\in Z$.
Then $y=T_p(x)$ and $z=T_q(x)$ for some $p$ and $q$ in $L$.
Since $Lq=L$, there exists
$r\in L$ such that $rq=p$. It follows that $T_r(z)=y$ and hence that $y\in Z$ as required.
(2) Let $Y$ be a subspace of $X$ which is minimal among all closed and invariant subsets of $X$ and let
$x\in Y$. Pick $y\in X$ such that $Y=\{T_p(y):p\in L\}$. Pick
$p\in L$ such that $x=T_p(y)$. By Lemma \ref{uniformrecurrence}(f), $x$ is uniformly recurrent.
(3) Let $x$ be a uniformly recurrent point of $X$ and let $Y=\{T_p(x):p\in \beta S\}$.
By Lemma \ref{uniformrecurrence}(b), pick $q\in L$ such that $T_q(x)=x$.
By (1) it suffices that $Y=\{T_p(x):p\in L\}$. To see this, let $y\in Y$ and
pick $p\in\beta S$ such that $y=T_p(x)$. Then $y=T_p\big(T_q(x)\big)=T_{pq}(x)$ and
$pq\in L$.
(4) Let $x$ be a uniformly recurrent point of $X$ and let $Y=\{T_p(x):p\in \beta S\}$.
By (3) $Y$ is minimal among all closed and invariant subsets of $X$ so (2) applies.
\end{proof}
We conclude this section with a few results about the dynamical system
$(\beta S,\langle \lambda_s\rangle_{s\in S})$. We observe that, if we define $\lambda_p:\beta S\to\beta S$
in this system by $\lambda_p(q)=\displaystyle\lim_{s\to p}\lambda_s(q)$, where $s$ denotes an element of $S$,
then $\lambda_p(q)=pq$ for every
$p$ and $q$ in $\beta S$. So this does not conflict with the previous definition of $\lambda_p$ given in the
introduction.
\begin{theorem}\label{charunifrec} Let $S$ be a semigroup and
let $x\in \beta S$. Statements (a) and (b) are equivalent and imply (c). If $\beta S$
has a left cancelable element, all three are equivalent.
\begin{itemize}
\item[(a)] $x\in K(\beta S)$.
\item[(b)] $x$ is uniformly recurrent in the
dynamical system $(\beta S,\langle \lambda_s\rangle_{s\in S})$.
\item[(c)] $\beta Sx$ is a minimal left ideal of $\beta S$.
\end{itemize}
\end{theorem}
\begin{proof} To see that (a) implies (b), let $x\in K(\beta S)$ and let
$u$ be the identity of the group in $K(\beta S)$ to which $x$ belongs.
Then $x=\lambda_u(x)$ so by Lemma {uniformrecurrence}(e), $x$ is uniformly recurrent.
To see that (b) implies (a), assume that $x$ is uniformly recurrent. By
Lemma \ref{uniformrecurrence}(f) pick $y\in \beta S$ and $q\in K(\beta S)$ such that
$\lambda_q(y)=x$. Then $x=qy\in K(\beta S)$.
To see that (a) implies (c), assume that $x\in K(\beta S)$ and pick the minimal left ideal $L$
of $\beta S$ such that $x\in L$. Then $\beta Sx$ is a left ideal of $\beta S$ contained in $L$
and so $\beta Sx=L$.
Now assume that $\beta S$ has a left cancelable element $z$ and
that $\beta Sx$ is a minimal left ideal of $\beta S$.
Pick an idempotent $u\in \beta Sx$. Then
$zx\in\beta Sx$ so by \cite[Lemma 1.30]{HS}, $zx=zxu$ and
therefore $x=xu\in\beta Sx\subseteq K(\beta S)$.
\end{proof}
\begin{corollary}\label{U(x)isbeta} Let $S$ be an infinite semigroup and
let $x\in K(\beta S)$. Then $U(x)=\beta S$ with respect to the dynamical system
$(\beta S,\langle \lambda_s\rangle_{s\in S})$.
\end{corollary}
\begin{proof} By Theorem \ref{charunifrec}, $x$ is uniformly recurrent, so by Lemma
\ref{basicfacts}(4), $U(x)=\beta S$.\end{proof}
\begin{corollary}\label{charqinU(p)} Let $S$ be a semigroup and
let $p,q\in \beta S$. Statements (a) and (b) are equivalent and imply statement (c).
If $\beta S$ has a left cancelable element,
then all three statements
are equivalent.
\begin{itemize}
\item[(a)] $qp\in K(\beta S)$.
\item[(b)] $q\in U(p)$ with respect to the
dynamical system $(\beta S,\langle \lambda_s\rangle_{s\in S})$.
\item[(c)] $\beta Sqp$ is a minimal left ideal of $\beta S$.
\end{itemize}
\end{corollary}
\begin{proof}
We have that $q\in U(p)$ if and only if $\lambda_q(p)$ is
uniformly recurrent and $\lambda_q(p)=qp$
so Theorem \ref{charunifrec} applies.
\end{proof}
It is an old and difficult problem to characterize when
$K(\beta S)$ is prime or when $\cl K(\beta S)$ is prime.
There are trivial situations where the answer is known.
For example if $S$ is left zero or right zero, then
so is $\beta S$ and thus $K(\beta S)=\beta S$, and
is necessarily prime. It is not known whether
$K(\beta \ben,+)$ is prime or
$\cl K(\beta \ben,+)$ is prime. (Some partial results
were obtained in \cite{HSa}.)
\begin{corollary}\label{prime} Let $S$ be a semigroup. The following statements
are equivalent.
\begin{itemize}
\item[(a)] There exists $p\in\beta S\setminus K(\beta S)$ such that,
with respect to the dynamical system
$(\beta S,\langle \lambda_s\rangle_{s\in S})$,
$K(\beta S)\subsetneq U(p)$.
\item[(b)] $K(\beta S)$ is not prime.
\end{itemize}
\end{corollary}
\begin{proof} This is an immediate
consequence of Corollary \ref{charqinU(p)}.
\end{proof}
\section{Dynamical systems with phase space $\xtotwo{R}$}
Throughout this section we assume that $R$ is a semigroup,
$S$ a subsemigroup of $R$, and
$(Z,\langle T_s\rangle_{s\in S})$ is the dynamical system of
Lemma {\rm \ref{stotwodyn}}. While our results are valid in this
generality, in practice we are interested in just two situations,
one in which $R=S$ and the other in which $R=S\cup\{e\}$ where
$e$ is a two sided identity adjoined to $S$.
Our first results in this section are aimed at
showing that for any semigroup $S$, there is a dynamical
system such that both $K(\beta S)$ and $\cl K(\beta S)$
are intersections of sets of the form $U(x)$.
\begin{definition} Given $x\in Z$ we denote the continuous
extension of $x$ from $\beta R$ to $\{0,1\}$ by $\widetilde x$. \end{definition}
Of course, for each $x\in Z$, each $p\in \beta S$ and each $t\in R$,
$T_p(x)(t)=\displaystyle p{-}\lim_{s\in S}T_s(x)(t)=p{-}\lim_{s\in S}x(ts)$
and so $T_p(x)(t)=\widetilde x(tp)$.
\begin{lemma}\label{charUx} Let $x\in Z$, let $p\in \beta S$, and let $L$ be a minimal left ideal
of $\beta S$. The following statements are equivalent:
\begin{itemize}
\item[(a)] $p\in U(x)$.
\item[(b)] There exists $q\in L$
such that $\widetilde x(tp)=\widetilde x(tqp)$ for all $t\in R$.
\item[(c)] There exists an idempotent $q\in L$ such that $\widetilde x(tp)=\widetilde x(tqp)$
for all $t\in R$.
\end{itemize}
\end{lemma}
\begin{proof} To see that (a) implies (c), assume that $T_p(x)$ is uniformly recurrent.
By Lemma \ref{uniformrecurrence}(c), pick an idempotent $q\in L$ such that
$T_q\big(T_p(x)\big)=T(p)(x)$. Then $T_{qp}(x)=T_p(x)$ so as noted above,
for all $t\in R$, $\widetilde x(tqp)=\widetilde x(tp)$.
Trivially (c) implies (b). To see that (b) implies (a), pick
$q\in L$ such that $\widetilde x(tp)=\widetilde x(tqp)$ for all $t\in R$.
Then $T_p(x)=T_{qp}(x)=T_q\big(T_p(x)\big)$, so by Lemma \ref{uniformrecurrence}(b),
$T_p(x)$ is uniformly recurrent.\end{proof}
\begin{lemma}\label{charUxp} Let $x\in Z$ and let $p\in\beta S$. Then
$p\in U(x)$ if and only if for every minimal left ideal $L$ of $\beta S$ and
every $F\in\pf(R)$, there exists $q_F\in L$ such that for all $t\in F$,
$\widetilde x(tp)=\widetilde x(tq_Fp)$.\end{lemma}
\begin{proof} The necessity is an immediate consequence of Lemma \ref{charUx}(b).
For the sufficiency, let $L$ be a minimal left ideal of $\beta S$. For
each $F\in\pf(R)$, pick $q_F\in L$ as guaranteed. Direct $\pf(R)$ by agreeing
that $F<G$ if and only if $F\subseteq G$. Pick a cluster point
$q\in L$ of the net $\langle q_F\rangle_{F\in\pf(R)}$. It is then
routine to show that for all $t\in R$, $\widetilde x(tqp)=
\widetilde x(tp)$ so that by Lemma \ref{charUx}(b), $p\in U(x)$.
\end{proof}
\begin{theorem}\label{bigcapUx} \begin{itemize}
\item[(1)] $K(\beta S)\subseteq \bigcap_{x\in Z}U(x)$.
\item[(2)] If $p\in \bigcap_{x\in Z}\,U(x)$, then, for every minimal left ideal $L$ of $\beta S$,
$\beta Sp=Lp$ and so $\beta Sp$ is a minimal left ideal of
$\beta S$.
\item[(3)] If $R$ contains a left cancelable element,
then $K(\beta S)=\bigcap_{x\in Z}U(x)$.
In particular, if $R$ has a left identity,
then $K(\beta S)=\bigcap_{x\in Z}U(x)$.
\end{itemize}
\end{theorem}
\begin{proof} (1) $K(\beta S)\subseteq\bigcap_{x\in Z}U(x)$ by Corollary \ref{U(x)ideal}(3).
(2) Assume that $p\in \bigcap_{x\in Z}\,U(x)$. Let $L$ be a minimal left ideal of $\beta S$.
We shall show that, for every $t\in R$, $tp\in tLp$. To see this, assume the contrary. Then for some
$t\in R$, there exists $A\subseteq R$ such that $A\in tp$ and $\overline A\cap tLp=\emptyset$.
Let $x=\cchi_A$. So $\widetilde x$ is the characteristic function of
$\overline A$. Since $p\in U(x)$, it follows from Lemma \ref{charUx}
that $\widetilde x(tp)=\widetilde x(tqp)$ for some $q\in L$. However, $\widetilde x(tp)=1$ and
$\widetilde x(tqp)=0$. This contradiction establishes that $tp\in tLp$ for every $t\in R$. In particular,
$\beta Sp=\cl_{\beta S}Sp\subseteq Lp$. So $\beta Sp\subseteq Lp$. By \cite[Theorem 1.46]{HS},
$Lp$ is a minimal left ideal of $\beta S$, and so $\beta Sp=Lp$.
(3) Now suppose that $R$ contains a left cancelable element $t$ and let $p\in \bigcap_{x\in Z}U(x)$
Since $t$ is left cancelable in $\beta R$ by
\cite[Lemma 8.1]{HS} and $tp=tqp$ for some $q\in L$, it follows that $p=qp\in K(\beta S)$.
\end{proof}
Recall that a subset $A$ of a semigroup $S$ is {\it piecewise syndetic\/} if and only if
there is some $G\in\pf(S)$ such that for every $F\in\pf(S)$, there is some
$x\in S$ with $Fx\subseteq\bigcup_{t\in G}t^{-1}A$. The important fact about
piecewise syndetic sets is that they are the subsets of $S$ whose closure meets
$K(\beta S)$, \cite[Theorem 4.40]{HS}.
\begin{definition}
$\Omega=\Omega_Z=\{x\in Z:\overline{x^{-1}[\{1\}]\cap S}\cap K(\beta S)=\emp\}$.
\end{definition}
Thus $\Omega=\{x\in Z:x^{-1}[\{1\}]\cap S\hbox{ is not piecewise syndetic in }S\}$.
Note that, since $K(\beta S)$ is usually not topologically closed,
we have by Theorem \ref{bigcapUx} that not all sets of the form
$U(x)$ are closed.
\begin{definition} Let $x\in Z$.
$N(x)=\{p\in \beta S:(\forall t\in R)(T_p(x)(t)=0)\}$.
\end{definition}
\begin{lemma}\label{UequalsN} Let $x\in Z$. Then $N(x)$ is closed
and $N(x)\subseteq U(x)$. If $N(x)=U(x)$, then $x\in \Omega$.
If $S$ is a left ideal of $R$, then $N(x)=U(x)$ if and only if $x\in\Omega$.
\end{lemma}
\begin{proof}
To see that $N(x)$ is closed, let $p\in\beta S\setminus N(x)$, pick
$t\in R$ such that $T_p(x)(t)=1\}$, and let $A=\{s\in S:T_s(x)(t)=1\}$. Then
$A\in p$ and $\overline A\cap N(x)=\emp$.
If $T_p(x)$ is constantly equal to $0$ on $R$, then $T_p(x)$ is uniformly
recurrent and thus $p\in U(x)$.
Let $A=x^{-1}[\{1\}]\cap S$.
First assume that $N(x)=U(x)$
and suppose that $x\notin\Omega$. Since
$\overline A\cap K(\beta S)\neq \emptyset $, pick $p\in\overline A\cap K(\beta S)$.
By Corollary \ref{U(x)ideal}(3), $p\in U(x)$ and
so for all $t\in R$, $T_p(x)(t)=0$. Since $K(\beta S)$ is a union of groups, there exists
$q\in K(\beta S)$ such that $qp=p$. Pick $t\in S$ such that $t^{-1}A\in p$.
Also $T_p(x)(t)=0$ so $\{s\in S:x(ts)=0\}\in p$. Pick $s\in t^{-1}A$ such that
$x(ts)=0$, a contradiction.
Now assume that $S$ is a left ideal in $R$. Let $x\in\Omega$ and let $p\in U(x)$. We claim that $p\in N(x)$.
To see this, suppose we have some $t\in R$ such that
$T_p(x)(t)=1$. By Lemma \ref{charUx}, there exists an idempotent $q\in K(\beta S)$ such that
$\widetilde x(tqp)=1$. By \cite[Theorem 2.17]{HS}, $\beta S$ is a left ideal of $\beta R$ so
$tqp\in\beta S$ and so $A\in tqp=tqqp$. Thus there is some $s\in S$ such that
$tsqp\in\overline A$. Since $ts\in S$, $tsqp\in K(\beta S)$, a contradiction.
\end{proof}
\begin{lemma}\label{bigcapclUx} Let $p\in \bigcap_{x\in\Omega}U(x)$ and let $t\in R$. If $tp\in \beta S$,
then $tp\in \cl K(\beta S)$. In particular, $\beta Sp\subseteq \cl K(\beta S)$.
\end{lemma}
\begin{proof}
Assume that $tp\in \beta S\setminus \cl(K\beta S)$.
We can choose $A\in tp$ such that $A\subseteq S$ and
$\overline A\cap K(\beta S)=\emptyset$. Let $x$ be the characteristic function of $A$
in $R$, so that $x\in \Omega$ and hence $p\in U(x)$. Observe that
$\wdd x$ is the characteristic funcion of $\cl_{\beta R}(A)$ in $\beta R$ and
that $\cl_{\beta R}(A)\subseteq \beta S$, because $\beta S$ is
clopen in $\beta R$. Since $\wdd x(tp)=1$, it follows from Lemma \ref{charUx}(b) that
there exists $q\in K(\beta S)$ such that $\wdd x(tqp)=1$, and so $A\in tqp$. Now
$\{r\in \beta S:tqr\in \beta S\}$ is non-empty and is a right ideal of $\beta S$.
There exists an idempotent $u$ in the intersection of this right ideal with the left ideal
$\beta Sq$ of $\beta S$. Since $q\in \beta Su$, $qu=q$. So $tqp=tquup\in K(\beta S)$,
because $tqu\in\beta S$ and $u\in\beta Sq\subseteq K(\beta S)$.
This contradicts the assumption that $\overline A\cap K(\beta S)=\emptyset$.
\end{proof}
\begin{corollary}\label{clKbS} Each of the following statements
implies that $\bigcap_{x\in \Omega}U(x)\subseteq \cl K(\beta S)$.
\begin{itemize}
\item[(a)] There exists $e\in R$ such that $es=s$ for every $s\in S$.
\item[(b)] $S$ contains a left cancelable element.
\end{itemize}
\end{corollary}
\begin{proof} It follows from Lemma \ref{bigcapclUx} that
(a) implies that $\bigcap_{x\in \Omega}U(x)\subseteq \cl K(\beta S)$.
So assume that $s$ is a left cancelable element in $S$ and let
$p\in\bigcap_{x\in \Omega}U(x)$. By \cite[Lemma 8.1]{HS}, $s$ is left cancelable in
$\beta S$.
By Lemma \ref{bigcapclUx}, $sp\in \cl K(\beta S)$. Now
$s\beta S=\overline{sS}$ is clopen in $\beta S$.
So $sp\in \cl(K(\beta S)\cap s\beta S)$.
We claim that, if $q\in K(\beta S)\cap s\beta S$, then $q\in sK(\beta S)$.
To see this, suppose that $q\in K(\beta S)$ and that $q=sv$ for some $v\in \beta S$.
There is an idempotent $u\in K(\beta S)$ for which $qu=q$. This
implies that $sv=svu$ and hence that $v=vu\in K(\beta S)$.
So $sp\in \cl\big(sK(\beta S)\big)= s\cl K(\beta S)$ and hence $p\in \cl K (\beta S)$.
\end{proof}
\begin{corollary}\label{bigcapUandclK} Assume that $S$ is a left ideal of $R$.
Then each of the hypotheses (a) and (b) of Corollary {\rm \ref{clKbS}} implies that
$\bigcap_{x\in \Omega}U(x)=\cl K(\beta S)$.
\end{corollary}
\begin{proof} Assume that one of the hypotheses of Corollary \ref{clKbS} holds.
Then $$\textstyle\bigcap_{x\in \Omega}U(x)\subseteq \cl K(\beta S)\,.$$
To see that $\cl K(\beta S)\subseteq\bigcap_{x\in \Omega}U(x)$, let
$x\in\Omega$ be given. By Lemma \ref{UequalsN}, $U(x)=N(x)$ and so $U(x)$ is closed.
By Corollary \ref{U(x)ideal}(3), $K(\beta S)\subseteq U(x)$ and hence $\cl K(\beta S)\subseteq U(x)$.
\end{proof}
For the statement of the following corollary we depart from
our standing assumptions about $R$, $S$, and $(Z,\langle T_s\rangle_{s\in S})$.
\begin{corollary} \label{bigcapequal} Let $S$
be a semigroup. There exist a dynamical system\break
$(X,\langle T_s\rangle_{s\in S})$ and a subset
$M$ of $X$ such that $K(\beta S)=\bigcap_{x\in X}U(x)$
and $\cl K(\beta S)\break =\bigcap_{x\in M}U(x)$.
\end{corollary}
\begin{proof} If $S$ has a left identity, let $R=S$.
Otherwise, let $R=S\cup\{e\}$ where $e$ is an identity
adjoined to $S$. The conclusion then follows from
Theorem \ref{bigcapUx} and Corollary\ref{bigcapUandclK}.
\end{proof}
In the proof of the above corollary, we could have simply
let $R=S\cup\{e\}$ where $e$ is an identity
adjoined to $S$, regardless of whether $S$ has a left identity,
as was done in \cite[Theorem 19.27]{HS} to produce a dynamical
system for any semigroup $S$ establishing the equivalence
of the notions of {\it central\/} and {\it dynamically central\/}.
We shall investigate the relationship between the systems
with phase space $X=\xtotwo{R}$ and $Y=\xtotwo{S}$ in the next
section.
We note that it is possible that
$\bigcap_{x\in Z}U(x)\neq K(\beta S)$ and there is no subset $M$
of $Z$ such that $\bigcap_{x\in M}U(x)=\cl K(\beta S)$. To see this,
let $S$ be an infinite zero semigroup. That is, there is an element
$0\in S$ such that $st=0$ for all $s$ and $t$ in $S$. Then
$pq=0$ for all $p$ and $q$ in $\beta S$ and so
$\cl K(\beta S)=K(\beta S)=\{0\}$. Let $R=S$. Given
$x\in T$, if $a=x(0)$, then for all $p\in \beta S$,
$T_p(x)$ is constantly equal to $a$ and so $T_p(x)$ is uniformly recurrent.
That is, for any $x\in Z$, $U(x)=\beta S$.
In \cite{DH} it was shown that $\cl K(\beta\ben)$ is the intersection
of all of the closed two sided ideals that strictly contain it. In a
similar vein, we would like to show that each $U(x)$ properly contains
$K(\beta S)$. One cannot hope for this to hold in general. For
example, as we have already noted, if $S$ is either left zero or right zero then so is
$\beta S$ and then $K(\beta S)=\beta S$.
Results establishing that $U(x)$ properly contains $K(\beta S)$
require some weak cancellation assumptions.
\begin{definition}\label{solutionsets} Let $S$ be a semigroup
and let $A\subseteq S$.
\begin{itemize}
\item[(a)] $A$ is a {\it left solution set\/} if and only if
there exist $u$ and $v$ in $S$ such that $A=\{x\in S:ux=v\}$.
\item[(b)] $A$ is a {\it right solution set\/} if and only if
there exist $u$ and $v$ in $S$ such that $A=\{x\in S:xu=v\}$.
\end{itemize}
\end{definition}
As is standard, we denote by $\omega$ the first infinite ordinal, which
is also the first infinite cardinal. That is, $\omega=\aleph_0$.
\begin{definition}\label{weakcanc} Let $S$ be a semigroup
with $|S|=\kappa\geq\omega$.
\begin{itemize}
\item[(a)] $S$ is {\it weakly left cancellative\/} if and only if
every left solution set in $S$ is finite.
\item[(b)] $S$ is {\it weakly right cancellative\/} if and only if
every right solution set in $S$ is finite.
\item[(c)] $S$ is {\it weakly cancellative\/} if and
only if $S$ is both weakly left cancellative and weakly right cancellative.
\item[(d)] $S$ is {\it very weakly left cancellative\/} if and only if
the union of any set of fewer than $\kappa$ left solution sets has cardinality
less than $\kappa$.
\item[(e)] $S$ is {\it very weakly right cancellative\/} if and only if
the union of any set of fewer than $\kappa$ right solution sets has cardinality
less than $\kappa$.
\item[(f)] $S$ is {\it very weakly cancellative\/} if and
only if $S$ is both very weakly left cancellative and very weakly right cancellative.
\end{itemize}
\end{definition}
Given a set $X$ and a cardinal $\kappa$, we let
$U_\kappa(X)$ be the set of $\kappa$-uniform ultrafilters on $X$.
That is, $U_{\kappa}(X)=\{p\in\beta X:(\forall A\in p)(|A|\geq\kappa)\}$.
\begin{theorem}\label{proper} Assume that $|R|=|S|=\kappa\geq\omega$,
$S$ is very weakly cancellative, and has the property that
$|\{e\in S:(\exists s\in S)(es=s)\}|<\kappa$. Then for all $x\in Z$,
$U(x)\cap U_\kappa(S)\setminus \cl K(\beta S)\neq\emp$.
\end{theorem}
\begin{proof} Let $E=\{e\in S:(\exists s\in S)(es=s)\}$.
Let $x\in Z$ and pick $q\in K(\beta S)$. Let $y=T_q(x)$. By
Corollary \ref{U(x)ideal}(3), $y$ is uniformly recurrent. For
each $F\in\pf(R)$, let $B_F=\big\{s\in S:(\forall t\in F)\big(x(ts)=y(t)\big)\big\}$.
Since $$\textstyle B_F=\{s\in S:T_s(x)\in\bigcap_{t\in F}\pi_t^{-1}[\{y(t)\}]\}\,,$$ we have
$B_F\in q$. By \cite[Lemma 6.34.3]{HS}, $K(\beta S)\subseteq U_\kappa(S)$ and so $|B_F|=\kappa$.
Note that if $F\subseteq H$, then $B_H\subseteq B_F$.
Enumerate $\pf(R)$ as $\langle F_\alpha\rangle_{\alpha<\kappa}$.
Choose $t_0\in B_{F_0}\setminus E$. Let $0<\alpha<\kappa$ and assume that
we have chosen $\langle t_\delta\rangle_{\delta<\alpha}$ satisfying
the following inductive hypotheses.
\begin{itemize}
\item[(1)] For each $\delta<\alpha$, $t_\delta\in B_{F_\delta}$.
\item[(2)] For each $\delta<\alpha$, $FP(\langle t_\beta\rangle_{\beta\leq\delta})\cap E=\emp$.
\item[(3)] For each $\delta<\alpha$, if $\delta>0$, then
$t_\delta\notin FP(\langle t_\beta\rangle_{\beta<\delta})$.
\item[(4)] For each $\delta<\alpha$, if $\delta>0$,
$s\in FP(\langle t_\beta\rangle_{\beta<\delta})$, and
$\gamma<\delta$, then $st_\delta\neq t_\gamma$.
\end{itemize}
The hypotheses are satisfied for $\delta=0$. Let
\begin{eqnarray*}
M_0&=&\textstyle \{t\in S:\big(\exists H\in\pf(\alpha)\big)\big((\prod_{\beta\in H}t_\beta)t\in E\big)\big\}\}
\hbox{\rm\ and let}\\
M_1&=& \{t\in S:\big(\exists s\in FP(\langle t_\beta\rangle_{\beta<\alpha})\big)
(\exists \gamma<\alpha)(st=t_\gamma)\}\,.
\end{eqnarray*}
Note that
$|FP(\langle t_\beta\rangle_{\beta<\alpha})|\leq |\pf(\alpha)|<\kappa$.
Also, given $H\in\pf(\alpha)$ and $s\in E$, $\{t\in S:(\prod_{\beta\in H}t_\beta)t=s\}$
is a left solution set so $|M_0|<\kappa$. Note also that,
given $s\in FP(\langle t_\beta\rangle_{\beta<\alpha})$ and $\gamma<\alpha$,
$\{t\in S:st=t_\gamma\}$ is a left solution set so $|M_1|<\kappa$. Thus
we may choose $t_\alpha\in B_{F_\alpha}\setminus(E\cup FP(\langle t_\beta\rangle_{\beta<\alpha})
\cup M_0\cup M_1)$. The induction hypotheses are satisfied for $\alpha$.
Let $B=\{t_\alpha:\alpha<\kappa\}$ and let $C=\bigcap_{\alpha<\kappa}\cl FP(\langle t_\beta\rangle_{\alpha<\beta<\kappa})$.
By \cite[Theorem 4.20]{HS}, $C$ is a compact subsemigroup of $\beta S$. We claim that
$\overline B\cap K(C)=\emp$. Suppose instead that we have
$p\in \overline B\cap K(C)$. Pick $r\in K(C)$ such that $p=pr$. (By \cite[Lemma 1.30]{HS}, an
idempotent in the minimal left ideal $L$ of $C$ in which $p$ lies will do.)
Let $D=\{s\in S:s^{-1}B\in r\}$. Then $D\in p$ so $D\cap B\neq\emp$ so
pick $\alpha<\kappa$ such that $t_\alpha^{-1} B\in r$.
Then $\overline{t_\alpha^{-1} B}\cap FP(\langle t_\beta\rangle_{\alpha<\beta<\kappa})\neq
\emp$ so pick finite $H\subseteq\{\beta:\alpha<\beta<\kappa\}$ such that
$\prod_{\beta\in H}t_\beta\in t_{\alpha}^{-1}B$. Pick
$\gamma<\kappa$ such that $t_\alpha\prod_{\beta\in H}t_\beta=t_\gamma$. Let
$\max H=\mu$ and let $K=H\setminus\{\mu\}$.
If $K=\emp$, then $t_\alpha t_\mu=t_\gamma$. If $K\neq\emp$, then
$t_\alpha(\prod_{\beta\in K}t_\beta) t_\mu=t_\gamma$.
If $\gamma>\mu$ we get
a contradiction to hypothesis (3). If $\mu=\gamma$ we either
get $t_\alpha\in E$ or $t_\alpha\prod_{\beta\in K}t_\beta\in E$, contradicting
hypothesis (2). If $\gamma<\mu$ we get a contradiction to hypothesis (4).
Thus $\overline B\cap K(C)=\emp$ as claimed.
Now we claim that $\overline B\cap K(\beta S)=\emp$. Suppose
instead we have $p\in\overline B\cap K(\beta S)$. By \cite[Lemma 6.34.3]{HS}
we have that $p\in U_\kappa(S)$ and consequently, $p\in C$.
Thus $K(\beta S)\cap C\neq\emp$ and so, by \cite[Theorem 1.65]{HS},
$K(C)=K(\beta S)\cap C$, contradicting the fact that $\overline B\cap K(C)=\emp$.
Since $\overline B$ is clopen, we thus have $\overline B\cap \cl K(\beta S)=\emp$.
Now let ${\cal C}=\{B_F:F\in\pf(S)\}\cup\{B\}$. We claim that
${\cal C}$ has the $\kappa$-uniform finite intersection property.
To see this, let ${\cal F}\in\pf\big(\pf(S)\big)$ and let
$H=\bigcup{\cal F}$. If $\delta<\kappa$ and
$H\subseteq F_\delta$, then $t_\delta\in B\cap\bigcap_{F\in {\cal F}}B_F$.
Since $|\{\delta<\kappa:H\subseteq F_\delta\}|=|\{F\in\pf(S):H\subseteq F\}|=\kappa$,
we have that $|\bigcap{\cal C}|=\kappa$ as required. Pick by \cite[Corollary 3.14]{HS}
$p\in U_\kappa(S)$ such that ${\cal C}\subseteq p$.
Since $B_F\in p$ for each $F\in\pf(R)$, we have $T_p(x)=y$ so $p\in U(x)$.
Since $B\in p$, $p\notin \cl K(\beta S)$.\end{proof}
\begin{corollary}\label{rightcanc}
Assume that $|R|=|S|=\kappa\geq\omega$ and that
$S$ is right cancellative and very weakly left cancellative.
Then for all $x\in Z$,
$U(x)\cap U_\kappa(S)\setminus \cl K(\beta S)\neq\emp$.
\end{corollary}
\begin{proof} Let $E=\{e\in S:(\exists s\in S)(es=s)\}$.
It suffices to show that $|E|<\kappa$. Pick
$x\in S$. Given $e\in E$ and $s\in S$ such that $es=s$,
we have that $xes=xs$ so $xe=x$. Thus
$E$ is contained in the left solution set $\{t\in S:xt=x\}$.
\end{proof}
\begin{corollary}\label{Kproper}
Assume that $|R|=|S|=\kappa\geq\omega$, that
$S$ is very weakly cancellative, that $S$ has a member $e$ such that $es=s$ for all $s\in S$,
and $|\{e\in S:(\exists s\in S)(es=s)\}|<\kappa$. Then $K(\beta S)=\bigcap_{x\in Z}U(x)$ and
for each $x\in Z$, $U(x)$ properly contains $K(\beta S)$.
\end{corollary}
\begin{proof} This is an immediate consequence of Theorems \ref{bigcapUx} and
\ref{proper}. \end{proof}
\begin{corollary}\label{clKproper}
Assume that $S$ is a left ideal
of $R$, $|R|=|S|=\kappa\geq\omega$, $S$ is very weakly cancellative,
$S$ has a member $e$ such that $es=s$ for all $s\in S$,
and $|\{e\in S:(\exists s\in S)(es=s)\}|<\kappa$. Then $\cl K(\beta S)=\bigcap_{x\in \Omega}U(x)$ and
for each $x\in \Omega$, $U(x)$ properly contains $\cl K(\beta S)$.
\end{corollary}
\begin{proof} By Corollary \ref{bigcapUandclK} $\cl K(\beta S)=\bigcap_{x\in \Omega}U(x)$.
By Theorem \ref{proper}, for each $x\in \Omega$, $U(x)$ properly contains $\cl K(\beta S)$.
\end{proof}
\section{Relations between systems with phase spaces $X$ and $Y$}
Throughout this section we will let $S$ be an arbitrary semigroup and
let $Q=S\cup\{e\}$, where $e$ is an identity adjoined to $S$, even
if $S$ already has an identity. We will let $(X,\langle T_{X,s}\rangle_{s\in S})$
be the dynamical system of Lemma {\rm \ref{stotwodyn}} determined by $R=Q$
let $(Y,\langle T_{Y,s}\rangle_{s\in S})$
be the dynamical system of Lemma {\rm \ref{stotwodyn}} determined by $R=S$.
For $x\in X$ we will let $U_X(x)=\{p\in\beta S:T_{X,p}(x)$ is uniformly recurrent$\}$
and let $U_Y(x)=\{p\in\beta S:T_{Y,p}(x)$ is uniformly recurrent$\}$.
We have from the results of the previous section that for any semigroup
$S$, $K(\beta S)=\bigcap_{x\in X}U_X(x)$ and $\cl K(\beta S)=\bigcap_{x\in \Omega_X}U_X(x)$.
We are interested in determining when the corresponding assertions hold with
respect to $Y$. Of course, the simplest situation in which they do is
when for each $x\in X$, $U_X(x)=U_Y(x_{|S})$ so we address this problem first,
beginning with the following simple observation.
\begin{lemma}\label{uxinuy} Let $x\in X$. Then $U_X(x)\subseteq U_Y(x_{|S})$.
\end{lemma}
\begin{proof} Let $y=x_{|S}$ and note that
$\widetilde y$ is the restriction of $\widetilde x$ to $\beta S$.
Let $L$ be a minimal left ideal of $\beta S$.
By Lemma \ref{charUx},
$p\in U_X(x)$ if and only if there exists $q\in L$, such
that $\widetilde x(tp)=\widetilde x(tqp)$ for all $t\in Q$. And $p\in U_Y(x_{|S})$
if and only if there exists $q\in L$ such that $\widetilde y(tp)=\widetilde y(tqp)$ for
all $t\in S$. \end{proof}
\begin{theorem}\label{uxisuy} The following statements are equivalent.
\begin{itemize}
\item[(a)] For all $x\in X$, $U_X(x)=U_Y(x_{|S})$.
\item[(b)] There do not exist $p\in\beta S$ and $x\in X$ such
that $T_{X,p}(x)$ is the characteristic function of $\{e\}$ in $X$.
\item[(c)] For every $p\in\beta S$, $p\in\beta Sp$.
\end{itemize}
\end{theorem}
\begin{proof} Assume that (a) holds and suppose we have
$p\in\beta S$ and $x\in X$ such
that $T_{X,p}(x)$ is the characteristic function of $\{e\}$ in $X$. Then $T_{Y,p}(x_{|S})$ is constantly
$0$ so $p\in U_Y(x_{|S})$. But
$V=\{u\in X:w(e)=1\}$
is a neighborhood of $w=T_{X,p}(x)$ in $X$, while
$\{s\in S:T_{X,s}(w)\in V\}=\emp$, so
$p\notin U_X(x)$.
To see that (b) implies (c), assume that (b) holds and suppose that
we have some $p\in\beta S$ such that $p\notin\beta Sp$. Since
$\beta Sp=\rho_p[\beta S]$, $\beta Sp$ is closed. Pick
$A\in p$ such that $\overline A\cap \beta Sp=\emp$. Let $x$ be the
characteristic function of $A$ in $X$.
First let $s\in S$.
Then $sp\notin \overline A$ so $s^{-1}(S\setminus A)\in p$ so to
see that $T_{X,p}(s)=0$, it
suffices to observe that $s^{-1}(S\setminus A)\subseteq \{t\in S:T_{X,t}(x)(s)=0\}$.
Since $A\in p$ and for $t\in A$, $T_{X,t}(x)(e)=x(t)=1$, we have that
$T_{X,p}(x)(e)=1$.
By Lemma \ref{uxinuy}, we have $U_X(x)\subseteq U_Y(x_{|S})$ for all $x\in X$, so
to show that (c) implies (a), it suffices to let $x\in X$, let $p\in U_Y(x_{|S})$, assume
that $p\in \beta Sp$, and show
that $p\in U_X(x)$. By Lemma \ref{charUxp}, it suffices to let $L$ be a
minimal left ideal of $\beta S$ and let $F\in\pf(Q)$ and show that there
is some $q\in L$ such that $\widetilde x(tp)=\widetilde x(tqp)$ for every
$t\in F$. For $t\in F$, let $B_t=\{s\in S:
x(ts)=\widetilde x(tp)\}$. Then $\bigcap_{t\in F}B_t\in p$
and $p\in\beta Sp=\cl(Sp)$ so pick $v\in S$ such that
$\bigcap_{t\in F}B_t\in vp$. Let $y=x_{|S}$. Now $Fv\in\pf(S)$ so pick by Lemma \ref{charUxp}
$q\in L$ such that for all $t\in F$,
$\widetilde y(tvp)=\widetilde y(tvqp)$.
Let $q'=vq$ and note that $q'\in L$. Let $t\in F$ be given.
Then $B_t\in vp$ so $\widetilde x(tvp)=\widetilde x(tp)$ and thus
$\widetilde x(tp)=\widetilde y(tvp)=\widetilde y(tvqp)=
\widetilde x(tq'p)$. \end{proof}
\begin{corollary}\label{corKisU} If for all $p\in\beta S$, $p\in\beta Sp$,
then $K(\beta S)=\bigcap_{x\in Y}U_Y(x)$ and
$\cl K(\beta S)=\bigcap_{x\in \Omega_Y}U_Y(x)$.
\end{corollary}
\begin{proof} The first assertion is an immediate consequence of Theorems \ref{bigcapUx} and \ref{uxisuy}.
The second assertion follows from Corollary \ref{bigcapUandclK} and Theorem \ref{uxisuy}.
\end{proof}
We have already mentioned the problem of determining whether
$K(\beta S)$ or $\cl K(\beta S)$ is prime. Recall that
an ideal $I$ in a semigroup is {\it semiprime\/} if and only if
whenever $ss\in I$, one must have $s\in I$.
\begin{corollary}\label{primeK} \begin{itemize}
\item[(1)] If $K(\beta S)\neq \bigcap_{x\in Y}U_Y(x)$, then
$K(\beta S)$ is not semiprime.
\item[(2)] If $\cl K(\beta S)\neq \bigcap_{x\in \Omega_Y}U_Y(x)$, then
$\cl K(\beta S)$ is not semiprime.
\end{itemize}
\end{corollary}
\begin{proof}
(1) If $p\in \bigcap_{x\in Y}U_Y(x)\setminus K(\beta S)$, then $pp\in\beta Sp$ and by
Theorem \ref{bigcapUx}, $\beta Sp\subseteq K(\beta S)$.
(2) If $p\in \bigcap_{x\in \Omega_Y}U_Y(x)\setminus \cl K(\beta S)$, then $pp\in\beta Sp$
and by Lemma \ref{bigcapclUx}, $\beta Sp\subseteq \cl K(\beta S)$.
\end{proof}
By virtue of Theorem \ref{uxisuy} we are interested in knowing when there
is some $p\in\beta S$ such that $p\notin \beta Sp$.
\begin{lemma}\label{lempnotin} Let $p\in\beta S$. Then $p\notin \beta Sp$ if and only if
there exists $A\subseteq S$ such that for all $x\in S$, $x^{-1}A\in p$
and $A\notin p$.
\end{lemma}
\begin{proof} Let $C(p)=\{A\subseteq S:(\forall x\in S)(x^{-1}A\in p)\}$.
By \cite[Theorem 6.18]{HS}, $p\in \beta Sp$ if and only if $C(p)\subseteq p$.
\end{proof}
\begin{theorem}\label{thmpnotin} Assume that
$|S|=\kappa\geq\omega$. There exists $p\in\beta S$ such that
$p\notin \beta Sp$ if and only if there exists
$\langle y_F\rangle_{F\in\pf(S)}$ in $S$ such that\hfill\break
$\{y_F:F\in\pf(S)\}\cap\bigcup\{Fy_F:F\in\pf(S)\}=\emp$.
\end{theorem}
\begin{proof} Necessity. Pick $p\in\beta S$ such that
$p\notin \beta Sp$. By Lemma \ref{lempnotin}, pick $A\subseteq S$ such that
for all $x\in S$, $x^{-1}A\in p$ and $A\notin p$. For
$F\in\pf(S)$ pick $y_F\in (S\setminus A)\cap\bigcap_{x\in F}x^{-1}A$.
Sufficiency. Let $A=\bigcup\{Fy_F:F\in\pf(S)\}$. Then
$\{S\setminus A\}\cup\{x^{-1}A:x\in S\}$ has the finite intersection
property so pick $p\in\beta S$ such that $\{S\setminus A\}\cup\{x^{-1}A:x\in S\}\subseteq
p$. By Lemma \ref{lempnotin}, $p\notin \beta Sp$. \end{proof}
One of the assumptions in the following corollary is that $S^*=\beta S\setminus S$ is a right ideal
of $\beta S$. A (not very simple) characterization of when $S^*$ is a right ideal
of $\beta S$ is given in \cite[Theorem 4.32]{HS}. By \cite[Corollary 4.33 and Theorem 4.36]{HS} it is
sufficient that $S$ be either right cancellative or weakly cancellative.
\begin{corollary}\label{corpnotin} Assume that $|S|=\kappa\geq\omega$
and assume that $$|S\setminus\{t\in S:(\exists s\in S)(st=t)\}|=\kappa\,.$$ If
either $S^*$ is a right ideal of $\beta S$ or $S$ is very weakly left cancellative,
then there exists $p$ in $\beta S$ such that $p\notin \beta Sp$.
\end{corollary}
\begin{proof} Assume first that $S^*$ is a right ideal of $\beta S$, and pick
$t\in S$ such that there is no $s\in S$ with $st=t$. Then $t\notin St$ and
$t\notin S^* t$.
Now assume that $S$ is very weakly left cancellative.
Enumerate $\pf(S)$ as $\langle F_\alpha\rangle_{\alpha<\kappa}$. By
Theorem \ref{thmpnotin}, it suffices to produce $\langle t_\alpha\rangle_{\alpha<\kappa}$ in
$S$ such that\break $\{t_\alpha:\alpha<\kappa\}\cap\bigcup\{F_\alpha t_\alpha:\alpha<\kappa\}=\emp$.
Let $E=\{t\in S:(\exists s\in S)(st=t)\}$. Pick $t_0\in S\setminus E$. Let $0<\alpha<\kappa$ and assume
we have chosen $\langle t_\delta\rangle_{\delta<\alpha}$ in $S\setminus E$ such that
if $\delta>0$, then $t_\delta\notin\bigcup_{\mu<\delta}F_{\mu} t_\mu$ and for each
$x\in F_\delta$, $xt_\delta\notin\{t_\mu:\mu<\delta\}$.
For $x\in S$ and $\mu<\alpha$, let $H_{x,\mu}=\{t\in S:xt=t_\mu\}$. Then
each $H_{x,\mu}$ is a left solution set so $|\bigcup\{H_{x,\mu}:x\in F_\alpha\hbox{ and }
\mu<\alpha\}|<\kappa$. Pick
$$\textstyle t_\alpha\in S\setminus (E\cup\bigcup\{H_{x,\mu}:x\in F_\alpha\hbox{ and }
\mu<\alpha\}\cup\bigcup_{\mu<\alpha}F_\mu t_\mu)\,.$$
Suppose we have some $\mu<\kappa$ such that
$t_\mu\in \bigcup\{F_\alpha t_\alpha:\alpha<\kappa\}$ and pick
$\alpha<\kappa$ and $x\in F_\alpha$ such that
$t_\mu=xt_\alpha$. Then $\alpha\neq\mu$ because $t_\alpha\notin E$.
If $\alpha<\mu$, we would have $t_\mu\in F_\alpha t_\alpha$.
So we must have $\mu<\alpha$. But then $t_\alpha\in H_{x,\mu}$, a contradiction.
\end{proof}
We conclude this section by exhibiting a sufficient condition which guarantees
$K(\beta S)=\bigcap_{x\in Y}U_Y(x).$ We shall see that this
does not require equality between $U_X(x)$ and $U_Y(x_{|S})$ for
all $x\in X$.
\begin{theorem}\label{sAimplies} Assume that for all
$p\in\bigcap_{x\in Y}U_Y(x)$ and all $A\in p$
the assumption that $\{t\in S:t^{-1}sA\in p\}$ is syndetic for every $s\in S$,
implies that $\{t\in S:t^{-1}A\in p\}\neq \emptyset$. Then
$K(\beta S)=\bigcap_{x\in Y}U_Y(x)$.\end{theorem}
\begin{proof} Assume that $p\in \bigcap_{x\in Y}U_Y(x)\setminus K(\beta S)$. By Theorem \ref{bigcapUx}(2),
$\beta Sp\subseteq K(\beta S)$ so $p\notin \beta Sp$. Pick $A\in p$ such that $\overline A\cap \beta Sp=\emptyset$.
Thus $\{t\in S:t^{-1}A\in p\}=\emptyset$.
We claim that for all $s\in S$, $\{t\in S:t^{-1}sA\}$ is syndetic. So let $s\in S$.
By \cite[Theorem 4.48]{HS} it suffices to let $L$ be a minimal left ideal of $\beta S$
and show that there is some $q\in L$ such that $\{t\in S:t^{-1}sA\in p\}\in q$.
By Theorem \ref{bigcapUx}(2), $sp\in Lp$ so pick $q\in L$ such that $sp=qp$.
Then $sA\in qp$ so $\{t\in S:t^{-1}sA\in p\}\in q$ as required.
\end{proof}
Note that by Theorem \ref{bigcapUx}(3),
$K(\beta \ben,+)=\bigcap_{x\in Y}U_Y(x)$ while $1\notin \beta \ben+1$ so by
Theorem \ref{uxisuy}, it is not the case that for all $x\in X$, $U_X(x)=U_Y(x_{|S})$.
On the other hand, given $p\in K(\beta\ben,+)$ one has $p=q+p$ for some
$p\in K(\beta\ben,+)$ so automatically for any $A\in p$,
$\{t\in \ben:-t+A\in p\}\neq \emptyset$ so the hypotheses of Theorem \ref{sAimplies} are valid.
\section{Recurrence and surjectivity of $T_p$}
So far in this paper we have been considering the notion
of uniform recurrence. We now introduce a notion which is usually weaker.
\begin{definition}\label{defrecur} Let $(X,\langle T_s\rangle_{s\in S})$ be a
dynamical system. The point $x\in X$ is {\it recurrent\/} if and only if
for each neighborhood $V$ of $x$ in $X$, $\{s\in S:T_s(x)\in V\}$ is infinite.
\end{definition}
If all syndetic subsets of a semigroup $S$ are infinite, then any
uniformly recurrent point of $X$ is recurrent. This is not always the
case. For example, if $S$ is a left zero semigroup and $x\in S$,
then $x$ is uniformly recurrent in the dynamical system $(\beta S, \langle \lambda_s\rangle_{s\in S})$
but is not recurrent. (We have that $\{x\}$ is a neighborhood of $x$ and
$\big\{s\in S:\lambda_s(x)\in\{x\}\big\}=\{x\}$, which is syndetic, but finite.)
The following characterization of recurrence is very similar to the
characterization of uniform recurrence in \cite[Theorem 19.23]{HS}.
Part of the results depend on the assumption that $S^*$ is a subsemigroup of $\beta S$.
There is a characterization of $S^*$ as a subsemigroup in \cite[Theorem 4.28]{HS}.
By \cite[Corollary 4.29 and Theorem 4.31]{HS} it is sufficient that
$S$ be right cancellative or weakly left cancellative.
\begin{theorem}\label{charrec} Let $(X,\langle T_s\rangle_{s\in S})$ be a
dynamical system. Statements (a) and (b) are equivalent and imply
statements (c) and (d), which are equivalent. If $S^*$ is a subsemigroup
of $\beta S$, then all four statements are equivalent.
\begin{itemize}
\item[(a)] There exists an idempotent $p\in S^*$ such that
$T_p(x)=x$.
\item[(b)] There exist $y\in X$ and an idempotent $p\in S^*$ such that
$T_p(y)=x$.
\item[(c)] There exists $p\in S^*$ such that
$T_p(x)=x$.
\item[(d)] $x$ is recurrent.
\end{itemize}\end{theorem}
\begin{proof} Trivially (a) implies (b) and (a) implies (c). To see that
(b) implies (a), pick $y\in X$ and an idempotent $p\in S^*$ such that
$T_p(y)=x$. Then $x=T_p(y)=T_{pp}(y)=T_p\big(T_p(y)\big)=T_p(x)$.
To see that (c) implies (d), pick $p\in S^*$ such that
$T_p(x)=x$. Let $V$ be a neighborhood of $x$. Then
$\{s\in S:T_s(x)\in V\}\in p$ so $\{s\in S:T_s(x)\in V\}$ is infinite.
To see that (d) implies (c), assume that $x$ is recurrent and for each neighborhood $V$ of $x$,
let $D_V=\{s\in S:T_s(x)\in V\}$. Then any finite subfamily of
$\{D_V:V$ is a neighborhood of $x\}$ has infinite intersection so
pick by \cite[Corollary 3.14]{HS} some $p\in S^*$ such that
$\{D_V:V$ is a neighborhood of $x\}\subseteq p$. Then $T_p(x)=x$.
Now assume that $S^*$ is a semigroup. To see that (c) implies (a), pick $p\in S^*$ such that
$T_p(x)=x$ and let $E=\{q\in S^*:T_q(x)=x\}$.
Since $S^*$ is a subsemigroup of $\beta S$, we have that
$E$ is a subsemigroup of $\beta S$. We claim that
$E$ is closed. To see this, let $q\in \beta S\setminus E$.
If $q\in S$, then $\{q\}$ is a neighborhood of $q$ missing $E$,
so assume that $q\in S^*$. Pick an open neighborhood $U$ of $T_q(x)$
such that $x\notin \cl U$ and let $A=\{s\in S:T_s(x)\in U\}$.
Then $\overline A$ is a neighborhood of $q$ which misses $E$.
Since $E$ is a compact right topological semigroup, there
is an idempotent in $E$.
\end{proof}
Recall that in any dynamical system, $(X,\langle T_s\rangle_{s\in S})$,
$K(\beta S)\subseteq\bigcap_{x\in X}U_X(x)$ and we have obtained
sufficient conditions for equality.
\begin{theorem}\label{qpinK} Let $(X,\langle T_s\rangle_{s\in S})$ be a dynamical system,
let $p\in \beta S$, and assume that $T_p:X\to X$ is surjective and
$K(\beta S)=\bigcap_{x\in X}U_X(x)$. Then for any $q\in\beta S$,
$qp\in K(\beta S)$ if and only if $q\in K(\beta S)$.
\end{theorem}
\begin{proof} Let $q\in\beta S$. The sufficiency is trivial, so
assume that $qp\in K(\beta S)$. It suffices to show
that $q\in \bigcap_{x\in X}U(x)$, so let $x\in X$ be given.
Pick $y\in X$ such that $T_p(y)=x$.
Then $T_q(x)=T_q\big(T_p(y)\big)=T_{qp}(y)$. Since
$qp\in U(y)$ we have $T_{qp}(y)$ is uniformly recurrent, and
so $T_q(x)\in U(x)$ as required.\end{proof}
\begin{definition}\label{defNS} Let $(X,\langle T_s\rangle_{s\in S})$ be a dynamical system.
Then $NS=NS_X=\{p\in\beta S:T_p$ is not surjective$\}$.
\end{definition}
We have seen that $U(x)$ is always a left ideal of $\beta S$.
\begin{lemma}\label{NSright} Let $(X,\langle T_s\rangle_{s\in S})$ be a dynamical system.
If $NS\neq\emp$, then $NS$ is a right ideal of $\beta S$.
\end{lemma}
\begin{proof} Given $p\in NS$ and $q\in \beta S$, the
range of $T_{pq}$ is contained in the range of $T_p$. \end{proof}
\begin{lemma}\label{idemNS} Let $(X,\langle T_s\rangle_{s\in S})$ be a dynamical system.
If there is some $x\in X$ such that $x$ is not recurrent, then
$\{p\in S^*:pp=p\}\subseteq NS$.
\end{lemma}
\begin{proof} Pick $x\in X$ such that $x$ is not recurrent and
let $p$ be an idempotent in $S^*$. We claim that
$x$ is not in the range of $T_p$, so suppose instead we have
$y\in X$ such that $T_p(y)=x$. Then by
Theorem \ref{charrec}, $x$ is recurrent.\end{proof}
We shall establish a strong connection between the surjectivity of $T_p$
and $p$ being right cancelable in $\beta S$. The
purely algebraic result in Theorem \ref{rpcancel} will be useful.
\begin{lemma}\label{satb} Let $S$ be a countable right cancellative and
weakly left cancellative semigroup and let $B$ be an infinite subset of $S$.
There is an infinite subset $D$ of $B$ with the property that whenever $s$ and $t$ are distinct
members of $S$, there is a finite subset $F$ of $D$ such that $sa\neq tb$ whenever $a,b\in D\setminus F$.
\end{lemma}
\begin{proof} Let $\Delta=\{(s,s):s\in S\}$ and enumerate $(S\times S)\setminus\Delta$ as
$\langle (s_n,t_n)\rangle_{n=1}^\infty$. Pick $a_1\in B$. Assume $n\in \ben$ and we
have chosen $\langle a_i\rangle_{i=1}^n$. Let $W_n=\{b\in S:$ there exist $i,j\in\nhat{n}$
such that $s_ia_j=t_ib$ or $s_ib=t_ia_j\}$. Then $W_n$ is the union of finitely many left solution sets, so
is finite. Pick $a_{n+1}\in B\setminus(W_n\cup\{a_1,a_2,\ldots,a_n\})$.
Let $D=\{a_n:n\in\ben\}$. Let $s$ and $t$ be distinct members of $S$ and pick $n$ such that
$(s,t)=(s_n,t_n)$. Let $F=\big\{a_i:i\in\nhat{n}\big\}$. To see that $F$ is as required,
let $a,b\in D\setminus F$ and suppose $sa=tb$. Then by right cancellation, $a\neq b$.
Pick $m>n$ and $r>n$ such that $a=a_m$ and $b=a_r$. If $m<r$, then $a_r\in W_{r-1}$.
If $r<m$, then $a_m\in W_{m-1}$.
\end{proof}
\begin{theorem}\label{rpcancel} Let $S$ be a countable cancellative semigroup. If $p\in \beta S\setminus K(\beta S)$,
then there exists an infinite $D\subseteq S$ such that for every $r\in D^*$, $rp$ is right cancelable
in $\beta S$.\end{theorem}
\begin{proof} Choose $q\in K(\beta S)$. We first claim that for each $s\in S$,
$sp\notin K(\beta S)$ and in particular, $sp\notin\beta Sqp$. So suppose
we have $sp\in K(\beta S)$. Then $sp$ is in a minimal left ideal $L$ of
$\beta S$. Pick an idempotent $r\in L$. By \cite[Lemma 1.30]{HS}, $sp=spr$.
By \cite[Lemma 8.1]{HS} $s$ is left cancelable in $\beta S$ so $p=pr$, and thus
$p\in K(\beta S)$. This contradiction establishes the claim. For each
$s\in S$, pick $U_s\in sp$ such that $\overline{U_s}\cap \beta Sqp=\emptyset$.
For each $s,t\in S$, there exists $V_{s,t}\in q$
such that $\overline {U_s}\cap t\overline {V_{s,t}}p=\emptyset$
because $\lambda_t\circ\rho_p(q)\in \beta S\setminus\overline {U_s}$.
By \cite[Theorem 3.36]{HS}, there exists an infinite subset $B$
of $S$ such that $B^*\subseteq \bigcap_{s,t\in S}\,\overline{V_{s,t}}$.
Then for every $r\in B^*$ and every $s,t\in S$,
$trp\notin \overline {U_s}$.
By Lemma \ref{satb} pick an infinite subset $D$ of $B$ such that, whenever $s$ and $t$ are distinct elements
of $S$, there is a finite subset $F$ of $D$ such that $sa\neq tb$ whenever $a,b\in D\setminus F$.
Enumerate $D$ as $\langle d_n\rangle_{n=1}^\infty$ and for each distinct $s$ and $t$ in $S$, pick
$n_{s,t}\in\ben$ such that $sd_m\neq td_n$ whenever $m,n>n_{s,t}$.
We claim that, for every $r\in D^*$, $rp$ is right cancelable in $\beta S$.
We shall apply \cite[Theorem 3.40]{HS} three times.
Assume that $q_1rp=q_2rp$, where $q_1$ and $q_2$ are distinct elements of $\beta S$. Let $A_1$ and $A_2$
be disjoint subsets of $S$ which are members of $q_1$ and $q_2$ respectively. Since $q_1rp\in cl(A_1rp)$ and
$q_2rp\in cl(A_2rp)$, an application of \cite[Theorem 3.40]{HS} shows that
either $A_1rp\cap cl(A_2rp)\neq \emp$ or $A_2rp\cap cl(A_1rp)\neq \emp$, and without loss of generality,
we may assume that the former holds. Thus we have some $s\in A_1$ and $q'\in \overline{A_2}$ such that
$srp=q'rp$. Now $srp\in\cl(sDp)$ and $q'rp\in\cl\big((S\setminus\{s\})rp\big)$, so either
$sDp\cap\cl\big((S\setminus\{s\})rp\big)\neq \emp$ or $(S\setminus\{s\})rp\cap\cl(sDp)\neq\emp$.
We thus have either
\begin{itemize}
\item[(i)] $sDp\cap \cl\big((S\setminus\{s\})rp\big)\neq \emp$, in which case
we choose $d\in D$ and $y\in\beta S$ such that $sdp=yrp$; or
\item[(ii)] $sDp\cap \cl\big((S\setminus\{s\})rp\big)= \emp$, in which case we pick
$t\in S\setminus\{s\}$ and $r'\in\overline{D}$ such that $sr'p=trp$. Since $sDp\cap \cl\big((S\setminus\{s\})rp\big)= \emp$,
we have $r'\in D^*$.
\end{itemize}
Suppose that (i) holds. Then $U_{sd}\in sdp$ so $\{v\in S:v^{-1}U_{sd}\in rp\}\in y$, so
pick $v\in S$ such that $U_{sd}\in vrp$. But $r\in V_{sd,v}$, so this is a contradiction.
Thus (ii) holds.
Now $sr'p\in\cl\{sd_mp:m>n_{s,t}\}$ and
$trp\in\cl\{td_mp:m>n_{s,t}\}$ so, essentially without loss of generality,
we have $\{sd_mp:m>n_{s,t}\}\cap\cl\{td_mp:m>n_{s,t}\}\neq\emp$. (We have distinguished
between $s$ and $t$ at this stage, but the arguments below with $s$ and $t$ interchanged
remain valid.) Thus either
\begin{itemize}
\item[(iii)] there exist $m,n>n_{s,t}$ such that $sd_mp=td_np$; or
\item[(iv)] there exist $m>n_{s,t}$ and $r''\in D^*$ such that $sd_mp=tr''p$.
\end{itemize}
If (iii) holds, then by \cite[Lemma 6.28]{HS}, $sd_m=td_n$, contradicting the choice of $n_{s,t}$.
So (iv) holds. But $r''\in V_{sd_m,t}$ so $tr''p\notin U_{sd_m}$, a contradiction.
\end{proof}
We now present several results about the dynamical systems considered in Section 3.
\begin{lemma}\label{AcapbetaSp} Let $S$ be a semigroup and let $p$ be
a right cancelable element of $\beta S$.
Then for any clopen subset
$E$ of $\beta Sp$, there is some $A\subseteq S$ such that
$E=\overline A\cap \beta Sp$.
\end{lemma}
\begin{proof} Let $E$ be a clopen subset of $\beta Sp$.
Let ${\cal D}=\{\overline D\cap\beta Sp:D\subseteq S
\hbox{ and }\overline D\cap\beta Sp\subseteq E\}$. Since $\{\overline D\cap\beta Sp:D\subseteq S\}$
is a basis for the topology of $\beta Sp$ and $E$ is open in $\beta Sp$, we have that
$E=\bigcup{\cal D}$. Since $E$ is compact, pick finite ${\cal F}\subseteq{\cal P}(S)$
such that $E=\bigcup_{D\in{\cal F}}(\overline D\cap\beta Sp)$ and let
$A=\bigcup{\cal F}$.\end{proof}
\begin{theorem}\label{rtcancimpsur} Let $S$ be a semigroup. Let
$(Y,\langle T_s\rangle_{s\in S})$
be the dynamical system of Lemma {\rm \ref{stotwodyn}} determined by $R=S$. Let $p\in\beta S$.
If $p$ is right cancelable in $\beta S$, then $T_p:Y\to Y$ is surjective.
\end{theorem}
\begin{proof} Note that since $\rho_p:\beta S\to\beta Sp$ is injective and
takes closed sets to closed sets, it is a homeomorphism.
To see that $T_p$ is surjective, let $z\in Y$, let
$B=\{s\in S:z(s)=1\}$, and let $E=\rho_p[\,\overline B\,]$.
Then $E$ is clopen in $\beta Sp$ so by Lemma \ref{AcapbetaSp} pick $A\subseteq S$ such that
$E=\overline A\cap \beta Sp$. Let $x$ be the characteristic function of
$A$ in $Y$.
We claim that $T_p(x)=z$. For this, it suffices that for each
$s\in S$, $\{t\in S:T_t(x)(s)=z(s)\}\in p$. So let
$s\in S$. Note that $\{t\in S:T_t(x)(s)=1\}=\{t\in S:x(st)=1\}=s^{-1}A$.
Also $s^{-1}A\in p$ if and only if $s\in \rho_p^{-1}[\,\overline A\cap \beta Sp]$
so $s\in B$ if and only if $s^{-1}A\in p$.
If $z(s)=1$, then $s\in B$ so $s^{-1}A\in p$ so
$\{t\in S:T_t(x)(s)=z(s)\}\in p$. If $z(s)=0$, then $s\notin B$ so $s^{-1}A\notin p$ so
$\{t\in S:T_t(x)(s)=z(s)\}\in p$.
\end{proof}
Notice that the hypotheses of the following corollary hold if $S$ has any
right cancelable element.
\begin{corollary}\label{rtcanciffpsurY} Let $S$ be a semigroup. Let
$(Y,\langle T_s\rangle_{s\in S})$
be the dynamical system of Lemma {\rm \ref{stotwodyn}} determined by $R=S$. Let $p\in\beta S$.
Assume that for whenever $q$ and $r$ are distinct elements of $\beta S$, there
exists $s\in S$ such that $sq\neq sr$.
Then $T_p:Y\to Y$ is surjective if and only if
$p$ is right cancelable in $\beta S$.\end{corollary}
\begin{proof} The necessity is Theorem \ref{rtcancimpsur}.
So assume that $T_p$ is surjective and suppose that we have distinct
$q$ and $r$ in $\beta S$ such that $qp=rp$. We claim that
$T_q=T_r$. To see this, let $x\in Y$ be given. Pick $z\in Y$ such that
$T_p(z)=x$. Then $T_q(x)=T_q\big(T_p(z)\big)=T_{qp}(z)=T_{rp}(z)=
T_r\big(T_p(z)\big)=T_r(x)$.
Pick $s\in S$ such that $sq\neq sr$,
pick $A\in sq\setminus sr$, and let $x$ be the characteristic function
of $A$ in $Y$. Then $A\subseteq\{t\in S:T_t(x)(s)=1\}$ so $T_q(x)(s)=1$
and $S\setminus A\subseteq\{t\in S:T_t(x)(s)=0\}$ so $T_r(x)(s)=1$.
\end{proof}
\begin{theorem}\label{rtcanciffpsurX} Let $S$ be a semigroup and
let $Q=S\cup\{e\}$ where $e$ is an identity adjoined to $S$. Let
$(X,\langle T_s\rangle_{s\in S})$
be the dynamical system of Lemma {\rm \ref{stotwodyn}} determined by $R=Q$ and
let $p\in\beta S$. Then $T_p:X\to X$ is surjective if and only if
$p$ is right cancelable in $\beta Q$.
\end{theorem}
\begin{proof} Sufficiency. Note that $\rho_p:\beta S\to\beta Sp$ is a homeomorphism.
Note also that $p\notin \beta Sp$. (If we had $p=qp$ for some $q\in\beta S$, then
we would have $ep=qp$.) Let $x\in X$ and let $B=\{s\in S:x(s)=1\}$. By Lemma
\ref{AcapbetaSp}, pick $A\subseteq S$ such that $\rho_p[\,\overline B\,]=
\overline A\cap \beta Sp$. Pick $P\in p$ such that $\overline P\cap \beta Sp=\emp$.
If $x(e)=1$, let $D=A\setminus P$. If $x(e)=0$, let $D=A\cup B$. Let
$z$ be the characteristic function of $D$ in $X$.
We claim that $T_p(z)=x$. As in the proof of Theorem \ref{rtcancimpsur}, we
see that for $s\in S$, $T_p(z)(s)=x(s)$.
Regardless of the value of $x(e)$, we have that
$P\subseteq\{s\in S:T_s(z)(e)=x(e)\}$, so $T_p(z)(e)=x(e)$.
Necessity. Suppose that $T_p$ is surjective and we have $q\neq r$ in $\beta Q$ such that
$qp=rp$. Assume first
that $e\in \{q,r\}$, so without loss of generality, $q=e$. Let
$x$ be the characteristic function of $S$ in $X$ and pick
$z\in X$ such that $T_p(z)=x$. Then
$0=x(e)=T_p(z)(e)=T_{rp}(z)(e)=T_r\big(T_p(z)\big)(e)
=T_r(x)(e)=1$, a contradiction.
So we can assume that $q$ and $r$ are in $\beta S$. Pick
$A\in q\setminus r$ and let $A$ be the characteristic function
of $A$ in $X$. Pick $z\in X$ such that $T_p(z)=x$. Then
$0=T_r(x)(e)=T_{rp}(z)(e)=T_{qp}(z)(e)=T_q\big(T_p(z)\big)(e)
=T_q(x)(e)=1$, a contradiction.
\end{proof}
\begin{theorem}\label{Singroup}
Let $S$ be a countable semigroup which can be embedded in a group and assume that
$S$ can be enumerated as $\langle s_t\rangle_{t=0}^\infty$ so that if $u,v\in S$, $i,j\in\omega$ with
$i<j$, and $s_iu=s_jv$, then $s_0s_i^{-1}s_j\in S$.
Let $(Y,\langle T_s\rangle_{s\in S})$
be the dynamical system of Lemma {\rm \ref{stotwodyn}} determined by $R=S$ and let $p\in\beta S$.
The $T_p$ is surjective if and only if there exists $x\in Y$ such that $T_p(x)$ is the characteristic
function of $\{s_0\}$ in $Y$.\end{theorem}
\begin{proof} The necessity is trivial.
Assume that we have $x\in Y$ such that $T_p(x)$ is the characteristic
function of $\{s_0\}$ in $Y$. For $m\in\ben$, let
$D_m=\{s_0s_i^{-1}s_j:i,j\in\ohat{m}\,,\, i<j\hbox{, and }
s_0s_i^{-1}s_j\in S\}$ and note that $s_0\notin D_m$. For each
$m\in\ben$, let
$$\textstyle B_m=\{s\in S:T_s(x)\in\pi_{s_0}^{-1}[\{1\}]
\cap\bigcap_{i=1}^m\pi_{s_i}^{-1}[\{0\}]\cap\bigcap_{r\in D_m}\pi_r^{-1}[\{0\}]\}\,,$$
and note that $B_m\in p$. We claim that if $m,k\in\ben$, $u\in B_m$,
$v\in B_k$, $i\in\ohat{m}$, $j\in\ohat{k}$, and $s_iu=s_jv$, then
$i=j$. Suppose instead we have such $m,k,u,v,i,j$ with $i\neq j$ and
assume without loss of generality that $i<j$. Then $u=s_i^{-1}s_jv$.
By assumption $s_0s_i^{-1}s_j\in S$ so $s_0s_i^{-1}s_j\in D_k$.
Since $u\in B_m$, $1=T_u(x)(s_0)=x(s_0u)$. Since $v\in B_k$ and
$s_0s_i^{-1}s_j\in D_k$, $0=T_v(x)(s_0s_i^{-1}s_j)=x(s_0s_i^{-1}s_jv)$,
a contradiction.
Now to see that $T_p$ is surjective, let $y\in Y$ be given. Define $w\in Y$ as
follows. If $m\in\ben$, $u\in B_m$, and $i\in\ohat{m}$, then
$w(s_iu)=y(s_i)$. For $s\in S$ which is not of the form
$s_iu$ for some $m\in\ben$, $u\in B_m$, and $i\in\ohat{m}$, define
$w(s)$ at will. To see that $T_p(w)=y$, let $U$ be a neighborhood of
$y$. Pick $m\in\ben$ such that $\bigcap_{i=0}^m\pi_i^{-1}[\{y(s_i)\}]\subseteq U$.
Then $B_m\subseteq U$.
\end{proof}
The following is an immediate corollary of Theorem \ref{Singroup}.
\begin{corollary}\label{ingroup} Let $S$ be a countable group with identity $e$, let $(Y,\langle T_s\rangle_{s\in S})$
be the dynamical system of Lemma {\rm \ref{stotwodyn}} determined by $R=S$, and let $p\in\beta S$.
The following statements are equivalent.
\begin{itemize}
\item[(a)] $T_p$ is surjective.
\item[(b)] For each $s\in S$, there exists $x\in Y$ such that $T_p(x)$ is the characteristic
function of $\{s\}$.
\item[(c)] There exists $x\in Y$ such that $T_p(x)$ is the characteristic
function of $\{e\}$.
\end{itemize}\end{corollary}
Notice that the hypotheses of the following theorem hold if $S$ is
very weakly left cancellative and right cancellative. If $\kappa$ is regular,
the assumption that for any subset $D$ of $S$ with fewer than
$\kappa$ members, $|\{e\in S:(\exists s\in D)(\exists t\in D\setminus\{s\})(se=te)\}|<\kappa$
can be replaced by the assumption that for all distinct $s$ and $t$ in $S$,
$|\{e\in S:se=te\}|<\kappa$.
\begin{theorem}\label{vwlckappa}
Let $S$ be a semigroup with $|S|=\kappa\geq\omega$
which is very weakly left cancellative and
has the property that for any subset $D$ of $S$ with fewer than
$\kappa$ members, $|\{e\in S:(\exists s\in D)(\exists t\in D\setminus\{s\})(se=te)\}|<\kappa$.
Let $(Y,\langle T_s\rangle_{s\in S})$
be the dynamical system of Lemma {\rm \ref{stotwodyn}} determined by $R=S$.
There is a dense open subset $W$ of $U_\kappa(S)$ such that
for every $p\in W$, $p$ is right cancelable in $\beta S$ and
$T_p:Y\to Y$ is surjective.\end{theorem}
\begin{proof} We show that for any $C\in [S]^\kappa$, there exists
$B\in [C]^\kappa$ such that for every $p\in \overline B\cap U_\kappa(S)$, $p$ is right cancelable in $\beta S$ and
$T_p:Y\to Y$ is surjective.
Enumerate $S$ as $\langle s_\gamma\rangle_{\gamma<\kappa}$.
Choose $t_0\in C$. Let $0<\alpha<\kappa$ and assume that
we have chosen $\langle t_\delta\rangle_{\delta<\alpha}$ in $C$ satisfying
the following inductive hypotheses:
\begin{itemize}
\item[(1)] If $\gamma<\delta$, then $t_\gamma\neq t_\delta$.
\item[(2)] If $\gamma<\delta$, $\mu<\beta\leq \delta$, and $\mu\neq\gamma$,
then $s_\gamma t_\delta\neq s_\mu t_\beta$.
\end{itemize}
The hypotheses are satisfied for $\delta=0$.
Let $E=\{e\in S:(\exists \mu<\beta\leq \alpha)(s_\mu e=s_\beta e)\}$.
For $\mu<\beta<\alpha$ and $\gamma<\alpha$ let
$A_{\gamma,\mu,\beta}=\{t\in S:s_\gamma t=s_\mu t_\beta\}$.
Then each $A_{\gamma,\mu,\beta}$ is a left solution set.
Pick $$\textstyle t_\alpha\in C\setminus(\{t_\gamma:\gamma<\alpha\}\cup
E\cup\{\bigcup_{\gamma<\alpha}\bigcup_{\beta<\alpha}\bigcup_{\mu<\beta}A_{\gamma,\mu,\beta})\,.$$
Hypothesis (1) is trivially satisfied and if $\mu<\beta<\alpha$ and
$\gamma<\alpha$, then $t_\alpha\notin A_{\gamma,\mu,\beta}$ so
$s_\gamma t_\alpha\neq s_\mu t_\beta$. If
$\mu<\beta=\alpha$ and $\gamma<\alpha$, then $t_\alpha\notin E$ so
$s_\gamma t_\alpha\neq s_\mu t_\beta$.
Let $B=\{t_\alpha:\alpha<\kappa\}$ and let $p\in\overline B\cap U_\kappa(S)$.
To see that $p$ is right cancelable in $\beta S$, let $q\neq r\in \beta S$ and suppose
that $qp=rp$. Pick subsets $C$ and $D$ of $S$ such that $C\cap D=\emp$ and
$C\in q$ and $D\in r$. Then $H=\{s_\gamma t_\alpha:\gamma<\alpha\hbox{ and }s_\gamma\in C\}
\in qp$. (To see this, let $s_\gamma\in C$. Then $\{t_\alpha:\gamma<\alpha<\kappa\}\subseteq s_{\gamma}^{-1}H$.)
Similarly, $\{s_\mu t_\beta:\mu<\beta\hbox{ and }s_\mu\in D\}
\in rp$. Since these sets are disjoint by hypothesis (2), we have a contradiction.
The fact that $T_p$ is surjective follows from Theorem \ref{rtcancimpsur}.
\end{proof}
\begin{lemma}\label{chianotrec} Let $S$ be a cancellative semigroup, let
$a\in S$, and let $(Y,\langle T_s\rangle_{s\in S})$
be the dynamical system of Lemma {\rm \ref{stotwodyn}} determined by $R=S$.
If $x$ is the characteristic function of $\{a\}$ in $Y$, then
$x$ is not a recurrent point.
\end{lemma}
\begin{proof} We claim that $|\{s\in S:T_s(x)(a)=1\}|\leq 1$.
Indeed, if $x(as)=1$, then $as=a$ so by left cancellation, $s$
is a left identity for $S$ and then by right cancellation, $s$
is a two sided identity for $S$.\end{proof}
We have seen that $U(x)$ is always a left ideal of $\beta S$ and that $NS$ is a right
ideal of $\beta S$ provided it is nonempty.
\begin{theorem}\label{NSnotleft} Let $S$ be a countable cancellative semigroup.
Let $(Y,\langle T_s\rangle_{s\in S})$
be the dynamical system of Lemma {\rm \ref{stotwodyn}} determined by $R=S$.
Then $NS_Y$ is not a left ideal of $\beta S$.
\end{theorem}
\begin{proof} By \cite[Corollary 6.33]{HS} pick an idempotent $p\in \beta S\setminus K(\beta S)$.
By Theorem \ref{rpcancel} pick $r\in\beta S$ such that $rp$ is right cancelable in $\beta S$.
By Lemma \ref{chianotrec} and Theorem \ref{idemNS}, $p\in NS$ and by
Theorem \ref{rtcancimpsur}, $rp\notin NS$.
\end{proof}
If $S$ is commutative, then by \cite[Exercise 4.4.9]{HS} and Theorem \ref{NSright},
if $NS\neq\emp$, then $\cl NS$ is a two sided ideal of $\beta S$. The following
theorem shows that this may fail if $S$ is not commutative.
\begin{theorem}\label{freesg} Let $S$ be the free semigroup on the alphabet $\{a,b\}$ (where $a\neq b$).
Let $(Y,\langle T_s\rangle_{s\in S})$
be the dynamical system of Lemma {\rm \ref{stotwodyn}} determined by $R=S$.
Then $NS\neq\emp$ and $\cl NS$ is not a left ideal of $\beta S$.
\end{theorem}
\begin{proof} Let $p$ be an idempotent in $\beta S$ with $\{a^n:n\in\ben\}\in p$.
By Lemma \ref{chianotrec} and Theorem \ref{idemNS}, $p\in NS$. We will show that
$bp\notin\cl NS$. Let $B=\{ba^n:n\in\ben\}$. Then $B\in bp$. We shall show that
$\overline B\cap NS=\emp$. So let $q\in \overline B$.
Let $s_0=a$ and let $\langle s_n\rangle_{n=1}^\infty$ enumerate $S\setminus\{a\}$ so that
if the length of $s_i$ is less than the length of $s_j$, then $i<j$.
By Theorem \ref{Singroup}, to see that $T_q$ is surjective, it suffices to show that
there is some $x\in Y$ such that $T_q(x)$ is the characteristic function of $\{a\}$.
Let $x$ be the characteristic function of $\{aba^n:n\in\ben\}$ in $Y$.
Let $U$ be a neighborhood of $\cchi_{\{a\}}$ and pick $F\in \pf(S\setminus\{a\})$
such that $\pi_a^{-1}[\{1\}]\cap\bigcap_{y\in F}\pi_{y}^{-1}[\{0\}]\subseteq U$.
It suffices to show that $B\subseteq\{w\in S:T_w(x)\in \pi_a^{-1}[\{1\}]\cap\bigcap_{y\in F}\pi_{y}^{-1}[\{0\}]\}$.
So let $ba^n\in B$. Then $T_{ba^n}(x)(a)=x(aba^n)=1$ and for
$y\in F$, $T_{ba^n}(x)(y)=x(yba^n)=0$.
\end{proof}
We remark that Theorem \ref{freesg} remains
valid if $S$ is the free semigroup on a countably infinite alphabet.
\bibliographystyle{plain}
|
\section{Introduction}
PAC-Bayesian analysis, where PAC stands for the Probably Approximately Correct frequentist learning model \cite{Val84}, analyzes prediction accuracy of \emph{randomized classifiers}. A randomized classifier is a classifier defined by a distribution $\rho$ over a hypothesis class ${\cal H}$. A randomized classifier predicts by drawing a hypothesis from ${\cal H}$ according to $\rho$ and applying it to make the prediction \cite{McA98}. In many applications randomized prediction is replaced by a $\rho$-weighted majority vote \cite{GLLM09}.
PAC-Bayesian analysis provides some of the tightest generalization bounds in statistical learning theory \cite{GLLM09}. PAC-Bayesian bounds have a form of a trade-off between empirical performance of $\rho$ and its complexity, measured by the Kullback-Leibler (KL) divergence (a.k.a.\ relative entropy) between $\rho$ and a prior distribution $\pi$. Most of PAC-Bayesian literature relies on cross-validation to tune the trade-off. Cross-validation is an extremely powerful and practical heuristic for selecting model parameters, but it can potentially be misleading \cite{KMNR97,KR99}. It is also computationally expensive, especially for computationally demanding models, such as kernel SVMs, since it requires training a large number of classifiers on almost the whole dataset. Derivation of theoretical results that would not require parameter cross-validation is a long-standing challenge for theoretical machine learning \cite{Lan05}.
The need to rely on cross-validation stems from several reasons:
\begin{itemize}
\item Not all of the existing PAC-Bayesian bounds are convex in the posterior distribution $\rho$. For example, the most widely used PAC-Bayes-kl bound due to \citet{See02} is non-convex. This makes it hard to minimize the bound with respect to the posterior distribution. In most papers the bound is replaced by a linear trade-off between empirical error and the KL divergence and the trade-off parameter is tuned by cross-validation.
While it is possible to achieve convexity in the posterior distribution $\rho$ by introducing an additional trade-off parameter \cite{Cat07,KMH11}, we are unaware of successful attempts to tune the additional trade-off parameter through rigorous bound minimization. In practice, the alternative bounds are replaced by the same linear trade-off mentioned above and tuned by cross-validation.
\item The second obstacle is that, in order to keep the KL divergence between the posterior and the prior tractable, the set of posterior and prior distributions is often restricted. A popular example are Gaussian posteriors and Gaussian priors \cite{LST02,McA03,Lan05}. Even if the initial bound is convex in the posterior distribution, the convexity may be broken by such a restriction or reparametrization, as it happens in the Gaussian case \cite{GLLM09}.
\item Even though PAC-Bayesian bounds are some of the tightest, we are unaware of examples, where their tightness is sufficient to compete with cross-validation in tuning the trade-off between complexity and empirical performance.
\end{itemize}
We propose a relaxation of Seeger's PAC-Bayes-kl inequality, which we name \emph{PAC-Bayes-$\lambda$ inequality} or \emph{PAC-Bayes-$\lambda$ bound} when referring to the right hand side of the inequality. The bound is convex in the posterior distribution $\rho$ and has a convex trade-off between the empirical loss and KL divergence. The inequality is similar in spirit to the one proposed by \citet{KMH11}, but it does not restrict the form of $\rho$ and $\pi$. We provide an alternating procedure for minimizing the bound. We show that the bound can be rewritten as a continuous one-dimensional function of the trade-off parameter $\lambda$ and that under certain sufficient conditions this function is strongly quasiconvex (it has a single global minimum and no other stationary points). This guarantees convergence of alternating minimization to the global optimum.
For infinite hypothesis spaces alternating minimization can be computationally intractable or require parametrization, which can break the convexity of the bound in the posterior distribution. We get around this difficulty by constructing a finite data-dependent hypothesis space. The hypothesis space is constructed by taking $m$ subsamples of size $r$ from the training data. Each subsample is used to train a weak classifier, which is then validated on the remaining $n-r$ points, where $n$ is the sample size. We adapt our PAC-Bayesian bound and minimization procedure to this setting. Our analysis and minimization procedure work for any $m$, $r$, and any split of the data, including overlaps between training sets and overlaps between validation sets. In particular, it can also be applied to aggregate models originating from a cross-validation split of the data. However, in cross-validation the training sets are typically large (of order $n$) and validation sets and the number of models are small. While the prediction accuracy is still competitive in this setting, the highest computational advantage from our approach is achieved when the relation is reversed and the training size $r$ is taken to be small, roughly of order $d$, where $d$ is the number of features, and the number of models $m$ is taken to be large, roughly of order $n$. The construction of hypothesis space can be seen as sample compression \cite{LM07}. However, unlike the common approach to sample compression, which considers all possible subsamples of a given size and thus computationally and statistically inefficient, we consider only a small subset of possible subsamples.
We provide experimental results on several UCI datasets showing that the prediction accuracy of our learning procedure (training $m$ weak classifiers and weighting their predictions through minimization of the PAC-Bayes-$\lambda$ bound) is comparable to prediction accuracy of kernel SVMs tuned by cross-validation. In addition, we show that when $r$ is considerably smaller than $n$ and $m$ is of order $n$, the comparable prediction accuracy is achieved at a much lower computation cost. The computational speed-up is achieved because of the super-quadratic training time of kernel SVMs, which makes it much faster to train many weak SVMs on small training sets than one powerful SVM on a big training set.
In the following, we provide a brief review of PAC-Bayesian analysis, then present the PAC-Bayesian bound and its minimization procedure in Section~\ref{sec:bound}, derive conditions for convergence of minimization procedure to the global minimum in Section~\ref{sec:quasi}, describe our construction of a hypothesis space and specialize our results to this construction in Section~\ref{sec:aggregation}, and provide experimental validation in Section~\ref{sec:experiments}.
\section{A Brief Review of PAC-Bayesian Analysis}
To set the scene we start with a brief review of PAC-Bayesian analysis.
\subsubsection*{Notations}
We consider a supervised learning setting with an input space ${\cal X}$ and an output space ${\cal Y}$. We let $S = \lrc{(X_1,Y_1),\dots,(X_n,Y_n)}$ denote an independent identically distributed (i.i.d.) sample of size $n$ drawn according to an unknown distribution ${\cal D}(X,Y)$. A hypothesis $h$ is a function from the input to the output space $h:{\cal X} \to {\cal Y}$. We use ${\cal H}$ to denote a hypothesis class. We let $\ell:{\cal Y}^2 \to [0,1]$ denote a bounded loss function. The loss of $h$ on a sample $(X,Y)$ is $\ell(h(X),Y)$ and the expected loss of $h$ is denoted by $L(h) = \mathbb E\lrs{\ell(h(X),Y)}$. We use $\hat L(h,S) = \frac{1}{n} \sum_{i=1}^n \ell(h(X_i),Y_i)$ to denote the empirical loss of $h$ on $S$.
A randomized prediction rule parametrized by a distribution $\rho$ over ${\cal H}$ is
defined in the following way. For each prediction on a sample point $X$ the rule
draws a new hypothesis $h \in {\cal H}$ according to $\rho$ and applies it to $X$. The expected loss of such prediction rule
is
\mathbb E_{h \sim \rho}\lrs{L(h)}$ and the empirical loss is
\mathbb E_{h\sim\rho}\lrs{\hat L(h,S)}$. We use $\KL(\rho\|\pi) = \mathbb E_{h\sim\rho}\lrs{\ln\frac{\rho(h)}{\pi(h)}}$ to denote the KL divergence between $\rho$ and $\pi$. For Bernoulli distributions with biases $p$ and $q$ we use $\kl(p\|q)$ as a shorthand for $\KL([p,1-p]\|[q,1-q])$, the KL divergence between the two distributions. Finally, we use $\mathbb E_\rho\lrs{\cdot}$ as a shorthand for $\mathbb E_{h\sim\rho}\lrs{\cdot}$ and $\mathbb E_S\lrs{\cdot}$ as a shorthand for $\mathbb E_{S\sim{\cal D}^n}\lrs{\cdot}$.
\subsubsection*{Change of Measure Inequality}
The majority of PAC-Bayesian bounds are based on the following lemma.
\begin{lemma}[Change of Measure Inequality] For any function $f: {\cal H} \times \lr{{\cal X} \times {\cal Y}}^n \to \mathbb R$ and for any distribution $\pi$ over ${\cal H}$ , such that $\pi$ is independent of $S$, with probability greater than $1-\delta$ over a random draw of $S$, for all distributions $\rho$ over ${\cal H}$ simultaneously:
\begin{equation}
\mathbb E_{h\sim\rho}\lrs{f(h,S)} \leq \KL(\rho\|\pi) + \ln \frac{1}{\delta} + \ln \mathbb E_{h\sim\pi}\lrs{\mathbb E_{S'}\lrs{e^{f(h,S')}}}.
\label{eq:com}
\end{equation}
\label{lem:com}
\end{lemma}
The lemma is based on Donsker-Varadhan's variational definition of the KL divergence \cite{DV75}, by which $\KL(\rho\|\pi) = \sup_f \lrc{\mathbb E_\rho[f(h)] + \ln \mathbb E_\pi \lrs{e^{f(h)}}}$, where the supremum is over all measurable functions $f:{\cal H}\to\mathbb R$. In the lemma, $f$ is extended to be a function of $h$ and $S$ and then Markov's inequality is used to bound the expectation with respect to $\pi$ by $\mathbb E_\pi \lrs{e^{f(h,S)}} \leq \frac{1}{\delta} \mathbb E_{S'}\lrs{\mathbb E_\pi \lrs{e^{f(h,S')}}}$ with probability at least $1-\delta$. Independence of $\pi$ and $S$ allows to exchange the order of expectations, leading to the statement of the lemma. For a formal proof we refer to \citet{TS13}.
\subsubsection*{PAC-Bayes-kl Inequality}
Various choices of the function $f$ in Lemma~\ref{lem:com} lead to various forms of PAC-Bayesian bounds \cite{SLCB+12}. The classical choice is $f(h,S) = n \kl(\hat L(h,S)\|L(h))$. The moment generating function of $f$ can be bounded in this case by $\mathbb E_S\lrs{e^{f(h,S)}} \leq 2 \sqrt n$ \cite{Mau04,GLL+15}. This bound is used to control the last term in equation \eqref{eq:com}, leading to the PAC-Bayes-kl inequality \cite{See02}.
\begin{theorem}[PAC-Bayes-kl Inequality] For any probability distribution $\pi$ over ${\cal H}$ that is independent of $S$ and any $\delta \in (0,1)$, with probability greater than $1-\delta$ over a random draw of a sample $S$, for all distributions $\rho$ over ${\cal H}$ simultaneously:
\begin{equation}
\label{eq:PBkl}
\kl\lr{\mathbb E_\rho\lrs{\hat L(h,S)}\middle\|\mathbb E_\rho\lrs{L(h)}} \leq \frac{\KL(\rho\|\pi) + \ln \frac{2 \sqrt n}{\delta}}{n}.
\end{equation}
\label{thm:PBkl}
\end{theorem}
\section{PAC-Bayes-$\lambda$ inequality and its Alternating Minimization}
\label{sec:bound}
By inversion of the $\kl$ with respect to its second argument, inequality \eqref{eq:PBkl} provides a bound on $\mathbb E_\rho\lrs{L(h)}$. However, this bound is not convex in $\rho$ and, therefore, inconvenient for minimization. We introduce a relaxed form of the inequality, which has an additional trade-off parameter $\lambda$. The inequality leads to a bound, which is convex in $\rho$ for a fixed $\lambda$ and convex in $\lambda$ for a fixed $\rho$, making it amenable to alternating minimization. Theorem~\ref{thm:PBlambda} is analogous to \citet[Theorem 1]{KMH11} and a similar result can also be derived by using the techniques from \citet{TS13}, as shown by \citet{Thi16}.
\begin{theorem}[PAC-Bayes-$\lambda$ Inequality] For any probability distribution $\pi$ over ${\cal H}$ that is independent of $S$ and any $\delta \in (0,1)$, with probability greater than $1-\delta$ over a random draw of a sample $S$, for all distributions $\rho$ over ${\cal H}$ and $\lambda \in (0,2)$ simultaneously:
\begin{equation}
\label{eq:PBlambda}
\mathbb E_\rho\lrs{L(h)} \leq \frac{\mathbb E_\rho\lrs{\hat L(h,S)}}{1 - \frac{\lambda}{2}} + \frac{\KL(\rho\|\pi) + \ln \frac{2 \sqrt n}{\delta}}{\lambda\lr{1-\frac{\lambda}{2}}n}.
\end{equation}
\label{thm:PBlambda}
\end{theorem}
We emphasize that the theorem holds for \emph{all} values of $\lambda \in (0,2)$ simultaneously. This is in contrast to some other parametrized PAC-Bayesian bounds, for example, the one proposed by \citet{Cat07}, which hold for a \emph{fixed} value of a trade-off parameter.
\begin{proof}
We use the following analog of Pinsker's inequality \citep[Lemma 8.4]{Mar96,Mar97,Sam00,BLM13}: for $p < q$
\begin{equation}
\label{eq:kl}
\kl(p\|q) \geq (q-p)^2/(2q).
\end{equation}
By application of inequality \eqref{eq:kl}, inequality \eqref{eq:PBkl} can be relaxed to
\begin{equation}
\mathbb E_\rho\lrs{L(h)} - \mathbb E_\rho\lrs{\hat L(h,S)} \leq \sqrt{2 \mathbb E_\rho\lrs{L(h)} \frac{\KL(\rho\|\pi) + \ln \frac{2 \sqrt n}{\delta}}{n}}
\label{eq:PBsqrt}
\end{equation}
\cite{McA03}. By using the inequality $\sqrt{xy} \leq \frac{1}{2}\lr{\lambda x + \frac{y}{\lambda}}$ for all $\lambda > 0$, we have that with probability at least $1-\delta$ for all $\rho$ and $\lambda > 0$
\begin{equation}
\mathbb E_\rho\lrs{L(h)} - \mathbb E_\rho\lrs{\hat L(h,S)} \leq \frac{\lambda}{2}\mathbb E_\rho\lrs{L(h)} + \frac{\KL(\rho\|\pi) + \ln \frac{2 \sqrt n}{\delta}}{\lambda n}
\end{equation}
\cite{KMH11}. By changing sides
\[
\lr{1 - \frac{\lambda}{2}} \mathbb E_\rho\lrs{L(h)} \leq \mathbb E_\rho\lrs{\hat L(h,S)} + \frac{\KL(\rho\|\pi) + \ln \frac{2 \sqrt n}{\delta}}{\lambda n}.
\]
For $\lambda < 2$ we can divide both sides by $\lr{1-\frac{\lambda}{2}}$ and obtain the theorem statement.
\end{proof}
Since $\mathbb E_\rho\lrs{\hat L(h,S)}$ is linear in $\rho$ and $\KL(\rho\|\pi)$ is convex in $\rho$ \cite{CT06}, for a fixed $\lambda$ the right hand side of inequality \eqref{eq:PBlambda} is convex in $\rho$ and the minimum is achieved by
\begin{equation}
\label{eq:rho}
\rho_\lambda(h) = \frac{\pi(h) e^{-\lambda n \hat L(h,S)}}{\mathbb E_\pi\lrs{e^{-\lambda n \hat L(h',S)}}},
\end{equation}
where $\mathbb E_\pi\lrs{e^{-\lambda n \hat L(h',S)}}$, a shorthand for $\mathbb E_{h'\sim\pi}\lrs{e^{-\lambda n \hat L(h',S)}}$, is a convenient way of writing the normalization factor, which covers continuous and discrete hypothesis spaces in a unified notation. Furthermore, for $t \in (0,1)$ and $c_1,c_2 \geq 0$ the function $\frac{c_1}{1-t} + \frac{c_2}{t(1-t)}$ is convex in $t$ \cite{TS13}. Therefore, for a fixed $\rho$ the right hand side of inequality \eqref{eq:PBlambda} is convex in $\lambda$ for $\lambda \in (0,2)$. The minimum is achieved by
\begin{equation}
\label{eq:lambda}
\lambda = \frac{2}{\sqrt{\frac{2n \mathbb E_\rho\lrs{\hat L(h,S)}}{\lr{\KL(\rho\|\pi) + \ln \frac{2 \sqrt n}{\delta}}} + 1} + 1}
\end{equation}
\citep{TS13}. Note that the optimal value of $\lambda$ is smaller than 1 and that for $n \geq 4$ it can be lower bounded as $\lambda \geq \frac{2}{\sqrt{2 n + 1} + 1} \geq \frac{1}{\sqrt n}$. Alternating application of the update rules \eqref{eq:rho} and \eqref{eq:lambda} monotonously decreases the bound, and thus converges to a local minimum.
Unfortunately, the bound is \emph{not} jointly convex in $\rho$ and $\lambda$ (this can be verified by computing the Hessian of the first term and taking large $n$, so that the second term can be ignored). Joint convexity would have been a sufficient condition for convergence to the global minimum, but it is \emph{not} a necessary condition. In the following section we show that under certain conditions alternating minimization is still guaranteed to converge to the global minimum of the bound despite absence of joint convexity.
\section{Strong Quasiconvexity of the PAC-Bayes-$\lambda$ Bound}
\label{sec:quasi}
We denote the right hand side of the bound in Theorem~\ref{thm:PBlambda} by
\[
{\cal F}(\rho,\lambda) = \frac{\mathbb E_\rho\lrs{\hat L(h,S)}}{1 - \lambda/2} + \frac{\KL(\rho\|\pi) + \ln \frac{2 \sqrt{n}}{\delta}}{n \lambda(1-\lambda/2)}.
\]
By substituting the optimal value of $\rho$ from equation \eqref{eq:rho} into ${\cal F}(\rho,\lambda)$ and applying the identity
\begin{align}
\KL(\rho_\lambda\|\pi) &= \mathbb E_{\rho_\lambda}\lrs{\ln \frac{\rho_\lambda(h)}{\pi(h)}} = \mathbb E_{\rho_\lambda}\lrs{\ln \frac{e^{-n\lambda \hat L(h,S)}}{\mathbb E_\pi\lrs{e^{-n\lambda \hat L(h',S)}}}}\notag\\ &= -n\lambda \E_{\rho_\lambda}\lrs{\hat L(h,S)} - \ln\E_\pi\lrs{e^{-n\lambda \hat L(h,S)}}
\label{eq:KL}
\end{align}
we can write ${\cal F}$ as a function of a single scalar parameter $\lambda$:
\begin{align*}
{\cal F}(\lambda) &= \frac{\mathbb E_{\rho_\lambda}\lrs{\hat L(h,S)}}{1 - \lambda/2} + \frac{\KL(\rho_\lambda\|\pi) + \ln \frac{2 \sqrt{n}}{\delta}}{n \lambda(1-\lambda/2)}\\
&= \frac{\mathbb E_{\rho_\lambda}\lrs{\hat L(h,S)}}{1 - \lambda/2} - \frac{\mathbb E_{\rho_\lambda}\lrs{\hat L(h,S)}}{1 - \lambda/2} + \frac{-\ln \E_\pi\lrs{e^{-n\lambda \hat L(h,S)}} + \ln \frac{2 \sqrt{n}}{\delta}}{n \lambda(1-\lambda/2)}\\
&= \frac{- \ln \E_\pi\lrs{e^{-n\lambda \hat L(h,S)}} + \ln \frac{2 \sqrt{n}}{\delta}}{n \lambda(1-\lambda/2)}.
\end{align*}
We note that ${\cal F}(\lambda)$ is not necessarily convex in $\lambda$. For example, taking ${\cal H} = \lrc{h_1,h_2}$, $\hat L(h_1,S) = 0$, $\hat L(h_2,S) = 0.5$, $\pi(h_1) = \pi(h_2) = \frac{1}{2}$, $n = 100$, and $\delta = 0.01$ produces a non-convex ${\cal F}$. However, we show that ${\cal F}(\lambda)$ is strongly quasiconvex under a certain condition on the variance defined by
\[
{\textrm{Var}}_{\rho_\lambda}\lrs{\hat L(h,S)} = \E_{\rho_\lambda}\lrs{\hat L(h,S)^2} - \E_{\rho_\lambda}\lrs{\hat L(h,S)}^2.
\]
We recall that a univariate function $f:\mathcal{I}\to\mathbb R$ defined on an interval $\mathcal{I} \subseteq \mathbb R$ is \emph{strongly quasiconvex} if for any $x,y \in \mathcal{I}$ and $t \in (0,1)$ we have $f(tx + (1-t)y) < \max\lrc{f(x),f(y)}$.
\begin{theorem}
${\cal F}(\lambda)$ is continuous and if at least one of the two conditions
\begin{equation}
2\KL(\rho_\lambda\|\pi) + \ln\frac{4n}{\delta^2} > \lambda^2 n^2 {\textrm{Var}}_{\rho_\lambda}\lrs{\hat L(h,S)} \label{eq:KL-cond}
\end{equation}
or
\begin{equation}
\E_{\rho_\lambda}\lrs{\hat L(h,S)} > (1-\lambda) n {\textrm{Var}}_{\rho_\lambda}\lrs{\hat L(h,S)} \label{eq:E-cond}
\end{equation}
is satisfied for all $\lambda \in \lrs{\sqrt{\frac{\ln \frac{2 \sqrt{n}}{\delta}}{n}}, 1}$, then ${\cal F}(\lambda)$ is strongly quasiconvex for $\lambda \in (0,1]$ and alternating application of the update rules \eqref{eq:rho} and \eqref{eq:lambda} converges to the global minimum of ${\cal F}$.
\label{thm:global-min}
\end{theorem}
\begin{proof}
The proof is based on inspection of the first and second derivative of ${\cal F}(\lambda)$. Calculation of the derivatives is provided in Appendix~\ref{app:derivatives}. The existence of the first derivative ensures continuity of ${\cal F}(\lambda)$. By inspecting the first derivative we obtain that stationary points of ${\cal F}(\lambda)$ corresponding to ${\cal F}'(\lambda) = 0$ are characterized by the identity
\[
2 (1 - \lambda) \lr{\KL(\rho_\lambda\|\pi) + \ln \frac{2\sqrt n}{\delta}} = \lambda^2 n \E_{\rho_\lambda}\lrs{\hat L(h,S)}.
\]
The identity provides a lower bound on the value of $\lambda$ at potential stationary points. Using the facts that $\E_{\rho_\lambda}\lrs{\hat L(h,S)} \leq 1$ and for $\lambda \leq \frac{1}{2}$ the complement $(1-\lambda) \geq \frac{1}{2}$, for $n \geq 7$ we have
\[
\lambda = \sqrt{\frac{2 (1 - \lambda) \lr{\KL(\rho_\lambda\|\pi) + \ln \frac{2\sqrt n}{\delta}}}{n\E_{\rho_\lambda}\lrs{\hat L(h,S)}}} \geq \min\lr{\sqrt{\frac{\KL(\rho_\lambda\|\pi)+\ln \frac{2 \sqrt{n}}{\delta}}{n}}~,~\frac{1}{2}} \geq \sqrt{\frac{\ln \frac{2 \sqrt{n}}{\delta}}{n}}.
\]
By expressing $\KL(\rho_\lambda\|\pi)$ via $\E_{\rho_\lambda}\lrs{\hat L(h,S)}$ (or the other way around) and substituting it into the second derivative of ${\cal F}(\lambda)$ we obtain that if either of the two conditions of the theorem is satisfied at a stationary point then the second derivative of ${\cal F}(\lambda)$ is positive there. Thus, if \eqref{eq:KL-cond} or \eqref{eq:E-cond} is satisfied for all $\lambda \in \lrs{\sqrt{\frac{\ln \frac{2 \sqrt{n}}{\delta}}{n}}, 1}$ then all stationary points of ${\cal F}(\lambda)$ (if any exist) are local minima. Since ${\cal F}(\lambda)$ is a continuous one-dimensional function it means that ${\cal F}(\lambda)$ is strongly quasiconvex (it has a single global minimum and no other stationary points). Since alternating minimization monotonously decreases the value of ${\cal F}(\lambda)$ it is guaranteed to converge to the global minimum.
\end{proof}
Next we show a sufficient condition for \eqref{eq:KL-cond} to be satisfied for a finite ${\cal H}$ for all $\lambda$.
\begin{theorem}
Let $m$ be the number of hypotheses in ${\cal H}$ and assume that the prior $\pi(h)$ is uniform. Let $a = \frac{\sqrt{\ln \frac{4n}{\delta^2}}}{n\sqrt 3}$, $b = \frac{\ln(3mn^2)}{\sqrt{n\ln \frac{2 \sqrt{n}}{\delta}}}$, and $K = \frac{e^2}{12}\ln\frac{4n}{\delta^2}$. Let $x_h = \hat L(h,S) - \min_h \hat L(h,S)$. If the number of hypotheses for which $x_h \in (a,b)$ is at most $K$ then ${\textrm{Var}}_{\rho_\lambda}\lrs{\hat L(h,S)} \leq \frac{\ln \frac{4n}{\delta^2}}{\lambda^2 n^2}$ for all $\lambda \in \lrs{\sqrt{\frac{\ln \frac{2 \sqrt{n}}{\delta}}{n}}, 1}$ and ${\cal F}(\lambda)$ is strongly quasiconvex and its global minimum is guaranteed to be found by alternating application of the update rules \eqref{eq:rho} and \eqref{eq:lambda}.
\label{thm:var}
\end{theorem}
The theorem splits hypotheses in ${\cal H}$ into ``good'', ``mediocre'', and ``bad''. ``Good'' hypotheses are those for which $x_h \leq a$, meaning that the empirical loss $\hat L(h,S)$ is close to the best. ``Mediocre'' are those for which $x_h \in (a,b)$. ``Bad'' are those for which $x_h \geq b$. The theorem states that as long as the number of ``mediocre'' hypotheses is not too large, ${\cal F}(\lambda)$ is guaranteed to be quasiconvex.
\begin{proof}
We have ${\textrm{Var}}_{\rho_\lambda}\lrs{\hat L(h,S)} = {\textrm{Var}}_{\rho_\lambda}\lrs{\hat L(h,S) - \min_h \hat L(h,S)}$. Under the assumption of a uniform prior $\rho_\lambda(h) = e^{-n\lambda x_h} / \sum_{h'} e^{-n\lambda x_{h'}}$. Since for $h^* = \arg\min_h \hat L(h,S)$ we have $x_{h^*} = 0$, the denominator satisfies $\sum_h e^{-n \lambda x_h} \geq 1$. Let ${\cal I}_{0a} = [0,a]$, ${\cal I}_{ab} = (a,b)$, and ${\cal I}_{b1} = [b,1]$ be the intervals corresponding to ``good'', ``mediocre'', and ``bad'' hypotheses. We have:
\begin{align*}
{\textrm{Var}}_{\rho_\lambda}\lrs{\hat L(h,S)} &\leq \mathbb E_{\rho_\lambda}\lrs{x_h^2}\\
&= \frac{\sum_h x_h^2 e^{-n\lambda x_h}}{\sum_h e^{-n\lambda x_h}}\\
&= \frac{\sum_{x_h\in {\cal I}_{0a}} x_h^2 e^{-n\lambda x_h}}{\sum_h e^{-n\lambda x_h}} + \frac{\sum_{x_h\in {\cal I}_{ab}} x_h^2 e^{-n\lambda x_h}}{\sum_h e^{-n\lambda x_h}} + \frac{\sum_{x_h\in {\cal I}_{b1}} x_h^2 e^{-n\lambda x_h}}{\sum_h e^{-n\lambda x_h}}\\
&\leq a^2 + \sum_{x_h\in {\cal I}_{ab}} x_h^2 e^{-n\lambda x_h} + \sum_{x_h\in {\cal I}_{b1}} x_h^2 e^{-n\lambda x_h}.
\end{align*}
We show a number of properties of the above expression. First, we recall that $\lambda \leq 1$. Therefore,
\[
a^2 = \frac{\ln \frac{4n}{\delta^2}}{3n^2} \leq \frac{\ln\frac{4n}{\delta^2}}{3\lambda^2 n^2}.
\]
For the second term, simple calculus gives $x^2 e^{-n\lambda x} \leq \frac{4}{e^2 n^2 \lambda^2}$. Since by the theorem assumption there are at most $K$ hypotheses falling in ${\cal I}_{ab}$,
\[
\sum_{h\in {\cal I}_{ab}} x_h^2 e^{-n\lambda x_h} \leq \frac{4K}{e^2 n^2 \lambda^2} \leq \frac{\ln\frac{4n}{\delta^2}}{3\lambda^2 n^2}.
\]
For the last term we have
\[
b > \frac{2}{\sqrt{n\ln \frac{2 \sqrt{n}}{\delta}}} = \frac{2}{n}\sqrt{\frac{n}{\ln \frac{2 \sqrt{n}}{\delta}}} \geq \frac{2}{\lambda n}
\]
and we note that for $x \geq 2 / \lambda n$ the function $x^2 e^{-n\lambda x}$ is monotonically decreasing in $x$. Since $\lambda \geq \sqrt{\frac{\ln \frac{2 \sqrt{n}}{\delta}}{n}}$ we obtain
\begin{equation}
\label{eq:b}
\sum_{x_h \in {\cal I}_{b1}} x_h^2 e^{-n\lambda x_h} \leq m b^2 e^{-n\lambda b} \leq m b^2 e^{-\sqrt{n\ln \frac{2 \sqrt{n}}{\delta}} b} \leq m e^{-\sqrt{n\ln \frac{2 \sqrt{n}}{\delta}} b} = m \frac{1}{3m n^2} = \frac{1}{3n^2} \leq \frac{\ln\frac{4n}{\delta^2}}{3\lambda^2 n^2}.
\end{equation}
By taking all three terms together we arrive at
\[
{\textrm{Var}}_{\rho_\lambda}\lrs{\hat L(h,S)} \leq \frac{\ln \frac{4n}{\delta^2}}{\lambda^2 n^2},
\]
which implies condition \eqref{eq:KL-cond} of Theorem~\ref{thm:global-min} since $\KL(\rho_\lambda\|\pi) \geq 0$.
\end{proof}
In Appendix~\ref{app:thm-var} we provide a couple of relaxations of the conditions in Theorem~\ref{thm:var}. The first allows to trade the boundaries $a$ and $b$ of the intervals with the value of $K$ and the second improves the value of $b$.
In our experiments presented in Section~\ref{sec:experiments}, ${\cal F}(\lambda)$ turned to be convex even when the sufficient conditions of Theorem~\ref{thm:var} (including the relaxations detailed in Appendix~\ref{app:thm-var}) were violated. This suggests that it may be possible to relax the conditions even further. At the same time, it is possible to construct artificial examples, where ${\cal F}(\lambda)$ is not quasiconvex. For example, taking $n = 200$, $\delta = 0.25$, and $m = \round\lr{e^{0.74 n\Delta}} + 1 \approx 2.7 \cdot 10^6$ hypotheses (where $\round$ is rounding to the nearest integer) with $\hat L(h_1, S) = 0$ and $\hat L(h_i, S) = \Delta = 0.1$ for all $i \in \lrc{2,\dots,m}$ and a uniform prior leads to ${\cal F}(\lambda)$ with two local minima. The artificial example requires $m$ to be of the order of $e^{n \lambda^\star \Delta}$, where $\lambda^\star$ is the value of $\lambda$ at a stationary point of ${\cal F}(\lambda)$ and $\Delta$ is the loss of suboptimal hypotheses (in the example $\Delta = 0.1$). Thus, quasiconvexity is not always guaranteed, but it holds in a wide range of practical scenarios.
\section{Construction of a Hypothesis Space}
\label{sec:aggregation}
Computation of the partition function (the denominator in \eqref{eq:rho}) is not always tractable, however, it can be easily computed when ${\cal H}$ is finite.
The crucial step is to construct a sufficiently powerful finite hypothesis space ${\cal H}$. Our proposal is to construct ${\cal H}$ by training $m$ hypotheses, where each hypothesis is trained on $r$ random points from $S$ and validated on the remaining $n-r$ points. This construction resembles a cross-validation split of the data. However, in cross-validation $r$ is typically large (close to $n$) and validation sets are non-overlapping. Our approach works for any $r$ and has additional computational advantages when $r$ is small. We do not require validation sets to be non-overlapping and overlaps between training sets are allowed. Below we describe the construction more formally.
Let $h \in \lrc{1,\dots,m}$ index the hypotheses in ${\cal H}$. Let $S_h$ denote the training set of $h$ and $S\setminus S_h$ the validation set. $S_h$ is a subset of $r$ points from $S$, which are selected independently of their values (for example, subsampled randomly or picked according to a predefined partition of the data). We define the validation error of $h$ by $\hat L^{\text{\normalfont val}}(h,S) = \frac{1}{n-r} \sum_{(X,Y) \in S\setminus S_h} \ell(h(X),Y)$. Note that the validation errors are $(n-r)$ i.i.d.\ random variables with bias $L(h)$ and, therefore, for $f(h,S) = (n-r) \kl(\hat L^{\text{\normalfont val}}(h,S)\|L(h))$ we have $\mathbb E_S\lrs{e^{f(h,S)}} \leq 2 \sqrt{n-r}$. The following result is a straightforward adaptation of Theorem~\ref{thm:PBlambda} to our setting. A proof sketch is provided in Appendix~\ref{app:PBaggregation}.
\begin{theorem} Let $S$ be a sample of size $n$. Let ${\cal H}$ be a set of $m$ hypotheses, where each $h \in {\cal H}$ is trained on $r$ points from $S$ selected independently of the composition of $S$. For any probability distribution $\pi$ over ${\cal H}$ that is independent of $S$ and any $\delta \in (0,1)$, with probability greater than $1-\delta$ over a random draw of a sample S, for all distributions $\rho$ over $\cal H$ and $\lambda \in (0,2)$ simultaneously:
\begin{equation}
\label{eq:PBaggregation}
\mathbb E_\rho\lrs{L(h)} \leq \frac{\mathbb E_\rho\lrs{\hat L^{\text{\normalfont val}}(h,S)}}{1 - \frac{\lambda}{2}} + \frac{\KL(\rho\|\pi) + \ln \frac{2 \sqrt{n-r}}{\delta}}{\lambda\lr{1-\frac{\lambda}{2}}(n-r)}.
\end{equation}
\label{thm:PBaggregation}
\end{theorem}
It is natural, but not mandatory to select a uniform prior $\pi(h) = 1/m$. The bound in equation \eqref{eq:PBaggregation} can be minimized by alternating application of the update rules in equations \eqref{eq:rho} and \eqref{eq:lambda} with $n$ being replaced by $n-r$ and $\hat L$ by $\hat L^{\text{\normalfont val}}$.
\section{Experimental Results}
\label{sec:experiments}
In this section we study how PAC-Bayesian weighting of weak classifiers proposed in Section~\ref{sec:aggregation} compares with ``strong'' kernel SVMs tuned by cross-validation and trained on all training data. The experiments were performed on eight UCI datasets \citep{Asuncion+Newman:2007} summarized in Table \ref{tbl:uci_sizes}. In our experiments we employed the SVM solver from LIBSVM \citep{libsvm}.
\begin{table
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|c|c|}
\hline
& Mushrooms & Skin & Waveform & Adult & Ionosphere & AvsB & Haberman & Breast\\ \hline
$|\text{S}|$ & 2000 & 2000 & 2000 & 2000 & 200 & 1000 & 150 & 340 \\ \hline
$|\text{T}|$ & 500 & 500 & 500 & 500 & 150 & 500 & 150 & 340 \\ \hline
$d$ & 112 & 3 & 40 & 122 & 34 & 16 & 3 & 10 \\ \hline
\end{tabular}
\end{center}
\caption{\textbf{Datasets summary.} $|S| = n$ refers to the size of the training set and $|T|$ refers to the size of the test set, $d$ refers to the number of features. ``Breast'' abbreviates ``Breast cancer'' dataset.}
\label{tbl:uci_sizes}
\end{table}
We compared the prediction accuracy and run time of our prediction strategy with a baseline of RBF kernel SVMs tuned by cross-validation. For the baseline we used 5-fold cross-validation for selecting the soft-margin parameter, $C$, and the bandwidth parameter $\gamma$ of the kernel $k(X_i,X_j)=\exp(-\gamma\|X_i-X_j\|^2)$. The value of $C$ was selected from a grid, such that $\log_{10}C \in \{-3,-2, \hdots, 3\}$. The values for the grid of $\gamma$-s were selected using the heuristic proposed in \citet{JDH99}. Specifically, for $i \in \{1,\hdots, n\}$ we defined
$G(X_i) = \min_{(X_j, Y_j) \in S \wedge Y_i \neq Y_j}\left\lVert X_i - X_j \right\rVert $.
We then defined a seed $\gamma_J$ by
$\gamma_{J} = \frac{1}{2\cdot\text{median}(G)^{2}}$.
Finally, we took a geometrically spaced grid around $\gamma_J$, so that $\gamma \in \{\gamma_{J}10^{-4}, \gamma_{J}10^{-2}, \hdots, \gamma_{J}10^{4}\}$
For our approach we selected $m$ subsets of $r$ points uniformly at random from the training set $S$. We then trained an RBF kernel SVM for each subset. The kernel bandwidth parameter $\gamma$ was randomly selected for each subset from the same grid as used in the baseline. In all our experiments very small values of $r$, typically up to $d+1$ with $d$ being the input dimension, were sufficient for successfully competing with the prediction accuracy of the baseline and provided the most significant computational improvement. For such small values of $r$ it was easy to achieve perfect separation of the training points and, therefore, selection of $C$ was unnecessary.
The performance of each weak classifier was validated on $n-r$ points not used in its training. The weighting of classifiers $\rho$ was then computed through alternating minimization of the bound in Theorem~\ref{thm:PBaggregation}.
In most of PAC-Bayesian literature it is common to replace randomized prediction with $\rho$-weighted majority vote. From a theoretical point of view the error of $\rho$-weighted majority vote is bounded by at most twice the error of the corresponding randomized classifier, however, in practice it usually performs better than randomized prediction \citep{GLLM09}. In our main experiments we have followed the standard choice of using the $\rho$-weighted majority vote. In Appendix~\ref{app:majority} we provide additional experiments showing that in our case the improvement achieved by taking the majority vote was very minor compared to randomized prediction. We use the term \emph{PAC-Bayesian aggregation} to refer to prediction with $\rho$-weighted majority vote.
\begin{figure}[t]
\centering
\subfigure[Ionosphere dataset. $L_{\text{\normalfont CV}} = 0.06$.]{
\includegraphics[width=0.5\textwidth]{heatmap_kernel_ionosphere200.png}
}%
\subfigure[Mushrooms dataset. $L_{\text{\normalfont CV}} = 0$]{
\includegraphics[width=0.5\textwidth]{heatmap_kernel_mushrooms2000.png}
}
\caption{\textbf{Prediction accuracy of PAC-Bayesian aggregation vs.\ cross-validated SVM across different values of $m$ and $r$.} The colors of the heatmap represent the difference between the zero-one loss of the $\rho$-weighted majority vote and the zero-one loss of the cross-validated SVM. The loss of the cross-validated SVM is given by $L_{\text{\normalfont CV}}$ in the caption.}
\label{fig:heatmap_kernel}
\end{figure}
\begin{figure}[th!]
\centering
\subfigure[Ionosphere dataset. $n = 200$,\newline$r = d+1 = 35$.]{
\includegraphics[width=0.48\textwidth]{subsample_kernel_ionosphere200.png}
}%
\subfigure[Waveform dataset. $n = 2000$,\newline$r = d+1 = 41$.]{
\includegraphics[width=0.48\textwidth]{subsample_kernel_waveform2000.png}
}
\subfigure[Breast cancer dataset. $n = 340$,\newline$r = d+1 = 11$.]{
\includegraphics[width=0.48\textwidth]{subsample_kernel_breast-cancer340.png}
}%
\subfigure[AvsB dataset. $n = 1000$, $r = d+1 = 17$.]{
\includegraphics[width=0.48\textwidth]{subsample_kernel_avsb1000.png}
}
\caption{\textbf{Comparison of PAC-Bayesian aggregation with RBF kernel SVM tuned by cross-validation.} The solid red, black, and blue lines correspond, respectively, to the zero-one test loss of the cross-validated SVM, the loss of $\rho$-weighted majority vote, where $\rho$ is a result of minimization of the PAC-Bayes-$\lambda$ bound, and PAC-Bayes-kl bound on the loss of randomized classifier defined by $\rho$. The dashed black line represents the training time of PAC-Bayesian aggregation, while the red dashed line represents the training time of cross-validated SVM. The prediction accuracy and run time of PAC-Bayesian aggregation and PAC-Bayes-kl bound are given as functions of the hypothesis set size $m$.}
\label{fig:runtime_kernel}
\end{figure}
In the first two experiments we studied the influence of $r$ and $m$ on classification accuracy and run time. The complexity term in Theorem~\ref{thm:PBaggregation} (the second term on the right hand side of \eqref{eq:PBaggregation}) decreases with the decrease of the training set sizes $r$ (because the size of the validation sets $n-r$ increases) and with the decrease of the number of hypotheses $m$ (because $\pi(h) = 1/m$ increases). From the computational point of view it is also desirable to have small $r$ and $m$, especially when working with expensive-to-train models, such as kernel SVMs, which have super-quadratic training time. What pushes $r$ and $m$ up is the validation error, $\mathbb E_\rho\lrs{\hat L^{\text{\normalfont val}}(h,S)}$.
In the first experiment we studied the influence of $r$ and $m$ on the prediction accuracy of PAC-Bayesian aggregation. We considered 20 values of $m$ in $[1, n]$ and 20 values of $r$ in $[2, d+1]$. For each pair of $m$ and $r$ the prediction accuracy of PAC-Bayesian aggregation was evaluated, resulting in a matrix of losses. To simplify the comparison we subtracted the prediction accuracy of the baseline, thus zero values correspond to matching the accuracy of the baseline. In Figure~\ref{fig:heatmap_kernel} we show a heatmap of this matrix for two UCI datasets and the results for the remaining datasets are provided in Appendix~\ref{app:more-figures}. Overall, reasonably small values of $m$ and $r$ were sufficient for matching the accuracy of SVM tuned by cross-validation.
In the second experiment we provide a closer look at the effect of increasing the number $m$ of weak SVMs when their training set sizes $r$ are kept fixed. We picked $r = d+1$ and ran our training procedure with 20 values of $m$ in $[1, n]$. In Figure~\ref{fig:runtime_kernel} we present the prediction accuracy of the resulting weighted majority vote vs.\ prediction accuracy of the baseline for four datasets. The graphs for the remaining datasets are provided in Appendix~\ref{app:more-figures}. We also show the running time of our procedure vs.\ the baseline. The running time of the baseline includes cross-validation and training of the final SVM on the whole training set, while the running time of PAC-Bayesian aggregation includes training of $m$ weak SVMs, their validation, and the computation of $\rho$. In addition, we report the value of PAC-Bayes-kl bound from Theorem~\ref{thm:PBkl} on the expected loss of the randomized classifier defined by $\rho$. The kl divergence was inverted numerically to obtain a bound on the expected loss $\mathbb E_\rho\lrs{L(h)}$. The bound was adapted to our construction by replacing $n$ with $n-r$ and $\mathbb E_\rho\lrs{\hat L(h,S)}$ with $\mathbb E_\rho\lrs{\hat L^{\text{\normalfont val}}(h,S)}$. We note that since the bound holds for any posterior distribution, it also holds for the distribution found by minimization of the bound in Theorem~\ref{thm:PBaggregation}. However, since Theorem~\ref{thm:PBaggregation} is a relaxation of PAC-Bayes-kl bound, using PAC-Bayes-kl for the final error estimate is slightly tighter. The bound on the loss of $\rho$-weighted majority vote is at most a factor of 2 larger than than the bound for the randomized classifier. In calculation of the bound we used $\delta = 0.05$. We conclude from the figure that relatively small values of $m$ are sufficient for matching or almost matching the prediction accuracy of the baseline, while the run time is reduced dramatically. We also note that the bound is exceptionally tight.
\begin{table
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|c|c|}
\hline
& Mushrooms & Skin & Waveform & Adult & Ionosphere & AvsB & Haberman & Breast\\ \hline
$m$ & 130 & 27 & 140 & 28 & 24 & 160 & 23 & 50\\
\hline
\end{tabular}
\end{center}
\caption{\textbf{Average maximal values of $m$ for which quasiconvexity was guaranteed by Theorem~\ref{thm:var}.}}
\label{tbl:quasiconvexity}
\end{table}
\begin{figure*
\centering
\subfigure[Breast cancer dataset with $|\text{S}| = 340$.]{
\includegraphics[width=0.29\textwidth]{f_lambda_convex_breast-cancer340.png}
}\quad
\subfigure[Adult dataset with $|\text{S}| = 2000$.]{
\includegraphics[width=0.29\textwidth]{f_lambda_convex_adult2000.png}
}\quad
\subfigure[Mushrooms dataset with $|\text{S}| = 2000$.]{
\includegraphics[width=0.29\textwidth]{f_lambda_convex_mushrooms2000.png}
}
\caption{\textbf{Empirical evaluation of the shape of ${\cal F}(\lambda)$}. The blue curve represents the value of ${\cal F}(\lambda)$ and the red cross marks $\lambda$ returned by alternating minimization.}
\label{fig:f_lambda_convex}
\end{figure*}
In our last experiment we tested quasiconvexity of ${\cal F}(\lambda)$. Theorem~\ref{thm:var} provided theoretical guarantee of quasiconvexity for small $m$ and numerical evaluation has further shown that ${\cal F}(\lambda)$ was convex for all values of $m$ used in our experiments. For testing the theoretical guarantees we increased $m$ in steps of 10 until the sufficient condition for strong quasiconvexity in Theorem~\ref{thm:var} was violated. The condition included adjustment of interval boundaries $a$ and $b$, as described in Appendix~\ref{app:thm-var-tune}, and improved value of $b$, as described in Appendix~\ref{app:thm-var-b}. The experiments were repeated 10 times for each dataset, where in each experiment the training sets were redrawn and a new set of hypotheses was trained, leading to a new set of validation losses $\hat L^{\text{\normalfont val}}$. In Table~\ref{tbl:quasiconvexity} we report the average over the 10 repetitions of the maximal values of $m$ with guaranteed quasiconvexity. Since ${\cal F}(\lambda)$ is always quasiconvex for $m \leq K(0,0) + 1$, where $K(0,0) = \frac{e^2}{4}\ln \frac{4n}{\delta^2}$ (see equation \eqref{eq:Kab} in Appendix~\ref{app:thm-var-tune}), we report the value of $K(0,0)+1$ whenever it was not possible to ensure quasiconvexity with larger $m$. When it was not possible to guarantee quasicovexity theoretically we tested the shape of ${\cal F}(\lambda)$ empirically. Figure~\ref{fig:f_lambda_convex} shows a few typical examples. The plots in Figure~\ref{fig:f_lambda_convex} were constructed in the following way: given a sample $S$, we trained $m$ weak SVMs. We then computed the corresponding vector of validation losses, $\hat L^{\text{\normalfont val}}(h, S)$. For each value in a grid of $\lambda$-s, we computed the corresponding $\rho$ according to equation \eqref{eq:rho}. Finally, we substituted the value of $\rho_\lambda$ and $\lambda$ into equation \eqref{eq:PBlambda} to get the value of the bound. In all our calculations we used a uniform prior and $\delta = 0.05$. The figure shows that ${\cal F}(\lambda)$ was convex in $\lambda$ in all the cases.
\section{Conclusion}
We have presented a new PAC-Bayesian inequality, an alternating procedure for its minimization, and a way to construct a finite hypothesis space for which the bound and minimization procedure work particularly well. We have derived sufficient conditions for the minimization procedure to converge to the global optimum of the bound. We have shown that the procedure is competitive with cross-validation in tuning the trade-off between complexity and empirical performance of $\rho$. In addition, it provides tight high-probability generalization guarantees and achieves prediction accuracies on par with kernel SVMs tuned by cross-validation, but at a considerably lower computation cost.
In our experiments the bound turned to be convex even when the sufficient conditions of Theorem~\ref{thm:var} were violated. It suggests that further relaxation of these conditions may be possible to achieve in future work.
\acks{We thank anonymous reviewers of this and earlier versions of the manuscript for valuable feedback. We also thank Oswin Krause for suggesting the use of the term ``quasiconvexity'' to describe the shape of ${\cal F}(\lambda)$. CI and YS acknowledge support by the Innovation Fund Denmark through the \emph{Danish Center for Big Data Analytics Driven Innovation} (DABAI). OW would like to thank the mathematical department of the university of Copenhagen for his guest professorship in 2015-2016.}
|
\section{Introduction}\label{intro}
Inasmuch as computability can be of utmost importance, one often
confines selection of stochastic differential equation (SDE) models to
those facilitating calculation and simulation.
This is perhaps best
exemplified in mathematical finance, where the popularity of the
inaccurate Black-Scholes model is only justifiable through the
evaluation ease of the resulting derivative product formulae.
Indeed, \citet[p. 272]{Kunita:1984} writes in his notes on
SDEs that ``It is an important problem
in applications that we can compute the output from the input
explicitly''.
We shall call such solutions \emph{explicit solutions}.
Filtering applications (see \citet{Kouritzin:1998}), option pricing
applications (see companion paper \citet{Kouritzin16}) and pedagogical considerations initially prompted
our classifications of which It\^o processes $X_t^{x,s}$, starting
at $(x,s)$, are representable as a time-dependent function of a simple
stochastic integral $\phi^{x,s}\left(\int_s^t U_{s,u} dW_u,t\right)$.
However,
our determination of $\phi^{x,s}$, $U_{s,u}$ also facilitates an effective means
of calculation and simulation. To simulate, one merely needs to
compute the Gauss-Markov process $\int_s^tU_{s,u} dW_u$ at discrete
times and substitute these samples into $\phi^{x,s}$, which is often
known in closed form and otherwise is the solution of differential
equations that can be solved numerically a priori.
The idea is applied to strong solutions here and extended to
weak solutions of a popular financial model
in \citet{Kouritzin16}.
$\int_s^tU_{s,u} dW_u=\int_s^tU_{s,u}(X_u) dW_u$ can depend upon $X$ but
not in a way that will destroy its Gaussian distribution nor
make simulation difficult and our explicit solutions are diffusion
solutions for all starting points $(x,s)$.
This Explicit Solution Simulation is without (Euler or Milstein) bias and is extremely efficient, often
orders of magnitude faster than Euler or Milstein
methods when our method is applicable and high accuracy is desired
(see \citet{Kouritzin16}).
Our representations also make properties of certain
stochastic differential equations readily discernible and simplifies
some filtering calculations. Finally, as demonstrated in
\citet[Proposition 5.2.24]{Karatzas/Shreve:1987}, explicit solutions
can be useful in establishing convergence for solutions of
stochastic differential equations.
\citet{Doss:1977} and \citet{Sussmann:1978} were apparently the
first to solve stochastic differential equations through use of
differential equations. In the multidimensional setting, Doss
imposed the Abelian condition on the Lie algebra generated by the
vector fields of coefficients and showed, in this case, that strong
solutions, $X_t^{x}$, of Fisk-Stratonovich equations
are representable as $X_t^{x}=\rho (\Phi (x,W_{\cdot })_t,W_t)$,
for some continuous $\rho $, $\Phi $ solving differential equations.
Under the restriction of $C^\infty $ coefficients,
\citet{Yamato:1979} extended the work of Doss by dispensing with
the Abelian assumption in favour of less restrictive $q$ step
nilpotency, whilst also introducing a simpler form for his explicit
solutions $X_t^{x}=u(x,t,(W_t^I)_{I\in F})$. Here, $u$ solves a
differential equation, and $(W_t^I)_{I\in F}$ are iterated
Stratonovich integrals with integrands and integrators selected
from $\left( t,W_t^1,...,W_t^d\right)$. Another substantial work
on explicit solutions to stochastic differential equations is due to
\citet{Kunita:1984}[Section III.3]. He considers representing
solutions to time-homogeneous Fisk-Stratonovich equations via flows
generated by the coefficients of the equation under a commutative
condition similar to ours, and, more generally, under solvability of
the underlying Lie algebra. Kunita's work therefore generalizes
\citet{Yamato:1979}. Perhaps, the two most distinguishing features
of our work are: We allow time-dependent coefficients and utilize a
different representation that is very useful in simulation and
other applications (see e.g.\ \cite{Kouritzin:1998}, \cite{Kouritzin16}).
We compare our results to \citet{Yamato:1979} and \citet{Kunita:1984} in
Section \ref{kun}.
In order to describe our method, we mention that the hitherto rather
ad hoc, state-space diffeomorphism mapping method has been used to
construct solutions to interesting stochastic differential equations
from solutions to simpler ones.
The idea of this method is to change
the infinitesimal generator $L$ of a simple It\^o process to the
generator $\mathcal L$ corresponding to a more complicated It\^o process via ${
\mathcal L}f(x)= \{L(f\circ \Lambda^{-1} )\}\circ \Lambda(x)$.
This corresponds to using It\^o's formula on $X_t = \Lambda^{-1}(\xi
_t)$, where
$\xi$ is a diffusion process with infinitesimal generator $L$.
For related
examples, we refer the reader to the problems in
\citet{Friedman:2006}[page 126] or \citet{Ethier/Kurtz:1986}[page
303].
Motivated by applications in filtering, \citet{Kouritzin/Li:2000}
and \citet{Kouritzin:2000} used differential equation methods to
study: ``When can global, time-dependent diffeomorphisms be used to
construct solutions to It\^o equations?", ``What scalar It\^o
equations can be solved via diffeomorphisms?", and ``How can one
construct these diffeomorphisms?". They considered scalar solutions
in an open interval $D$
to the time-homogeneous stochastic differential equation
\begin{equation} \label{SDE1}
dX_t=b(X_t)dt+\sigma (X_t)dW_t,\ X_0=x,
\end{equation}
which are of the form $\phi^{x} \left(\int_0^tU_udW_u,t\right)$,
and showed that all nonsingular solutions of this form were actually
(time-dependent) diffeomorphisms $\Lambda^{-1} _t(\xi _t)$ with $\xi $
satisfying
$$
d\xi _t=( \chi -\kappa \xi_t)dt+dW_t,\ \xi _0 =\Lambda_0(x).
$$
Nonsingular in this scalar case was interpreted as finiteness of $
\int_\lambda ^y\sigma ^{-1}(x)dx$ for some fixed point $\lambda $
and all $ y\in D$.
(Their methods involve non-stochastic differential equations
that can continue to hold in the singular situations when global
diffeomorphisms fail.)
For our current work, we suppose henceforth that
$D\subset \mathbb R^p$ is a bounded convex domain, $T>0,$
and define
$$
D_T=\left\{\begin{array}{ll} D& \mbox{ if } \sigma,b \mbox{ do not depend on t}\\
D\times [0,T)& \mbox{ if either do}
\end{array} \right.
$$
so $(x,s)\in D_T$ means $x\in D$ when $D_T=D$.
Then, we resolve the question: ``When can
we explicitly solve vector-valued It\^o equations
\begin{equation} \label{VSDE1}
dX_t=b(X_t,t)dt+\sigma (X_t,t)dW_t,\ X_s=x,
\end{equation}
with the dimensions of $X_t, W_t$ being $p,d$ respectively, through
representations of the form $X_t^{x,s}= \phi^{x,s}\left(\int_s^t
U_{s,u} dW_u,t\right)$?''. This question is more precisely broken into
two separate important questions: ``For which $\sigma$ and $b$ does
such a strong-local-solution representation exists?" and ``What
conditions are required on $\phi$ and $U$ for such representations
with $\int_s^tU_{s,u} dW_u=\int_s^tU_{s,u}(X_u) dW_u$ still being Gauss-Markov?"
Equivalently, we consider ``When can the solutions to the
Fisk-Stratonovich equation
\begin{equation}\label{Strato}
dX_t^{x}=h(X_t^{x},t)dt+\sigma (X_t^{x},t)\bullet dW_t,
\end{equation}
with
\begin{equation}
h=b-{\frac 12}\sum_{j=1}^d\{\nabla_\varphi \sigma _j\}\sigma _j\mbox{ on }D_T
\label{bh}
\end{equation}
and $\sigma _j$ denoting the $j^{\rm th}$ column of the matrix $\sigma$, be
locally represented in this manner?" It follows from, for example,
\citet{Kunita:1984}[p. 239] that the unique local solutions to these
(\ref{VSDE1}) and (\ref{Strato}) are equal if (\ref{bh}) holds and
$\sigma $ is twice continuously differentiable or satisfies the
Fisk-Stratonovich acceptable condition in $D$, the latter being
discussed in \citet{Protter:2004}[Chapter 5]. We work with It\^o
equations to avoid these stronger assumptions on $\sigma $ but still
relate $b$ and $h$ through (\ref{bh}). Also, to obtain simple,
concrete necessary and sufficient conditions for such a
representation, we consider all solutions starting from each
$(x,s)\in D_T$. Actually, assuming natural regularity conditions
and using differential form techniques, we obtain very satisfying
answers to these question by showing the equivalence of the
following three conditions: 1) The SDEs (\ref{VSDE1}) have our
local-solution-representations for all starting points $(x,s)\in
D_T$. 2) The representation pair $\phi^{x,s}, U_{s,t}$ satisfy a
system of differential equations. 3) The SDE coefficients
$\sigma$ and $h$ satisfy simple commutator conditions.
In the process of establishing this three-way equivalence, we also answer
the question ``When is (\ref{VSDE1}) locally diffeomorphic to an
SDE with a simple diffusion coefficient?" i.e.\ ``When will it have
a representation as in (\ref{ODE1},\ref{SDE3}) to follow?".
It turns out that this representation facilitates explicit weak solution
of the important financial Heston model as is shown in
\cite{Kouritzin16}.
Given precise conditions of when an It\^{o} equation has such a
representation, the next natural questions we answer are: ``What form
do the solutions have?" and ``How do you construct such solutions?"
In order to include as many interesting examples as
possible we will only require \emph{local} representation $X_t^{x,s}=\phi^{x,s}
\left(\int_s^tU_{s,u} dW_u,t\right)$ and allow $\sigma $ to have rank
less than $\min(p,d)$. The first opportunity borne out of allowing
the rank of $\sigma (x)$ to be less than $p$ is the ability to
handle time-dependent coefficients, treating time as an extra state.
The second advantage from allowing lesser rank than $\min(p,d)$ is
the extra richness afforded by appending a deterministic equation
into the diffeomorphism solution.
A third, important benefit of this general rank condition is
the possibility of producing explicit \emph{weak} solutions to
SDEs where no explicit strong solution exists (see \cite{Kouritzin16}).
In our construction results, we show that $\phi$ is constructed via
a time-dependent diffeomorphism $\Lambda_t$, which in turn is defined in terms of $\sigma$.
The diffeomorphism separates a representable SDEs into deterministic and stochastic differential equations:
$\Lambda_t(X_t)=(\overline X_t,\widetilde X_t)$, where
$\widetilde X_t \in \mathbb R^{p-r}$ is deterministic and satisfies
the differential equation
\begin{equation}\label{ODE1}
\frac{d}{dt}\widetilde X_t = \widetilde h(\widetilde X_t,t),
\end{equation}
while $\overline X_t$ is a
Gauss-Markov process satisfying
\begin{equation}\label{SDE3}
d\overline X_t = (\overline \theta(\widetilde X_t,t)+\overline \beta(\widetilde X_t,t)\overline X_t) dt+ \left(I_r\Big{|}\; \overline \kappa(\widetilde X_t,t)\right)dW_t.
\end{equation}
$\overline \kappa$ is determined (within an equivalence class) by $\sigma$ while
$\overline \theta$, $\widetilde h$ and
and $\overline \beta$ can be anything (subject to dimensional and differentiability regularity conditions).
These parameters allow us to handle a whole class of \emph{nonlinear} drift coefficients $b$ for a given $\sigma$
in the SDE (\ref{VSDE1}) for $X_t=\Lambda_t^{-1}(\overline X_t,\widetilde X_t)$.
In the next section, we introduce notation and state the main existence
results. In Section 3, we build off of these existence results to give our
construction results, illustrated with simple applications.
We compare our work to prior work of Yamato and Kunita in Section 4.
The proofs of all main results are postponed to Section 5.
\section{Notation and Existence Results} \label{main}
Let $(W_t)_{t\ge 0}$ be a standard $d$-dimensional Brownian motion
with respect to filtration $\{{ \mathcal F}_t\}_{t\ge 0}$ satisfying
the usual hypotheses on a complete probability space $(\Omega,{ \mathcal F},P)$.
We will use $\phi$ to denote a representation function and $x$ to denote a starting point as in the introduction.
On the other hand, $\varphi$ will denote a variable with the same dimension $p$ as $\phi$ and $x$.
For functions of time or paths of a stochastic process, we use $Z_t$ and $Z(t)$ interchangeably.
For a matrix $V$, $V_j$ will denote its $j^{\rm th}$ column vector and $V_{i,j}$ the
$i^{\rm th}$ element of this $j^{\rm th}$ column.
$B_z(\delta)$ denotes an open Euclidean ball centered at $z$ with radius $\delta>0$.
Suppose $m,r\in \mathbb N$, $O\subset \mathbb R^m$ is open and $I\subset [0,T)$ is an interval.
Then, $C(I)$ is the continuous functions on $I$ and $C^r(O)$ denotes the continuous functions whose partial derivatives up to order $r$
exist and are continuous on $O$.
Moreover, $C^{r,1}(O\times I)$ denotes the continuous functions $g(\varphi,t)$ whose mixed
partial derivatives in $\varphi\in O$ up to order $r$
and in $t\in I$ up to order $1$ all exist and are continuous functions on $O\times I$.
$C^{1}(O\times I)=C^{1,0}(O\times I)\cap C^{0,1}(O\times I)$.
(We only require one-sided derivatives in time to exist at interval endpoints.)
For such functions of both $\varphi$ and $t$,
$\nabla_\varphi g $ is the Jacobian matrix
of vector function $g$, that is $\left( \nabla_\varphi g \right)
_{i,j}=\partial _{\varphi_j}g_i$, while $\nabla g$ will include the
time derivative as the last column.
The purpose of our representations is to simulate a large class of processes
in an efficient manner, which leads to a dilemma.
We would like to allow $U_{s,t}$ to depend upon $X^{x,s}$ for generality
but not in a way that would destroy the ease of simulation.
Our approach to this dilemma is to allow $U_{s,t}$ to act as an operator on the functions
$\phi^{x,s }(y_u,u)\big|_{u\in[s,t]}$ but then impose the condition
that the result $U_{s,t}\phi^{x,s }(y_\cdot,\cdot)$ can not depend upon $y$.
As we will expose below, this basically allows $U_{s,t}$ to depend upon
some hidden deterministic part of $X$ but not the purely stochastic part, saving
the Gaussian nature of
\begin{equation}\label{Ydefn}
Y^s_t=\int_s^t U_{s,u}\phi(Y_\cdot^s,\cdot)dW(u)=\int_s^t U_{s,u}\phi(0,\cdot)dW(u)
\end{equation}
so it can be computed off-line, which is the point of this work.
Then, $\phi$ must be differentiable enough to apply
It\^{o}'s formula and allow room for random process $Y^{s}_t$ to move.
Finally, we want $U_{s,t}\phi$ to satisfy some type of simple state
equation so it is easy to compute.
The precise regularity conditions for potential representations
$X_t^{x,s}= \phi^{x,s}\left(Y^s_t,t\right)$ now follow:
\begin{itemize}
\item[$\mathcal{C}_1$: ] For each $(x,s)\in D_T$, there is a $t_0=t_0^{x,s}>s$ and
a convex neighbourhood ${\mathcal N}^{x,s}\subset \mathbb R^d$ of $0$
such that $\phi^{x,s } \in C^{2,1}({\mathcal N}^{x,s}\times [s,t_0); \mathbb R^p)$ and
$t\rightarrow U_{s,t}\phi^{x,s }(y_\cdot,\cdot) \in C^{1}( [s,t_0);\mathbb R^{d\times d})$.
\item[$\mathcal{C}_2$: ] $\phi ^{x,s},U_{s,t}$ start correctly
\begin{equation}
\phi ^{x,s}(0,s) = x,\ \ U_{s,s}\phi^{x,s}(0,s)=I_d\ \forall (x,s)\in D_T. \label{initphi}
\end{equation}
\item[$\mathcal{C}_3$: ] $U_{s,t}\phi^{x,s}$ is non-singular on ${\mathcal N}^{x,s}\times [s,t_0)$ (with matrix inverse denoted $U^{-1}_{s,t}\phi^{x,s}$) and satisfies
\begin{equation}\label{Urestrict}
U_{s,t}\phi^{x,s }(y_u,u)=U_{s,t}\phi^{x,s }(0,u)
\end{equation}
as well as
\begin{equation}\label{Ugroup}
\ \ \ \ \ \ \ \ \ U^{-1}_{s,t}\phi^{x,s}(y_t,t)\frac{d}{dt}U_{s,t}\phi^{x,s}(y_u,u)\big|_{u=t}
=\frac{d}{dt}U_{u,t}\phi^{\phi^{x,s}(y_u,u),u}(y_u,u)\big|_{u=t}.
\end{equation}
\end{itemize}
Then, (\ref{initphi},\ref{Ugroup}) imply
\begin{equation}\label{Ugroup1}
U^{-1}_{s,t}\phi^{x,s}(y_t,t)\frac{d}{dt}U_{s,t}\phi^{x,s}(y_u,u)\big|_{u=t}
=U^{-1}_{t,t}\phi^{\phi^{x,s}(y_t,t),t}\frac{d}{dt}U_{u,t}\phi^{\phi^{x,s}(y_u,u),u}\big|_{u=t}
\end{equation}
and therefore that $U$ is a (two parameter) semigroup.
We use (\ref{Urestrict}) to economize the notation $U_{s,t}\phi^{x,s }(y_\cdot,\cdot)$
to $U_{s,t}\phi^{x,s}$.
Now, define the $\mathcal{F}_t$-stopping time
\[
\tau^{x,s} = \min\left( t_0^{x,s},\inf\{t > s: \;
Y^{s}_t\notin{\mathcal N}^{x,s} \mbox{ or } (\phi^{x,s}(Y^{s}_t,t),t) \not \in D_T
\}\right)
\]
and let
\begin{equation} \label{Rdefn}
\mathcal R^{x,s}=\mathop{\cup}\limits_{t\ge 0}\left\{(y,t):P((Y^{s}_t,t)\in B_{(y,t)}(\delta),t\le\tau^{x,s} )>0\ \forall \delta>0\right\}.
\end{equation}
There is structure that can be imposed upon $\phi,U$ that will turn out to be equivalent to the existence of our explicit strong local solutions.
\begin{defn}\label{RepPair}
An $(x,s,\sigma,h)$-representation is a pair $\phi^{x,s},U_{s,t}$ satisfying ($\mathcal{C}_1, \mathcal{C}_2, \mathcal{C}_3$) such that the following system of differential equations:
\begin{eqnarray}
\nabla _y\phi^{x,s}(y,t)& = & \sigma (\phi^{x,s}(y,t),t) U^{-1}_{s,t}\phi^{x,s}, \label{grad} \\
\partial _t\phi^{x,s}(y,t) & = & h(\phi^{x,s}(y,t),t) \label{dtphi}
\end{eqnarray}
hold for all $(y,t)\in{\mathcal R}^{x,s}$ and $\partial_s \nabla_y \phi^{x,s}(0,s)$,
$\partial_s \partial_t \phi^{x,s}(0,s)$, $\partial_{x_i} \nabla_y \phi^{x,s}(0,s)$
and $\partial_{x_i} \partial_t \phi^{x,s}(0,s)$ exist as continuous functions.
Here and below, $\partial_t \phi^{x,s}(0,s)$ means $\partial_t \phi^{x,s}(0,t)\big|_{t=s}$.
\end{defn}
Now, our explicit solutions are:
\begin{equation} \label{rep}
X^{x,s}_t=\phi (Y_t,t)=\phi^{x,s}(Y^{s}_t,t) \mbox{ on }
[s,\tau^{x,s}).
\end{equation}
Our first main result establishes two necessary and sufficient conditions
for all $X^{x,s}$, defined in (\ref{rep}), to be strong local solutions to
\begin{equation} \label{SDE4}
dX_t=b(X_t,t)dt+\sigma (X_t,t)dW_t,\qquad X_s=x
\end{equation}
on $[s,\tau^{x,s})$.
The function $h$ is
always related to $b$ through (\ref{bh}) and $U_{s,t}\phi^{x,s}$ comes into the necessary and sufficient commutator conditions through generator
\begin{equation} \label{SecondA}
A(x,s)=\frac{d}{dt} U_{s,t}\phi^{x,s}\big|_{t=s}.
\end{equation}
It follows from (\ref{Urestrict}) that $A$ does not depend upon $y$.
\medskip
\begin{thm}\label{r1}
The following are
equivalent:
\begin{enumerate}
\item[a)]
$\sigma\in C^1(D_T;\mathbb R^{p\times d})$, $h\in C^1(D_T;\mathbb R^p)$,
there is a unique strong solution to (\ref{SDE4}) on $[s,\tau^{x,s})$
for each $(x,s)\in D_T$, and this solution has explicit form
$\phi^{x,s}(Y_t^s,t)$ with $\phi^{x,s},U_{s,t}$ satisfying $\mathcal{C}_1,\mathcal{C}_2,\mathcal{C}_3$.
\item[b)] There is a $(x,s,\sigma,h)$-representation $\phi^{x,s},U_{s,t}$ for each $(x,s)\in D_T$.
\item[c)]
$\sigma\in C^1(D_T;\mathbb R^{p\times d})$, $h\in C^1(D_T;\mathbb R^p)$ and
the following commutator conditions hold on $D_T$:
\begin{equation}\label{braket2}
(\nabla_\varphi \sigma _k)\sigma _j=(\nabla_\varphi \sigma _j)\sigma _k,
\mbox{ for all } j,k\in \{1,...,d\},
\end{equation}
\begin{equation} \label{h}
(\nabla_\varphi h)\sigma _j
=(\nabla_\varphi \sigma _j)h+\partial_t\sigma_j-\sigma A_j,\mbox{ for all } 1\le j\le d.
\end{equation}
\end{enumerate}
\end{thm}
\begin{RM}\label{timeinvcase}
Theorem \ref{r1} simplifies in the time-invariant $h,\sigma$ coefficient case.
Clearly, one only needs to check the commutator conditions on $D$ versus $D_T$.
However, the second commutator condition actually changes in form to:
\begin{equation} \label{h-timeinv}
(\nabla_\varphi h)\sigma _j
-(\nabla_\varphi \sigma _j)h=\sigma B_j,\mbox{ for all } 1\le j\le d,
\end{equation}
where $B(\varphi)=-A(\varphi,0)$.
Indeed, the left hand side of (\ref{h-timeinv}) does not depend on time
so the right side can not either.
\end{RM}
\begin{RM}
Theorem \ref{r1} also simplifies when $d=1$, which corresponds to appending a deterministic equation and
allowing time dependence to the case considered in \citet{Kouritzin:2000}.
In this $d=1$ case, (\ref{braket2}) is automatically true and (\ref{h}) becomes
\begin{equation} \label{hd1}
(\nabla_\varphi h)\sigma
=(\nabla_\varphi \sigma )h+\partial_t\sigma-\sigma A.
\end{equation}
\end{RM}
Often, we are interested in establishing the representation for a given
stochastic differential equation.
In this case, the commutator conditions can be used quickly to determine if such
a representation is possible.
The easiest way to ensure (\ref{braket2}) is to have each column a constant multiple of another
$\sigma_j=c_j\sigma_1$ for all $j$ say.
However, there are other possibilities.
\begin{ex}
Let $p=d=2$ and $D\subset\mathbb R$ be a domain.
Suppose $a,e,f,g,m,n$ are $C^2(D)$-functions and our Fisk-Stratonovich equation has time-invariant coefficients:
\begin{equation}
h(\varphi_1,\varphi_2)=\left(\begin{matrix}f\left(\varphi_{1}\right)g\left(\varphi_{2}\right)\\m\left(\varphi_{1}\right)n\left(\varphi_{2}\right)\end{matrix}\right), \
\sigma \left(\varphi_{1},\varphi_{2}\right)=
\left(\begin{matrix}a\left(\varphi_{1}\right)&0\\e\left(\varphi_{2}\right)&e\left(\varphi_{2}\right)\end{matrix}\right).
\end{equation}
Moreover, suppose $a(\varphi_1)$ and $e(\varphi_2)$ are never $0$.
Then, $\sigma$ is always non-singular and it follows by (\ref{grad})
as well as the mean value theorem that for any $u\in[s,t]$
\[
\phi^{x,s}(y,u)-\phi^{x,s}(\hat y,u)=\sigma(\phi^{x,s}( y^*,u))
U^{-1}_{s,u}\phi^{x,s}\cdot(y-\hat y)
\]
with $y^*\in\mathcal N^{x,s}$ for $y,\hat y\in\mathcal N^{x,s}$ and any possible representation $\phi^{x,s},U_{s,t}$.
Hence, $\phi^{x,s}(y,u)=\phi^{x,s}(\hat y,u)\leftrightarrow y=\hat y$.
Therefore, it follows from (\ref{Urestrict}) that
$U_{s,u}$ can not depend upon $\phi^{x,s}( y,u)$
for any $u\in[s,t]$ and
$B$ in (\ref{h-timeinv}) is constant by (\ref{SecondA}).
Now,
\begin{equation}
\nabla_\varphi h =\left(\begin{matrix}f'(\varphi_1)g(\varphi_2)&f(\varphi_1)g'(\varphi_2)\\m'(\varphi_1)n(\varphi_2)&m(\varphi_1)n'(\varphi_2)\end{matrix}\right)
\end{equation}
and
\begin{equation}
\nabla_\varphi\sigma_2 =\left(\begin{matrix}0&0\\0&e'\left(\varphi_{2}\right)\end{matrix}\right),\ \nabla_\varphi\sigma_1 =\left(\begin{matrix}a'(\varphi_1)&0\\0&e'\left(\varphi_{2}\right)\end{matrix}\right)
\end{equation}
so the first commutator condition (\ref{braket2}) is fine since
\begin{equation}
\nabla_\varphi\sigma_1\sigma_2 =\left(\begin{matrix}0\\e'\left(\varphi_{2}\right)e(\varphi_2)\end{matrix}\right)=\nabla_\varphi\sigma_2 \sigma_1.
\end{equation}
Moreover,
\begin{equation}
\nabla_\varphi h\sigma_2-\nabla_\varphi \sigma_2 h =\left(\begin{matrix}e(\varphi_2)f(\varphi_1)g'(\varphi_2)\\m(\varphi_1)(e(\varphi_2)n'(\varphi_2)-e'(\varphi_2)n(\varphi_2))\end{matrix}\right)
\end{equation}
and
\begin{equation}
\nabla_\varphi h\sigma_1-\nabla_\varphi \sigma_1 h =\left(\begin{matrix}af'g+efg'-a'fg\\am'n+emn'-e'mn\end{matrix}\right).
\end{equation}
On the other hand, denoting $B=\left(\begin{matrix}b_{11}&b_{12}\\b_{21}&b_{22}\end{matrix}\right)$, we have
\begin{equation}
\sigma B =\left(\begin{matrix}ab_{11}&ab_{12}\\eb_{11}+eb_{21}&eb_{12}+eb_{22}\end{matrix}\right).
\end{equation}
Hence, by (\ref{h-timeinv}) there is an explicit solution if and only if
\begin{equation}
\!\!\!\left(\begin{matrix}af'g+efg'-a'fg&efg'\\am'n+emn'-e'mn&m(en'-e'n)\end{matrix}\right)=\left(\begin{matrix}ab_{11}&ab_{12}\\eb_{11}+eb_{21}&eb_{12}+eb_{22}\end{matrix}\right)
\end{equation}
for constants $b_{11},b_{12},b_{21},b_{22}$.
If $f=c_1a$, $n=c_2e$, $eg'=c_3$ and $m'a=c_4$ for some constants $c_1,c_2,c_3,c_4$, then it is easy to show that this condition
is met with $b_{22}=-c_1c_3$, $b_{21}=c_2c_4-c_1c_3$ and $b_{11}=b_{12}=c_1c_3$ so the representation holds for
\begin{equation}
h(\varphi_1,\varphi_2)=\left(\begin{matrix}\alpha\frac{g\left(\varphi_{2}\right)}{m'\left(\varphi_{1}\right)}\\\beta \frac{m\left(\varphi_{1}\right)}{g'\left(\varphi_{2}\right)}\end{matrix}\right), \
\sigma \left(\varphi_{1},\varphi_{2}\right)=\left(\begin{matrix}\frac{\gamma}{m'\left(\varphi_{1}\right)}&0\\\frac{\delta}{g'\left(\varphi_{2}\right)}&\frac{\delta}{g'\left(\varphi_{2}\right)}\end{matrix}\right),
\end{equation}
where $\alpha=c_1c_4,\beta=c_2c_3,\gamma=c_4,\delta=c_3$ are any constants and $g,m$ are $C^2$-functions with $\frac{1}{m'\left(\varphi_{1}\right)},\frac{1}{g'\left(\varphi_{1}\right)}\in C^1(D)$.
\end{ex}
\begin{ex}
In a similar manner, it follows that
\begin{equation}
h(\varphi_1,\varphi_2)=\left(\begin{matrix}\alpha\frac{g\left(\varphi_{2}\right)}{m'\left(\varphi_{1}\right)}\\\beta \frac{m\left(\varphi_{1}\right)}{g'\left(\varphi_{2}\right)}\end{matrix}\right), \
\sigma \left(\varphi_{1},\varphi_{2}\right)=\left(\begin{matrix}\frac{\gamma}{m'\left(\varphi_{1}\right)}&0\\0&\frac{\delta}{g'\left(\varphi_{2}\right)}\end{matrix}\right),
\end{equation}
for any constants $\alpha,\beta,\gamma,\delta$, also has a representation.
\end{ex}
There was significant work done in the previous examples and we still did not have
the representation functions.
The next example is the key to solving for complete representations and will be used in the following section.
\begin{ex}\label{CoreEG}
Suppose
$\sigma(\varphi,t) = \left(\begin{array}{cc} I_r & \overline \kappa(\varphi,t)\\
0 & 0 \end{array}\right) \in \mathbb R^{p\times d}$ satisfies (\ref{braket2}).
We will find the possible $h,b$ satisfying (\ref{h}) and the corresponding representations $U_{s,t},\phi^{x,s}$ by Theorem \ref{r1}.\\
{\bf Notation:}
As always, $\varphi$ is a variable and $\phi$ is the representation function.
Further, let $\overline{x}=(x_1,...,x_r)$, $\widetilde{x}=(x_{r+1},...,x_d)$,
$\overline{\varphi}=(\varphi_1,...,\varphi_r)$, $\widetilde{\varphi}=(\varphi_{r+1},...,\varphi_d)$,
$\widetilde D = \{\widetilde \varphi:(\overline \varphi,\widetilde \varphi) \in D\ \text{for some}\ \overline \varphi\}$,
$\widetilde D_T =\widetilde D \times[0,T)$,
\begin{equation}
\phi^{x,s}(y,t)=\left(\begin{array}{c}\overline\phi^{x,s}(y,t)\\\widetilde\phi^{x,s}(y,t)\end{array}\right),\ h=\left(\begin{array}{c}\overline h\\\widetilde h\end{array}\right)\ \text{and}\ A=\left(\begin{array}{cc}A_{11}&A_{12}\\A_{21}&A_{22}\end{array}\right),
\end{equation}
where $A_{11}\in\mathbb R^{r\times r}$.
Finally, we let
\begin{equation}\label{betadef}
\overline \beta(\varphi,t)=-A_{11}(\varphi,t)-\overline \kappa(\varphi,t)A_{21}(\varphi,t),
\end{equation}
which will appear often below.\\
{\bf Step 1:} Interpret (\ref{grad}) and $\mathcal C_2$ condition (\ref{Urestrict}) on $U_{s,t},A$.\\
Suppose $u\in[s,t]$.
By (\ref{grad}) as well as the mean value theorem
\begin{equation}
\left(\!\begin{array}{c}\overline\phi^{x,s}(y,u)-\overline\phi^{x,s}(\hat y,u)\\\widetilde\phi^{x,s}(y,u)-\widetilde\phi^{x,s}(\hat y,u)\end{array}\!\right)=\left(\!\begin{array}{cc} I_r & \overline \kappa(\phi^{x,s}(y^*,u),u)\\
0 & 0 \end{array}\!\right)
U^{-1}_{s,u}\phi^{x,s}\cdot(y-\hat y)
\end{equation}
with $y^*\in\mathcal N^{x,s}$ for $y,\hat y\in\mathcal N^{x,s}$ and any possible representation $\phi^{x,s}$.
Hence, $\overline\phi^{x,s}(y,u)\neq\overline\phi^{x,s}(\hat y,u)$ implies $y\neq\hat y$.
Therefore, it follows from (\ref{Urestrict}) that
$U_{s,t}\phi^{x,s}$ can not depend upon $\overline\phi^{x,s}( y,u)$
for any $u\in[s,t]$,
which implies $U_{s,t}\phi\doteq U_{s,t}\widetilde \phi$ only depends on $\widetilde \phi,t$.
This also means by (\ref{SecondA}) that
\begin{equation}\label{FirstA}
A(\varphi,t)=\frac{d}{dt} U_{u,t}\widetilde \phi^{\varphi,u}\big|_{u=t}.
\end{equation}
{\bf Step 2:} Interpret commutator conditions on $\overline \kappa,h$.\\
Let $e_i$ denote the $i^{th}$ column of $I_{p}$ so $\sigma_i=e_i$ for $i\le r$.
We have by (\ref{braket2}), that
\begin{equation}
\left(\begin{array}{c}\nabla_\varphi\overline \kappa_{j-r}\\0\end{array}\right)e_i=0\ \ \forall i\in\{1,2,...,r\},j\in{r+1,...,d},
\end{equation}
which establishes that $\overline \kappa(\widetilde \varphi,t)$ can only depend upon $\widetilde \varphi,t$.
This is the only restriction on $\overline \kappa$ from (\ref{braket2}).
By (\ref{h}), we find
\begin{equation}
\nabla_\varphi\left(\begin{array}{c}\overline h\\\widetilde h\end{array}\right)\sigma_j-\nabla_\varphi\sigma_j\left(\begin{array}{c}\overline h\\\widetilde h\end{array}\right) =\left(\begin{array}{cc}\overline \beta&\partial_t\overline \kappa-A_{12}-\overline \kappa A_{22}\\0&0\end{array}\right)_j
\end{equation}
so $\nabla_{\overline \varphi} \widetilde h=0$, implying $\widetilde h(\varphi)\in C^1(\widetilde D_T,\mathbb R^{p-r})$ only depends upon $\widetilde \varphi,t$, and
\begin{equation}\label{SpecCom1}
\nabla_{\overline{\varphi}} \overline h=\overline \beta,
\end{equation}
\begin{equation}\label{SpecCom2}
\overline \beta\, \overline \kappa=[(\nabla_{\widetilde \varphi}\overline \kappa_1)\widetilde h,...,(\nabla_{\widetilde \varphi}\overline \kappa_{d-r})\widetilde h] +\partial_t\overline \kappa-A_{12}-\overline \kappa A_{22}.
\end{equation}
Hence, it follows from (\ref{grad},\ref{dtphi},\ref{initphi}) that $\widetilde\phi^{x,s}$ satisfies
\begin{eqnarray}
\nabla _y\widetilde\phi^{x,s}(y,t)& = & 0, \label{vargrad} \\
\partial _t\widetilde\phi^{x,s}(y,t) & = & \widetilde h(\widetilde\phi^{x,s}(y,t),t), \label{dtivarphi}\\
\widetilde\phi ^{x,s}(0,s) & = & \widetilde x, \label{initvarphi}
\end{eqnarray}
which implies that $\widetilde \phi$ does not depend upon $\overline \phi$ nor $y$.
Moreover, by (\ref{FirstA}) and (\ref{betadef}), we conclude that $A(\varphi,t)\doteq A(\widetilde \varphi,t)$
and $\overline \beta(\varphi,t)\doteq\overline \beta(\widetilde \varphi,t)$ only depend on $\widetilde \varphi,t$.\\
{\bf Step 3:} Determine possible $h,b$.\\
By (\ref{SpecCom1}), we find
\begin{equation}\label{h1}
\overline h(\overline \varphi,\widetilde \varphi,t)= \overline \beta(\widetilde \varphi,t)\overline \varphi+\overline \theta(\widetilde \varphi,t)
\end{equation}
for some $C^1$-function $\overline \theta$.
Hence, the possible $h(\overline \varphi,\widetilde \varphi,t)=\left(\begin{array}{c}\overline h(\overline \varphi,\widetilde \varphi,t)\\\widetilde h(\widetilde \varphi,t)\end{array}\right)$ are:
\begin{equation}\label{possibleh}
\begin{array}{c}\widetilde h\in C^1(\widetilde D_T,\mathbb R^{p-r}),\\
\overline h\in \left\{\overline \theta(\widetilde \varphi,t)+
\overline \beta(\widetilde \varphi,t)\overline \varphi:\overline \beta\in C^1(\widetilde D_T,\mathbb R^{r\times r});\overline \theta\in C^1(\widetilde D_T,\mathbb R^r)\!\right\}
\end{array}.
\end{equation}
From (\ref{bh}) and fact $\overline \kappa(\widetilde \varphi,t)$ only depends on $\widetilde \varphi,t$, we find that
\begin{equation}
b=h+{\frac 12}\sum_{j=1}^d\{\nabla_\varphi \sigma _j\}\sigma _j=h.
\end{equation}
{\bf Free Parameters:} $A_{21}$, $A_{22}$, $\overline \kappa$, $\overline \beta$, $\overline \theta$ and $\widetilde h$ can be anything (subject to dimensionality and dependency on only $\widetilde \varphi,t$).
$A_{12}$ is then determined by (\ref{SpecCom2})
and $A_{11}$ by (\ref{betadef}).
$\overline \beta$ and $\overline \theta$ also determine the possible $\overline h$ above and $\phi^{x,s}$ below.
Different choices of $\overline \kappa$, $\overline \beta$, $\overline \theta$ and
$\widetilde h$ will result in different solutions.
However, there is no loss in generality in taking $A_{21},A_{22}$ to
be zero.\\
{\bf Step 4:} Interpret differential system for $\phi^{x,s}$.\\
Since $\phi^{x,s}=\left(\begin{array}{c}\overline \phi\\\widetilde \phi\end{array}\right)$ satisfies (\ref{dtphi},\ref{initphi}), $\widetilde \phi$ must be of the form
\begin{equation}\label{phi2evolve}
\partial_t\widetilde \phi = \widetilde h(\widetilde \phi,t),\ \ \text{s.t.}\ \widetilde \phi(s)=\widetilde x.
\end{equation}
We let $\widetilde X_t$ denote the solution of this differential equation.
Next, since $\phi$ satisfies (\ref{grad}), $\overline \phi$ must be of the form
\begin{equation}\label{phiphi}
\overline \phi^{x,s}(y,t) = \overline c(t) + \left[ I_r \ \ \overline \kappa( \widetilde X_t,t)
\right]U^{-1}_{s,t}\widetilde \phi^{x,s}y,
\end{equation}
for some $\overline c \in C^1([0,T);\mathbb R^r)$.
Differentiating in $t$, noting by (\ref{FirstA}) (with $\widetilde \varphi=\widetilde X_t$) that
\begin{equation}\label{AAsim}
A(\widetilde X_t,t)=
\frac{d}{dt} U_{u,t}\widetilde \phi^{\widetilde X_u,u}\big|_{u=t},
\end{equation}
and using (\ref{phiphi},\ref{AAsim},\ref{Ugroup},\ref{phi2evolve},\ref{SpecCom2},\ref{betadef}), one has (with $U^{-1}_{s,t}=U^{-1}_{s,t}\widetilde \phi^{x,s}$) that
\begin{eqnarray}\label{partialtphi}
\!\!&\!\!\!&\!\!\!\partial_t \overline \phi(y,t)\\\nonumber
&\!\!\!=&\!\!\! \overline c' (t) -\left[ I \ \
\overline \kappa( \widetilde X_t,t)\right]A(\widetilde X_t,t) U^{-1}_{s,t} y
\\
&\!\!\!+&\!\!\!\left[ 0 \ \partial_t\overline \kappa( \widetilde X_t,t)+\nabla_{\widetilde \varphi}\overline \kappa_1(\widetilde X_t,t)\, \widetilde h(\widetilde X_t,t),...,\nabla_{\widetilde \varphi}\overline \kappa_{d-r}(\widetilde X_t,t)\, \widetilde h(\widetilde X_t,t)
\right]U^{-1}_{s,t} y\ \ \nonumber\\
&\!\!\!=&\!\!\!\overline c' (t) +\overline \beta(\widetilde X_t,t)[I\ \ \overline \kappa(\widetilde X_t,t)]U^{-1}_{s,t} y\nonumber\\
&\!\!\!=&\!\!\!\overline c' (t) +\overline \beta(\widetilde X_t,t)(\overline \phi(y,t)-\overline c(t))\nonumber.
\end{eqnarray}
On the other hand, by (\ref{dtphi}) and (\ref{h1})
\begin{equation}\label{phipartialt}
\partial_t \overline \phi(y,t)= \overline \theta(\widetilde X_t,t)+\overline \beta(\widetilde X_t,t)\overline \phi(y,t).
\end{equation}
Comparing (\ref{partialtphi}) and (\ref{phipartialt}), one has that
\begin{equation}\label{ceqn}
\overline c'(t)=\overline \theta(\widetilde X_t,t)+\overline \beta(\widetilde X_t,t)\overline c(t)\ \ \text{subject to }\overline c(s)=\overline{x}.
\end{equation}
{\bf Step 5:} Determine $U$ in terms of $\overline \kappa$, $\overline \beta$ and $\theta$.\\
We just need $A$ to satisfy (\ref{betadef},\ref{SpecCom2}) so there is no loss of generality in taking
\begin{equation}\label{Aderived}
\!\!\!\!\!\!\left(\!\!\begin{array}{cc}A_{11}&A_{12}\\A_{21}&A_{22}\end{array}\!\!\right)(\widetilde\varphi,t)
=\left(\!\begin{array}{cc} -\overline \beta & {
[(\nabla_{\widetilde\varphi}\overline \kappa_1)\widetilde h,...,(\nabla_{\widetilde\varphi}\overline \kappa_{d-r})\widetilde h]+\partial_t\overline \kappa
-\overline \beta\,\overline \kappa
} \\0& 0 \end{array}\!\right)(\widetilde\varphi,t).
\end{equation}
By (\ref{AAsim}), (\ref{Ugroup}) and (\ref{Aderived}), we know
\begin{eqnarray}\label{MatUdef}
\!\!\partial_tU_{s,t}\widetilde X&\!\!\!=&\!\!\!(U_{s,t}\widetilde X)A(\widetilde X_t,t)\\
&\!\!\!=&\!\!\!U_{s,t}\widetilde X\!\nonumber
\left(\!\!\begin{array}{cc} -\overline \beta & \{[(\nabla_{\widetilde \phi}\overline \kappa_1)\widetilde h,...,(\nabla_{\widetilde \phi}\overline \kappa_{d-r})\widetilde h]
+\partial_t\overline \kappa-\overline \beta\,\overline \kappa\}\! \\0 & 0 \end{array}\!\!\right)\!(\widetilde X_t,t)
\end{eqnarray}
subject to $U_{s,s}\widetilde X=U_{s,s}\widetilde x=I_d$.
Now, suppose that $T_{u,t}$ is the two parameter semigroup:
\begin{equation}
\frac{d}{dt}T_{u,t}=-T_{u,t}\,\overline \beta(\widetilde X_t,t),\ \ \forall t\ge u\ \ \text{subject to }T_{u,u}=I_r.
\end{equation}
Then, the solution of (\ref{MatUdef}) is
\begin{equation}
\!\!\!U_{s,t}\widetilde X=
\left(\!\begin{array}{cc} T_{s,t} & T_{s,t}\overline \kappa(\widetilde X_t,t)-\overline \kappa(\widetilde X_s,s) \\0 & I_{d-r} \end{array}\!\right),
\end{equation}
and so
\begin{equation}
U_{s,t}^{-1}\widetilde X=
\left(\!\begin{array}{cc} T^{-1}_{s,t} & T^{-1}_{s,t}\overline \kappa(\widetilde X_s,s)-\overline \kappa(\widetilde X_t,t) \\0 & I_{d-r} \end{array}\!\right).
\end{equation}
Moreover, it follows by (\ref{ceqn}) that $\overline c$ can also be expressed in
terms of $T_{s,t}^{-1}$.
{\bf Step 6:} Solution Algorithm.\\
\begin{enumerate}
\item[a:]
Check $\overline \kappa$ only depends upon $\widetilde \varphi,t$.
This must be true by Step 2.
\item[b:]
Choose any functions $\overline \beta\in C^1(\widetilde D_T,\mathbb R^{r\times r});\overline \theta\in C^1(\widetilde D_T,\mathbb R^r)$ and $\widetilde h\in C^1(\widetilde D_T,\mathbb R^{p-r})$ for drift of the form
$b(\overline \varphi,\widetilde \varphi,t)=h(\overline \varphi,\widetilde \varphi,t)=
\left(\begin{array}{c}\overline \theta(\widetilde \varphi,t)+\overline \beta(\widetilde \varphi,t)\overline \varphi\\\widetilde h(\widetilde \varphi,t)\end{array}\right)$.
These are the only possible drifts by Step 3.
\item[c:]
Solve
\begin{eqnarray*}
\widetilde X'_t&\!\! =&\!\!\widetilde h(\widetilde X_t,t)\ \ \text{subject to }\widetilde X_s = \widetilde x
\end{eqnarray*}
\item[d:]
Solve
\begin{equation}
\frac{d}{dt}T_{s,t}=-T_{s,t}\, \overline\beta(\widetilde X_t,t),\ \ \forall t\ge s\ \ \text{subject to }T_{s,s}=I_r.
\end{equation}
Then, set
\begin{eqnarray}
\!\!U_{s,t}\widetilde X&\!\!=&
\!\!\left(\!\begin{array}{cc} T_{s,t} & T_{s,t}\overline \kappa(\widetilde X_t,t)-\overline \kappa(\widetilde X_s,s) \\0 & I_{d-r} \end{array}\!\right),\\
\!\!U_{s,t}^{-1}\widetilde X&\!\!=&
\!\!\left(\!\begin{array}{cc} T^{-1}_{s,t} & T^{-1}_{s,t}\overline \kappa(\widetilde X_s,s)-\overline \kappa(\widetilde X_t,t) \\0 & I_{d-r} \end{array}\!\right),\\
\overline c(t) &\!\!=& T^{-1}_{s,t}\overline{x}
+ T^{-1}_{s,t} \int_s^tT_{s,u}\overline \theta(\widetilde X_u,u)du.
\end{eqnarray}
\item[e:]
Divide $\phi=\left(
\begin{array}{c} \overline \phi \\ \widetilde \phi \end{array}
\right)$ and set
$ \widetilde \phi(t) =\widetilde X_t $,
\begin{eqnarray*}
\overline \phi(y,t)&=&\overline c(t) + \left[ I_r \ \ \overline \kappa(\widetilde X_t,t)
\right](U^{-1}_{s,t}\widetilde X)y.
\end{eqnarray*}
\end{enumerate}
\end{ex}
The preceding example was intuitively pleasing:
We showed you could indeed represent \emph{linear} SDEs using a single Gaussian stochastic integral.
Further, we showed that we could \emph{append} an ordinary differential equation
($d\widetilde X_t = \widetilde h(\widetilde X_t)dt$) and use its solution
within the coefficients of the stochastic differential equation.
Finally, we showed how to construct the solution.
While none of this is surprising, it does explain our necessary and sufficient conditions.
In the next section, we will show how to combine this example with diffeomorphisms to handle the general case
with nonlinear coefficients.
\section{Construction Results and Examples} \label{appli}
When one explicit solution exists, there will be a whole class of
such solutions corresponding to distinct $b$'s. We now identify the
$b$'s, $\phi$'s and $U$'s for these solutions corresponding to a
given $\sigma$.
This is done by using local diffeomorphisms to convert the
general case to the case of Example \ref{CoreEG}.
The idea is based upon the following simple lemma.
\begin{lem}\label{dc}
Suppose $D \subset \mathbb R^p$ is a domain, $T>0$, $D_T=D\times[0,T)$,
$\widehat\Lambda\doteq\left(\begin{array}{c}\Lambda_t\\t\end{array}\right):
D_T\rightarrow\widehat\Lambda(D_T)\subset\mathbb R^{p+1}$ is a $C^2$-diffeomorphism
and $\sigma, b, h, \{\phi^{x,s}\}_{(x,s)\in D_T}$,
$\{U_{s,t}\phi^{x,s}\}_{(x,s)\in D_T,s\le t<T}$, $A$ satisfy
Conditions $\mathcal C_1, \mathcal C_2, \mathcal C_3$ as well as equations (\ref{bh},\ref{SecondA}).
Let $\widehat D_T=\widehat\Lambda(D_T)$,
\begin{eqnarray*}
\widehat{\sigma} &=& \{(\nabla_\varphi\Lambda_t)\sigma\}\circ\widehat\Lambda^{-1},\ \ \widehat{h} = \{(\nabla_\varphi\Lambda_t)h\}\circ\widehat\Lambda^{-1},\\
\widehat{b} &=& \left\{(\nabla_\varphi\Lambda_t)b+\frac12\sum_{j=1}^d\sum_{i,k=1}^p
(\partial_{\varphi_i}\partial_{\varphi_k}\Lambda_t)\sigma_{i,j}\sigma_{k,j}\right\}\circ\widehat\Lambda^{-1},\ \
\\
\widehat{\phi}^{x,s}(y,t) &=& \Lambda_t\circ\phi^{\widehat\Lambda^{-1}(x,s)}(y,t),\\
\widehat{U}_{s,t}\widehat{\phi}^{x,s} &=& U_{s,t}\phi^{\widehat\Lambda^{-1}(x,s)},\\
\widehat{A}&=&A\circ\widehat\Lambda^{-1}.
\end{eqnarray*}
Then, $\widehat \sigma, \widehat b, \widehat h, \{\widehat \phi^{x,s}\}_{(x,s)\in \widehat D_T}, \widehat U, \widehat A$ satisfy
Conditions $\mathcal C_1, \mathcal C_2, \mathcal C_3$ as well as equations (\ref{bh},\ref{SecondA}) on $\widehat D_T$.
Moreover,
\begin{itemize}
\item[i)] $\widehat \phi,\widehat U$ is a $(x,s,\widehat \sigma,\widehat h)$-representation for each $(x,s)\in\widehat D_T$ if and only if $\phi,U$ is a $(x,s,\sigma,h)$-representation for each $(x,s)\in D_T$.
\item[ii)] (\ref{braket2}) holds if and only
\begin{equation}\label{braket2hat}
(\nabla_\varphi \widehat \sigma _k)\widehat \sigma _j=(\nabla_\varphi \widehat \sigma _j)\widehat \sigma _k,
\mbox{ on } \widehat D_T\mbox{ for all } j,k\in \{1,...,d\}.
\end{equation}
\item[iii)] (\ref{h}) holds if and only
\begin{equation} \label{hhat}
(\nabla_\varphi \widehat h)\widehat \sigma _j
=(\nabla_\varphi \widehat \sigma _j)\widehat h+\partial_t\widehat \sigma_j-\widehat \sigma\widehat A_j,\mbox{ on } \widehat D_T\mbox{ for all } 1\le j\le d.
\end{equation}
\end{itemize}
\end{lem}
\begin{RM}
In the time-homogeneous case, we can deal with $B$ instead of $A$ and set
$\widehat{B}=B\circ\Lambda^{-1}_0$.
\end{RM}
\begin{proof}
This lemma follows by direct calculation.
Perhaps, the fastest way to verify the commutator conditions is to think of (\ref{Strato}) as a time-homogeneous
equation
\[
d\left[\begin{array}{c}X_t\\t\end{array}\right]=\left[\begin{array}{c}h(X_t,t)\\1\end{array}\right]dt+\left[\begin{array}{c}\sigma (X_t,t)\\0\end{array}\right]\bullet dW_t,\qquad \left[\begin{array}{c}X_s\\s\end{array}\right]=\left[\begin{array}{c}x\\s\end{array}\right]
\]
on $[s,\tau^{x,s})$, by appending the trivial equation $t=t$ and thinking of $t$
as an additional state variable.
Then, verifying (\ref{h}) is equivalent to (\ref{hhat}) is the same as
verfying
\begin{eqnarray*}
&\!\!&\!\!\left(\nabla\left[\begin{array}{c}h\\1\end{array}\right]\right)\left[\begin{array}{c}\sigma_j \\0\end{array}\right]
=\left(\nabla\left[\begin{array}{c}\sigma_j \\0\end{array}\right]\right)\left[\begin{array}{c}h\\1\end{array}\right]-\left[\begin{array}{c}\sigma \\0\end{array}\right]A_j\\
&\!\!\leftrightarrow&\!\!
\left(\nabla\left[\begin{array}{c}\widehat h\\1\end{array}\right]\right)\left[\begin{array}{c}\widehat \sigma_j \\0\end{array}\right]
=\left(\nabla\left[\begin{array}{c}\widehat \sigma_j \\0\end{array}\right]\right)\left[\begin{array}{c}\widehat h\\1\end{array}\right]-\left[\begin{array}{c}\widehat \sigma \\0\end{array}\right]\widehat A_j,
\end{eqnarray*}
which avoids $\partial_t\sigma_j$ and $\Lambda_t$ if we express $(\widehat h^T,1)^T$ and
$(\widehat \sigma_j^T,0)^T$ in terms of $\widehat\Lambda$.
\end{proof}
The idea behind this lemma is that with some diffeomorphism $\widehat{\sigma}=
\left(\begin{array}{cc} I_r & \overline \kappa\\ 0 & 0 \end{array}\right)$
so we can use Example \ref{CoreEG} to solve for the possible $\hat h$
and the representations $\widehat{\phi}^{x,s}$, $\widehat{U}^{x,s}$.
Unfortunately, it is sometimes impossible to have a single diffeomorphism
for all of $D_T$ and, even when it is possible, we may not know that until
after local diffeomorphisms are constructed and one of them is extendable to all
of $D_T$.
\begin{defn}
Suppose $(x,s)\in D_T$. Then, an $({x},s)$-local
diffeomorphism $(O^{x,s},\widehat{\Lambda}^{x,s})$ is a bijection
$\widehat{\Lambda}^{x,s}:O^{x,s}\rightarrow
\widehat{\Lambda}^{x,s}(O^{x,s})$ such that $\widehat{\Lambda}^{x,s}\in
C^2(O^{x,s};\mathbb R^{p+1})$, where $O^{x,s}\subset D_T$ is a
(relatively open) neighbourhood of $x,s$.
We define $\n \widehat{\Lambda}^{-1}(\widehat\Lambda(\varphi,t))$ to be
$\left[\nabla\widehat\Lambda (\varphi,t)\right]^{-1}$ for $(\varphi,t)\in O^{
x,s}$.
\end{defn}
We imposed sufficient differentiability on our local diffeomorphisms for
our uses to follow.
Our $(x,s)$-local diffeomorphisms will take the form
$\widehat\Lambda=\left(\begin{array}{c}\Lambda_t\\t\end{array}\right)$
with $\Lambda_t$ being constructed from $\sigma$ under the conditions:
\begin{itemize}
\item[$D$: ]
Let $D\subset\mathbb R^p$ be a bounded convex domain, $T>0$ and
$D_T=D\times[0,T)$.
\item[$\partial_1$: ]
$\sigma\in C^{ 1}(D_T;{\mathbb R}^{p\times d})$.
\item[$H_r$: ]
The rank of $\sigma$ is $r$ on $D_T$ with the first $r$ rows having full
row rank.
\item[$B$: ]
$(\nabla_\varphi \sigma_j)\sigma_k -(\nabla_\varphi \sigma_k)\sigma_j =0 $ on $D_T$,
for $1\le j,k \le d$ and $(x,s)\in D_T$.
\end{itemize}
To ensure the row rank part of $H_r$, we can just permute the rows of (\ref{VSDE1}),
amounting to relabeling the $\{X^i_t\}_{i=1}^p$.
\begin{prop}\label{diff1}
Suppose [$D$, $\partial_1$, $H_r$, $B$] hold.
Then,
there exists an $(x,s)$-local diffeomorphism $(O^{x,s},\widehat\Lambda^{x,s})$
and a constant permutation matrix $\pi$ such that
$$
\widehat \sigma \doteq \{(\nabla_\varphi \Lambda_t)\sigma \pi\}\circ \widehat\Lambda ^{-1}=
\left(\begin{array}{cc} I_r & \overline \kappa\\ 0 & 0 \end{array}\right) \in \mathbb R^{p\times d} \mbox{ on } \widehat\Lambda(O^{x,s}),
$$
where $\overline \kappa \in C^1(\widehat\Lambda(O^{x,s});{\mathbb R}^{r\times (d-r)})$
does not depend on $\varphi_1, \ldots , \varphi_r$.
\end{prop}
\begin{proof}
Provided in the Appendix.
\end{proof}
\begin{RM}
The permutation matrix $\pi$ permutes the columns of $\sigma$.
We label the permuted diffusion coefficient $\sigma^\pi=\sigma\pi$ and note that
\[
dX_t=b(X_t)dt+\sigma(X_t)dW_t=b(X_t)dt+\sigma^\pi(X_t)dW^\pi_t,
\]
where $W^\pi=\pi^{-1}W$ is a permutation of the Brownian motions $W$.
Also, the Stratonovich drift $h$ remains the same by (\ref{bh}).
\end{RM}
\begin{RM}\label{diffform}
It follows from the proof in the Appendix that the diffeomorphism can have the form
$\widehat\Lambda=\widehat\Lambda_r\circ\cdots\circ\widehat\Lambda_2\circ\widehat\Lambda_1$ for any diffeomorphisms
$\widehat\Lambda_i:\widehat\Lambda_{i-1}\circ\cdots\circ\widehat\Lambda_2\circ\widehat\Lambda_1(D_T)\rightarrow \mathbb R^{p+1}$ satisfying
$\{\nabla\widehat\Lambda_{i}\cdots \nabla\widehat\Lambda_2\nabla \widehat\Lambda_1\sigma^\pi_i\}\circ\widehat\Lambda_1^{-1}\circ\widehat\Lambda_2^{-1}\circ\cdots\circ\widehat\Lambda_i^{-1}=e_i$,
where $(e_1\, e_2\,\ldots\,e_p\,e_{p+1})=I_{p+1}$ is the identity matrix.
However, as will be seen below in Remark \ref{Remark7}, this does not
uniquely define the diffeomorphism.
\end{RM}
Proposition \ref{diff1} immediately provides us our second main theorem.
\begin{thm}\label{r2}
Suppose [$D$, $\partial_1$, $H_r$, $B$] hold, $h\in C^1(D_T;{\mathbb R}^p)$, $(x,s)\in D_T$ and
$W$ is an $\mathbb R^d$-valued standard Brownian motion.
Then, there exists a stopping time $\tau>s$, a permutation matrix $\pi$ and an $(x,s)$-local diffeomorphism $(O^{x,s},\widehat \Lambda^{x,s})$,
as in Proposition \ref{diff1} and Remark \ref{diffform}, such that
$$
\mbox{i) }\ \widehat \sigma \doteq \{(\nabla_\varphi \Lambda_t)\sigma^\pi\}\circ \widehat\Lambda ^{-1}=
\left(\begin{array}{cc} I_r & \overline \kappa\\ 0 & 0 \end{array}\right) \in \mathbb R^{p \times d} \mbox{ on } \widehat\Lambda(O^{x,s}),
$$
with $\overline \kappa \in C^1(\Lambda(O^{x,s});{\mathbb R}^{r\times (d-r)})$
not depending on $\varphi_1, \ldots , \varphi_r$ and ii) the Stratonovich SDE
\(
dX_t =h(X_t)dt +\sigma(X_t)\bullet dW_{t},\ X_s=x
\)
has a solution
$X_t=\Lambda^{-1}_t\left(\begin{array}{c}\overline X_t\\\widetilde X_t\end{array}\right)$ on $[0,\tau]$
if and only if the simpler SDE
\[
\!\!d\left[\begin{array}{c}\overline X_t\\\widetilde X_t\end{array}\right]
= \widehat h\left(\!\begin{array}{c}\overline X_t\\\widetilde X_t\end{array}\!\right)dt
+\left(\begin{array}{cc} I_r & \overline \kappa\\ 0 & 0 \end{array}\right) dW^\pi_t,\ \left[\begin{array}{c}\overline X_s\\\widetilde X_s\end{array}\right]=\Lambda_s(x)
\]
has a solution on $[0,\tau]$, where
$\widehat h=(\nabla_\varphi \Lambda_t h+\partial_t\Lambda_t)\circ\widehat\Lambda^{-1}$.
\end{thm}
We stated the simpler SDE in terms of It\^{o} integration.
However, it follows by (\ref{bh}) and the nature of $\overline\kappa$ that
this equation would have exactly the same form in terms of Stratonovich integration.
In this theorem we do not have a commutator condition for $h$ so we can not
guarantee the simple form of $\widehat h$ as in Example \ref{CoreEG}.
This means that $\widetilde X$ is not in general deterministic nor is $\overline X$
necessarily Gaussian.
We also impose slightly stronger conditions on $\sigma$ compared to
Theorem \ref{r1} but gain
information about the representation as local diffeomorphisms.
\cite{Kouritzin16} solves for a local diffeomorphism $\widehat \Lambda$ of
the form stated in Remark \ref{diffform} corresponding
to the (extended) Heston model, shows that it exists globally,
finds the corresponding $\widehat h$, and solves the SDEs.
The use of the extended model means that our explicit Heston SDE solutions are weak not strong
because the real Heston model corresponds to just part of the extended model
that includes \emph{extra randomness}.
Also, this approach only works when a condition is imposed on the Heston parameters.
When this condition is not true, one can still obtain an explicit weak solution
by using Likelihoods and Girsanov's theorem to convert to the case where the
condition is true.
For our final main result, we add back the commutator
condition for $h$, and characterize all the solutions $X^{x,s}_t=\phi^{x,s}(Y_t,t)$ to
(\ref{SDE4}) via Example \ref{CoreEG}.
We do this through our basic set of parameters for $(x,s)$:
\begin{defn}\label{def3}
Let ${\mathcal P} = {\mathcal P}^{x,s}_{\sigma}$ be the set of all
$(\widehat \Lambda, \overline \kappa, \overline \beta, \overline \theta, \widetilde h,\pi)$ such that
\begin{enumerate}
\item[P0)] $\pi$ is a constant permutation matrix.
\item[P1)] $(O^{x,s},\widehat \Lambda^{x,s})$ is a
$(x,s)$-local diffeomorphism, where $\widehat \Lambda(\varphi,t) =
\mat{\Lambda_t(\varphi)}{t}$.
For convenience, we let $\Lambda_t = \mat{\overline\Lambda_t}{\widetilde\Lambda_t}$ with
$\overline\Lambda_t \in \mathbb R^r$;
\item[P2)] $\overline \kappa \in C^1(\widehat \Lambda(O);{\mathbb R}^{r\times(d- r)}) $ depends only on
$\varphi_{r+1}, \ldots, \varphi_p$, and $t$;
\item [P3)] $\{(\n_\varphi \Lambda_t )\sigma^\pi\}\circ (\widehat\Lambda)^{-1} =
\left(\begin{array}{rl} I_r & \overline \kappa \\
0 & 0\end{array}\right)$ on $\widehat\Lambda(O)$;
\item[P4)] $\overline \beta \in C^1(\widehat \Lambda(O);{\mathbb R}^{r\times r}) $ depends only
on $\varphi_{r+1}, \ldots, \varphi_p$, and $t$;
\item[P5)] $\overline \theta \in
C^1(\widehat \Lambda(O);{\mathbb R}^r)$ depends only on
$\varphi_{r+1}, \ldots, \varphi_p, t$;
\item[P6)]
$\widetilde h\in C^1(\widehat \Lambda(O);{\mathbb R}^{p- r})$ depends only on
$\varphi_{r+1}, \ldots, \varphi_p, t$.
\end{enumerate}
\end{defn}
To each $(\widehat \Lambda, \overline \kappa, \overline \beta, \overline \theta, \widetilde h,\pi) \in {\mathcal P}$,
we
associate the following functions:
\begin{equation}\label{frel}
\begin{cases}
\widetilde X=\widetilde X^{x,s} \in {\mathbb R}^{p-r} \mbox{ uniquely solves } \frac{d}{dt} \widetilde X_t =
\widetilde h (\widetilde X_t,t),\;
\widetilde X_s =\widetilde \Lambda_s(x);\\
G(t) = \left(I_r \Big{|}\; \overline \kappa( \tilde X_t,t)\right)\in {\mathbb R}^{r\times d};\\
\frac{d}{du}T_{s,u}=-T_{s,u}\,\overline \beta(\widetilde X_u,u),\ \ \forall u\ge s\ \ \text{subject to }T_{s,s}=I_r;\\
\!U_{s,u}\widetilde X=
\left(\!\begin{array}{cc} T_{s,u} & T_{s,u}\overline \kappa(\widetilde X_u,u)-\overline \kappa(\widetilde X_s,s) \\0 & I_{d-r} \end{array}\!\right);\\
\!U^{-1}_{s,u}\widetilde X=
\left(\!\begin{array}{cc} T^{-1}_{s,u} & T^{-1}_{s,u}\overline \kappa(\widetilde X_s,s)-\overline \kappa(\widetilde X_u,u) \\0 & I_{d-r} \end{array}\!\right);\\
\overline c_s(t) = T^{-1}_{s,t}\overline{\Lambda}_s(x)
+ T^{-1}_{s,t} \int_s^tT_{s,u}\overline \theta(\widetilde X_u,u)du.
\end{cases}
\end{equation}
\medskip
The following theorem follows from Theorem \ref{r2}, Theorem \ref{r1} (so the explicit solution
implies $B$ above) and Example \ref{CoreEG}.
In particular, we must have
\begin{equation}\label{possibleh2}
(\nabla_\varphi \Lambda_t h+\partial_t\Lambda_t)\circ\widehat\Lambda^{-1}
=\left(\begin{array}{c}\overline h(\overline \varphi,\widetilde \varphi,t)\\\widetilde h(\widetilde \varphi,t)\end{array}\right)
=
\left(\begin{array}{c}
\overline \theta(\widetilde \varphi,t)+
\overline \beta(\widetilde \varphi,t)\overline \varphi\\\widetilde h(\widetilde \varphi,t)
\end{array}\right),
\end{equation}
which gives our possible drifts $h$ in the following theorem.
\medskip
\begin{thm}\label{r3}
Suppose [$D$, $\partial_1$, $H_r$] hold, $(x,s)\in D_T$
and $X_t^{x,s} = \phi^{x,s}\left(\int_s^t U_{s,u}\phi^{x,s}dW^\pi_u,t
\right)$, with $\phi,U$ satisfying $\mathcal C_1$, $\mathcal C_2$, $\mathcal C_3$, solves (\ref{SDE4}) up to
some stopping time
$\tau^{x,s}>s$.
Then, there exists $((O^{x,s},\widehat \Lambda^{x,s}), \overline \kappa, \overline \beta, \overline \theta, \widetilde h,\pi)
\in {\mathcal
P}^{x,s}_{\sigma}$, and related functions $\widetilde X, G, U, \overline c$ defined
by (\ref{frel}), such that
\begin{equation} \label{hdef}
h= [\n_\varphi \Lambda_t]^{-1}\left\{
\left[\begin{array}{c}\overline\theta (\widetilde X_t,t)\\
\widetilde h (\widetilde X_t,t)\end{array}\right] -\partial_t \Lambda_t
+ \mat{\overline\beta(\widetilde X_t,t)\ \overline \Lambda_t}{0} \right\} \mbox{ on } O^{x},
\end{equation}
\begin{equation} \label{fdef}
\phi^{x,s}(y,t) = \phi_{(\widehat\Lambda, \overline \kappa, \overline\beta,\overline \theta, \widetilde h)}(y,t) =
\Lambda_t^{-1}\left( \mat{\overline c_s(t) +G(t)(U^{-1}_{s,t}\widetilde X)y}{\widetilde X_t}\right)
\end{equation}
on ${\mathcal N}^{x} = \left\{(y,t):
\mat{\overline c_s(t) +G(t)U^{-1}_{s,t}\widetilde Xy}{\widetilde X_t} \in \Lambda_t(O^{x,s})\right\}$.
Finally, if $\breve\pi$, $\breve\Lambda$ and $\breve\kappa$ also satisfies P0--P3,
then there exist $\breve\beta, \breve\theta,\breve h$ such that
$ (\breve\Lambda,\breve \kappa,\breve\beta, \breve\theta,\breve h, \breve\pi) \in {\mathcal P}$,
$b_{(\breve\Lambda, \breve\kappa,\breve\beta,\breve \theta,\breve h, \breve\pi)}
= b_{(\widehat\Lambda, \overline\kappa, \overline\beta, \overline\theta,\widetilde h)}$, and
$\phi_{(\breve\Lambda, \breve\kappa,\breve\beta, \breve\theta,\breve h, \breve\pi)}
= \phi_{(\widehat\Lambda, \overline\kappa, \overline\beta, \overline\theta,\widetilde h)}$.
\end{thm}
\begin{RM}\label{DiffFind}
For the sake of brevity in the examples below, we will just give local diffeomorphisms
satisfying P3) above.
However, as is shown in our companion paper \citet{Kouritzin16}, it is
often possible to solve for them using the technique used in the proof
of Proposition \ref{diff1} herein.
\end{RM}
\begin{RM}\label{Remark7}
To illustrate the need of the final statement of Theorem \ref{r3}, we take for example,
$\sigma (x) = x \in {\mathbb R}^p$. Then, any $L \in C^1({\mathbb R}^p)$ depending on $x_2/x_1,
\ldots, x_p/x_1$ satisfies
$(\nabla L)\sigma =0$. Therefore, $\widehat\Lambda $ and hence the parameter set
is not unique but we can create the same $b,\phi$ from any consistent
$\overline \kappa ,\widehat\Lambda .$
\end{RM}
\subsection{One Dimensional Case}
Suppose $d=p=r=1$, $D\subset\mathbb R$ and $x\in D$.
Then, $\overline\kappa,\widetilde h$ do not exist and
$\overline\beta,\overline\theta$ only depend on $t$.
Moreover, $U_{s,t}=T_{s,t}=e^{-\int_s^t \overline\beta(u)du}$,
$\overline c_s(t)=T_{s,t}^{-1}\overline{\Lambda}_s(x)+T_{s,t}^{-1}\int_s^t T_{s,u}\overline\theta(u)du$
and the diffeomorphism can be taken as
$\Lambda_t(\varphi) = \int\frac{1}{\sigma(\varphi ,t)}d\varphi $.
One then finds by (\ref{bh},\ref{frel},\ref{hdef},\ref{fdef}) that
the corresponding diffusion drift $b$ and explicit solutions are
\begin{equation}\label{scalardrift}
b(\varphi,t) = \sigma(\varphi,t)\left\{\overline\theta(t) +\overline\beta(t)\Lambda_t(\varphi) - \partial_t \Lambda_t\right\}+\frac{1}{2}\sigma(\varphi,t)\partial_\varphi \sigma(\varphi,t)
\end{equation}
\begin{equation}\label{scalarsoln}
X_t = \Lambda_t^{-1}\left[\left\{\Lambda_s(x)+ \int_s^t
T_{s,u} \overline\theta(u)du + \int_s^t T_{s,u} dW_u\right\}\Big{/}T_{s,t}\right].
\end{equation}
\begin{ex}[Time-varying Cox-Ingersoll-Ross model]
Suppose $\overline\theta, \overline\beta$ and continuously differentiable
$s(t)>0$ are chosen and
$\sigma(\varphi,t) = s(t) \sqrt{\varphi}$.
Then,
\(
\Lambda_t(\varphi) = \frac{2\sqrt{\varphi}}{s(t)}
\),
\(
\Lambda_t^{-1}(z)=\left(\frac{zs(t)}2\right)^2
\)
and the possible It\^{o} drifts are
\[
b(\varphi,t)=\overline\theta(t)s(t)\sqrt{\varphi}+2\left(\overline\beta(t)+\frac{\dot{s}(t)}{s(t)}\right)\varphi+\frac{s^2(t)}4.
\]
The explicit solutions are then
\begin{eqnarray}\label{CIRsoln}
\!\!X_t^{x,s} &\!\!= &\!\!\bigg|\frac{s(t)}{s(s)}e^{\int_s^t \overline\beta(v)dv}\sqrt{x}\\
\nonumber
&\!\!+&\!\!\frac{s(t)}2\left\{ \int_s^t
e^{\int_u^t \overline\beta(v)dv} \overline\theta(u)du + \int_s^t e^{\int_u^t \overline\beta(v)dv} dW_u\right\}\,\bigg|^2.
\end{eqnarray}
In the case $s(t)=\sigma,\overline \theta$ and $\overline\beta$ are taken constant,
we get
$$
\!X_t^{x,s} = \frac{1}{4}\left\{
2e^{\overline \beta (t-s)}\sqrt{x}
+ \frac{\overline \theta \sigma}{\overline \beta} (e^{\overline \beta (t-s)}-1) +
\sigma\! \int_s^t\! e^{\overline \beta(t-u)}dW_u
\right\}^2
$$
solves
$$
dX_t^{x,s} = \left(\sigma^2/4 +2\overline \beta X_t^{x,s} +\sigma \overline\theta \sqrt{X_t^{x,s}}
\right)dt+ \sigma \sqrt{X_t^{x,s}} dW_t, \ X_s=x
$$
as long as $X_t^{x,s}>0$.
This solves the usual CIR model
\begin{equation}\label{eq:feller}
dX_t = \alpha \left(\beta -X_t \right)dt+ \sigma \sqrt{X_t} dW_t.
\end{equation}
when $\overline\theta =0$, $\alpha = 2\overline \beta$, $\beta = \sigma^2/(8\overline \beta)$.
Now, set $Y_t = \sqrt{X_t}$, where $X$ solves \eqref{eq:feller}
with $\sigma^2 = 4\alpha\beta$, and $\tau =
\inf\{t >0; X_t = 0\}$.
It is well known that $P(\tau<\infty)=1$.
Then,
\begin{eqnarray}
dY_t &=& \frac{1}{8Y_t}\left(4\alpha\beta - \sigma^2\right)dt -
\frac{\alpha}{2} Y_t dt+
\frac{\sigma}{2} dW_t \nonumber\\
&=& - \frac{\alpha}{2} Y_t dt+ \frac{\sigma}{2}
dW_t,\label{eq:OU}
\end{eqnarray}
by It\^o's formula.
However, since \eqref{eq:OU}
defines a Gaussian process and $Y$ must be non-negative, one cannot have $Y_t$ defined by
\eqref{eq:OU} unless $t < \tau$.
This explains why we first look for explicit \emph{local} solutions.
\end{ex}
\subsection{Square Non-Singular Case}
\label{square}
Suppose that $d=p=r$, $\sigma = \sigma(\varphi,t)$ is a $d\times d$ non-singular
continuously-differentiable matrix satisfying (\ref{braket2}), $D\subset\mathbb R^p$
and $x\in D$.
Again, we apply Theorem \ref{r3} and find $\overline\kappa,\widetilde h$ do not exist while
$\overline\beta,\overline\theta$ only depend on $t$.
Also, there is a local diffeomorphism
$\widehat \Lambda=\left( \begin{array}{c}\Lambda_t\\t\end{array}\right)$ such that $ \n_\varphi \Lambda_t(\varphi) = [\sigma(\varphi,t)]^{-1}$, and
all explicit solutions are of the form $\phi^{x,s} (t,y) =
\Lambda_t^{-1}\left(\overline c_s(t) + U^{-1}_{s,t}y\right)$, where
\[
U_{s,t} =
-\int_s^t U_{s,u} \overline\beta(u)du +I \text{ and } \overline c_s(t)
=U^{-1}_{s,t}\left\{\Lambda_s(x)+ \int_s^t U_{s,u}\overline\theta(u)du \right\}
\]
for some $\overline\theta \in C([0,T);\mathbb R^d)$ and $\overline\beta\in
C^1([0,T),\mathbb R^{d\times d})$.
The resulting drift is
$$
b(\varphi,t) = \sigma(\varphi,t)\left\{\overline\theta(t) + \overline\beta(t)\Lambda_t(\varphi) - \partial_t \Lambda_t(\varphi)\right\}
+\frac{1}{2}\sum_{j=1}^d (\n_\varphi \sigma_j(\varphi,t))\sigma_j(\varphi,t).
$$
\begin{ex}\emph{Geometric Brownian motions:}
Take $\sigma_{ij}(\varphi) = \varphi_i
\gamma_{ij}$ with $\gamma$ non-singular and $D = (0,\infty)^d$.
Then, $\sigma$ satisfies the
commutation condition (\ref{braket2}) since $ [ (\n_\varphi \sigma_j)\sigma_k]_i = \varphi_i
\gamma_{ij}\gamma_{ik}$, and the diffeomorphism can be chosen as
$\Lambda(\varphi) = \Lambda_t(\varphi) =\gamma^{-1} \mat{\log{\varphi_1} \\ \vdots}{\log{\varphi_d}}$.
$\Lambda$'s image is $\mathbb R^d$, so
$\Lambda^{-1}(z) = \mat{e^{(\gamma z)_1}\\ \vdots}{e^{(\gamma z)_d}}$ is defined everywhere
and $\phi_i^{x,s}(y,t) = \exp\left[ \gamma \{\overline c_s(t) + U^{-1}_{s,t}y\}\right]_i$.
The possible drifts satisfy
$$
b_i(\varphi,t) = \varphi_i \left\{\alpha_i(t) -\sum_{j=1}^d B_{ij}(t)
\log{\varphi_j}\right\},
$$
for $1\le i \le d$,
where $B(t) = \gamma\overline\beta(t) \gamma^{-1}$,
and $\alpha_i(t)=\frac12[\gamma
\gamma^\top]_{ii}+ [\gamma\overline\theta(t)]_i$.
\end{ex}
\begin{ex}
\emph{Diffeomorphism example:}
In the previous examples, we started with $\sigma$.
Suppose instead we had a diffeomorphism
$$
\Lambda (\varphi_1,\varphi_2) = \Lambda_t(\varphi_1,\varphi_2) = \mat{\frac{\pi}{2} + \arcsin( \log{\varphi_1
\varphi_2}-1)}{\frac{\pi}{2} + \arcsin( \frac{2\varphi_2}{\varphi_1}-1)}
$$
on $ 1 < \varphi_1 \varphi_2 < e$, $ \varphi_2 \le \varphi_1$.
Then, the possible full rank $\sigma$'s satisfy $\sigma=(\nabla_\varphi \Lambda)^{-1}$ i.e.\
\begin{equation}
\sigma(\varphi_1,\varphi_2) = \left(\!\begin{array}{cc}
\frac{\varphi_1}{2}\sqrt{2\log{\varphi_1 \varphi_2}-(\log{\varphi_1 \varphi_2})^2}
& -\frac{\varphi_1}{2\varphi_2}\sqrt{\varphi_2(\varphi_1-\varphi_2)}\\
& \\
\frac{\varphi_2}{2}\sqrt{2\log{\varphi_1 \varphi_2}-(\log{\varphi_1 \varphi_2})^2} &
-\frac{1}{2}\sqrt{\varphi_2(\varphi_1-\varphi_2)}
\end{array}\!\right)
\end{equation}
so $(\n \Lambda)\sigma = I_2$ and $\sigma$ satisfies (\ref{braket2}) by Lemma \ref{dc} ii).
The possible Stratonovich (time-dependent) drifts $h(\varphi_1, \varphi_2,t)$ are
\begin{equation}
\sigma(\varphi_1, \varphi_2)\!\left(\!\!\begin{array}{c}
\overline\theta_1(t)+\overline\beta_{11}(t)(\frac{\pi}{2} \! +\! \arcsin( \log{\varphi_1
\varphi_2}-\!1) )
-\!\overline\beta_{12}(t)(\frac{\pi}{2} \! + \!\arcsin( \frac{2\varphi_2}{\varphi_1}-\!1)\!)\\
\overline\theta_2(t)+\overline\beta_{21}(t)(\frac{\pi}{2} \! +\! \arcsin( \log{\varphi_1
\varphi_2}-\!1) )
-\!\overline\beta_{22}(t)(\frac{\pi}{2} \! +\! \arcsin( \frac{2\varphi_2}{\varphi_1}-\!1)\!)
\end{array}\!\!\!\right)
\end{equation}
while $U_{s,t},\overline c_s$ satisfy the equations at the start of Subsection
\ref{square}.
\end{ex}
\subsection{Non-Square Case}
Our most important example is probably the \emph{Extended Heston}
model of our companion paper \cite{Kouritzin16}.
It is non-square.
However, we provide a second interesting non-square example herein.
\begin{ex}[Heisenberg group]
Let $\overline x\in\mathbb R^d$ and $\widetilde x\in\mathbb R$ be the components
of the starting point, $A=A(t)$ be a $\mathbb R^{d\times d}$ continuously differentiable
matrix function and
$\sigma(\varphi,t) = \sigma(\xi,z,t) =
\mat{I_d}{(A(t)\xi)^\top}$, where $\xi\in \mathbb R^d$, $z\in \mathbb R$.
Then, $\sigma$ has rank $r=d$.
The solution to $dX_t = \sigma(X_t,t)dW_t$ is known as the Brownian motion
on the Heisenberg group.
Moreover,
$$
(\n_\varphi\sigma_j)\sigma_k- (\n_\varphi\sigma_k)\sigma_j =
\mat{0}{A_{jk}-A_{kj}}.
$$
Therefore, (\ref{braket2}) holds true if and only if $A$
is symmetric.
In this case, one can solve for an explicit solution for an
arbitrary starting point $(\overline x,\widetilde x,s)$.
The diffeomorphism $\widehat
\Lambda(\xi,z,t) = \mat{\Lambda_t(\xi,z)}{t}$ is solved
$\Lambda_t(\xi,z) = \mat{\xi}{g}$ with $g(\xi,z,t) = z-\frac{1}{2}\xi^\top A(t)\xi$
following the proof of Proposition \ref{diff1} in the Appendix (see
\cite{Kouritzin16} for details on a more involved example).
Hence, $\pi=I_d$, $\widehat\sigma= \left[ \begin{array}{cc}I_d \\0\end{array}\right]$,
$\overline\kappa$ does not exist so $G(t)=I_d$ and
\(
[\nabla\Lambda_t]^{-1}=
\left[
\begin{array}{cc}
I_d &0\\ \xi^\top A(t)&1
\end{array}
\right].
\)
Now, we can take any functions $\overline\theta\in\mathbb R^d$, $\overline\beta\in\mathbb R^{d\times d}$, $\widetilde h\in\mathbb R$
satisfying the differentiability conditions in Definition \ref{def3}
and let $\widetilde X_t,\ U_{s,t}\widetilde X,\ \overline c_s(t)$ satisfy:
\begin{eqnarray*}
\frac{d}{dt} \widetilde X_t &\!\!= &\!\!\widetilde h(\widetilde X_t,t)\ \text{s.t.\ }\widetilde X_s=\widetilde x-\frac{1}{2}\overline x^\top A(s)\overline x
\\
\frac{d}{du}U_{s,u}\widetilde X&\!\!=&\!\!-(U_{s,u}\widetilde X)\,\overline \beta(\widetilde X_u,u)\ \text{s.t.\ }U_{s,s}\widetilde X=I_d
\\
\overline c_s(t) &\!\!=&\!\! U^{-1}_{s,t}\left\{\overline x +\disp \int_0^t
U_{s,u}\overline \theta(\widetilde X_u,u)du \right\}.
\end{eqnarray*}
From Theorem \ref{r3} and (\ref{bh}), drift $b$ must be of the (quadratic) form
$$
b(\xi,z,t)\! =\! \mat{\overline\theta(\widetilde X_t,t)-\overline\beta(\widetilde X_t,t)\xi}
{\widetilde h(\widetilde X_t,t) +
\xi^\top A(t)\overline\theta(\widetilde X_t,t)-\xi^\top A(t) \overline\beta(\widetilde X_t,t)\xi +\frac12 \xi^\top \frac{d}{dt}A(t)\xi +\frac{1}{2}{\rm Tr}\{A(t)\} }
$$
for some $\overline\theta$, $\overline\beta$, $\widetilde h$.
Finally, the corresponding $\phi$ is given by
$$
\phi(y,t) = \mat{\overline c_s(t)+(U^{-1}_{s,t}\widetilde X)y}{\widetilde X_t
+\frac{1}{2}(\overline c_s(t)+(U^{-1}_{s,t}\widetilde X)y)^\top
A(t)(\overline c_s(t)+(U^{-1}_{s,t}\widetilde X)y)}.
$$
\end{ex}
\section{Comparison with the works of Yamato and Kunita} \label{kun}
Now, we compare our existence results to those appearing in
\citet{Yamato:1979} and \citet{Kunita:1984}. In Section III.3 of
Kunita's treatise, he considers representations of time-homogeneous
Fisk-Stratonovich equations
\begin{equation} \label{N2.6a}
dX^{x}_t = h(X^{x}_t)dt+\sigma(X^{x}_t)\bullet dW_t
\end{equation}
in terms of the flows generated by the vector
fields
\begin{equation}\label{VectField}
{\mathfrak X}_0(y)= \disp \sum_{i=1}^ph_i(y)\frac \partial
{\partial y_i} \text{ and }{\mathfrak X}_k(y)= \disp \sum_{i=1}^p \sigma
_{ik}(y)\frac \partial {\partial y_i}, k=1,...,d,
\end{equation}
under
conditions imposed on the Lie algebra $L_0({\mathfrak X}_0,
{\mathfrak X}_1, \ldots, {\mathfrak X}_d)$ generated by
${\mathfrak X}_k$, $0\le k\le d$. In the special case where these
vector fields commute, i.e. the Lie bracket $[{\mathfrak
X}_k,{\mathfrak X}_j]=0 $ for each $j,k=0,...,d$, and the
coefficients $h_i$, $\sigma _{ik}$ are respectively in
$C^3_\alpha$, $C^4_\alpha$ (the locally four times
continuously differentiable functions whose fourth derivative is $\alpha
-H\"older continuous), his work gives rise to the composition formula
\begin{eqnarray}
\left( X_t^{x}\right) _i &=& Exp\left( t{\mathfrak X}_0\right) \circ Exp\left(
W_t^1{\mathfrak X}_1\right) \circ \cdots \circ Exp\left( W_t^d{\mathfrak X}_d\right)
\circ \chi_i(x), \label{Kunita} \\
\ &=&\phi _i(W_t,t) \nonumber
\end{eqnarray}
locally. Here, $\chi _i$ is the function taking $x$ to its $i^{th}$
component and $Exp\left( u{\mathfrak X}_k\right) $ is the one parameter group of
transformations generated by vector field ${\mathfrak X}_k$, i.e. the unique solution to
\begin{equation}
\frac d{du}(f\circ \varphi _u)={\mathfrak X}_kf(\varphi _u),\;\varphi _0=x\qquad \forall
f\in C^\infty . \label{Kde}
\end{equation}
In fact, to use (\ref{Kunita}), one must solve (\ref{Kde}) for
$k=0,...,d$ and $f=\chi _i$, $i=1,...,d$. Kunita also goes beyond
commutability, even surpassing \citet{Yamato:1979} in generality by
considering the situation where $L_0({\mathfrak X}_0,...,{\mathfrak
X}_d)$ is only solvable, but the expression replacing (\ref{Kunita})
necessarily becomes more unwieldy.
Our characterization of $\phi $ provided by Theorem \ref{r3} provides an
alternative to (\ref{Kunita}) that is more amenable to direct
calculation.
Corollary \ref{corr3} (to follow) supplies a converse to
(\ref{Kunita}) in the sense that if $X_t^{x,s}$ were to have such a
functional representation $\phi^{x,s}(W_t,t)$ in terms of Brownian
motions only, then the vector fields must commute. This was
previously established in Theorem 4.1 of \citet{Yamato:1979} under
$C^\infty$ conditions on both $\phi$ and the coefficients.
The other advantages of our representations over Kunita's results
are:
\begin{itemize}
\item We allow time dependent vector fields.
\item We decrease the regularity assumptions by imposing weaker
differentiability
on $h$ and on $\sigma$ when $r$ is small. The looser
regularity on the
coefficients requires eschewing Fisk-Stratonovich equations in
favour of It\^o processes.
\item We remove the nilpotency assumptions (for our representations).
\end{itemize}
To validate the final claim, we take $p=2$, $d=1$,
\[
{\mathfrak
X}_0 = \{\overline \theta(x_2)-B(x_2)x_1\} \dx{1} + \widetilde \theta(x_2)\dx{2},
\]
and ${\mathfrak X}_1 = \dx{1}$. Then $[{\mathfrak X}_0,{\mathfrak
X}_1] = B\dx{1}$. Moreover, if $ {\mathfrak X}_k = [{\mathfrak
X}_0,{\mathfrak X}_{k-1}]$, $k\ge 2$, then $ {\mathfrak X}_k =
a_k(x_2)\dx{1}$, where $a_{k+1} = \widetilde \theta (\dx{2} a_k)+ Ba_k$,
$k\ge 1$ and $a_1 =1$.
In general, the $a_k$'s will not vanish
and thereby the Lie algebra contains an infinite number of
linearly independent vector fields. This algebra is solvable but
is not nilpotent.
\medskip
Using Theorem \ref{r1}, we can also give the converse to Kunita's result,
Example III.3.5 in \citet{Kunita:1984}, that is valid under the mild
regularity on $b,\sigma,h$ given at the beginning of the section.
\bigskip
\begin{cor}\label{corr3}
Suppose that there exists a domain $\widetilde{D}$ such that the
coefficients $\sigma$ and $h$ are time-homogeneous and
Fisk-Stratonovich acceptable on
$\tilde D_T = \tilde D\times (0,T)$.
Further, assume that the solution to the Fisk-Stratonovich equation
(\ref{N2.6a}) has a unique local solution
$$
\left( X_t^{x}\right) _i=Exp\left( t{\mathfrak X}_0\right) \circ Exp\left( W_t^1{\mathfrak X}_1\right)
\circ \cdots \circ Exp\left( W_t^d{\mathfrak X}_d\right) \circ \chi _i(x)
$$
on $0\leq t<\tau _x$ for some positive stopping time $\tau _x$ and
each $x\in \widetilde{D}$, where ${\mathfrak X}_k$, $k=0,1,\ldots,
d$ are the vector fields defined in (\ref{VectField}).
Then,
$$
[{\mathfrak X}_k,{\mathfrak X}_j]=0 \mbox{ on }\widetilde{D}\mbox{ for each }j,k=0,\ldots,d.
$$
\end{cor}
\begin{proof}
We find that $X^{x}_t=\phi(Y_t,t)$ with $U_{s,t}=I$ so it follows
from Theorem \ref{r1} that $\sigma A=0$.
The condition $[{\mathcal
X}_k,{\mathcal X}_j]=0$ then follows from (\ref{braket2},\ref{h}).
\end{proof}
\section{Proofs of the main results}\label{proofs}
We note that $b,\sigma$ are Lipschitz on
any compact, convex subset of $D_T$ by our $C^1$-conditions and use
the proof of \citet{Kunita:1984}[Theorem II.5.2] for
uniqueness of (strong) local solutions to the SDE until they leave such a compact subset.
\subsection{Proof of Theorem \ref{r1} a) is equivalent to b).}\label{Prop1proof}
\begin{proof}
Using (\ref{Ydefn}) and It\^o's formula for $X_t=\phi(Y_t,t)$, one finds that for any
$1\le i\le p$,
\begin{eqnarray}
\!\!&\!\!\!&\!\!\!d(X_t)_i =\sum_{m=1}^d\sum_{j=1}^d\partial
_{y_m}\phi_i(Y_t,t)(U_{s,t}\phi)_{mj}dW_t^j\label{ctscoef}\\
&+&\!\!\!\!\left[ \partial _t\phi_i(Y_t,t)+ \nonumber
\frac{1}{2}\sum_{j=1}^d\sum_{k=1}^d\partial _{y_j}\partial
_{y_k}\phi_i(Y_t,t)(U_{s,t}\phi\ (U_{s,t}\phi)^{\top })_{jk}\right]\! dt.
\end{eqnarray}
Now, starting with b) implies a) and using (\ref{grad},\ref{dtphi}) on (\ref{ctscoef}),
we find
\begin{eqnarray}
\!\!d(X_t)_i \!\!\!&=&\!\!\!\sigma_i(\phi(Y_t,t),t) dW_t+h_i(\phi(Y_t,t),t)dt\label{ctscoef1}\\
&+&\!\! \nonumber
\frac{1}{2}\sum_{j=1}^d\sum_{k=1}^d\partial _{y_j}\partial
_{y_k}\phi_i(Y_t,t)(U_{s,t}\phi\ (U_{s,t}\phi)^{\top } )_{jk} dt.
\end{eqnarray}
Moreover,
$$
\dy{m} \{\sigma_{ij}(\phi,t)\} = \sum_{n=1}^p\{\partial _{\varphi_n}\sigma
_{ij}\}(\phi,t
)\dy{m}\phi_n
$$
and if (\ref{grad}) is true, one obtains
$$
\dy{m} \{\sigma_{ij}(\phi,t)\} =\sum_{l=1}^d\dy{m}\dy{l}\phi_i\;(U_{s,t}\phi )_{lj}.
$$
Abbreviating notation $U_{m k}(\phi,t)=(U_{s,t}\phi) _{mk}$,
multiplying the last two equalities by $U_{m k}$, summing over $m$ and using (\ref{grad}) again,
one finds that
\begin{equation}\label{iden}
\sum_{n=1}^p\{\partial _{\varphi_n}\sigma _{ij}\}(\phi,t )\sigma
_{nk}(\phi,t )=\sum_{m=1}^d\sum_{l=1}^d \partial _{y_m}\partial
_{y_l}\phi_i\;U_{lj}(\phi,t)U_{mk}(\phi,t),
\end{equation}
and, taking $k=j$ and summing over $j$, one has that
\begin{equation}\label{corrreduce}
\!\sum_{j=1}^d \{\nabla_\varphi \sigma_j\}(\phi,t)\sigma_j (\phi,t) =\!
\sum_{l=1}^d\sum_{m=1}^d \;(U(\phi,t)U^{\top }(\phi,t))_{lm} \dy{m}\dy{l} \phi .
\end{equation}
Therefore, if (\ref{grad},\ref{dtphi},\ref{initphi}) are satisfied, then clearly $X_t$ is a local strong solution to
(\ref{SDE4}) by (\ref{bh}).
Moreover, letting $t\searrow s$, we find by (\ref{grad},\ref{dtphi},\ref{initphi}) that
\[
\sigma(x,s)=\nabla_y \phi^{x,s}(0,s)\ \text{ and } h(x,s)=\partial_t \phi^{x,s}(0,s)
\]
so $\sigma,h\in C^1$ by the last part of Definition \ref{RepPair}.
To show a) implies b), we suppose $X_t$ is a strong solution to
(\ref{SDE4}) on $(s,\tau^{x,s})$.
Then, since continuous finite-variation martingales are
constant, the (continuous) It\^o process
$\phi(Y_t,t)$ from (\ref{ctscoef}) matches (\ref{SDE4}) if
and only if
\begin{equation} \label{sigma0}
\sigma _{ij}(\phi,t )=\sum_{m=1}^d\partial _{y_m}\phi _i\;(U_{s,t}\phi)_{mj}\;\forall 1\le i\leq
p,\;1\le j\le d,
\end{equation}
and
\begin{equation} \label{b0}
b_i(\phi,t )=\partial _t\phi_i+{\frac
12}\sum_{j=1}^d\sum_{k=1}^d\partial _{y_j}\partial
_{y_k}\phi_i\;(U_{s,t}\phi (U_{s,t}\phi)^{\top })_{jk}\;\forall 1\le i\le p
\end{equation}
for all $t\in(s,\tau^{x,s})$.
Rewriting (\ref{sigma0}) in matrix form, one finds
\begin{equation}
\sigma (\phi(Y_t,t),t )=\{\nabla _y\phi(Y_t,t) \}U_{s,t}\phi \label{crosscross},
\end{equation}
and (\ref{grad}) is true.
Now, we can use (\ref{corrreduce}) (which was just shown to be a consequence of (\ref{grad}))
to find (\ref{b0}) is equivalent to
\begin{equation} \label{N4.7a}
\partial _t\phi =b(\phi,t )-{\frac 12}\sum_{k=1}^d\{\nabla_x \sigma _k\}(\phi,t
)\;\sigma _k(\phi,t )=h(\phi,t ),
\end{equation}
using (\ref{bh}).
Now, (\ref{dtphi}) follows by continuity and (\ref{Rdefn}).
Letting $t\searrow s$ in (\ref{crosscross}) and (\ref{N4.7a}), one finds
\[
\sigma(x,s)=\nabla_y \phi^{x,s}(0,s)\ \text{ and } h(x,s)=\partial_t \phi^{x,s}(0,s)
\]
so the last part of Definition \ref{RepPair}
follows from the $C^1$ property of $h,\sigma$.
\end{proof}
\subsection{Proof of Theorem \ref{r1} b) is equivalent to c).}\label{Th1proof}
{\bf Idea:}
Below we show that the existence of a representation without the commutator conditions leads
to a contradiction and the commutator conditions yield a representation.
\begin{proof}
By exactness of differential 1-forms, the existence of our
function $\phi^{x,s}$ satisfying ((\ref{grad}), (\ref{dtphi}) and
(\ref{initphi})) is equivalent to the following two conditions:
\begin{equation}\label{exact1}
\partial_{y_j}\{\sigma(\phi,t)(U^{-1}_{s,t}\phi)_k \}=
\partial_{y_k}\{\sigma(\phi,t)(U_{s,t}^{-1}\phi)_j\}
\end{equation}
and
\begin{equation}\label{exact2}
\frac{d}{dt}\{\sigma(\phi,t)(U_{s,t}^{-1}\phi)_k \}=
\partial_{y_k}h(\phi,t).
\end{equation}
We show (\ref{exact1}) and (\ref{exact2}) for all starting points $(x,s)$ are equivalent to (\ref{braket2}) and (\ref{h}) respectively.\\
{\bf Step 1:} Show that (\ref{braket2}) implies (\ref{exact1}) (under (\ref{grad})).\\
It follows by (\ref{grad}) and $\mathcal C_2$, $\mathcal C_3$ that
\begin{eqnarray}\label{N4.7b}
\partial_{y_j}\!\!\!\!\!\!\!&\!\!\!\!\!\!\!\!\!\!\!\!&\!\!\!\!\!\!\!\!\{\sigma(\phi,t)(U_{s,t}^{-1}\phi)_k \}\\
&=&
\sum_m \{\partial_{y_j} \sigma_m(\phi,t)\}(U_{s,t}^{-1}\phi)_{mk}\ \ \nonumber\\
&=&\sum_m \nabla_\phi \sigma_m(\phi,t)\sigma(\phi,t)(U_{s,t}^{-1}\phi)_j (U_{s,t}^{-1}\phi)_{mk}\ \ \nonumber\\
&=&\sum_m\sum_n \nabla_\phi \sigma_m(\phi,t)\sigma_n(\phi,t)(U_{s,t}^{-1}\phi)_{nj} (U_{s,t}^{-1}\phi)_{mk}\ \ \nonumber
\end{eqnarray}
and similarly
\begin{eqnarray}\label{Nov25}
\partial_{y_k}\!\!\!\!\!\!\!&\!\!\!\!\!\!\!\!\!\!\!\!&\!\!\!\!\!\!\!\!\{\sigma(\phi,t)(U_{s,t}^{-1}\phi)_j \}\\
&=&\sum_n \sum_m\nabla_\phi \sigma_n(\phi,t)\sigma_m(\phi,t)(U_{s,t}^{-1}\phi)_{mk} (U_{s,t}^{-1}\phi)_{nj}\nonumber
\end{eqnarray}
Hence, (\ref{exact1}) holds when (\ref{braket2}) holds.\\
{\bf Step 2:} Show that (\ref{exact1}) implies (\ref{braket2}) (under (\ref{grad})).\\
Letting $t\searrow s$ in (\ref{N4.7b}) and (\ref{Nov25}), one finds by (\ref{exact1})
that for all $1\le j,k \le d$,
\begin{eqnarray}
&&\sum_m\sum_n \nabla_x \sigma_m(x,s)\sigma_n(x,s)(U_{s,s}^{-1}\phi)_{nj} (U_{s,s}^{-1}\phi)_{mk}\\
&=&\lim_{t\searrow s}\partial_{y_j}\{\sigma(\phi,t)(U_{s,t}^{-1}\phi)_k \}\ \ \nonumber\\
&=&\lim_{t\searrow s}\partial_{y_k}\{\sigma(\phi,t)(U_{s,t}^{-1}\phi)_j \}\ \ \nonumber\\
&=&\sum_m\sum_n\nabla_x \sigma_n(x,s) \sigma_m(x,s)(U_{s,s}^{-1}\phi)_{nj} (U_{s,s}^{-1}\phi)_{mk}.\nonumber
\end{eqnarray}
However, $U_{s,s}^{-1}\phi=I$
so we have that
$$
(\nabla_x \sigma_q)(x,s)\sigma_p (x,s)= (\nabla_x
\sigma_p)(x,s)\sigma_q(x,s).
$$
Hence, (\ref{braket2})
holds when (\ref{exact1}) does.\\
{\bf Step 3:} Show that (\ref{exact2}) implies (\ref{h}) (under (\ref{grad},\ref{dtphi})).\\
One gets by (\ref{grad}) that
\begin{equation}
\partial_{y_k}h(\phi,t) = \nabla_\phi h(\phi,t)\partial_{y_k}\phi(y,t)=\nabla_\phi h(\phi,t)\sigma(\phi,t)(U_{s,t}^{-1}\phi)_k
\end{equation}
and by (\ref{exact2}), (\ref{dtphi}) that
\begin{eqnarray}
\partial_{y_k}h(\phi,t)\!\! &=&\!\! \frac{d}{dt}\{\sigma(\phi,t)(U_{s,t}^{-1}\phi)_k \}\\
&=& \!\!\sum_m\nabla_\phi \sigma_m(\phi,t)h(\phi,t)(U_{s,t}^{-1}\phi)_{mk}+\partial_{t}\sigma(\phi,t)(U_{s,t}^{-1}\phi)_k\nonumber\\
&-&\!\!\sigma(\phi,t)U_{s,t}^{-1}\phi\sum_m\frac{d}{dt}(U_{s,t}\phi)_m(U_{s,t}^{-1}\phi)_{mk}.\nonumber
\end{eqnarray}
Combining these equations, multiplying by $(U_{s,t}\phi)_{kn}$ and summing, we get
\begin{eqnarray}
\nabla_\phi h(\phi,t)\sigma_n(\phi,t)&=&\nabla_\phi
\sigma_n(\phi,t)h(\phi,t)+\partial_{t}\sigma_n(\phi,t)\\
&-&\sigma(\phi,t)U^{-1}_{s,t}\phi^{x,s}\frac{d}{dt}(U_{s,t}\phi^{x,s})_n\nonumber
\end{eqnarray}
so, letting $t\searrow s$ and using (\ref{SecondA},\ref{initphi}), one arrives at (\ref{h}).\\
{\bf Step 4:} Show that (\ref{h}) implies (\ref{exact2}) (under (\ref{grad},\ref{dtphi})).\\
Using (\ref{dtphi}) and (\ref{Ugroup}), we get that
\begin{eqnarray*}
&&\frac{d}{dt}\{\sigma(\phi,t)(U^{-1}_{s,t}\phi)_k\}\\
&\!\!=&\!\! \sum_n\Big[\{\nabla_\phi \sigma_n(\phi,t)\}h(\phi,t)
+\{\partial_t\sigma_n(\phi,t)\}\Big](U^{-1}_{s,t}\phi)_{nk}\\
&\!\!&\!\!\quad -\sum_n\sigma(\phi,t)U^{-1}_{s,t}\phi\left(\frac{d}{dt}(U_{s,t}\phi)_n\right)(U^{-1}_{s,t}\phi)_{nk}\\
&\!\!=&\!\! \sum_n\Big[\{\nabla_\phi \sigma_n(\phi,t)\}h(\phi,t)
+\partial_t\sigma_n(\phi,t)\\
&\!\!&\!\!-\sigma(\phi,t)\left(\frac{d}{dt}U_{u,t}\phi^{\phi_u,u}\right)_n\big|_{u=t}\Big]
(U^{-1}_{s,t}\phi)_{nk},
\end{eqnarray*}
where $\phi_u$ is short for $\phi^{x,s}(y_u,u)$.
Hence, by (\ref{SecondA}), (\ref{h}) applied at $\varphi=\phi$ and (\ref{grad})
\begin{eqnarray*}
\frac{d}{dt}\{\sigma(\phi,t)(U^{-1}_{s,t}\phi)_k\}
&\!\!=&\!\!\{\nabla_\phi h(\phi,t)\} \sigma(\phi,t) (U^{-1}_{s,t}\phi)_k \\
&\!\!=&\!\!\partial_{y_k}h(\phi,t)
\end{eqnarray*}
and we have (\ref{exact2}).
\end{proof}
\subsection{Proof of Proposition \ref{diff1}}
Our methods
are motivated in part by \citet{Brickell/Clark:1970}[Propositions 8.3.2 and 11.5.2].\\
We let
\[
(q,D^2_T) = \left\{\begin{array}{ll}
(p+1,D\times (-T,T)) & \mbox{if $\sigma$ or $h$ depend on $t$}\\
(p,D) & \mbox{otherwise}
\end{array} \right.,
\]
take $\sigma_{p+1}=0$ if $q>p$,
set $\partial_t\sigma(x,t)=\partial_t\sigma(x,0)$,
$\partial_{x_i}\sigma(x,t)=\partial_{x_i}\sigma(x,0)$ for $t<0$, $i=1,2,...,q$
and use exactness of the corresponding $1$-form
to extend $\sigma$ uniquely to $D^2_T$ such that
$\sigma\in C^1(D^2_T;\mathbb R^{q\times d})$.
By reducing $T>0$ if necessary, we can find a permuation $\pi$ such that the first
$r$ columns of $\sigma^\pi=\sigma\pi$ are linearly independent on $D_T^2$.
\begin{proof}
Fix $\widehat x=(\widehat x_1,...,\widehat x_q)\in D_T$.
The $C^1$-diffeomorphism $\Lambda$ will have form:
\begin{eqnarray}
\!\!\Lambda&\!\!=&\!\!\Lambda^{r,1},\ \ \text{where }\ \Lambda^{i,1}=\Lambda^{i}\circ\Lambda^{i-1}\circ\cdots\circ\Lambda^{2}\circ\Lambda^{1},\\\label{DiffForm}
\Lambda^{i}&\!\!=&\!\!\sum_{j=1}^{i-1}x_je_j+\left[\begin{array}{c}H^i(x_i,...,x_q)\\L^i(x_i,...,x_q)\end{array} \right]\ \ \text{and }\ H^i(x_i,...,x_q)\in\mathbb R^{i-1},\
\end{eqnarray}
and $\widehat \sigma_i$ will be defined as
$\widehat \sigma_i\doteq\{(\nabla_\varphi \Lambda)\sigma^\pi_i\}\circ \Lambda^{-1}$.
Here,
$\Lambda^i$ is a $C^1$-diffeomorphism
on a neighborhood $O^{\widehat x^{i-1}}$ of $\widehat x^{i-1}=
\Lambda^{i-1,1}(\widehat x)$
so $\Lambda:O^{\widehat x}\rightarrow \mathbb R^q$.
To construct $\Lambda^i$ recursively starting with $\Lambda^1$, we
suppose $\widehat \sigma_j=e_j$ for $j<i$ and
\begin{equation}\label{alphadef}
\alpha_i\doteq\{\nabla\Lambda^{i-1,1}\sigma^\pi_i\}\circ(\Lambda^{i-1,1})^{-1}
\end{equation}
does not depend upon $x_1,...,x_{i-1}$, which are certainly true when $i=1$.
Moreover, without loss of generatlity, we assume the $i^{th}$ component of $\alpha_i$ satisfies
$\alpha_{i,i}\ne 0$ (or else we change $\pi$ by permuting columns $i,...,d$
of $\sigma^\pi$).
Set $\psi^i(x) = \theta(x_i -\widehat x_{i}^{i-1}; x_1,...,x_{i-1},\widehat x_i^{i-1},x_{i+1},...,x_q)$,
where $\theta$ satisfies
$\frac{d}{dt}\theta(t;x)=\alpha_i(\theta(t;x))$, $\theta(0;x)=x$
for $t\in I^x$, an open interval containing $0$, and $x$
in a neighborhood containing $\widehat x^{i-1}$.
Then, $\dx{i} \psi^i = \alpha_i(\psi^i)$.
For $j\ne i$, we have $\partial_{x_j}\psi^i(x)=\partial_{x_j}\theta(x_i -\widehat x_{i}^{i-1};x_1,...,x_{i-1},\widehat x_i^{i-1},x_{i+1},...,x_q)$ and
$$
\partial_t\partial_{x_j}\theta(t;x)=\partial_{x_j}\alpha_i(\theta(t;x))\ \text{ s.t. }\ \partial_{x_j}\theta(0,x)=e_j
$$
so $\nabla \psi^i(\widehat x^{i-1})$ has determinant
$\alpha_{i,i}(\widehat x^{i-1}) \neq 0$.
Thus, $\psi^i$ has inverse $\Lambda^i\in C^2(O^{\widehat x^{i-1}},\mathbb R^{q})$
and $\n \Lambda^i = [\n \psi^i]^{-1}(\Lambda^i)$ on neighborhood
$O^{\widehat x^{i-1}}$
of $\widehat x^{i-1}$ by the Inverse Function Theorem.
Hence, $\nabla \Lambda^i((\Lambda^i)^{-1}) \nabla \psi^i = I$ and
\begin{equation}\label{OneCol}
\widehat \sigma_i=\{ \nabla \Lambda^i\alpha_i \}(\Lambda^i)^{-1} =e_i\in\mathbb R^{q}.
\end{equation}
Moreover, $\Lambda^i$ has the form (\ref{DiffForm}) if $\psi^i$ has similar form.
$\psi^i$ has this form by its definition
as well as the facts $\alpha_i$ is locally Lipschitz and does not depend upon $x_1,...,x_{i-1}$.
Next,
\begin{equation}\label{TransCom}
(\nabla \widehat \sigma_j) \widehat \sigma_k - (\nabla \widehat \sigma_k)\widehat \sigma_j = (\nabla \sigma^\pi_j)
\sigma^\pi_k -(\nabla \sigma^\pi_k)\sigma^\pi_j=0\ \ \forall\ \ 1\le k , j\le d
\end{equation}
by Lemma \ref{dc}.
Now, since $\widehat \sigma_k=e_k\in\mathbb R^q$ for $1\le k\le i$, (\ref{TransCom}) implies
$$
(\nabla \widehat \sigma_j) e_k = (\nabla \widehat \sigma_j) e_k - (\nabla e_k)\widehat \sigma_j = 0\ \ \forall\ 1\le k \leq i<j
$$
on a neighborhood $O$ of $\widehat x$.
Therefore, $\widehat \sigma_j$ and (by a similar argument) $\alpha_{i+1}$ can not
depend upon $x_1, \ldots, x_i$ so we can take $i=r$ by induction and
$$
\widehat \sigma =
\{(\nabla \Lambda) \sigma^\pi\}\circ \Lambda^{-1} =
\left(\begin{array}{cc} I_r & \overline \kappa\\ 0 & \widetilde \kappa \end{array}\right)
\in \mathbb R^{q\times d} \mbox{ on } \Lambda(O\cap D_T),
$$
where $\overline \kappa \in {\mathbb R}^{r\times (d-r)}$ and
$\widetilde \kappa \in {\mathbb R}^{(q-r)\times (d-r)}$ do
not depend on the variables
$x_1, \ldots, x_r$.
Since $\widehat \sigma$ has also rank $r$, it follows that $\widetilde \kappa =0$.
\end{proof}
\bibliographystyle{apalike}
|
\section{\@startsection{section}{1}%
\z@{1.75\linespacing\@plus\linespacing}{1\linespacing}%
{\normalfont\bf\centering}}
\makeatother
\mathcode`l="8000
\begingroup
\makeatletter
\lccode`\~=`\l
\DeclareMathSymbol{\lsb@l}{\mathalpha}{letters}{`l}
\lowercase{\gdef~{\ifnum\the\mathgroup=\m@ne \ell \else \lsb@l \fi}}%
\endgroup
\newcommand{\varphi}\newcommand{\eps}{\varepsilon}{\varphi}\newcommand{\eps}{\varepsilon}
\newcommand{\smath}[1]{\hbox{$#1$}}
\newcommand{\sfrac}[2]{\smath{\frac{#1}{#2}}}
\newcommand{\tilde\ph_\eps}{\tilde\varphi}\newcommand{\eps}{\varepsilon_\eps}
\newcommand{\tilde L}{\tilde L}
\newcommand{\setminus}{\setminus}
\newcommand{{\mathbb N}}{{\mathbb N}}
\newcommand{\skipit}[2][{\par\noindent[...]\medskip\par}]{\medskip #1}
\RequirePackage{color}\definecolor{RED}{rgb}{1,0,0}\definecolor{BLUE}{rgb}{0,0,1}
\newcommand{\K}[1][...]{{\color{blue}\textbf{[[}\emph{#1}\textbf{]]}}}
\def\langle{\langle}
\def\rangle{\rangle}
\def{\rm End}{{\rm End}}
\def\text{deg}\,{\text{deg}\,}
\def\wedge{\wedge}
\def\bar\partial{\bar\partial}
\def{\mathbb R}{{\mathbb R}}
\def{\mathbb C}{{\mathbb C}}
\def{\mathbb P}{{\mathbb P}}
\def\C^n{{\mathbb C}^n}
\def\leftarrow{\leftarrow}
\def{\rm Hom\, }{{\rm Hom\, }}
\def{\mathcal O}{{\mathcal O}}
\def{\mathcal Q}{{\mathcal Q}}
\def{\rm Re\, }{{\rm Re\, }}
\DeclareMathOperator{\Id}{Id}
\DeclareMathOperator{\supp}{supp}
\DeclareMathOperator{\ann}{ann}
\DeclareMathOperator{\rank}{rank}
\DeclareMathOperator{\tr}{tr}
\DeclareMathOperator{\codim}{codim}
\def\begin{equation}{\begin{equation}}
\def\end{equation}{\end{equation}}
\newtheorem{thm}{Theorem}[section]
\newtheorem{lma}[thm]{Lemma}
\newtheorem{cor}[thm]{Corollary}
\newtheorem{prop}[thm]{Proposition}
\theoremstyle{definition}
\newtheorem{df}[thm]{Definition}
\newtheorem{preremark}[thm]{Remark}
\newtheorem{preex}[thm]{Example}
\newenvironment{remark}{\begin{preremark}}{\end{preremark}}
\newenvironment{ex}{\begin{preex}}{\end{preex}}
\numberwithin{equation}{section}
\hypersetup{
pdftoolbar=true,
pdfmenubar=true,
pdftitle={Chern forms of singular metrics on vector bundles},
pdfauthor={},
pdfsubject={Preprint},
pdfkeywords={},
pdfborder= 0 0 .1,
bookmarksnumbered=false,
}
\begin{document}
\title[]{Chern forms of singular metrics on vector bundles}
\date{\today}
\author{Richard L\"ark\"ang, Hossein Raufi, Jean Ruppenthal, Martin Sera}
\address{R. L\"ark\"ang, H. Raufi, M. Sera, Department of Mathematics\\Chalmers University of Technology and the University of Gothenburg\\412 96 G\"oteborg, Sweden.}
\email{<EMAIL>, <EMAIL>, <EMAIL>}
\address{J. Ruppenthal, Department of Mathematics, University of Wuppertal, Gau{\ss}str. 20, 42119 Wuppertal, Germany.}
\email{<EMAIL>}
\thanks{The first author was supported by a grant from the Swedish Research Council, the second author by a grant from the
Olle Engkvist foundation, and the last three authors by the Deutsche Forschungsgemeinschaft
(DFG, German Research Foundation), grant RU 1474/2 within DFG's Emmy Noether Programme.}
\begin{abstract}
We study singular hermitian metrics on holomorphic vector bundles, following Berndtsson-P{\u{a}}un. Previous work by Raufi has shown that for such metrics, it is in general not possible to define the curvature as a current with measure coefficients. In this paper we show that despite this, under appropriate codimension restrictions on the singular set of the metric, it is still possible to define Chern forms as closed currents of order 0 with locally finite mass, which represent the Chern classes of the vector bundle.
\end{abstract}
\maketitle
\section{Introduction}
\noindent Let $X$ be a complex manifold of dimension $n$, let $E\to X$ be a rank $r$ holomorphic vector bundle over $X$, and let $h$ denote a hermitian metric on $E$. The classical differential geometric study of $X$ through $(E,h)$, revolves heavily around the notion of the curvature associated with $h$. This approach requires the metric to be smooth (i.e.\ twice differentiable). However, for line bundles Demailly in \cite{Dem4} introduced the notion of \emph{singular} hermitian metrics, and in a series of influential papers he and others showed how these are a fundamental tool for giving complex algebraic geometry an analytic interpretation.
In \cite{BP} Berndtsson and P{\u{a}}un introduced the following notion of singular metrics for vector bundles:
\begin{df}\label{df:sing}
Let $E\to X$ be a holomorphic vector bundle over a complex manifold $X$. A \emph{singular hermitian metric} $h$ on $E$ is a measurable map from the base space $X$ to the space of hermitian forms on the fibers. The hermitian forms are allowed to take the value $\infty$ at some points in the base (i.e.\ the norm function $\|\xi\|_h$ is a measurable function with values in $[0,\infty]$), but for any fiber $E_x$ the subset $E_0:=\{\xi\in E_x\ ;\ \|\xi\|_{h(x)}<\infty\}$ has to be a linear subspace, and the restriction of the metric to this subspace must be an ordinary hermitian form.
\end{df}
They also defined what it means for these types of metrics to be curved in the sense of Griffiths:
\begin{df}\label{df:Gr}
Let $E\to X$ be a holomorphic vector bundle over a complex manifold $X$ and let $h$ be a singular hermitian metric. We say that $h$ is \emph{Griffiths negative} if $\|u\|^2_h$ is plurisubharmonic for any (local) holomorphic section $u$. Furthermore, we say that $h$ is \emph{Griffiths positive} if the dual metric $h^*$ is Griffiths negative.
\end{df}
\begin{remark}
(i) Strictly speaking, \cite{BP} define $h$ to be Griffiths negative if $\log\|u\|_h$ is plurisubharmonic for any holomorphic section $u$. It is, however, not too difficult to show that these two definitions are equivalent (see e.g. \cite{R}, section 2).\vspace{0.1cm}\\
(ii) Any singular hermitian metric on a vector bundle $E$ induces a dual metric on the dual bundle $E^\ast$ (see Lemma \ref{lma:SingularDual} below). This justifies the notion of Griffiths positivity in Definition \ref{df:Gr} in terms of duality.
\qed
\end{remark}
\noindent Definition~\ref{df:Gr} is very natural as these conditions are well-known equivalent properties for smooth metrics.
Although Definition~\ref{df:sing} is very liberal, as it basically puts no restriction on the metrics, it turns out that Definition~\ref{df:Gr} rules out most of the pathological behaviour. For example, we have the following proposition (\cite{R}, Proposition 1.3 (ii)):
\begin{prop}\label{prop:psh}
Let $h$ be a singular, Griffiths negative, hermitian metric. If $\det h\not\equiv0$, then $i\partial\bar\partial\log\det h$ is a closed, positive $(1,1)$-current.
\end{prop}
The proof uses the well-known fact that if $h$ is a metric on $E$, then $\det h$ is a metric on $\det E$. For smooth metrics it is also well-known that the curvature of $\det h$, i.e.\ $-\partial\bar\partial\log\det h = \partial \bar\partial \log \det h^*$, is the trace of the curvature of $h$, i.e.\ $2\pi i$ times the first Chern form $c_1(E,h)$. Thus, a simple consequence of Proposition~\ref{prop:psh} is that for a singular metric which is curved in the sense of Definition~\ref{df:Gr}, it is possible to define the first Chern form in a meaningful way as a closed, positive or negative $(1,1)$-current.
However, despite this, one of the main results in \cite{R} (Theorem 1.5) is a counter-example that shows that the curvature requirement of Definition~\ref{df:Gr} is not enough to define the curvature of a singular metric as a current with measure coefficients. This rather surprising fact, given the existence of the first Chern form, leads to the question of which differential geometric concepts one can obtain from Definition~\ref{df:Gr}.
The main purpose of this note is to show that under appropriate codimension restrictions on the singular set, it is possible to define Chern forms, $c_k(E,h)$, $k=1,\ldots,\min(r,n)$, associated with a singular metric which is curved as in Definition~\ref{df:Gr}. For metrics that are also continuous outside of the singular set, we can prove the following theorem:
\begin{thm}\label{thm:ContThm}
Let $E\to X$ be a rank $r$ holomorphic vector bundle over a complex manifold $X$, and let $h$ denote a singular, Griffiths positive or negative hermitian metric on $E$. Assume that there is some subvariety $V$ of $X$ with $codim(V)\geq k$ such that $h$ is continuous and non-degenerate outside of $V$.
Then, there exists a unique, closed $(k,k)$-current, $c_k(E,h)$, of order 0 with locally finite mass in $X$ such that for any local regularizing sequence $\{h_\varepsilon\}$ of $h$, with $h_\varepsilon\to h$ locally uniformly outside of $V$, we have that
\begin{equation}\label{eq:weak_conv}
c_k(E,h_{\varepsilon})\to c_k(E,h)
\end{equation}
in the sense of currents.
More generally, let $E_1,\dots,E_m$ be holomorphic vector bundles on $X$, and for $i=1,\dots,m$, let $h^i$ be a singular Griffiths positive or negative hermitian metric on $E_i$, $k_i \in \mathbb{N}\cup\{0\}$, and let $k := k_1 + \dots + k_m$.
Assume that there exists a subvariety $V$ of $X$ such that $\codim(V) \geq k$, and such that $h^1,\dots,h^m$ are
continuous and non-degenerate outside of $V$.
Then, there exists a unique, closed $(k,k)$-current,
\begin{equation*}
c_{k_1}(E_1,h^1) \wedge \dots \wedge c_{k_m}(E_m,h^m),
\end{equation*}
of order $0$ with locally finite mass in $X$ such that for any local regularizing sequences $\{h^i_\varepsilon\}$, with $h^i_\varepsilon \to h^i$
locally uniformly outside of $V$ for $i=1,\dots,m$, we have that
\begin{equation*}
c_{k_1}(E_1,h^1_{\varepsilon}) \wedge \dots \wedge c_{k_m}(E_m,h^m_{\varepsilon}) \to
c_{k_1}(E_1,h^1) \wedge \dots \wedge c_{k_m}(E_m,h^m)
\end{equation*}
in the sense of currents.
\end{thm}
\begin{remark}
(i) Let $L$ be a trivial line bundle with a possibly singular positive metric $e^{-\varphi}$, i.e.\ $\varphi$ is a plurisubharmonic function. Let $E$ be the trivial rank $r$ bundle $E = L \oplus \cdots \oplus L$. Equip $E$ with the metric $h$ induced by $e^{-\varphi}$, which is then a Griffiths positive singular hermitian metric. If $\varphi$ is smooth, then $c_k(E,h) = C_k (dd^c \varphi)^k$ for some constant $C_k$. For an arbitrary plurisubharmonic function $\varphi$, one does not have a natural meaning of the product $(dd^c \varphi)^k$, without any condition like for example that the set where $\varphi$ is unbounded is contained in an analytic subset of codimension $k$. Thus, since $h$ degenerates precisely where $\varphi$ is $-\infty$, this shows that the codimension requirements on the degeneracy set of $h$ in Theorem \ref{thm:ContThm} can not be relaxed in general.\vspace{0.1cm}\\
(ii) A singular metric $h$ which is Griffiths positive can always be locally approximated by an increasing sequence $\{h_\varepsilon\}$ of (smooth) Griffiths positive metrics. This is the content of \cite{BP}, Proposition 3.1 and \cite{R}, Proposition 6.2 (see Proposition \ref{prop:appr} below), and the regularizing sequence is obtained through convolution with an approximate identity. Thus, if the singular metric is continuous outside its degeneracy set, these propositions yield that we can always obtain a regularizing sequence which converges locally uniformly on this set.\vspace{0.1cm}\\
(iii) To be precise, from the proof of Theorem~\ref{thm:ContThm} it follows that $c_k(E,h)$ can be written as the difference of two positive, closed $(k,k)$-currents. It is natural to ask whether it is actually positive. However, even when $h$ is a smooth Griffiths positive metric on $E$, it is to the best of our knowledge an open question whether $c_k(E,h)$ is a positive form in general, and only known when $k=1$, which follows from the fact that $c_1(E,h)$ equals the first Chern class of the positive line bundle $(\det E,\det h)$, and when $k=2$ and $\dim X = 2$, which is due to Griffiths, \cite{GrPos} (Appendix to \S 5 (b)). If it is indeed the case that $c_k(E,h)$ is positive for all smooth Griffiths positive metrics on $E$, then it would follow by weak convergence that if $h$ is a singular Griffiths positive metric on $E$, then $c_k(E,h)$ is a positive current.
\end{remark}
\begin{ex}
There are indeed non-trivial singular metrics on vector bundles satisfying the conditions
of Theorem~\ref{thm:ContThm}. One such class can for example be found in \cite{Hos}.
Namely, let $E$ be a vector bundle on a complex manifold $X$. If $E$ has global sections $s_1,\dots,s_m$,
then these sections induce a morphism $s : E^* \to X \times {\mathbb C}^m$, where $X \times {\mathbb C}^m$ is a trivial rank $m$ bundle,
which we equip with a trivial metric. Through the morphism $s$, we can define a singular hermitian metric $h^*$ on $E^*$ by
\begin{equation*}
\langle \xi,\eta \rangle_{h^*} := \langle s(\xi), s(\eta) \rangle.
\end{equation*}
This is a singular Griffiths negative metric on $E^*$ since if $\xi$ is a holomorphic section of $E^*$, then $s(\xi)$ is a holomorphic section of the trivial rank $m$ bundle, and hence $\|\xi\|^2_{h^*} = \|s(\xi)\|^2$ is plurisubharmonic.
We thus obtain a singular Griffiths positive metric $h = h^{**}$ on $E$ (cf. Lemma~\ref{lma:SingularDual}).
At the points where $s_1,\dots,s_m$ span $E$, the metric $h$ is smooth and non-degenerate,
and thus, if $s_1,\dots,s_m$ span $E$ on $X \setminus V$ for some subvariety $V$ of $X$ with $\codim(V) \geq k$,
then the conditions of Theorem~\ref{thm:ContThm} are satisfied to define the Chern current $c_k(E,h)$.
\end{ex}
Using the Chern currents of Theorem \ref{thm:ContThm} it is straightforward to define Chern characters associated with our singular metrics.
For a smooth metric $h$ on a holomorphic vector bundle $E$, the full Chern character can be defined as $ch(E,h) = \tr \exp( (i/2\pi)\Theta(E,h))$.
The Chern character has the advantage that the formula for the Chern character of a tensor product is very simple,
$ch(E\otimes F, h\otimes g) = ch(E,h) \wedge ch(F,g)$. In general, the full Chern character can be expressed as a polynomial in the Chern forms
with integer coefficients.
Since for singular metrics we need to restrict the degree of the products of the Chern forms, we will consider the part $ch_k(E,h)$
of the Chern character of a fixed degree $(k,k)$, which can thus be expressed as:
\begin{equation} \label{eq:ch_k}
ch_k(E,h) := \tr \left( \frac{(i \Theta)^k}{(2\pi)^k k!} \right) = \sum_K b_K c_{k_1}(E,h) \wedge \dots \wedge c_{k_m}(E,h),
\end{equation}
where $b_K$ are integers, and $K = (k_1,\dots,k_m)$ runs over all partitions of $k$, i.e.\ $k_1,\dots,k_m$ are positive integers and $k_1 + \cdots + k_m = k$.
\begin{df}\label{df:ChernCharacter}
Let $h$ be a singular, Griffiths positive or negative, hermitian metric on a holomorphic vector bundle $E\to X$. Assume that
there exists a subvariety $V$ of $X$ with $\codim(V) \geq k$ such that $h$ is continuous and non-degenerate outside of $V$.
Then the $k$:th \emph{Chern character} $ch_k(E,h)$ is defined by \eqref{eq:ch_k}, where the products of Chern forms in
the sum in the right-hand side of \eqref{eq:ch_k} have meaning by Theorem~\ref{thm:ContThm}.
\end{df}
The requirement in Theorem~\ref{thm:ContThm} that $h$ should be continuous outside of its degenerate set can be dropped for the existence and uniqueness (see Theorem \ref{thm:main} below), but then we will no longer know that $c_k(E,h_{\varepsilon})$ converges to $c_k(E,h)$. This property is important since an immediate corollary is that basic (local) properties for smooth Chern forms also hold in the singular setting:
\begin{cor}\label{cor:Properties}
Let $X$ be a complex manifold and let $(E,h)\to X$ and $(F,g)\to X$ be two holomorphic vector bundles, where $h$ and $g$ both are singular, Griffiths positive or negative hermitian metrics.
If there exists a subvariety $V$ of $X$ with $codim(V)\geq k$ such that $h$ is continuous and non-degenerate outside of $V$, then:\vspace{0.1cm}\\
(a) The Chern current $c_k(E^*,h^*)$ (where $h^*$ denotes the dual metric on the dual bundle $E^*$) can be defined as in Theorem \ref{thm:ContThm}, and
$$c_k(E^*,h^*)=(-1)^kc_k(E,h).$$
(b) For a complex manifold $Y$ and any holomorphic submersion $f:Y\to X$, the Chern current $c_k(f^*E,f^*h)$ can be defined as in Theorem \ref{thm:ContThm} and
$$c_k(f^*E,f^*h)=f^*c_k(E,h).$$
(c) If there exists a subvariety $V$ of $X$ with $codim(V)\geq k$ such that both $h$ and $g$ are continuous and non-degenerate outside of $V$, then $c_k(E\oplus F,h\oplus g)$ can be defined as in Theorem \ref{thm:ContThm}, and
$$c_k(E\oplus F,h\oplus g)=\sum_{j=0}^k c_j(E,h) \wedge c_{k-j}(F,g),$$
where the products in the sum are defined by Theorem~\ref{thm:ContThm}. \\
(d) If there exists a subvariety $V$ of $X$ with $codim(V)\geq k$ such that both $h$ and $g$ are continuous and non-degenerate outside of $V$, then
$ch_k(E\otimes F,h\otimes g)$ can be defined as in Definition~\ref{df:ChernCharacter}, and
$$ch_k(E\otimes F,h\otimes g)=\sum_{j=0}^k ch_j(E,h) \wedge ch_{k-j}(F,g),$$
where the products in the sum are defined by \eqref{eq:ch_k} and Theorem~\ref{thm:ContThm}
\end{cor}
\bigskip
\begin{remark}
By arguing as in \cite{PT}, Lemma 2.3.2 (a), we get that for any complex manifold $Y$, and any holomorphic mapping $f:Y\to X$, the pullback $f^* h$ of a singular, Griffiths positive (negative), hermitian metric $h$, is also Griffiths positive (negative). In Corollary \ref{cor:Properties} (b), we nevertheless need to restrict ourselves to submersions, since otherwise $f^*c_k(E,h)$, the pullback of a \emph{current}, is not well-defined. Also $f$ being a submersion ensures that the codimension requirements needed to define $c_k(f^*E,f^*h)$ are fulfilled.
\end{remark}
As already mentioned, the continuity requirements of Theorem \ref{thm:ContThm} can be relaxed, at the price of losing the weak convergence (\ref{eq:weak_conv}). Thus, for a singular hermitian metric $h$ (under appropriate codimension requirements on the degeneracy set), it is possible to define Chern currents by only requiring the metric to be curved as in Definition~\ref{df:Gr}.
This is achieved by expressing the Chern currents in terms of so called Segre currents. In the smooth setting these are defined as the inverses of the Chern forms (see section 2 below) and can be expressed recursively through,
\begin{equation}\label{eq:recursive}
s_k(E,h)+s_{k-1}(E,h) \wedge c_1(E,h)+\ldots+c_k(E,h)=0,\quad k=1,\ldots,n.
\end{equation}
Hence, any Chern form can be expressed as a sum of products of Segre forms
\begin{equation} \label{eq:ChernSegre}
c_k(E,h) = \sum_K a_K s_{k_1}(E,h) \wedge \dots \wedge s_{k_m}(E,h)
\end{equation}
where $a_K$ are integers, and where $K = (k_1,\dots,k_m)$ runs over all partitions of $k$, i.e.\ $k_1,\dots,k_m$ are positive integers and $k_1 + \cdots + k_m = k$.
Now as already mentioned, a singular Griffiths positive metric $h$ can always be locally approximated by an increasing sequence $\{h_\varepsilon\}$ of Griffiths positive metrics (see Proposition \ref{prop:appr}). The advantage of Segre forms over Chern forms in the singular setting is that we can show that for products of Segre forms with different regularizations, under appropriate codimension restrictions on the singular set of $h$, there exists subsequences such that the iterated limit
\begin{equation}\label{eq:IteratedLimit}
\lim_{\varepsilon^m_\nu \to 0} \cdots \lim_{\varepsilon^1_\nu \to 0} s_{k_1}(E,h_{\varepsilon^1_\nu}) \wedge \cdots \wedge s_{k_m}(E,h_{\varepsilon^m_\nu})
\end{equation}
exists in the sense of currents. This limit is furthermore independent of the regularizations, and will thus yield a global current which we will denote by $s_{k_1}(E,h)\wedge\cdots\wedge s_{k_m}(E,h)$. These currents can then be used to define the Chern current $c_k(E,h)$ through (\ref{eq:ChernSegre}). The precise statement of the conditions under which this construction works is as follows.
\begin{thm}\label{thm:main}
Let $E\to X$ be a holomorphic vector bundle over a complex manifold $X$, and let $h$ denote a singular, Griffiths positive, hermitian metric on $E$.
Assume that there is some subvariety $V$ of $X$ with $codim(V)\geq k$ such that $L(\log\det h^*)\subseteq V$, where $L(\log\det h^*)$ denotes the unbounded locus of $\log\det h^*$ (see Definition~\ref{df:ULocus} and Remark~\ref{remark:UEx} below).
Then the $k$:th Chern current of $E$ associated with $h$, $c_k(E,h)$, can be defined through (\ref{eq:ChernSegre}), where the Segre products $s_{k_1}(E,h)\wedge\cdots\wedge s_{k_m}(E,h)$ are defined by the iterated limit (\ref{eq:IteratedLimit}). The Chern current $c_k(E,h)$ will be a closed $(k,k)$-current of order 0 with locally finite mass in $X$.
If $h$ is instead a singular, Griffiths negative, hermitian metric, then the same result holds if $L(\log \det h^*)$ is replaced by $L(\log \det h)$ throughout the statement.
\end{thm}
Although Theorem~\ref{thm:ContThm} is not formulated in terms of Segre forms, its proof also uses the formula \eqref{eq:ChernSegre} for expressing Chern forms of a smooth regularization of the metric in terms of Segre forms. The key difference between the proof of Theorem~\ref{thm:ContThm} and Theorem~\ref{thm:main} is that when $h$ is continuous outside of the non-degeneracy set, then one can use the same regularization in each of the Segre forms and can just take a single limit, while in the setting of Theorem~\ref{thm:main}, one needs to have different regularizations of the metric in each of the Segre forms, and take an iterated limit as in \eqref{eq:IteratedLimit} to see that the limits exist.
\begin{remark}
If the conditions of Theorem~\ref{thm:ContThm} are fulfilled to define the current $c_k(E,h)$, then the conditions of Theorem~\ref{thm:main} are also fulfilled, and one would a priori obtain two different currents $c_k(E,h)$. However, it follows from the respective proofs that these currents indeed coincide.
\end{remark}
Finally, in the smooth setting the most salient feature of Chern forms is of course that they are closed forms whose cohomology classes (in the de Rham cohomology group of smooth forms) are invariants of the vector bundle, i.e.\ independent of the metric. As these forms also define cohomology classes in the de Rham cohomology group of currents, it is natural to ask whether or not the cohomology classes of the Chern currents of Theorem \ref{thm:main} coincide with the usual Chern classes.
Since this cohomology class invariance is a global property, the singular counterpart is not a direct consequence of the weak convergence in either Theorem~\ref{thm:ContThm} or Theorem \ref{thm:main}.
Our next result nevertheless shows that the Chern currents lie in the right cohomology classes when the manifold is compact by using Demailly's regularization in \cite{DemaillyReg}
\begin{thm}\label{thm:Cohomology}
Let $h$ be a singular hermitian metric on a holomorphic vector bundle $E\to X$ satisfying the assumptions of Theorem \ref{thm:main} so that the Chern current $c_k(E,h)$ can be defined.
If $X$ is compact, then
$$[c_k(E,h)]=c_k(E)$$
where $c_k(E)=[c_k(E,h_0)]$ is the usual Chern class defined by any smooth metric $h_0$ on $E$.
\end{thm}
\begin{remark} Chern classes of certain singular hermitian metrics have been studied earlier by Mumford in \cite{Mum}:
Let $X$ be a Zariski open set of a projective variety $\bar X$ such that $\bar X\setminus X$ is a divisor with normal crossings, and let $\bar E$ be a vector bundle on $\bar X$.
In \cite{Mum}*{\S\,1}, Mumford studies smooth hermitian metrics on $E=\bar E|_X$
which satisfy certain growth conditions when approaching $\bar X\setminus X$.
These metrics could then be considered as singular metrics on $\bar E$.
Mumford then uses the classical Bott-Chern theory and the growth assumptions to show that trivial extensions of the Chern forms on $X$ to all of $\bar X$ represent the Chern classes of $\bar E$ (in all degrees).
There is no apparent relation between the growth assumptions in \cite{Mum} and the positivity assumptions that we consider.
Our approach to defining Chern currents is different to the one in \cite{Mum}, where ours is based on using the positivity of the
curvature to be able to use pluripotential theoretical methods in the spirit of Bedford-Taylor-Demailly. This approach
might in particular produce currents with mass on the singular set of the metric, in contrast to the case in \cite{Mum}.
\end{remark}
The paper is organized as follows. Section 2 contains a short introduction to Segre forms. In section 3 we collect a few preliminary results that will be needed in the proofs of Theorem \ref{thm:ContThm} and Theorem \ref{thm:main}, and some basic facts from pluripotential theory. Section 4 is devoted to the proofs of Theorem \ref{thm:ContThm} and Theorem \ref{thm:main}. Finally, in section 5 we end by showing that on a compact manifold, the Chern currents of Theorem \ref{thm:main} are in the right cohomology class.
\section*{Acknowledgements}
\noindent It is a pleasure to thank Bo Berndtsson and Mihai P{\u{a}}un for expressing interesting in our work and for valuable comments.
We also thank the anonymous referee for helpful suggestions which helped to improve the paper.
\section{Segre forms}
\label{sec:segre-forms}
\noindent Let $E \to X$ be a holomorphic vector bundle with a smooth metric $h$. By definition, the \emph{total Segre form} $s(E,h)=1+s_1(E,h)+\cdots+s_n(E,h)$ associated with $(E,h)$ is the multiplicative inverse of the total Chern form $c(E,h)=1+c_1(E,h)+\cdots+c_n(E,h)$. Hence, as mentioned in the introduction, they can be expressed recursively through,
$$s_k(E,h)+s_{k-1}(E,h)\wedge c_1(E,h)+\cdots+c_k(E,h)=0,\quad k=1,\ldots,n.$$
The first three Segre forms, for example, are
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
s_1(E,h)&=&-c_1(E,h),\nonumber\\
s_2(E,h)&=&c_1(E,h)^2-c_2(E,h),\nonumber\\
s_3(E,h)&=&-c_1(E,h)^3+2c_1(E,h) \wedge c_2(E,h)-c_3(E,h),\nonumber
\end{eqnarray}}
\!\!and expressing the first three Chern forms in terms of these Segre forms yields,
{\setlength\arraycolsep{2pt}
\begin{eqnarray}
c_1(E,h)&=&-s_1(E,h),\nonumber\\
c_2(E,h)&=&s_1(E,h)^2-s_2(E,h),\nonumber\\
c_3(E,h)&=&-s_1(E,h)^3+2s_1(E,h)\wedge s_2(E,h)-s_3(E,h).\nonumber
\end{eqnarray}}
Our interest in the Segre forms stems from the fact that they turn out to be closely related to the projectivized bundle, $\pi:{\mathbb P}(E)\to X$, associated with $E$. Recall that this fiber bundle is constructed by letting ${\mathbb P}(E)_x:={\mathbb P}(E_x^*)$ for each $x\in X$ (the projectivization of an $r$-dimensional vector space, where $E^*$ denotes the dual bundle of $E$). The pullback bundle $\pi^*E^*\to{\mathbb P}(E)$ will then carry a tautological sub-bundle ${\mathcal O}_{{\mathbb P}(E)}(-1)$, where the notation is justified by the fact that fiberwise this is nothing but ${\mathcal O}(-1)$ over ${\mathbb P}^{r-1}$. The global holomorphic sections of ${\mathcal O}_{{\mathbb P}(E)}(1)$ (the dual of ${\mathcal O}_{{\mathbb P}(E)}(-1)$), over any fiber are in one-to-one correspondence with linear forms on $E_x^*$, i.e.\ with the elements of $E_x$ (this is the reason for projectivizing $E^*$ instead of $E$).
Now if $h$ is a smooth metric on $E$, then ${\mathcal O}_{{\mathbb P}(E)}(1)$ can be equipped with a natural metric which we will denote by $e^{-\varphi}$. We let $\Phi$ denote the first Chern form of this metric, i.e.
$$\Phi:=\frac{i}{2\pi}\Theta\big({\mathcal O}_{{\mathbb P}(E)}(1),e^{-\varphi}\big).$$
The following proposition is at the center of our interest in Segre forms:
\begin{prop}\label{prop:Segre}
Let $h$ be a smooth hermitian metric on a holomorphic vector bundle $E \to X$.
Then
$$s_k(E,h) = (-1)^k\pi_*\big(\Phi^{k+r-1}\big).$$
\end{prop}
To our knowledge this result first appears in \cite{Mou} (see also \cite{G} and \cite{Diverio}). We will use Proposition~\ref{prop:Segre} to define Segre forms in the singular setting. Before we turn to this we first need to establish a few preliminary results and recall some facts from pluripotential theory.
\section{Preliminaries}
\noindent Throughout the paper, we are interested in singular hermitian metrics that are Griffiths positive as in Definition \ref{df:Gr}. Since this condition is formulated in terms of the dual metric we first of all need to check that this dual metric is also a singular hermitian metric in the sense of Definition \ref{df:sing}. Already in \cite{BP} (section 3) this is claimed in passing, but since no argument is presented, for the sake of completeness, we include a proof here.
\begin{lma}\label{lma:SingularDual}
Any singular hermitian metric $h$ on a vector bundle $E$ induces a canonical dual singular hermitian metric $h^*$ on the dual bundle $E^*$
such that $h^{**} = h$ under the natural isomorphism $E^{**} \cong E$.
\end{lma}
\begin{proof}
As the statement is pointwise, it is enough to consider a fixed fiber $E_x=:V$. Recall that by Definition \ref{df:sing} a singular hermitian form $h$ on $V$ is a map
\[V\rightarrow [0,\infty],\ \|\xi\|_h:=\begin{cases}\|\xi\|_{h_0}&\hbox{if } \xi\in V_0 \\\ \infty & \hbox{otherwise} \end{cases}\]
where $V_0$ is the subspace of $V$ where $h$ is finite, and $h_0$ is a hermitian form (with values in $[0,\infty)$) on $V_0$.
We want to show that any singular hermitian form $h$ on a vector space $V$ given by the subspace $V_0$ and the hermitian form $h_0$ on $V_0$ induces a singular dual hermitian form on the dual of $V$.
Let $N$ denote the linear subspace of $V_0$ where the hermitian form $h_0$ degenerates, i.\,e., $N$ is the eigenspace of the eigenvalue $0$. Then, $V_0/N$ admits a (canonical) hermitian form $\tilde h_0$ induced by $h_0$ which is non-degenerate (given by $\|\xi+N\|_{ \tilde h_0}:=\|\xi\|_{h_0}$). In particular, there exits a dual hermitian form ${\tilde h_0}{\!\!\big.}^\ast$ on $(V_0/N)^\ast$.
We let $W := V^*$ and $W_0:=N^o=\{\eta\in W:\eta|_N=0\}$ be the annihilator of $N$.
Each $\eta_0:=\eta|_{V_0}$ with $\eta\in W_0$ can be seen as an element $\tilde \eta_0$ of $(V_0/N)^\ast$ defined by $\tilde \eta_0 (\xi+N) := \eta_0(\xi)$,
Hence, we obtain the dual singular hermitian form
\[\|\eta\|_{h^\ast}:= \begin{cases}\big\|\widetilde{\eta|_{V_0}}\big\|_{{\tilde h_0}{\!\!\!\big.}^\ast} &\hbox{if } \eta\in W_0 \\ \ \infty & \hbox{otherwise}\end{cases} \]
on $W=V^\ast$ associated to $h$.
If we let $M$ denote the linear subspace of $W_0$ where $h^*$ degenerates, then by definition of $h^*$, one obtains that $M = V_0^o=\{ \eta \in W_0 : \eta|_{V_0} = 0 \}$ is the annihilator of $V_0$.
Thus, it follows that $W_0/M \cong (V_0/N)^*$.
If we let $U := W^* \cong V$ and do the above construction for $(W,h^*)$, then one obtains, using the isomorphism $U \cong V^{**}$, that
\begin{equation*}
U_0 = M^o\cong \{ v : \eta(v) = 0 \text{ for all $\eta \in W_0$ such that $\eta|_{V_0} = 0$} \} = V_0.
\end{equation*}
Finally if $P$ is the null space of the induced metric $h^{**}$, then
\begin{equation*}
P=W_0^o \cong \{ v \in V_0 : \eta(v) = 0 \text{ for all $\eta \in W_0$ such that $\eta|_N = 0$} \} = N,
\end{equation*}
and one readily verifies that the metric $h^{**}$ on $U_0 \cong V_0$ equals $h$.
\end{proof}
In the previous section we mentioned the close connection between Segre forms and the Serre line bundle ${\mathcal O}_{{\mathbb P}(E)}(1)$ over the projective bundle ${\mathbb P}(E)$ (Proposition \ref{prop:Segre}), and how we will use this connection to define Segre currents in the singular setting. For a smooth metric $h$ it is well-known that if $h$ is Griffiths positive, then the induced metric $e^{-\varphi}$ on ${\mathcal O}_{{\mathbb P}(E)}(1)$ will be positive as well (see e.g. \cite{Z}, Example~7.10). Before we can apply techniques from pluripotential theory to define the push-forward in Proposition \ref{prop:Segre} for singular metrics, we need to check that $e^{-\varphi}$ is positive in the singular setting as well.
\begin{lma} \label{lma:inducedPositive}
Let $E$ be a vector bundle, and let $h$ denote a singular, Griffiths positive, hermitian metric on $E$. Let $e^{-\varphi}$ be the induced metric on the line bundle ${\mathcal O}_{{\mathbb P}(E)}(1)$. Then $e^{-\varphi}$ is a singular positive metric.
\end{lma}
\begin{proof}
The Griffiths positivity of $h$ implies (by definition) that the dual metric $h^*$ is Griffiths negative on $E^*$. By arguing as in \cite{PT}, Lemma 2.3.2 (a) the pullback metric $\pi^* h^*$ on $\pi^* E^*$ will also be Griffiths negative, and since negativity is preserved when restricting to subbundles, this in turn implies that the induced metric on ${\mathcal O}_{{\mathbb P}(E)}(-1)$ is negative as well. Thus, the metric $e^{-\varphi}$ on ${\mathcal O}_{{\mathbb P}(E)}(1)$ is positive.
\end{proof}
The standard way of defining powers of currents of the form $dd^cu$, where $u$ is a plurisubharmonic function, is through the inductive approach of Bedford-Taylor \cite{BT} (here $d^c=\frac{i}{4\pi}(\bar\partial-\partial)$). However, in our case there are some extra twists. First off, for the classic theorem of Bedford-Taylor one needs the plurisubharmonic functions to be locally bounded, which is not the case for us. Fortunately, it turns out that this assumption can be replaced by requiring the \textit{unbounded locus} to be of appropriate codimension. Let us quickly recall this important concept.
\begin{df}\label{df:ULocus}
Let $u$ be a plurisubharmonic function on a complex manifold $X$. The \textit{unbounded locus} of $u$, $L(u)$, is defined to be the set of points $x\in X$ such that $u$ is unbounded in every neighborhood of $x$.
\end{df}
Note that if $h = e^{-\varphi}$ is a singular positive hermitian metric on a line bundle, then in a local frame, a local representative of $\varphi$ is plurisubharmonic, so $L(\varphi)$ is defined locally with respect to this frame. However, since two different local representatives of $\varphi$ as defined with respect to two different frames will differ by a locally bounded (harmonic) function, $L(\varphi)$ is well-defined, independent of the local frame.
\begin{remark}\label{remark:UEx}
Although $L(u)$ is always closed and contains the closure of the pole set of $u$, these two set are different in general. In \cite{D2}, Chapter III.4, the function
$$u(z)=\sum\frac{1}{k^2}\log\Big(\big|z-\frac{1}{k}|+e^{-k^3}\Big)$$
is provided as an example of a function which is everywhere finite in ${\mathbb C}$, but with $L(u)=\{0\}$.\qed
\end{remark}
The requirements we impose in the main theorems are inspired by Demailly's variant of the Bedford-Taylor
theorem that makes it possible to define powers of currents $dd^cu$,
where $u$ is a not necessarily locally bounded plurisubharmonic function.
\begin{thm}\label{thm:PPT}
Let $u$ be a plurisubharmonic function, and let $T$ denote a positive $(q,q)$-current. If $L(u) \cap \supp T$ is contained in some
variety $V$ with $\codim(V)\geq k+q$, then there exists a well-defined closed positive $(k+q,k+q)$-current $(dd^cu)^k\wedge T$,
which can be defined locally as the limit of
$$ (dd^c u_\nu)^k \wedge T$$
for any sequence $\{u_\nu\}$ of smooth plurisubharmonic functions decreasing pointwise to $u$.
\end{thm}
We refer to \cite{D2}, Chapter III, Theorem 4.5 and Proposition 4.9, or \cite{D1}, Proposition 2.2.3 for a proof of (more general variants of) this theorem.
\begin{remark}\label{remark:codim}
Looking at Proposition~\ref{prop:Segre} with $\Phi=dd^c\varphi$ for some plurisubharmonic function $\varphi$, at first glance it might seem tempting to use Theorem~\ref{thm:PPT} to define $(dd^c\varphi)^{k+r-1}$. For such a direct application of the theorem, however, we would need to assume that $L(\varphi)$ is contained in some subvariety $W$ with
$$\codim(W)\geq k+r-1,$$
a requirement which is much too strong to be of any practical use, and which in view of Lemma \ref{lma:UnbLoc} does not follow from the
codimension requirements in Theorem~\ref{thm:ContThm} or Theorem~\ref{thm:main}. Instead, we will aim at defining the push-forward directly.
Thus, the main part of the proofs of these theorems will basically be an adjustment of the proof of Theorem~\ref{thm:PPT} for the push-forward.
\end{remark}
\begin{lma}\label{lma:UnbLoc}
Let $E$ be a holomorphic vector bundle with a singular Griffiths positive metric $h$, and let
$e^{-\varphi}$ denote the singular positive metric on ${\mathcal O}_{{\mathbb P}(E)}(1)$ as in Lemma~\ref{lma:inducedPositive}.
Then,
\begin{equation} \label{eq:unbounded}
L(\varphi) \subseteq \pi^{-1}(L(\log \det h^*))
\end{equation}
where $h^*$ denotes the induced dual metric on the dual bundle $E^*$.
\end{lma}
\begin{proof}
We take $x\in L(\log\det h^*)^c$ so that $\log\det h^*$ is bounded in
a neighborhood $U_x$ of $x$. We are going to show that the eigenvalues $\lambda_1,\ldots,\lambda_r$
of $h^*$ are uniformly bounded from above and below in $U_x$, by strictly positive constants.
This will show that $\pi^{-1}(x)\in L(\varphi)^c$ since if $e$ is a local frame for ${\mathcal O}_{{\mathbb P}(E)}(-1)$,
then $\varphi=\log\|e\|^2_{\pi^*h^*}$.
If $h^*$ is expressed as a matrix with respect to a holomorphic frame $u_1,\ldots,u_r$ for $E^*$, then
$$\tr(h^*)=\|u_1\|^2_{h^*}+\cdots+\|u_r\|^2_{h^*}.$$
By the Griffiths negativity of $h^*$, all the norms are plurisubharmonic, and so in particular (after
possibly shrinking $U_x$) bounded from above on $U_x$. Let $M>0$ be such that $\lambda_1,\ldots,\lambda_r\leq M$ on $U_x$.
On the other hand, $x\in L(\log\det h^*)^c$ implies that $\det h^*>C$ on $U_x$, for some $C>0$. Thus,
the smallest eigenvalue $\lambda_k$ of $h^*$ is uniformly bounded from below on $U_x$ by
$$\lambda_k>\frac{C}{\lambda_1\cdots\hat{\lambda}_k\cdots\lambda_r}\geq\frac{C}{M^{r-1}}>0.$$
\end{proof}
As mentioned above, the proofs of our main theorems will be an adaption of the proof of Theorem~\ref{thm:PPT}. The basic idea behind this proof is to reduce to the locally bounded situation, where one can apply the original Bedford-Taylor result. We will use the same basic approach in our proofs, but once the reduction has been carried out we will be in the setting of singular metrics on line bundles, and so we will need the following variant of the Bedford-Taylor theorem.
\begin{thm}\label{thm:PPT2}
Let $e^{-\varphi}$ be a non-degenerate singular positive metric on a line bundle $L$, let
$e^{-\varphi_\nu}$ be a sequence of smooth positive metrics on $L$, and let $e^{-\psi}$ be a fixed smooth metric on $L$. Let furthermore $\{ T_\nu \}$ be a sequence of closed positive currents
converging to a current $T$. Assume that either:\vspace{0.1cm}\\
\noindent(i) $\varphi_\nu$ converges locally uniformly to $\varphi$, or \\
\noindent(ii) $\varphi_\nu$ decreases pointwise to $\varphi$ and $T_\nu = T$.\vspace{0.1cm}
Then, the limits of
$$ (\varphi_\nu-\psi) (dd^c \varphi_\nu)^k \wedge T_\nu \text{ and } (dd^c \varphi_\nu)^k \wedge T_\nu,$$
as $\nu \to \infty$ exist in the sense of currents. These limits are independent of the sequences
$\{ \varphi_\nu \}$ and $\{ T_\nu \}$, and can thus be used as definitions of the $(k+q,k+q)$-currents
$(\varphi-\psi)(dd^c \varphi)^k\wedge T$ and $(dd^c \varphi)^k\wedge T$.
\end{thm}
A proof of this in the case when $L$ is a trivial line bundle and $\psi = 0$ can be found in for example \cite{D2},
Chapter III, in Corollary 3.6 under the assumption (i), and in Theorem 3.7 under the assumption (ii).
The case of singular metrics on line bundles then follows directly from this case since
in a local frame, $\varphi_\nu - \psi$ and $\varphi - \psi$ are the sum of a plurisubharmonic and a smooth function, whose sum is independent of the local trivialization, and in addition $dd^c \varphi_\nu$ is independent of the local trivialization.
\section{Segre Currents}
\label{sec:Segre_currents}
\noindent As mentioned previously, the existence of Chern currents will be achieved by expressing them in terms of Segre currents, as we do for forms in section~\ref{sec:segre-forms}. Hence, we will first need to prove the existence of such Segre currents and furthermore that it is possible to define the wedge products of these. A key result for achieving this is the fact that Definition~\ref{df:Gr} provides us with a notion of positivity that is easy to approximate, see \cite{BP}, Proposition 3.1 and \cite{R}, Proposition 6.2.
\begin{prop}\label{prop:appr}
Let $E$ be a trivial holomorphic vector bundle over a polydisc, and let $h$ be a singular Griffiths positive (negative) hermitian metric on $E$. Then on any smaller polydisc, there exists a sequence of smooth hermitian metrics $\{h_\varepsilon\}$ with positive (negative) Griffiths curvature, increasing (decreasing) to $h$ pointwise.
\end{prop}
The key ingredient in proving the local existence of Segre currents, and their wedge products, is Lemma \ref{lma:locallyBounded} below. Before we get to this result however, we first need to introduce the following concept:
\begin{df}
We say that a smooth $(p,p)$-form $\beta$ is a \emph{bump form} at a point $x$ if it is strongly positive, and such that for some (or equivalently for any) K\"ahler form $\omega$ defined near $x$, there exists a constant $C > 0$ such that $C\omega^p \leq \beta$ as strongly positive forms in a neighborhood of $x$.
\end{df}
We will also use in the proof of the following lemma, as well as later on, the basic formula
that if $\pi : X \to Y$ is a proper submersion, and if $\gamma$ is either a smooth
form or a current on $Y$, and $\eta$ is a smooth form on $X$, then
\begin{equation} \label{eq:gammaeta}
\gamma \wedge \pi_* \eta = \pi_*( \pi^* \gamma \wedge \eta).
\end{equation}
\begin{lma}\label{lma:locallyBounded}
Let $E$ be a trivial, rank $r$ vector bundle over a polydisc $P\subset{\mathbb C}^n$, let $h$ denote a singular Griffiths positive hermitian metric on $E$, and let $\{ h_\varepsilon \}$ be a family of smooth hermitian metrics on $E$. Let furthermore $\{T_\varepsilon\}$ be a sequence of closed positive $(q,q)$-currents converging to a current $T$.
Assume that $L(\log\det h^*)$ is contained in some subvariety $V$ of $P$ such that $\codim(V)\geq k+q$. Assume also that either:\vspace{0.1cm}\\
\noindent(i) Outside of $V$, $h$ is non-degenerate and continuous and $h_\varepsilon \to h$ locally uniformly, and on all of $P$, $T_\varepsilon \to T$, or \\
\noindent (ii) On all of $P$, $h_\varepsilon$ increases pointwise to $h$ and $T_\varepsilon = T$.
Then, for any point in $P$, there exists an $(n-k-q,n-k-q)$ bump form $\beta$ with arbitrarily small support
such that the limit of
\begin{equation} \label{eq:locallyBounded}
(-1)^k \int T_\varepsilon \wedge s_k(E,h_\varepsilon)\wedge \beta
\end{equation}
as $\varepsilon$ tends to 0 exists, and is independent of the sequences $\{h_\varepsilon\}$ and $\{T_\varepsilon\}$.
\end{lma}
\begin{proof}
The idea is to express the Segre forms $s_k(E,h_\varepsilon)$ as in Proposition~\ref{prop:Segre}, and then use a similar procedure
as the proof of Theorem~\ref{thm:PPT}.
First of all, by Lemma~\ref{lma:inducedPositive}, the Griffiths positivity of $h$ implies that the metric $e^{-\varphi}$ on ${\mathcal O}_{{\mathbb P}(E)}(1)$ is positive.
By \eqref{eq:unbounded}, if $W := \pi^{-1}(V)$, then $L(\varphi) \subseteq W$, and we note that since $\pi$ is a submersion,
$W$ has codimension $\geq k+q$ in ${\mathbb P}(E)$.
For a subset $I \subseteq \{1,\dots,n\}$ consisting of $n-k-q$
distinct elements $I_1 < \dots < I_{n-k-q}$, we define $p_I : {\mathbb C}^n \to {\mathbb C}^{n-k-q}$ to be the projection onto the
coordinates $(z_{I_1},\dots,z_{I_{n-k-q}})$, and similarly, we let $p_{I^c} : {\mathbb C}^n \to {\mathbb C}^{k+q}$ be the projection
onto the remaining coordinates. After any generic linear change of coordinates, for a subvariety $V$ of codimension $k+q$
with $0 \in V$, the projections $p_I$ and $p_{I^c}$ satisfy that the origin is an isolated point of $p_I^{-1}(0)$ in $V$,
and there exist small balls $B_I'\subset{\mathbb C}^{n-k-q}$ and $B_I''\subset{\mathbb C}^{k+q}$ and $0 < \delta_I < 1$ such that
$$V \cap p_I^{-1}(\bar B_I') \cap p_{I^c}^{-1}(\overline{B}_I'' \setminus (1-\delta_I) B_I '') = \emptyset.$$
Since this holds for a generic projection, we can in fact assume that it holds simultaneously for all
the projections $p_I$ and $p_{I^c}$ for all subsets $I \subseteq \{1,\dots,n\}$ consisting of $n-k-q$ elements.
We take $\chi_{I,1}$ and $\chi_{I,2}$ to be positive cut-off functions with compact support on $B_I'$ and $B_I''$
respectively, and which are both strictly positive at $0$, and such that $\chi_{I,2}$ is constant on $(1-\delta_I) B_I''$,
and we let
\begin{equation*}
\beta := \sum_I (p_I^* \chi_{I,1}) (p_{I^c}^* \chi_{I,2}) idz_{I_1} \wedge d\bar{z}_{I_1} \wedge \dots \wedge i dz_{I_{n-k-q}} \wedge d\bar{z}_{I_{n-k-q}},
\end{equation*}
where the sum is over all ordered subsets $I = \{ I_1,\dots,I_{n-k-q} \mid I_1 < \dots < I_{n-k-q} \}$ of $\{1,\dots,n\}$. Then, $\beta$ is a bump form at $\{ 0 \}$,
and if $B_I'$ and $B_I''$ were chosen small enough above, then $\beta$ can be chosen to have arbitrarily small support.
By linearity, it is enough to prove \eqref{eq:locallyBounded} for each single term in the sum above, and without loss of generality,
we can assume that the term is when $I = \{1,\dots,n-k-q\}$. From now on, for convenience of notation, we write $\chi_1 = \chi_{I,1}$,
$\chi_2 = \chi_{I,2}$, $B' = B_I'$ and $B'' = B_I''$, $\delta = \delta_I$,
and write the coordinates on ${\mathbb C}^n$ as $(z',z'') \in {\mathbb C}^{n-k-q} \times {\mathbb C}^{k+q}$.
We will thus prove \eqref{eq:locallyBounded} for
\begin{equation} \label{eq:simpler_beta}
\beta = \chi_1(z') \chi_2(z'') \beta_0, \text{ where } \beta_0 := i dz_1 \wedge d\bar{z}_1 \wedge \dots \wedge idz_{n-k-q} \wedge d\bar{z}_{n-k-q}.
\end{equation}
The approximating sequence $\{h_\varepsilon\}$ yields an approximating sequence $\{e^{-\varphi_\varepsilon}\}$ on ${\mathcal O}_{{\mathbb P}(E)}(1)$,
such that either $\varphi_\varepsilon$ converges locally uniformly to $\varphi$ outside of $W$ or decreases pointwise to $\varphi$, depending
on whether we are under assumption (i) or (ii). As these metrics are smooth, we can apply Proposition~\ref{prop:Segre}, and combining this with
\eqref{eq:gammaeta}, we deduce that
$$(-1)^k\int_P T_\varepsilon \wedge s_k(E,h_\varepsilon) \wedge \beta =\int_P T_\varepsilon \wedge\pi_*\big((dd^c\varphi_\varepsilon)^{k+r-1}\big)\wedge \beta
=\int_{P \times {\mathbb P}^{r-1}}\pi^*T_\varepsilon \wedge (dd^c\varphi_\varepsilon)^{k+r-1}\wedge \pi^* \beta,$$
where $\pi^* T_\varepsilon$ exists as a closed positive $(q,q)$-current since $\pi: P \times{\mathbb P}^{r-1}\to P$ is a proper submersion.
The main point with introducing $B'$ and $B''$ is that $V$ is disjoint from the compact cylinder,
$$K_\delta:=\overline{B}' \times \big(\overline{B}'' \setminus (1-\delta)B''\big)$$
and thus, $\varphi$ is locally bounded on $\pi^{-1}(K_\delta)$.
By performing integration by parts, we want to move to an integral on $\pi^{-1}(K_\delta)$, which then basically
reduces the problem to the well-known locally bounded setting.
Since we are dealing with metrics (not functions), we cannot directly perform integration by parts,
but we first need to add and subtract a smooth reference metric, say $(dd^c\varphi_1)^{k+r-1}$ with $\varepsilon=1$ fixed,
\begin{equation*
\begin{split}
&\int_{P\times{\mathbb P}^{r-1}} \pi^*T_\varepsilon \wedge(dd^c\varphi_\varepsilon)^{k+r-1} \wedge \pi^* \beta=
\int_{P\times{\mathbb P}^{r-1}} \pi^*T_\varepsilon\wedge (dd^c\varphi_1)^{k+r-1} \wedge \pi^* \beta+\\
&\quad+\int_{P\times{\mathbb P}^{r-1}} \hspace{-1em}\pi^*T_\varepsilon\wedge dd^c(\varphi_\varepsilon-\varphi_1)\wedge \Big(\sum_{j=0}^{k+r-2}(dd^c\varphi_\varepsilon)^{k+r-2-j}\wedge(dd^c\varphi_1)^{j}\Big)\wedge \pi^* \beta=:I+II.
\end{split}\end{equation*}
The integral $I$ converges as $\varepsilon \to 0$ since $T_\varepsilon \to T$. Since $\chi_1$ only depends on $z'$, while $\pi^*\beta$
is already of full degree in the $z'$-variables, we obtain that $d\chi_1 \wedge \beta_0 = 0$ and $d^c \chi_1 \wedge \beta_0 = 0$.
Using this in combination with that $\pi^* T_\varepsilon$ is $d$ and $d^c$-closed, we can formally
perform integration by parts in $II$, and obtain that it equals
\begin{equation*
II = \int_{P \times{\mathbb P}^{r-1}}\pi^*T_\varepsilon \wedge(\varphi_\varepsilon-\varphi_1)\Big(\sum_{j=0}^{k+r-2}(dd^c\varphi_\varepsilon)^{k+r-2-j}\wedge(dd^c\varphi_1)^{j}\Big)\wedge\pi^* \chi_1 \pi^*dd^c \chi_2 \wedge \pi^*\beta_0.
\end{equation*}
Since $\varphi$ is locally bounded on $\pi^{-1}(K_\delta)$, and $\supp dd^c\chi_2 \subset \overline{B}''\setminus (1-\delta)B''$,
we get by Theorem~\ref{thm:PPT2} that the limit of $II$ when $\varepsilon$ tends to $0$
exists, and is independent of the choice of regularizing sequences $\{\varphi_\varepsilon\}$ and $\{T_\varepsilon\}$.
\end{proof}
From the proof, it follows that where $h$ is non-degenerate, one can form the product of a closed positive current and a Segre current.
More precisely:
\begin{lma}\label{lma:UniqueLim}
Let $E$, $h$, $\{h_\varepsilon\}$, $T$ and $\{T_\varepsilon\}$ be as in Lemma~\ref{lma:locallyBounded}.
If we are under either the assumptions (i) or (ii) of this lemma, then
outside of $L(\log \det h^*)$, the limit of
\begin{equation*}
T_\varepsilon \wedge s_k(E,h_\varepsilon)
\end{equation*}
as $\varepsilon$ tends to $0$ exists in the sense of currents, and the limit is independent of the sequences $\{h_\varepsilon\}$ and $\{T_\varepsilon\}$.
\end{lma}
\begin{proof}
If we fix $\varepsilon$, then $h_\varepsilon$ is smooth, and we denote by $e^{-\varphi_\varepsilon}$ the
induced metric on ${\mathcal O}_{{\mathbb P}(E)}(1)$. Since $\varphi_\varepsilon$ is smooth, we get by Proposition~\ref{prop:Segre}
together with \eqref{eq:gammaeta} that
\begin{equation} \label{eq:UniqueLim}
T_\varepsilon \wedge s_k(E,h_\varepsilon) = (-1)^k \pi_* ( \pi^* T_\varepsilon \wedge (dd^c \varphi_\varepsilon)^{k+r-1} ).
\end{equation}
By Theorem~\ref{thm:PPT2}, \eqref{eq:UniqueLim} has a limit as $\varepsilon$ tends to 0,
and the limit is independent of the sequences $\{\varphi_\varepsilon\}$ and $\{T_\varepsilon\}$.
\end{proof}
As mentioned above, the local existence of Segre currents will follow from Lemma~\ref{lma:locallyBounded}. For the transition from local to global, we will use the following lemma.
\begin{lma}\label{lma:Unique}
Let $S$ and $T$ be two closed, positive $(k,k)$-currents such that $S = T$ outside a subvariety $A$ with $\codim(A)\geq k$, and assume that for each point of $p \in A$, there exists an $(n-k,n-k)$ bump form $\beta$ at $p$ with arbitrarily small support such that
\begin{equation*}
\int S \wedge \beta = \int T \wedge \beta.
\end{equation*}
Then $S = T$ everywhere.
\end{lma}
\begin{proof}
The difference $U := S-T$ is a closed $(k,k)$-current of order $0$ with support on $A$. By \cite{D2}, Chapter III, Corollary 2.14, $U = \sum \lambda_i [A_i]$, where $A_i$ are the irreducible components of codimension $k$ of $A$, and $\lambda_i \in \mathbb{C}$. We want to show that $\lambda_i = 0$ for all $i$.
Fix some $i$, take a point in $A_i$ which does not belong to any of the other $A_j$'s, and take a $(n-k,n-k)$ bump form $\beta$ whose support still does not intersect any of the other $A_j$'s. Multiplying $S-T$ with $\beta$ and integrating, we get that
\begin{equation*}
\lambda_i \int_{A_i} \beta = \int (S-T) \wedge \beta = 0.
\end{equation*}
The integral on the left-hand side is non-zero, so $\lambda_i = 0$.
\end{proof}
With these lemmas at our disposal, we can now define Segre currents and their wedge products.
\begin{prop} \label{prop:segre_products}
Let $E\to X$ be a rank $r$ holomorphic vector bundle over an $n$-dimensional complex manifold $X$, and let $h$ denote a singular Griffiths positive hermitian metric on $E$.
Take $k_1,\ldots,k_m\in\mathbb{N}$ and let $k := k_1 + \dots + k_m$. If $L(\log\det h^*)$ is contained in some subvariety $V$ of $X$ with
$\codim(V)\geq k$, then the wedge product
$$(-1)^k s_{k_1}(E,h)\wedge\cdots\wedge s_{k_m}(E,h)$$
can be defined as a closed, positive current on $X$, through regularization by an iterated limit as in (\ref{eq:IteratedLimit}).
\end{prop}
\begin{proof}
The proof is by induction over $m$, starting with the trivial case $m = 0$, when the current is simply $1$. We thus assume by induction that $T := (-1)^{k-k_m} s_{k_1}(E,h) \wedge \dots \wedge s_{k_{m-1}}(E,h)$ exists as a well-defined positive
closed current, and we start by proving existence locally, which we will achieve with the help of Lemma~\ref{lma:locallyBounded}. Hence, assume that $E$ is a trivial vector bundle over a polydisc, let $\{h_\varepsilon\}$ denote an approximating sequence to $h$ (as in Proposition~\ref{prop:appr}), and let $\{s_{k_m}^\varepsilon\}$ denote the corresponding sequence of Segre forms.
Since $(-1)^{k_m} s_{k_m}(E,h_\varepsilon)$ is a smooth and strongly positive $(k_m,k_m)$-form, the product $S_\varepsilon := (-1)^{k_m} T \wedge s_{k_m}^\varepsilon$ is a closed positive current, and its mass is dominated by its trace measure,
$$\| S_\varepsilon \| \leq \int S_\varepsilon \wedge (i\partial\bar\partial|z|^2)^{n-k}.$$
By Lemma~\ref{lma:locallyBounded}, this integral is uniformly bounded in $\varepsilon$ on a small enough neighborhood of $0$
since $(i\partial\bar\partial|z|^2)^{n-k}$ is dominated by a constant times any bump form at $\{ 0 \}$ in such a neighborhood.
Hence, it follows from the Banach-Alaoglu theorem that we can find a sequence $\varepsilon_j\rightarrow 0$ such that
$S_{\varepsilon_j}$ converges weakly to a closed, positive $(k,k)$-current $S$.
From combining Lemma~\ref{lma:UniqueLim} and Lemma~\ref{lma:locallyBounded} (under the assumption (ii)) with Lemma~\ref{lma:Unique},
we obtain that the limit $S$ is unique and only depends on $T$ and $h$.
We then in fact obtain that the full sequence $\{ S_{\varepsilon} \}$ converges to $S$
as $\varepsilon \to 0$, since otherwise, we could find a sequence $\{ S_{\varepsilon_j}\}$ with $\varepsilon_j \to 0$,
and a test-form $\phi$ such that the limit of $\int S_{\varepsilon_j} \wedge \phi$ exists, but is not equal to $\int S \wedge \phi$.
After passing to a subsequence, we could assume that $\{S_{\varepsilon_j}\}$ is indeed convergent,
and obtain a contradiction to the uniqueness of the limiting current $S$.
We take the limiting current $S$ as the (local) \emph{definition} of $(-1)^k s_{k_1}(E,h) \wedge \dots \wedge s_{k_m}(E,h)$,
and since it is independent of the regularizing sequence, this locally defined product of Segre currents in fact defines a global
current on $X$.
\end{proof}
\begin{remark}\label{remark:wedge_product_of_currents}
Since we will need a bit more general result in the next section, we note that the proof of Proposition \ref{prop:segre_products}
actually gives the following statement:
\emph{
Let $T$ be a closed positive $(q,q)$-current, let $\{\theta_\eps\}$ be a sequence of closed smooth $(k,k)$-forms, and let $V$ be a variety of codimension at least $k+q$. Moreover, we assume that the following three conditions are satisfied: \\
(i) $T\wedge\theta_\eps$ are positive (which is the case for instance when $\theta_\eps$ are strongly positive). \\
(ii) For any point in $X$, there exists a bump form $\beta$ such that $\int\! T\wedge\theta_\eps\wedge\beta$ converges. \\
(iii) Outside of $V$, $T\wedge \theta_\eps$ converges to a (closed positive) current $\tilde S$. \\
Then, the limit $S := \lim_{\eps\to 0} T\wedge\theta_\eps$ exists and is a closed positive current with $S|_{X\setminus V}=\tilde S$. Furthermore, $S$ is uniquely determined by $\tilde S$ in (iii) and the limits of $\int\! T\wedge\theta_\eps\wedge\beta$ in (ii).
}
\end{remark}
Theorem~\ref{thm:main} is an immediate consequence of Proposition~\ref{prop:segre_products} as outlined in the introduction. It is however not enough for Theorem~\ref{thm:ContThm}. For this latter theorem we need the following lemma, whose proof
is similar to the proof of Proposition~\ref{prop:segre_products}.
\begin{lma}\label{lma:UniqueLimCh}
Let $E\to X$ be a rank $r$ holomorphic vector bundle over an $n$-dimensional complex manifold $X$,
and let $h$ denote a singular Griffiths positive hermitian metric on $E$. Let also $\{ T_\varepsilon \}$
be a family of closed positive $(q,q)$-currents on $X$ converging to a current $T$.
Assume that $L(\log \det h^*)$ is contained in some subvariety $V$ of $X$ with $\codim(V) \geq k+q$.
If $\{ h_\varepsilon \}$ is a family of smooth hermitian metrics on $E$ converging locally uniformly to $h$ outside of $V$,
then there exist closed positive $(k+q,k+q)$ currents $S_{\varepsilon}^+$ and $S_{\varepsilon}^-$ such that
\begin{equation*}
T_\varepsilon \wedge c_k(E,h_\varepsilon) = S_{\varepsilon}^+ - S_{\varepsilon}^-,
\end{equation*}
and the limits of both $S_{\varepsilon}^+$ and $S_{\varepsilon}^-$ as $\varepsilon$ tends to $0$ exist in the sense
of currents, and the limits are independent of the sequences $\{ h_\varepsilon \}$ and $\{ T_\varepsilon \}$.
\end{lma}
\begin{proof}
By \eqref{eq:ChernSegre}, we get that $T_\varepsilon \wedge c_k(E,h_\varepsilon)$
can be written as a finite sum of terms
\begin{equation} \label{eq:induction_terms}
(-1)^k a_K T_\varepsilon \wedge s_{k_1}(E,h_\varepsilon) \wedge \dots \wedge s_{k_m}(E,h_\varepsilon),
\end{equation}
where $K = (k_1,\dots,k_m) \in \mathbb{N}^m$ and $k_1 + \dots + k_m = k$, and $a_K$ is an integer.
We let $S_{\varepsilon}^+$ be the sum of all terms \eqref{eq:induction_terms} with $a_K$ positive,
and $S_{\varepsilon}^-$ the sum of the remaining terms.
In order to prove that $S_{\varepsilon}^+$ and $S_{\varepsilon}^-$ have limits, which are independent of
the sequences, it is thus enough to prove this for each term \eqref{eq:induction_terms}. The proof of this
is by induction over $m$, and is essentially the same as the proof of local existence and uniqueness in
the proof of Proposition~\ref{prop:segre_products}, only that the basic case $m=0$ here becomes the fact that
$T_\varepsilon$ tends to $T$, and when applying Lemma~\ref{lma:locallyBounded} and Lemma~\ref{lma:UniqueLim},
assumption (i) is used instead of assumption (ii).
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:ContThm}]
We want to prove that the limit of
\begin{equation} \label{eq:chern_regularization}
c_{k_1}(E_1,h^1_\varepsilon) \wedge \dots \wedge c_{k_m}(E_m,h^m_\varepsilon)
\end{equation}
as $\varepsilon$ tends to $0$ exists, and is independent of the regularizations $h^1_\varepsilon,\dots,h^m_\varepsilon$.
In order to prove this, we will show by induction over $m$ that
\begin{equation} \label{eq:induction_cont}
c_{k_1}(E_1,h^1_\varepsilon) \wedge \dots \wedge c_{k_{m}}(E_{m},h^{m}_\varepsilon) = S_{m,\varepsilon}^+ - S_{m,\varepsilon}^-,
\end{equation}
where $S_{m,\varepsilon}^+$ and $S_{m,\varepsilon}^-$ are smooth closed positive forms, which have limits as currents as $\varepsilon$ tends to $0$,
independent of the regularizing sequences of the metrics on $E_1,\dots,E_{m}$. The basic case $m = 0$ is trivial when we let $S_{0,\varepsilon}^+ = 1$
and $S_{0,\varepsilon}^- = 0$. Thus by induction, we assume that such $S_{m-1,\varepsilon}^+$ and $S_{m-1,\varepsilon}^-$ exist.
Applying Lemma~\ref{lma:UniqueLimCh} with $T_\varepsilon$ either $S_{m-1,\varepsilon}^+$ or $S_{m-1,\varepsilon}^-$,
and the Chern form $c_{k_m}(E_m,h^m_\varepsilon)$, we get in total four terms, two of which together define $S_{m,\varepsilon}^+$
and two which define $S_{m,\varepsilon}^-$, and by Lemma~\ref{lma:UniqueLimCh}, these terms have limits
as $\varepsilon$ tends to $0$, and which are independent of the regularizing sequences.
\end{proof}
\section{Cohomology Check}
\noindent
Locally, a positive singular hermitian metric $h$ on a vector bundle can be approximated by positive smooth metrics (see Lemma \ref{prop:appr}). On the projectivization of the vector bundle, this induces an approximation of the positive singular hermitian metric $e^{-\varphi}\newcommand{\eps}{\varepsilon}$ induced by $h$
with smooth positive metrics $e^{-\varphi}\newcommand{\eps}{\varepsilon_\eps}$ over preimages of small enough open sets.
However, to prove the cohomology result (Theorem \ref{thm:Cohomology}), we need a global approximation of $e^{-\varphi}\newcommand{\eps}{\varepsilon}$. In general, we do not have a global smooth approximation with positive metrics.
Instead, we have the following global regularization with smooth uniformly quasi-positive metrics (i.e.\ bounded from below by something negative but smooth) by Demailly:
\begin{thm}[Main Theorem 1.1 in \cite{DemaillyReg}]\label{thm:Demailly-regularization}
Let $M$ be a compact hermitian manifold, and
let $e^{-\varphi}\newcommand{\eps}{\varepsilon}$ be a singular positive metric on a line bundle $L$ over $M$.
Then, there exist a sequence of smooth metrics $\{e^{-\tilde\ph_\eps}\}$ increasing to $e^{-\varphi}\newcommand{\eps}{\varepsilon}$
and a smooth positive real $(1,1)$-form $u$ such that
\[dd^c\tilde\ph_\eps +u\geq 0.\]
\end{thm}
Actually, Demailly proves a much more general result but the version above is sufficient for our purpose. For the reader's convenience, let us show how the statement from above is deduced from Demailly's theorem.
\begin{proof
Let us use the same notation as in \cite{DemaillyReg}: We set $T:=dd^c \varphi}\newcommand{\eps}{\varepsilon$ which is positive by assumption.
For a smooth metric $e^{-\varphi}\newcommand{\eps}{\varepsilon_0}$ on $L$, the smooth real $(1,1)$-form $\alpha:=dd^c\varphi}\newcommand{\eps}{\varepsilon_0$ is in the same $dd^c$-class as $T$. Furthermore, $T=\alpha+dd^c\psi$ for the quasi-plurisubharmonic function $\psi:=\varphi}\newcommand{\eps}{\varepsilon-\varphi}\newcommand{\eps}{\varepsilon_0$.
Let $g$ be a smooth hermitian metric on the tangent bundle $TM$, and let $e^{-\sigma}$ denote the associated metric on ${\mathcal O}_{{\mathbb P}(TM)}(1)$ over the projectivization of $TM$.
Although $g$ does not need to be positive, we know that $\sigma$ is positive with respect to tangent vectors along fibres of $\pi\colon{\mathbb P}(TM)\rightarrow M$. In fact,
if $\Theta:=\Theta(TM,g)$ denotes the curvature matrix of $g$, then
\[dd^c \sigma(p,[v])=-\sfrac i{2\pi\cdot \|v\|^2_g} \langle\Theta v,v\rangle_g + \omega_{\textrm{FS},[g(p)]}\]
for $(p,[v])\in{\mathbb P}(TM)$ (see for instance \cite{G}*{Section 2}).
Let $\{U_j\}$ be an open Stein covering of $M$ such that $TM$ is trivial on $U_j$, and let $\chi_j$ be cut-off functions on $X$ with support in $U_j$ such that $\chi_j=1$ on slightly smaller open sets $\tilde U_j\subset U_j$. We may assume that $\{\tilde U_j\}$ still covers $M$.
On $U_j$, there exist strictly plurisubharmonic functions $a_j$ such that $\frac i\pi\langle\Theta v,v\rangle_g +dd^c a_j\geq 0$ for all $v\in TM$ with $\|v\|_g=1$. For $u:=\sum_j \chi_j dd^c a_j$, we get
\[dd^c\sigma+u \geq -\sfrac i{2\pi\cdot \|v\|^2_g} \langle\Theta v,v\rangle_g +\omega_{\textrm{FS},[g(p)]} + dd^c\pi^\ast a_l\geq 0\]
on $\tilde U_l$ for every $l$.
Furthermore, we may assume that $u\geq\omega$ where $\omega$ denotes the hermitian form associated to $g$.
We conclude that all the assumptions of Main Theorem 1.1 in \cite{DemaillyReg} are satisfied.
That means that for $c>\max_{x\in M} \nu(T,x)$, there exist smooth quasi-plurisubharmonic functions $\psi_\eps:=\psi_{c,\eps}$ decreasing to $\psi$ such that
\[T_\eps:= \alpha+dd^c\psi_\eps \geq -c u- \omega.\]
Replacing $u$ by $(c+1) u$, we obtain that $T_\eps\geq -u$.
In particular, the smooth quasi-positive metrics $\tilde\ph_\eps:=\varphi}\newcommand{\eps}{\varepsilon_0+\psi_\eps$ on $L$ satisfy the claimed: $dd^c\tilde\ph_\eps=\alpha+dd^c\psi_\eps =T_\eps\geq-u$ and $\tilde\ph_\eps$ decreases pointwise to $\varphi}\newcommand{\eps}{\varepsilon_0+\psi=\varphi}\newcommand{\eps}{\varepsilon$.
\end{proof}
We would like to use Demailly's regularization although the approximating sequence $\{\tilde\ph_\eps\}$ is only quasi-positive. For this, we need to consider slightly more general results than in section \ref{sec:Segre_currents} which are given by the following two propositions. First, we consider a variation of Proposition \ref{prop:segre_products}:
\begin{prop} \label{prop:quasi_segre_products_1}
Let $E\to P$ be a trivial holomorphic vector bundle of rank $r$ over a polydisc $P\subset{\mathbb C}^n$, and
let $e^{-\psi}$ denote a singular positive metric on a holomorphic line bundle $L$ over ${\mathbb P}(E)$.
Let $\alpha$ be a closed positive $(1,1)$-form on ${\mathbb P}(E)$, and
let $T$ be a closed positive $(q,q)$-current on $P$.
If $\pi(L(\psi))$ is contained in some subvariety $V\subset P$ with $\codim(V)\geq k+q$, then for every $j=0{,}...,k+r-1$,
there is a well-defined closed positive current
\begin{equation} \label{eq:alpha_segre_product}
T\wedge\pi_\ast (\alpha^j\wedge(dd^c\psi)^{k+r-1-j}),
\end{equation}
which is defined as the limit of
\begin{equation*}
T\wedge\pi_\ast (\alpha^j\wedge(dd^c\psi_\eps)^{k+r-1-j}),
\end{equation*}
as $\eps$ tends to $0$, for any sequence $\{ e^{-\psi_\eps} \}$ of smooth positive metrics on $L$ increasing pointwise to $e^{-\psi}$.
\end{prop}
\begin{proof}
Replacing the form $s_k(E,h_\eps)$ by $\pi_\ast(\alpha^j\wedge(dd^c \varphi}\newcommand{\eps}{\varepsilon_\eps)^{k+r-1-j})$, we see that the analogous statements of Lemma \ref{lma:locallyBounded} and Lemma \ref{lma:UniqueLim} (under the assumptions (ii)) hold true.
Since $\alpha^j\wedge (dd^c\psi_\eps)^{k+r-1-j}$ are strongly positive,
we get that
\[T\wedge\pi_\ast (\alpha^j\wedge (dd^c\psi_\eps)^{k+r-1-j})=\pi_\ast\big(\pi^\ast T\wedge\alpha^j\wedge (dd^c\psi_\eps)^{k+r-1-j}\big)\]
are positive currents for all $\eps>0$.
Hence, the result follows from the statement in Remark~\ref{remark:wedge_product_of_currents} since the conditions (i--iii) in the remark are all satisfied.
\end{proof}
\begin{prop}\label{prop:quasi_segre_products_2}
Let $E\to P$ be a trivial holomorphic vector bundle of rank $r$ over a polydisc $P\subset{\mathbb C}^n$, and let $e^{-\varphi}\newcommand{\eps}{\varepsilon}$ denote a singular positive hermitian metric on ${\mathcal O}_{{\mathbb P}(E)}(1)$ induced by a positive hermitian metric on $E$.
Moreover, let $\alpha=dd^c a$, where $e^{-a}$ is a smooth positive metric on a holomorphic line bundle $L$ over ${\mathbb P}(E)$, and
let $T$ be a closed positive $(q,q)$-current on $P$.
Assume that $\pi(L(\varphi}\newcommand{\eps}{\varepsilon))$ is contained in some subvariety $V\subset P$ such that $\codim(V)\geq k+q$.
If $\{e^{-\tilde\varphi}\newcommand{\eps}{\varepsilon_\eps}\}$ is a sequence of smooth metrics increasing pointwise to $e^{-\varphi}\newcommand{\eps}{\varepsilon}$ and $dd^c\tilde\varphi}\newcommand{\eps}{\varepsilon_\eps+\alpha\geq 0$,
then the limit of
\[T\wedge\pi_\ast ((dd^c\tilde\varphi}\newcommand{\eps}{\varepsilon_\eps)^{k+r-1})\]
as $\eps$ tends to 0 exists and equals $T\wedge\pi_\ast ((dd^c\varphi}\newcommand{\eps}{\varepsilon)^{k+r-1})$, as defined in \eqref{eq:alpha_segre_product}.
\end{prop}
\begin{proof} Let $l:=k+r-1$.
Let $\{ e^{-\varphi}\newcommand{\eps}{\varepsilon_\eps} \}$ be a sequence of smooth positive metrics on ${\mathcal O}_{{\mathbb P}(E)}(1)$ increasing pointwise to $e^{-\varphi}\newcommand{\eps}{\varepsilon}$, which exists since $\varphi}\newcommand{\eps}{\varepsilon$ was induced by a positive metric on $E$, see Lemma \ref{prop:appr}.
Proposition \ref{prop:quasi_segre_products_1} (for $j=0$) then implies
\begin{equation}\label{eq:quasi-stuff_definition}
T\wedge\pi_\ast ((dd^c\varphi}\newcommand{\eps}{\varepsilon)^{l}) = \lim_{\eps\rightarrow 0} T\wedge\pi_\ast ((dd^c\varphi}\newcommand{\eps}{\varepsilon_\eps)^{l}).
\end{equation}
The proposition claims that the same holds when $\varphi_\eps$ is replaced by $\tilde\ph_\eps$. To prove this, we account for the lack of positivity of
$\tilde\ph_\eps$ by twisting:
We define $\psi:=\varphi}\newcommand{\eps}{\varepsilon+a$, $\psi_\eps:=\tilde\ph_\eps+a$ which give singular respectively smooth positive metrics $e^{-\psi}$ and $e^{-\psi_\eps}$ on ${\mathcal O}_{{\mathbb P}(E)}(1)\otimes L$.
By assumption, $\psi_\eps$ decreases pointwise to $\psi$.
Thus, Proposition \ref{prop:quasi_segre_products_1} gives
\begin{equation}\label{eq:quasi-stuff_first_approximation}
T\wedge\pi_\ast (\alpha^j\wedge(dd^c\psi)^{l-j}) = \lim_{\eps\rightarrow 0} T\wedge\pi_\ast (\alpha^j\wedge(dd^c\psi_\eps)^{l-j}).
\end{equation}
On the other hand, $\{\varphi}\newcommand{\eps}{\varepsilon_\eps+a\}$ is another sequence of positive smooth metrics on ${\mathcal O}_{{\mathbb P}(E)}(1)\otimes L$
which decreases to $\varphi}\newcommand{\eps}{\varepsilon+a=\psi$. By Proposition \ref{prop:quasi_segre_products_1}, we get again:
\begin{equation}\label{eq:quasi-stuff_second_approximation}
T\wedge \pi_\ast (\alpha^j\wedge(dd^c\psi)^{l-j}) = \lim_{\eps\rightarrow 0} T\wedge\pi_\ast (\alpha^j\wedge(dd^c(\varphi}\newcommand{\eps}{\varepsilon_\eps+a))^{l-j}).
\end{equation}
Since
\[(dd^c\varphi}\newcommand{\eps}{\varepsilon_\eps)^l=(dd^c(\varphi}\newcommand{\eps}{\varepsilon_\eps+a)-\alpha)^l=\sum{\Big.\!}_{j} \smath{\binom{l}{j}} (-\alpha)^j \wedge (dd^c(\varphi}\newcommand{\eps}{\varepsilon_\eps+a))^{l-j},\]
we obtain: \newcommand{\widereq}{\makebox[2.2em]{=}}%
\begin{equation*}\begin{split}
T\wedge\pi_\ast ((dd^c\varphi}\newcommand{\eps}{\varepsilon)^{l})
&\overset{\eqref{eq:quasi-stuff_definition}}{\widereq} \lim_{\eps\rightarrow 0} T\wedge\pi_\ast ((dd^c\varphi}\newcommand{\eps}{\varepsilon_\eps)^{l})\\
&\widereq \sum{\Big.\!}_{j} \smath{\binom{l}{j}} \lim_{\eps\rightarrow 0} T\wedge\pi_\ast \big((-\alpha)^j \wedge (dd^c(\varphi}\newcommand{\eps}{\varepsilon_\eps+a))^{l-j}\big)\\
&\overset{\eqref{eq:quasi-stuff_second_approximation}}{\widereq} \sum{\Big.\!}_{j} \smath{\binom{l}{j}}\ T\wedge\pi_\ast \big((-\alpha)^j \wedge (dd^c\psi)^{l-j}\big)\\
&\overset{\eqref{eq:quasi-stuff_first_approximation}}{\widereq} \sum{\Big.\!}_{j} \smath{\binom{l}{j}} \lim_{\eps\rightarrow 0} T\wedge\pi_\ast \big((-\alpha)^j \wedge (dd^c\psi_\eps)^{l-j}\big)\\
&\widereq \lim_{\eps\rightarrow 0} T\wedge\pi_\ast ((dd^c\psi_\eps-\alpha)^{l}) \widereq \lim_{\eps\rightarrow 0} T\wedge\pi_\ast ((dd^c\tilde\ph_\eps)^{l}).
\qedhere
\end{split}\end{equation*}%
\end{proof}
\medskip
\begin{proof}[Proof of Theorem~\ref{thm:Cohomology}]
The metric $e^{-\varphi}$ on ${\mathcal O}_{{\mathbb P}(E)}(1)$, induced by the singular metric $h$ on $E$, is positive by Lemma~\ref{lma:inducedPositive}.
Therefore, we can apply Theorem \ref{thm:Demailly-regularization} and obtain a smooth positive $(1,1)$ form $u$ and
a smooth sequence of metrics $\{ e^{-\tilde\ph_\eps} \}$ increasing pointwise to $e^{-\varphi}\newcommand{\eps}{\varepsilon}$ on ${\mathbb P}(E)$ such that $dd^c\tilde\ph_\eps+u\geq 0$ for all $\eps>0$.
We set $T^\varepsilon_k:=(-1)^k\pi_*((dd^c\tilde\varphi_\varepsilon)^{k+r-1})$.
For all small enough open sets $U$ of $X$,
there is a closed positive $(1,1)$-form $\alpha\geq u$ on $\pi^{-1}(U)$ such that $\alpha=dd^c a$ for a smooth positive metric $a$ on ${\mathcal O}_{{\mathbb P}(E|_U)}(\mu)$, $\mu\in{\mathbb N}$.
For instance, let $a$ be a $\mu$-fold tensor power (with $\mu$ big enough) of a (strictly) positive metric
on ${\mathcal O}_{{\mathbb P}(E|_U)}(1)$ which is induced by a smooth positive metric on $E|_U$.
By an iterated application of Proposition \ref{prop:quasi_segre_products_2}, we obtain that
\begin{equation}\begin{split}\label{eq:weak_converges_of_T_eps}
\lim_{\varepsilon_m\to0}\cdots\lim_{\varepsilon_1\to0}T^{\varepsilon_1}_{k_1}\wedge\cdots\wedge T^{\varepsilon_m}_{k_m}
&= (-1)^{k}\pi_\ast (dd^c\varphi}\newcommand{\eps}{\varepsilon)^{k_1+r-1}\wedge\cdots\wedge\pi_\ast (dd^c\varphi}\newcommand{\eps}{\varepsilon)^{k_m+r-1}\\
&=s_{k_1}(E,h)\wedge\cdots\wedge s_{k_m}(E,h)
\end{split}\end{equation}
for any positive integers $k_1,\ldots,k_m$ with $k_1+\cdots+k_m=k$.
Note that even though we apply Proposition~\ref{prop:quasi_segre_products_2} locally, \eqref{eq:weak_converges_of_T_eps} holds globally
on $X$ since all the currents are globally defined.
\smallskip
For an arbitrary smooth metric $h_0$ on $E$, let $e^{-\varphi_0}$ denote the induced metric on ${\mathcal O}_{{\mathbb P}(E)}(1)$. The Chern classes being invariants of ${\mathcal O}_{{\mathbb P}(E)}(1)$ implies that
$$[dd^c\tilde\varphi_\varepsilon]=[dd^c\varphi_0]$$
which yields
$$[(dd^c\tilde\varphi_\varepsilon)^{j+r-1}]=[(dd^c\varphi_0)^{j+r-1}]$$
for all integers $j$.
Since the push-forward $\pi_*$ and exterior differentiation commute, this combined with Proposition \ref{prop:Segre} gives
$$[T^\varepsilon_j]=[s_j(E,h_0)]=s_j(E).$$
More generally, this exact same argument yields that
\begin{equation}\label{eq:Smooth_Cohomology}
[T^{\varepsilon_1}_{k_1}\wedge\cdots\wedge T^{\varepsilon_m}_{k_m}]=[s_{k_1}(E,h_0)\wedge\cdots\wedge s_{k_m}(E,h_0)].
\end{equation}
It remains to show that this implies that
$$[s_{k_1}(E,h)\wedge\cdots\wedge s_{k_m}(E,h)]=[s_{k_1}(E,h_0)\wedge\cdots\wedge s_{k_m}(E,h_0)].$$
This, however, is a straightforward consequence of \eqref{eq:weak_converges_of_T_eps}, \eqref{eq:Smooth_Cohomology} and Poincar\'e duality,
which yields that the exactness of $s_{k_1}(E,h)\wedge\cdots\wedge s_{k_m}(E,h)-s_{k_1}(E,h_0)\wedge\cdots\wedge s_{k_m}(E,h_0)$ is equivalent to showing that
$$\int_X\big(s_{k_1}(E,h)\wedge\cdots\wedge s_{k_m}(E,h)-s_{k_1}(E,h_0)\wedge\cdots\wedge s_{k_m}(E,h_0)\big)\wedge\beta=0$$
for all smooth, closed $2(n-k)$-forms $\beta$.
\end{proof}
\begin{bibdiv}
\begin{biblist}
\bib{BP}{article}{
author={Berndtsson, Bo},
author={P{\u{a}}un, Mihai},
title={Bergman kernels and the pseudoeffectivity of relative canonical
bundles},
journal={Duke Math. J.},
volume={145},
date={2008},
number={2},
pages={341--378},
}
\bib{BT}{article}{
author={Bedford, Eric},
author={Taylor, B. A.},
title={A new capacity for plurisubharmonic functions},
journal={Acta Math.},
volume={149},
date={1982},
number={1-2},
pages={1--40},
}
\bib{D1}{book}{
author={Demailly, Jean-Pierre},
title={Analytic methods in algebraic geometry},
series={Surveys of Modern Mathematics},
volume={1},
publisher={International Press, Somerville, MA; Higher Education Press,
Beijing},
date={2012},
pages={viii+231},
}
\bib{D2}{article}{
author={Demailly, Jean-Pierre},
title={Complex Analytic and Differential Geometry},
status={Monograph},
Year={2012},
eprint={http://www-fourier.ujf-grenoble.fr/~demailly},
}
\bib{DemaillyReg}{article}{
author={Demailly, Jean-Pierre},
TITLE = {Regularization of closed positive currents and intersection theory},
JOURNAL = {J. Algebraic Geom.},
VOLUME = {1},
YEAR = {1992},
NUMBER = {3},
PAGES = {361--409},
ISSN = {1056-3911},
}
\bib{Dem4}{article}{
author={Demailly, Jean-Pierre},
title={Singular Hermitian metrics on positive line bundles},
conference={
title={Complex algebraic varieties},
address={Bayreuth},
date={1990},
},
book={
series={Lecture Notes in Math.},
volume={1507},
publisher={Springer, Berlin},
},
date={1992},
pages={87--104},
}
\bib{Diverio}{article}{
author={Diverio, Simone},
title={Segre forms and Kobayashi--L\"ubke inequality},
journal={Math. Z.},
volume={283},
date={2016},
number={3-4},
pages={1033--1047},
}
\bib{GrPos}{article}{
author={Griffiths, Phillip A.},
title={Hermitian differential geometry, Chern classes, and positive
vector bundles},
conference={
title={Global Analysis (Papers in Honor of K. Kodaira)},
},
book={
publisher={Univ. Tokyo Press, Tokyo},
},
date={1969},
pages={185--251},
}
\bib{G}{article}{
author={Guler, Dincer},
title={On Segre forms of positive vector bundles},
journal={Canad. Math. Bull.},
volume={55},
date={2012},
number={1},
pages={108--113},
}
\bib{Hos}{article}{
author={Hosono, Genki},
title={Approximations and examples of singular Hermitian metrics on vector bundles},
status={Preprint},
eprint={arXiv:1505.02396 [math.CV]},
url={http://arxiv.org/abs/1505.02396}
}
\bib{Mou}{article}{
author={Mourougane, Christophe},
title={Computations of Bott-Chern classes on ${{\mathbb P}}(E)$},
journal={Duke Math. J.},
volume={124},
date={2004},
number={2},
pages={389--420},
}
\bib{Mum}{article}{
AUTHOR = {Mumford, David},
TITLE = {Hirzebruch's proportionality theorem in the noncompact case},
JOURNAL = {Invent. Math.},
VOLUME = {42},
YEAR = {1977},
PAGES = {239--272},
ISSN = {0020-9910},
DOI = {10.1007/BF01389790},
URL = {http://dx.doi.org/10.1007/BF01389790},
}
\bib{PT}{article}{
author={P{\u{a}}un, Mihai},
author={Takayama, Shigeharu},
title={Positivity of twisted relative pluricanonical bundles and their direct images},
date={2014},
status={Preprint},
eprint={arXiv:1409.5504 [math.AG]},
url={http://arxiv.org/abs/1409.5504}
}
\bib{R}{article}{
author={Raufi, Hossein},
title={Singular hermitian metrics on holomorphic vector bundles},
journal={Ark. Mat.},
volume={53},
date={2015},
number={2},
pages={359--382},
}
\bib{Z}{book}{
author={Zheng, Fangyang},
title={Complex differential geometry},
series={AMS/IP Studies in Advanced Mathematics},
volume={18},
publisher={American Mathematical Society, Providence, RI; International
Press, Boston, MA},
date={2000},
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction}
\label{secIntroduction}
Model selection is a key problem in statistics and data mining, and the MDL approaches \citep{Rissanen78} to model selection have been extensively studied in the literature \citep{Grunwald07}, with successful applications in many practical problems.
Simple models such as Bernoulli or mainly multinomial distributions are important because they are easier to analyze theoretically and useful in many applications.
For example, the multinomial distribution has been used as a building block in more complex models, such as naive Bayes classifiers \citep{MononenEtAl07}, Bayesian networks \citep{RoosEtAl08}, decision trees \citep{VoisineEtAlAKDM09} or coclustering models \citep{BoulleHOPR10,GuigouresEtAlECML15}. These models involve up to thousands of multinomials blocks, some of them with potentially very large numbers of occurrences and outcomes. For example, the text $\times$ word coclustering of the 20-newsgroup dataset described in \citep{BoulleHOPR10} exploits a main multinomial block with around two millions words (occurrences) distributed on 200,000 coclusters (outcomes). In \citep{GuigouresEtAlECML15}, half a billion call detail records (occurrences) are distributed on one million coclusters (outcomes).
These various and numerous applications critically rely on the use of effective and efficient MDL code lengths to get a robust and accurate summary of the data.
The MDL approaches come with several flavors, ranging from theoretical but not computable to practical but sub-optimal.
Ideal MDL \citep{VitanyiEtAl00} relies on the Kolmogorov complexity, that is the ability of compressing data using a computer program. However, it suffers from large constants depending on the description method used and cannot be computed, not even approximated in the case of two-part codes \citep{AdriaansEtAl07}.
Practical MDL leverages description methods that are less expressive than general-purpose computer languages.
It has been employed to retrieve the best model given the data in case of families of parametrized statistical distributions.
Crude MDL is a basic MDL approach with appealing simplicity.
In two-part crude MDL, you just have to encode the model parameters and the data given the parameter, with a focus on the code length only. However, crude MDL suffers from arbitrary coding choices.
Modern MDL relies on universal coding resulting in Refined MDL \citep{Grunwald07}, with much stronger foundations and interesting theoretical properties.
In this paper, we investigate the enumerative two-part crude MDL code for the Bernoulli and multinomial models, exhibit a strong connection with the NML approach, with surprising impacts on the estimation of the model complexity and superior compression performance.
The rest of the paper is organized as follows.
For self-containment reasons, Section~\ref{secBernoulli} presents standard codes for the Bernoulli distribution: one simplistic two-part crude MDL code as well as a refined MDL code based on the NML approach.
Section~\ref{secRevisitedEnumerative} describes a particular two-part crude MDL code based on enumerations and establishes the connection of its parameter coding length with its NML parametric complexity.
Section~\ref{secComparison} proceeds with a deep comparison between this enumerative MDL code and the NML code presented in Section~\ref{secBernoulli}.
Section~\ref{secMultinomial} suggests an extension of the enumerative two-part crude MDL code to multinomial distributions and Section~\ref{secComparisonM} compares this code with the alternative NML code.
Finally, Section~\ref{secConclusion} summarizes this paper.
\section{Standard MDL codes for Bernoulli strings}
\label{secBernoulli}
We briefly present one simplistic example of two-part crude MDL code for encoding binary strings using the Bernoulli model, as well as a modern MDL code based on NML. This has been presented many times in the literature, e.g. \citep{HansenEtAl01b, Grunwald07, RooijEtAl09}.
Let us consider the Bernoulli model with $\theta \in [0, 1]$ in the case of binary sequences $x^n \in X^n$ of size $n$. Let $k(x^n)$ be the number of ones in $x^n$.
\subsection{Simplistic two-part crude MDL approach}
\label{secSimplistic}
Using a two-part version of the MDL principle \citep{Grunwald07}, we select the best hypothesis $H$ that explains the data $D$ by minimizing the sum $L(H) + L(D|H)$, where $L(H)$ is the coding length of the hypothesis and $L(D|H)$ is the coding length of the data encoded with the help of the hypothesis.
In the case of the Bernoulli model, we have to encode the parameter $\theta$ and the data $x^n$ given $\theta$.
The number of ones in the binary string $x^n$ is between 0 and $n$.
The $\theta$ parameter can thus be chosen among $(n+1)$ values $\theta=\frac {0} {n}, \frac {1} {n}, \frac {2} {n}, \ldots, \frac {n} {n}$, and be encoded using $L(\theta) = \log (n+1)$ bits.
For $\theta \in \{0, 1 \}$, the string $x^n$ is degenerated with only zeros or ones, and its coding length given $\theta$ is $L(x^n|\theta) = 0$.
For $\theta = \frac {k} {n},\; 0 < k < n$, every symbol of the string $x^n$ can be encoded using $-\log \frac {k} {n}$ bit for a one and $-\log \frac {n-k} {n}$ bit for a zero, leading to $L(x^n|\theta) = -k \ln \frac {k} {n} - (n-k) \ln \frac {n-k} {n}$.
This gives a total code length of
\begin{equation}
L(\theta = \frac {k} {n}, x^n) = \log(n+1) + \left(-k \log \frac {k} {n} - (n-k) \log \frac {n-k} {n}\right).
\end{equation}
Equivalently, the likelihood of the whole string $x^n$ can be estimated as $P(x^n|\theta = \frac {k} {n}) = (\frac {k} {n})^k (\frac {n-k} {n})^{n-k}$, with $L(x^n|\theta) = - \log P(x^n|\theta = \frac {k} {n})$.
Using the Shannon entropy $H(\frac {k} {n}) = -\frac {k} {n} \log (\frac {k} {n}) - \frac {n-k} {n} \log (\frac {n-k} {n})$, we also have $L(x^n|\theta) = n H(\theta)$.
\subsection{Standard NML Approach}
\label{secNML}
The simplistic two-part MDL code defined previously suffers from some arbitrary choices and may be suboptimal at best, with arbitrary bad behavior for small sample sizes \citep{Grunwald07}.
In the case of the Bernoulli model, this is pointed out in \citep{RooijEtAl09},
\begin{quotation}
``Example 5. $\ldots$
A uniform code uses $L(\theta) = \log(n + 1)$ bits to identify an element of this set. Therefore the resulting regret is always exactly $\log(n + 1)$. By using a slightly more clever discretisation we can bring this regret down to about $\frac {1} {2} \log n + O(1)$, which we mentioned is usually achievable for uncountable single parameter models.''
\end{quotation}
Using universal coding, a much more grounded approach is proposed to better evaluate the model complexity, based on the Shtarkov NML code, which provides strong theoretical guarantees \citep{Rissanen00}.
It exploits the following NML distribution $\overline{P}_{nml}^{(n)}$ on $X^n$:
\begin{equation}
\label{eqnNML}
\overline{P}_{nml}^{(n)}(x^n) = \frac { P_{\widehat{\theta}(x^n)}(x^n)}
{\sum_{y^n \in X^n} {P_{\widehat{\theta}(y^n)}(y^n)}}
\end{equation}
where $\widehat{\theta}(x^n)$ is the model parameter that maximizes the likelihood of $x^n$.
The log of the denominator stands for the \textit{parametric complexity} $COMP^{(n)}(\theta)$ of the model whereas the negative log of the numerator is the \textit{stochastic complexity} of the data given the model.
The sum of both terms provides the NML code. It is noteworthy that the NML code is a one-part rather than two-part code: data is encoded with the help of all the model hypotheses rather than the best hypothesis.
\medskip
In the case of the Bernoulli model, $\widehat{\theta}(x^n) = k(x^n)/n$.
We have
\begin{eqnarray}
COMP^{(n)}(\theta) &=& \log \sum_{y^n \in X^n} {P_{\widehat{\theta}(y^n)}(y^n)}, \\
&=& \log \sum_{k=0}^n {{{n}\choose{k}} \left(\frac {k}{n}\right)^k \left(\frac {n-k}{n}\right)^{n-k}}.
\end{eqnarray}
Using the Stirling' formula together with the Fisher information provides the following accurate approximation \citep{Rissanen96}:
\begin{eqnarray}
COMP^{(n)}(\theta) &=& \frac{1}{2} \log \frac{n}{2 \pi} + \int_{\theta} {\sqrt{ det I(\theta)}} +o(1),\\
&=& \frac{1}{2} \log \frac{n \pi}{2} + o(1).
\end{eqnarray}
Remarkably, this is in line with the classical BIC regularization term $\frac{1}{2}\log n$.
\section{Revisiting enumerative two-part crude MDL}
\label{secRevisitedEnumerative}
We present the enumerative two-part crude MDL code for Bernoulli distributions, suggest a finite data sample Bayesian interpretation and show a connection with the NML approach.
\subsection{Enumerative two-part crude MDL}
\label{secEnumerativeB}
We present an alternative type of two-part crude MDL code for Bernoulli distributions. It has already been proposed in the past literature, under the names of \emph{index} or \emph{enumerative} code (see for example \citet{Grunwald07} Example~10.1 \emph{Coding by Giving an Index}).
First, like in Section~\ref{secSimplistic}, we enumerate all possible $\theta = \frac{i}{n}$ parameter values given the sample size $n$. We then use $\log (n+1)$ bits to encode $\theta$.
Second, given $\widehat{\theta}(x^n) = \frac{k(x^n)}{n}$, we enumerate all the ${{n}\choose{k}}$ binary sequences with $k$ ones and encode the data $x^n$ using $\log {{n}\choose{k}}$ bits.
This gives a total code length of
\begin{equation}
L(\widehat{\theta}(x^n), x^n) = \log(n+1) + \log \frac{n!} {k! (n-k)!},
\end{equation}
Interestingly, this crude MDL approach results in the same code length as that obtained in \citep{HansenEtAl01b} using \emph{Predictive Coding} or \emph{Mixture Coding} with a uniform prior.
For the case of the Bayes mixture model with uniform prior $w(\theta) = 1,\; \theta \in [0,1]$,
we have
\begin{eqnarray}
P_{Bayes}(x^n) &=& \int_0^1{w(\theta) P_{\theta}(x^n)d\theta}, \\
&=& \int_0^1{\theta^{k} (1-\theta)^{n-k} d\theta},\\
&=& \frac{1}{n+1} \frac{k! (n-k)!} {n!} .
\end{eqnarray}
The negative log of $P_{Bayes}(x^n)$ actually corresponds to the code length of the enumerative code.
This code has also been studied by \citet{Grunwald07} (Chapter~10, Section~10.2) under the name \emph{Conditional Two-Part Universal Code}, which suggests that at least for the Bernoulli model, this code is strictly preferable to the ordinary two-part code.
\subsection{Bayesian interpretation}
\label{secEnumBayesian}
Let $\mathcal{M} = \{ P_\theta \, | \, \theta \in [0,1] \}$ be the class of all Bernoulli distributions.
We propose to focus on the family of models
$\mathcal{M}^{(n)} = \{P_\theta \,|\, \theta = \frac{i}{n},\; 0 \leq i \leq n\}$ that are models of description for finite size data samples.
$\mathcal{M}^{(n)}$ is related to the set of all the possible maximum likelihood estimates of $\theta$ (from $\mathcal{M}$) for binary strings of size $n$.
The interest of using $\mathcal{M}^{(n)}$ is that the number of model parameters is now finite instead of uncountable infinite.
Using a uniform prior on the model parameters in $\mathcal{M}^{(n)}$, we get $P(\theta = \frac{i}{n}) = 1/{|\mathcal{M}^{(n)}|}$, leading to $L(\theta) = \log (n+1)$.
Given $\theta = \frac{i}{n} \in \mathcal{M}^{(n)}$, we now have to encode the data $x^n$.
If $k(x^n)/n \neq \theta$, we cannot encode the data and $P(x^n|\theta) = 0$.
If $k(x^n)/n = \theta$, the observed data is consistent with the model parameter, and we assume that all the possible observable data are uniformly distributed.
The number of binary strings with $k$ ones is the binomial coefficient ${{n}\choose{k}}$. Thus the probability of observing one of them is $P(x^n|\widehat{\theta}(x^n)) = 1/{{n}\choose{k}}$.
We have a discrete likelihood that concentrates the probability mass on binary strings that can be observed given the model parameter. As a result, coding lengths are defined only for strings that are consistent with the model parameter.
This gives a total code length of
\begin{equation}
L(\widehat{\theta}(x^n), x^n) = \log(n+1) + \log \frac{n!} {k! (n-k)!},
\end{equation}
defined only when $\theta = \widehat{\theta}(x^n)$.
\paragraph{Generative model for the enumerative Bernoulli distribution.}
Given a sequence length $n$ and $\theta = \frac{i}{n} \in \mathcal{M}^{(n)}$, we can formulate these models as generative models of sequences with exactly $i$ ones and $n-i$ zeros. For example, from a sequence of $n$ zeros, we randomly choose $i$ times without replacement a zero in the sequence and replace it with a one.
For this generative model, we have the following likelihood, as seen previously:
\begin{equation}
P(x^n|\theta = \frac{i}{n}) = {\mathbb{1}_{\left\{ \frac{i}{n} = \frac{k(x^n)}{n}\right\}}} 1/{{n}\choose{k(x^n)}}.
\end{equation}
For the case of the Bayes mixture model with uniform prior $w(\theta) = \frac{1}{n+1}, \theta = \frac{i}{n},\; 0 \leq i \leq n$,
we have
\begin{eqnarray}
P_{Bayes}(x^n) &=& \sum_{i=0}^n{w(\frac{i}{n}) P(x^n|\theta = \frac{i}{n})}, \\
&=& \frac{1}{n+1} \frac{k(x^n)! (n-k(x^n))!} {n!} .
\end{eqnarray}
The negative log of this probability actually corresponds to the code length of the enumerative code.
Interestingly, the standard Bernoulli model and the enumerative one are related to slightly different generative models, but their Bayes mixture under the uniform prior leads to the same distribution.
In Section~\ref{NMLinterpretation}, we will see that on the opposite, their normalized maximum likelihood distribution is not the same.
\paragraph{Cardinality of models spaces.}
Let us consider the union of the $\mathcal{M}^{(n)}$ models for all the sample sizes:
\begin{equation}
\mathcal{M}^{(\mathbb{N})} = \cup_{n \in \mathbb{N}}{\mathcal{M}^{(n)}}.
\end{equation}
Interestingly, $\mathcal{M}^{(\mathbb{N})}$ is very close to $\mathcal{M}$, with $\theta \in \mathbb{Q}$ rather than $\theta \in \mathbb{R}$. Thus, the number of model parameters in $\mathcal{M}^{(\mathbb{N})}$ is countable infinite rather than uncountable infinite, which provides a significant simplification.
\subsection{NML interpretation}
\label{NMLinterpretation}
Let us compute the NML parametric complexity of this enumerative code, on the basis of the discrete likelihood presented in Section~\ref{secEnumBayesian}.
We have
\begin{eqnarray}
COMP^{(n)}(\theta) &=& \log \sum_{y^n \in X^n} {P_{\widehat{\theta}(y^n)}(y^n)}, \\
&=& \log \sum_{k=0}^n {{{n}\choose{k}} \left({1} /{{n}\choose{k}}\right)},\\
&=& \log (n+1).
\end{eqnarray}
Interestingly, we find exactly the same complexity term $\log (n+1)$ as the coding length of the best hypothesis in the enumerative two-part crude MDL code presented in Section~\ref{secEnumerativeB}.
This shows that the enumerative code is both a two-part and a one-part code. It is parametrization invariant and optimal w.r.t. the NML approach, with minimax regret guarantee.
Surprisingly, its parametric complexity is asymptotically twice that of the NML code described in Section~\ref{secNML}.
We further investigate on the comparison between the enumerative and NML codes in next section
\section{Code comparison for the Bernoulli distribution}
\label{secComparison}
In this section, we compare the NML code (Section~\ref{secNML}) and enumerative two-part crude MDL codes (Section~\ref{secRevisitedEnumerative}) for the Bernoulli distribution.
\subsection{Notation}
Let us use the names \emph{simplistic}, \emph{NML} and \emph{enumerative} for the specific MDL codes presented in Sections \ref{secSimplistic}, \ref{secNML} and \ref{secEnumerativeB}.
We also consider the \emph{random} code as a baseline: it corresponds to a direct encoding of each binary string $x^n$ with a coding length of $n \log 2$. The likelihood of each string $x^n$ is $1/2^n$, and as $\sum_{k=0}^n {{{n}\choose{k}} 1/2^n} = 1$, we have $COMP_{random}^{(n)}(\emptyset) = 0$ and
$L_{random}\left(x^n|\emptyset \right) = n \log 2$.
Table~\ref{tableCodes} reminds the parametric and stochastic complexity of each considered code.
\begin{table}[htbp!]
\caption{Parametric and stochastic complexity per code.}
\label{tableCodes}
\centering
\renewcommand{\arraystretch}{1.5}
\begin{tabular}[10pt]{ccc}\hline
Code name & $COMP_{name}^{(n)}$ & $L_{name}\left(x^n|\widehat{\theta}(x^n) \right)$ \\\hline
\emph{enumerative} & $\log (n+1)$ & $\log \frac {n!} {k! (n-k)!}$ \\
\emph{NML} & $\frac{1}{2} \log \frac{n \pi}{2} + o(1)$ & $\log \frac {n^n} {k^k (n-k)^{n-k}}$ \\
\emph{simplistic} & $\log (n+1)$ & $\log \frac {n^n} {k^k (n-k)^{n-k}}$ \\
\emph{random} & $0$ & $n \log 2$ \\\hline
\end{tabular}
\end{table}
As for the simplistic code, the coding length of the parameter is presented in place of the parametric complexity. The total coding length of the simplistic code has an overhead of about $\frac{1}{2} \log n$ compared to the NML code.
This confirms that the simplistic code is dominated by the NML code, as expected.
\subsection{Stochastic complexity term}
The stochastic complexity term of the enumerative code is always smaller than that of the NML code for non-degenerated binary strings:
\begin{equation}
\label{sc_Bernoulli}
\forall n, \forall x^n \in X^n, \; 0 <k(x^n) < n, \;
L_{enum}\left(x^n|\widehat{\theta}(x^n)\right) < L_{nml}\left(x^n|\widehat{\theta}(x^n)\right).
\end{equation}
An intuitive proof relies on the fact that the enumerative MDL likelihood assigns the same probability to all binary strings having the same number of ones, with a null probability for all the other strings. The NML likelihood also assigns the same probability to all binary strings having the same number of ones, but with a non-null probability for the other strings. Then they have to share a smaller probability mass, resulting in a smaller probability per string and a strictly greater coding length.
To gain further insights, let us approximate the difference of coding length:
$$\delta L\left(x^n|\widehat{\theta}(x^n)\right) = L_{nml}\left(x^n|\widehat{\theta}(x^n)\right) -L_{enum}\left(x^n|\widehat{\theta}(x^n)\right).$$
Using the approximation given in \citep{Grunwald07} (formula~4.36) with the Bernoulli parameter $\theta = \widehat{\theta}(x^n)$ , we have
\begin{eqnarray}
L_{enum}\left(x^n|\theta\right) &=& \log {{n}\choose{\theta n}},\\
&=& n H(\theta) - \log \sqrt{2 \pi n \mathrm{var} (\theta)} + O(1/n),\\
&=& L_{nml}\left(x^n|\theta\right) -\frac{1}{2} \log (2 \pi n \mathrm{var}(\theta)) + O(1/n).
\end{eqnarray}
We get
\begin{equation}
\label{deltaL}
\delta L\left(x^n|\widehat{\theta}(x^n)\right) = \frac{1}{2} \log (2 \pi n \mathrm{var}(\theta)) + O(1/n).
\end{equation}
The difference of coding length is always positive but not uniform:
\begin{itemize}
\item for $k(x^n) = 0$, $\delta L\left(x^n|\widehat{\theta}(x^n)\right) = 0$,
\item for $k(x^n) \approx n/2$, $\delta L\left(x^n|\widehat{\theta}(x^n)\right) \approx \frac{1}{2} \log (\frac {n \pi}{2})$.
\end{itemize}
\medskip
These results demonstrate that the enumerative code provides a better encoding of the data with the help of the model for any binary strings, all the more for strings with equidistributed zeros and ones. The gain in coding length compared to the NML code grows as the logarithm of the sample size.
\subsection{Parametric complexity term}
Using inequality~\ref{sc_Bernoulli} and as the parametric complexity of code is the sum of the stochastic complexity over all possible strings, we get:
\begin{equation}
\label{pc_Bernoulli}
\forall n > 1, COMP_{enum}^{(n)} > COMP_{NML}^{(n)}.
\end{equation}
Both terms are equal for $n=1$ and asymptotically, the parametric complexity of the enumerative code is twice that of the NML code (see Table~\ref{tableCodes}).
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.85\textwidth]{CompareBernoulliCOMP.pdf}
\caption{Parametric complexity for the Bernoulli model.}
\label{fig_comp}
\end{center}
\end{figure}
We now focus on the non-asymptotic behavior of the parametric complexity terms and their approximations.
Figure~\ref{fig_comp} shows the value of the parametric complexity of the Bernoulli model, using the enumerative code, the NML code (exact numerical computation and approximation), as well as the related BIC penalization term.
The approximation of the NML parametric complexity is very good as soon as $n$ is beyond 100, but less accurate for small sample sizes.
Asymptotically, the parametric complexity of the enumerative code is twice that of the NML approach. It is always greater, but for very small sample sizes, the difference becomes smaller and smaller.
\subsection{Overall code length}
Both the enumerative and NML codes exploit universal distributions on all binary strings $x^n \in X^n$.
The compression of the data with the help of the model is better for the enumerative distribution, at the expense of a worse parametric complexity.
Adding the parametric complexity and stochastic complexity terms, previous sections show that the NML code is much shorter for degenerated binary strings:
\begin{eqnarray}
\mathrm{for} \; k(x^n) = 0 \;\, \mathrm{or} \;\, k(x^n) = n, \quad \quad&&\\ \nonumber
L_{enum}\left(\widehat{\theta}(x^n), x^n\right) &\approx& \log n\\ \nonumber
L_{nml}\left(\widehat{\theta}(x^n), x^n\right) &\approx& \frac{1}{2} \log n + \frac{1}{2} \log \frac{\pi}{2},
\end{eqnarray}
whereas the enumerative code is slightly shorter for equidistributed binary strings (where $H(\widehat{\theta}(x^n)) \approx \log 2$), with a margin of $\log \frac{\pi}{2}$:
\begin{eqnarray}
\label{eqn_mixture}
\mathrm{for} \; k(x^n) \approx n/2, \quad \quad&&\\ \nonumber
L_{enum}\left(\widehat{\theta}(x^n), x^n\right) &\approx& \frac{1}{2} \log n - \frac{1}{2} \log \frac{\pi}{2} + n \log 2,\\ \nonumber
L_{nml}\left(\widehat{\theta}(x^n), x^n\right) &\approx& \frac{1}{2} \log n + \frac{1}{2} \log \frac{\pi}{2} + n \log 2.
\end{eqnarray}
\subsection{Expectation of the coding length of all binary strings}
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.85\textwidth]{CompareBernoulliCodingLength.pdf}
\caption{Expected overhead of coding length w.r.t. random model.}
\label{fig_codinglength}
\end{center}
\end{figure}
Let us now estimate the expectation of the coding length for all binary strings under the uniform distribution, where $\forall x^n \in X^n, p(x^n) = 1/{2^n}$.
\begin{eqnarray}
\label{eqn_codinglength}
\mathrm{E}\left(L\left(\widehat{\theta}(x^n), x^n\right)\right) &=& \frac{1} {2^n} \sum_{x^n \in X^n} {L\left(\widehat{\theta}(x^n), x^n\right)},\\ \nonumber
&=& \frac{1} {2^n} \sum_{k=0}^n {{{n}\choose{k}} L\left(\widehat{\theta}(x^n), x^n\right)}.
\end{eqnarray}
We perform an exact numerical calculation using the exact value of the NML model complexity term, for all $n, 1 \leq n \leq 1000$.
Figure~\ref{fig_codinglength} reports the expected coding length for the enumerative and NML codes minus that of the random code ($n \log 2$).
The results show that both codes have an average overhead of about $1/2 \log n$ compared to the direct encoding of the binary strings, and that,
under the uniform distribution,
the enumerative code always compresses the data better on average that the NML code, especially in the non-asymptotic case.
Actually, averaging on all binary strings is the same as considering exhaustively all the binary string outcomes of a Bernoulli distribution with parameter $\theta = 1/2$.
Using the central limit theorem, the proportion of binary strings $x^n$ where $k(x^n)/n \approx 1/2$ goes to 1 as $n$ goes to infinity, which explains why the shorter coding lengths obtained with the enumerative code for binary strings with $k(x^n)/n \approx 1/2$ provide the main contribution in the expectation.
Using Formula~\ref{eqn_mixture}, the expected coding length of the enumerative code is asymptotically better than that of the NML code by a margin of $\log \frac{\pi}{2}$.
\subsection{Percentage of compressible binary strings}
\label{percentCompressibleBernoulli}
We now focus on the percentage $p_{compressible}$ of compressible binary strings using both the enumerative and NML codes, that is the percentage of binary strings with coding length shorter than $n \log 2$:
\begin{eqnarray}
\label{eqn_percentcompressible}
p_{compressible} &=& \frac{1} {2^n} \sum_{x^n \in X^n}
{\mathbb{1}_{\left\{L\left(\widehat{\theta}(x^n), x^n\right) \leq n \log 2\right\}}},\\ \nonumber
&=& \frac{1} {2^n} \sum_{k=0}^n {{{n}\choose{k}}
\mathbb{1}_{\left\{L\left(\widehat{\theta}(x^n), x^n\right) \leq n \log 2\right\}}}.
\end{eqnarray}
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.85\textwidth]{CompareBernoulliPercentCompressible.pdf}
\caption{Percentage of compressible binary strings.}
\label{fig_percentcompressible}
\end{center}
\end{figure}
As previously, we perform an exact numerical calculation for all $n, 1 \leq n \leq 1000$.
Figure~\ref{fig_percentcompressible} shows that the percentage of compressible strings decreases at a rate of $O(1/\sqrt{n})$ for both codes, as expected.
However, the enumerative code always compresses more binary strings than the NML code. Due to the discrete decision threshold in formula~\ref{eqn_percentcompressible}, the exact computed percentage values are not smooth like in Figure~\ref{fig_codinglength}, especially in the non-asymptotic case for small string sizes, but the overall tendency appears clearly for large sample sizes.
In the asymptotic case, around $60\%$ more strings can be compressed using the enumerative code (empirical evaluation).
\subsection{Distribution of compression rates}
\label{compressionrateBernoulli}
We now focus on the distribution of the compression rates, that is the ratio of the coding length of a binary string using the Bernoulli versus the random model:
\begin{equation}
\label{eqn_compressionrate}
\%compression = \frac{L\left(\widehat{\theta}(x^n), x^n\right)} {n \log 2}.
\end{equation}
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.7\textwidth]{CompareBernoulliCompressionRate10.pdf}\\
\includegraphics[width=0.7\textwidth]{CompareBernoulliCompressionRate100.pdf}\\
\includegraphics[width=0.7\textwidth]{CompareBernoulliCompressionRate1000.pdf}
\caption{Inverse cumulative distribution of compression rates for strings of size 10, 100, 1000.}
\label{fig_compressionrate}
\end{center}
\end{figure}
Figure~\ref{fig_compressionrate} shows the inverse cumulative distribution of compression rates for binary strings of size 10, 100, 1000, using the NML or enumerative code .
For example, from the 1024 ($2^{10}$) strings of size 10, only $0.2\%$ (the two ``pure'' strings) can be compressed better with the NML than with the enumerative code.
All the other strings are better compressed with the enumerative code. For both codes, $11\%$ of the strings are compressible ($\%compression < 1$).
For string of size 100, only $3.0\; 10^{-15}\%$ of the strings are better compressed with the NML code.
However, $2.1\%$ of the strings are compressible using the NML code, which is less that the $3.5\%$ obtained using the enumerative code.
The tendency is the same for string of size 1000. A tiny portion of the strings, those with almost only zeros or only ones, are better compressed using the NML code.
All the other string are better compressed using the enumerative code, with difference growing for balanced strings. This results in a greater number of compressible strings using the enumerative code.
Let us evaluate the asymptotic value of the Bernoulli parameter $\theta$ for which both codes achieve the same compression rate.
Using Table~\ref{tableCodes} and Formula~\ref{deltaL} in the asymptotic case, we get :
\begin{eqnarray}
\label{eqn_compareCodes}
\delta\left(COMP^{(n)} + L\left(x^n|\widehat{\theta}(x^n)\right) \right) = 0
&\Leftrightarrow& \frac{1}{2} \log \frac{n \pi}{2} - \log(n+1) +\frac{1}{2} \log (2 \pi n \mathrm{var}(\theta)) = 0 \nonumber \\
&\Leftrightarrow& \log (2 \pi n \mathrm{var}(\theta)) = \log \frac {(n+1)^2}{n \pi /2} \nonumber \\
&\Leftrightarrow& \mathrm{var}(\theta) = \frac {(n+1)^2}{n^2 \pi^2} \approx \frac {1}{\pi^2} \nonumber \\
&\Leftrightarrow& \theta(1-\theta) = \frac {1}{\pi^2}
\end{eqnarray}
Equation~\ref{eqn_compareCodes} has two solutions: $\theta = 1/2 (1 \pm \sqrt{1-4/{\pi^2}})$, that is $\theta \approx 0.114$ and $\theta \approx 0.886$. Thus asymptotically, the enumerative code better compresses the strings for $\theta \in [0.114, 0.886]$, that is around $77\%$ of the values of $\theta$.
Overall, both the NML and enumerative codes have the same asymptotic behavior, with tiny differences in compression rates. However, the enumerative code allows to better compress far more strings, both in the non-asymptotic and asymptotic cases.
\subsection{Detection of a biased coin}
\label{biasedCoin}
We apply the previous Bernoulli codes to the problem of detection of a biased coin.
A fair coin is a randomizing device with two states named \emph{heads} and \emph{tails} that are equally likely to occur. It can be modeled using a Bernoulli process with $\theta_{fair} = \frac{1}{2}$.
For a biased coin, the heads and tails are not equally likely to occurs, and the related Bernoulli parameter is $\theta_{bias} \neq \frac{1}{2}$.
The problem is to determine whether a coin is biased given a limited sample of Bernoulli trials.
Given a sample $x^n$ of trials, we compute the coding length of this sample using either the NML or the enumerative code and decide that the coin is biased if its coding length is shorter than that of the random code ($n\log 2$). For a given size $n$ and a code (e.g. enumerative or NML), we compute the probability of detecting the biased coin by averaging the detection over all the possible samples of size $n$.
Formally, for each code, we thus compute:
\begin{eqnarray}
\label{eqn_biasedCoin}
prob^D (\theta_{bias}, n)
&=& \mathrm{E}_{B(\theta_{bias})}
\left( \mathbb{1}_{ \left\{ L\left(\widehat{\theta}(x^n), x^n\right) < n \log 2 \right\} } \right) \\
&=& \sum_{k=0}^n {{{n}\choose{k}} \theta_{bias}^{k} (1-\theta_{bias})^{n-k}
\mathbb{1}_{ \left\{ L\left(\widehat{\theta}(x^n), x^n\right) < n \log 2 \right\} } }.
\end{eqnarray}
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.7\textwidth]{StudyBiasedCoin40DetectionThreshold.pdf}\\
\caption{Probability of detection of of a biased coin where $\theta_{bias} = 0.40$.}
\label{fig_BiasedCoinProbDetection}
\end{center}
\end{figure}
The issue is to be able to detect the biased coin with the minimum number of trials.
Using Formula~\ref{eqn_biasedCoin}, we then determine the first value of $n$ where the probability of detecting the biased coin is beyond $50\%$.
For example, Figure~\ref{fig_BiasedCoinProbDetection} shows the probability of detection of a biased coin ($\theta_{bias} = 0.40$) for sample size rangin from 1 to 1000, using the enumerative and NML codes.
The horizontal gray line represents a probability of $50\%$ of detecting the bias.
As Formula~\ref{eqn_biasedCoin} is not strictly increasing with $n$ and unstable for tiny $n$ (for reasons similar as in Section~\ref{percentCompressibleBernoulli}), we collect the two following lower and upper thresholds of sample sizes:
\begin{eqnarray}
\label{eqn_thresholdBiasedCoin}
\underline{n}^{\; D} (\theta_{bias})
&=& \min_{n \geq 10} \{prob^D (\theta_{bias}, n) \geq 50\% \},\\
\overline{n}^{\; D} (\theta_{bias})
&=& \max_{n \geq 10} \{prob^D (\theta_{bias}, n) \leq 50\% \}.
\end{eqnarray}
In Figure~\ref{fig_BiasedCoinProbDetection} for example, we have
$\underline{n}^{\; D} (\theta_{bias}) = 96$ and $\overline{n}^{\; D} (\theta_{bias}) = 115$ for the enumerative code and
$\underline{n}^{\; D} (\theta_{bias}) = 126$ and $\overline{n}^{\; D} (\theta_{bias}) = 145$ for the NML code, that thus needs around $10\%$ more trials to detect the biased coin.
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.7\textwidth]{StudyBiasedCoinDetectionSizeThreshold.pdf}\\
\caption{Minimum sample size to detect a biased coin with probability greater than $50\%$.}
\label{fig_thresholdBiasedCoin}
\end{center}
\end{figure}
Figure~\ref{fig_thresholdBiasedCoin} shows the detection thresholds computed using the NML or enumerative codes for $\theta_{bias} \in [0.35; 0.5]$.
As expected, the min sample size necessary to detect a biased coin increases quickly when $\theta_{bias}$ becomes close to $\frac{1}{2}$. For $\theta_{bias} \approx 0.46$, around 1,000 trials are necessary to detect the bias, and for $\theta_{bias} \approx 0.495$, around 100,000 trials are necessary.
Although all the thresholds are quite close, the enumerative code always needs smaller sample sizes to detect the biased coin.
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.7\textwidth]{StudyBiasedCoinDetectionSizeThreshold_ratio.pdf}\\
\caption{Minimum sample size to detect a biased coin with probability greater than $50\%$.}
\label{fig_thresholdBiasedCoinRatio}
\end{center}
\end{figure}
To better compare the threshold without being hampered by the logarithmic scale of the sample size, Figure~\ref{fig_thresholdBiasedCoinRatio} shows all the detection thresholds normalized by the enumerative upper threshold. The lower and upper thresholds converge both for the enumerative and NML codes. However, the difference between code does not vanish with the sample size. At least withing the range explored in this experiment, up to 100,000 trials, the enumerative code always needs around $10\%$ less samples on average than the NML code to detect a biased coin.
\paragraph{False versus true positive rate.}
The probability of detecting a bias when a coin is actually biased can be interpreted as a true positive rate, and when the coin is fair as a false positive rate.
Given this, the enumerative code needs less samples than the NML code to detect a bias with a true positive rate greater than $50\%$.
In the case of a fair coin, the false positive rate of both codes decreases at a rate of O$(1/\sqrt n)$, as shown in the experiment related to the percentage of compressible strings (cf. Section~\ref{percentCompressibleBernoulli}: formula~\ref{eqn_percentcompressible} is the same as formula~\ref{eqn_thresholdBiasedCoin} for $\theta_{bias} = \frac{1}{2}$). Still, the false positive rate is about $60\%$ higher for the enumerative code than for the NML code.
Overall, the enumerative code compresses most binary strings slightly better than the NML code, resulting in a better sensitivity to biased coins at the expense of more false detections in case of fair coins.
\subsection{Biased versus fair coin classification}
\label{secCoinClassification}
To further investigate on the comparison between the NML and enumerative codes, we suggest a classification experiment where the objective is to predict whether a coin if fair or biased.
Let $\theta_{bias} \in [0;1]$ and $n \in \mathbb{N}^*$ be fixed parameters.
The instances to classify are sequences $x^n$ of $n$ trials generated with equal probability ($p_F=p_B = \frac{1}{2}$) either from a fair coin ($\theta = \frac{1}{2}$) or from a biased coin ($\theta = \theta_{bias}$).
The objective is to predict whether the coin that produced each sequence was fair or biased.
As in Section~\ref{biasedCoin}, we evaluate both the NML and enumerative codes as classifiers by predicting a bias if they can encode a sequence with a coding length shorter than that of the random code ($n\log 2$), and predicting fair otherwise.
\begin{table}[htbp!]
\caption{Coin classification results.}
\label{coinContigencyTable}
\centering
\begin{tabular}[10pt]{|c|c|c|}\hline
{\scriptsize Real $\downarrow$ Predicted $\rightarrow$} & \qquad Bias \quad \quad & \qquad Fair \quad \quad \\\hline
Bias & \qquad TP \quad \quad & \qquad FN \quad \quad \\\hline
Fair & \qquad FP \quad \quad & \qquad TN \quad \quad \\\hline
\end{tabular}
\end{table}
The result can be analyzed in terms of a contingency table, as illustrated in Table~\ref{coinContigencyTable}:
\begin{itemize}
\item true positive (TP): detecting bias correctly,
\item false positive (FP): detecting bias when there is none,
\item true negative (TN): detecting fair correctly,
\item false negative (FN): detecting fair when the coin is biased.
\end{itemize}
In this experiment, we focus on the correct detections:
\begin{itemize}
\item sensitivity or true positive rate $TPR= TP/(TP+FN)$ for the correct detection of bias,
\item specificity or true negative rate $TNR= TN/(TN+FP)$ for the correct detection of fair,
\item accuracy $ACC = (TP+TN)/(TP+FP+FN+TN)$ for the global rate of correct detection.
\end{itemize}
For given $\theta_{bias}$ and $n$ parameters and for each code, we compute the expectation of the indicators by integrating other the distribution of all the sequences issued from the generation process.
\begin{eqnarray}
\label{eqn_TPR}
E(TPR)
&=& \mathrm{E}_{B(\theta_{bias})}
\left( \mathbb{1}_{ \left\{ L\left(\widehat{\theta}(x^n), x^n\right) < n \log 2 \right\} } \right),\\
&=& \sum_{k=0}^n {{{n}\choose{k}} \theta_{bias}^{k} (1-\theta_{bias})^{n-k}
\mathbb{1}_{ \left\{ L\left(\widehat{\theta}(x^n), x^n\right) < n \log 2 \right\} } },\\
E(TNR)
&=& \mathrm{E}_{B(1/2)}
\left( \mathbb{1}_{ \left\{ L\left(\widehat{\theta}(x^n), x^n\right) \geq n \log 2 \right\} } \right), \\
&=& \frac{1}{2^n} \sum_{k=0}^n {{{n}\choose{k}}
\mathbb{1}_{ \left\{ L\left(\widehat{\theta}(x^n), x^n\right) \geq n \log 2 \right\} } },\\
E(ACC)
&=& p_B E(ETR) + p_F E(TNR),\\
&=& \frac{E(ETR) + E(TNR)}{2}.
\end{eqnarray}
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.45\textwidth]{StudyCoinDetectionAccuracy_0501.pdf}
\includegraphics[width=0.45\textwidth]{StudyCoinDetectionAccuracy_0501_delta.pdf}\\
\includegraphics[width=0.45\textwidth]{StudyCoinDetectionAccuracy_051.pdf}
\includegraphics[width=0.45\textwidth]{StudyCoinDetectionAccuracy_051_delta.pdf}\\
\includegraphics[width=0.45\textwidth]{StudyCoinDetectionAccuracy_055.pdf}
\includegraphics[width=0.45\textwidth]{StudyCoinDetectionAccuracy_055_delta.pdf}\\
\includegraphics[width=0.45\textwidth]{StudyCoinDetectionAccuracy_060.pdf}
\includegraphics[width=0.45\textwidth]{StudyCoinDetectionAccuracy_060_delta.pdf}\\
\includegraphics[width=0.45\textwidth]{StudyCoinDetectionAccuracy_075.pdf}
\includegraphics[width=0.45\textwidth]{StudyCoinDetectionAccuracy_075_delta.pdf}\\
\includegraphics[width=0.45\textwidth]{StudyCoinDetectionAccuracy_090.pdf}
\includegraphics[width=0.45\textwidth]{StudyCoinDetectionAccuracy_090_delta.pdf}
\caption{Classification of coins using the NML and Enum codes for different biases.
Accuracy (left) and difference $(Enum - NML)$ for the true positive, false negative and accuracy (right).}
\label{fig_CoinClassification}
\end{center}
\end{figure}
We perform the coin classification experiment for $\theta_{bias} \in \{0.501, 0.51, 0.55, 0.75, 0.60, 0.90\}$ and $n$ ranging from 1 to $10,000$ using the enumerative and the NML codes ($n$ up to $10,000,000$ for $\theta_{bias} = 0.501$).
Figure~\ref{fig_CoinClassification} reports the accuracy results (left) as well as the difference $(Enum - NML)$ of the three indicators.
Overall, both codes exhibit a similar behavior w.r.t. the coin classification problem, with accuracy increasing from 0.5 for small $n$ to 1 for large $n$, and a slow increase rate for small bias and a fast one for large bias. Except in the tiny samples with $n \leq 20$, the difference between any of the three indicators never exceeds around 15\%.
However, there are some interesting differences.
As noticed in Section~\ref{biasedCoin}, the enumerative code has a better sensitivity at the expense to a worse specificity, and the aggregated accuracy result exhibits a variety of behaviors.
When the bias is small ($\theta_{bias}$ close from $\frac{1}{2}$), the enumerative code is far more sensitive while being a little less specific, resulting in more accurate predictions in the non-asymptotic case.
When the bias is large ($\theta_{bias}$ far from $\frac{1}{2}$), both codes get almost the same sensitivity while the enumerative code remains less specific, resulting in slightly less accurate predictions.
In all cases, the differences between both codes get tiny for large $n$, in the asymptotic case.
\section{The case of multinomial distribution}
\label{secMultinomial}
Let us consider the multinomial model with parameter $\theta = (\theta_1, \ldots, \theta_m), \; \sum_{j=1}^m {\theta_j}=1, \forall j, \theta_j> 0$, such that $P_{\theta}(X=j) = \theta_j$, in the case of m-ary sequences $x^n \in X^n$ of size $n$.
For a given sequence $x_n$, $P_{\theta}(x_n) = \prod_{j=1}^m {\theta_j^{n_j}}$, where $n_j$ is the number of occurrences of outcome $j$ in sequence $x^n$.
\subsection{Standard NML approach}
\label{secMultinomialNML}
As pointed out in \citep{RooijEtAl09},
\begin{quotation}
``The NML distribution has a number of significant practical problems.
First, it is often undefined, because for many models the denominator in (\ref{eqnNML}) is infinite, even for such simple models as the Poisson or geometric distributions.
Second, $X^n$ is exponentially large in $n$, so calculating the NML probability exactly is only possible in special cases such as the Bernoulli model above, where the number of terms is reduced using some trick. Something similar is possible for the more general multinomial model
(...
, but in most cases
[it]
has to be approximated, which introduces errors that are hard to quantify.''
\end{quotation}
The parametric complexity of the NML universal model with respect to a k-parameter exponential family model is usually approximated by $\frac{k}{2} \log \frac{n}{2 \pi}$ \citep{Grunwald07}.
In the case of the multinomial distribution with $(m-1)$ free parameters, this gives $\frac{m-1}{2} \log \frac{n}{2 \pi}$.
A better approximation based on Rissanen's asymptotic expansion \citep{Rissanen96} is presented in \citep{Kontkanen2009}:
\begin{equation}
\label{compM_R}
COMP_{nml}^{(n)}(\theta) = \frac{m-1}{2} \log \frac{n}{2 \pi} + \log \frac{\pi^{m/2}}{\Gamma (m/2)} + o(1),
\end{equation}
where $\Gamma(.)$ is the Euler gamma function.
Still in \citep{Kontkanen2009}, a sharper approximation based on Szpankowski's approximation is presented. This last approximation, far more complex is very accurate w.r.t. $n$, with $o(\frac{1}{n^{3/2}})$ precision.
We present below its first terms until $o(\frac{1}{\sqrt n})$, which actually are the same that in Rissanen's approximation:
\begin{equation}
\label{compM_S}
COMP_{nml}^{(n)}(\theta) = \frac{m-1}{2} \log \frac{n}{2} + \log \frac{\sqrt \pi}{\Gamma (m/2)} + o(\frac{1}{\sqrt n}),
\end{equation}
Finally, \citep{KontkanenEtAl07} propose an exact computation of the multinomial stochastic complexity, at the expense of sophisticated algorithms with quasilinear computation time.
\subsection{Enumerative two-part crude MDL}
\label{secEnumerativeM}
We apply the same approach as in the case of the Bernoulli model (Sections~\ref{secEnumerativeB} and \ref{secEnumBayesian}).
Given a sample size $n$, the number of tuples $(n_1, n_2, \ldots, n_m)$ such that $\sum_{j=1}^m {n_j} = n$ is ${{n+m-1}\choose{m-1}}$.
We then encode the multinomial model parameter using a uniform prior
$$P\left(\theta = \left(\frac{n_1}{n}, \frac{n_2}{n}, \ldots, \frac{n_m}{n}\right) \right) = 1 / {{n+m-1}\choose{m-1}},$$ leading to $L(\theta) = \log {{n+m-1}\choose{m-1}}$.
Second, we have to encode the data $x^n$ at best given the $\theta$ parameter.
We suggest using a probability distribution for encoding the finite size data sample $x^n$, with the following likelihood.
For $\theta \neq \left(\frac{n_1(x^n)}{n}, \frac{n_2(x^n)}{n}, \ldots, \frac{n_m(x^n)}{n}\right)$, we cannot encode the data and $P(x^n|\theta) = 0$.
For $\theta = \widehat{\theta}(x^n) = \left(\frac{n_1(x^n)}{n}, \frac{n_2(x^n)}{n}, \ldots, \frac{n_m(x^n)}{n}\right)$, the observed data is consistent with the model parameter and we assume that all the possible observable data are uniformly distributed.
The number of m-ary strings where the number of occurrences of outcome $j$ is $n_j$ is given by the multinomial coefficient $\frac{n!}{n_1! n_2! \ldots n_m!}$. Thus the probability of observing one particular m-ary string is $P(x^n|\widehat{\theta}(x^n)) = 1/\frac{n!}{n_1! n_2! \ldots n_m!}$.
This gives a total code length of
\begin{equation}
L(\widehat{\theta}(x^n), x^n) = \log {{n+m-1}\choose{m-1}} + \log \frac{n!}{n_1! n_2! \ldots n_m!},
\end{equation}
defined only when $\theta = \widehat{\theta}(x^n)$.
\subsection{NML interpretation}
Let us compute the NML parametric complexity of this enumerative code on the basis of the discrete likelihood.
We have
\begin{eqnarray}
COMP^{(n)}(\theta) &=& \log \sum_{y^n \in X^n} {P_{\widehat{\theta}(y^n)}(y^n)}, \\
&=& \log \sum_{\{n_1+\ldots+n_m=n\}} {\frac{n!}{n_1! n_2! \ldots n_m!} \left({1} / \frac{n!}{n_1! n_2! \ldots n_m!} \right)},\\
&=& \log {{n+m-1}\choose{m-1}}.
\end{eqnarray}
Interestingly, we find exactly the same complexity term as the coding length of the best hypothesis in the enumerative approach, that simply relies on counting the possibilities for the model parameters.
Like in the Bernoulli case, this shows that the enumerative code is both a two-part and a one-part code, optimal w.r.t. the NML approach and parametrization invariant.
We have an exact formula for the complexity term, very simple to compute.
Using Stirling's approximation $\log n! = n \log n - n + \frac{1}{2} \log {2 \pi n} + O(1/n)$, we get the following asymptotic approximation:
\begin{equation}
\label{compEnum_M}
COMP^{(n)}(\theta) = (m-1)(\log n - \log (m -1) +1) -\frac{1}{2} \log {2 \pi (m-1)} + o(\frac{1}{n}).
\end{equation}
Once again, this asymptotic model complexity is twice that of the alternative classical NML code or the standard BIC regularization term $(m-1)\log n$.
\section{Code comparison for the multinomial distributions}
\label{secComparisonM}
In this section, we compare the NML code (Section~\ref{secMultinomialNML}) and enumerative two-part crude MDL codes (Section~\ref{secEnumerativeM}) for the multinomial distribution.
\subsection{Notation}
\label{secNotationM}
Let us use the names \emph{NML} and \emph{enumerative} for the specific MDL codes presented in Sections \ref{secMultinomialNML} and \ref{secEnumerativeM}.
We also consider the \emph{random} code as a baseline: it corresponds to a direct encoding of each binary string $x^n$ with a coding length of $n \log m$. The likelihood of each string $x^n$ is $1/m^n$, and as $\sum_{\{n_1+\ldots+n_m=n\}} {\frac{n!}{n_1! n_2! \ldots n_m!} 1/m^n} = 1$, we have $COMP_{random}^{(n)}(\emptyset) = 0$ and
$L_{random}\left(x^n|\emptyset \right) = n \log m$.
Table~\ref{tableCodesM} reminds the parametric and stochastic complexity of each considered code for the multinomial distribution.
\begin{table}[htbp!]
\caption{Parametric and stochastic complexity per code.}
\label{tableCodesM}
\centering
\renewcommand{\arraystretch}{1.5}
\begin{tabular}[10pt]{ccc}\hline
Code name & $COMP_{name}^{(n)}$ & $L_{name}\left(x^n|\widehat{\theta}(x^n) \right)$ \\\hline
\emph{enumerative} & $\log {{n+m-1}\choose{m-1}}$ & $\log \frac {n!} {n_1! \ldots n_m!}$ \\
\emph{NML} & $\frac{m-1}{2} \log \frac{n}{2} + \log \frac{\sqrt \pi}{\Gamma (m/2)} + o(\frac{1}{\sqrt n})$ & $\log \frac {n^n} {n_1^{n_1} \ldots n_m^{n_m}}$ \\
\emph{random} & $0$ & $n \log m$ \\\hline
\end{tabular}
\end{table}
\subsection{Stochastic complexity term}
\label{secStochasticM}
The stochastic complexity term of the enumerative code is always smaller than that of the NML code for non-degenerated m-ary strings:
\begin{equation}
\label{sc_Multinomial}
\forall n, \forall x^n \in X^n \; \mbox{such that} \; (\max_{1 \leq j \leq m} n_j ) < n \; \mbox{ then} \;
L_{enum}\left(x^n|\widehat{\theta}(x^n)\right) < L_{nml}\left(x^n|\widehat{\theta}(x^n)\right).
\end{equation}
An intuitive proof relies on the fact that the enumerative MDL likelihood assigns the same probability to all m-ary strings having the same number of occurrence per outcome $j$, with a null probability for all the other strings. The NML likelihood also assigns the same probability to these m-ary strings, but with a non-null probability for the other strings. Then they have to share a smaller probability mass, resulting in a smaller probability per string and a strictly greater coding length.
\medskip
To gain further insights, let us approximate the difference of coding length for the stochastic complexity term:
$$\delta L_{SC}\left(x^n|\widehat{\theta}(x^n)\right) = L_{nml}\left(x^n|\widehat{\theta}(x^n)\right) -L_{enum}\left(x^n|\widehat{\theta}(x^n)\right).$$
We assume that $\forall j, n_j > 0$ and $n_j \approx n \widehat{\theta}_j$.
Using Stirling's approximation $\log n! = n \log n - n + \frac{1}{2} \log {2 \pi n} + O(1/n)$, we get
\begin{eqnarray*}
L_{enum}\left(x^n|\widehat{\theta}(x^n)\right)
&=& \log \frac{n!}{n_1! n_2! \ldots n_m!} \\
%
&=& n \log n -n + \frac{1}{2} \log {2 \pi n} + O(1/n) \\
&& -\sum_{j=1}^m {\left(n_j \log n_j - n_j + \frac{1}{2} \log {2 \pi n_j} + O(1/n_j)\right)} \\
%
&=& \log \frac {n^n} {n_1^{n_1} \ldots n_m^{n_m}} - \frac{m-1}{2} \log {2 \pi n}
- \frac{1}{2} \log \prod_{j=1}^m {\widehat{\theta}_j} + O(1/n) \\
\end{eqnarray*}
We get
\begin{equation}
\label{deltaM_SC}
\delta L_{SC}\left(x^n|\widehat{\theta}(x^n)\right) = \frac{m-1}{2} \log {2 \pi n}
+ \frac{1}{2} \log \prod_{j=1}^m {\widehat{\theta}_j} + O(1/n).
\end{equation}
It is noteworthy that the $\mathrm{var}(\widehat{\theta})$ term in the Bernoulli case (see Formula~\ref{deltaL}) generalizes to a $\prod_{j=1}^m {\widehat{\theta}_j}$ term in the multinomial case.
\medskip
These results demonstrate that the enumerative code provides a better encoding of the data with the help of the model for any m-ary strings. The gain in coding length compared to the NML code is always positive and grows asymptotically as $(m-1)/2$ times the logarithm of the sample size.
\subsection{Parametric complexity term}
\label{secParametricM}
Using inequality~\ref{sc_Multinomial} and as the parametric complexity of code is the sum of the stochastic complexity over all possible strings, we get:
\begin{equation}
\label{pc_Multinomial}
\forall n > 1, COMP_{enum}^{(n)} > COMP_{NML}^{(n)}.
\end{equation}
Both terms are equal for $n=1$ and asymptotically, the parametric complexity of the enumerative code is twice that of the NML code (see Formulas~\ref{compM_S} and \ref{compEnum_M}).
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.85\textwidth]{CompareMultinomialCOMP10.pdf}\\
\includegraphics[width=0.85\textwidth]{CompareMultinomialCOMP100.pdf}
\caption{Parametric complexity for the multinomial model with 10 or 100 categories.}
\label{fig_compM}
\end{center}
\end{figure}
We now focus on the non-asymptotic behavior of the parametric complexity terms and their approximations.
Figure~\ref{fig_compM} shows the value of the parametric complexity of the multinomial model, using the enumerative code, the NML code (exact numerical computation and Rissanen or Szpankowski based approximations: see Section~\ref{secMultinomialNML}), as well as the related BIC penalization term.
The BIC approximation is very bad, all the more as $m$ increases.
The Rissanen approximation of the NML parametric complexity is very good as soon as $n$ is about 100 times the number $m$ of categories, but less accurate for small sample sizes.
As expected, the Szpankowski based approximation if much sharper, being accurate as soon as $n$ is beyond $m$, but with bad accuracy for $n << m$.
\medskip
Let us now compute an asymptotic approximation of the difference of parametric complexity between the two codes:
$$\delta L_{PC} COMP^{(n)}(\theta) = COMP_{nml}^{(n)}(\theta) - COMP_{enum}^{(n)}(\theta).$$
Using previous approximations presented in Formulas~\ref{compM_S}, ~\ref{compEnum_M} and the Stirling's approximation of the Gamma function, we get:
\begin{eqnarray}
\label{eqn_diffPC_M}
\delta L_{PC} COMP^{(n)}(\theta)
&=& \frac{m-1}{2} \log \frac{n}{2} + \log \sqrt \pi
-(\frac{m}{2} \log \frac{m}{2} - \frac{m}{2} -\frac{1}{2} \log \frac{m}{4 \pi})\\
&& - ((m-1)(\log n - \log (m -1) +1) -\frac{1}{2} \log {2 \pi (m-1)}) \\
&& + o(\frac{1}{\sqrt n}),\\
&=& -\frac{m-1}{2} \log n -\frac{m-1}{2} \log 2 + \frac{1}{2}\log \pi \\
&&-\frac{m}{2} \log m + \frac{m}{2} \log 2 + \frac{m}{2} + \frac{1}{2} \log m - \frac{1}{2} \log \pi - \log 2\\
&& + (m-1)\log (m -1) - (m-1) \\
&& + \frac{1}{2} \log 2 + \frac{1}{2} \log \pi + \frac{1}{2} \log (m-1) + o(\frac{1}{\sqrt n}),\\
&=& -\frac{m-1}{2} \log {n m e} + (m - \frac{1}{2}) \log (m-1) + \frac{1}{2} \log {\pi e} + o(\frac{1}{\sqrt n}).
\end{eqnarray}
We obtain
\begin{eqnarray}
\label{deltaM_PC}
\delta L_{PC} COMP^{(n)}(\theta)
&=& -\frac{m-1}{2} \log n + \frac{m}{2}\log{\frac{m}{e}} \\ \nonumber
&& +\log \frac{e}{\sqrt \pi} + (m-\frac{1}{2})\log(1-\frac{1}{m}) + o(\frac{1}{\sqrt n})
\end{eqnarray}
and for $n >> m$
\begin{eqnarray}
\label{deltaM_PC2}
\delta L_{PC} COMP^{(n)}(\theta)
&=& -\frac{m-1}{2} \log n + \frac{m}{2}\log{\frac{m}{e}}
-\log \sqrt \pi + o(\frac{1}{m}) + o(\frac{1}{\sqrt n}).
\end{eqnarray}
This result demonstrates that the difference of parametric complexity increases as the logarithm of the sample size. The speed of increase is with a factor $(m-1)/2$, but for small sample sizes (typically $n \leq m$), the difference remains small.
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.85\textwidth]{StudyMultinomialCOMP.pdf}
\caption{Ratio of enumerative to the NML parametric complexities for the multinomial model with up to 100,000 outcomes.}
\label{fig_ratio_compM}
\end{center}
\end{figure}
We illustrate this behavior in the non-asymptotic case.
Figure~\ref{fig_ratio_compM} focuses on the ratio of the exact parametric complexity terms for the enumerative and NML codes. This ratio always increases from 1 for $n=1$ to 2 when $n$ goes to infinity, with the speed of convergence decreasing as the number $m$ of outcomes increases.
\subsection{Overall code length}
\label{secOverallM}
Both the enumerative and NML codes exploit universal distributions on all m-ary strings $x^n \in X^n$.
The compression of the data with the help of the model is better for the enumerative distribution, at the expense of a worse parametric complexity.
The overall code length is the sum of the parametric and stochastic complexities.
Using previous approximations presented in Formulas~\ref{deltaM_SC}, ~\ref{deltaM_PC}, we obtain the following approximation of the difference of overall code lengths between the two codes:
\begin{eqnarray}
\label{eqn_mixtureM}
\Delta L\left(\widehat{\theta}(x^n), x^n\right)
&=& \delta L_{PC} COMP^{(n)}(\theta) + \delta L_{SC}\left(x^n|\widehat{\theta}(x^n)\right),\\
&=& -\frac{m-1}{2} \log n + \frac{m}{2}\log{\frac{m}{e}}
+\log \frac{e}{\sqrt \pi} + (m-\frac{1}{2})\log(1-\frac{1}{m})\\
&& + \frac{m-1}{2} \log {2 \pi n}
+ \frac{1}{2} \log \prod_{j=1}^m {\widehat{\theta}_j} + o(\frac{1}{\sqrt n}).
\end{eqnarray}
We obtain
\begin{footnotesize}
\begin{eqnarray}
\label{eqn_mixtureM2}
\Delta L\left(\widehat{\theta}(x^n), x^n\right)
&=& \frac{m}{2} \log \frac{m 2 \pi}{e} + \frac{1}{2} \log \prod_{j=1}^m {\theta_j}
+\log \frac{e}{\sqrt 2} + (m-\frac{1}{2})\log(1-\frac{1}{m})
+ o(\frac{1}{\sqrt n}),
\end{eqnarray}
\end{footnotesize}
and for $n >> m$
\begin{eqnarray}
\label{eqn_mixtureM3}
\Delta L\left(\widehat{\theta}(x^n), x^n\right)
&=& \frac{m}{2} \log \frac{m 2 \pi}{e}
+ \frac{1}{2} \log \prod_{j=1}^m {\theta_j}
- \log \sqrt 2 + o(\frac{1}{m}) + o(\frac{1}{\sqrt n}) .
\end{eqnarray}
Asymptotically, the difference in overall code length does not depend on the size $n$ of the string.
Both codes differ by a margin that depends essentially on the number $m$ of outcomes and of the multinomial parameter $\theta$.
\paragraph{Case of balanced multinomial distributions.}
The term $\log \prod_{j=1}^m {\theta_j}$ is minimal for equidistributed multinomial distribution ($\theta_j=1/m$).
For such distributions, we get
\begin{eqnarray}
\label{eqn_mixtureM4}
\Delta L\left(\widehat{\theta}(x^n), x^n\right)
&=& \frac{m}{2}\log \frac{2 \pi}{e} + \log \frac{e}{\sqrt 2} + o(\frac{1}{m}) + o(\frac{1}{\sqrt n}),
\end{eqnarray}
which means that the enumerative code compresses the strings better than the NML code with a margin that increases linearly with $m$.
\paragraph{Case of degenerated multinomial distributions.}
In case of multinomial distributions with one single observed outcome (e.g. $\widehat{\theta} = (1, 0, \ldots, 0)$), both the NML and enumerative codes have a null stochastic complexity and the difference between the coding lengths reduces to the difference between the parametric complexity terms (see Formula~\ref{deltaM_PC}). In this extreme case, the NML code compresses the string better with a margin that grows as $\frac{m-1}{2}$ times the logarithm of the sample size.
\paragraph{Case of unbalanced multinomial distributions.}
Let us study the boundary between balanced distributions and degenerated distributions, where the enumerative code dominates the NML code and conversely.
We are seeking for distributions where both codes achieve approximately the same compression.
For that purpose, let us consider peaked multinomial distributions, with most of the probability mass for the first outcome ($\theta_1 = \theta_{max}$) and the rest of the probability mass equistributed for the other outcomes ($\theta_j = \theta_{min}, 2 \leq j \leq m$, with $\theta_{min} = \frac {1 - \theta_{max}} {m-1}$).
Using Formula~\ref{eqn_mixtureM3}, we thus try to find the peaked distribution $\theta = (\theta_{max}, \theta_{min}, \ldots, \theta_{min})$ such that $\Delta L\left(\widehat{\theta}(x^n), x^n\right) = o(1)$.
The solution is obtained for
\begin{equation}
\label{peakThresholdMax}
\theta_{max} = 1-\frac{m-1}{m + \log m}\frac{e}{2 \pi} > 0.56,
\end{equation}
\begin{equation}
\label{peakThresholdMin}
\theta_{min} = \frac{e}{2 \pi (m + \log m)} < 0.44\frac{1}{m},
\end{equation}
leading to
\begin{equation}
\Delta L\left(\widehat{\theta}(x^n), x^n\right) = \alpha + o(\frac{1}{m}) + o(\frac{1}{\sqrt n})
\end{equation}
with $\alpha \approx 0.37$. This peaked distribution is at the limit where the NML code compresses the data better than the enumerative code.
Numerical experiments, not reported here, confirm the accuracy of Formulas~\ref{peakThresholdMax} and \ref{peakThresholdMin} and show that the asymptotic value of the peak probability $\theta_{max}$ behaves as a lower bound of the obtained probability in the non asymptotic case.
For Bernoulli distributions, we had $\theta_{max} \approx 0.886$ (see Formula~\ref{eqn_compareCodes}), and not surprisingly, $\theta_{max}$ decreases with $m$ (see Formula~\ref{peakThresholdMax}).
Interestingly, $\theta_{max}$ is always greater than $0.56$ whatever $m$. This means that when $m$ increases, the ratio $\theta_{max}/\theta_{min}$ grows linearly with $m$ and the fraction of multinomial distributions where the NML code dominates the enumerative code becomes negligible.
\paragraph{Synthesis.}
The overall difference of coding length is in favor of the enumerative code for balanced strings with a margin that increases linearly with $m$. The NML code is better only for heavily unbalanced strings, where the most frequent outcome occurs more that half of the times, whatever be $m$.
Under the uniform distribution, such unbalanced strings are far less frequent than balanced strings, and the enumerative code compresses most strings better than the NML code.
\subsection{Expectation of the coding length of all m-ary strings}
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{CompareMultinomial5CodingLength.pdf}
\includegraphics[width=0.49\textwidth]{CompareMultinomial10CodingLength.pdf}
\caption{Expected overhead of coding length w.r.t. random model for $m=5$ (left) and $m=10$ (right).}
\label{fig_codinglengthM}
\end{center}
\end{figure}
Let us estimate the expectation of the coding length for all m-ary strings under the uniform distribution, where $\forall x^n \in X^n, p(x^n) = 1/{m^n}$.
\begin{eqnarray}
\label{eqn_codinglengthM}
\mathrm{E}\left(L\left(\widehat{\theta}(x^n), x^n\right)\right) &=& \frac{1} {m^n} \sum_{x^n \in X^n} {L\left(\widehat{\theta}(x^n), x^n\right)},\\ \nonumber
&=& \frac{1} {m^n} \sum_{\{n_1+\ldots+n_m=n\}} {\frac{n!}{n_1! n_2! \ldots n_m!} L\left(\widehat{\theta}(x^n), x^n\right)}.
\end{eqnarray}
We perform an exact numerical calculation using the exact value of the NML parametric complexity term, for $m=5$ ($1 \leq n \leq 100$) and $m=10$ ($1 \leq n \leq 50$). Let us notice that the sum in Formula~\ref{eqn_codinglengthM} involves more than ten billions terms for $m=10$ and $n=50$.
Figure~\ref{fig_codinglengthM} reports the expected coding length for the enumerative and NML codes minus that of the random code ($n \log m$).
The results show that both codes have an asymptotic overhead that grows as $(m-1)/2 \log n$, compared to the direct encoding of the binary strings.
Under the uniform distribution, the enumerative code always compresses the data better than the NML code, especially in the non-asymptotic case.
As for the Bernoulli codes, most possible m-ary strings are almost equidistributed and their shorter coding lengths obtained with the enumerative provide the main contribution in the expectation of the coding length.
Following Formula~\ref{eqn_mixtureM3}, the enumerative code compresses the m-ary strings better than the NML code with a margin that grows linearly with $m$.
\subsection{Percentage of compressible m-ary strings}
\label{percentCompressibleMultibomial}
We now focus on the percentage $p_{compressible}$ of compressible m-ary strings using both the enumerative and NML codes, that is the percentage of m-ary strings with coding length shorter than $n \log m$:
\begin{eqnarray}
\label{eqn_percentMcompressible}
p_{compressible} &=& \frac{1} {m^n} \sum_{x^n \in X^n}
{\mathbb{1}_{\left\{L\left(\widehat{\theta}(x^n), x^n\right) \leq n \log m\right\}}},\\ \nonumber
&=& \frac{1} {m^n} \sum_{\{n_1+\ldots+n_m=n\}} {\frac{n!}{n_1! n_2! \ldots n_m!}
\mathbb{1}_{\left\{L\left(\widehat{\theta}(x^n), x^n\right) \leq n \log m\right\}}}.
\end{eqnarray}
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{CompareMultinomial3PercentCompressible.pdf}
\includegraphics[width=0.49\textwidth]{CompareMultinomial5PercentCompressible.pdf}
\caption{Percentage of compressible m-ary strings (left: $m=3$; right: $m=5$).}
\label{fig_percentMcompressible}
\end{center}
\end{figure}
As previously, we perform an exact numerical calculation for all $n, 1 \leq n \leq n_{max}$, with $n_{max} = 5000$ for $m=3$ and $n_{max} = 500$ for $m=5$.
Figure~\ref{fig_percentMcompressible} shows that empirically, beyond the non-asymptotic case, the percentage of compressible strings decreases at a rate of $O(1/n^{(m-1)/2})$ for both codes.
However, the enumerative code always compresses more binary strings than the NML code.
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.39\textwidth]{CompareBernoulliRatioCompressible.pdf}
\includegraphics[width=0.29\textwidth]{CompareMultinomial3RatioCompressible.pdf}
\includegraphics[width=0.26\textwidth]{CompareMultinomial5RatioCompressible.pdf}
\caption{Ratio of compressible binary strings using the enumerative rather than the NML code (left: $m=2$; center: $m=3$; right: $m=5$).}
\label{fig_ratiocompressible}
\end{center}
\end{figure}
To better characterize the behavior of each code, especially in the non-asymptotic case, we report in
Figure~\ref{fig_ratiocompressible} the ratio of the number of compressible strings of the enumerative to the NML code, for $m=2$ (Bernoulli), $m=3$ and $m=5$.
Empirically, beyond the non-asymptotic case, this ratio converges to a constant that increases with $m$: around 1.6 for $m=2$, 2.5 for $m=3$ and above 5 for $m=5$.
We expect that this empirical behavior generalizes for larger $m$, but empirical evaluation is not feasible for large $m$, even for small $n$.
On the other hand, studying the asymptotic behavior of this ratio is a non trivial task, beyond the scope of this paper.
\subsection{Detection of a biased dice}
\label{biasedDice}
We apply the previous multinomial codes to the problem of detection of a biased dice.
A fair dice is a randomizing device with $m$ outcomes that are equally likely to occur, which can be modeled using a multinomial process with equidistributed $\theta_j = \frac{1}{m}$.
Among all the possibilities of bias, we choose a simple family of peaked multinomial distributions, like those presented in Section~\ref{secOverallM}.
A peak biased dice is then determined by one single parameter $\theta_{bias} > \frac{1}{m}$, with $\theta_1 = \theta_{bias}$ and $\theta_j = \frac{1-\theta_{bias}}{m-1}, \forall j, 1 \leq j \leq m$.
The problem is to determine whether a dice is biased given a limited sample of multinomial trials.
Given a sample $x^n$, we compute the coding length of this sample using either the NML or the enumerative code and decide that the dice is biased if its coding length is shorter than that of the random code ($n\log m$). For a given size $n$ and a code (e.g. enumerative or NML), we compute the probability of detecting the biased dice by averaging the detection over all the possible samples of size $n$.
Formally, we thus compute:
\begin{eqnarray}
\label{eqn_biasedDice}
prob^D (\theta_{bias}, n)
&=& \mathrm{E}_{M(\theta_{bias})}
\left( \mathbb{1}_{ \left\{ L\left(\widehat{\theta}(x^n), x^n\right) < n \log m \right\} } \right) \\
&=& \sum_{\{n_1+\ldots+n_m=n\}} {\frac{n!}{n_1! n_2! \ldots n_m!} \theta_{bias}^{n_1} (\frac{1-\theta_{bias}}{m-1})^{n-n_1}
\mathbb{1}_{ \left\{ L\left(\widehat{\theta}(x^n), x^n\right) < n \log m \right\} } }. \nonumber
\end{eqnarray}
The issue is to be able to detect a biased dice with the minimum sample size.
Using Formula~\ref{eqn_biasedDice}, we then determine the first value of $n$ where the probability of detecting the biased dice is beyond $50\%$:
\begin{eqnarray}
\label{eqn_thresholdBiasedDice}
\underline{n}^{\; D} (\theta_{bias})
&=& \min_{n \geq 10} \{prob^D (\theta_{bias}, n) \geq 50\% \}.
\end{eqnarray}
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{StudyBiasedDice3DetectionSizeThreshold.pdf}
\includegraphics[width=0.49\textwidth]{StudyBiasedDice5DetectionSizeThreshold.pdf}
\caption{Minimum sample size to detect a biased dice (left: $m=3$; right: $m=5$) with probability greater than $50\%$.}
\label{fig_thresholdBiasedDice}
\end{center}
\end{figure}
Figure~\ref{fig_thresholdBiasedDice} shows the detection thresholds computed using the NML or enumerative codes for dices with $m=3$ and $m=5$.
As expected, the minimum number of trials necessary to detect a biased dice increases when $\theta_{bias}$ decreases.
Although all the thresholds are quite close, the enumerative code always needs smaller sample sizes to detect the biased dice. For example, for $m$=5, the enumerative code needs around $40\%$ less samples than the NML code to detect a biased dice with $\theta_{bias} \approx 0.3$.
According to the experiment, the relative difference decreases as the detection threshold increases, but this could not be studied further for heavy computational reasons.
\subsection{Biased versus fair dice classification}
\label{secDiceClassification}
Like in the case of Bernoulli distributions, the enumerative code compresses most m-ary strings slightly better than the NML code, resulting in a better sensitivity to biased dices at the expense of more false detections in case of fair dices.
Interestingly, the difference of behavior between the two codes increases for larger $m$.
We do not extend the coin classification experiment (see Section~\ref{secCoinClassification}) to dices, because there are multiple free parameters to define biased dice and because the calculation of the expectation of accuracy is too computationally intensive.
Still, we expect that the results might be similar, with overall a similar behavior w.r.t. the detection of biased dice, but better detection for the enumerative code in the non asymptotic case for small bias.
\section{Conclusion}
\label{secConclusion}
In this paper, we have revisited the enumerative two-part crude MDL code for the Bernoulli model, which compares favorably with the alternative standard NML code.
We have suggested a Bayesian interpretation of the enumerative code, that relies on models for finite size samples and results in a discrete definition of the likelihood of the data given the model parameter.
We have shown that the coding length of the model parameter is exactly the same as the model complexity computed by applying the NML formula using the definition of the enumerative maximum likelihood. This means that the enumerative code is both a one-part and two part code, which brings parametrization independence, optimality and simplicity.
Surprisingly, the obtained parametric complexity is twice that of the alternative classical NML code or the standard BIC regularization term.
The enumerative code has a direct interpretation in terms of two part codes for finite sample data. The model parameter is encoded using a uniform prior w.r.t. the sample size and the data are also encoded using a uniform prior among all the binary strings of given size that can be generated using the model parameter.
This explains why the enumerative code provides a more parsimonious encoding of the data given the parameter, which compensates the larger model complexity term.
Experimental comparisons between the enumerative and NML codes show that they are very similar, with small differences only. Under the uniform distribution, the enumerative code compresses most individual sequences slightly better, resulting in a slightly better compression on average. An application to the detection of biased coins demonstrates that the enumerative code has a better sensitivity to biased coins at the expense of more false detections in case of fair coins, but the differences are small and vanish asymptotically.
Extension to the multinomial model is also presented. Using the same approach, we obtain a very simple and interpretable analytic formula for the parametric complexity term, that once again is approximately twice that of the alternative classical NML code or the standard BIC regularization term.
The resulting code, both one-part and two-part, is optimal w.r.t. NML approach and parameterization invariant, with a much simpler parametric complexity term.
It compresses most strings better than the ``standard'' NML code with a constant margin and extremely few heavily unbalanced strings with a margin logarithmic in the sample size.
Experimental comparisons extend the results obtained with Bernoulli distributions.
Both codes are very similar, with small differences that roughly increase linearly with the number of model parameters.
Altogether, the theoretical and experimental results suggest that one might use the enumerative code rather than NML in practice, for Bernoulli and multinomial distributions.
\bibliographystyle{apalike}
|
\section{Introduction Version2}
\section*{Conventions}
We write $P$ or $P_{x}$ for the polynomial ring $k[x_1, \dots, x_n]$ and $\mathfrak{m}_{0}$ for the ideal $(x_1, \dots, x_n) \subset P$. We write $P_{y}$ for $k[y_1, \dots, y_n]$. We write $\widetilde{P}$ for the graded ring $k[X_0, \dots, X_n]$ with grading $\operatorname{deg}(X_i)=1$. We then have $\mathbf{P}_{k}^{n} = \operatorname{Proj} \widetilde{P}$
If $f \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$ is a polynomial function, then we write $f_1, \dots, f_n \in P_{x}$ for the components of $f$. We say a polynomial $f \in P$ has an isolated zero at a closed point $x \in \mathbf{A}^{n}_{k}$ if the local algebra $Q_{x}(f) := P_{\mathfrak{m}_{x}}/(f_1, \dots, f_n)$ has finite length. We say that a closed point $x$ of $\mathbf{A}^n_k$ is isolated in its fiber $f^{-1}(f(x))$ if $x$ is a connected component $f^{-1}(f(x))$, or equivalently, if there is a Zariski open neighborhood of $U$ of $x$ in $\mathbf{A}^n_k$ such that $f$ maps $U -\{x\}$ to $\mathbf{A}^n_k - \{f(x) \}$. Note that if $f$ has an isolated zero at the origin, then $Q_{0}(f)$ has dimension $0$, which implies that the connected component of $f^{-1}(0) \cong \operatorname{Spec} P/(f_1, \dots, f_n)$ containing $0$ contains no other points, whence $0$ is isolated in its fiber.
Using homogeneous coordinates $[X_0, X_1, \ldots, X_n]$ for $\mathbf{P}^n$, we use $\mathbf{A}^n$ to denote the open subscheme of $\mathbf{P}^n$ where $X_0 \neq 0$, and $\mathbf{P}^{n-1}$ to denote its closed complement isomorphic to $\mathbf{P}^{n-1}$.
For a vector bundle $\mathcal{E}$ on a smooth scheme $X$, let $\operatorname{Th}(\mathcal{E})$ denote the Thom space of $\mathcal{E}$ of Section 3 Definition 2.16 of \cite{morelvoevodsky1998}, i.e., $\operatorname{Th}(\mathcal{E})$ is the pointed sheaf $$\operatorname{Th}(\mathcal{E}) = \mathcal{E}/(\mathcal{E} - z(X)),$$ where $z: X \to \mathcal{E}$ denotes the zero section.
It will be convenient to work in the stable $\mathbf{A}^1$-homotopy category $\cat{Spt}(B)$ of $\mathbf{P}^1$-spectra over $B$, where $B$ is a finite type scheme over $k$. Most frequently, $B=L$, where $L$ is a field extension of $k$. The notation $[-,-]_{\cat{Spt}(B)}$ will be used for the morphisms. $\cat{Spt}(B)$ is a symmetric monoidal category under the smash product $\wedge$, with unit $1_B$, denoting the sphere spectrum. Any pointed simplicial presheaf $X$ determines a corresponding $\mathbf{P}^1$-suspension spectrum $\Sigma^{\infty} X$. For example, $\Sigma^{\infty} \operatorname{Spec} L_+ \cong 1_L$ and $\Sigma^{\infty} (\mathbf{P}^1_L)^{\wedge n}$ is a suspension of $1_L$. When working in $\cat{Spt}(L)$, we will identify pointed spaces $X$ with their suspension spectra $\Sigma^{\infty} X$, omitting the $\Sigma^{\infty}$. We will use the six operations $(p^*, p_*, p_!, p^!, \wedge, \operatorname{Hom})$ given by Ayoub \cite{Ayoub_sixop1} and developed by Ayoub, and Cisinksi-D\'eglise \cite{CD-triang_cat_mixed_motives}. There is a nice summary in \cite[\S 2]{Hoyois_lef}. We use the following associated notation and constructions. When $p: X \to Y$ is smooth, $p^*$ admits a left adjoint, denoted $p_{\sharp}$, induced by the forgetful functor $\mathrm{Sm}_{X} \to \mathrm{Sm}_{Y}$ from smooth schemes over $X$ to smooth schemes over $Y$. For $p:X\to \operatorname{Spec} L$ a smooth scheme over $L$, the suspension spectrum of $X$ is canonically identified with $p_!p^!1_L$ as an object of $\cat{Spt}(L)$. For a vector bundle $p:E \to X$, the Thom spectrum $\Sigma^{\infty} \operatorname{Th}(E)$ (or just $\operatorname{Th}(E)$) is canonically identified with $s^*p^! 1_X$. Let $\Sigma^E = s^* p^!: \cat{Spt}(X) \to \cat{Spt}(X)$. Let $e: E \to X$ and $d: D \to Y$ be two vector bundles over smooth schemes $p: X \to \operatorname{Spec} L$ and $q:Y \to \operatorname{Spec} L$ over $L$. Given a map $f: Y \to X$ and a monomorphism $\phi: D \hookrightarrow f^* E$, there is an associated natural transformation $$\operatorname{Th}_f \phi : q_! \Sigma^D q^! \to p_! \Sigma^E p^!$$ of endofunctors on $\cat{Spt}(L)$ inducing the map on Thom spectra. The natural transformation $\operatorname{Th}_f \phi $ is defined as the composition \begin{equation}\label{thfphi=comp}\operatorname{Th}_f \phi = \operatorname{Th}_f {1_{f^*E}} \circ \operatorname{Th}_{1_Y} \phi.\end{equation} The natural transformation $ \operatorname{Th}_{1_Y} \phi$ is the composition $$ t^*d^! \cong t^* \phi^!e^!\to t^* \phi^* e^! \cong s^* e^!,$$ where $t: Y \to D$ denotes the zero section of $D$, $s: X \to E$ denotes the zero section of $E$, and the middle arrow is induced by the exchange transformation $ \phi^! \cong 1^* \phi^* \to 1^! \phi^* \cong \phi^*$. The natural transformation $\operatorname{Th}_f {1_{f^*E}}$ is the composition \begin{equation}\label{Thf1_as_counit} \operatorname{Th}_f 1: q_! \Sigma^{f^* E} q^! \cong p_! f_! \Sigma^{f^*E} f^! p^! \cong p_!\Sigma^E f_! f^! p^! \stackrel{\epsilon}{\rightarrow} p_! \Sigma^E p^!,\end{equation} where $\epsilon: f_! f^! \to 1$ denotes the counit.
\section{The Grothendieck--Witt class of Eisenbud--Khimshiashvili--Levine} \label{Section: LocalForm}
In this section we recall the definition of the Grothendieck--Witt class $w_{0}(f)$ studied by Eisenbud--Khimshiashvili--Levine. We compute the class when $f$ has a nondegenerate zero and when $f$ is the gradient of the equation of an ADE singularity. Here $f \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$ is a polynomial function with an isolated zero at the origin. We write $f_1, \dots, f_n \in P$ for the components of $f$.
\begin{df} \label{Definition: LocalAlgbra}
Suppose that $x \in \mathbf{A}^{n}_{k}$ is a closed point such that $y=f(x)$ has residue field $k$. Writing the maximal ideal of $x$ and $y$ respectively as $\mathfrak{m}_{x}$ and $\mathfrak{m}_{y}= (y_1-\overline{b}_{1}, \dots, y_n-\overline{b}_{n})$, we define the \textbf{local algebra} $Q_{x}(f)$ of $f$ at a closed point $x$ to be $P_{\mathfrak{m}_{x}}/(f_1-\overline{b}_{1}, \dots, f_n-\overline{b}_{n})$. We also write $Q$ for $Q_{0}(f)$, the local algebra at the origin.
The \textbf{distinguished socle element} at the origin $E=E_{0}(f) \in Q_{0}(f)$ is
\[
E_{0}(f) := \det \begin{pmatrix} a_{i, j} \end{pmatrix}
\]
for $a_{i, j} \in P$ polynomials satisfying
\[
f(x) = f(0)+\sum_{j=1}^{n} a_{i, j} x_{j}.
\]
The \textbf{Jacobian element} at the origin $J=J_{0}(f) \in Q_{0}(f)$ is
\[
J_{0}(f) := \det( \frac{\partial f_i}{\partial x_j}).
\]
\end{df}
\begin{rmk}
The elements $E$ and $J$ are related by $J = (\operatorname{rank}_{k} Q_{0}(f)) \cdot E$ \cite[(4.7)~Korollar]{scheja}. From this, we see that the two elements contain essentially the same information when $k=\mathbf{R}$ (the case studied in \cite{eisenbud77}), but $E$ contains more information when the characteristic of $k$ divides the rank of $Q_{0}(f)$.
\end{rmk}
\begin{lm} \label{Lemma: SocleElement}
If $f$ has an isolated zero at the origin, then the socle of $Q_{0}(f)$ is generated by $E$.
\end{lm}
\begin{proof}
Since $Q$ is Gorenstein (by e.g.~Lemma~\ref{Lemma: RelativeCompleteIntersection}) with residue field $k$, the socle is $1$-dimensional, so it is enough to prove that $E$ is nonzero and in the socle. This follows from the proof of \cite[(4.7)~Korolllar]{scheja}. In the proof, Scheja--Storch show that $E = \Theta(\pi)$ for $\pi \colon Q \to k$ the evaluation function $\pi(a) = a(0)$ and $\Theta \colon \operatorname{Hom}_{k}(Q, k) \cong Q$ a certain isomorphism of $Q$-modules (for the $Q$-module structure on $\operatorname{Hom}_{k}(Q, k)$ defined by $(a \cdot \phi)(b)=\phi(a \cdot b)$). The maximal ideal $\mathfrak{m}_{0}=(x_{1}, \dots, x_{n})$ of $Q$ must annihilate $E$ since the ideal annihilates $\pi$, and so $E$ lies in the socle. Furthermore, $E$ is nonzero since $\pi$ is nonzero.
\end{proof}
\begin{df}
If $\phi \colon Q_{0}(f) \to k$ is a $k$-linear function, then we define a symmetric bilinear form $\beta_{\phi} \colon Q \times Q \to k$ by $\beta_{\phi}(a_1, a_2) := \phi(a_1 \cdot a_2)$.
\end{df}
\begin{lm} \label{Lemma: WhenFormsEqual}
If $\phi_1$ and $\phi_2$ are $k$-linear functions satisfying $\phi_1(E) = \phi_2(E) \text{ in $k/(k^{\ast})^{2}$}$, then $\beta_{\phi_{1}}$ is isomorphic to $\beta_{\phi_{2}}$. Furthermore, if $\phi(E) \ne 0$, then $\beta_{\phi}$ is nondegenerate.
\end{lm}
\begin{proof}
Since $E$ generates the socle, the result follows from \cite[Propositions~3.4, 3.5]{eisenbud77}.
\end{proof}
\begin{df} \label{Definition: EKLform}
The Grothendieck--Witt \textbf{class of Eisenbud--Khimshiashvili--Levine} or the \textbf{EKL class} is $w = w_{0}(f) \in \operatorname{GW}(k)$ is the Groethendieck--Witt class of $\beta_{\phi}$ for any $k$-linear function $\phi \colon Q \to k$ satisfying $\phi(E) = 1$.
\end{df}
Recall that the Grothendieck--Witt group $\operatorname{GW}(k)$ of $k$ is the groupification of the monoid of nondegenerate symmetric bilinear forms \cite[Definition~1.1]{lam05}. The Grothendieck--Witt class $w_{0}(f)$ is independent of the choice of $\phi$ by Lemma~\ref{Lemma: WhenFormsEqual}.
In this paper we focus on the class $w_{0}(f)$, but in work recalled in Section~\ref{Section: Family}, Scheja--Storch constructed a distinguished symmetric bilinear form $\beta_0$ that represents $w_{0}(f)$. This symmetric bilinear form encodes more information than $w_{0}(f)$ when $f$ is a polynomial in $1$ variable, and we discuss this topic in greater detail in \cite[Section~4]{wickelgren16}.
To conclude this section, we explicitly describe some ELK classes. The descriptions are in terms of the following classes.
\begin{df}
Given $\alpha_1, \dots, \alpha_{m} \in k^{\ast}$, we define $\langle \alpha_{1}, \dots, \alpha_{m} \rangle \in \operatorname{GW}(k)$ to be the class of the symmetric bilinear form
\begin{gather*}
\beta \colon k^{\oplus m} \times k^{\oplus m} \to k,\\
\beta((a_{1}, \dots, a_{m}), (b_1, \dots, b_m)) = \alpha_{1} \cdot a_{1} b_{1} + \dots + \alpha_{m} \cdot a_{m} b_{m}.
\end{gather*}
The standard hyperbolic form $\mathbf{H}$ is the symmetric bilinear form
\begin{gather*}
\beta \colon k^{\oplus 2} \times k^{\oplus 2} \to k,\\
\beta((a_{1}, a_{2}), (b_1, b_2)) = a_{1} b_{2} + a_{2} b_{1}.
\end{gather*}
The class of $\mathbf{H}$ equals $\langle 1, -1 \rangle$ in $\operatorname{GW}(k)$.
\end{df}
The following lemma describes $w_{0}(f)$ when $f$ has a simple zero.
\begin{lm}
If $f$ has a simple zero at the origin, then $w_{0}(f) = \left\langle \det \frac{\partial f_i}{\partial x_j}(0) \right\rangle$.
\end{lm}
\begin{proof}
We have $Q_{0}(f) = k$ and $E = \det \frac{\partial f_i}{\partial x_j}(0)$. The element $E \in Q_{0}(f)$ is then a $k$-basis, and $w_{0}(f)$ is represented by the form $\beta_{\phi}$ satisfying
\begin{align*}
\beta_{\phi}(E, E) =& \phi( \det \frac{\partial f_i}{\partial x_j}(0) \cdot E) \\
=& \det \frac{\partial f_i}{\partial x_j}(0).
\end{align*}
\end{proof}
When $f$ has an arbitrary isolated zero, the following procedure computes $w_{0}(f)$.
\begin{enumerate}
\item Compute a standard basis for $(f_1,\dots, f_n)$ and a $k$-basis for $Q_{0}(f)$.
\item Express $E$ in terms the $k$-basis by performing a local division.
\item Define an explicit $k$-linear function $\phi \colon Q_{0}(f) \to k$ satisfying $\phi(E)=1$ using the $k$-basis.
\item For every pair $b_i, b_j$ of basis elements, express $b_i \cdot b_j$ in terms of the $k$-basis by performing a local division and then use that expression to evaluate $\phi( b_{i} \cdot b_{j})$.
\end{enumerate}
The matrix with entries $\phi( b_{i} \cdot b_{j} )$ is the Gram matrix of a symmetric bilinear form that represents $w_{0}(f)$.
Table~\ref{Table: ADEsing} describes some classes that were computing using this procedure. The table should be read as follows. The second column displays a polynomial $g$, namely the polynomial equation of the ADE singularity named in the first column. The associated gradient $\operatorname{grad}(g) := (\frac{\partial g}{\partial x}, \frac{\partial g}{\partial y})$ is a polynomial function $\mathbf{A}^2_{\mathbf{Q}} \to \mathbf{A}^{2}_{\mathbf{Q}}$ with an isolated zero at the origin, and the third column is its ELK class $w_{0}(\operatorname{grad}(g)) \in \operatorname{GW}(\mathbf{Q})$. (We consider $g$ as a polynomial with rational coefficients.)
\begin{table}[htdp] \label{Table: ADEsing}
\caption{ELK classes for ADE singularities}
\begin{center}
\begin{tabular}{l l l}
Singularity & Equation $g$ & $w_{0}(\operatorname{grad}(g)) \in \operatorname{GW}(\mathbf{Q})$ \\
\hline \hline \\
$A_{n}$, $n$ odd & $x_1^2+x_2^{n+1}$ & $ \frac{n-1}{2} \cdot \mathbb{H}+\left \langle 2(n+1) \right \rangle$ \\
$A_{n}$, $n$ even & $x_1^2+x_2^{n+1}$ & $\frac{n}{2} \cdot \mathbb{H}$ \\
$D_{n}$, $n$ even & $x_{2} (x_{1}^{2}+x_{2}^{n-2})$ & $\frac{n-2}{2} \cdot \mathbb{H}+\left\langle -2, 2(n-1) \right\rangle$ \\
$D_{n}$, $n$ odd & $x_{2} (x_{1}^{2}+x_{2}^{n-2})$ & $\frac{n-1}{2} \cdot \mathbb{H}+\langle -2 \rangle$ \\
$E_{6}$ & $x_{1}^{3}+x_{2}^{4}$ & $3 \cdot \mathbb{H}$ \\
$E_{7}$ & $x_{1} (x_{1}^{2}+x_{2}^{3})$ & $3 \cdot \mathbb{H}+\left\langle -3 \right\rangle $ \\
$E_{8}$ & $x_{1}^{3}+x_{2}^{5}$ & $4 \cdot \mathbb{H}$
\end{tabular}
\end{center}
\end{table}
\section{Local $\mathbf{A}^1$-Brouwer degree} \label{Section: LocalDegree}
Morel's $\mathbf{A}^1$-Brouwer degree homomorphism \begin{equation*}
\deg: [(\mathbf{P}_k^1)^{\wedge n},(\mathbf{P}_k^1)^{\wedge n}] \to \mathrm{GW}(k)
\end{equation*} gives rise to a notion of local degree, which we describe in this section. We then show that the degree is the sum of local degrees under appropriate hypotheses (Proposition \ref{pr_deg_is_sum_local_deg}), and that when $f$ is \'etale at $x$, the local degree is computed by $\deg^{\mathbf{A}^1}_x f =\operatorname{Tr}_{k(x)/k} \langle J (x) \rangle$, where $J(x)$ denotes the Jacobian determinant $J = \det \begin{pmatrix} \frac{\partial f_i}{\partial x_j} \end{pmatrix}$ evaluated at $x$ (Proposition \ref{loc_degree_etale_point}). For endomorphisms of $\mathbf{P}^1_k$, these notions and properties are stated in \cite{Morel_motivicpi0_sphere} \cite{morel06}, and build on ideas of J. Lannes.
To motivate the definition, recall that to define the local topological Brouwer degree of $f: \mathbf{R}^n \to \mathbf{R}^n$ at a point $x$, one can choose a sufficiently small $\epsilon>0$ and take the $\mathbf{Z}$-valued topological degree of the map $$\xymatrix{S^{n-1} \cong \{ z : \| z - x \| = \epsilon \} \ar[rrr]^{\frac{f - f(x)}{\| f - f(x) \|}} &&& \{ z : \| z \| =1 \} \cong S^{n-1} }.$$ By translation and scaling, the map $\frac{f - f(x)}{\| f - f(x) \|}$ can be replaced by the map induced by $f$ from the boundary $\partial B(x, \epsilon)$ of a small ball $B(x, \epsilon)$ centered at $x$ to a boundary $\partial B(f(x), \epsilon')$ of a small ball centered at $f(x)$. The suspension of this map can be identified with map induced by $f$ \begin{equation}\label{loc_deg_ball_relative_map}f: \frac{B(x, \epsilon)}{\partial B(x, \epsilon)} \to \frac{B(f(x), \epsilon')}{\partial B(f(x), \epsilon')},\end{equation} from the homotopy cofiber of the inclusion $\partial B(x, \epsilon) \to B(x, \epsilon) $ to the analogous homotopy cofiber. As $\frac{B(x, \epsilon)}{\partial B(x, \epsilon)}$ is also the homotopy cofiber of $B(x, \epsilon) -\{x\}\to B(x, \epsilon) $, we are free to use the latter construction for the (co)domain in \eqref{loc_deg_ball_relative_map}: \begin{equation}\label{loc_deg_ball_rel2}f: \frac{B(x, \epsilon)}{B(x, \epsilon)-\{x\}} \to \frac{B(f(x), \epsilon')}{B(f(x), \epsilon') - \{f(x)\}}.\end{equation}
In $\mathbf{A}^1$-algebraic topology, the absence of small balls around points whose boundaries are spheres makes the definition of local degree using the map $\frac{f - f(x)}{\| f - f(x) \|}$ problematic. However, the map \eqref{loc_deg_ball_rel2} generalizes to a map between spheres by Morel and Voevodsky's Purity Theorem. This allows us to define a local degree when $x$ and $f(x)$ are both rational points, as in the definition of $f_x'$ given below. When $x$ is not rational, we precompose with the collapse map from the sphere $\mathbf{P}^n_k / \mathbf{P}^{n-1}_k \to \mathbf{P}^n_k /\mathbf{P}^n_k -\{x\}$ to obtain Definition \ref{def:local_degree_f(x)_rational}. This is shown to be compatible with the former definition (Proposition \ref{local_degree_alternate_def}).
We now give Definition \ref{def:local_degree_f(x)_rational}, first introducing the necessary notation.
By \cite[Proposition 2.17 numbers 1 and 3, p. 112]{morelvoevodsky1998}, there is a canonical $\mathbf{A}^1$-weak equivalence $(\mathbf{P}_k^1)^{\wedge n} \cong \mathbf{P}_k^n/\mathbf{P}_k^{n-1}$ as both can be identified with the Thom space $\operatorname{Th}(\mathcal{O}_k^n)$ of the trivial rank $n$ bundle on $\operatorname{Spec} k$. Thus we may take the degree of a map $ \mathbf{P}_k^n/\mathbf{P}_k^{n-1} \to \mathbf{P}_k^n/\mathbf{P}_k^{n-1}$ in the homotopy category.
Let $x$ be a closed point of $\mathbf{A}^n_k$, and let $f: \mathbf{A}^n_k \to \mathbf{A}^n_k$ be a function such that $x$ is isolated in its fiber $f^{-1}(f(x))$. Choose a Zariski open neighborhood $U$ of $x$ such that $f$ maps $U - \{x\} $ into $\mathbf{A}^n_k - \{ f(x)\}$. The Nisnevich local homotopy push-out diagram $$\xymatrix{ U -\{x\} \ar[r] \ar[d] & \mathbf{P}_k^{n} - \{x\} \ar[d] \\ U \ar[r] & \mathbf{P}_k^n} $$ induces a canonical homotopy equivalence $U/(U - \{ x\}) \to \mathbf{P}_k^n/\mathbf{P}_k^{n} - \{x\}$.
There is a trivialization of $T_x \mathbf{P}^n_k$ coming from the isomorphism $T_x \mathbf{P}_k^n \cong T_x \mathbf{A}_k^n$ and the canonical trivialization of $T_x \mathbf{A}^n$. Purity thus induces an $\mathbf{A}^1$-weak equivalence $ \mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{x\}) \cong \operatorname{Th}(\mathcal{O}^n_k) $. As above, \cite[Proposition 2.17 number 3, p.112]{morelvoevodsky1998} gives a canonical $\mathbf{A}^1$-weak equivalence $\operatorname{Th}(\mathcal{O}^n_k) \cong \mathbf{P}_k^n/(\mathbf{P}_k^{n-1})$. Let $r: \mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{x\}) \stackrel{\cong}{\to} \mathbf{P}_k^n/(\mathbf{P}_k^{n-1})$ denote the composite $\mathbf{A}^1$-weak equivalence.
\begin{lm}\label{crush=id}
For any $k$-point $x$ of $\mathbf{A}^n_k$, the composition $$ c_x: \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \to \mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{x\}) \cong \mathbf{P}_k^n/(\mathbf{P}_k^{n-1})$$ of the collapse map with $r$ is $\mathbf{A}^1$-homotopy equivalent to the identity.
\end{lm}
For $n = 1$, this is \cite[Lemma 5.4]{Hoyois_lef}, and the proof generalizes to the case of larger $n$, the essential content being \cite[Lemma 2.2]{Voevodsky_MCZ2}.
\begin{proof}
Suppose $x$ has homogeneous coordinates $[1, a_1, \ldots, a_n]$, and let $[X_0, X_1, \ldots, X_n]$ denote homogeneous coordinates on $\mathbf{P}^n_k.$ Let $f: \mathbf{P}_k^n \to \mathbf{P}_k^n$ be the automorphism $$f([X_0, X_1, \ldots, X_n] ) = [X_0, X_1 + a_1 X_0, \ldots, X_n + a_n X_0].$$ The diagram $$ \xymatrix{ \mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{0\}) \ar[d]_f \ar[r]_{\cong}^{r} & \ar[d]^{1} \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \\
\mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{x\})\ar[r]_{\cong}^{r} & \mathbf{P}_k^n/(\mathbf{P}_k^{n-1})}$$ commutes by naturality of Purity \cite[Lemma 2.1]{Voevodsky_MCZ2} and the compatibility of the trivializations of $T_x \mathbf{P}^n_k$ and $T_0 \mathbf{P}^n_k$. The diagram $$\xymatrix { \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \ar[d]_{\overline{f}} \ar[r] &\mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{0\}) \ar[d]^f \\
\mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \ar[r] &\mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{x\}) } $$ comparing collapse maps via the maps induced by $f$ commutes by definition. Since $$[X_0, X_1, \ldots, X_n] \times t \mapsto [X_0, X_1 + a_1 tX_0, \ldots, X_n + a_nt X_0] $$ defines a naive homotopy between $\overline{f}$ and the identity, it suffices to show the lemma when $x$ is the origin. This case follows from \cite[Lemma 2.2]{Voevodsky_MCZ2} and \cite[Proposition 2.17 proof of number 3, p. 112]{morelvoevodsky1998}.
\end{proof}
In particular, the collapse map $\mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \to \mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{x\}) $ is a homotopy equivalence.
\begin{df}\label{def:local_degree_f(x)_rational}
The {\em local degree} (or local $\mathbf{A}^1$-Brouwer degree) $\deg^{\mathbf{A}^1}_x f$ of $f$ at $x$ is Morel's $\mathbf{A}^1$-degree homomorphism applied to a map $$ f_x: \mathbf{P}_k^n/\mathbf{P}_k^{n-1} \to \mathbf{P}_k^n/\mathbf{P}_k^{n-1} $$ in the homotopy category, where $f_x$ is defined to be the composition
$$ \mathbf{P}_k^n/\mathbf{P}_k^{n-1} \to \mathbf{P}_k^n/\mathbf{P}_k^{n} - \{x\} \stackrel{\cong}{\leftarrow} U/(U - \{ x\}) \stackrel{f\vert_U}{\longrightarrow} \mathbf{P}_k^n/(\mathbf{P}_k^n - \{f(x) \}) \stackrel{\cong}{\leftarrow} \mathbf{P}_k^n/\mathbf{P}_k^{n-1} $$
\end{df}
When $x$ is a $k$-point, it is perhaps more natural to define the local degree in the following equivalent manner: the trivialization of the tangent space of $\mathbf{A}^n_k$ gives canonical $\mathbf{A}^1$-weak equivalences $U/(U - \{ x\}) \cong \operatorname{Th}(\mathcal{O}_k^n)$ and $ \mathbf{A}_k^n/(\mathbf{A}^n_k -\{ f(x)\})\cong \operatorname{Th}(\mathcal{O}_k^n)$ by Purity \cite[Theorem 2.23, p. 115]{morelvoevodsky1998}. As above, we have a canonical $\mathbf{A}^1$-weak equivalence $\operatorname{Th}(\mathcal{O}_k^n) \cong \mathbf{P}_k^n/(\mathbf{P}_k^{n-1})$. The local degree of $f$ at $x$ is the degree of the map in the homotopy category $$f_x': \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \cong U/(U -\{ x\}) \stackrel{f\vert_U}{\longrightarrow} \mathbf{A}_k^n/(\mathbf{A}^n_k -\{ f(x)\})\cong \mathbf{P}_k^n/(\mathbf{P}_k^{n-1})$$ as we now show.
\begin{pr}\label{local_degree_alternate_def}
$$\deg^{\mathbf{A}^1}_x f = \deg f_x'$$
\end{pr}
\begin{proof}
Let $ c_{f(x)}^{-1}$ denote the inverse in the homotopy category of $c_{f(x)}$ as defined in Lemma \ref{crush=id}. The definitions produce the equality $ c_{f(x)}^{-1} f_x' c_x = f_x$, which implies the result by Lemma \ref{crush=id}.
\end{proof}
\begin{lm}\label{collapse_is_unit}
Let $x$ be a closed point of $\mathbf{A}_k^n$. The collapse map $$ \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \to \mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{x\}) \cong \operatorname{Th} T_x \mathbf{P}_k^n \cong \operatorname{Th} T_x \mathbf{A}^n \cong \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \wedge \operatorname{Spec} k(x)_+$$ is $ \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \wedge (-)$ applied to the canonical map \begin{equation}\label{collapse_6_functors}\eta: 1_k \to p_* p^* 1_k \cong p_* 1_{k(x)}\end{equation} in $\cat{Spt}(k)$, where $p:\operatorname{Spec} k(x) \to \operatorname{Spec} k$ is the structure map, and the last equivalence is from \cite[3. Prop 2.17, p. 112]{morelvoevodsky1998}.
\end{lm}
For $n = 1$, this is \cite[Lemma 5.5]{Hoyois_lef}, and Hoyois's proof generalizes to higher $n$ as follows.
\begin{proof}
As above, consider the trivialization of $T_x \mathbf{P}^n_k$ coming from the isomorphism $T_x \mathbf{P}_k^n \cong T_x \mathbf{A}^n$ and the canonical trivialization of $T_x \mathbf{A}^n$. The closed immersion $x: \operatorname{Spec} k(x) \to \mathbf{P}_n^k$ and this trivialization determine a Euclidean embedding in the sense of Hoyois \cite[Definition 3.8]{Hoyois_lef}. This Euclidean embedding determines an isomorphism $ \mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{x\}) \cong \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \wedge p_! 1_{k(x)}$ in $\cat{Spt}(k)$ by \cite[3.9]{Hoyois_lef} and the identification $\operatorname{Th} \mathcal{O}_k^n \cong \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) $ of \cite[3. Prop 2.17, p. 112]{morelvoevodsky1998}. Since $p$ is finite \'etale, there is a canonical equivalence $p_! 1_{k(x)}\cong \operatorname{Spec} k(x)_+$, and these identifications agree with the isomorphism $\mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{x\}) \cong \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \wedge \operatorname{Spec} k(x)_+ $ in the statement of the lemma. By \cite[Proposition 3.14]{Hoyois_lef}, it thus suffices to show that a certain composition \begin{equation}\label{comp_collapse_unit_pf}
\mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \wedge \operatorname{Spec} k(x)_+ \to \mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{x\}) \wedge \operatorname{Spec} k(x)_+ \stackrel{h}{\to} \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \wedge \operatorname{Spec} k(x)_+
\end{equation} of the collapse map $ \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \to \mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{x\})$ smash $\operatorname{Spec} k(x)_+$ with a map $h$ is the identity in $\cat{Spt}(k)$.
To define $h$, introduce the following notation. Let $$x_{k(x)}: \operatorname{Spec} k(x) \otimes k(x) \hookrightarrow \mathbf{P}^n_{k(x)}$$ be the base change of $x$. Let $$\tilde{x} = x_{k(x)} \circ \Delta: \operatorname{Spec} k(x) \to \mathbf{P}_{k(x)}^n$$ be the composition of the diagonal with $x_{k(x)}$. Let $$r: \mathbf{P}_{k(x)}^n/(\mathbf{P}_{k(x)}^{n} - \{\tilde{x} \}) \stackrel{\cong}{\rightarrow} \mathbf{P}_{k(x)}^n/\mathbf{P}_{k(x)}^{n-1}$$ be as Lemma \ref{crush=id} with $k$ replaced by $k(x)$. Using the identifications $\mathbf{P}_k^n/(\mathbf{P}_k^{n} - \{x\}) \wedge \operatorname{Spec} k(x)_+ \cong \mathbf{P}_{k(x)}^n/(\mathbf{P}_{k(x)}^{n} - x_{k(x)})$ and $\mathbf{P}_{k(x)}^n/(\mathbf{P}_{k(x)}^{n-1}) \cong \mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \wedge \operatorname{Spec} k(x)_+$, we can view $h$ as a map $$h:\mathbf{P}_{k(x)}^n/(\mathbf{P}_{k(x)}^{n} - x_{k(x)}) \to \mathbf{P}_{k(x)}^n/(\mathbf{P}_{k(x)}^{n-1}).$$ Then $h$ is the composition \begin{equation*}
\mathbf{P}_{k(x)}^n/(\mathbf{P}_{k(x)}^{n} - x_{k(x)}) \to \mathbf{P}_{k(x)}^n/(\mathbf{P}_{k(x)}^{n} - \{\tilde{x} \}) \stackrel{r}{\to} \mathbf{P}_{k(x)}^n/(\mathbf{P}_{k(x)}^{n-1})
\end{equation*}
Then the composition \eqref{comp_collapse_unit_pf} is identified with $p_{\sharp}$ applied to the composition in Lemma \ref{crush=id} of the collapse map with $r$ for the rational point $\tilde{x}: \operatorname{Spec} k(x) \to \mathbf{A}_{k(x)}^n$, completing the proof by Lemma \ref{crush=id}
\end{proof}
The degree of an endomorphism of $\mathbf{P}_k^n/(\mathbf{P}_k^{n-1})$ is the sum of local degrees under the following hypotheses.
\begin{pr} \label{pr_deg_is_sum_local_deg}
Let $f: \mathbf{P}_k^n \to \mathbf{P}_k^n$ be a finite map such that $f^{-1}(\mathbf{A}_k^n) = \mathbf{A}_k^n$, and let $\overline{f}$ denote the induced map $\mathbf{P}_k^n/(\mathbf{P}_k^{n-1}) \to \mathbf{P}_k^n/(\mathbf{P}_k^{n-1})$. Then for any $k$-point $y$ of $\mathbf{A}_k^n$, $$\deg \overline{f} = \Sigma_{x \in f^{-1}(y)} \deg_x f .$$
\end{pr}
\begin{proof}
By Purity \cite[Theorem 2.23, p. 115]{morelvoevodsky1998}, $\mathbf{P}_k^n/(\mathbf{P}_k^n - f^{-1} \{y \})$ is the Thom space of the normal bundle to $f^{-1} \{y \} \hookrightarrow \mathbf{A}_k^n$. The Thom space of a vector bundle on a disjoint union is the wedge sum of the Thom spaces of the vector bundle's restrictions to the connected components. It follows that the quotient maps $$ \mathbf{P}_k^n/(\mathbf{P}_k^n - f^{-1} \{y \}) \to \mathbf{P}_k^n/(\mathbf{P}_k^n - \{x \})$$ for $x$ in $f^{-1}(y)$ determine an $\mathbf{A}^1$-weak equivalence $$ \mathbf{P}_k^n/(\mathbf{P}_k^n - f^{-1} \{y \}) \to \vee_{x \in f^{-1}(y)} \mathbf{P}_k^n/ (\mathbf{P}_k^n - \{x\}) .$$ There is a commutative diagram \begin{equation*}\xymatrix{ & \vee_{x \in f^{-1}(y)} \mathbf{P}_k^n/ (\mathbf{P}_k^n - \{x\}) \ar[rrd]&& \\ \mathbf{P}_k^n/ (\mathbf{P}_k^n - \{x\}) \ar[ur]^{k_x} & \ar[l] \ar[u]^{\cong} \mathbf{P}_k^n/(\mathbf{P}_k^n - f^{-1} \{y \}) \ar[rr] && \mathbf{P}_k^n/(\mathbf{P}_k^n -\{y \}) \\ & \ar[ul]^{\cong} \ar[u] \mathbf{P}_k^n/ \mathbf{P}_k^{n-1} \ar[rr]_{\overline{f}} && \ar[u]^{\cong} \mathbf{P}_k^n/ \mathbf{P}_k^{n-1} }\end{equation*} Apply $[\mathbf{P}_k^n/ \mathbf{P}_k^{n-1}, -]_{\cat{Spt}(k)}$ to the above diagram, and let $\overline{f}_*$ be the induced map $$\overline{f}_*: [\mathbf{P}_k^n/ \mathbf{P}_k^{n-1}, \mathbf{P}_k^n/ \mathbf{P}_k^{n-1}]_{\cat{Spt}(k)}\to [\mathbf{P}_k^n/ \mathbf{P}_k^{n-1}, \mathbf{P}_k^n/ \mathbf{P}_k^{n-1}]_{\cat{Spt}(k)} .$$ Because the wedge and the product are stably isomorphic, $$[\mathbf{P}_k^n/ \mathbf{P}_k^{n-1}, \vee_{x \in f^{-1}(y)} \mathbf{P}_k^n/ (\mathbf{P}_k^n - \{x\})]_{\mathbf{A}^1} \cong \oplus_{x \in f^{-1}(y)} [\mathbf{P}_k^n/ \mathbf{P}_k^{n-1}, \mathbf{P}_k^n/ (\mathbf{P}_k^n - \{x\})]_{\cat{Spt}(k)}$$ and on the right hand side, $k_x$ induces the inclusion of the summand indexed by $x$. The image of the identity map under $\overline{f}_*$ can be identified with $\deg \overline{f}$. Using the outer composition in the commutative diagram, we see that the image of the identity map under $\overline{f}_*$ can also be identified with $ \Sigma_{x \in f^{-1}(y)} \deg_x f.$
\end{proof}
We now give a computation of the local degree at points where $f$ is \'etale.
\begin{pr}\label{loc_degree_etale_point}
Let $f: \mathbf{A}_k^n \to \mathbf{A}_k^n$ be a morphism of schemes and $x$ be a closed point of $\mathbf{A}_k^n$ such that $f(x) = y$ is $k$-rational and $x$ is isolated in $f^{-1}(y)$. If $f$ is \'etale at $x$, then the local degree is computed by $$\deg^{\mathbf{A}^1}_x f =\operatorname{Tr}_{k(x)/k} \langle J (x) \rangle,$$ where $J(x)$ denotes the Jacobian determinant $J = \det \begin{pmatrix} \frac{\partial f_i}{\partial x_j} \end{pmatrix}$ evaluated at $x$, and $k(x)$ denotes the residue field of $x$.
\end{pr}
\begin{proof}
We work in $\cat{Spt}(k)$. Let $p: \operatorname{Spec} k(x) \to \operatorname{Spec} k$ denote the structure map.
Since $f$ is \'etale at $x$, the induced map of tangent spaces $df(x): T_x \mathbf{A}_k^n \to f^* T_{f(x)} \mathbf{A}^n_k$ is a monomorphism. Thus $df(x)$ induces a map on Thom spectra, which factors as in the following commutative diagram (see Conventions \eqref{thfphi=comp}): \begin{equation}\label{Thdf_CD_in_loc_deg_et_pt_pf}\xymatrix{ \operatorname{Th} T_x \mathbf{A}_k^n \ar[rrrrr]^{\operatorname{Th}_f (df(x))} \ar[d]^{\cong}&&&&& \operatorname{Th} T_{f(x)} \mathbf{A}^n_k \ar[d]^{\cong} \\ \ar[u] \operatorname{Th} \mathcal{O}^n_{\operatorname{Spec} k(x)} \ar[rrr]^{\operatorname{Th}_{1_{\operatorname{Spec} k(x)}} \begin{pmatrix} \frac{\partial f_i}{\partial x_j} \end{pmatrix}} &&& \ar[rr]^{\operatorname{Th}_p 1_{p^*\mathcal{O}^n_{\operatorname{Spec} k} }} \operatorname{Th} \mathcal{O}^n_{\operatorname{Spec} k(x)} && \ar[u] \operatorname{Th} \mathcal{O}^n_{\operatorname{Spec} k}} \end{equation}
The naturality of the Purity isomorphism \cite[Lemma 2.1]{Voevodsky_MCZ2} gives the commutative diagram \begin{equation}\label{purity_in_loc_deg_etale_pt_pf}
\xymatrix{ U/(U-\{x\}) \ar[rr]^{f\vert_U} && \mathbf{P}^n_k / (\mathbf{P}^n_k - \{ f(x)\}) \\
\operatorname{Th} T_x U \ar[rr]^{\operatorname{Th}_f (df(x))} \ar[u]^{\cong}&& \operatorname{Th} T_{f(x)} \mathbf{P}^n_k \ar[u]^{\cong}.}
\end{equation}
The isomorphisms $T_x U \cong T_x \mathbf{A}_k^n$ and $\operatorname{Th} T_{f(x)} \mathbf{A}^n_k \cong \operatorname{Th} T_{f(x)} \mathbf{P}^n_k $ allow us to stack Diagram \eqref{purity_in_loc_deg_etale_pt_pf} on top of Diagram \eqref{Thdf_CD_in_loc_deg_et_pt_pf}. We then expand the resulting diagram to express the map $f_x$ from Definition \ref{def:local_degree_f(x)_rational} in terms of $\begin{pmatrix} \frac{\partial f_i}{\partial x_j} \end{pmatrix}$.
\begin{equation}
\xymatrix{ \mathbf{P}^n_k/\mathbf{P}^{n-1}_k \ar@/^2pc/[rrrrr]^-{f_x} \ar[r] \ar[rrddd]_{ \mathbf{P}^n_k/\mathbf{P}^{n-1}_k\wedge \eta }&\mathbf{P}^n_k/\mathbf{P}^n_k - \{ x\} &\ar[l]^{\cong }U/(U-\{x\}) \ar[rr]^{f\vert_U} && \mathbf{P}^n_k / (\mathbf{P}^n_k - \{ f(x)\}) & \ar[l]^{\cong} \mathbf{P}^n_k/\mathbf{P}^{n-1}_k \ar[dddl]^1\\
& & \operatorname{Th} T_x U \ar[rr]^{\operatorname{Th}_f (df(x))} \ar[u]^{\cong}&& \operatorname{Th} T_{f(x)} \mathbf{P}^n_k \ar[u]^{\cong} &\\
& &\operatorname{Th} \mathcal{O}^n_{\operatorname{Spec} k(x)} \ar[u]^{\cong} \ar[r] & \ar[r] \ldots& \operatorname{Th} \mathcal{O}^n_{\operatorname{Spec} k} \ar[u]^{\cong} &\\
& & \mathbf{P}^n_k/\mathbf{P}^{n-1}_k \wedge \operatorname{Spec} k(x)_+ \ar[u]^{\cong}&& \mathbf{P}^n_k/\mathbf{P}^{n-1}_k \ar[u]^{\cong} &.}
\end{equation}
We have applied Lemma \ref{collapse_is_unit} to identify the diagonal maps.
We furthermore have an identification (see Conventions \eqref{Thf1_as_counit}) of $\operatorname{Th}_p 1_{p^*\mathcal{O}^n_{\operatorname{Spec} k} }$ with the composition \begin{equation*}
p_! \Sigma^{p^* \mathcal{O}^n_{\operatorname{Spec} k}} 1_{k(x)} \to \Sigma^{\mathcal{O}^n_{\operatorname{Spec} k}} p_! 1_{k(x)} \stackrel{\epsilon}{\rightarrow} \Sigma^{\mathcal{O}^n_{\operatorname{Spec} k}} 1_k .
\end{equation*} Since $p$ is \'etale, there is a canonical identification $p_! \cong p_{\sharp}$.
We may therefore identify $f_x$ with the composition
$$\mathbf{P}^n_k/ \mathbf{P}^{n-1}_k \stackrel{\mathbf{P}^n_k/ \mathbf{P}^{n-1}_k \wedge \eta}{\rightarrow} \mathbf{P}^n_{k(x)}/ \mathbf{P}^{n-1}_{k(x)} \cong \operatorname{Th} \mathcal{O}^n_{\operatorname{Spec} k(x)} \stackrel{\begin{pmatrix} \frac{\partial f_i}{\partial x_j} \end{pmatrix}}{\to} \operatorname{Th} \mathcal{O}^n_{\operatorname{Spec} k(x)} \cong \mathbf{P}^n_{k(x)}/ \mathbf{P}^{n-1}_{k(x)} \stackrel{\mathbf{P}^n_k/ \mathbf{P}^{n-1}_k \wedge \epsilon}{\rightarrow} \mathbf{P}^n_k/ \mathbf{P}^{n-1}_k$$
By \cite[Lemma 5.3]{Hoyois_lef}, we therefore have that $f_x $ is the trace of the endomorphism $\begin{pmatrix} \frac{\partial f_i}{\partial x_j} \end{pmatrix}$ of $\operatorname{Th} \mathcal{O}^n_{\operatorname{Spec} k(x)}$ in $\cat{Spt}(k)$. By \cite[Theorem 1.9]{Hoyois_lef}, it follows that $\deg f_x = \operatorname{Tr}_{k(x)/k} \langle J \rangle$.
\end{proof}
\section{Some finite determinacy results } \label{Section: FinitelyDetermined}
Here we prove a finite determinacy result and then use that result to prove a result, Proposition~\ref{Prop: ReduceToRegSeq_enhanced}, that allows us to reduce the proof of the Main Theorem to a case where $f$ is \'{e}tale at $0$. In this section we fix a polynomial function $f \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$ that has an isolated zero at the origin and write $f_1, \dots, f_n \in P$ for the component functions.
The finite determinacy lemma is:
\begin{lm} \label{Lemma: FiniteDeterminacy}
Given a polynomial function $f \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$ with an isolated zero at the origin, there exists an integer $b>0$ such that if $g \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$ satisfies $f_i-g_i \in \mathfrak{m}_{0}^{b+1}$ for $i=1, \dots, n$, then
\begin{enumerate}
\item $g$ has an isolated zero at the origin;
\item we have $Q_{0}(f)=Q_{0}(g)$ and $E_{0}(f) =E_{0}(g)$;
\item $\deg^{\mathbf{A}^1}_{0}(f) = \deg^{\mathbf{A}^1}_{0}(g)$.
\end{enumerate}
\end{lm}
\begin{proof}
Since $Q_{0}(f)$ is a finite length quotient of $P_{\mathfrak{m}_{0}}$, its defining ideal must contain a power of the maximal ideal, say $\mathfrak{m}_{0}^{b} \subset (f_1, \dots f_n)$. We will prove that any $g$ satisfying $f_i - g_i \in \mathfrak{m}_{0}^{b+1}$ satisfies the desired conditions. To begin, we show the ideals $(f_1, \dots, f_n)$ and $(g_1, \dots, g_n)$ are equal. By the choice of $b$, we have $(g_1, \dots, g_n) \subset (f_1, \dots, f_n)$, and to see the reverse inclusion, we argue as follows. The elements $g_1, \dots, g_n$ generate $(g_1, \dots, g_n) + \mathfrak{m}_{0}^{b}$ modulo $\mathfrak{m}_{0}^{b+1}$ (since, modulo $\mathfrak{m}_{0}^{b+1}$, the $g_i$'s equal the $f_i$'s), so by Nakayama's lemma, the $g_i$'s generate $(g_1, \dots, g_n) + \mathfrak{m}_{0}^{b}$. In particular, $(g_1, \dots, g_n) \supset \mathfrak{m}_{0}^{b}$, and we conclude $(g_1, \dots, g_n) \supset (f_1, \dots, f_n)$ as before.
We immediately deduce $Q_{0}(f)= Q_{0}(g)$. To see that $E_{0}(f) = E_{0}(g)$, observe that if we write $g_i = f_i + \sum b_{i,j} x_j$ with $b_{i, j} \in \mathfrak{m}_{0}^{b}$ and $f_i= \sum a_{i, j} x_j$, then $g_{i} = \sum (a_{i, j} + b_{i, j}) x_{j}$ and
\[
E_{0}(g) = \det( a_{i, j} + b_{i, j}).
\]
Since $a_{i, j}$ equals $a_{i, j} + b_{i, j}$ modulo $\mathfrak{m}^{b}_{0}$, these elements are equal modulo $(f_1, \dots, f_n) = (g_1, \dots, g_n)$. Taking determinants, we conclude $E_{0}(f) = E_{0}(g)$. This proves (1) and (2).
We prove (3) by exhibiting an explicit naive $\mathbf{A}^1$-homotopy between the maps $f'_{0}$ and $f'_{0}$ on Thom spaces. Write
\[
g_i = \sum n_{i, j} f_{j} \text{ in $P_{\mathfrak{m}_{0}}$.}
\]
By definition, $f_i = g_i \text{ modulo $\mathfrak{m}_{0}^{b+1}$}$, hence modulo $\mathfrak{m}_{0} \cdot (f_1, \dots, f_n)$. Moreover, $f_1, \dots, f_n$ is a basis for the $k$-vector space $(f_1, \dots, f_n)/ \mathfrak{m}_{0} \cdot (f_1, \dots, f_n)$, so the matrix $(n_{i, j})$ must reduce to the identity matrix modulo $\mathfrak{m}_{0}$, allowing us to write $(n_{i, j}) = \operatorname{id}_{n} +(m_{i, j})$ with $m_{i, j} \in \mathfrak{m}_{0}$. Let $V \subset \mathbf{A}^{n}_{k}$ be a Zariski neighborhood of the origin such that the entries of the matrix $ \operatorname{id}_{n} + (m_{i, j})$ are restrictions of elements of $H^{0}(V, { \mathcal O})$ that we denote by the same symbols. Now consider the matrix
\[
M(x,t) := \operatorname{id}_{n} + ( t \cdot m_{i, j}(x))
\]
and the map
\begin{gather*}
H \colon V \times_{k} \mathbf{A}_{k}^{1} \to \mathbf{A}^{n}_{k}, \\
(x,t) \mapsto M(x, t) \cdot f(x).
\end{gather*}
The preimage $H^{-1}(0)$ contains $\{ 0 \} \times_{k} \mathbf{A}^{1}_{k}$ as a connected component. Indeed, to see this is a connected component, it is enough to show that the subset is open. To show this, observe that the complement set is
\begin{displaymath}
\left( \{ \det(M)=0 \} \cup (f^{-1}(0) - 0) \times_{k} \mathbf{A}^{1}_{k} \right) \cap H^{-1}(0).
\end{displaymath}
The subset $\{ \det(M) =0 \}$ is closed in $V \times_{k} \mathbf{A}^{1}_{k}$ as $\det(M)$ is a regular function, and $f^{-1}(0) - \{ 0 \} \subset \mathbf{A}^{n}_{k}$ is closed as $0 \in f^{-1}(0)$ is a connected component by hypothesis.
The map $H$ induces a map on quotient spaces
\begin{equation} \label{Eqn: ConstructHomotopy}
H \colon \frac{V \times \mathbf{A}^{1}_{k}}{V \times \mathbf{A}^{1}_{k} - H^{-1}(0)} \to \frac{\mathbf{A}^{n}_{k}}{\mathbf{A}^{n}_{k} - 0}.
\end{equation}
The quotient $\frac{V \times \mathbf{A}^{1}_{k}}{V \times \mathbf{A}^{1}_{k} - 0 \times \mathbf{A}^{1}_{k}}$ naturally includes into $\frac{V \times \mathbf{A}^{1}_{k}}{V \times \mathbf{A}^{1}_{k} - H^{-1}(0)}$ because, as $0 \times \mathbf{A}^{1}$ is a connected component of $H^{-1}(0)$, $\frac{V \times \mathbf{A}^{1}_{k}}{V \times \mathbf{A}^{1}_{k} - H^{-1}(0)}$ is canonically identified with the wedge sum of $\frac{V \times \mathbf{A}^{1}_{k}}{ V \times \mathbf{A}^{1}_{k} - 0 \times \mathbf{A}^{1}_{k}}$ and $\frac{V \times \mathbf{A}^{1}_{k}}{ V \times \mathbf{A}^{1}_{k} - W}$, where $W$ is the complement of $0 \times \mathbf{A}^{1}_{k}$ in $H^{-1}(0)$. Consider now the composition of the inclusion with $H$:
\begin{equation} \label{Eqn: ConstructHomotopy2}
H \colon \frac{V \times \mathbf{A}^{1}_{k}}{V \times \mathbf{A}^{1}_{k} - 0 \times_{k} \mathbf{A}^{1}_{k}} \to \frac{\mathbf{A}^{n}_{k}}{\mathbf{A}^{n}_{k} - 0}.
\end{equation}
The spaces appearing in this last equation are identified with the Thom spaces of normal bundles by the purity theorem, and these Thom spaces, in turn, are isomorphic to a smash product with $\operatorname{Th}({ \mathcal O}^{\oplus n}_{\operatorname{Spec}{k}})$ because the relevant normal bundles are trivial:
\begin{align*}
\frac{\mathbf{A}^{n}_{k}}{\mathbf{A}^{n}_{k} - 0} =& \operatorname{Th}(\{ 0 \} ) \\
=& \{ 0 \}_{+} \wedge \operatorname{Th}({ \mathcal O}_{\operatorname{Spec}(k)}^{\oplus n}), \\
\frac{V \times \mathbf{A}^{1}_{k}}{V \times \mathbf{A}^{1}_{k} - 0 \times \mathbf{A}^{1}_{k}} =& \operatorname{Th}(0 \times \mathbf{A}^{1}_{k}) \\
=& (0 \times \mathbf{A}^{1}_{k})_{+} \wedge \operatorname{Th}({ \mathcal O}_{\operatorname{Spec}(k)}^{\oplus n}).
\end{align*}
These identifications identify \eqref{Eqn: ConstructHomotopy2} with a naive $\mathbf{A}^1$-homotopy
\[
(0 \times \mathbf{A}^{1}_{k})_{+} \wedge \operatorname{Th}({ \mathcal O}_{\operatorname{Spec}(k)}^{\oplus n}) \to \{ 0 \}_{+} \wedge \operatorname{Th}({ \mathcal O}_{\operatorname{Spec}(k)}^{\oplus n})
\]
from $f_x'$ to $g_x'$.
\end{proof}
\begin{df}
If $b$ is an integer satisfying the conditions of Lemma~\ref{Lemma: FiniteDeterminacy}, then we say that $f$ is \textbf{determined of order $b$}. If we are additionally given $g$ satisfying the conditions of the lemma, then we say that $g$ is \textbf{equivalent} to $f$ at the origin.
\end{df}
\begin{rmk}
If $f$ is a polynomial function in $1$ variable that contains a nonzero monomial of degree $b$, then $f$ is $b$-determined. Indeed, we can assume $b$ is the least such integer and then write $f = u \cdot x^{b}$ for $u \in k[x]$ a unit in $(P_{x})_{\mathfrak{m}_{0}}$. The ideal $(x^b)$ lies in $(f)$, so the proof of Lemma~\ref{Lemma: FiniteDeterminacy} shows that $f$ is $b$-determined. We make use of this fact in the companion paper \cite{wickelgren16}.
\end{rmk}
In the proof of the Main Theorem, we use Proposition~\ref{Prop: ReduceToRegSeq_enhanced} to reduce the proof to the special case where the following assumption holds:
\begin{assumption} \label{Assumption}
The polynomial function $f $ is the restriction of a morphism $F \colon \mathbf{P}^{n}_{k} \to \mathbf{P}^{n}_{k}$ such that
\begin{enumerate}
\item $F$ is finite, flat, and with induced field extension $\operatorname{Frac} F_{*}{ \mathcal O}_{\mathbf{P}^{n}_{k}} \supset \operatorname{Frac} { \mathcal O}_{\mathbf{P}^{n}_{k}}$ of degree coprime to $p$;
\item $F$ is \'etale at every point of $F^{-1}(0) - \{ 0 \}$;
\item $F$ satisfies $F^{-1}(\mathbf{A}^n_{k}) \subset \mathbf{A}^{n}_{k}$.
\end{enumerate}
\end{assumption}
To reduce to the special case, we need to prove that, after possible passing from $k$ to a field extension, every $f$ is equivalent to a polynomial function satisfying Assumption~\ref{Assumption}, and we conclude this section with a proof of this fact. The proof we give below is a modification of \cite[Theorem 4.1]{Becker}, a theorem about real polynomial functions due to Becker--Cardinal--Roy--Szafraniec.
We will show that if $f$ is a given polynomial function with an isolated zero at the origin, then for a general $h \in P$ that is homogeneous and of degree sufficiently large and coprime to $p$, the sum $f+h$ satisfies Assumption~\ref{Assumption}. This result is Proposition~\ref{Prop: ReduceToRegSeq_enhanced}. We prove that result as a result about the affine space $H^{d}_{k}$ parameterizing polynomial maps $h \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$ given by $n$-tuples of homogeneous degree $d$ polynomials. We show that the locus of $h$'s such that $f+h$ fails to satisfy Assumption~\ref{Assumption} is not equal to $H^{d}_{k}$ by using the following three lemmas.
\begin{lm}\label{modify_to_give_endo_Pn}
The subset of $H^{d}_{k}$ corresponding to $h$'s such that $h^{-1}(0)=0$ is a nonempty Zariski open subset.
\end{lm}
\begin{proof}
The subset in question is nonempty since it contains e.g.~$(x_1^d, \dots, x_n^d)$. To see that it is open, consider image $I \subset \mathbf{P}( H^{d}_{k})$ of the incidence variety
\[
\{ ([h_1, \dots, h_n], [x_1, \dots, x_{n}]) \colon h_{1}(x_1, \dots, x_n) = \dots = h_{n}(x_1, \dots, x_n)=0 \} \subset \mathbf{P}(H^{d}_{k}) \times \mathbf{P}^{n-1}.
\]
under projection $\mathbf{P}(H^{d}_{k}) \times \mathbf{P}_{k}^{n-1} \to \mathbf{P} (H^{d}_{k})$. The subset $I$ is closed by the properness of the projection morphism. The preimage of $I$ under the natural morphism $H^{d}_{k}-0 \to \mathbf{P}(H^{d}_{k})$ is closed in $H^{d}_{k}-0$, hence the union of the preimage of $I$ and $0$ is closed in $H^{d}_{k}$. The complement of this closed subset is the subset of $h$'s such that $h^{-1}(0)=0$.
\end{proof}
The following lemma is used in Lemma~\ref{modify_to_have_regular_values} to bound a dimension.
\begin{lm} \label{Lemma: BoundFibers}
Let $f \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$ be a nonzero polynomial function satisfying $f(0)=0$ and $a=(a_1, \dots, a_n) \in \mathbf{A}^{n}_{k}(k)$ a $k$-point that is not the origin $0$. If $\sum \frac{\partial f_1}{\partial x_i}(a) \cdot a_i \ne d \cdot f_{1}(a)$, then the subset of $H_{k}^{d}$ consisting of $h$'s satisfying
\begin{align}
f_{i}(a) + h_{i}(a) =& 0 \text{ for $i=1, \dots, n$,} \label{Eqn: RegSeqOne} \\
\det \begin{pmatrix} \frac{\partial(f_i + h_i) }{\partial x_j}(a) \end{pmatrix} =& 0 \label{Eqn: RegSeqTwo}
\end{align}
is a Zariski closed subset of codimension $n+1$.
\end{lm}
\begin{proof}
It is enough, by the Krull principal ideal theorem, to show that \eqref{Eqn: RegSeqOne} and \eqref{Eqn: RegSeqTwo}, considered as regular functions on $H_{k}^{n}$, form a regular sequence. Writing
\begin{gather*}
h_1 = \sum c_{\underline{i}}(1) x_1^{i_1} x_{2}^{i_2} \dots x_{n}^{i_n}, \\
h_2 = \sum c_{\underline{i}}(2) x_1^{i_1} x_{2}^{i_2} \dots x_{n}^{i_n}, \\
\dots \\
h_n = \sum c_{\underline{i}}(n) x_1^{i_1} x_{2}^{i_2} \dots x_{n}^{i_n}, \\
\end{gather*}
the coefficients $\{ c_{\underline{i}}(1), \dots, c_{\underline{i}}(n) \}$ are coordinates on the affine space $H_{k}^{d}$. As polynomials in these coefficients, the elements $f_{1}(a)+ h_{1}(a), \dots, f_{n}(a) + h_{n}(a)$ from \eqref{Eqn: RegSeqOne} are affine linear equations, and distinct linear equations involve disjoint sets of variables, so we conclude that the first set of elements form a regular sequence with quotient equal to a polynomial ring. In particular, the quotient is a domain, so to prove the lemma, it is enough to show that \eqref{Eqn: RegSeqTwo} has nonzero image in the quotient ring.
To show this, first make a linear change of variables so that $a = (1, 0, \dots, 0)$. Setting $\underline{i}_1 = (d, 0, \dots, 0)$, the elements \eqref{Eqn: RegSeqOne} take the form
\begin{displaymath} \label{Eqn: FirstElements}
f_1(a)+c_{\underline{i}_{1}}(1), \dots, f_{n}(a)+c_{\underline{i}_{1}}(n),
\end{displaymath}
and if we let
\begin{gather*}
\underline{i}_{2} = (d-1, 1, 0, \dots, 0, 0), \\
\underline{i}_{3} = (d-1, 0, 1, \dots, 0, 0), \\
\dots \\
\underline{i}_{n} = (d-1, 0, 0, \dots, 0, 1), \\
\end{gather*}
then \eqref{Eqn: RegSeqTwo} can be rewritten as
\[
\det\begin{pmatrix}
d \cdot c_{\underline{i}_{1}}(1) + \frac{\partial f_1}{\partial x_1}(a) & d \cdot c_{\underline{i}_{2}}(1) + \frac{\partial f_1}{\partial x_2}(a) & \dots & d \cdot c_{\underline{i}_{n}}(1) + \frac{\partial f_1}{\partial x_n}(a) \\
d \cdot c_{\underline{i}_{1}}(2) + \frac{\partial f_2}{\partial x_1}(a) & d \cdot c_{\underline{i}_{2}}(2) + \frac{\partial f_2}{\partial x_2}(a) & \dots & d \cdot c_{\underline{i}_{n}}(2) + \frac{\partial f_2}{\partial x_n}(a) \\
\vdots & \vdots & \ddots & \vdots \\
d \cdot c_{\underline{i}_{1}}(n) + \frac{\partial f_n}{\partial x_1}(a) & d \cdot c_{\underline{i}_{2}}(n) + \frac{\partial f_n}{\partial x_2}(a) & \dots & d \cdot c_{\underline{i}_{n}}(n) + \frac{\partial f_n}{\partial x_n}(a)
\end{pmatrix}.
\]
This determinant is essentially the determinant of the general $n$-by-$n$ matrix $\det(x_{\alpha, \beta})$.
More precisely, identify ${ \mathcal O}_{H_{k}^{d}}$ with $k[ x_{\alpha, \beta}]$ by setting $d \cdot c_{\underline{i}_{\beta}}(\alpha) + \frac{\partial f_{\alpha}}{\partial x_{\beta}}(a) = x_{\alpha, \beta}$ (and, say, arbitrarily matching the remaining variables $c_{\underline{i}_{\beta}}(\alpha)$, $i_{\beta} \ne \underline{i}_{1}, \dots, \underline{i}_{n}$, in ${ \mathcal O}_{H_{k}^{d}}$ with the remaining variables in $k[x_{\alpha, \beta}]$). This identification identifies the determinant under consideration with the determinant $\det(x_{\alpha, \beta})$ of the general $n$-by-$n$ matrix and identifies the elements \eqref{Eqn: RegSeqOne} with linear polynomials, say $A_{1} x_{1, 1} + B_{1}, A_{2} x_{2, 1} + B_{2}, \dots, A_{n} x_{n, 1} + B_{n}$.
Now consider $\det(x_{\alpha, \beta})$ as a function $\det(v_1, \dots, v_n)$ of the column vectors. Under the identification of \eqref{Eqn: RegSeqTwo} with $\det(x_{\alpha, \beta})$, the image of \eqref{Eqn: RegSeqTwo} in the quotient ring is identified with $\det(\overline{v}_1, v_{2}, \dots, v_n)$ for $\overline{v}_{1} = (-B_{1}/A_{1}, -B_{2}/A_{2}, \dots, -B_{n}/A_{n})$. By the hypothesis $\sum \frac{\partial f_1}{\partial x_i}(a) \cdot a_i \ne d \cdot f_{1}(a)$, so $-B_{1}/A_{1} \ne 0$ and $\overline{v}_{1}$ is not the zero vector. We conclude that $\det(\overline{v}_1,v_{2}, \dots, v_n)$ is nonzero because e.g.~we can extend $\overline{v}_{1}$ to a basis $\overline{v}_{1}, \dots, \overline{v}_{n}$ and then $\det(\overline{v}_{1}, \overline{v}_{2} \dots, \overline{v}_{n}) \ne 0 $ by the fundamental property of the determinant.
\end{proof}
\begin{lm}\label{modify_to_have_regular_values}
Let $f \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$ be a nonzero polynomial map satisfying $f(0)=0$ and $d$ an integer greater than the degrees of $f_1, \dots, f_n$ and coprime to $p$. Then the subset $S \subset H^{d}_{k}$ of $h$'s such that $f+h$ is \'{e}tale at every point of $(f+h)^{-1}(0) - 0$ contains a nonempty Zariski open subset.
\end{lm}
\begin{proof}
We prove the lemma by proving that the Zariski closure of the complement of $S$ in $H^{d}_{k}$ has dimension strictly smaller than $\dim H^{d}_{k}$, hence the complement of $S$ cannot be Zariski dense.
Consider
\[
\Delta = \left\{ (h, a) \colon (f+h)(a)=0, \det \begin{pmatrix} \frac{\partial(f_i + h_i) }{\partial x_j}(a) \end{pmatrix} = 0, a \ne 0 \right\} \subset H_{k}^{d} \times (\mathbf{A}^{n}_{k}-0).
\]
The subset $S$ is the image $\pi_{1}(\Delta)$ of $\Delta$ under the first projection $\pi_1 \colon H_{k}^{d} \times \mathbf{A}^{n}_{k} \to H_{k}^{d}$. We bound dimensions by analyzing the second projection $\pi_2 \colon H_{k}^{d} \times \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$.
To bound the dimension, we argue as follows. Some $f_i$ is nonzero since $f$ is nonzero, and without loss of generality, we can assume $f_1 \ne 0$. Because $d$ is coprime to $p$ and strictly larger than the degree of $f_1$, the polynomial $\sum \frac{\partial f_1}{\partial x_i}(x) \cdot a_i - d \cdot f_{1}(x)$ is nonzero. We conclude that
\[
B := \{ a \in \mathbf{A}^{n}_{k} \colon \sum \frac{\partial f_1}{\partial x_i}(a) \cdot a_i = d \cdot f_{1}(a) \}.
\]
has codimension $1$ in $\mathbf{A}^{n}_{k}$. We separately bound $\Delta \cap \pi_{2}^{-1}(B)$ and $\Delta \cap \pi_2^{-1} (\mathbf{A}^n_k - B)$.
The fibers of $\pi_{2} \colon \Delta - \pi_{2}^{-1}(B) \to \mathbf{A}^{n}_{k}-B$ have codimension $n+1$ by Lemma~\ref{modify_to_give_endo_Pn}, so by \cite[Proposition~5.5.2]{egaIV_2}, we have
\begin{align*}
\dim \Delta \cap \pi_2^{-1} (\mathbf{A}^n_k - B) \leq& \dim (\mathbf{A}^n_k - B) + \dim \Delta_a \\
=& n + \dim H_k^d - (n+1)\\
<& \dim H_k^d.
\end{align*}
By similar reasoning
\begin{align*}
\dim \pi_2^{-1}(B) \cap \Delta \leq& \dim B + \dim \Delta_{a} \\
=& n-1 + \dim H_k^d - n \\
<& \dim H_k^d.
\end{align*}
We conclude that $\pi_{1} \colon \Delta \to H_{k}^{d}$ cannot be dominant for dimensional reasons \cite[Theorem~4.1.2]{egaIV_2}. The complement of the closure of $\pi_{1}(\Delta)$ can thus be taken as the desired Zariski open subset.
\end{proof}
\begin{pr} \label{Prop: ReduceToRegSeq_enhanced}
Let $f \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$ be a nonzero polynomial function satisfying $f(0)=0$. Then there exists an odd degree extension $L/k$ such that $f \otimes_{k} L$ is equivalent to a function satisfying Assumption~\ref{Assumption}. If $k$ is infinite, we can take $L=k$.
\end{pr}
\begin{proof}
The function $f$ is finitely determined by Lemma~\ref{Lemma: FiniteDeterminacy}, so say it is $b$-determined for $b \in \mathbf{Z}$. Choose $d$ to be an integer coprime to $p$ and larger than both $b$ and the degrees of the $f_i$'s. We claim that there exists an odd degree field extension $L/k$ and a degree $d$ homogeneous polynomial function $h \in H_{k}^{d}(L)$ such that $h^{-1}(0) = \{ 0 \}$ and $g := (f \otimes_{k} L)+h$ is \'{e}tale at every point of $g^{-1}(0)-0$. To verify the claim, observe that Lemmas \ref{modify_to_give_endo_Pn} and \ref{modify_to_have_regular_values} imply that the subset of all such $h$'s contains a nonempty Zariski open subset $U \subset H_{k}^{d}$. If $k$ is an infinite field, $U(k)$ must be non-empty, so we take $L=k$. Otherwise, $k$ is a finite field, say $k=\mathbf{F}_{q}$. For $n$ a sufficiently large odd number, $U(\mathbf{F}_{q^n})$ is nonempty since the union $\cup \mathbf{F}_{q^n}$ over all odd $n$ is an infinite field. In this case, we take $L = \mathbf{F}_{q^n}$ for any sufficiently large odd $n$. The function $f \otimes_{k} L $ is also $b$-determined, and we complete the proof by showing that $g := (f\otimes_{k}L) +h$ satisfies Assumption~\ref{Assumption}.
To ease notation, we only give the proof in the case $L=k$ (the general case involves only notational changes). Define degree $d$ homogeneous polynomials
\begin{align*}
G_0 :=& X_0^d, \\
G_1 :=& X_0^{d}\cdot f_1(X_1/X_0, \dots, X_n/X_0)+h_1(X_1, \dots, X_n), \\
\dots \\
G_n :=& X_0^{d}\cdot f_n(X_1/X_0, \dots, X_n/X_0)+h_n(X_1, \dots, X_n).
\end{align*}
The only solution $(X_0, \dots, X_n)$ to $G_0 = G_1 = \dots = G_n=0$ over $\overline{k}$ is $(0, \dots, 0)$ because from the first equation we deduce $X_{0}=0$ and then from the remaining equations we deduce that $(X_1, \dots, X_n)$ lies in $h^{-1}(0) = \{ 0 \}$. It follows from general formalism that $G = [G_0, G_1, \dots, G_n]$ defines a morphism $G \colon \mathbf{P}^{n}_{k} \to \mathbf{P}^{n}_{k}$ such that $G^{*}({ \mathcal O}(1)) = { \mathcal O}(d)$. Moreover, by construction $G$ is an extension of $g$ that satisfies $G^{-1}(\mathbf{A}^{n}_{k}) \subset \mathbf{A}^{n}_{k}$.
To complete the proof, we need to show that $G$ is finite, flat, and induces a field extension $\operatorname{Frac} { \mathcal O}_{\mathbf{P}^{n}_{k}} \subset \operatorname{Frac} G_{*}{ \mathcal O}_{\mathbf{P}^{n}_{k}}$ of degree coprime to $p$. To see that $G$ is finite, observe that the pullback $G^{-1}(\mathcal{H})$ of a hyperplane $\mathcal{H} \subset \mathbf{P}^{n}_{k}$ has nonnegative degree on every curve (since the associated line bundle is $G^{*} { \mathcal O}(1) = { \mathcal O}(d)$, an ample line bundle). We conclude that a fiber of $G$ cannot contain a curve since $G^{-1}(\mathcal{H})$ can be chosen to be disjoint from a given fiber. In other words, $G$ has finite fibers. Being a morphism of projective schemes, $G$ is also proper and hence finite by Zariski's main theorem. This implies that $G$ is flat since every finite morphism $\mathbf{P}^{n}_{k} \to \mathbf{P}^{n}_{k}$ is flat by \cite[Corollary to Theorem~23.1]{Matsumura_CRT}.
Finally, we complete the proof by noting that the degree of $\operatorname{Frac} { \mathcal O}_{\mathbf{P}^{n}_{k}} \subset \operatorname{Frac} G_{*}{ \mathcal O}_{\mathbf{P}^{n}_{k}}$ equals $d^n$, the top intersection number $G^{*}(\mathcal{H}^{n}) = \int c_{1}({ \mathcal O}(d))^{n}$. (To deduce the equality, observe that ${ \mathcal H}^{n}$ is the class of a $k$-point $\overline{y} \in \mathbf{P}^{n}_{k}(k)$, so the top intersection number is the $k$-rank of ${ \mathcal O}_{G^{-1}(\overline{y})}$, the stalk of $G_{*}{ \mathcal O}_{\mathbf{P}^{n}_{k}}$ at $\overline{y}$. The rank of that stalk is equal to the rank of any other stalk of $G_{*}{ \mathcal O}_{\mathbf{P}^{n}_{k}}$ since $G_{*}{ \mathcal O}_{\mathbf{P}^{n}_{k}}$, being finite and flat, is locally free. In particular, that rank equals the rank of the generic fiber, which is the degree of $\operatorname{Frac} G_{*}{ \mathcal O}_{\mathbf{P}^{n}_{k}} \supset \operatorname{Frac} { \mathcal O}_{\mathbf{P}^{n}_{k}}$.)
\end{proof}
\section{The family of symmetric bilinear forms} \label{Section: Family}
In this section we construct, for a given finite polynomial map $f \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$, a family of symmetric bilinear forms over $\mathbf{A}^{n}_{k}$ such that the fiber over the origin contains a summand that represents the ELK class $w_{0}(f)$. This family has the property that the stable isomorphism class of the fiber over $\overline{y} \in \mathbf{A}^{n}_{k}(k)$ is independent of $\overline{y}$, and we use this property in Section~\ref{Section: PrOfMainThm} to compute $w_{0}(f)$ in terms of a regular value. Finally, we compute the stable isomorphism class of the family over an \'{e}tale fiber.
Throughout this section $f$ denotes a finite polynomial map, except in Remark~\ref{Remark: WhyFiniteAssumption} where we explain what happens if the finiteness condition is weakened to quasi-finiteness.
The basic definition is the following.
\begin{df} \label{Def: FamilyOfForms}
Define the \textbf{family of algebras} $\widetilde{Q} = \widetilde{Q}(f)$ associated to $f$ by
\begin{align*}
\widetilde{Q} :=& f_{*}{ \mathcal O}_{\mathbf{A}^{n}_{k}}.
\end{align*}
\end{df}
Concretely $\widetilde{Q}$ is the ring $P_{x}$ considered as a $P_{y}$-algebra by the homomorphism $y_1 \mapsto f_{1}(x), \dots, y_n \mapsto f_{n}(x)$ or equivalently the algebra $P_y[x_1, \dots, x_n]/(f_1(x)-y_1, \dots, f_{n}(x)-y_n)$. Given $(\overline{y}_{1}, \dots, \overline{y}_{n}) = \overline{y} \in \mathbf{A}^{n}_{k}(L)$ for some field extension $L/k$, the fiber $\widetilde{Q} \otimes k(\overline{y})$ is the $L$-algebra $L[x_1, \dots, x_n]/(f_{1}(x)-\overline{y}_{1}, \dots, f_{n}(x)-\overline{y}_{n})$. This algebra decomposes as
\[
\widetilde{Q} \otimes k(\overline{y}) = Q_{x_1}(f) \times \dots \times Q_{x_m}(f),
\]
where $Q_{x}(f)$ is as in Definition~\ref{Section: LocalForm} and the product runs over all closed points $x \in f^{-1}(\overline{y})$.
The algebra $\widetilde{Q}$ has desirable properties because we have assumed that $f$ is finite.
\begin{lm} \label{Lemma: RelativeCompleteIntersection}
The elements $f_1(x)-y_1, \dots, f_{n}(x) - y_n \in P_{y}[x_1, \dots, x_n]$ form a regular sequence, and $\widetilde{Q}$ is $P_{y}$-flat.
\end{lm}
\begin{proof}
It is enough to show that, for any maximal ideal $\mathfrak{m} \subset P_{y}$, the images of $y_1-f_{1}(x),\dots, y_n - f_{n}(x)$ in $(P_{y}/\mathfrak{m})[x_1, \dots, x_n]$ form a regular sequence. The quotient of $P_{y}[x_1, \dots, x_n]$ by the sequence is the structure ring of $f^{-1}(\mathfrak{m})$, which is $0$-dimensional by hypothesis. In particular, the images of $y_1-f_1(x), \dots, y_n-f_{n}(x)$ generate a height $n$ ideal and hence they form a regular sequence by \cite[Theorem~17.4(i)]{Matsumura_CRT}.
\end{proof}
As we mentioned in Section~\ref{Section: LocalForm}, Scheja--Storch have constructed a distinguished symmetric bilinear form $\beta_0$ on $Q_{0}(f)$ that represents $w_{0}(f)$. In fact, they construct a family $\widetilde{\beta}$ of symmetric bilinear forms on $\widetilde{Q}$ such that the fiber over $0$ contains a summand that represents $w_{0}(f)$. The family is defined as follows.
\begin{df} \label{Definition: SSForm}
Let $\widetilde{\eta} \colon \widetilde{Q} \to P_{y}$ be the generalized trace function, the $P_y$-linear function defined on \cite[page~182]{scheja}. Let $\widetilde{\beta}$ be the symmetric bilinear form $\widetilde{\beta} \colon \widetilde{Q} \to P_{y}$ defined by $\widetilde{\beta}(\widetilde{a}_{1}, \widetilde{a}_{2}) = \widetilde{\eta}(\widetilde{a}_{1} \cdot \widetilde{a}_{2})$. Given $\overline{y} \in \mathbf{A}^{n}_{n}(L)$ and $x \in f^{-1}(\overline{y})$, we write $\eta_{x}$ and $\beta_{x}$ for the respective restrictions to $Q_{x}(f) \subset \widetilde{Q} \otimes k(\overline{y})$. We write $w_{x}(f) \in \operatorname{GW}(k)$ for the isomorphism class of $(Q_{x}(f), \beta_{x})$.
\end{df}
\begin{rmk}\label{rmk_eta_comments}
We omit the definition of $\widetilde{\eta}$ because it is somewhat involved and we do not make direct use of it. We do make use of two properties of $\widetilde{\eta}$. First, the homomorphism has strong base change properties. Namely, for a noetherian ring $A$ and an $A$-finite quotient $B$ of $A[x_1, \dots, x_n]$ by a regular sequence, Scheja--Storch construct a distinguished $A$-linear function $\eta_{B/A} \colon B \to A$ in a manner that is compatible with extending scalars by an arbitrary homomorphism $A \to \overline{A}$ \cite[page~184, first paragraph]{scheja}. Second, the restriction $\eta_0 \colon Q_{0}(f) \to k$ of $\widetilde{\eta}$ satisfies the condition from Definition~\ref{Definition: EKLform} (by Lemma~\ref{Lemma: EtaSatisfiesEKL} below). In particular, the definition of $w_{0}(f)$ in Definition~\ref{Definition: EKLform} agrees with the definition of $w_{x}(f)$ from Definition~\ref{Definition: SSForm} when $x=0$.
The reader familiar with \cite{eisenbud77} may recall that in that paper, where $k=\mathbf{R}$, the authors do not make direct use of Scheja--Storch's work but rather work directly with the functional on $\widetilde{Q}$ defined by $a \mapsto \operatorname{Tr}(a/J)$. Here $\operatorname{Tr}$ is the trace function of the field extension $\operatorname{Frac} \widetilde{Q} / \operatorname{Frac} P_{y}$. We do not use the function $a \mapsto \operatorname{Tr}(a/J)$ because it is not well-behaved in characteristic $p>0$ since e.g.~the trace can be identically zero.
\end{rmk}
We now describe the properties of the family $(\widetilde{Q}, \widetilde{\beta})$.
\begin{lm} \label{Lemma: EtaSatisfiesEKL}
The distinguished socle element $E$ satisfies $\eta_{0}(E)=1$.
\end{lm}
\begin{proof}
By the construction of $\eta_0$ in \cite{scheja}, $\eta_{0} = \Theta^{-1}(1)$ for a certain explicit $Q_{0}(f)$-linear homomorphism $\Theta \colon \operatorname{Hom}_{k}(Q_{0}(f), k) \to Q_{0}(f)$. Furthermore, Scheja--Storch prove $\pi = \Theta^{-1}(E)$ for $\pi \colon Q_{0}(f) \to k$ the evaluation map $\pi(a) = a(0)$ by \cite[proof of (4.7)~Korollar]{scheja}. By linearity, we have
\begin{align*}
\Theta(\pi) =& E \\
=& E \cdot 1 \\
=& E \cdot \Theta(\eta_{0}) \\
=& \Theta( E \cdot \eta_{0}),
\end{align*}
so $\pi = E \cdot \eta_{0}$. Evaluating $\pi(1)$, we deduce
\begin{align*}
1 =& (E \cdot \eta_{0})(1) \\
=& \eta_{0}(E).
\end{align*}
\end{proof}
\begin{lm} \label{Lemma: FormOrthogonal}
Let $\widetilde{Q} \otimes k(\overline{y}) = Q_1 \times Q_2$ be a decomposition into a direct sum of rings. Then $Q_1$ is orthogonal to $Q_2$ with respect to $\widetilde{\beta} \otimes k(\overline{y})$.
\end{lm}
\begin{proof}
By definition $a_1 \cdot a_2=0$ for all $a_1 \in Q_1$, $a_2 \in Q_2$, hence $(\widetilde{\eta} \otimes k(\overline{y})) (a_1 \cdot a_2)=0$.
\end{proof}
\begin{lm} \label{Lemma: SpecializedNondegenerate}
The bilinear symmetric form $\widetilde{\beta} \otimes k(\overline{y})$ is nondegenerate.
\end{lm}
\begin{proof}
It is enough to verify the claim after extending scalars, so we can assume $L$ and the residue field of every point of $f^{-1}(\overline{y})$ is $k$. We then have
\[
\widetilde{Q} \otimes k(\overline{y}) = Q_{x_1}(f) \times \dots \times Q_{x_m}(f)
\]
where $x_1, \dots, x_m$ are the points of $f^{-1}(\overline{y})$. Repeated applications of Lemma~\ref{Lemma: FormOrthogonal} show that this is an orthogonal decomposition, and the restriction of $\widetilde{\beta} \otimes k(\overline{y})$ to $Q_{x_i}(f)$ is nondegenerate by Lemmas~\ref{Lemma: WhenFormsEqual} and \ref{Lemma: EtaSatisfiesEKL} (translate to identify $Q_{x_i}(f)$ with an algebra supported at the origin).
\end{proof}
\begin{co}
The symmetric bilinear form $\widetilde{\beta}$ is nondegenerate.
\end{co}
\begin{proof}
The $P_{y}$-module $\widetilde{Q}$ is finite (by hypothesis) and flat (by Lemma~\ref{Lemma: RelativeCompleteIntersection}), so it is free by the solution to Serre's problem. Fix a basis $\widetilde{e}_1, \dots, \widetilde{e}_{m}$ and consider the discriminant $\mathfrak{D} := \det( \widetilde{\beta}( \widetilde{e}_{i}, \widetilde{e}_{j})) \in P_{y}$. Since the formation of both the determinant and $(\widetilde{Q}, \widetilde{\beta})$ are compatible with specialization, we deduce from Lemma~\ref{Lemma: SpecializedNondegenerate} that $\mathfrak{D}$ is not contained in a maximal ideal of $P_{y}$ and thus is a unit.
\end{proof}
We now prove that, for $\overline{y}_{1}, \overline{y}_{2} \in \mathbf{A}^{n}_{k}(k)$, the restriction $\widetilde{\beta}$ to the fiber over $\overline{y}_{1}$ is stably isomorphic to restriction to the fiber over $\overline{y}_2$. This result follows easily from the following form of Harder's theorem.
\begin{lm}[Harder's theorem] \label{Lemma: HarderLemma}
Suppose that $( \widetilde{Q}, \widetilde{\beta})$ is a pair consisting of a finite rank, locally free module $\widetilde{Q}$ on $\mathbf{A}^1_k$ and a nondegenerate symmetric bilinear form $\widetilde{\beta}$ on $\widetilde{Q}$. Then $(\widetilde{Q}, \widetilde{\beta}) \otimes k(\overline{y}_{1})$ is stably isomorphic to $(\widetilde{Q}, \widetilde{\beta}) \otimes k(\overline{y}_{2})$ for any $\overline{y}_1, \overline{y}_{2} \in \mathbf{A}^1_{k}(k)$.
\end{lm}
\begin{proof}
When $\operatorname{char} k \ne 2$, the stronger claim that $( \widetilde{Q}, \widetilde{\beta})$ is isomorphic to a symmetric bilinear form defined over $k$ is \cite[Theorem~3.13]{lam05}. When $\operatorname{char} k = 2$, the claim can be deduced from loc.~cit.~as follows. By \cite[Remark~3.14]{lam05}, the pair ($\widetilde{Q}, \widetilde{\beta})$ is isomorphic to an orthogonal sum of a symmetric bilinear form defined over $k$ and a sum of symmetric bilinear forms defined by Gram matrices of the form
\begin{equation} \label{Eqn: Char2Matrix}
\begin{pmatrix} a(y) & 1 \\
1 & 0
\end{pmatrix}
\end{equation}
for $ a(y) \in k[y]$, so it is enough to prove the lemma in the special case of a symmetric bilinear form defined by \eqref{Eqn: Char2Matrix}. Given $\overline{y} \in \mathbf{A}^{1}_{k}(k)$, if $a(\overline{y})=0$, the specialization of \eqref{Eqn: Char2Matrix} is $\mathbf{H}$, and otherwise the following matrix identity
\[
\begin{pmatrix}
a(\overline{y}) & 0 & 0 \\
0 & 0 & 1 \\
0 & 1 & 0
\end{pmatrix}
=
\begin{pmatrix}
1 & 1 & 0 \\
0 & 1 & 0 \\
a(\overline{y}) & a(\overline{y}) & 1
\end{pmatrix}^{\operatorname{T}}
\cdot
\begin{pmatrix}
a(\overline{y}) & 0 & 0 \\
0 & a(\overline{y}) & 1 \\
0 & 1 & 0
\end{pmatrix} \cdot
\begin{pmatrix}
1 & 1 & 0 \\
0 & 1 & 0 \\
a(\overline{y}) & a(\overline{y}) & 1
\end{pmatrix}
\]
shows that the specialization is stably isomorphic to $\mathbf{H}$.
\end{proof}
\begin{co} \label{Co: TwoFibersSame}
The sum
\begin{equation} \label{Eqn: ConservedSum}
\sum_{x \in f^{-1}(\overline{y})} w_{x}(f) \in \operatorname{GW}(k)
\end{equation}
is independent of $\overline{y} \in \mathbf{A}^{n}_{k}(k)$.
\end{co}
\begin{proof}
The sum \eqref{Eqn: ConservedSum} is the class of $\widetilde{\beta} \otimes k(\overline{y})$ by Lemma~\ref{Lemma: FormOrthogonal}, so since any two $k$-points of $\mathbf{A}^{n}_{k}$ lie on a line, the result follows from Lemma~\ref{Lemma: HarderLemma}.
\end{proof}
\begin{rmk} \label{Remark: WhyFiniteAssumption}
Corollary~\ref{Co: TwoFibersSame} becomes false if the hypothesis that $f$ is finite is weakened to the hypothesis that $f$ is quasi-finite (i.e.~has finite fibers). Indeed, under this weaker assumption, Scheja--Storch construct a nondegenerate symmetric bilinear form $\beta \otimes k(\overline{y})$ on $\widetilde{Q} \otimes k(\overline{y})$ for every $\overline{y} \in \mathbf{A}^{n}_{k}(k)$, but the class of $\sum w_{x}(f)$ can depend on $\overline{y}$. For example, consider $k=\mathbf{R}$ (the real numbers) and $f := (x_{1}^{3} x_{2}+x_{1}-x_{1}^{3}, x_{2})$. A computation shows
\[
\sum_{x \in f^{-1}(\overline{y})} w_{x}(f) = \begin{cases}
\langle 1 \rangle & \text{ if $\overline{y}_{2}=1$;} \\
\langle 1/(\overline{y}_{2}-1) \rangle+\mathbf{H} & \text{ otherwise,}
\end{cases}
\]
so the rank of $\sum w_{x}(f)$ depends on $\overline{y}$.
The morphism $f$ fails to be finite, and we recover finiteness by passing to the restriction $f \colon f^{-1}(U) \to U$ over $U := \mathbf{A}_{k}^2 - \{ y_{2}=1 \}$. Over $U$, the rank is constant, but the isomorphism class still depends on $\overline{y}$ because
\[
\text{signature of $\sum_{x \in f^{-1}(\overline{y})} w_{x}(f)$} = \begin{cases}
+1 & \text{ if $\overline{y}_{2}>1$;}\\
-1 & \text{ if $\overline{y}_{2}<1$.}
\end{cases}
\]
\end{rmk}
We now compute $w_{x}(f)$ when $f$ is \'etale at $x$.
\begin{lm} \label{Lemma: DescentForEtale}
Let $f \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$ be finite and and $\overline{y} \in \mathbf{A}^{n}_{k}(k)$ be a $k$-rational point. Suppose that $f$ is \'{e}tale at $x \in f^{-1}(\overline{y})$ and $L$ is a finite Galois extension of $k$ such that $k(x)$ embeds into $L$. Let $S$ be the set
\[
S = \{ \overline{x} \in \mathbf{A}^{n}_{k}(L) \colon \overline{x} (\operatorname{Spec} L)= \{ x\} \},
\]
of $L$-points whose image in $\mathbf{A}^n_k$ is $x$. Define $V(S)$ to be the $L$-algebra of functions $S \to L$, with point-wise addition and multiplication. Define $\phi \colon S \to k$ by $\phi(\overline{x}) = 1/J(\overline{x})$, and $\beta_{\phi} \colon V(S) \times V(S) \to L$ to be the bilinear form
\begin{align*}
\beta_{\phi}(v_1, v_2)= \sum_{s \in S} \phi(s) v_{1}(s) \cdot v_{2}(s).
\end{align*}The Galois action on $S$ induces an action on $V(S)$, which is descent data.
The $k$-bilinear space described by the descent data on the $L$-bilinear space $V(S)$ is isomorphic to the bilinear space $w_{x}(f)=(Q_{x}(f), \beta_x)$.
\end{lm}
\begin{proof}
The Galois action on $V(S)$ is descent data, because $J$ is a polynomial with coefficients in $k$. To prove the lemma, it is sufficient to construct a $k$-linear map $Q_{x}(f) \to V(S)$ that respects the bilinear pairings and realizes $Q_{x}(f)$ as the equalizer
\[
Q_{x}(f) \to V(S) \rightrightarrows \prod_{\sigma \in G} V(S).
\] where $G=\operatorname{Gal}(L/k)$.
Let $V(\phi): V(S) \to L$ denote the linear functional $V(\phi)(v) = \sum_{s \in S} \phi(s) V(s)$. An element of $Q_{x}(f)$ is a polynomial function on $S$, so there is a tautological inclusion $Q_{x}(f) \to V(S)$, and we will show that this inclusion has the desired properties. The inclusion respects the bilinear forms because the functional $V(\phi)$ restricts to the residue functional $\eta$ (to see this, extend scalars to $L$ and then observe that, for every summand $L$ of $Q_{x}(f)$, both $\eta$ and $\phi$ map the Jacobian element to $1$).
Unraveling the definition of the descent data, we see that the equalizer is the subset of $G$-invariant functions (i.e.~functions $v \colon S \to L$ satisfying $v(\sigma s) = v(s)$ for all $\sigma \in G$, $s \in S$). Because $S$ is finite, every function on $S$ is a polynomial function, and a polynomial function is $G$-invariant if and only if it can be represented by a polynomial with coefficients in $k$, i.e.~lies in $Q_{x}(f)$.
\end{proof}
\section{Proof of the Main Theorem} \label{Section: PrOfMainThm}
We first prove the Main Theorem in the case where $f$ is \'etale.
\begin{lm} \label{Lemma: etalevalue}
Let $f \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$ be a polynomial function that satisfies Assumption~\ref{Assumption} and $\overline{y} \in \mathbf{A}^{n}_{k}(k)$ be a $k$-rational point. Suppose that $f$ is \'{e}tale at $x \in f^{-1}(\overline{y})$. Then $$w_{x}(f)= \deg^{\mathbf{A}^1}_x f.$$
\end{lm}
\begin{proof}
Let $(V(S), \beta_{\phi} )$ be the bilinear space of Lemma \ref{Lemma: DescentForEtale}. By Lemma \ref{Lemma: DescentForEtale} and Proposition \ref{loc_degree_etale_point}, it suffices to show that the Galois descent datum on $V(S)$ determines the isomorphism class of $\operatorname{Tr}_{k(x)/k} \langle J(x) \rangle$.
The equality $\langle J \rangle = \langle 1/J \rangle$ in the Grothendieck-Witt group shows that $\operatorname{Tr}_{k(x)/k} \langle J(x) \rangle$ has representative bilinear form $B: k(x) \times k(x) \to k$ defined $B(x,y) = \operatorname{Tr}_{k(x)/k}(xy/J)$. The proposition is equivalent to the statement that there is a $\operatorname{Gal}(L/k)$-equivariant isomorphism $ L \otimes_k k(x) \cong V(S)$ respecting the bilinear forms.
Note that $S$ is in bijective correspondence with the set of embeddings $k(x) \hookrightarrow L$, and that we may therefore view $s$ in $S$ as a map $s: k(x) \to L$. Let $\Theta: L \otimes_k k(x) \to V(S)$ denote the $L$-linear isomorphism defined by $$\Theta (l \otimes q) (s) =l (s q).$$ By definition, $$\beta_{\phi}(\Theta(1 \otimes q_1), \Theta(1 \otimes q_1)) = \sum_{s \in S} (1/J(s)) s(q_1) s(q_2),$$ where $J(s)$ denotes the Jacobian determinant evaluated at the point $s$, and $s(q_i)$ denotes the image of $q_i$ under the embedding $k(x) \to L$ corresponding to $s$. Since $J$ is defined over $k$, we have $J(s) = s(J)$. Thus $$\sum_{s \in S} (1/J(s)) s(q_1) s(q_2) = \operatorname{Tr}_{L/k}(q_1 q_2/J) = B(q_1,q_2),$$ showing that $\Theta$ respects the appropriate bilinear forms.
\end{proof}
We now use the previous results to prove the Main Theorem.
\begin{proof}[Proof of Main Theorem]
Recall that, after possibly passing from $k$ to an odd degree field extension $L/k$ when $k$ is a finite field, we can assume that $f$ satisfies Assumption~\ref{Assumption} by Proposition~\ref{Prop: ReduceToRegSeq_enhanced}. This allows us to reduce to the case where the assumption is satisfied because the natural homomorphism
\begin{equation} \label{Eqn: GWExtendScalars}
\operatorname{GW}(k) \to \operatorname{GW}(L), w \mapsto w \otimes_{k} L
\end{equation}
is injective. Injectivity holds for a somewhat general $L/k$, but we only need the result in the simple case of a finite field. For finite fields, injectivity can be established as follows. When $\operatorname{char} k = 2$, the only invariant of an element of $\operatorname{GW}(k)$ is its rank, so injectivity follows from the observation that extending scalars preserves the rank. When $\operatorname{char} k \ne 2$, an element of $w \in \operatorname{GW}(k)$ is completely determined by its rank and discriminant. Thus to show injectivity, we need prove that if $\operatorname{disc}(w \otimes_{k} L) \in (L^{\ast})^{2}$, then $\operatorname{disc}(w) \in (k^{\ast})^{2}$, and this result is e.g.~a consequence of \cite[Corollary~2.6]{lam05}.
Since the formation of $w_0(f)$ is compatible with field extensions and similarly with $\deg^{\mathbf{A}^1}_{0}(f)$, we conclude that it is enough to prove the theorem when $f$ satisfies Assumption~\ref{Assumption}.
After possibly passing to another odd degree field extension, we can further assume that there exists $\overline{y}_0 \in \mathbf{A}^{n}_{k}(k)$ such that $f$ is \'{e}tale at every point of $x \in f^{-1}(\overline{y}_{0})$. (To see this: Since any field extension of degree prime to $p$ is separable, $f$ is \'etale at the generic point \cite[17.6.1c']{egaIV_4}. Therefore $f$ is \'etale when restricted to a Zariski neighborhood of the generic point \cite[D\'efinition 17.3.7]{egaIV_4}. Let $Z$ denote the complement of this open neighborhood. Since finite maps are closed, $f(Z)$ is closed a closed subset of $\mathbf{A}^n_k$ not containing the generic point. Thus $\mathbf{A}_k^n - f(Z)$ is a non-empty open subset of $\mathbf{A}^n_k$, which therefore contains a rational point after passing to any infinite field such as the composite of all odd degree finite extensions.)
By these assumptions, $f$ is the restriction of $F \colon \mathbf{P}^{n}_{k} \to \mathbf{P}^{n}_{k}$, and $F$ induces a map $\mathbf{P}(F): \mathbf{P}^n_k/\mathbf{P}^{n-1}_k \to \mathbf{P}^n_k/\mathbf{P}^{n-1}_k$ of motivic spheres that has degree
\begin{equation} \label{pmt:local_deg}
\deg^{\mathbf{A}^1}( \mathbf{P}(F)) = \sum_{x \in f^{-1}(\overline{y})} \deg^{\mathbf{A}^1}_{x}(F) \text{ for any $\overline{y} \in \mathbf{A}^{n}_{k}(k)$}
\end{equation}
by the local degree formula (Proposition~\ref{pr_deg_is_sum_local_deg}). In particular, the right-hand side is independent of $\overline{y}$ since the left-hand side is.
Analogously we proved in Section~\ref{Section: Family} that the sum
\[
\sum_{x \in f^{-1}(\overline{y})}w_{x}(f) \in \operatorname{GW}(k)
\]
is independent of $\overline{y} \in \mathbf{A}^{n}_{k}(k)$ (Corollary~\ref{Co: TwoFibersSame}).
The local terms $w(\beta_{x})$ and $\deg^{\mathbf{A}^1}_{x}(F)$ are equal when $F$ is \'{e}tale at $x$ by the Lemma~\ref{Lemma: etalevalue}. As a consequence
\begin{equation} \label{Eqn: SumsAreEqual}
\sum_{x \in f^{-1}(\overline{y})} \deg^{\mathbf{A}^1}_{x}(F) = \sum_{x \in f^{-1}(\overline{y})}w_{x}(f)
\end{equation}
for $\overline{y}=\overline{y}_{0}$ and hence (by independence) all $\overline{y} \in \mathbf{A}^{n}_{k}(k)$. In particular, equality holds when $\overline{y}=0$. By Assumption~\ref{Assumption}, the morphism $F$ is \'{e}tale at every $x \in f^{-1}(0)$ not equal to the origin, so subtracting off these terms from \eqref{Eqn: SumsAreEqual}, we get
\[
\deg^{\mathbf{A}^1}_{0}(f) = w_{0}(f).
\]
\end{proof}
\section{Application to singularity theory} \label{Section: SingThy}
Here we use the local $\mathbf{A}^1$-degree to count singularities arithmetically, as proposed in the introduction. We assume in this section
\[
\operatorname{char} k \ne 2,
\]
but see Remark~\ref{Rmk: CharTwo} for a discussion of $\operatorname{char} k =2$.
Specifically, given the equation $f \in P_{x}$ of an isolated hypersurface singularity $X \subset \mathbf{A}^{n}_{k}$ at the origin, we interpret the following invariant as a counting invariant:
\begin{df}
If $f \in P_{x}$ is a polynomial such that $\operatorname{grad} f$ has an isolated zero at the closed point $x \in \mathbf{A}^{n}_{k}$, then we define $\mu^{\mathbf{A}^1}_{x}(f) := \deg^{\mathbf{A}^1}_{x}(\operatorname{grad} f)$. When $x$ is the origin, we write $\mu^{\mathbf{A}^1}(f)$ for this class and call it the \textbf{arithmetic Milnor number} (or \textbf{$\mathbf{A}^1$-Milnor number}).
\end{df}
Two remarks about this definition:
\begin{rmk}
The condition that $\operatorname{grad} f$ has an isolated zero at $x$ implies that $\operatorname{Spec}( P_{x}/f)$ has an isolated singularity at $x$, and the converse is true in characteristic $0$ but not in characteristic $p>0$, as the example of $f(x) = x_1^p+x_2^2$ shows.
\end{rmk}
\begin{rmk}
When $k=\mathbf{C}$, the arithmetic Milnor number is determined by its rank, which is the classical Milnor number $\mu(f) = \operatorname{rank} Q_{0}(f)$. The classical Milnor number $\mu(f)$ is not only an invariant of the equation $f$ but in fact is an invariant of the singularity $0 \in \operatorname{Spec}(P_{x}/f)$ defined by $f$. When $k$ is arbitrary, the invariant properties of $\mu^{\mathbf{A}^1}(f)$ are more subtle, especially in characteristic $p>0$. In particular, the rank of $\mu^{\mathbf{A}^1}(f)$ is not an invariant of the singularity in characteristic $p>0$ . For example, both $f(x) = x_1^2 + x_2^p + x^{p + 1}$ and $g(x) = x_1^2 + x_2^p + x_2^{2 p +1}$ both define the $A_{p-1}$ singularity (in the sense that $\widehat{P}_{x}/f$, $\widehat{P}_{x}/g$, and $\widehat{P}_{x}/(x_1^2 + x_2^p)$ are isomorphic), but the ranks of $w_{0}(f)$ and $w_{0}(g)$ are respectively $p$ and $2p$.
\end{rmk}
We now recall the definition of a node.
\begin{df} \label{Def: Node}
The algebra of the \textbf{standard node} (or standard $A_1$-singularity) is
\begin{equation} \label{Eqn: StdNode}
\widehat{P}/ x_1^2 + \dots x_n^2.
\end{equation}
Given a hypersurface $X = \operatorname{Spec}( P/f) \subset \mathbf{A}^{n}_{k}$, we say that a closed point $x \in X$ is a (possibly nonstandard) \textbf{node} (or $A_1$-singularity) if, for every closed point $\widetilde{x} \in X \otimes_{k} \overline{k}$ mapping to $x$, the completed local ring $\widehat{{ \mathcal O}}_{X \otimes \overline{k}, \widetilde{x}}$ is isomorphic to the algebra of the standard node.
We say that $f \in P$ is \textbf{the equation of a node} at a closed point $x \in \mathbf{A}^{n}_{k}$ if $x \in \operatorname{Spec}(P/f)$ is a node. Suppose $L/k$ is a field extension, $x \in \mathbf{A}^{n}_{k}$ a closed point, and $f \in P \otimes_{k} L$ an equation that defines a node at every point $\widetilde{x} \in \mathbf{A}^{n}_{L}$ in the preimage of $x$. If $\operatorname{grad}(f)$ lies in $P$, then we define \textbf{arithmetic type} of the equation at $x$ to be $\deg^{\mathbf{A}^1}_{x} \operatorname{grad}(f) \in \operatorname{GW}(k)$.
\end{df}
The arithmetic type just defined is the invariant discussed in the introduction. Indeed, if $x \in \mathbf{A}^{n}_{k}$ is the origin and $f = u_1 x_1^2 + \dots + u_n x_n^2$, then $\operatorname{grad}(f) = (2 u_1 x_1, \dots, 2 u_n x_n)$, so the arithmetic type of $f$ at $x$.
\begin{align*}
\left\langle \det \frac{\partial^2 f}{\partial x_i \partial x_j}(0) \right\rangle =& \langle 2^{n} u_1 \dots u_n \rangle.
\end{align*}
Somewhat more generally, the arithmetic type of $f \in P$ at a point $x \in \mathbf{A}^{n}_{k}$ with residue field $k$ is given by the Hessian.
When $f$ has coefficients in $k$ but $k(x)$ is a nontrivial extension of $k$, the arithmetic type can be described as follows. If $L/k$ splits $k(x)$, then every point $\widetilde{x}$ of $\mathbf{A}^{n}_{L}$ mapping to $x$ has residue field $L$, so as an element of $\operatorname{GW}(L)$, the arithmetic type of $f \otimes L$ at $\widetilde{x}$ is the value of the Hessian. The symmetric bilinear form
\[
(\text{arithmetic type of $f \otimes L$ at $\widetilde{x}_{1}$}) \oplus \dots \oplus (\text{arithmetic type of $f \otimes L$ at $\widetilde{x}_{m}$})
\]
defined by summing over all $\widetilde{x}$'s carries natural descent data, and this symmetric bilinear form over $L$ with descent data corresponds to the arithmetic type of $f$ at $x$.
Finally, the arithmetic type is also defined for a closed point $x \in \mathbf{A}^{n}_{k}$ and an equation $f \in P \otimes_{k} L$ with coefficients in $L$ provided certain hypotheses are satisfied. An example where the hypotheses are satisfied is when $k=\mathbf{Q}_{5}$ (the $5$-adic numbers), $L=\mathbf{Q}_{5}(\sqrt{-5})$,
\begin{gather}
f(x,y) = x_1^2 + x_2^3 + 2 \cdot 3 x_2 - 10 \sqrt{-5}~~ \text{ and } \label{Eqn: LastTypeExample} \\
x = (x_1, x_2^2+5). \notag
\end{gather}
Observe that $f$ is an equation obtained from an equation with coefficients in $k$ (namely $x_1^2 + x_2^3 + 2 \cdot 3 x_2$) and adding an element of $L$, and such equations naturally arise when interpreting the arithmetic Milnor number as a count of nodes. We demonstrate how such equations arise in the example following Corollary~\ref{Corollary: SingularityCount}.
When $n$ is odd, the arithmetic type is an invariant of the equation $f$ at $x$ and not the pointed subscheme $x \in \operatorname{Spec}(P/f) \subset \mathbf{A}^{n}_{k}$ the equation defines. Indeed, both $f = x_1^2 + \dots + x_n^2$ and $-f$ define the standard node, but the arithmetic types are respectively $\langle 2^n \rangle$ and $\langle (-2)^n \rangle$. For $n$ is odd, these two classes are equal only when $-1$ is a perfect square.
When $n$ is even, the arithmetic type has better invariance properties. In this case, the arithmetic type is an invariant of the pointed $k$-scheme $x \in \operatorname{Spec}(\widehat{P}/f)$, as the following lemma shows.
\begin{lm} \label{Lemma: ContactInv}
Assume $n$ is even. If
\begin{align*}
x \in X =& \operatorname{Spec}(P_{x}/f), \\
y \in Y =& \operatorname{Spec}(P_{y}/g)
\end{align*}
are nodes with the property that the local algebras $\widehat{{ \mathcal O}}_{X, x}$ and $\widehat{{ \mathcal O}}_{Y, y}$ are isomorphic, then
\[
\text{ the arithmetic type of $f$ at $x$} = \text{ the arithmetic type of $g$ at $y$.}
\]
\end{lm}
\begin{proof}
By a descent argument, we can reduce to the case where both residue fields $k(x)$ and $k(y)$ equal $k$. Under this assumption, we can further assume $x=y$ is the origin, so the two local degrees $\deg^{\mathbf{A}^{1}}_{x}(\operatorname{grad}(f))$ and $\deg^{\mathbf{A}^{1}}_{y}(\operatorname{grad}(g))$ are respectively the classes of the Hessians $\det \begin{pmatrix} \frac{\partial^{2} f}{\partial x_i \partial x_j}(0) \end{pmatrix}$ and $\det \begin{pmatrix} \frac{\partial^{2} g}{\partial y_i \partial y_j}(0) \end{pmatrix}$. We have assumed that the local algebras $\widehat{{ \mathcal O}}_{X, x} = \widehat{P}_{x}/f$ and $\widehat{{ \mathcal O}}_{Y, y} = \widehat{P}_{y}/g$ are isomorphic, so suppose $y_1 \mapsto a_1(x_1, \dots, x_n), \dots, y_n \mapsto a_{n}(x_1, \dots, x_n)$ defines an isomorphism. The series $a_1, \dots, a_n$ must satisfy
\begin{equation} \label{Eqn: ContactEquivalence}
f= u \cdot g(a_1, \dots, a_n) \text{ for some $u \in (\widehat{P}_{x})^{\ast}$.}
\end{equation}
Computing the Hessian of $f$ using \eqref{Eqn: ContactEquivalence}, we deduce
\[
\det \begin{pmatrix} \frac{\partial^{2} f}{\partial x_i \partial x_j}(0) \end{pmatrix} = u(0)^n \cdot \det \begin{pmatrix} \frac{\partial^{2} a}{\partial x_i \partial x_j}(0) \end{pmatrix}^2 \cdot \det \begin{pmatrix} \frac{\partial^{2} g}{\partial y_i \partial y_j}(0) \end{pmatrix}
\]
Since $n$ is even, this shows that the Hessian of $f$ at $0$ differs from the Hessian of $g$ at $0$ by a square in $k$, so the two Hessians represent the same element of $\operatorname{GW}(k)$.
\end{proof}
While the arithmetic type is an invariant of a node when $n$ is even, it does not, in general, determine the isomorphic class of a node, as the following example shows.
\begin{expl}
The equations $f=x^2+y^2+z^2+w^2 \in \mathbf{R}[w, x, y, z]$ and $g=x^2+y^2-z^2-w^2 \in \mathbf{R}[w, x, y, z]$ both define a node at the origin with arithmetic type $\langle +1 \rangle$. The algebras $\mathbf{R}[[w, x, y, z]]/f$ and $\mathbf{R}[[w, x, y, z]]/g$ are not isomorphic because e.g.~the projectivized tangent cones are not isomorphic (as $\operatorname{Proj}(\mathbf{R}[W, X, Y, Z]/ X^2+Y^2-Z^2-W^2)$ has infinitely many $\mathbf{R}$-points, but $\operatorname{Proj}(\mathbf{R}[W, X, Y, Z]/ X^2+Y^2+Z^2+W^2)$ has none).
\end{expl}
We now use the notion of arithmetic type to identify $\mu^{\mathbf{A}^1}(f)$ as a count of nodes.
\begin{lm} \label{Lemma: ConserveDegree}
Suppose that $g \colon \mathbf{A}^n_{k} \to \mathbf{A}^{n}_{k}$ is a finite morphism that is \'{e}tale at every point $x \in g^{-1}(0) - \{0 \}$. For any $\overline{y} \in \mathbf{A}^{n}_{k}(k)$ such that $g$ is \'{e}tale at every point of $x \in g^{-1}(\overline{y})$, we have
\begin{equation} \label{Eqn: LocalDegreeConserved}
\sum_{x \in g^{-1}(0)} \deg^{\mathbf{A}^1}_{x}(g) = \sum_{x \in g^{-1}(\overline{y})} \deg^{\mathbf{A}^1}_{x}(g).
\end{equation}
\end{lm}
\begin{proof}
By the Lemma~\ref{loc_degree_etale_point} and Main Theorem, we have $\deg_{x}^{\mathbf{A}^1}(g) = w_{x}(g)$ for all $x$'s appearing in Equation~\eqref{Eqn: LocalDegreeConserved}, and the analogous result for $w_{x}(g)$ is Corollary~\ref{Co: TwoFibersSame}.
\end{proof}
\begin{co} \label{Corollary: SingularityCount}
Let $X = \operatorname{Spec}(P_{x}/f)$ be a hypersurface defined by an equation $f$ such that $\operatorname{grad}(f) \colon \mathbf{A}^{n}_{k} \to \mathbf{A}^{n}_{k}$ is a finite morphism that is \'{e}tale at every point of $x \in g^{-1}(0)$. If $(a_1, \dots, a_n) \in \mathbf{A}^{n}_{k}(k)$ has the property that the family
\begin{gather}
\mathbf{A}^{n}_{k} \to \mathbf{A}^{1}_{k}, \label{Eqn: NodalFamily} \\
x \mapsto f(x) + \sum a_i x_i \notag
\end{gather}
has only nodal fibers, say $x_1 \in X_1, \dots, x_m \in X_m$, then
\begin{equation} \label{Eqn: XBifurcatesToNodesTwo}
\sum_{x \in X} \mu_{x}^{\mathbf{A}^1}(f) = \sum (\text{arithmetic type of $f$ at $x_i \in X_i$}) \text{ in $\operatorname{GW}(k)$.}
\end{equation}
If additionally $f$ is separable, then there exists a nonempty Zariski open subset $U \subset\mathbf{A}^{n}_{k}$ such that \eqref{Eqn: NodalFamily} has only nodal fibers for $(a_1, \dots, a_n) \in U(k)$.
\end{co}
\begin{proof}
Equation~\eqref{Eqn: XBifurcatesToNodesTwo} is a restatement of Lemma~\ref{Lemma: ConserveDegree}.
To construct $U$, observe that since $\operatorname{grad} f$ is separable, the locus of points $V \subset \mathbf{A}^n_{k}$ where $\operatorname{grad} f$ is \'{e}tale contains the generic point of $\mathbf{A}^n_k$ and hence is a nonempty Zariski open subset. The subset $\operatorname{grad} f(\mathbf{A}^{n}_{k} - V) \subset \mathbf{A}^{n}_{k}$ is closed because $\operatorname{grad} f$ is proper and so $U := \mathbf{A}^{n}_{k} - \operatorname{grad} f(\mathbf{A}^{n}_{k} - V)$ has the desired properties.
\end{proof}
A special case of Corollary~\ref{Corollary: SingularityCount} is the result about the cusp (or $A_2$) singularity discussed at the end of the introduction. From Table~\ref{Table: ADEsing}, we see that the equation $f$ of the cusp satisfies $\mu^{\mathbf{A}^1}(f) = \mathbf{H}$. When $k=\mathbf{Q}_5$ (the $5$-adic numbers), a comparison of discriminants shows $\mu^{\mathbf{A}^1}(f) \ne \langle 1 \rangle + \langle 2 \rangle$, so the cusp cannot bifurcate to the split node $\{ x_1^2 + x_2^2 = 0 \}$ and the nonsplit node $\{ x_1^2+ 2 \cdot x_2^2=0\}$, in the sense of Corollary~\ref{Corollary: SingularityCount}.
Somewhat more generally, up to isomorphism, the nodes over $k=\mathbf{Q}_5$ with residue field $k$ are
\begin{equation} \label{Eqn: 5adicNodes}
x_1^2 + x_2^2, x_1^2 + 2 \cdot x_2^2, x_1^2 + 5 \cdot x_2^2,\text{ and } x_1^2 + 5 \cdot 2 \cdot x_2^2.
\end{equation}
If the cusp bifurcates to two of these nodes, then, by Corollary~\ref{Corollary: SingularityCount}, the two nodes must be isomorphic. There are no further obstructions: the family $x \mapsto f(x) + -u^2/3 \cdot x$ contains two nodes which are isomorphic to $\{ x_1^2 + u \cdot x_2^2=0\}$.
The nodes in \eqref{Eqn: 5adicNodes} are all nodes with residue field $k$, and the cusp can bifurcate to more complicated nodes. For example, the only singular fiber of $x \mapsto f(x) + 3 \cdot 5 \cdot x_2$ is the fiber over the closed point $(t^2 + 4 \cdot 5^3)$, a closed point with residue field $\mathbf{Q}_{5}(\sqrt{-5})$. This family illustrates the reason for defining the arithmetic type as we have done: the arithmetic type of the nodal fiber is the arithmetic type of an equation with coefficients in a nontrivial extension of $k$, namely the equation in \eqref{Eqn: LastTypeExample}.
Additional examples describing the collections of nodes that a singularity can bifurcate to can be found in \cite{oberwolffach}.
\begin{rmk} \label{Rmk: CharTwo}
We conclude with a remark about the assumption that $\operatorname{char} k \ne 2$. When $\operatorname{char} k = 2$, the equation $x_1^2 + \dots + x_n^2 = (x_1+\dots+x_n)^2$ does not define an isolated singularity, and instead the standard node should be defined as the hypersurface singularity defined by
\[
f(x) = \begin{cases}
x_{1}^2 + x_{2} x_{3} + \dots + x_{x-1} x_{n} & \text{$n$ odd;} \\
x_{1}^{2} + x_{1} x_{2} + \dots + x_{n-1} x_{n} & \text{$n$ even.}
\end{cases}
\]
Taking this as the definition of the standard node, one can then define arbitrary nodes and their arithmetic types as before, but this arithmetic type is a weak invariant. For example, consider the node that is defined by $f(x) = x_1^2 + x_1 x_2 + u x_2^2$ for $u \in k$. The gradient function $\operatorname{grad} f(x)=(x_2, x_1)$ does not depend on $u$ so the arithmetic type, and any other invariant obtained from the gradient, does not depend on $u$. The isomorphism class of the node does depend on $u$: the isomorphism class is classified by the image of $u$ in $k/\{ v^2 + v \colon v \in k \}$.
\end{rmk}
\section{Acknowledgements}
TO BE ADDED AFTER THE REFEREE PROCESS.
Jesse Leo Kass was partially sponsored by the National Security Agency under Grant Number H98230-15-1-0264. The United States Government is authorized to reproduce and distribute reprints notwithstanding any copyright notation herein. This manuscript is submitted for publication with the understanding that the United States Government is authorized to reproduce and distribute reprints.
Kirsten Wickelgren was partially supported by National Science Foundation Award DMS-1406380 and DMS-1552730.
}
\bibliographystyle{alpha}
|
\section{Introduction}
A large fraction of stars ($\sim{}10$ per cent, \citealt{raghavan2010,riddle2015}) are found to be a part of hierarchical triple systems (HTs) whereby the inner binary is orbited by a tertiary body with a much larger semi-major axis. HTs exhibit interesting dynamical effects; if the tertiary body is inclined with respect to the inner binary it can induce oscillations in the orbital parameters of the inner binary, the so-called eccentric Kozai-Lidov (KL) resonance \citep{kozai1962,lidov1962}. Notably, the tertiary body causes the eccentricity of the inner binary to oscillate between two extremes, with a maximum value given by $e_{\rm max} \approx \sqrt{1-5/3 \, \cos{}^2 i}$, where $i$ is the inclination of the tertiary orbit with respect to the inner binary (in the limit of a test particle secondary when the three-body Hamiltonian is expanded to quadrupole order). These oscillations occur over the timescale \citep{holman1997,kiseleva1998,antognini2014}
\begin{equation} \label{eq:KL}
T_{\rm K} = \frac{2P_t^2}{3\,{}\pi \,{}P_b} (1-e_t^2)^{3/2}\,{} \frac{m_b+m_t}{m_t},
\end{equation}
where $P$ is the period, $e$ the eccentricity, $m$ the mass and the subscripts $b$ and $t$ denote the inner binary pair and tertiary, respectively.
If the inner binary system is composed of two compact objects, KL cycles can drive the inner binary to merge via gravitational wave (GW) emission. The timescale for such a merger can be approximated as \citep{peters1964}
\begin{eqnarray}\label{eq:peters}
t_{\rm GW} \approx 2 \, \, {\rm Gyr} \left (\frac{a_b}{0.01AU} \right)^4 (1-e_b^2)^{7/2} \left( \frac{1\,{}M_{\odot}}{m_1} \right) \nonumber\\ \left( \frac{1\,{}M_{\odot}}{m_2} \right)\left( \frac{2\,{}M_{\odot}}{m_1+m_2} \right)
\end{eqnarray}
where $a_b$ is the semi-major axis of the inner binary, $e_b$ the eccentricity of the binary, and $m_1$ and $m_2$ the respective masses of the two members of the binary pair. It is therefore clear that should resonance effects lead to the eccentricity of the inner binary to approach unity, the binary may merge in a very short time.
KL cycles can trigger mergers of compact-object binaries (e.g. \citealt{miller2002,naoz2013,antonini2012,antonini2014,antognini2014,antognini2015,antognini2016,antonini2016}), enhance the formation of blue straggler stars \citep{naoz2014,antonini2016}, affect the orbits of planets (e.g. \citealt{martin2016}), and drive the formation of low-mass X-ray binaries \citep{naoz2015}. Moreover, KL cycles can enhance double white-dwarf mergers enough to explain the type~Ia SN rate \citep{thompson2011}. These effects are particularly important in globular clusters \citep{antonini2016,naoz2015}, where the high central stellar density favours the formation of stellar triples.
Quantifying the impact of KL resonance on the merger rate of compact objects is of extreme importance in light of the recent first direct detection of GWs by the Advanced LIGO detectors \citep{abbott2016,LIGO2016,LIGO2016b,GW151226a,GW151226b}. Two GW events have been observed by Advanced LIGO so far, GW150914 \citep{abbott2016} and GW151226 \citep{GW151226a}. Both events are associated with the merger of double black hole (BH) systems. A third GW signal, LVT151012, might be associated with the merger of another double BH system, but has low significance.
In this paper, we focus on young dense star clusters (SCs) and open clusters, and we investigate the impact of KL resonances on the merger of BH binaries. We re-simulate HTs that have been found to form self consistently in $N$-body simulations \citep{ziosi2014}, accounting for the effects of KL oscillations so as to discern the effect of KL resonance on the merger rate of BH-BH systems. The paper is organized as follows: in Section 2 we outline the methods and simulations; in Section 3 we describe and discuss our results.
The conclusions are presented in Section~4.
\section{Numerical methods and setup}
The initial conditions were drawn from a set of 600 direct $N$-body realisations of young SCs, already discussed in \cite{mapelli2014} and in \cite{ziosi2014}. We extracted 570 HTs composed of three BHs or two BHs and a neutron star (NS) from these simulations and then re-simulated them with a regularized algorithm including the 2.5 post-Newtonian (PN) term, but not the 1PN term. In the following, we discuss the details of the initial conditions, of the regularized code, and of the simulations.
\subsection{The initial conditions}
The 600 direct $N$-body simulations we start from were run using the {\sc starlab}\footnote{\tt http://www.sns.ias.edu/$\sim{}$starlab/} public software environment \citep{portegieszwart2001}. {\sc kira}, the direct $N$-body integrator included in {\sc starlab}, implements a Hermite 4th order integration algorithm \citep{makino1992} and a neighbours--perturbers scheme to integrate tight binaries and multiple systems. This means that each binary (or multiple system) is treated in a different way according to the strength of the interactions with other stars. If the binary is unperturbed (i.e. it has no strong perturbers according to the criterion discussed in \citealt{portegieszwart2001}), it is integrated analytically as two-body motion: it is seen by the other stars in the system as a point mass. If the binary has close perturbers, it is resolved into its components and integrated through direct summation. No regularization schemes are applied.
We included stellar and binary evolution, using the modified version of {\sc starlab} described in \cite{mapellibressan2013} and \cite{mapelli2016}. Stars evolve in radius, temperature and luminosity at different metallicities according to the polynomial fitting formulae by \cite{hurley2000}. A treatment of stellar winds is included for both main sequence and post-main sequence stars (\citealt{portegieszwart1996,vink2001,vink2005,belczynski2010}; see \citealt{mapelli2013} for a complete description).
The formation of stellar remnants is implemented as described in \cite{mapelli2013}. In particular, BH masses at various metallicities follow the distribution described in fig. 1 of \cite{mapelli2013} (see also \citealt{fryer1999,fryer2001, fryer2012,mapelli2009,mapelli2010,spera2015,spera2016}). If the final mass $m_{\rm fin}$ of the progenitor star (i.e. the mass bound to a star immediately before the collapse) is $>40$ M$_\odot$, we assume that the supernova (SN) fails and that the star collapses quietly to a BH.
The mass of a BH born from direct collapse is similar to the final mass of the progenitor star. Thus, BHs with mass up to $\sim{}80$ M$_\odot$ ($\sim{}40$ M$_\odot$) can form if the metallicity of the progenitor is $Z\sim{}0.01$ Z$_\odot$ ($Z\sim{}0.1$ Z$_\odot$).
NSs are assumed to receive a natal kick drawn from the distribution of \cite{hartman1997}. BHs that form from quiet collapse are assumed to receive no natal kick \citep{fryer2012}. For BHs that form from a SN explosion, the natal kicks were drawn from the same distribution as NSs but scaled as $m_{\rm NS}/m_{\rm BH}$ (where $m_{\rm NS}=1.3$ M$_\odot{}$ is the typical mass of a NS, while $m_{\rm BH}$ is the mass of the considered BH), to preserve linear momentum.
The 600 runs we employ to derive the initial conditions include $N=5500$ particles each, corresponding to an average total mass $M\sim{}3.5\times{}10^3$ M$_\odot{}$. The SCs are modelled as King models \citep{king1966} with virial radius $r_{\rm v}=1$ pc, core radius $r_{\rm c}\sim{}0.4$ pc, half-mass radius $r_{\rm hm}\sim{}0.8$ pc, and central dimensionless potential $W_0=5$. These values are typical of intermediate-mass SCs in the Milky Way, such as the Orion Nebula Cluster \citep{portegieszwart2010}. We generate star masses according to a Kroupa initial mass function (IMF, \citealt{kroupa2001}) with minimum mass $m_{\rm low}=0.1$ M$_\odot$ and maximum mass $m_{\rm up}=150$ M$_\odot$. We include a fraction $f_{\rm PB}=0.18$ of stars in primordial binaries. This is a lower limit to the binary fraction observed in young SCs (e.g. \citealt{li2013}), but higher primordial binary fractions are prohibitive for the integration time. Additional binaries form by dynamical encounters. We consider three different metallicities: $Z=$ 1, 0.1, and 0.01 Z$_\odot$, respectively (here we assume Z$_\odot{}=0.02$). We simulate 200 different realizations for each of these metallicities, to increase the statistics and filter out statistical fluctuations. Each SC has been simulated in isolation for $t=100$ Myr.
\cite{ziosi2014} analysed the formation of BH-BH binaries in these 600 simulations. They found that the vast majority ($\sim{}97$ per cent) of BH-BH binaries come from dynamical exchanges. Thus, SC dynamics is extremely important for the formation of BH-BH binaries (see also \citealt{downing2010,downing2011,oleary2006,sadowski2008,clausen2013,morscher2015,rodriguez2015,oleary2016,chatterjee2016,giersz2015,mapelli2016,hurley2016} for similar conclusions). However, most BH-BH binaries are too loose to emit detectable gravitational waves: only 7 out of 2096 BH-BH binaries are expected to merge within a Hubble time in the sample of \cite{ziosi2014}. This implies an expected merger rate of a few $\times{}10^{-3}$ Gpc$^{-3}$ yr$^{-1}$. The minimum merger rate consistent with the two detected events, GW150914 and GW151226, is $\sim{}9\times{}10^{-3}$ Gpc$^{-3}$ yr$^{-1}$ \citep{abbott2016,GW151226a}.
570 HTs composed of three BHs or two BHs and a NS form in the simulations analysed by \cite{ziosi2014}, approximately one per simulated SC. Triples hosting NSs are a minority: the inner binary is a BH-NS binary only in 5 cases, while the NS is the tertiary body in 3 additional systems. In contrast, systems composed of three BHs are a striking majority ($\sim{}99$ per cent of all simulated compact-object HTs). Since 2096 BH-BH binaries form in all simulations by \cite{ziosi2014}, this means that $\sim{}27$ per cent of all BH-BH binaries went through the formation of a HT with a third BH (or NS) at least once during the simulations. This indicates that the formation of triple BH systems is a common feature in SCs.
\begin{figure*}
\begin{center}
\includegraphics[width=15cm]{fig1.eps}
\caption{\label{fig:fig1}
In the top row, the left-hand panel shows the distribution of the masses of the members of the inner binary ($m_i$, with $i=1,\,{}2$), while the right-hand panel shows the distribution of the masses of the tertiary ($m_{\rm t}$). In the bottom row, the left-hand panel and the right hand panel show the chirp mass $m_{\rm c}=(m_1\,{}m_2)^{3/5}\,{}(m_1+m_2)^{-1/5}$ and the total mass $m_{\rm b}=m_1+m_2$ of the inner binary, respectively. In all panels, diagonally hatched blue histogram: $Z=0.01$ Z$_\odot$; vertically hatched black histogram: $Z=0.1$ Z$_\odot$; horizontally hatched red histogram: $Z={\rm Z}_\odot$.
}
\end{center}
\end{figure*}
The dynamical integrator included in {\sc starlab} is not sufficiently accurate to evolve HTs accounting for KL oscillations, because no regularization scheme is implemented in {\sc kira}. This issue is particularly severe for high eccentricity binaries, such as those where KL oscillations are effective. Moreover, PN terms might be very important in the final evolution of some KL systems. Thus, we selected all triple BH systems from our {\sc starlab} simulations, and we re-simulated them with a new regularized $N$-body code including the 2.5PN term.
\subsection{The regularization code and post-Newtonian terms}
All HTs are evolved by means of a fully regularized $N$-body code that implements the Mikkola's algorithmic regularization (MAR, \citealt{mikkola1999a,mikkola1999b}). This code is particularly suitable for studying the dynamical evolution of few-body systems in which strong gravitational encounters are very frequent and the mass ratio between the interacting objects is large. The MAR scheme removes the singularity of the two-body gravitational potential for $r\rightarrow{}0$, by means of a transformation of the time coordinate (see \citealt{mikkola1999a} for the details). Our implementation uses a leapfrog scheme in combination with the Bulirsh-Stoer extrapolation algorithm \citep{stoer1980} to increase the accuracy of the numerical results.
We also included the dissipative relativistic correction to the Newtonian forces that accounts for the quadrupole gravitational radiation. This correction, indicated as 2.5PN, comes from the post-Newtonian approximation theory
and is of order $1/c^5$, where $c$ is the speed of light (e.g. \citealt{kupi2006}). Since the 2.5PN term depends on the velocity of the particles, we modified the leapfrog scheme into a generalized midpoint method to maintain the correct symmetry of the integration scheme in the case of velocity dependent perturbations \citep{mikkola2006,mikkola2008}. In this way, we can successfully embed the integrator inside the Bulirsh-Stoer scheme to increase the accuracy of the integration. Other post-Newtonian terms are not included in the current version of the code.
The code integrates the equations of motion employing relative coordinates by means of the so called chain structure. This change of coordinates reduces round-off errors significantly \citep{aarseth2003}.
At present, this code is a sub-module of the direct $N$-body code HiGPUs-R which
is still under development (Spera, in preparation; see \citealt{capuzzo2013} for the current non-regularized version of HiGPUs). Still, it can be used as a stand-alone tool to study the dynamical evolution of few-body systems with very high precision.
\subsection{Simulations and {\it caveats}}\label{sec:simulations}
We take each of the 570 HTs at the time it first appears in the $N$-body simulation. Then, we simulate it in isolation with the regularized code for a Hubble time ($t_{\rm H}=14$ Gyr). To identify genuine KL oscillations in our simulations, we search for periodic peaks over a timescale of the same order as the theoretical Kozai timescale (equation~\ref{eq:KL}). If such peaks are found, KL oscillations are considered important.
Before analyzing the results of our simulations, we briefly discuss their main issues. Simulating the HTs in isolation, we neglect possible perturbers. Perturbers are expected to be important for binaries and multiple systems in dense SCs. A simple estimate of the three-body encounter rate (based on the geometrical cross-section of a binary corrected by gravitational focusing, e.g. \citealt{colpi2003}) gives
\begin{eqnarray}
R_{\rm en}=\frac{2\,{}\pi{}G\,{}m_{\rm b}\,{}a_{\rm b}\,{}n_\ast{}}{\sigma{}_\ast{}}\sim{}\nonumber{}\\0.1\,{}{\rm Myr^{-1}}\left(\frac{m_{\rm b}}{50\,{}{\rm M}_\odot}\right)\,{}\left(\frac{a_{\rm b}}{100\,{}{\rm AU}}\right)\,{}\left(\frac{n_\ast}{10^3 {\rm pc}^{-3}}\right)\,{}\left(\frac{5 {\rm km \,{}s}^{-1}}{\sigma{}_\ast}\right),
\end{eqnarray}
where $m_{\rm b}$ and $a_{\rm b}$ are the binary mass and semi-major axis; $n_{\ast}$ and $\sigma_\ast$ are the number density and velocity dispersion of stars close to the binary, respectively. If the SC survives tidal disruption for $\sim{}100$ Myr, a binary in its core will thus undergo $\sim{}10$ encounters. This value depends on the SC lifetime, on the local density and on the velocity dispersion in the neighborhoud of a binary system.
Close encounters can either break the HT (whose stability is extremely precarious, see e.g. \citealt{aarseth2001}) or re-orient it. It is even possible that re-orientation favours the merger of the inner binary. Moreover, weak encounters
with distant perturbers have less dramatic effects but are far more frequent, and can also contribute to harden or soften binaries \citep{heggie1975}.
These {\it caveats} must be considered when interpreting the results of our work. A further study is needed, to investigate the impact of strong and weak perturbations. On the other hand, the HTs where the inner binary merges triggered by KL oscillations (see Section~\ref{sec:results}) are not affected by strong perturbers during the entire $N$-body simulation performed with {\sc starlab}.
Another important {\it caveat} is represented by relativistic effects. In our simulations, we include the 2.5PN term, but we neglect other post-Newtonian terms. However, general relativistic precession (described by 1PN) can stop eccentricity oscillations by destroying the Kozai resonance (e.g. \citealt{holman1997,hollywood1997,blaes2002,antonini2012}). KL oscillations are not suppressed by relativistic precession only if \citep{blaes2002}
\begin{eqnarray}\label{eq:blaes}
\frac{a_{\rm t}}{a_{\rm b}}<910\,{}\left(\frac{a_{\rm b}}{1000\,{}{\rm AU}}\right)^{1/3}\,{}\left(\frac{m_{\rm b}}{50\,{}{\rm M}_\odot{}}\right)^{-1/3}\nonumber{}\\\times{}\,{}\left(\frac{2\,{}m_{\rm t}}{m_{\rm b}}\right)^{1/3}\, \left(\frac{1-e_{\rm b}^2}{1-e_{\rm t}^2}\right)^{1/2},
\end{eqnarray}
where $a_{\rm b}$ ($a_{\rm t}$) is the semi-major axis of the inner (outer) binary, $e_{\rm b}$ ($e_{\rm t}$) is the eccentricity of the inner (outer) binary, $m_{\rm b}$ is the total mass of the inner binary and $m_{\rm t}$ is the mass of the tertiary body. In the following analysis, we will use this simple analytic formalism to predict whether relativistic precession can influence the merger of simulated systems.
\begin{figure*}
\begin{center}
\includegraphics[width=15cm]{fig2.eps}
\caption{\label{fig:fig2}
In the top row, the left-hand and the right-hand panel show the distribution of the semi-major axis of the inner binary ($a_{\rm b}$) and of the outer binary ($a_{\rm t}$), respectively. In the bottom row, the left-hand and the right-hand panel show the distribution of the eccentricity of the inner binary ($e_{\rm b}$) and of the outer binary ($e_{\rm t}$), respectively. In all panels, the black diagonally hatched histogram refers to the initial conditions, while the red open histogram indicates the end of the simulations.
}
\end{center}
\end{figure*}
\section{Results}\label{sec:results}
\subsection{Orbital properties of simulated HTs}
Figure~\ref{fig:fig1} shows the relevant masses of the simulated systems, divided by metallicity. The masses depend on the metallicity by construction, according to the adopted recipes of stellar evolution and stellar winds \citep{mapelli2013}. On average, the tertiary is slightly less massive than both members of the binary. In fact, if the tertiary was more massive than any member of the binary, a dynamical exchange would occur very fast, because the probability of an exchange strongly depends on the mass ratio of the involved bodies (\citealt{mapelli2014} and references therein). A dynamical exchange leads to the ejection of one of the former members of the binary and prevents the formation of a HT.
The chirp mass of the inner binary, defined as $m_{\rm c}=(m_1\,{}m_2)^{3/5}(m_1+m_2)^{-1/5}$, where $m_1$ and $m_2$ are the masses of the two components of the binary, is particularly important for GWs, since it determines how fast the binary sweeps, or chirps, through a frequency band (the amplitude and the frequency of GWs scale as $m_{\rm c}^{5/3}$ and $m_{\rm c}^{-5/8}$, respectively). The range of chirp masses is $\sim{}5-25$ M$_\odot$, $\sim{}5-36$ M$_\odot$, and $\sim{}5-65$ M$_\odot$ at $Z=1$, 0.1 and 0.01 Z$_\odot$, respectively. Correspondingly, the range of total masses of the binary $m_{\rm b}=m_1+m_2$ is $\sim{}10-55$ M$_\odot$, $\sim{}10-80$ M$_\odot$, and $\sim{}10-150$ M$_\odot$ at $Z=1$, 0.1 and 0.01 Z$_\odot$, respectively. This range of chirp and total masses is consistent with the events that were recently detected by Advanced LIGO \citep{LIGO2016}.
Figure~\ref{fig:fig2} shows the distribution of initial eccentricity and semi-major axis of the inner and outer binary at the start and the end of the simulations, across all systems. We found no global change of the distribution of semi-major axis, consistent with the fact that KL resonance does not affect the energy of the binary. Only exchanges and PN terms can change the semi-major axis, but they affect only a small minority of our runs. The Kolmogorov-Smirnov (KS) test indicates that the probability that the initial and final semi-major axes are drawn from the same distribution is $\sim{}1$ and $\sim{}0.7$ for the inner and outer binary, respectively. The majority of inner binaries have a semi-major axis $a_{\rm b}\sim{}10^2-10^3$ AU, while the semi-major axis of the outer binary is much higher ($\sim{}10^4-10^5$ AU).
As to the eccentricity distribution, we find no apparent change of eccentricity distribution, consistent with the fact that KL oscillations do not change the average eccentricity in time. The KS test indicates that the probability that the initial and final eccentricities are drawn from the same distribution is $\sim{}0.4$ and $\sim{}0.2$ for the inner and outer binary, respectively. These probabilities are quite lower than for the semi-major axis, but we cannot reject the hypothesis that these distributions come from the same one.
Finally, Figure~\ref{fig:fig3} shows the distribution of inclinations $i$ between the orbital plane of the inner and outer binary. The distribution is nearly flat and does not change significantly during the integration (top panel of Fig.~\ref{fig:fig3}), consistently with the behaviour of KL resonance. The KS test indicates that the probability that the initial and final inclinations are drawn from the same distribution is $\sim{}1$. In Figures~\ref{fig:fig2} and \ref{fig:fig3}, we do not distinguish the three metallicities because, as already shown by \cite{ziosi2014}, the orbital properties do not depend on the metallicity significantly.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{fig3.eps}
\caption{\label{fig:fig3}
Distribution of the inclinations between the plane of the inner binary and that of the outer binary. Top panel: inclinations at the beginning (black diagonally hatched histogram) and at the end of the simulations (red open histogram) for all simulated systems. Bottom panel: initial inclinations of systems in class~A (dark green open histogram), class~B (turquoise horizontally hatched histogram), class~C (orange diagonally hatched histogram), class~A1 (violet vertically hatched histogram) and class~A2 (violet filled histogram). Class~A2 corresponds to the three merging systems. See Table~\ref{tab:table1} and Section~\ref{sec:statistics} for details on the classification of HTs in these groups.
}
\end{center}
\end{figure}
\subsection{Statistics of KL systems}\label{sec:statistics}
\begin{figure}
\begin{center}
\includegraphics[height=12cm]{fig4.eps}
\caption{\label{fig:fig4}
Properties of the inner binary of a HT showing regular KL cycles. From top to bottom: time dependence of semi-major axis ($a_{\rm b}$), eccentricity ($e_{\rm b}$), inclination with respect to the outer binary ($i$), and coalescence timescale $t_{\rm GW}$ (eq.~\ref{eq:peters}). Blue dotted line in the bottom panel: Hubble time $t_{\rm H}$. This simulation was run with the 2.5PN term.}
\end{center}
\end{figure}
\begin{table}
\begin{center}
\caption{\label{tab:table1}
Classification of the HTs.}
\leavevmode
\begin{tabular}[!h]{c|ccccc}
\hline
Class & A & A1 & A2 & B & C \\
\hline
Frequency & 502 & 30 & 3 & 10 & 58 \\
Percentage & 88.0 & 5.3 & 0.5 & 1.8 & 10.2 \\
\noalign{\vspace{0.1cm}}
\hline
\end{tabular}
\end{center}
\footnotesize{Class~A: systems where KL oscillations are effective; class~A1: sub-sample of class~A where $t_{\rm GW}<t_{\rm H}$ for a fraction of the KL cycle; class~A2: systems that merge during the simulations (sub-sample of class~A1); class~B: exchanges; class~C: other systems (where Kozai is not important).}
\end{table}
Table~\ref{tab:table1} summarizes our classification of the 570 simulated systems. We classify each system as either (i) a system that undergoes KL oscillations (hereafter, class~A); (ii) a system that undergoes a dynamical exchange, i.e. the tertiary body becomes member of the inner binary and, usually, the former member of the inner binary is ejected from the system (class~B); (iii) any other system (class~C). These are all systems that do not undergo an exchange, and in which the oscillations do not match the Kozai timescale and/or their amplitude is $<1$ per cent of the eccentricity.
Among the systems that undergo KL oscillations (class~A) we then consider the sub-sample of systems for which $t_{\rm GW}<t_{\rm H}$ for at least a fraction of the KL cycle (where $t_{\rm GW}$ is defined in equation~\ref{eq:peters} and $t_{\rm H}$ is the Hubble time). These systems are labelled as class~A1. A further sub-sample of class~A1 is represented by systems that merge during the simulations as an effect of KL oscillations and GW emission (these will be indicated as class~A2).
From Table~\ref{tab:table1} it is apparent that a large fraction of HTs ($\sim{}88$ per cent) exhibit KL oscillations (class~A), whilst $\sim{}2$ per cent undergo an exchange (class~B). In $\sim{}10$ per cent of HTs KL resonance is not efficient (class~C).
Furthermore, 30 inner binaries, i.e $\sim{}5$ per cent of all simulated systems, have $t_{\rm GW}<t_{\rm H}$ during a fraction of their KL cycles, because the eccentricity becomes $\sim{}1$ (class~A1). However, in 27 out of 30 systems (among which, the system shown in Figure~\ref{fig:fig4}), the time spent in the configuration with $t_{\rm GW}<t_{\rm H}$ is not sufficiently long to produce effective orbital decay by GW emission (which, in our simulations, is modelled as 2.5PN term).
Only $3$ inner binaries ($0.5$ per cent of the simulated systems) are observed to merge, as a consequence of KL oscillations (class~A2). All three merging systems are triple BH systems; no NS-BH system merges.
The orbital decay induced by GW emission is effective in the three merging systems. Figure~\ref{fig:fig5} shows the time dependence of semi-major axis, eccentricity, inclination, and coalescence timescale ($t_{\rm GW}$) of the three merging systems. For all of them, we compare the evolution with the 2.5PN term to the evolution without 2.5PN term.
The three merging systems (class~A2) all have inclination $i\sim{}90$ DEG, as it is shown in the bottom panel of Figure~\ref{fig:fig3}. Similarly, most systems in class~A1 have inclination $i\sim{}90$ DEG. In contrast, exchanged (class~B) and `other' systems (class~C) tend to have either smaller or larger inclinations. This is consistent with the expectation that KL oscillations are maximally efficient for inclinations close to $\sim{}90$ DEG.
For the three merging systems we also checked the effect of relativistic precession by using equation~\ref{eq:blaes}. We found that the ratio $a_{\rm t}/a_{\rm b}$ is smaller than the critical value for relativistic precession to stop KL oscillations for the entire simulation, with the exception of the very last epoch of the merger phase (when $e_{\rm b}>0.99$ and when 2.5PN corrections are very effective, and $a_{\rm b}$ drops in few seconds). Thus, we expect that relativistic precession cannot stop the merger of these three systems.
\begin{figure*}
\begin{center}
\includegraphics[height=12cm]{fig5a}
\includegraphics[height=12cm]{fig5b}
\includegraphics[height=12cm]{fig5c}
\caption{\label{fig:fig5}
Properties of the inner binary in the three HTs merging within a Hubble time. For each system, from top to bottom: time dependence of semi-major axis ($a_{\rm b}$), eccentricity ($e_{\rm b}$), inclination ($i$) with respect to the outer binary, and coalescence timescale $t_{\rm GW}$ (eq.~\ref{eq:peters}). Red dashed line: simulation with 2.5PN term; black solid line: same simulation run without 2.5PN term. Blue dotted line in the bottom panels: Hubble time $t_{\rm H}$.
}
\end{center}
\end{figure*}
\subsection{Properties of merging systems}
Table~\ref{tab:table2} shows the properties of the three systems that merge in a Hubble time because of KL oscillations and GW emission (hereafter, we will call them `K16 merging systems', for brevity). For comparison, we list also the properties of the events detected by Advanced LIGO \citep{abbott2016,GW151226a}, and those of the seven simulated BH-BH systems that were found to merge in a Hubble time by \cite{ziosi2014} (hereafter, `Z14 merging systems'). The seven Z14 merging systems come from the same set of simulations as the BH-BH systems we are considering in this paper, but their merger is not triggered by KL resonance.
In Table~\ref{tab:table2}, we also show the quantity $\tilde{t}_{\rm GW}$ (effective coalescence time), defined as
\begin{equation}\label{eq:tildetgw}
\tilde{t}_{\rm GW}=t_{\rm bin}+t_{\rm GW},
\end{equation}
where $t_{\rm bin}$ is the time elapsed from the formation of the binary till the end of the simulation, and $t_{\rm GW}$ the coalescence timescale defined in eq.~\ref{eq:peters}, estimated at the end of the simulation. For the Z14 merging systems $t_{\rm GW}>t_{\rm bin}$ (because \citealt{ziosi2014} simulated the evolution of SCs for 100 Myr).
In the case of the three K16 merging systems, accounting for the 2.5PN term slows down the integration and makes the simulation stall: $t_{\rm bin}$ is the duration of the simulation till it stalls, and we evaluate $t_{\rm GW}$ at the stalling time. For the three K16 merging systems $t_{\rm GW}\lesssim{}1$ yr, hence $\tilde{t}_{\rm GW}\sim{}t_{\rm bin}$.
Even if the statistics of merging systems is low, we can do several interesting considerations.
\begin{itemize}
\item[i)] In both \cite{ziosi2014} and this paper no BH-BH system is expected to merge within a Hubble time from solar-metallicity progenitors. All merging systems form from stars with metallicity $\lesssim{}0.1$ Z$_\odot$. The first observed merger event, GW150914, is also thought to be associated with low metallicity ($<0.5$ Z$_\odot$), because of the large BH mass \citep{abbott2016}.
\item[ii)] The masses of K16 merging systems are generally higher than the masses of Z14 merging systems. In particular, the mass of the primary BH ($39$ M$_\odot$) in the K16 merging system at $Z=0.1$ Z$_\odot$ is consistent with the mass of the primary BH in GW150914 ($36.2^{+5.2}_{-3.8}$ M$_\odot$, \citealt{abbott2016}).
In contrast, the BH masses of most Z14 merging systems are fairly consistent with those of GW151226 (Table~\ref{tab:table2}).
\item[iii)] All three K16 merging systems involve binaries that formed from a dynamical exchange, while $\sim{}70$ per cent of Z14 merging systems occur in primordial binaries. Thus, dynamics is very important for all K16 merging systems.
\item[iv)] For all simulated merging systems, both Z14 and K16, $\tilde{t}_{\rm GW}<4$ Gyr, much shorter than the Hubble time.
\end{itemize}
Figure~\ref{fig:fig6} shows the systems listed in Table~\ref{tab:table2} in the plane of $m_1,\,{}m_2$. All but one Z14 merging system have relatively low masses (5--20 M$_\odot$), matching the mass of GW151226.
The mass of one of the primary members in K16 matches the mass of the primary member of GW150914, while the mass of the secondary members is slightly lower. The mass of the primary BH in K16 merging systems is always $\gtrsim{}40$ M$_\odot$. This might indicate that Kozai-induced mergers are skewed toward larger BH-BH masses. This consideration is consistent with the fact that HTs form dynamically, and dynamically born systems tend to have larger masses than primordial binaries (e.g. \citealt{mapelli2014} and references therein). However, our sample is far too small to draw any definitive conclusion. A large parameter-space investigation is requested to have statistically significant results.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{fig6.eps}
\caption{\label{fig:fig6}
Mass of the primary member of a merging binary ($m_1$) versus the mass of the secondary member ($m_2$). Black crosses: Z14 merging systems; blue filled pentagons: K16 merging systems; red star with error bars: GW150914; magenta star with error bars: GW151226; turquoise star with error bars: LVT151012.}
\end{center}
\end{figure}
\subsection{Impact on the merger rate}
From the results of Table~\ref{tab:table2} we can determine the merger rate for BH-BH binaries using a Drake-like equation \citep{esposito2015}:
\begin{eqnarray}\label{eq:drake}
R\sim{} 5\,{}{\rm Gpc}^{-3}\,{}{\rm yr}^{-1}\,{}\left(\frac{t_{\rm life}}{10^8\,{}\text{yr}} \right)\,\left(\frac{\rho_{\rm SF}}{0.015\, \text{ M}_{\odot} \text{ yr}^{-1} \text{ Mpc}^{-3}}\right)\nonumber\\\,{}\left(\frac{f_{\rm SF}}{0.8}\right)\,{}\left(\frac{600}{n_{\rm SC}}\right)\,{}\left( \frac{3500\text{ M}_\odot{}}{M_{\rm SC}}\right)\,{}\left(\frac{\sum_{\rm i} \tilde{t}_{\rm GW, i}^{-1}}{8\,{}{\rm Gyr^{-1}}}\right),
\end{eqnarray}
where $\rho_{\rm SF}$ is the cosmological star formation rate density at redshift zero ($1.5 \times 10^{-2} \text{ M}_{\odot} \text{ yr}^{-1} \text{ Mpc}^{-3}$, \citealt{hopkins2006}), $t_{\rm life}$ is the average lifetime of a simulated SC ($=100$ Myr), $f_{\rm SF}$ is the fraction of star formation that occurs in SCs (=0.8 from \citealt{lada2003}), $n_{\rm SC}$ is the number of simulated SCs, $M_{\rm SC}$ is the average mass of the SCs ($=3500 \,{}\text{M}_{\odot}$),
and $\tilde{t}_{\rm GW, i}$ is the effective coalescence timescale of the i-th binary, as given by equation~\ref{eq:tildetgw}. Using the values of $\tilde{t}_{\rm GW}$ given in Table~\ref{tab:table2} for the three K16 merging systems, we find $R\sim{}5$ Gpc$^{-3}$ yr$^{-1}$, if we assume that all metallicities are equally likely.
Applying the same equation (eq.~\ref{eq:drake}) to Z14 merging systems, we find $R\sim{}11$ Gpc$^{-3}$ yr$^{-1}$ (\citealt{ziosi2014} give an estimate of $\sim{}3.5$ Gpc$^{-3}$ yr$^{-1}$ because they assume a model for the metallicity evolution in the local Universe and use a more approximate formula than equation~\ref{eq:drake}). Combining Z14 and K16 merging systems, the rate becomes $R\sim{}16$ Gpc$^{-3}$ yr$^{-1}$. Under the simplest assumptions, KL oscillations can increase the merger rate by $\approx{}50$ per cent.
\begin{table}
\begin{center}
\caption{\label{tab:table2} List of the merging BH-BH systems and their associated parameters at the time of merging.}
\leavevmode
\begin{tabular}{cccccccc}
\hline
Ref. & $Z$ & $\tilde{t}_{\rm GW}$ & $m_1$ & $m_2$ & $m_{\rm c}$ & $m_{\rm t}$ & Bin. \\
& (Z$_\odot$) & (Gyr) & (M$_{\odot}$) & (M$_{\odot}$) & (M$_{\odot}$) & (M$_\odot$) &Type \\
\hline
K16 & 0.01 & 0.16 & 61 & 20 & 30 & 5 & E\\
K16 & 0.01 & 3.5 & 72 & 20 & 32 & 16 & E\\
K16 & 0.1 & 0.5 & 39 & 18 & 23 & 15 & E\\
\hline
Z14 & 0.01 & 0.09 & 50 & 40 & 39 & -- & E \\%& Z001n154idsa1746b4480+14480\\
Z14 & 0.01 & 1.34 & 12 & 8 & 8 & -- & P \\%& Z001n052idsa3370b13370\\
Z14 & 0.01 & 1.76 & 9 & 9 & 8 & -- & P\\%& Z001n194idsa2870b12870\\
Z14 & 0.01 & 2.06 & 15 & 12 & 11 & -- & P \\%Z001n075idsa350b10350
Z14 & 0.1 & 0.20 & 10 & 5 & 6 & -- & E \\%& Z010n200idsa3555b180\\
Z14 & 0.1 & 0.67 & 11 & 7 & 8 & -- & P \\%& Z010n185idsa4340b14340\\
Z14 & 0.1 & 1.49 & 10 & 9 & 8 & -- & P\\%& Z010n162idsa2070b12070\\
\hline
A16a & $<0.5$ & $<14$ & $36.2^{+5.2}_{-3.8}$ & $29.1^{+3.7}_{-4.4}$ & $28.1^{+1.8}_{-1.5}$ & -- & U \\
A16b & -- & $<14$ & $14.2^{+8.3}_{-3.7}$ & $7.5^{+2.3}_{-2.3}$ & $8.9^{+0.3}_{-0.3}$ & -- & U \\
A16b & -- & $<14$ & $23^{+18}_{-6}$ & $13^{+4}_{-5}$ & $15.1^{+1.4}_{-1.1}$ & -- & U \\
\noalign{\vspace{0.1cm}}
\hline
\end{tabular}
\begin{flushleft}
\footnotesize{Column 1 (Ref.): reference of the paper from which the system was taken. K16 means this paper; Z14 indicates \cite{ziosi2014}; A16a and A16b indicate \cite{abbott2016} and \cite{GW151226a}, respectively. The last three rows show the properties of the two GW events (GW150914, GW151226) and the third possible signal (LVT151012) detected by Advanced LIGO \citep{abbott2016,GW151226a}, for comparison. Column 2 ($Z$): metallicity; column 3 ($\tilde{t}_{\rm GW}$): effective coalescence time since the beginning of the simulation (eq.~\ref{eq:tildetgw}); column 4 ($m_1$): mass of the primary member of the inner binary; column 5 ($m_2$): mass of the secondary member of the inner binary; column 6 ($m_{\rm c}$): chirp mass of the inner binary; column 7 ($m_{\rm t}$): mass of the tertiary, when present; column 8 (Binary type): origin of the binary. `P' and `E' stand for primordial and exchanged binary, respectively. In the last column, `U' means unknown origin.}
\end{flushleft}
\end{center}
\end{table}
We can also calculate the detection rate for Advanced LIGO/Virgo, $R_{\rm det}$, by extrapolating the merger rate to a volume set by the radius out to which the mergers produce a GW signal within the instrumental range of Advanced LIGO/Virgo (this range depends on the chirp mass of each binary):
\begin{multline}\label{eq:detec}
R_{\rm det} = \frac{4\pi}{3} \frac{t_{\rm life} \,\rho_{\rm SF}\, f_{\rm SF}}{n_{\rm SC}\,{} M_{\rm SC}} \,{} \left( \frac{d_{H}}{f} \right)^3\frac{1}{m_{\rm c, 10}^{15/6}} \sum_{\rm i} \frac{m_{\rm c, i}^{15/6}}{\tilde{t}_{\rm GW, i}},
\end{multline}
where $f=2.26$ is a correction factor accounting for the random location on the sky and orientation of sources, as well as the non-uniform pattern of detector sensitivity \citep{finn1996,belczynski2013}, $m_{\rm c, 10}=8.7$ M$_\odot$ is the chirp mass of a BH-BH binary composed of two 10 M$_\odot$ BHs, $d_{H}=1$ Gpc is the approximate instrumental range of Advanced LIGO and Virgo \citep{abadie2010} for a binary with $m_{\rm c}=m_{\rm c, 10}=8.7$ M$_\odot$, and $m_{\rm c, i}$ is the chirp mass of i-th merging BH-BH system.
Adopting the same values for $t_{\rm life}$, $\rho{}_{\rm SF}$, $f_{\rm SF}$, $n_{\rm SC}$ and $M_{\rm SC}$ as discussed before, and using $t_{\rm GW, i}$ and $m_{\rm c, i}$ from Table~\ref{tab:table2}, we find $R_{\rm det}\sim{}40$ yr$^{-1}$. For comparison, the detection rate of the seven Z14 merging systems, derived from eq.~\ref{eq:detec} is $R_{\rm det}\sim{}100$ yr$^{-1}$. Combining Z14 and K16 merging systems, the detection rate becomes $R_{\rm det}\sim{}140$ yr$^{-1}$. We note that $d_{H}=1$ Gpc will be reached by Advanced LIGO approximately in the O2 run and that equation~\ref{eq:detec} assumes that our simulated mergers are homogeneously distributed in a sphere of $1$ Gpc. Thus, our estimate is likely an upper limit.
Given the low statistics, these results have large uncertainties and should be considered only as preliminary predictions of the importance of Kozai merging systems with respect to the other merging systems in young SCs. Thus, we find that including the effect of KL resonances can increase the detection rate by $\sim{}40$ per cent. This is a mild effect but is extremely important if we consider the properties of systems that merge by KL resonances. As we discussed in the previous Section, Kozai merging systems have generally higher masses than the other BH-BH systems. Thus, they can be detected at larger distances, and are consistent with the mass of the observed GW150914 event. Distinguishing between Kozai triggered mergers and mergers of primordial binaries can give us insights into the mechanisms that lead to the formation of a BH-BH binary.
\section{Summary}
HTs, i.e. triple systems where an inner binary is orbited by a tertiary body further out, are very common in the local Universe. If the orbital plane of the tertiary is inclined with respect to that of the inner binary, the system might undergo KL resonance: the eccentricity of the inner binary and the inclination between the two orbital planes start oscillating. Since the coalescence timescale of a double compact object binary strongly depends on its eccentricity, the periodic increase of eccentricity due to KL oscillations can affect the merger rate significantly.
In \cite{ziosi2014}, we simulated the formation and the dynamical evolution of double compact-object binaries in young SCs. We found that $\sim{}27$ per cent of all BH-BH and BH-NS binaries (corresponding to 570 systems) are members of a HT with another BH or, in few cases, with a NS. However, the algorithms adopted by \cite{ziosi2014} were not sufficiently accurate to study the development of KL oscillations in these 570 HTs. In this paper, we re-simulate the HTs found in \cite{ziosi2014} by using a new code based on the Mikkola's algorithmic regularization scheme. We also included the 2.5PN term.
We find that $\sim{}88$ per cent of the simulated HTs develop KL oscillations. Less than $\sim{}2$ per cent of the simulated HTs undergo an exchange during the simulation, while the remaining $\sim{}10$ per cent of simulated HTs do not show significant KL oscillations or other interesting dynamical features.
In three runs (i.e. $\sim{}0.5$ per cent of all runs) KL oscillations lead to the merger of the inner binary, if the 2.5PN term is included. Even if this is a small sample to make any strong statements, we note that all three merging systems originated from dynamical exchanges (i.e. none comes from a primordial binary). In contrast, \cite{ziosi2014} reported seven systems merging in a Hubble time, 70 per cent of which were born from primordial binaries and only 30 per cent from dynamical exchanges. Moreover, most of the merging systems simulated by \cite{ziosi2014} have lower masses than the three systems merging by KL oscillations. The BH masses of merging systems found by \cite{ziosi2014} are fairly consistent with those of GW151226 (the second LIGO event), while the three systems merging by KL oscillations have higher masses, similar to those of GW150914 (the first LIGO event).
Dynamical interactions trigger the formation of BH-BH binaries more massive than BH-BH binaries originating from primordial binaries \citep{ziosi2014,morscher2015,giersz2015,mapelli2016}. However, most of these massive dynamically formed BH-BH binaries are not expected to merge, according to the results of \cite{ziosi2014}, who did not account for KL oscillations. In our paper, we show that
KL oscillations might enhance the merger rate of massive dynamically formed BH-BH binaries. This result is particularly important to interpret the formation channel of GW150914: while it is highly controversial whether GW150914 comes from a primordial binary or from a dynamically formed system (e.g. \citealt{LIGO2016b}), such high-mass BH-BH binary can result from the evolution of primordial binaries only with fine tuning (e.g. \citealt{belczynski2016,belczynski2016b,marchant2016}). Our simulations indicate that dynamical mechanisms (e.g. exchanges) trigger the formation of massive BH-BH binaries \citep{ziosi2014}, and KL oscillations can boost the merger rate of such massive dynamically formed BH-BH binaries.
We estimate that the merger rate $R$ of BH-BH binaries from \cite{ziosi2014} is $R\sim{}11$ Gpc$^{-3}$ yr$^{-1}$, while the merger rate of the three systems merging by KL oscillations is $\sim{}5$ Gpc$^{-3}$ yr$^{-1}$. Combining these results, we find that a total merger rate $R\sim{}16$ Gpc$^{-3}$ yr$^{-1}$ is expected from the simulations of \cite{ziosi2014} when accounting for Kozai resonances. While this result is affected by large uncertainties, it is fairly consistent with the constraints posed by GW150914 and GW151226 ($R\sim{}9-240$ Gpc$^{-3}$ yr$^{-1}$, \citealt{GW151226a}).
Our results suffer from a number of issues. First, we simulated the HTs in isolation. While we checked that none of the merging HTs is affected by strong perturbers during the $N$-body simulations of \cite{ziosi2014}, we cannot exclude that even weak perturbations can prevent any of such systems from merging. A full $N$-body simulation with a regularization scheme is requested to check this issue. Moreover, we neglected non-dissipative PN terms. They do not contribute to GW decay but can induce precession of the inner binary and affect the dynamical evolution of the HT. Their contribution will be considered in a forthcoming study. Taking into account all these {\it caveats}, our results indicate that HT formation and Kozai resonance might have a crucial impact on the demographics of GW sources. In particular, KL oscillations can trigger the coalescence of massive BH-BH systems in dynamically formed binaries, leading to an increase of the merger rate by $\approx{}50$ per cent.
\section*{Acknowledgments}
We thank the anonymous referee for their useful comments, which improved this work significantly. We also thank Mauro D'Onofrio, Alessandro Bressan, Ugo Niccol\`o Di Carlo and Enrico Montanari for useful discussions. TK acknowledges financial support from the Erasmus+ programme. MM, MS and BMZ acknowledge financial support from the Italian Ministry of Education, University and Research (MIUR) through grant FIRB 2012 RBFR12PM1F, and from INAF through grant PRIN-2014-14. MM acknowledges financial support from the MERAC Foundation.
|
\section{Introduction}
The search for exotic phases of correlated quantum systems have become one of the most promising -- but also very challenging -- topic within condensed matter theory. It is believed that the interplay of strong quantum fluctuations and frustrated interactions plays the most important role in stabilizing such unconvential phases. However, the most common quantum phases in condensed matter physics display spontaneous symmetry breaking and long-range order. Exotic phases are then often close in parameter space to these ordered phases and it is a notable challenge and an important task to develop strategies and tools allowing to extract quantitative information of such more conventional phases in order to characterize real systems and to guide experiments toward more exotic phases.
A perfect microscopic test bed for this question is the antiferromagnetic Heisenberg model at zero temperature in two dimensions, since it is expected to display various ordered and disordered ground states depending on the degree of geometric frustration and on the value of the total spin. At the same time it is the relevant microscopic description for many experiments. Important examples are the unfrustrated spin-1/2 Heisenberg model on the square lattice which shows long-range N\'{e}el order~\cite{Manousakis91,Reger88,Chakravarty89} as well as the same model on the geometrically frustrated triangular lattice which has been studied extensively over the last decades and exhibits a three-sublattice $120^\circ$ ordered ground state~\cite{Bernu92,Capriotti99,White07}. In both cases the SU(2) symmetry is spontaneously broken and one has gapless spin-wave exitations according to the Goldstone Theorem~\cite{Auerbach94}. In contrast, on the highly frustrated kagome lattice, the spin-1/2 Heisenberg model is believed to realize a quantum spin liquid ground state with exotic topological order \cite{Yan11,Depenbrock12} while the ground state of the spin-$1$ cousin is most likely a spontaneous trimerized phase with long-range singlet order~\cite{Changlani15,Liu15,Li15,Ghosh15,Oitmaa15,Ixert15} and gapped excitations. For larger spins $S>1$, the ground-state of the Heisenberg model on the kagome lattice is again magnetically ordered~\cite{Huse92,Henley09,Goetze11,Oitmaa15}.
Numerically, there are several techniques which can be applied to two-dimensional quantum many-body systems which all have strengths as well as complications. One promising tool which is under active development in recent years are so-called non-perturbative linked-cluster expansions (NLCEs)\cite{Rigol06,Rigol07_1,Rigol07_2,yang11,Tang13,coester15}, which are non-perturbative variants of perturbative linked-cluster expansions (LCEs) where high-order series expansions (SE) are derived in the thermodynamic limit using a full-graph decomposition and the linked-cluster theorem. In \mbox{NLCEs}, perturbation theory on graphs is replaced by non-perturbative tools like exact diagonalization (ED), density matrix renormalization group~\cite{stoudenmire14} or continuous unitary transformations \cite{yang11,coester15}.
All current NLCEs are real-space approaches. As a consequence, one expects convergence in gapped quantum phases due to the finite correlation length and complications for gapless systems. However, as shown recently \cite{Ixert15}, even in gapped phases like the trimerized phase of the spin-$1$ Heisenberg model on the kagome lattice, the NLCE data sequences can become erratic due to peculiar quantum criticalities of lower-dimensional subsets of graphs. Here the trimerized order of the two-dimensional system is not reflected in the calculation on one-dimensional graphs leading to the wrong assignment of fluctuations which are not present in the two-dimensional system. It is therefore necessary to extend the NLCEs in such a way that each cluster ``remembers'' the quantum order out of which it is taken from. For the ordered phase of the transverse-field Ising model on the square lattice this has been already realized successfully in Ref.~\onlinecite{Kallin13}.
In this work we present an extended scheme which also realizes this line of reasoning. We consider clusters to be taken out of a long-range ordered reference state. This state can either be a classical spin-ordered configuration like the N\'{e}el state or a valence bond solid breaking translational symmetry like a trimerized state. These reference states introduce symmetry-breaking edge-fields in the ED of clusters which enable us to perform NLCEs in gapped and gapless ordered quantum phases. We test our scheme on various lattices using either quantum spins 1/2 or 1.
This paper is organized as follows: In Sect.~\ref{Sect:nlce} we explain briefly the main idea behind NLCEs. The edge-field NLCE for ordered quantum systems is introduced in Sect.~\ref{Sect:models} and the tested scaling behaviors of NLCE data is given in Sect.~\ref{Sect:scaling}, followed by the results for the spin-$1/2$ and spin-$1$ Heisenberg models on the square and triangular lattice in Sections~\ref{Sect:results:squareOneHalf} to~\ref{Sect:results:triangularOne} displaying long-range magnetic order. Finally, in Sect.~\ref{Sect:spin1kagome}, the edge-field approach is applied to the trimerized and gapped spin-$1$ kagome Heisenberg model. The paper concludes with a brief summary of the main results.
\section{NLCE}
\label{Sect:nlce}
The essential idea of NLCEs is to exploit the linked-cluster theorem so that actual
numerical calculations are done on finite linked clusters, but the final results are valid
directly in the thermodynamic limit. Generically, NLCEs consist of three steps: i) choosing and generating the families of clusters or topologically distinct graphs used in the LCE, ii) performing numerical calculations on graphs extracting the physical quantities of interest, and iii) determining the reduced contributions specific to each graph and embed these contributions into the infinite lattice.
The details of the choice of graphs (i) will be given below. Concerning ii), we are using ED with the Lanczos algorithm~\cite{Lanczos50} to determine the ground-state energy $E^{\mathcal{G}_\nu}_0$ (as well as the sublattice magnetization for spin-ordered phases) on each graph. Apart from the exponential increase of graphs with the number of sites $N$, the memory needed for ED is the limiting factor of NLCEs.
In the third step iii) we concentrate on the ground-state energy per site $e_0$ and the appropriate order parameter in the thermodynamic limit, e.g.~the sublattice magnetization in long-range spin-ordered phases. In this technical part we only discuss $e_0$ which can be expressed as
\begin{equation}
e_0 = \sum_{j} \nu^{\mathcal{G}_j} \, e_{0,{\rm red}}^{\,\mathcal{G}_j} \quad ,
\label{eq:e0}
\end{equation}
where the sum runs over all linked graphs $\mathcal{G}_j$. The integer number $\nu^{\mathcal{G}_j}$ is the so-called embedding factor specifying the combinatorical number how often graph $\mathcal{G}_j$ can be embedded into the infinite lattice. The reduced contribution $e_{0,{\rm red}}^{\,\mathcal{G}_j}$ specific to graph $\mathcal{G}_j$ results from subtracting all contributions of subgraphs in order to avoid double counting. One has
\begin{equation}
e_{0,{\rm red}}^{\,\mathcal{G}_j} = E_0 ^{\mathcal{G}_j} - \sum_{\mathcal{G}^\prime_{j}\subset \mathcal{G}_{j}} e_{0,{\rm red}}^{\,\mathcal{G}^\prime_{j}} \quad .
\label{eq:e0_tilde}
\end{equation}
Note thate the sum runs over all connected subgraphs of $\mathcal{G}_j$ without identifying topologically equivalent contributions. Depending on the lattice, different resummation schemes could be useful. Here we apply the following four different expansions:
\subsection{Full-graph expansion}
In the full-graph expansion all topological distinct graphs with up to $N$ (effective) sites are generated. We define the order of the expansion as the maximum number of sites $N$ taken into account. Note that the sites can be either single spins or a collection of spins like dimers or trimers which function then as effective sites for the graph expansion. In this scheme the number of graphs grows quickly with the order but the Hilbert spaces for ED are typically small for the considered $N$. At the same time it seems impossible to assign typical length scales to graphs which complicates a scaling towards the infinite-order limit.
\subsection{Rectangular-graph expansion}
In contrast to the full-graph expansion where the number of graphs grows very quickly, the rectangular-graph expansion~\cite{deNeef77,Enting77,Dusuel10,coester15,Kallin13} is restricted to graphs which are rectangular with linear dimensions $L_x$ and $L_y$ so that \mbox{$N=L_x \times L_y$}. The number of this type of graphs grows much less fast with the number of sites $N$. At the same time the subtraction and embedding procedure becomes trivial for rectangular graphs. The rectangular-graph expansion is by construction especially useful for the square lattice, since rectangular graphs respect the space-symmetries of this lattice. The order is still defined as the number of sites $N$, but now the limiting factor is not the number of graphs, but the ED. In contrast to the full-graph expansion, as shown in Ref.~\onlinecite{Kallin13}, it is possible to introduce the typical length scale $\mathcal{L}_{\rm rect}=\sqrt{N}$ for each order which allows scalings to the infinite-order limit. However, as we detail below, this scaling is still complicated, since the behavior of the rectangular-graph expansion is not smooth with the NLCE order.
\subsection{Arithmetic expansion}
Another scheme, which we call arithmetic expansion, uses again only rectangular graphs, but defines the order as the spatial dimension along the diagonal in a Manhattan-distance sense. The order of an $L_x \times L_y$ rectangular graph is then defined as $L_x+L_y-2$. The $-2$ is chosen, so that the smallest graph, which is a single two-site chain segment, has the order $1$. Therefore the arithmetic order counts the number of links along two of the edges.
For all scalings to the infinite-order limit we introduce the length scale $\mathcal{L}_{\rm arith}$ for the arithmetic expansion. Since the number of sites is not the same for each graph within each order of the arithmetic expansion, we use
\begin{align}
\mathcal{L}_{\rm arith} &= \frac{L_x+L_y}{2}
\end{align}
for all scalings, since this corresponds to the average linear extension of the largest quadratic graph in each NLCE order. As we will see below, only even NLCE orders are taken into account in the scalings and therefore $\mathcal{L}_{\rm arith}$ is an integer number.
\subsection{Square-graph expansion}
Within the rectangular-graph expansion the graphs with the largest spatial extension in one direction are very long chain graphs. This can become problematic for two-dimensional systems, since these one-dimensional graphs might direct the NLCE in the wrong direction. The arithmetic expansion corrects this to a certain extent. Here the maximal chain graph in a given order has the same length as the manhattan distance of the more quadratic graphs. The square-graph expansion goes one step further. Here the order is defined as the maximum number $L_x$ (or $L_y$) of sites into one direction of the rectangular graphs. In order to perform a valid graph expansion all subgraphs of the largest graph must also be taken into account. As a consequence, there are still chain graphs contained in the square-graph expansion, but their length will never be greater than the length of the largest quadratic graph.
Again, as in the arithmetic expansion, the number of sites is not the same for each graph in the expansion. But now it is clear that the quadratic graph with $L \times L$ sites has the maximum number of sites in a given order. Therefore we use $\mathcal{L}_{\rm sq}=L$ as the appropriate length scale for the square-graph expansion when performing scalings to the infinite-order limit.
\section{Edge-fields in NLCE}
\label{Sect:models}
One problem within NLCEs is, that the graphs themself are not ``aware'' of the ordered ground state on the infinite lattice. While the full symmetries of the lattice are restored through the embedding process, single graphs (or subsets of graphs) could show an entirely different behavior compared to the physics of the full lattice, as it has been shown in Ref.~\onlinecite{Ixert15} for the spin-$1$ kagome antiferromagnet. There one-dimensional subsets of graphs behave differently, since the one-dimensional subsystem undergoes a different quantum phase transition in a relevant parameter regime. This situation can be expected generically in quantum many-body systems, especially in the presence of geometric frustration where many different phases compete with each other. As a consequence, it can happen that any partial finite NLCE order displays an erratic behavior and the NLCE essentially breaks down.
To overcome this issue we want to incorporate the following line of reasoning. The NLCE is intended to expand non-perturbatively about the expected quantum phase of a given model which in our case is two-dimensional and long-range ordered. Each graph treated in the NLCE should then contribute the fluctuations specfic to this graph {\it being part} of the ordered phase for the physical quantities of interest, e.g.~the ground-state energy per site. This is only possible if in the calculation on graphs one includes the couplings to the expected ordered state {\it outside} the graph.
In practice, we assume that the Hamiltonian can be expressed in the following form
\begin{equation}
\label{Eq::setting}
\mathcal{H}=\mathcal{H}_0+\lambda\mathcal{V}
\end{equation}
so that $\mathcal{H}_0$ has a symmetry broken ground state \mbox{$\ket{0}\equiv\cdots\ket{0_i}\ket{0_j}\cdots$} which is a product state and is adiabatically connected to the ordered quantum phase expected to be present at finite values of $\lambda$ (often $\lambda=1$ is targeted). Let us mention that the parameter $\lambda$ might be already part of the original model under study or is introduced ``by hand'' corresponding to a deformation of the Hamiltonian to the desired form.
For the NLCE, the elementary site is then chosen according to the product-state structure of $\mathcal{H}_0$. The graphs $\mathcal{G}$ are considered to be surrounded by the state $\ket{0}$. All couplings between a site $i$ of $\mathcal{G}$ and a site $\nu$ outside $\mathcal{G}$ are called the ``edge-couplings'' $\mathcal{O}_\nu\otimes\mathcal{O}_i$ of $\mathcal{G}$. These edge-couplings are illustrated in Fig.~\ref{fig:EdgeFieldLattice} for a particular graph on the square lattice. All couplings within the graph are treated as usual during the ED. However, the operator $\mathcal{O}_\nu\otimes\mathcal{O}_i$ of an edge-coupling reduces to the edge-field
\begin{equation}
\bra{0_\nu}\mathcal{O}_\nu\ket{0_\nu}\mathcal{O}_i
\end{equation}
on site $i$ of $\mathcal{G}$. As a consequence, the correct symmetry-breaking is incorporated inside the NLCE scheme. The strength of the total edge-field influencing the results on a graph scales with the perimeter of the graph. As a result, the influence on one-dimensional chain graphs is maximal while it is minimal on the most two-dimensional graphs. This is exactly inline with the above reasoning.
Let us stress that due to the subtraction and embedding procedure within NLCE the impact of edge-fields becomes less and less for increasing graph sizes, since the size of the bulk of a graph scales much faster with the number of sites $N$ (again, one-dimensional chain graphs are special). In practice, however, the treated graphs are typically not in this limit. We therefore introduce the parameter $K\in\{0,\infty\}$ so that the edge-field
\begin{equation}
F_i = -K \mathcal{O}_i
\end{equation}
can be tuned in a flexible fashion. Physically, a value $K\neq \bra{0_\nu}\mathcal{O}_\nu\ket{0_\nu}$ corresponds to a different mean-field product state $\ket{\bar{0}}$. In the limit $K\rightarrow\infty$ the edge-field is so strong that no fluctuations take place on graphs and the system remains in the product state $\ket{0}$ of $\mathcal{H}_0$. In the opposite limit $K\rightarrow 0$, the standard NLCE without edge-fields is recovered with the above mentioned problems. In practice, one expects $K<\bra{0_\nu}\mathcal{O}_\nu\ket{0_\nu}$, since the true ground state of $\mathcal{H}$ contains quantum fluctuations giving smaller values compared to $\bra{0_\nu}\mathcal{O}_\nu\ket{0_\nu}$.
It is important to note, that the contribution of these edge-fields must be subtracted from the ground-state energy within the NLCE. Additionally, as the clusters are getting bigger the contribution of the edge-fields is getting smaller since they are only present at the boundary of the graphs. Therefore the edge-field contribution to the ground-state energy is sub-extensive.
In this work we consider two types of ordered phases which we treat with the edge-field NLCE. The first class are magnetically ordered phases. The quantum ground state can then be considered as a dressed version of the associated classical order, both having the same kind of order parameter corresponding to a finite sublattice magnetization. In this case one can always perform an appropriate sublattice rotation so that the classically ordered state is given by the perfect polarized state where all spins point in $z$-direction. After the sublattice rotation and assuming two-site interactions (a generalization is straightforward), the Hamiltonian can then be written as
\begin{equation}
\label{Eq::setting_spin_order}
\mathcal{H}=-J\sum_{i,j} S_i^{z} S_j^{z}+\lambda\mathcal{V}
\end{equation}
which is of the desired form Eq.~\eqref{Eq::setting} with \mbox{$\ket{0}\equiv\ket{\uparrow\cdots\uparrow}$.} The edge-field is therefore a local magnetic field operator on the sites $i$ of the edge of any graph
\begin{align}
F^{\rm mag}_{i} &= -K S_i^{z} \; . \label{eq:edgeFieldTerm}
\end{align}
The second class of quantum phases are non-magnetic valence bond solids (VBS) which break the translational symmetry of the system. In the most common form the system dimerizes, i.e.~pairs of spins form dimers which themselves order on the lattice. A generalization of this is a trimerization which is important for the spin-$1$ antiferromagnetic Heisenberg model on the kagome lattice.~\cite{Changlani15,Liu15,Li15,Ghosh15,Oitmaa15,Ixert15} Here three spins on a triangle build a low-energy singlet state and it is the trimer entity which is the elementary building block of the ordered state. In both cases (dimerization and trimerization) it is always possible to introduce a parameter $\lambda$ so that for $\lambda=0$ one gets a Hamiltonian $\mathcal{H}_0$ having a ground state of decoupled singlet dimers or trimers. Therefore, the elementary ``sites'' in the NLCE are chosen to be these dimers or trimers.
In contrast to the edge-field NLCE in the spin-ordered phases described above, the Hamiltonian $\mathcal{H}_0$ is completely local in terms of sites. As a consequence, there are no edge-couplings of the form $\mathcal{O}_\nu\otimes\mathcal{O}_i$ for $\lambda=0$. In this case it is mandatory to consider the limit of infinitely small $\lambda$ where edge-couplings appear naturally in second-order perturbation theory in $\lambda$ of the form
\begin{align}
\left(\ket{s_{\nu}}\bra{s_{\nu}}\right) \otimes \left(\ket{s_{i}}\bra{s_{i}} \right) \quad .
\end{align}
Here $\ket{s_{i}}\bra{s_{i}}$ is the singlet projector on dimer or trimer $i$ at the edge of a given graph. The edge-fields in the NLCE are then given by
\begin{align}
F^{\rm VBS}_{i} &=-K \ket{s_{i}}\bra{s_{i}} \quad .
\end{align}
\begin{figure}
\begin{center}
\includegraphics[width=0.65\columnwidth]{Fig1.pdf}
\end{center}
\caption{(Color online) Illustration of edge-couplings (blue) for one graph (black) on the square lattice (gray).}
\label{fig:EdgeFieldLattice}
\end{figure}
The implementation of edge-fields is now straightforward: we only take ``strong'' graphs~\cite{Oitmaa06} into account. ``Strong'' graphs are graphs where all possible couplings between the sites are set. This is necessary to make the edge-field configuration unique for each graph and not embedding-dependent. For each graph the outgoing couplings of site $i$ then lead to the above defined edge-fields. The next step involves the calculation of extensive ground-state quantities (e.g.~the energy or the sublattice magnetization) via ED. In the last step the considered quantities are reduced and embedded onto the infinite lattice. Remember, that it is important to subtract the edge-field contributions from the ground-state energy. For example, if the reference state is the polarized ferromagnet the subtraction is quite easy, since each edge-coupling leads to a subtraction of $-S \cdot K$, where $S$ is the spin of a single site. This can be seen as a subtraction of single-site contributions, because the edge-fields are local terms which act only on single sites.
The question that remains is, what is the ``optimal'' value of the edge-field strength $K$? In contrast to Ref.~\onlinecite{Kallin13}, in this work we apply the following two different schemes: i) the first is based on a self-consistent solution for the order parameter, e.g.~the sublattice magnetization for the spin-ordered phases. ii) the second targets $K$-values with a minimal sensitivity with respect to the ground-state energy.
i) Physically, it is reasonable to assume that a good choice for $K$ is of the order of the true order parameter of the quantum phase. And indeed, we find that one obtains very good results if $K$ is determined self-consistently as the value of the sublattice magnetization $m_0$ for the spin-ordered phases. We will call this scheme the fixpoint-method (abbreviated by ``fix''), since we are looking for the fixpoint $K=m_0(K)$.
To find this fixpoint we use the secant-method of root finding for each NLCE order, i.e.~for a fixed order we choose two starting values $K_1$ and $K_2$. For both $K$-values we calculate the sublattice magnetization of all graphs which contribute to this NLCE order. Then we subtract and embed them according to the linked-cluster theorem to obtain the values $m_{0}^{(i)}\coloneqq m_0(K_i)$ of the sublattice magnetization. In the final step the secant-method is used to construct an improved choice of $K$. In the \mbox{$(i+1)$th} iteration one has
\begin{align}
K_{i+1} &=K_i - \frac{K_i-K_{i-1}}{m_{0}^{(i)}-K_i-m_{0}^{(i-1)}+K_{i-1}} \cdot \left( m_{0}^{(i)}-K_i\right)
\end{align}
from which we calculate the corresponding $m_{0}^{(i+1)}$. We stop this iterative process if $\left|K_{i+1}-m_{0}^{(i+1)}\right| < 0.0001$.
ii) The second approach to determine a proper $K$ is to use the criterium of minimal sensitivity with respect to the ground-state energy per site. Physically, this is based on the observation that the effect of $K$ reduces for increasing graph sizes, since it is a subextensive quantity. So in principle the NLCE should converge for each value of $K$ in the limit of large graphs. As a consequence, it makes sense to choose the $K$ in each NLCE order so that the ground-state energy per site depends only minimally on $K$. We therefore check in each NLCE order for a local minimum in the function $e_0(K)$. At this minimum the variation with respect to $K$ is minimal. Note that there exist NLCE orders which do not display a local minimum. In this case case we discard this NLCE order for any further extrapolation to the infinite-order limit. This scheme is called the minimum-method abbreviated by ``min''.
\section{NLCE scaling}
\label{Sect:scaling}
The NLCE is performed directly in the thermodynamic limit, which is one important aspect compared to other numerical tools. Therefore there are no finite-size effects but each NLCE order corresponds to a different truncation of the real-space fluctuations in the infinite system. Nevertheless, one is interested in a proper scaling of the NLCE data towards the infinite-order limit. Such scalings of NLCEs are a challenging task and a priori no scaling laws are known for NLCEs to the best of our knowledge. In this section we list two different scaling behaviors which we test below: (i) The first is known from conventional finite-size scaling of gapless spin-ordered Heisenberg models \cite{Sandvik97}. (ii) The second stems heuristically from the coupled-cluster method (CCM) \cite{Bishop00} which is, similar to NLCEs, a tool which works directly in the thermodynamic limit.
(i) If we consider a long-range ordered antiferromagnetic ground state in two dimensions with gapless spin-wave excitations, then the finite-size scaling for the ground-state energy per site $e_0$ and the sublattice magnetization $m_0$ is given by \cite{Sandvik97}
\begin{align}
e_0 (\mathcal{L}) &=e_0 +\frac{a_1}{\mathcal{L}^3}+\frac{a_2}{\mathcal{L}^4}+\ldots\\
m_0^2 (\mathcal{L}) &=m_0^2 +\frac{b_1}{\mathcal{L}}+\frac{b_2}{\mathcal{L}^2}+\ldots\quad .
\end{align}
Here $\mathcal{L}$ is the linear length of the considered finite system, $a_1,a_2,b_1,b_2$ are fitting parameters, and $\ldots$ denotes higher-order terms in $1/\mathcal{L}$.
(ii) In CCM, the following heuristic scaling laws have been deduced \cite{Bishop00}
\begin{align}
e_0 (\mathcal{L}) &=e_0 +\frac{\bar{a}_1}{\mathcal{L}^4}+\frac{\bar{a}_2}{\mathcal{L}^8}+\ldots\\
m_0 (\mathcal{L}) &=m_0 +\frac{\bar{b}_1}{\mathcal{L}^2}+\frac{\bar{b}_2}{\mathcal{L}^4}+\ldots\quad ,
\end{align}
which allows an optimal scaling of CCM data obtained directly in the thermodynamic limit. Here $\bar{a}_1,\bar{a}_2,\bar{b}_1,\bar{b}_2$ are fitting parameters, and $\ldots$ denotes again higher powers in $1/\mathcal{L}$.
In the following we will test these scaling laws of type (i) and (ii) for the edge-field NLCE data squences by identifying the length scale $\mathcal{L}$ with the typical length scales $\mathcal{L}_{\rm rect}$, $\mathcal{L}_{\rm arith}$, and $\mathcal{L}_{\rm sq}$ of the different NLCEs.
\section{Results}
\label{Sect:results}
In the following we present the results obtained by our edge-field NLCE for various models and lattices. We start by benchmarking our approach for the N\'eel-ordered ground state of the spin-1/2 Heisenberg square lattice, which is well-studied and allows us to identify the optimal set ups concerning the type of cluster expansion and the choice of the edge-field $K$. Afterwards, we concentrate on the optimal set ups and discuss the results for the other considered systems.
\subsection{Spin-$1/2$ square lattice Heisenberg model}\label{Sect:results:squareOneHalf}
The spin-$1/2$ Heisenberg model on the square lattice is given by
\begin{align}
\mathcal{H} &= J \sum_{\langle i,j\rangle} {\mathbf S}_{i}\cdot{\mathbf S}_{j} \; , \label{eq:heisenbergModel}
\end{align}
where $J>0$ is chosen antiferromagnetic and the sum runs over all pairs of nearest neighbors.
The zero-temperature ground state is the long-range ordered N\'eel state \cite{Reger88,Chakravarty89} which represents a true challenge for any real-space approach due to the infinite correlation length and the gapless spin-wave excitations. On the other side the system is geometrically unfrustrated which is expected to help for the convergence of the NLCE.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig2.pdf}
\end{center}
\caption{(Color online) Ground-state energy per site $e_0/J$ for the spin-$1/2$ square lattice Heisenberg model as a function of the edge-field $K$ using the (a) full-graph, (b) rectangular-graph, (c) arithmetic, (d) square-graph NLCE. Different symbols correspond to different NLCE orders. The lines between symbols are guide to the eyes. The horizontal black line illustrates the QMC-value from Ref.~\onlinecite{Sandvik97}.}
\label{fig:NLCE_squareOneHalf}
\end{figure}
As in standard spin-wave calculations, we perform a sublattice rotation to obtain the following Hamiltonian
\begin{align}
\mathcal{H}_{\text{rot}} &= J \sum_{\langle i,j\rangle} \left(- S_i^z S_j^z - S_i^x S_j^x + S_i^y S_j^y \right) \; .
\end{align}
This form is well-suited for our edge-field NLCE. We introduce the parameter $\lambda$ which interpolates between the Ising ($\lambda=0$) and the Heisenberg ($\lambda=1$) limit. This yields the XXZ-Hamiltonian
\begin{align}
\label{eq::xxz}
\mathcal{H}_{\text{rot}}^{\rm XXZ} &= J \sum_{\langle i,j\rangle} \left[- S_i^z S_j^z+\lambda \left(S_i^y S_j^y - S_i^x S_j^x \right)\right] \; ,
\end{align}
which is exactly of the form Eq.~\eqref{Eq::setting}. The reference state is then chosen as one of the fully polarized Ising ground states along the $z$-direction and the definition of the edge-fields~\eqref{eq:edgeFieldTerm} are straightforward.
We focus first on the most challenging gapless case $\lambda=1$ for the NLCE. The simpler case of the gapped XXZ-model for $\lambda<1$ is discussed in the next subsection. We start the discussion with the ground-state energy per site $e_0$ and we use the value obtained by quantum Monte Carlo (QMC) simulations~\cite{Sandvik97} $e_0^{\text{QMC}}=-0.669437(5) J$ to gauge our edge-field approach. Note that the QMC value has been obtained by the scaling law of type (i) introduced in Sect.~\ref{Sect:scaling}.
In Fig.~\ref{fig:NLCE_squareOneHalf} (a) we show the ground-state energy per site for the full-graph expansion up to $N=8$ sites as a function of $K/J$. One clearly observes that the NLCE diverges for small $K/J$ (including zero), which already signals the importance to include the effects of the long-range order in the NLCE. In contrary, large values of $K/J$ stabilize the expansion, as expected. If one compares the results to the QMC value, it can be seen that the edge-field NLCE gets closer to the QMC-value with increasing order (which equals to the number of sites $N$ for the full-graph expansion). Still, the expansion is not yet well converged and for $K\lesssim0.6$ an odd-even effect is present. The curves of even order possess no optimal $K$-value, e.g.~no local minima exist. This is different for the odd orders where well-defined minima are present which can serve as the optimal $K$-values.
Next we turn to the results of the rectangular-graph expansion which are displayed in Fig.~\ref{fig:NLCE_squareOneHalf} (b). This expansion is much better converged, which, however, is mainly due to the larger number of sites included in the higher order clusters. Well-defined minima are present as a function of $K$ for odd and even orders alike. If one takes a closer look at these minima in Fig.~\ref{fig:NLCE_squareOneHalf} (b), the minima converge well to the QMC-value. A problematic feature of the rectangular-graph expansion is a proper scaling or extrapolation in the order, which has already been observed in other NLCE studies using rectangular graphs \cite{Kallin13,coester15}. Since $\mathcal{L}_{\rm rect}=\sqrt{N}$ is used to define the typical length scale in the rectangular-graph expansion, energy ``plateaus'' are visible as can be seen in the inset of Fig.~\ref{fig:NLCE_squareOneHalf_scaling}, i.e.~different orders have almost the same value for the energy and only at certain orders the value changes from one plateau to the one at lower energy. This behavior results due to the fact that certain orders, e.g.~12 and 13, only differ in the contribution from one chain graph (in the example the one with 13 sites) which typically have a very small reduced contribution to the two-dimensional energy per site. As a consequence, the extrapolation to the infinite-order limit is difficult for the rectangular-graph expansion. Nevertheless, a scaling of type (i) (see solid line in the inset of Fig.~\ref{fig:NLCE_squareOneHalf_scaling}) captures the correct magnitude of the energy in the thermodynamic limit.
The arithmetic expansion gets rid of these energy plateaus and leads to very good results as can be seen in Fig.~\ref{fig:NLCE_squareOneHalf} (c). An odd-even effect, as in the full-graph expansion, is again visible and minima as a function of $K$ exist only for the even orders in this expansion. Remarkably, the even orders are the orders in which a quadratic graph is added to the expansion. These graphs are expected to capture the physics of the two-dimensional model in an optimally fashion and are therefore considered to be most important. Consequently, even better results are obtained for the square-graph expansion as shown in Fig.~\ref{fig:NLCE_squareOneHalf} (d). Already the bare minimal values of the ground-state energy per site for order $4$ and order $5$ are very close to the QMC-value. The minimal bare value of order $5$ is $e_0^{\text{min}}=-0.66874 J$, which deviates from the QMC-value by only $0.0007 J$. If one ignores the order $2$ result (which corresponds to the quite small bond and single plaquette clusters) we again expect the presence of an odd-even effect, i.e.~odd and even orders are converging separately to the thermodynamic value. Unfortunately, the order $6$ calculation requires an open cluster with $N=36$ spin-1/2's, which is not possible without the considerable use of symmetries.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig3.pdf}
\end{center}
\caption{(Color online) Upper panel (a) (lower panel (b)) shows scalings of type (i) (of type (ii)) of the ground-state energy per site $e_0/J$ as a function of $1/\mathcal{L}$ for the spin-$1/2$ square lattice as obtained by the minimum and fixpoint method for the arithmetic (red symbols) and square-graph expansion (green symbols). Squares (cirlces) correspond to the values from the fixpoint (minimum) method. The dashed (solid) lines are scalings through data points obtained by the fixpoint (minimum) method. The black diamond depicts the QMC value from Ref.~\onlinecite{Sandvik97}. {\it Inset}: Ground-state energy per site $e_0/J$ as a function of $1/\mathcal{L}$ for the rectangular-graph expansion. Solid line correspond to a scaling of type (i).}
\label{fig:NLCE_squareOneHalf_scaling}
\end{figure}
\setlength{\tabcolsep}{10pt}
\begin{table*}[ht]
\begin{tabular}{llllll}
&&\textbf{Spin-$1/2$ square lattice}&&&\\[4pt]
Method & Extrapolation & $e_0/J$ & $|e_0-e_0^\text{QMC}|/J$ & $m_0$ & $m_0-m_0^\text{QMC}$ \\ \hline\\[-6pt]
arith. (fix.) & scaling (i) ($4$, $6$, $8$) & $-0.669567$ & $0.00013$ & $0.3077$ & $0.0007$ \\[4pt]
arith. (fix.) & scaling (ii) ($4$, $6$, $8$) & $-0.669486$ & $0.000049$ & $0.3220$ & $0.015$ \\[4pt]
arith. (fix.) & Wynn ($4$, $6$, $8$) & $-0.669081$ & $0.00036$ & $0.3319$ & $0.025$ \\[4pt]
square (fix.) & scaling (i) ($3$, $5$) & $-0.669692$ & $0.00025$ & $0.3038$ & $0.0032$ \\[4pt]
square (fix.) & scaling (ii) ($3$, $5$) & $-0.669134$ & $0.00030$ & $0.3336$ & $0.027$ \\[4pt]
square (fix.) & Wynn ($3$-$5$) & $-0.668482$ & $0.00096$ & $0.3476$ & $0.041$ \\[4pt]
arith. (min.) & scaling (i) ($4$, $6$, $8$) & $-0.669588$ & $0.00015$ & $0.3059$ & $0.0011$ \\[4pt]
arith. (min.) & scaling (ii) ($4$, $6$, $8$) & $-0.669301$ & $0.00014$ & $0.3316$ & $0.025$ \\[4pt]
arith. (min.) & Wynn ($4$, $6$, $8$) & $-0.669102$ & $0.00034$ & $0.3277$ & $0.021$ \\[4pt]
square (min.) & scaling (i) ($3$, $5$) & $-0.670022$ & $0.00058$ & $0.2877$ & $0.019$ \\[4pt]
square (min.) & scaling (ii) ($3$, $5$) & $-0.669433$ & $0.000004$ & $0.3252$ & $0.018$ \\[4pt]
square (min.) & Wynn ($3$-$5$) & $-0.668765$ & $0.00067$ & $0.3435$ & $0.037$ \\[4pt]\hline\hline\\
&&\textbf{Spin-$1$ square lattice}&&&\\[4pt]
Method & Extrapolation & $e_0/J$ & $|e_0-e_0^\text{CCM}|/J$ & $m_0$ & $m_0-m_0^\text{CCM}$ \\ \hline\\[-6pt]
square (fix.) & scaling (i) ($3$, $4$) & $-2.32779$ & $0.0020$ & $0.8112$ & $0.017$ \\[4pt]
square (fix.) & scaling (ii) ($3$, $4$) & $-2.32645$ & $0.0033$ & $0.8387$ & $0.045$ \\[4pt]
square (min.) & scaling (i) ($3$, $4$) & $-2.33133$ & $0.0016$ & $0.7630$ & $0.031$ \\[4pt]
square (min.) & scaling (ii) ($3$, $4$) & $-2.32992$ & $0.00017$ & $0.8057$ & $0.012$ \\[4pt]\hline\hline
\end{tabular}
\caption{Comparison of ground-state energies per site $e_0/J$ and sublattice magnetizations $m_0$ on the spin-$1/2$ and spin-$1$ square lattice. The determination method for $K$ is denoted in brackets after the used graph-expansion method (min. for the minimum method and fix. for the fixpoint method). The used orders for the extrapolation are stated after the extrapolation method.}
\label{tab:squareOneHalfEnergy}
\end{table*}
Next we take the $K$-values of the well-defined minima as well as the fixpoint $K$-values for each NLCE order to perform scalings of type (i) and (ii) as well as to apply the Wynn-algorithm (see e.g. Ref.~\onlinecite{Tang13} and~\onlinecite{Guttmann89}) in order to obtain even better estimates of the ground-state energy per site in the infinite-order limit. Note that another option to extrapolate NLCE data sequences has been recently formulated in Ref.~\onlinecite{Coester2016}. However, this scheme is most powerful to extract critical properties which we do not consider in this work. The obtained values for the arithmetic and square-graph NLCE for the different extrapolations are listed in Tab.~\ref{tab:squareOneHalfEnergy} and are illustrated in Fig.~\ref{fig:NLCE_squareOneHalf_scaling}. In accordance with the explanations given above, the quality of the full- and the rectangular-graph expansion is not as high as the other two NLCEs. In the full-graph expansion the maximal order is clearly not competitive and it would be also interesting to push this expansion to higher orders. In contrast, in the rectangular-graph expansion the plateau-effect complicates a proper extrapolation in the order. We therefore do not show these results. We want to stress that not every order has a well-defined fixpoint, since sometimes such points are absent as can be seen in Fig.~\ref{fig:NLCE_squareOneHalf_mag}. Note that the orders which have a well-defined minimum of the magnetization also have a fixpoint.
The scalings of type (i) and (ii) are explicitly shown in Fig.~\ref{fig:NLCE_squareOneHalf_scaling}. Here we scaled through the ``best'' NLCE orders. ``Best'' implies that the involved orders should be as large as possible by respecting odd-even effects. Here we took the three largest even orders $4$, $6$, and $8$ in the arithemtic expansion while we took orders $3$ and $5$ for the square-graph expansion. The used orders are also given in Tab.~\ref{tab:squareOneHalfEnergy}.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig4.pdf}
\end{center}
\caption{(Color online) Sublattice magnetization $m_0$ for the spin-$1/2$ square lattice Heisenberg model as a function of the edge-field $K$ using the (a) full-graph, (b) rectangular-graph, (c) arithmetic, (d) square-graph NLCE. Different symbols correspond to different NLCE orders. The lines between symbols are guide to the eyes. The horizontal black line illustrates the QMC-value from Ref.~\onlinecite{Sandvik97}.}
\label{fig:NLCE_squareOneHalf_mag}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig5.pdf}
\end{center}
\caption{(Color online) Upper panel (a) (lower panel (b)) shows scalings of type (i) (of type (ii)) of the sublattice magnetization $m_0$ as a function of $1/\mathcal{L}$ for the spin-$1/2$ square lattice as obtained by the minimum and fixpoint method for the arithmetic (red symbols) and square-graph expansion (green symbols). Squares (circles) correspond to the values from the fixpoint (minimum) method. The dashed (solid) lines are scalings through data points obtained by the fixpoint (minimum) method. The black diamond depicts the QMC value from Ref.~\onlinecite{Sandvik97}.}
\label{fig:NLCE_squareOneHalf_mag_scaling}
\end{figure}
Globally, all schemes give values close to the QMC value improving the best bare NLCE values by roughly one order of magnitude. We observe that the fixpoint scheme as well as the local $K$-minima approach give similarly good results for both NLCEs. At the same time both type of scalings perform slightly better than the values from Wynn extrapolation. However, the quality of both scaling types is almost the same so that a comparison is complicated.
Next we discuss the behavior of the ground-state sublattice magnetization $m_0$ which corresponds to the order parameter of the N\'eel state. We stress that this is only possible if one performs the NLCE inside the symmetry-broken phase which is realized in our approach due to the presence of the edge-fields. As a reference, we again take the QMC value from Ref.~\onlinecite{Sandvik97} which is $m_0^{\text{QMC}}=0.3070(3)$. The obtained edge-field NLCE results for $m_0$ as a function of $K$ are displayed in Fig.~\ref{fig:NLCE_squareOneHalf_mag} for the four different cluster expansions.
All general features discussed for the ground-state energy per site are also present for the sublattice magnetization. Most importantly, well-defined minima exist at roughly the same $K$-values as for the ground-state energy per site. The convergence of the bare NLCE values to the QMC-value is not as good as for the ground-state energy, but the general trend is the same. We observe that the bare values at the local minima and the fixpoints are typically larger than the QMC-value. This originates from the fact that the fully classical reference state likely yields a too large edge-field for the considered clusters.
Along the same lines as for the ground-state energy per site, we performed scalings of type (i) and (ii) as shown in Fig.~\ref{fig:NLCE_squareOneHalf_mag_scaling}. Note that for the scaling we took the $K$-values where energy is minimal and not the local minima of the magnetization in order to get a consistent scheme. As before, for the fixpoint-method the scaled values are of course taken from the sublattice magnetization. The obtained scaled values for $m_0$ are also listed in Tab.~\ref{tab:squareOneHalfEnergy} together with their deviation from the QMC-value. These differences are larger compared to the ones for the ground-state energy per site, which is however expected, since the sublattice magnetization is considerably more sensitive. Typically, we obtain a satisfactory agreement with QMC having a difference of the order $0.01 J$. We find that the arithmetic expansions gives better results compared to the square-graph expansion. However, this is likely due to the fact that only two data points are included in the square-graph expansion which enhances the uncertainty of the scaling procedure. If one compares the two types of scaling, then it is apparent that type (i) yields more convincing results than type (ii), which typically overshoots the QMC value.
Altogether, the edge-field NLCE gives quantitative results for the ground-state energy per site and the sublattice magnetization for the ordered N\'eel state of the square lattice spin-1/2 Heisenberg model having an infinite correlation length. We have seen that the fixpoint and the local $K$-minima approaches give both comparable and convincing results. In the following we will focus on the scaling behavior in all other applications of the edge-field NLCE below. We therefore do not apply the Wynn algorithm anymore, but restrict the discussion to the two types of scaling laws.
\subsection{Spin-$1/2$ XXZ-model on the square lattice}
In this subsection we discuss the edge-field NLCE for $\lambda\in [0,1 ]$ in Eq.~\eqref{eq::xxz} corresponding to the XXZ model on the square lattice. This model interpolates between the Ising limit $\lambda=0$ and the just discussed Heisenberg model for $\lambda=1$. We stress that for $\lambda<1$ the system is gapped and the correlation length is finite. As a consequence, one expects the edge-field approach to converge faster with the order compared to the gapless Heisenberg model and to agree with perturbative SE for small $\lambda$. This is exactly what we find.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig6.pdf}
\end{center}
\caption{(Color online) Ground-state energy per site $e_0/J$ for the spin-$1/2$ XXZ-model on the square lattice for $\lambda=0.85$ using the square-graph NLCE. Different colors correspond to different orders of graph expansion. The horizontal black line illustrates the bare value obtained by SE~\cite{Singh89, Weihong91, Dusuel10} up to order $14$.
\textit{Left inset:} Scaling of type (i) (red solid line) and type (ii) (red dashed line) of the $e_0$-values obtained by the minimum- (cyan circles) or fixpoint-method (red circles) as a function of $1/\mathcal{L}$. \textit{Right inset:} Scalings of type (i) (solid lines) and type (ii) (dashed lines) of the corresponding $m_0$-values as a function of $1/\mathcal{L}$. Cyan (red) symbols/lines correspond to the values from the minimum-method (fixpoint-method), whereas Diamonds refer to the bare SE result.}
\label{fig:NLCE_squareOneHalf_XXZ_lambda0.85}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig7.pdf}
\end{center}
\caption{(Color online) (a) Ground-state energy per site $e_0/J$ and (b) sublattice magnetization $m_0$ for the XXZ-model on the square lattice obtained by scalings of type (i) (squares) and type (ii) (triangles) of the square-graph expansion. Cyan symbols result from the minimum-method and red symbols from the fixpoint-method. Circles correspond to the bare order $5$ NLCE, the solid line is the bare high-order SE in $\lambda$ up to order $14$ from Refs.~\onlinecite{Singh89, Weihong91, Dusuel10}, and the diamond refers to the QMC-value~\cite{Sandvik97} for the Heisenberg point $\lambda=1$.}
\label{fig:NLCE_squareOneHalf_XXZ}
\end{figure}
In this part we focus on the square-graph expansion, since this expansion performed best for the Heisenberg case. A representative plot of the ground-state energy per site and the sublattice magnetization for the specific value $\lambda=0.85$ is shown in Fig.~\ref{fig:NLCE_squareOneHalf_XXZ_lambda0.85}. One clearly observes that already the bare NLCE values are well converged. Again, as for the Heisenberg model, scalings of type (i) and type (ii) are performed through the orders $3$ and $5$ minimal values of the ground-state energy per site which further improves the quality of the NLCE data (see insets in Fig.~\ref{fig:NLCE_squareOneHalf_XXZ_lambda0.85}).
Scaled NLCE results for $e_0$ and $m_0$ on the full $\lambda$-axis are presented in Fig.~\ref{fig:NLCE_squareOneHalf_XXZ}. These results are compared to high-order SE in $\lambda$ from Refs.~\onlinecite{Singh89, Weihong91, Dusuel10}. Note that the highest NLCE order 5 of the square-graph expansion contains the order eight perturbation theory in $\lambda$ exactly (this is actually true for all four different NLCEs). It is therefore no surprise but in fact mandatory that SE and NLCE are in quantitative agreement for small $\lambda$. Genercally, for larger values of $\lambda$ this must not be the case, since the bare SE need not converge while the NLCE could still be convergent. However, for the unfrustrated XXZ model on the square lattice, all SE are monotonous and therefore even the bare perturbative series yields satisfactory results. This will be different for the same model on the triangular lattice, where one finds alternating series due to the geometric frustration.
Overall, the obtained NLCE results using edge-fields give convergent and satisfactory results on the full $\lambda$-axis for the unfrustrated XXZ model on the square lattice. Next we investigate the same model for larger spins one.
\subsection{Spin-$1$ square lattice Heisenberg model}
The edge-field NLCE of the spin-1/2 XXZ-model on the square lattice is a systematic and non-perturbative expansion about the long-range ordered classical N\'eel state. All quantum fluctuations contained on the clusters in a given NLCE order are taken fully into account. As discussed above, the most challenging case for the NLCE is the gapless Heisenberg model for $\lambda=1$ due to the diverging correlation length. In this subsection we focus again on the Heisenberg model but enlarge the spin value to one. The system is therefore still long-range ordered and gapless. The N\'eel-ordered reference state is unchanged and one expects that the edge-field NLCE converges faster, since larger spins display smaller quantum fluctuations and the edge-fields themselves originate from a mean-field decoupling assuming the classical N\'eel-ordered reference state outside the clusters under investigation. At the same time the ED on clusters is harder, since the local Hilbert space of spins one is larger. As a consequence, we do not reach the same cluster sizes as for spins 1/2.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig8.pdf}
\end{center}
\caption{(Color online) (a) Ground-state energy per site $e_0/J$ and (b) sublattice magnetization $m_0$ for the spin-$1$ Heisenberg model on the square lattice obtained with edge-field NLCE using the square-graph expansion. Different colored symbols correspond to different orders of the graph expansion. The lines between them are guide to the eyes. The insets show scalings of type (i) (solid lines) and type (ii) (dashed lines) through the NLCE orders $3$ and $4$ (cyan circles for the minimum- and red circles for the fixpoint-method) of the ground-state energy per site or the sublattice magnetization as a function of $1/\mathcal{L}$. The black diamond depicts the CCM value from Ref.~\onlinecite{Li11}.}
\label{fig:NLCE_squareOne_squareExpansion}
\end{figure}
In the following we focus on the square NLCE, which was the best NLCE in the spin-$1/2$ case. Due to the larger spin-one Hilbert space, we reach order four in the square NLCE, i.e.~the maximal cluster contains 16 sites. The corresponding NLCE results for the ground-state energy per site $e_0/J$ and the sublattice magnetization $m_0$ are displayed in Fig.~\ref{fig:NLCE_squareOne_squareExpansion}. A scaling of the bare NLCE data has also been performed as for the spin-1/2 case and yields again convincing results which are listed in the lower panel of Tab.~\ref{tab:squareOneHalfEnergy}. As a reference, we compare our NLCE results to the values $e_0^{\text{CCM}}=-2.32975 J$ and $m_0^{\text{CCM}}=0.7938$ obtained with the CCM in Ref.~\onlinecite{Li11}. The reduced quantum fluctuations in the spin-one case can be readily seen from the rather large value of the sublattice magnetization.
One observes that already the bare NLCE orders converge very well. The only exception, as for the spin-1/2 case, is the lowest order two which does not fit in the general trend of the NLCE. As a consequence, we discard this order from any scaling of the NLCE data. Already the bare NLCE order four yields the value $e_0^{\text{min}}=-2.32744 J$, which differs only by $0.00231 J$ from the CCM-value~\cite{Li11}. This means that the order four NLCE with up to $16$ sites for spin one is as good as the order five NLCE for spins-1/2 with up to $25$ sites. For the spin-$1$ case, the quality of the NLCE is only slightly improved due to scaling (see Tab.~\ref{tab:squareOneHalfEnergy}), which is mainly due to the lower order of the NLCE.
As for the spin-1/2 case, the quality of the sublattice magnetization is slightly lower as can be seen from Fig.~\ref{fig:NLCE_squareOne_squareExpansion}~(b) as well as from Tab.~\ref{tab:squareOneHalfEnergy} where we list the scaled values of the different NLCEs. The scaling of the sublattice magnetization is again performed as described for the spin-$1/2$ case (see inset of Fig.~\ref{fig:NLCE_squareOne_squareExpansion}~(b)). Again, as for the ground-state energy per site, the bare NLCE values converge smoothly to the CCM-value. We observe that the scaling of type (i) results in a too low value compared to the CCM-value \cite{Li11} while the value deduced from scaling of type (ii) is too large. These discrepancies originate from the uncertainty of the scaling, since we only have the two points from order three and four, and we expect also an even-odd effect in the NLCE as for the spin-1/2 NLCE data.
Altogether, the edge-field NLCE performs also well for the spin-$1$ Heisenberg model despite the fact that the maximal NLCE order is reduced compared to the spin-1/2 case due to the larger Hilbert space.
\subsection{Spin-$1/2$ triangular lattice Heisenberg model}
The Heisenberg model on the square lattice is unfrustrated for any value of the spin. This is expected to help the NLCE to converge with increasing NLCE order, since the bare reference state of $\mathcal{H}_0$ is unfrustrated on each cluster and therefore no competing ground states are close in energy for the vast majority of clusters. In contrast, on a geometrically frustrated system like the antiferromagnetic Heisenberg model on the triangular lattice the situation is more complex and it is a priori not clear how the edge-field NLCE performs.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig9}
\end{center}
\caption{(Color online) Ground-state energy per site $e_0/J$ for the (a) arithmetic expansion and (b) square-graph expansion for the spin-$1/2$ triangular lattice. Different colored symbols correspond to differend orders of the graph expansion. The lines between them are guide to the eyes. Insets show scalings of type (i) (solid lines) and type (ii) (dashed lines) as a function of $1/\mathcal{L}$ through the minimal (cyan circles) and fixpoint (red circles) NLCE values. The black diamond depicts the GFQMC value~\cite{Capriotti99} and the star corresponds to the SE value~\cite{Zheng06}.}
\label{fig:NLCE_triangularOneHalf_arithmeticSquare}
\end{figure}
As for the Heisenberg model on the square lattice, the ground state on the triangular lattice breaks the continuous SU(2) symmetry and the system possesses gapless Goldstone modes. The specific order is given by the so-called three-sublattice $120^\circ$ order for any value of the spin. It is therefore again possible to view the quantum ground state of the triangular Heisenberg model as a dressed version of the classical $120^\circ$ order of the corresponding Ising model on the same lattice. As a consequence, the edge-field NLCE for the $120^\circ$ quantum order can be applied along the same lines as for the N\'eel order on the square lattice.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig10.pdf}
\end{center}
\caption{(Color online) Sublattice magnetization $m_0$ for the (a) arithmetic expansion and (b) square-graph expansion for the spin-$1/2$ triangular lattice. Different colored symbols correspond to differend orders of the graph expansion. The lines between them are guide to the eyes. Insets show scalings of type (i) (solid lines) and type (ii) (dashed lines) as a function of $1/\mathcal{L}$ through the minimal (cyan circles) and fixpoint (red circles) NLCE values. For the arithmetic expansion we additionally display a linear fit as a dashed-dotted line in the upper inset. The black diamond depicts the GFQMC value~\cite{Capriotti99} and the star corresponds to the SE value~\cite{Zheng06}.}
\label{fig:NLCE_triangularOneHalf_arithmeticSquare_mag}
\end{figure}
If one performs the appropriate sublattice rotation for the $120^\circ$ ordered state to the Heisenberg model on the triangular lattice, one obtains the following rotated Hamiltonian
\begin{align}
\mathcal{H}_{\text{rot}} = J \sum_{\langle i,j\rangle} S_i^y S_j^y + &\cos\left(\theta_i-\theta_j\right)\left(S_i^z S_j^z+S_i^x S_j^x\right) \nonumber \\
+&\sin\left(\theta_i-\theta_j\right)\left(S_i^z S_j^x-S_i^x S_j^z\right) \; ,
\end{align}
where $\theta_i$ is $\theta_\text{A}=0$, $\theta_\text{B}=2\pi/3$, $\theta_\text{C}=4\pi/3$ depending on wether the site $i$ belongs to sublattice A, B or C.
Again, we introduce the parameter $\lambda$ in front of all terms which are not diagonal with respect to the rotated ferromagnetic reference state. The final Hamiltonian is then given by
\begin{align}
\label{Eq:XXZ_triangular}
\mathcal{H}_{\text{rot}}^{\rm XXZ} = J \sum_{\langle i,j\rangle} - \frac12 S_i^z S_j^z &+ \lambda \left(S_i^y S_j^y - \frac12 S_i^x S_j^x \right. \nonumber \\
&+\left. \frac{\sqrt{3}}{2} S_i^z S_j^x -\frac{\sqrt{3}}{2} S_i^x S_j^z \right)\; ,
\end{align}
which is of the desired form Eq.~\eqref{Eq::setting}. The order of $i$ and $j$ in the above equation is chosen such that the term $S_i^z S_j^x$ has a positive prefactor. As a reference, we compare our results to the values from Green's function quantum Monte Carlo (GFQMC)~\cite{Capriotti99} and SE~\cite{Zheng06}.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig11.pdf}
\end{center}
\caption{(Color online) (a) Ground-state energy per site $e_0/J$ and (b) sublattice magnetization $m_0$ for the XXZ-model on the triangular lattice obtained by scalings of type (i) (squares) and type (ii) (triangles) using the square-graph expansion and the minimum- (cyan symbols) as well as the fixpoint-method (red symbols). Blue circles are the bare results of order 5 and the solid line is the bare order-$13$ SE result~\cite{Zheng06}. At the Heisenberg point $\lambda=1$, the star depicts the extrapolated SE~\cite{Zheng06} whereas the diamond corresponds to the GFQMC-value~\cite{Capriotti99}.}
\label{fig:NLCE_triangularOneHalf_XXZ}
\end{figure}
We start with the discussion of the ground-state energy per site for the Heisenberg case $\lambda=1$. The values from GFQMC and extrapolated SE are $e_0^{\rm GFQMC} = −0.5458(1) J$ and $e_0^{\text{SE}}=-0.5502(4) J$. Our edge-field NLCE data using the arithmetic and the square-graph expansion are displayed in Fig.~\ref{fig:NLCE_triangularOneHalf_arithmeticSquare}. Similarly to the spin-$1/2$ square lattice Heisenberg model, the edge-field is again crucial to get a meaningful NLCE for both expansions. In the square-graph expansion, all NLCE orders display a well-defined minimum as a function of $K$. Notably, these minima converge very well and, in contrast to the square lattice case, monotonously with the NLCE order. Already the bare NLCE results are rather good. The minimal value of the order-$5$ square-graph expansion is \mbox{$e_0^{\text{min}}=-0.548258 J$} which differs only by about $0.002 J$ from $e_0^{\text{SE}}$ and $e_0^{\rm GFQMC}$. This is different for the arithmetic expansion where the bare value of the highest order is at slightly higher values and the convergence with increasing order is less compared to the square-graph expansion. Interestingly, the bare order-$5$ value from the square-graph expansion is already below the GFQMC value which suggests that the extrapolated value from SE is likely to be trusted more than the one from GFQMC, since the values from the edge-field NLCE monotonously decrease with the NLCE order for the square-graph expansion. These conclusions are further strengthened when performing scalings of type (i) and (ii) as shown in the insets of Fig.~\ref{fig:NLCE_triangularOneHalf_arithmeticSquare} and Tab.~\ref{tab:triangularEnergy}. For the arithmetic expansion, both scalings yield different values. While type (i) is very close to the value from SE, the scaling of type (ii) gives a considerably too low value of the energy. In contrast, for the square-graph expansion, both scalings work very well and yield a value within the error estimation of the SE extrapolation~\cite{Zheng06}.
Next we turn to the sublattice magnetization $m_0$ of the $120^\circ$ long-range ordered ground state of the triangular lattice Heisenberg model. The ordered moment obtained from SE and GFQMC is $m_0^{\text{SE}}=0.19$ and \mbox{$m_0^{\text{GFQMC}}=0.205(10)$} which is lower than in the square lattice Heisenberg model due to the geometric frustration. As for the ground-state energy per site, the edge-fields are essential to regularize the NLCE. The arithmetic and the square-graph NLCE display well defined minima as a function of $K/J$ (see Figs.~\ref{fig:NLCE_triangularOneHalf_arithmeticSquare} and~\ref{fig:NLCE_triangularOneHalf_arithmeticSquare_mag}). These minima decrease monotonously towards the values from SE and GFQMC, but, as for the square lattice Heisenberg model, the bare edge-field NLCE sublattice magnetization is not as close to SE and GFQMC as the ground-state energy per site, e.g.~the bare order-$5$ NLCE minimum is $m_0^{\text{min}}\approx 0.27$ for the square-graph expansion. Again, we have performed scalings of type (i) and (ii) shown in the insets of Fig.~\ref{fig:NLCE_triangularOneHalf_arithmeticSquare_mag}. It can be clearly observed that the quality of both scalings is rather poor for the arithmetic expansion. We therefore show also a linear fit in $1/\mathcal{L}$ which gives a reasonable value of the sublattice magnetization in the thermodynamic limit. In contrast, the scaling for the square-graph expansion works very well yielding values close to the ones from SE and GFQMC. The only exception is the scaling of type (i) for the fixpoint method which reflects the sensitivity due to the limited number of data points.
In our opinion the different quality of the arithmetic and square-graph expansion can be understood as follows. In the arithmetic expansion a given NLCE order contains several graphs with very similar typical length scales $\mathcal{L}_{\rm arith}$, but with different aspect ratios $L_x/L_y$. It is therefore reasonable that a scaling in $1/\mathcal{L}_{\rm arith}$ is complicated. This is different for the square-graph expansion. Here each NLCE order is clearly dominated by the defining length scale $\mathcal{L}_{\rm sq}$ of the largest quadratic cluster and one therefore expects a better scaling behavior in this length scale. The same reasoning holds also for the square lattice Heisenberg model, but here these properties are likely not observed due to the annoying even-odd effect which is absent on the triangular lattice.
\subsection{Spin-$1/2$ XXZ-model on the triangular lattice}
Next we discuss the full $\lambda$-axis between 0 and 1 in the XXZ-model on the triangular lattice as given by Eq.~\eqref{Eq:XXZ_triangular} after the appropriate sublattice rotation corresponding to the 120$^\circ$ order. As for the same model on the square lattice, one expects that the edge-field NLCE converges better when decreasing $\lambda$ from 1 to 0, since a gap opens for $\lambda<1$ introducing a finite correlation length. This is indeed the case as illustrated in Fig.~\ref{fig:NLCE_triangularOneHalf_XXZ} showing the results for the square-graph expansion.
For small values of $\lambda$ the edge-field NLCE is indistinguishable to the perturbative high-order SE. In contrast to the unfrustrated XXZ-model on the square lattice, the series is alternating in $\lambda$ and it becomes therefore problematic for $\lambda>0.6$. The SE can be improved by extrapolating the alternating series with Pad\'e extrapolants. The bare order-5 edge-field NLCE behaves smoothly on the full $\lambda$-axis and captures already well the global behavior. Scalings of the bare data become important for $\lambda>0.6$ which corresponds to the same large $\lambda$-regime where the bare SE is unreliable. Altogether, the edge-field NLCE with square graphs captures the physics of the XXZ-model well. It is only the scaling of the sublattice magnetization which becomes challenging close to $\lambda=1$.
We want to note, that we restricted $b_1$ and $\bar{b}_1$ in the scalings to positive values for $\lambda<1$.
\subsection{Spin-$1$ triangular lattice Heisenberg model}\label{Sect:results:triangularOne}
In this subsection we discuss our edge-field NLCE results using square graphs for the spin-$1$ triangular lattice Heisenberg model shown in Fig.~\ref{fig:NLCE_triangularOne_squareExpansion} and Tab.~\ref{tab:triangularEnergy}. As a reference, we compare to the CCM data from Ref.~\onlinecite{Li11} which are $e_0^{\text{CCM}}=-1.83968 J$ for the ground-state energy per site and $m_0^\text{CCM}=0.7086$ for the sublattice magnetization.
The anticipated trend from the spin-$1/2$ Heisenberg model on the triangular lattice is clearly visible. The edge-field NLCE displays well-defined minima which approach the values from the CCM monotoneously with increasing NLCE order. The only exception is the order-2 curve for the sublattice magnetization which we attribute to the low order as in the square lattice case. Unfortunately, as for the spin-$1$ square lattice Heisenberg model, the maximal NLCE order is reduced to four compared to the spin-$1/2$ case due to the larger Hilbert space. We also observe that the bare order-4 NLCE is not as close to the CCM as for the unfrustrated square lattice case. The reason might be that the square-graph NLCE does not converge so well due to the geometric frustration, which, however, is also present in the spin-$1/2$ case where quantum fluctuations are even stronger and the geometric frustration is also present.
The scaled values, along with their deviation from the value obtained by series expansion, are shown in Tab.~\ref{tab:triangularEnergy}.
We remark that, as for the spin-$1/2$ case on the triangular lattice, the scaling of type (i) performs better for the sublattice magnetization.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig12.pdf}
\end{center}
\caption{(Color online) (a) Ground-state energy per site $e_0/J$ and (b) sublattice magnetization $m_0$ for the spin-$1$ triangular lattice using the square-graph expansion. Different colored symbols correspond to differend orders of the graph expansion. The lines between them are guide to the eyes. Insets show scalings of type (i) (solid lines) and type (ii) (dashed lines) as a function of $1/\mathcal{L}$ through the minimal (cyan circles) and fixpoint (red circles) NLCE values. The black diamond depicts the CCM value from Ref.~\onlinecite{Li11}.}
\label{fig:NLCE_triangularOne_squareExpansion}
\end{figure}
\setlength{\tabcolsep}{10pt}
\begin{table*}[ht]
\begin{tabular}{llllll}
&&\textbf{Spin-$1/2$ triangular lattice}&&&\\[4pt]
Method & Extrapolation & $e_0/J$ & $|e_0-e_0^\text{SE}|/J$ & $m_0$ & $m_0-m_0^\text{SE}$ \\ \hline\\[-6pt]
arith. (fix.) & scaling (i) ($4$, $6$, $8$) & $-0.55301$ & $0.0028$ & -- & -- \\[4pt]
arith. (fix.) & scaling (ii) ($4$, $6$, $8$) & $-0.54778$ & $0.0024$ & $0.0438$ & $0.15$ \\[4pt]
square (fix.) & scaling (i) ($3$-$5$) & $-0.55178$ & $0.0016$ & $0.0898$ & $0.10$ \\[4pt]
square (fix.) & scaling (ii) ($3$-$5$) & $-0.55110$ & $0.00090$ & $0.2104$ & $0.020$ \\[4pt]
arith. (min.) & scaling (i) ($4$, $6$, $8$) & $-0.55451$ & $0.0043$ & -- & -- \\[4pt]
arith. (min.) & scaling (ii) ($4$, $6$, $8$) & $-0.54877$ & $0.0014$ & $0.0646$ & $0.13$ \\[4pt]
square (min.) & scaling (i) ($3$-$5$) & $-0.55015$ & $0.00005$ & $0.1949$ & $0.0049$ \\[4pt]
square (min.) & scaling (ii) ($3$-$5$) & $-0.55049$ & $0.00029$ & $0.2196$ & $0.030$ \\[4pt]\hline\hline\\
&&\textbf{Spin-$1$ triangular lattice}&&&\\[4pt]
Method & Extrapolation & $e_0/J$ & $|e_0-e_0^\text{CCM}|/J$ & $m_0$ & $m_0-m_0^\text{CCM}$ \\ \hline\\[-6pt]
square (fix.) & scaling (i) ($3$, $4$) & $-1.84689$ & $0.0072$ & $0.6961$ & $0.012$ \\[4pt]
square (fix.) & scaling (ii) ($3$, $4$) & $-1.83704$ & $0.0026$ & $0.7862$ & $0.078$ \\[4pt]
square (min.) & scaling (i) ($3$, $4$) & $-1.84180$ & $0.0021$ & $0.7175$ & $0.0089$ \\[4pt]
square (min.) & scaling (ii) ($3$, $4$) & $-1.83502$ & $0.0047$ & $0.7790$ & $0.070$ \\[4pt]\hline\hline
\end{tabular}
\caption{Comparison of ground-state energies per site $e_0/J$ and sublattice magnetizations $m_0$ on the spin-$1/2$ and spin-$1$ triangular lattice. The determination method for $K$ is denoted in brackets after the used graph-expansion method (min. for the minimum method and fix. for the fixpoint method). The used orders for the extrapolation are stated after the extrapolation method.}
\label{tab:triangularEnergy}
\end{table*}
\subsection{Spin-$1$ kagome Heisenberg model}
\label{Sect:spin1kagome}
In contrast to all other cases discussed so far, which were long-range spin-ordered states with a finite sublattice magnetization, here we apply the edge-field NLCE to the trimerized ground state of the spin-$1$ Heisenberg model on the kagome lattice \cite{Changlani15,Liu15,Li15,Ghosh15,Oitmaa15,Ixert15}. In the trimerized ground state the symmetry between up and down triangles on the kagome lattice is sponteaously broken. It is therefore possible to adiabatically connect the trimerized ground state to the limit of isolated up (or down) triangles where the ground state is given by the exact product state of singlets on the up (or down) triangles. The elementary excitations above the trimerized ground state of the kagome Heisenberg model have a finite gap \cite{Changlani15} and it is therefore expected that the NLCE converges better compared to the gapless Heisenberg points discussed above on the square and triangular lattice.
We introduce the real parameter $\lambda\in\left[0,1\right]$ which allows to interpolate between the limit of isolated up-triangles $\lambda=0$ and the isotropic Heisenberg model at $\lambda=1$. The Hamiltonian is then defined as
\begin{equation}
\mathcal{H} = J \sum_{\langle i,j\rangle\in\Delta } {\mathbf S}_{i}\cdot{\mathbf S}_{j} + \lambda J \sum_{\langle i,j\rangle\in\nabla } {\mathbf S}_{i}\cdot{\mathbf S}_{j}\; ,
\label{eq:ham}
\end{equation}
so that the first (second) sum runs over all nearest neighbor sites on up-triangles (down-triangles). In the following we focus on an antiferromagnetic exchange constant $J>0$.
We have performed the edge-field NLCE up to order six using a full-graph expansion in terms of elementary up-triangles. The NLCE order of a graph is defined as the number of triangles $N_{\rm tr}$ of this graph. As we have shown in Ref.~\onlinecite{Ixert15}, the NLCE without edge-fields shows an almost erratic behavior for increasing NLCE order. This behavior can be traced back to a different quantum critical behavior of the one-dimensional unfrustrated chain graphs (in terms of triangles) occuring at values $\lambda\approx 0.8$ well before the isotropic point $\lambda=1$. As a consequence, the reduced contributions of these chain graphs do decay only algebraically with the NLCE order for $\lambda\approx 0.8$ while the embedding factor grows exponentially with the NLCE order leading to a partially diverging NLCE data sequence. Let us note that in Ref.~\onlinecite{Ixert15} we also formulated a reorganized expansion in terms of highly symmetric clusters which gets rid of this issue. Here we will show that the edge-field NLCE also removes the diverging subseries for the full-graph expansion and leads to even better results.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig13.pdf}
\end{center}
\caption{(Color online) Ground-state energy per site $e_0/J$ for the spin-$1$ kagome lattice and $\lambda=0.82$. Different colored symbols correspond to different orders of the graph expansion. The lines between them are guide to the eyes. \textit{Inset:} Linear scaling (orange line) through the minimal values (orange circles) of order $5$ and $6$ as a function of $1/N_{\rm tr}^2$. The black diamond depicts the iPEPS result from Ref.~\onlinecite{Liu15}.}
\label{fig:NLCE_kagomeOne_lambda0_82}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{Fig14.pdf}
\end{center}
\caption{(Color online) Ground-state energy per site $e_0/J$ for the spin-$1$ kagome lattice as a function of $\lambda$. Triangles are the results without edge-fields from Ref.~\onlinecite{Ixert15}, whereas circles are results with a small edge-field. The corresponding lines are guide to the eyes. Black squares depict the scaled values of order $5$ and $6$, red stars are the best results of the reorganized graph expansion from Ref.~\onlinecite{Ixert15} and the solid black line illustrates the iPEPS result with bond dimension $D^{*}=5$ from Ref.~\onlinecite{Liu15}. \textit{Inset:} Difference between the ground-state energy per site from the scaled edge-field NLCE or the reorganized expansion from Ref.~\onlinecite{Ixert15} and the ground-state energy per site from iPEPS \cite{Liu15}.}
\label{fig:NLCE_kagomeOne}
\end{figure}
We begin our discussion with the specific case $\lambda=0.82$. This value is close to the quantum critical point of the spin-$1$ triangle chain \cite{Ixert15} and should therefore be most problematic for the full-graph expansion. Furthermore, we can compare our NLCE data directly with the iPEPS-value from Ref.~\onlinecite{Liu15}. The NLCE ground-state energy per site for $\lambda=0.82$ is shown in Fig.~\ref{fig:NLCE_kagomeOne_lambda0_82} as a function of $K/J$ for all NLCE orders up to six. Clearly, the NLCE without (or very small) edge-fields varies strongly for different NLCE orders. In contrast, the NLCE orders 4 to 6 display well-defined minima in $K/J$ which approach monotonously the iPEPS value from below. Although there is no obvious length scale, which can be well defined in the full-graph expansion, we observe heuristically that our NLCE data scales almost linearly in $1/N_{\rm tr}^2$. A linear fit through orders $5$ and $6$ in $1/N_{\rm tr}^2$ leads to a scaled value which is in very good agreement with the iPEPS-value, as shown in the inset of Fig.~\ref{fig:NLCE_kagomeOne_lambda0_82}. The edge-fields are therefore well suited to regularize the NLCE in this most problematic $\lambda$-regime. The same is true on the full $\lambda$-axis. This can be already seen in Fig.~\ref{fig:NLCE_kagomeOne} where we compare the NLCE without edge-fields to the case of a finite field $K/J=0.05$. It is clearly visible that the NLCE data without edge-fields are erratic (especially around $\lambda\approx 0.8$) while the edge-field smoothens the NLCE for all values of $\lambda$.
As for $\lambda=0.82$, we have performed scalings for various values of $\lambda$ shown as black squares in Fig.~\ref{fig:NLCE_kagomeOne}. Remarkably, they agree very well with the iPEPS-data. The differences between the energies from scaled edge-field NLCE and iPEPS is even smaller than the same difference using the reorganized NLCE from Ref.~\onlinecite{Ixert15} as can be seen in the inset of Fig.~\ref{fig:NLCE_kagomeOne}. Interestingly, the reorganized NLCE displays the largest difference to the iPEPS energy for $\lambda\approx 0.8$, i.e.~it still ``feels'' the instability observed in the full-graph expansion without edge-fields (although on a much smaller energy scale). In contrast, the edge-field NLCE shows an almost constant difference to the iPEPS values with constant bond dimension $D^{*}=5$ for $\lambda\in [0.5,1.0]$ which we attribute mostly to uncertainties in the scaling (as well as to the finite $D^{*}$). For $\lambda=1$ -- corresponding to the isotropic Heisenberg model on the kagome lattice -- our best result from Ref.~\onlinecite{Ixert15} using a reorganized graph expansion was $e_{1,\text{even}}^{(2)}=1.4114 J$. The full-graph expansion with edge-fields yields the scaled value $e_0^{\text{scaled}}=-1.41203 J$ which is nearly within the error estimations of the value $e_0^{\text{iPEPS}}=-1.4116(4)J$ obtained by the iPEPS calculations.
Furthermore it also compares well to the DMRG-Value $e_0^{\text{DMRG}}=−1.410(2)J$ from Ref.~\onlinecite{Changlani15}.
\section{Conclusions}
\label{Sect:conclusions}
In this work we have presented a generic scheme to perform NLCEs in long-range ordered quantum phases. The essential idea is to incorporate the effect of the long-range order into the exact diagonalization on graphs. Then each graph ``remembers'' the ordered quantum phase out of which it is taken from and the correct quantum fluctuations are captured in the NLCE. This is achieved by assuming an appropriate ordered reference state outside each graph so that the edge-couplings give rise to an effective edge-field. The edge-field breaks the relevant symmetry, which corresponds to the underlying long-range order, and it ensures that the NLCE is carried out for each graph inside the correct quantum phase. The field is subextensive, since it scales with the perimeter of the graphs. As a consequence, the NLCE is regularized and fluctuations from different quantum phases and different quantum critical behavior on lower-dimensional subgraphs are excluded.
We have implemented four different NLCEs. The full-graph expansion as well as the rectangular-graph expansion are complicated to extrapolate to the infinite-order limit, since in the full-graph expansion no proper length scale can be defined while the NLCE data squences are not smooth for different NLCE orders in the rectangular-graph expansion. Notably, we found heuristically for the full-graph expansion in the trimerized ground state of the spin-$1$ Heisenberg model on the kagome lattice, that the NLCE data sequence of the ground-state energy is approximately linear in $1/N^2$. This suggests that scalings might be also valuable for NLCEs using a full-graph expansion. In contrast, both the arithmetic as well as the square-graph expansion give satisfactory results and scalings towards the infinite-order limit can be performed successfully. However, we have observed that the scaling is more complicated in the arithmetic expansion which can be traced back to the fact that different clusters of the same NLCE order have almost the same typical length scale but different aspect ratios. This is different for the square-graph expansion where each order is clearly dominated by the length scale of the largest quadratic cluster. In this work we have also investigated two different scaling laws for the NLCE data. Although the quality of both scaling types is very similar (especially for the ground-state energy per site). We are strongly convinced that it is an important but challenging task to formulate proper scaling laws for NLCEs from first principles.
We have applied the edge-field NLCE to the long-range spin-ordered phases of the spin-1/2 and spin-1 Heisenberg model on the square and triangular lattice as well as to the trimerized valence bond crystal of the spin-1 Heisenberg model on the kagome lattice. In all cases we have found that edge-fields are essential to obtain a well-behaved NLCE for the ground-state energy per site and the sublattice magnetization. We have chosen the unfrustrated spin-1/2 Heisenberg model on the square lattice to test four different cluster expansions and we formulated two strategies to determine the optimal value of the edge-field. It is found that all NLCEs compare reasonably well with state-of-the-art numerical QMC data. Furthermore, the arithmetic and the square-graph expansion perform best and both approaches to determine the edge-field yield similarly good results. We find it interesting that the square-graph expansion works, in contrast to the arithmetic expansion, even better for the geometrically frustrated spin-1/2 triangular lattice compared to the unfrustrated square lattice in the sense that no even-odd effect is present and therefore scalings can be performed through more NLCE data points. It would be nice if one could formulate a reorganized NLCE with different elementary clusters which does not show this even-odd effect on the square lattice. Overall, we are convinced that the edge-field NLCE can be extended successfully into several directions which we hope to tackle in future works. First, it would be interesting to apply other numerical tools to extract quantitatively physical quantities for larger open clusters. This would allow to improve the quality of the infinite-order scaling considerably resulting in even better NLCE results. Second, one should apply the edge-field NLCE to systems where several quantum phases compete with each other. A comparison of different NLCE energies and their convergence behavior should allow the determination of the ground-state phase diagram. Third, one might speculate whether a formulation of edge-field NLCEs can be achieved where the mean-field reference product state is replaced by more complex entangled tensor networks. This would open the fascinating perspective to tackle also exotic quantum disordered phases with NLCEs.
\begin{acknowledgments}
We thank W. Li for providing numerical data. This research has been supported by the Virtual Institute VI-521 of the Helmholtz association.
\end{acknowledgments}
|
\section{Introduction}
The Standard Model of particle physics (SM) describes with exquisite precision the interactions of all known elementary particles. In spite of intensive searches, no significant deviation from the SM has been detected in collider or other particle physics experiments {\cite{Olive:2016xmw}.} However, several long-standing problems indicate that new physics beyond the SM is needed to achieve a complete description of Nature. First of all, there is overwhelming evidence, ranging from the cosmic microwave background (CMB) to the shapes of the rotation curves of spiral galaxies, that nearly 26\% of the Universe is made of yet unidentified dark matter (DM) \cite{Ade:2015xua}. Moreover, the SM cannot generate the primordial inflation needed to solve the horizon and flatness problems of the Universe, as well as to explain the statistically isotropic, Gaussian and nearly scale invariant fluctuations of the CMB \footnote{If the Higgs field has a large coupling to the curvature $R$, inflation might be obtained \cite{Bezrukov:2007ep}. However, as we will argue, this requires physics beyond the SM.}. The SM also lacks enough CP violation to explain why the Universe contains a larger fraction of baryonic matter than of anti-matter. Aside from these three problems at the interface between particle physics and cosmology, the SM suffers from a variety of intrinsic naturalness issues. In particular, the neutrino masses are disparagingly smaller than any physics scale in the SM and, similarly, the strong CP problem states that the $\theta$-parameter of quantum chromodynamics (QCD) is constrained from measurements of the neutron electric dipole moment {\cite{Afach:2015sja,Schmidt-Wellenburg:2016nfv}} to lie below an unexpectedly small value: {$|\theta|\lesssim 10^{-10}$.}
In this Letter we show that these problems may be intertwined in a remarkably simple way, with a solution pointing to a unique new physics scale around $10^{11}$ GeV. The SM extension we consider consists just of {a KSVZ-like} axion model \cite{Kim:1979if,Shifman:1979if} and three right-handed (RH) heavy SM-singlet neutrinos \footnote{One may also chose alternatively the DFSZ axion model. The inflationary predictions in this model stay the same, but the window in the axion mass will move to larger values \cite{Kawasaki:2014sqa}. Importantly, in this case the PQ symmetry is required to be an accidental rather than an exact symmetry in order to avoid the overclosure of the Universe due to domain walls \cite{Ringwald:2015dsf}}. This extra matter content was recently proposed in \cite{Dias:2014osa}, where it was emphasised that in addition to solving the strong CP problem, providing a good dark matter candidate (the axion), explaining the origin of the small SM neutrino masses (through an induced seesaw mechanism) and the baryon asymmetry of the Universe (via thermal leptogenesis), it could also stabilise the effective potential of the SM at high energies thanks to a threshold mechanism \cite{Lebedev:2012zw,EliasMiro:2012ay}. This extension also leads to successful primordial inflation by using the modulus of the KSVZ SM singlet scalar field \cite{Fairbairn:2014zta}. Adding a cosmological constant to account for the present acceleration of the Universe, this Standard Model Axion Seesaw Higgs portal inflation (SMASH) model offers a self-contained description of particle physics from the electroweak scale to the Planck scale and of cosmology from inflation until today. Although some parts of our SMASH model have been considered separately \cite{Langacker:1986rj,Shin:1987xc,Shaposhnikov:2006xi,Bezrukov:2007ep,Lerner:2009xg,Lebedev:2011aq,Fairbairn:2014zta,Boucenna:2014uma,Bertolini:2014aia,Salvio:2015cja,Kahlhoefer:2015jma,Clarke:2015bea,Ahn:2015pia}, a model incorporating all of them simultaneously had not been proposed {until now.}
{Remarkably, SMASH} {can accommodate} {the} {constraints from cosmological observations and Higgs stability, successfully reheat the Universe, provide the correct dark matter abundance and explain the origin of the baryon asymmetry.}
{In this Letter, we present the most important aspects and predictions of SMASH}. {Further details are given in \cite{Ballesteros:2016xej}.}
\section{The SMASH model}
We extend the SM with a new complex singlet scalar field $\sigma$ {and a Dirac fermion $\mathcal{Q}$, which can be split in} two Weyl fermions $Q$ and $\tilde Q $ in the $\mathbf 3$ and $\mathbf{\bar{3}}$ representations of $SU(3)_c$ {with charges} $-1/3$ and $1/3$ under $U(1)_Y$. {This ensures that $\mathcal{Q}$ can coannihilate and decay into SM} {quarks, thereby evading possible overabundance problems~\cite{Nardi:1990ku,Berezhiani:1992rk}.}
We also add three RH fermions $N_i$.
{The model is endowed with a new Peccei-Quinn (PQ) global $U(1)$ symmetry \cite{Peccei:1977hh}, which also plays the role of lepton number {in our case.} {Using left-handed Weyl spinors, we denote by $q_i$, $u_i$ and $d_i$ the SM quark doublet and the conjugates of the right-handed quarks of each generation $i=1,2,3$; and by $L_i$ and $E_i$ the corresponding lepton doublet and the conjugate of the right-handed lepton. Denoting the Higgs by $H$,} {the charges under the PQ symmetry are: $q(1/2)$, $u(-1/2)$, $d(-1/2)$, $L(1/2)$, $N(-1/2)$, $E(-1/2)$, $Q(-1/2)$, $\tilde Q(-1/2)$, $\sigma(1)$, $H(0)$.} {The most general Yukawa couplings involving the new fields are:} {${\cal L}\supset -[F_{ij}L_i\epsilon H N_j+\frac{1}{2}Y_{ij}\sigma N_i N_j+y\, \tilde Q \sigma Q+z_i\,\sigma Q d_i+h.c.]$, where $\epsilon$ is the two-component antisymmetric symbol.} {The Yukawa couplings $F$ and $Y$} realise the seesaw mechanism once $\sigma$ acquires a vacuum expectation value (VEV) {$\langle\sigma\rangle=v_\sigma/\sqrt{2}$,} giving a neutrino mass matrix of the form {$m_\nu=-FY^{-1}F^Tv^2/(\sqrt{2}v_\sigma)$,} with $v=246$ GeV. The strong CP problem is solved as in the standard KSVZ scenario, with the role of the axion decay constant, $f_A$, played by $v_\sigma=f_A$. Due to non-perturbative QCD effects, the angular part of {$\sigma=(\rho+v_\sigma)\exp(iA/f_A)/\sqrt{2}$,} the axion field $A$ \cite{Weinberg:1977ma,Wilczek:1977pj}, gains a potential with an absolute minimum at $A=0$. At energies above the QCD scale, the axion-gluon coupling is $\mathcal{L}\supset -(\alpha_s/8\pi)(A/f_A) G\tilde G$, solving the strong CP problem when
$\langle A\rangle$ relaxes to zero\footnote{{Since the $\tilde Q,Q$ quarks have hypercharge and the PQ charge assignments are different than in the standard KSVZ scenario, the axion has a non-standard coupling to the photon, as well as a coupling to neutrinos \cite{Ballesteros:2016xej}}}. The latest lattice computation of the axion mass gives $m_A=(57.2\pm0.7)(10^{11}{\rm GeV}/f_A)\,\mu{\rm eV} $~\cite{Borsanyi:2016ksw}.
\section{Inflation}
{Given the symmetries of SMASH, the most general renormalisable tree-level potential is}
\begin{align}
\label{scalar_potential} \nonumber
V(H,\sigma ) & = \lambda_H \left( H^\dagger H - \frac{v^2}{2}\right)^2
+\lambda_\sigma \left( |\sigma |^2 - \frac{v_{\sigma}^2}{2}\right)^2\\ &+
2\lambda_{H\sigma} \left( H^\dagger H - \frac{v^2}{2}\right) \left( |\sigma |^2 - \frac{v_{\sigma}^2}{2}\right)\,.
\end{align}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.475\textwidth]{log_contours_2}
\caption{\label{fig:r_vs_ns} \small {The tensor-to-scalar ratio, $r$, vs the scalar spectral index, $n_s$, at} {$k_0=0.002$} {Mpc$^{-1}$ for the inflationary potential \eq{genpotential}, assuming} {$|\lambda_{H\sigma}|\ll\lambda_H$.} {We show lines of constant $\xi_\sigma$ (dashed) and constant number of e-folds from the time the scale} {$k_0=0.002$} {Mpc$^{-1}$ exits the horizon to the end of inflation (thin solid).
In SMASH, the EOS of the Universe is like $w=1/3$ immediately after inflation, which allows to predict $N$ (thick red line).}
Coloured regions show observational constraints at 68\% and 95\% CL from \cite{Ade:2015lrj}.}
\end{center}
\end{figure}
In the unitary gauge, there are two scalar fields that could drive inflation: $h$, the neutral component of the Higgs doublet $H^t=(0\,,h)/\sqrt{2}$, and the modulus of the new singlet, {$\rho^2=2\,|\sigma|^2$.} In the context of the SM, it was proposed in \cite{Bezrukov:2007ep} that $h$ could be the inflaton if it is non-minimally coupled to the scalar curvature $R$ through a term $\mathcal{L}\supset -\sqrt{-g}\,\xi_H\,H^\dagger H\,R$ \cite{Salopek:1988qh}, with $\xi_H\sim 10^4$. Such a large value of $\xi_H$ is required by the constraint $\xi_H\sim 10^5 \sqrt{\lambda_H}$ to fit the amplitude of primordial fluctuations and it implies that perturbative unitarity breaks down at the scale {$\Lambda_U=M_P/\xi_H\ll M_P$} \cite{Burgess:2009ea, Barbon:2009ya}, where $M_P=1/\sqrt{8\pi\,G}$ is the reduced Planck mass. This raises a serious difficulty for Higgs inflation, which requires Planckian values of $h$ and an energy density of order $\Lambda_U^2$. Since new physics is expected at or below $\Lambda_U$ to restore unitarity, the predictivity of Higgs inflation is lost, because the effect of this new physics on inflation is undetermined. This issue affects some completions of the SM such as the $\nu$MSM \cite{Asaka:2005an,Asaka:2005pn} and the model proposed in \cite{Salvio:2015cja}.
Instead, inflation in SMASH is mostly driven by $\rho$, with a non-minimal coupling $\mathcal{L}\supset -\sqrt{-g}\,\xi_\sigma\,\sigma^* \sigma\,R$, {where $\xi_\sigma\lesssim 1$ ensures that the scale of perturbative unitarity breaking is at $M_P$ (provided that also $\xi_H\lesssim 1$).}
Neglecting $\xi_H$ \footnote{{Taking into account radiative corrections to $\xi_H$ and $\xi_\sigma$ one can check that the window $2\times 10^{-3}\lesssim\xi_\sigma\lesssim 1$ ensures that $\xi_H$ can be neglected with respect to $\xi_\sigma$}}, predictive slow-roll inflation in SMASH can happen along two directions in field space: the $\rho$-direction for $\lambda_{H\sigma}>0$ and the {line} $h/\rho=\sqrt{-\lambda_{H\sigma}/\lambda_H}$ for $\lambda_{H\sigma}<0$. We call them hidden scalar inflation (HSI) and Higgs-hidden scalar inflation (HHSI), {respectively. In both cases, inflation} can be described {in the Einstein frame} by a single canonically normalised field $\chi$ with potential
\begin{equation}
\label{genpotential}
\tilde V(\chi) = \frac{\lambda }{4}\rho(\chi)^4\left(1+\xi_{\sigma}\frac{\rho(\chi)^2}{M_P^2}\right)^{-2}\,,
\end{equation}
where $\lambda$ {stands for $\lambda_\sigma$ in HSI and for
$\tilde\lambda_\sigma=\lambda_\sigma-\lambda_{H\sigma}^2/\lambda_{H}$ in HHSI. }
The field $\chi$ is the solution of $\Omega^2\,d\chi/d\rho\simeq (b\,\Omega^2+6\,\xi_\sigma^2\,\rho^2/M_P^2)^{1/2}$, {being} $\Omega\simeq 1+\xi_\sigma\,\rho^2/M_P^2$ the Weyl transformation into the Einstein frame; and {$b=1$ (for HSI) or $b=1+|\lambda_{H\sigma}/\lambda_H|$ (for HHSI). The small value of $|\lambda_{H\sigma}|$ required for stability (see below) typically means that $b\sim1$ in HHSI, which makes impossible distinguishing in practice between HSI and HHSI from the inflationary potential. However, even a small Higgs component in the inflaton is relevant for reheating, as we will later discuss. The predictions of the potential \eq{genpotential}} in the case $\lambda=\lambda_\sigma$ {(or $b\rightarrow 1$ in HHSI)} for { the tensor-to-scalar ratio $r$} vs the scalar spectral index $n_s$ are shown in FIG. \ref{fig:r_vs_ns} for various values of $\xi_\sigma$.
{In SMASH, the equation of state} {(EOS)} {of the Universe after inflation is $w=1/3$ (like radiation) uninterruptedly until the} {standard epoch of matter-radiation equality is reached; see the reheating section below.} {This allows to compute the number of e-folds of inflation, $N(k)$, for any comoving scale, $k$, matching precisely the predictions for the inflationary spectrum with the observations of the CMB \cite{Liddle:2003as}. {This} determines the {thick line} of FIG.\ \ref{fig:r_vs_ns} as the SMASH prediction for $r(n_s)$ and {$N(k_0)$} at the {fiducial scale $k_0 \equiv 0.002$ Mpc$^{-1}$, which we use through the Letter for all the primordial inflationary parameters.}
The prediction spans {$N\sim(59,62)$}, depending on $n_s$, and its width {($\sim 1$ e-fold)} {quantifies the small} uncertainty on the transient regime from the end of inflation to {radiation domination.}}
{Note that the the} {condition $\xi_\sigma\lesssim 1$ corresponds to {$r\gtrsim 0.004$,} {which is within the planned sensitivities of {PIXIE \cite{Kogut:2011xw}, LiteBird \cite{Matsumura:2013aja}, CMB-S4 \cite{Abazajian:2016yjj} and COrE+ (which will measure $r$ with an error of $\Delta r\sim 4\times 10^{-4}$).}
The {joint} {constraints} of the Planck satellite and the BICEP/Keck array \cite{Ade:2015xua,Array:2015xqh}
{give} {$r< 0.07$ at 95\% CL,} {corresponding in SMASH to $\xi_\sigma\gtrsim 6\times 10^{-3}$.}
{Taking into account the former constraints, }the spectral index $n_s$ {at $k=k_0$ lies in the interval $(0.962, 0.966)$, and its running $\alpha=d\, n_s/d \ln k$ {lies} in {the range} $(- 7,-6)\times 10^{-4}$, which may be} probed e.g.\ by future observations of the 21 cm emission line of Hydrogen \cite{Mao:2008ug}. {Since inflation is effectively single-field slow-roll, non-Gaussian features are suppressed by $\sim(1-n_s)$ \cite{Acquaviva:2002ud,Maldacena:2002vr}.} These values of the primordial parameters are perfectly compatible with the latest CMB data, and the amount of inflation that is produced solves the horizon and flatness problems.
Given the current bounds on $r$ and $n_s$, {and the fact that fitting the amplitude of primordial scalar fluctuations requires $\xi_\sigma \sim 10^5 \sqrt{\lambda}$,} fully consistent (and predictive) inflation in SMASH occurs if {$5\times 10^{-13}\lesssim \lambda \lesssim 5\times 10^{-10}$.}
\section{Stability}
For the measured central values of the Higgs and top quark masses {\cite{Olive:2016xmw},} the Higgs quartic coupling of the SM becomes negative at $h=\Lambda_I\sim 10^{11}$ GeV \footnote{$\Lambda_I$ is very sensitive to small variations of the top mass, to the extent that the potential may be completely stable for sufficiently low (but still allowed) values.}. If no new physics changes this behaviour, Higgs inflation is not viable, since it requires a positive potential at Planckian field values. Moreover, {this instability is} a problem even if another field drives inflation. This is because {scalars that are light compared} to the Hubble scale, $\mathcal H$, acquire fluctuations of order {$\sim\mathcal{H}/2\pi$.} These {can make the Higgs field move} into the instability region of the potential, {which would contradict} the present electroweak vacuum \footnote{The fluctuations can be suppressed if $\xi_H\gtrsim 1$, which induces a large effective mass for the Higgs during inflation. In that case it would still be necessary to check that the classical trajectory of the fields does not fall in the negative region of the potential}.
Remarkably, {the} Higgs portal term $\propto \lambda_{H\sigma}$ in \eq{scalar_potential} allows {stability of the SMASH potential} via the threshold-stabilisation mechanism {of \cite{Lebedev:2012zw,EliasMiro:2012ay}, {which relies on a} {nontrivial matching with the SM potential} {at low energies.} The {matched Higgs quartic in the SM is} {$\tilde\lambda_H\equiv\lambda_H-\delta$, where the threshold correction is $\delta\equiv\lambda^2_{H\sigma}/\lambda_\sigma$. Even if the running of $\tilde\lambda_H$ in the SM makes it negative, the actual Higgs quartic coupling} {in the UV theory,} $\lambda_H$, {can remain positive provided that $\delta$ is large enough.} }
{A more detailed analysis \footnote{{At large field values, stability demands positivity of quartic couplings, but at intermediate field values, where negative quadratic interactions are important, one has to ensure
positivity of the potential along the} {potential} {energy valleys}} shows that, for} $\lambda_{H\sigma}>0$, absolute stability requires \cite{Ballesteros:2016xej}
\begin{eqnarray}
\label{stabilitycondition1}
\left\{\begin{array}{cc}
\tilde \lambda_H, \tilde \lambda_\sigma>0, \quad\text{for}\quad h < \sqrt{2}\Lambda_h\\
\lambda_H, \lambda_\sigma>0, \quad\text{for}\quad h > \sqrt{2}\Lambda_h \\
\end{array}\right.,
\end{eqnarray}
{where $\Lambda_h^2\equiv\lambda_{H\sigma}\,v_\sigma^2/\lambda_H$ and all the couplings run with the {beta functions} of SMASH, not the SM. The scale $\sqrt{2}\Lambda_h$ arises as the divide between large and small field values of $h$, for which $v_\sigma$ cannot be neglected and the quadratic interactions are relevant, as can be seen from \eq{scalar_potential}.} Instead, for $\lambda_{H\sigma}<0$, the stability condition is just
$\tilde \lambda_H, \tilde \lambda_\sigma>0$, for all $h$.
The Higgs direction is the one most prone to be destabilised (from top loops) and the potential must remain positive beyond the $h\sim M_P$ values needed for inflation. A one-loop analysis shows that a value of $\delta$ above $10^{-3}$--$10^{-1}$ (depending on the top mass, see FIG.\ \ref{fig:delta}) ensures stability up to $h_{30}\approxeq 30 M_P$ for a Higgs mass of $125.09$~GeV. }{
Finally, in SMASH, instabilities could also originate in the {direction of $\rho$} due to quantum corrections from {$N_i$ and $Q,\tilde Q$. Stability in this direction, requires $6 y^4+\sum Y^4_{ii}\lesssim 16\pi^2\lambda_\sigma/\log\(h_{30}/\sqrt{2\lambda_\sigma}v_\sigma\)$} \footnote{{In this expression and in FIG.\ \ref{fig:delta}, we demand stability up to a large RG scale $\mu\sim h_{30}$, which is sufficiently higher than $h$ during inflation.}.}.}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.43\textwidth]{deltamin_suma.pdf}
\caption{\label{fig:delta} \small Minimum value of the threshold correction to the Higgs quartic coupling, $\delta=\lambda_{H\sigma}^2/\lambda_{\sigma}$, for stable SMASH potentials at RG scales $\mu=m_\rho$ (solid) and $\mu=30 M_P$ (dashed), for $\lambda_{H\sigma}>0$ (black) and $\lambda_{H\sigma}<0$ (blue).}
\end{center}
\end{figure}
\section{Reheating}
{SMASH provides a complete model of cosmology for which the evolution after inflation can be calculated. The PQ symmetry is spontaneously broken during inflation by the large evolving value of $\rho$.} Slow-roll inflation ends at $\rho_{\rm end}\sim \mathcal{O}(M_P)$, where the effect of $\xi_\sigma$ is negligible. Since $\rho_{\rm end} \gg v_\sigma$, the inflaton starts to undergo Hubble-damped oscillations in a quartic potential.
{The first oscillations of the inflaton constitute a phase of so-called preheating \cite{Kofman:1997yn}, during which} fluctuations of $\sigma$ in the direction orthogonal to the inflaton increase exponentially.} The post-inflationary background can be understood as {a} homogeneous condensate of particles with energy given by the oscillation frequency $\omega(t)\sim\sqrt{\lambda}\rho_{\rm end}/a(t)$, {where $a(t)$ is the scale factor of the Universe and $t$ denotes cosmic time \footnote{We use natural units for which $\hbar=c=1$.}.} In SMASH, $\lambda$ is the weakest coupling and thus SM particles coupled to the inflaton have effective masses $\propto \rho(t)$, which are much larger than $\omega(t)$ except when $\rho(t)\sim 0$. {Higgs particles and} electroweak bosons could in principle be produced by parametric resonance \cite{Greene:1997fu} {at these crossings} but they either have large self-interactions or decay very efficiently into SM fermions. In contrast, the effective mass of $\sigma$ excitations is $\sim \sqrt{\lambda}\rho(t)\sim \omega(t)$, {which allows them to grow} by parametric resonance. The growth of fluctuations of a complex {inflaton field} in a quartic potential was studied {analytically in~\cite{Greene:1997fu} and} numerically in \cite{Tkachev:1998dc}. Our own numerical simulations~\cite{Ballesteros:2016xej} corroborate their results. {After the first $\sim 14$ oscillations after inflation,} {the fluctuations of $\sigma$ become} {as large as the inflaton amplitude $\langle|\sigma|^2\rangle \sim \rho_{\rm end}^2/a^2$, so the PQ symmetry is non-thermally restored.} Only if $v_\sigma$ were larger than $\sim 10^{-2}M_P$ would the field $\rho$ get trapped around its minimum $\rho=v_\sigma$ before the non-thermal restoration can occur. However, such high values of $v_\sigma$ are ruled out by CMB axion isocurvature constraints~\cite{Fairbairn:2014zta} \footnote{
Reference \cite{Fairbairn:2014zta} does not mention parametric resonance and non-thermal restoration of the PQ symmetry, because it {is} model dependent. The authors of \cite{Fairbairn:2014zta} showed that a large inflaton VEV suppresses the isocurvature constraints. We have redone the analysis finding stronger bounds~\cite{Ballesteros:2016xej}.
Probably, ref. \cite{Fairbairn:2014zta} used the Jordan frame VEV instead of the effective axion
decay constant in the Einstein frame, $~\chi$.
In any case, $v_\sigma \gtrsim 10^{-2}M_P$ is safely excluded. }.
{Aside from these common features, reheating progresses differently for HSI and HHSI. The reason is that the small Higgs component of the inflaton in HHSI (which is lacking in HSI) accelerates in that case the production of SM particles. We will now discuss the two cases separately.}\\
{\bf Reheating for {HSI} ($\lambda_{H\sigma}>0$):}
During preheating, Higgs bosons are {non-resonantly} produced during inflaton crossings because of the large value of the Higgs self-coupling \cite{Anisimov:2008qs}, as well as the fast decay of Higgses into tops and gauge bosons. {When the PQ symmetry is non-thermally restored, the} induced Higgs mass $\sqrt{\lambda_{H\sigma}}\sqrt{\langle |\sigma|^2\rangle}$ stabilises around a large value $\sqrt{\lambda_{H\sigma}} \rho_{\rm end}/a(t)\gg \omega(t)$, thus blocking Higgs production. Efficient reheating has to wait until the {spontaneously symmetry breaking (SSB)} of the PQ symmetry, i.e.\ when $\langle |\sigma|^2\rangle$ becomes $\sim v_\sigma^2$. We have simulated numerically the phase transition, finding that the energy initially stored in $\sigma$ fluctuations becomes equipartitioned into axions and $\rho$ particles. The latter can soon decay into Higgses and reheat the SM sector. The corresponding reheating temperature is $T_R\sim v_{11} \lambda_{10}^{3/8} \delta_3^{-1/8}\,10^7 \, {\rm GeV}$, where we introduce SMASH benchmark values: $v_{11}= v_\sigma/(10^{11}$~GeV), $\lambda_{10}=10^{10}\lambda_\sigma$, $\delta_3=\delta/0.03$ \footnote{We will see that in this scenario axion DM requires $v_\sigma=f_A\sim 10^{11}$ GeV.}. The accompanying axions are relativistic and remain decoupled from such a low temperature SM thermal bath~\cite{Graf:2010tv}. They contribute to the late Universe expansion rate as extra {(relativistic)} neutrino species. We estimate $\Delta N_\nu^{\rm eff}\sim 0.96\, (\lambda_{10}/\delta_3 v_{11})^{1/6}$ above the SM value $N_\nu^{\rm eff}({\rm SM})=3.046$ \cite{Mangano:2001iu}. {Current CMB and baryon acoustic oscillation data give $N_\nu^{\rm eff} = 3.04 \pm 0.18$ at 68\% CL~\cite{Ade:2015xua}, disfavouring HSI.}\\
{\bf Reheating for {HHSI} ($\lambda_{H\sigma}<0$):}
As in {HSI,} the direct production of Higgs excitations stops when the PQ symmetry is non-thermally restored. {However,} the Higgs component of the inflaton {continues} to oscillate around {$h\sim 0$ so that $W$ and $Z$ gauge bosons} can still be produced during crossings. The fast decay of $W,Z$ into light fermions when $h$ moves away from zero prevents {their exponential accumulation but makes} the comoving energy in light fermions increase.
When light particles thermalise, a population of $W,Z$ bosons is created by the thermal bath during crossings (when their mass is below the temperature) and decays when their mass grows {with $h$.} This mechanism enhances {the drain} of energy from the inflaton to the SM bath.
Using Boltzmann equations with thermal and non-thermal sources, and accounting for the energy loss of the background {fields,} we have calculated numerically the reheating temperature, {finding $T_R\sim{\cal O}(10^{10} $GeV) for the values of $\lambda$ and $\delta$ satisfying the requirements for inflation and stability.
The critical temperature for the PQ phase transition is ${T_c}\simeq 2\sqrt{6\lambda_\sigma}\,v_\sigma/\sqrt{8(\lambda_\sigma+\lambda_{H\sigma})+\sum_i Y^2_{ii}+6 y^2}$ \cite{Ballesteros:2016xej}. For SMASH benchmark values $|\lambda_{H\sigma}|\gg \lambda_\sigma$, and requiring the previous stability bound on the Yukawa couplings of the new fermions, $T_c\sim 0.01\,v_\sigma< T_R$. Therefore, the PQ symmetry{, which had been non-thermally restored by preheating, is also restored thermally at the end of reheating. A few Hubble times after,} {the temperature} {drops below $T_c$ and the PQ symmetry becomes spontaneously broken, this time for good.}
We thus predict a thermal abundance of axions, which decouple at ${\rm min}\{T_c,T_A^{\rm dec}\}$ where $T_A^{\rm dec} \simeq 2\times 10^9\ {\rm GeV} v_{11}^{2.246}$ \cite{Masso:2002np,Graf:2010tv,Salvio:2013iaa}. Considering $g_\ast = 427/4$ relativistic degrees of freedom at axion decoupling we get $\triangle N_\nu^{\rm eff}\simeq 0.03$, which is much smaller than in HSI and in good agreement with current data. This small value of $\triangle N_\nu^{\rm eff}$ could be probed with future CMB polarisation experiments \cite{Abazajian:2013oma,Errard:2015cxa}. As discussed in \cite{Baumann:2016wac}, a non-detection of new thermal relics with future CMB probes reaching $ \Delta N^{\rm eff}_\nu\sim 0.01$ will imply that if such relics exist they were never in thermal equilibrium with the SM.}
{Finally, we remark that the EOS of the Universe is $w=1/3$ both in the period of inflaton oscillations in a quartic potential \cite{Shtanov:1994ce} and the non-thermally PQ restored phase because the evolution is conformal in a quartic potential. This is so both for HHSI and HSI. However, in HSI, there is a small period of matter domination before the $\rho$ particles decay to reheat the SM, whose effects on $N$ are within the uncertainties. }
\section{Dark matter}
{At the spontaneous breaking of the PQ symmetry, a network of cosmic strings is formed} {both in
{HHSI} and HSI.} {In the first case, this happens by the standard Kibble mechanism in thermal equilibrium} {\cite{Kibble:1980mv}} {and in the second, non-thermally~\cite{Tkachev:1998dc}. }
{The evolution of the network} leads to a population of low-momentum axions that together with those arising from the realignment mechanism \cite{Preskill:1982cy,Abbott:1982af,Dine:1982ah} constitute the dark matter in SMASH.
Requiring that all the DM is made of axions demands
\begin{equation}
\label{classic_window_v}
3\times 10^{10}\, {\rm GeV}\lesssim v_\sigma \lesssim 1.2\times 10^{11}\, {\rm GeV},
\end{equation}
which translates into the mass window
\begin{equation}
\label{classic_window_m}
50\, \mu{\rm eV}\lesssim m_A\lesssim 200\, \mu{\rm eV},
\end{equation}
where we have updated the results of \cite{Kawasaki:2014sqa} with the latest axion mass data~\cite{Borsanyi:2016ksw}.
The main uncertainty {arises} from the string contribution~\cite{Kawasaki:2014sqa,Fleury:2016xrz}, {which we estimate as 3-4 times larger than the misalignment one;} the uncertainty is expected to be diminished in the near future~\cite{Moore:2016itg,Fleury:2015aca}. {The SMASH axion mass window \eq{classic_window_m} will be probed in the upcoming decade by direct detection experiments such as {MADMAX~\cite{MADMAX,TheMADMAXWorkingGroup:2016hpc}} and ORPHEUS~\cite{Rybka:2014cya}. A sizeable part of the DM in this scenario may be in the form of axion miniclusters \cite{Hogan:1988mp}, which offer interesting astrophysical signatures~\cite{Kolb:1995bu,Tinyakov:2015cgg}.}
\section{Baryogenesis}
The origin of the baryon asymmetry of the Universe is explained in SMASH from thermal leptogenesis \cite{Fukugita:1986hr}. This requires {the} massive RH neutrinos, {$N_i$, acquiring} equilibrium abundances and then decaying when their production rates become Boltzmann suppressed. {As we have seen, in HHSI}, $T_R>T_c$ for stable models in the DM window \eqref{classic_window_m}. The RH neutrinos become massive after the PQ {SSB}, and those with masses $M_i<T_c$ retain an equilibrium abundance. The stability bound on the Yukawa couplings $Y_{ii}$ enforces $T_c > M_1$, so that at least the lightest RH neutrino stays in equilibrium. Moreover, the annihilations of the RH neutrinos tend to be suppressed with respect to their decays. This allows for vanilla leptogenesis from the decays of a single RH neutrino, which demands $M_1\gtrsim 5\times10^8$ GeV \cite{Davidson:2002qv,Buchmuller:2002rq}. However, for $v_\sigma$ as in \eqref{classic_window_v}, this is just borderline compatible with stability. Nevertheless, leptogenesis can occur with a mild resonant enhancement \cite{Pilaftsis:2003gt} for a less hierarchical RH neutrino spectrum, which relaxes the stability bound and ensures that all the RH neutrinos remain in equilibrium after the {PQ SSB.}
\section{Future perspectives}
{
SMASH provides very clear predictions, which will be tested by the next generation of CMB, large scale structure and axion DM experiments. The model predicts a correlation between $r$, $n_s$ and a small negative value of $\alpha$, as well as tiny non-Gaussianities. It also implies the existence of a cosmic background of relativistic axions which may be detected with future CMB polarisation experiments. In SMASH, the totality of the DM in the Universe is made of cold axions with mass in the range \eq{classic_window_m}, which will be explored in the next decade. If all these features are met simultaneously, {it will be a very compelling hint in favor of SMASH}. If only one is not, the model will be ruled out. {We recall that the cosmological predictions of SMASH are reliable; as opposed to those of incomplete models such as Higgs inflation, which suffers from an early breaking of perturbative unitarity.}
SMASH provides an explanation for five of the most pressing problems in particle physics and cosmology: inflation, DM, baryogenesis, the strong CP problem and the smallness of neutrino masses; some of which are naturalness issues. However, the model does not solve the hierarchy problem nor the cosmological constant problem. It would be interesting to explore if e.g.\ some relaxation mechanism along the lines of \cite{Abbott:1984qf,Graham:2015cka,Alberte:2016izw,Arvanitaki:2016xds} could be embedded in SMASH to solve also these problems while maintaining its minimality.
}\\
\mysections{Acknowledgments}
We thank F.\ Bezrukov, A.\ G.\ Dias, J.\ R.\ Espinosa, D.\ Figueroa, {F.\ Finelli,} J.\ Garcia-Bellido, {J.\ Jaeckel, F.\ Kahlh\"ofer, B.\ Kniehl, J.\ Lesgourgues,} K.\ Saikawa, M.\ Shaposhnikov, B.\ Shuve, S.\ Sibiryakov and A.\ Westphal for discussions. The work of G.B.\ is funded by the European Union's Horizon 2020 research and innovation programme under the Marie Sk\l{}odowska-Curie grant agreement number 656794 and was partially supported by the German Science Foundation (DFG) within the Collaborative Research Center SFB 676 “Particles, Strings and the Early Universe.” G.B. thanks the DESY Theory Group and the CERN Theory Department for hospitality.
J.R. is supported by the Ramon y Cajal Fellowship 2012-10597 and FPA2015-
65745-P (MINECO/FEDER). G.B. and C.T. thank the Mainz Institute for Theoretical Physics for hosting them during a workshop. C.T. thanks MIAPP for hospitality while attending a programme.
|
\section{Introduction}
\label{sec:intro}
There have been many phenomenological studies of the minimal
supersymmetric (SUSY) extension of the Standard Model (MSSM) that assume some
degree of universality for the soft supersymmetry-breaking scalar and
gaugino masses, $m_0$ and $m_{1/2}$, and the trilinear soft
supersymmetry-breaking parameters $A_0$. Scenarios in which these
parameters are universal at the supersymmetric grand unification (GUT)
scale, $M_{GUT}$, called the constrained MSSM (CMSSM) \cite{funnel,
cmssm, efgo, cmssmwmap, eo6, ehow+}, have been particularly intensively
studied, usually assuming that the lightest supersymmetric particle
(LSP) is a neutralino, which is stable because of the conservation of
$R$-parity \cite{EHNOS}, and provides (all or some of) the cosmological cold
dark matter. These and other GUT-universal models are under strong
pressure from LHC data \cite{eo6, ehow+, mc75, mc8, mc9, ELOS, post-mh,
fp, eoz, elo, yal, Buchmueller:2015uqa, Ellis:2015rya}, in particular, the notable
absence of missing transverse energy signals at the LHC
\cite{ATLAS20, CMS20}, with the measurement of the Higgs mass \cite{lhch, 125}, $m_h$,
providing an additional important constraint.
Fewer studies have been performed for scenarios in which the soft
supersymmetry-breaking parameters are universal at some other scale
$M_{in} \ne M_{GUT}$, which might be either below the GUT scale
(so-called sub-GUT or GUT-less scenarios \cite{subGUT,ELOS,Ellis:2015rya}) or above the GUT
scale (so-called super-GUT scenarios \cite{superGUT,emo,dlmmo}). For
example, in our current state of confusion about the possible mechanism
of supersymmetry breaking, and specifically in the absence of a
convincing dynamical origin at $M_{GUT}$, one could well imagine that
the universality scale $M_{in}$ might lie closer to the Planck or string
scale: $M_{in} > M_{GUT}$.
When studying such super-GUT scenarios, there appear additional
ambiguities beyond those in the conventional CMSSM. What is $M_{in}$?
Which GUT model to study? What are its additional parameters? How much
additional freedom do they introduce? In parallel, once one commits to a
specific GUT model, one must also consider the constraint imposed by the
absence (so far) of proton decay \cite{Takhistov:2016eqm}. In order to
minimize the ambiguities and the number of additional GUT parameters,
we study here the minimal supersymmetric SU(5) GUT
\cite{Dimopoulos:1981zb}.
It is well known that the length of the proton lifetime
is a significant challenge for this model \cite{Goto:1998qg, mp}, and one of the
principal new ingredients in this paper, compared to previous studies of
super-GUT CMSSM models, is the incorporation of this constraint in our
exploration of the model parameter space. Another improvement on
previous super-GUT CMSSM studies is the incorporation of LHC constraints,
of which the measurement of the Higgs mass turns out to be the most
relevant.
We find regions of the soft supersymmetry-breaking parameters $m_0$,
$m_{1/2}$, $A_0$ and the unknown coefficients in the SU(5)
superpotential that are compatible with these and other phenomenological
constraints such as the density of cold dark matter. As usual, we assume
that this is provided by the LSP, which we assume to be the lightest
neutralino. The Higgs mass and proton lifetime constraints both favor
$m_0$ and $m_{1/2}$ in the multi-TeV region, and proton stability
favours a value $\lesssim 6$ for the ratio of supersymmetric Higgs
vacuum expectation values (VEVs), $\tan \beta$. The cosmological
constraint on the cold dark matter density typically favors narrow
strips of parameter space where coannihilation with the lighter stop
brings the LSP density into the cosmological range. All these
constraints can be reconciled for suitable values of the unknown SU(5)
superpotential couplings.
The layout of this paper is as follows. In Section~\ref{sec:supergut}
we review our set-up of the super-GUT CMSSM, with particular attention
to the model parameters and the matching to the relevant parameters
below the GUT scale. Section~\ref{pdecay} then reviews our treatment
of proton decay, paying particular attention to the potential implications
of unknown GUT-scale phases. Our results are presented and explained in
Section~\ref{sec:results}, and Section~\ref{sec:discussion} then
summarizes our conclusions. An Appendix reviews details of our
nucleon decay calculations.
\section{Super-GUT CMSSM Models}
\label{sec:supergut}
\subsection{Minimal SUSY SU(5)}
We first review briefly the minimal supersymmetric SU(5) GUT~\cite{Dimopoulos:1981zb},
specifying our notation. This model is the
simplest supersymmetric extension of the original SU(5) GUT model due to
Georgi and Glashow \cite{Georgi:1974sy}. In this model, the right-handed
down-type quark and left-handed lepton chiral superfields, $\overline{D}_i$
and ${L}_i$, respectively, reside in $\bf{\overline{5}}$
representations, ${\Phi}_i$, while the left-handed quark doublet, right-handed
up-type quark, and right-handed charged-lepton chiral superfields,
${Q}_i$, $\overline{U}_i$, and $\overline{E}_i$, respectively, are in
$\bf{10}$ representations, ${\Psi}_i$, where the index $i = 1,2,3$ denotes the
generations. The MSSM Higgs chiral superfields
$H_u$ and $H_d$ are embedded into ${\bf 5}$ and $\overline{\bf 5}$
representations, $H$ and $\overline{H}$, respectively, where they are accompanied by
the ${\bf 3}$ and $\overline{\bf 3}$ coloured Higgs superfields $H_C$ and $\overline{H}_C$, respectively.
The SU(5) GUT gauge symmetry is assumed to be spontaneously broken down to the Standard
Model (SM) gauge group by the vacuum expectation value (vev) of a ${\bf 24}$ chiral superfield,
$\Sigma \equiv \sqrt{2}\Sigma^A T^A$, where the $T^A$ ($A=1, \dots, 24$) are the
generators of SU(5) normalized so that ${\rm Tr}(T^A T^B) =
\delta_{AB}/2$. The renormalizable superpotential for this model is then
given by
\begin{align}
W_5 &= \mu_\Sigma {\rm Tr}\Sigma^2 + \frac{1}{6} \lambda^\prime {\rm
Tr} \Sigma^3 + \mu_H \overline{H} H + \lambda \overline{H} \Sigma H
\nonumber \\
&+ \left(h_{\bf 10}\right)_{ij} \epsilon_{\alpha\beta\gamma\delta\zeta}
\Psi_i^{\alpha\beta} \Psi^{\gamma\delta}_j H^\zeta +
\left(h_{\overline{\bf 5}}\right)_{ij} \Psi_i^{\alpha\beta} \Phi_{j \alpha}
\overline{H}_\beta ~,
\label{W5}
\end{align}
where Greek sub- and superscripts denote SU(5) indices, and $\epsilon$ is the totally
antisymmetric tensor with $\epsilon_{12345}=1$.
The adjoint Higgs $\Sigma$ is assumed to have a vev of the form
\begin{equation}
\langle \Sigma \rangle = V \cdot {\rm diag} \left(2,2,2,-3,-3\right) ~,
\end{equation}
where $V\equiv 4 \mu_\Sigma/\lambda^\prime$.
In this case, the GUT gauge bosons acquire masses $M_X = 5 g_5 V$,
where $g_5$ is the SU(5) gauge coupling. In order to realize the
doublet-triplet mass splitting in $H$ and $\overline{H}$, we need to
impose the fine-tuning condition $\mu_H -3\lambda V \ll V$, which we
discuss in Section~\ref{sec:matchingcond}. In this case, the masses of the
color and weak adjoint components of $\Sigma$ are equal to $M_\Sigma =
5\lambda^\prime V/2$, while the singlet component of $\Sigma$ acquires a
mass $M_{\Sigma_{24}} = \lambda^\prime V/2$. The color-triplet Higgs
states have masses $M_{H_C} = 5\lambda V$.
\subsection{Planck-scale suppressed higher-dimensional operators}
\label{eq:highdimops}
In supersymmetric GUTs, gauge-coupling unification predicts that the unification scale
is ${\cal O}(10^{16})$~GeV. Since the unification scale is
fairly close to the reduced Planck mass $M_P = 2.4\times 10^{18}$~GeV,
interactions of gravitational strength may give rise to sizable effects. We
accommodate these effects by considering higher-dimensional effective
operators suppressed by powers of $M_P$.
We may expect that such effective
operators play significant roles in the minimal SUSY SU(5) GUT. For example, in minimal SU(5) GUTs
fthe down-type Yukawa couplings are predicted to be equal to the
corresponding lepton Yukawa couplings at the GUT scale, since they both
originate from $h_{\overline{\bf 5}}$. Nevertheless, in most of the
parameter space we consider, this Yukawa unification is imperfect. For
the third generation, the deviation is typically at the ${\cal O}(10)$\%
level. For the first two generations, on the other hand, there are ${\cal
O}(1)$ differences. These less successful predictions can be rectified
if one considers the following dimension-five effective operators
that are suppressed by the Planck scale~\cite{nro, Bajc:2002pg}:
\begin{equation}
W_{\rm eff}^{\Delta h} = \frac{c^{\Delta h}_{ij}}{M_P} \Phi_{i \alpha}
\Sigma^\alpha_{~\beta} \Psi^{\beta\gamma} \overline{H}_\gamma ~.
\label{eq:weffdelh}
\end{equation}
These operators induce non-universal contributions to the effective Yukawa couplings that are ${\cal O}(V/M_P)$ after the
adjoint Higgs acquires a VEV~\footnote{There is another class of
dimension-five operators of the form $\Psi_i^{\alpha\beta} \Phi_{j\alpha}
\Sigma^\gamma_{~\beta} \overline{H}_\gamma$. However, they do not spoil
Yukawa unification, but only modify the overall sizes of the down-type
quark and charged-lepton Yukawa couplings by ${\cal O}(V/M_P)$.},
which is sufficient to account for the observed deviations~\footnote{One
may also use higher-dimensional Higgs representations to explain the
observed differences between down-type and lepton Yukawa
couplings \cite{GM}. However, in this paper we focus on the minimal SU(5) GUT,
and do not consider this alternative.}.
There are several other dimension-five operators that one may consider. Among them is
\begin{equation}
W_{\rm eff}^{\Delta g} = \frac{c}{M_P} {\rm Tr}\left[
\Sigma {\cal W} {\cal W}
\right] ~,
\label{eq:SigmaWW}
\end{equation}
where ${\cal W}\equiv T^A {\cal W}^A$ denotes the
superfields corresponding to the field strengths of the SU(5) gauge vector bosons
${\cal V} \equiv {\cal V}^A T^A$. The term (\ref{eq:SigmaWW}) can have a significant effect, since it
changes the matching conditions of the gauge coupling constants
after $\Sigma$ develops a VEV \cite{Ellis:1985jn, Hill:1983xh,
Tobe:2003yj}. This operator also modifies the matching conditions for
gaugino masses, thereby modifying gaugino mass unification
\cite{Ellis:1985jn, Tobe:2003yj, Anderson:1996bg}. We discuss these
effects in detail in Section~\ref{sec:matchingcond}.
We may also have terms of the form~\cite{Bajc:2002pg}
\begin{equation}
W_{\rm eff}^\Sigma = \frac{a}{M_P} \left({\rm Tr} \Sigma^2\right)^2
+\frac{b}{M_P} {\rm Tr} \Sigma^4 ~.
\end{equation}
These operators can split the masses of the color and SU(2)$_L$ adjoint
components in $\Sigma$, $M_{\Sigma_8}$ and $M_{\Sigma_3}$ by ${\cal
O}(V^2/M_P)$. This mass difference induces threshold corrections to gauge
coupling constants of $\sim
\ln(M_{\Sigma_3}/M_{\Sigma_8})/(16\pi^2)$. This effect is negligible
for $\lambda^\prime \gg (a,b)V/M_P$ but could be significant for very small $\lambda^\prime$.
However, in order to simplify our analysis, we neglect the effects of these operators in this
paper.
\subsection{Soft supersymmetry-breaking mass parameters}
The soft supersymmetry-breaking terms in the minimal supersymmetric SU(5) GUT are
\begin{align}
{\cal L}_{\rm soft} = &- \left(m_{\bf 10}^2\right)_{ij}
\widetilde{\psi}_i^* \widetilde{\psi}_j
- \left(m_{\overline{\bf 5}}^2\right)_{ij} \widetilde{\phi}^*_i
\widetilde{\phi}_j
- m_H^2 |H|^2 -m_{\overline{H}}^2 |\overline{H}|^2 - m_\Sigma^2 {\rm Tr}
\left(\Sigma^\dagger \Sigma\right)
\nonumber \\
&-\biggl[
\frac{1}{2}M_5 \widetilde{\lambda}^{A} \widetilde{\lambda}^A
+ A_{\bf 10} \left(h_{\bf 10}\right)_{ij}
\epsilon_{\alpha\beta\gamma\delta\zeta} \widetilde{\psi}_i^{\alpha\beta}
\widetilde{\psi}^{\gamma\delta}_j H^\zeta
+ A_{\overline{\bf 5}}\left(h_{\overline{\bf 5}}\right)_{ij}
\widetilde{\psi}_i^{\alpha\beta} \widetilde{\phi}_{j \alpha} \overline{H}_\beta
\nonumber \\
&+ B_\Sigma \mu_\Sigma {\rm Tr} \Sigma^2 +\frac{1}{6} A_{\lambda^\prime
} \lambda^\prime {\rm Tr} \Sigma^3 +B_H \mu_H \overline{H} H+
A_\lambda \lambda \overline{H} \Sigma H +{\rm h.c.}
\biggr]~,
\end{align}
where $\widetilde{\psi}_i$ and $\widetilde{\phi}_i$ are the scalar
components of $\Psi_i$ and $\Phi_i$, respectively,
the $\widetilde{\lambda}^A$ are the SU(5) gauginos, and we
use the same symbols for the scalar components of the Higgs fields as for the
corresponding superfields.
In the super-GUT CMSSM model, we impose the following universality
conditions for the soft-mass parameters at a soft supersymmetry-breaking mass input scale
$M_{in} > M_{\rm GUT}$:
\begin{align}
\left(m_{\bf 10}^2\right)_{ij} =
\left(m_{\overline{\bf 5}}^2\right)_{ij}
&\equiv m_0^2 \, \delta_{ij} ~,
\nonumber \\[3pt]
m_H = m_{\overline{H}} = m_\Sigma &\equiv m_0 ~,
\nonumber \\[3pt]
A_{\bf 10} = A_{\overline{\bf 5}} = A_\lambda = A_{\lambda^\prime}
&\equiv A_0 ~,
\nonumber \\[3pt]
M_5 &\equiv m_{1/2} ~.
\label{eq:inputcond}
\end{align}
The bilinear soft SUSY-breaking therms $B_\Sigma$ and $B_H$ are determined from the other
parameters, as we shall see in the following. Note that, if we set $M_{in} =
M_{GUT}$, the above conditions are equivalent to those in the CMSSM.
These parameters are evolved down to $M_{GUT}$ using the
renormalization-group equations (RGEs) of the minimal supersymmetric SU(5) GUT,
which can be found in~\cite{pp,Baer:2000gf,emo}, with appropriate changes
of notation. During the evolution, the GUT parameters in Eq.~\eqref{W5}
affect the running of the soft supersymmetry-breaking parameters, which results
in non-universality in the soft parameters at $M_{GUT}$. In particular,
the $\lambda$ coupling enters into the RGEs for the soft masses of the
${\bf 5}$ and $\overline{\bf 5}$ Higgs fields, and can have
significant effects on their evolution. These effects become
particularly important in the vicinity of the focus-point region at
large $m_0$, since it is very close to the boundary of consistent
electroweak symmetry breaking (EWSB). In addition, $\lambda$ contributes
to the running of the Yukawa couplings and the corresponding $A$-terms.
On the other hand,
$\lambda^\prime$ affects directly only the running of $\lambda$,
$m_\Sigma$, and $A_\lambda$ (besides $\lambda^\prime$ and
$A_{\lambda^\prime}$), and thus can affect the MSSM soft mass parameters
only at higher-loop level. Both of $\lambda$ and
$\lambda^\prime $ contribute to the RGEs of the soft masses of
matter multiplets only at higher-loop level, and thus their effects on
these parameters are rather small. Thus, the low-energy phenomenology is rather
insensitive to the value of $\lambda^\prime$. The $\mu$ parameters $\mu_\Sigma$ and
$\mu_H$, as well as the corresponding bilinear parameters $B_\Sigma$ and
$B_H$, do not enter into RGEs of the rest of the parameters, and thus
their values give no effects on the running of the parameters in
Eq.~\eqref{eq:inputcond}. We note in passing that, if we set $M_{in} =
M_{GUT}$, we obtain the CMSSM and there is no effect from the running
above the GUT scale on the low-energy spectrum \footnote{However, we find
that the GUT-scale matching condition on the $B$ parameter gives a
constraint on the model parameter space even though $M_{in} = M_{GUT}$,
as we see below. }.
\subsection{GUT-scale matching conditions}
\label{sec:matchingcond}
At the unification scale $M_{GUT}$, the SU(5) GUT parameters are matched
onto the MSSM parameters. In this Section, we summarize these matching conditions and
discuss the constraints on the parameters from the low-energy
observables.
The matching conditions for the Standard Model gauge couplings at one-loop level in
the $\overline{\rm DR}$ scheme are given by
\begin{align}
\frac{1}{g_1^2(Q)}&=\frac{1}{g_5^2(Q)}+\frac{1}{8\pi^2}\biggl[
\frac{2}{5}
\ln \frac{Q}{M_{H_C}}-10\ln\frac{Q}{M_X}
\biggr]+\frac{8cV}{M_P} (-1)
~, \label{eq:matchg1} \\
\frac{1}{g_2^2(Q)}&=\frac{1}{g_5^2(Q)}+\frac{1}{8\pi^2}\biggl[
2\ln \frac{Q}{M_\Sigma}-6\ln\frac{Q}{M_X}
\biggr]+\frac{8cV}{M_P} (-3)
~, \\
\frac{1}{g_3^2(Q)}&=\frac{1}{g_5^2(Q)}+\frac{1}{8\pi^2}\biggl[
\ln \frac{Q}{M_{H_C}}+3\ln \frac{Q}{M_\Sigma}-4\ln\frac{Q}{M_X}
\biggr]+\frac{8cV}{M_P} (2)~,
\end{align}
where $g_1$, $g_2$, and $g_3$ are the U(1), SU(2), and SU(3) gauge
couplings, respectively, and $Q$ is a renormalization scale taken in our analysis to be
the unification scale: $Q = M_{GUT}$. The last terms in
these equations represent the contribution of the dimension-five
operator \eqref{eq:SigmaWW}. Since $V/M_P \simeq 10^{-2}$, these terms
can be comparable to the one-loop threshold corrections, and thus should
be taken into account when discussing gauge-coupling unification~\cite{Tobe:2003yj}.
From these equations, we have
\begin{align}
\frac{3}{g_2^2(Q)} - \frac{2}{g_3^2(Q)} -\frac{1}{g_1^2(Q)}
&=-\frac{3}{10\pi^2} \ln \left(\frac{Q}{M_{H_C}}\right)
-\frac{96cV}{M_P}
~,\label{eq:matchmhc} \\[3pt]
\frac{5}{g_1^2(Q)} -\frac{3}{g_2^2(Q)} -\frac{2}{g_3^2(Q)}
&= -\frac{3}{2\pi^2}\ln\left(\frac{Q^3}{M_X^2 M_\Sigma}\right) ~,
\label{eq:matchmgut}
\\[3pt]
\frac{5}{g_1^2(Q)} +\frac{3}{g_2^2(Q)} -\frac{2}{g_3^2(Q)}&= -\frac{15}{2\pi^2} \ln\left(\frac{Q}{M_X}\right) + \frac{6}{g_5^2(Q)} -\frac{144cV}{M_P} ~,\label{eq:matchg5}
\end{align}
We note that there is no contribution to (\ref{eq:matchmgut}) from the
dimension-five operator~\footnote{This feature can be understood as
follows. The contributions of the color-triplet Higgs multiplets to the
gauge coupling beta functions are given by $(b_1^{H_C}, b_2^{H_C},
b_3^{H_C}) =(2/5, 0, 1)$. In this notation, the matching conditions may be rewritten as
\begin{equation}
\frac{1}{g_i^2 (Q)} = \frac{1}{g_5^2(Q)} +\frac{1}{8\pi^2}
\left[b_i^{H_C} \ln\left(\frac{Q}{M_{H_C}}\right)+ \dots\right]
+\frac{8cV}{M_{P}}\left(-3 + 5b_i^{H_C}\right) ~.
\end{equation}
Since $5b_1^{H_C} -3b_{2}^{H_C}-2b^{H_C}_3 = 0$ and $5-3-2=0$,
neither $\ln (M_{H_C})$ nor $V/M_P$ appears in (\ref{eq:matchmgut}).}.
By running the gauge couplings up from their low-energy values, we can
determine the combination $M_X^2 M_\Sigma$ via
(\ref{eq:matchmgut})~\cite{Hisano:1992mh, Hisano:1992jj,
Hisano:2013cqa}. Notice that without the dimension-five operator ($c=0$),
$M_{H_C}$ is also determined from the values of the gauge couplings at
the GUT scale via Eq.~\eqref{eq:matchmhc}. The contribution of this
operator relaxes this constraint, and allows us to regard $M_{H_C}$ as a
free parameter. The last matching condition, Eq. (\ref{eq:matchg5}), will be used to determine $g_5$ and $M_{H_C}$ as will be discussed below.
For the Yukawa couplings, we use the tree-level matching
conditions. However, we note here that there is an ambiguity in the determination
of the GUT Yukawa couplings. As we mentioned in
Section~\ref{eq:highdimops}, Yukawa unification in the MSSM is imperfect in
most of the parameter space. Although this is cured by the
higher-dimensional operators in (\ref{eq:weffdelh}), they introduce
additional contributions to the matching conditions for the Yukawa
couplings. With this in mind, in this paper, we use
\begin{equation}
h_{{\bf 10}, 3} = \frac{1}{4}f_{u_3}~, ~~~~~~
h_{\overline{\bf 5}, 3} = \frac{f_{d_3} + f_{e_3}}{\sqrt{2}} ~,
\end{equation}
for the third-generation Yukawa couplings,
where $h_{{\bf 10}, i}$, $h_{\overline{\bf 5}, i}$, $f_{u_i}$,
$f_{d_i}$, and $f_{e_i}$ are eigenvalues of $h_{\bf 10}$,
$h_{\overline{\bf 5}}$, the MSSM up-type Yukawa couplings, the MSSM
down-type Yukawa couplings, and the MSSM lepton Yukawa couplings,
respectively. This condition is the same as that used in
Ref.~\cite{emo}. For the first- and second-generation Yukawa couplings,
on the other hand, we use
\begin{equation}
h_{{\bf 10}, i} = \frac{1}{4} f_{u_i} ~, ~~~~~~
h_{\overline{\bf 5}, i} = \sqrt{2} f_{d_i} ~.
\end{equation}
We chose the down-type Yukawa
couplings for the $h_{\overline{\bf 5}}$ matching condition, rather than
the lepton Yukawa couplings, since it results in longer proton decay lifetimes
and thus gives a conservative bounds on the model parameter space
\cite{Ellis:2015rya, evno}.
Next we obtain the matching conditions for the soft supersymmetry-breaking terms. To
this end, we first note that in the presence of soft supersymmetry-breaking terms
the VEV of $\Sigma$ deviates from $V$ by ${\cal O}(M_{\rm SUSY})$, where
$M_{\rm SUSY}$ denotes the supersymmetry-breaking scale \cite{Hall:1983iz}. In addition, $\langle
\Sigma \rangle$ develops a non-vanishing $F$-term. We find that
\begin{equation}
\langle \Sigma \rangle =
\left[
V + \frac{V(A_{\lambda^\prime} - B_\Sigma)}{2\mu_\Sigma}
+ F_\Sigma\, \theta^2
\right]\cdot {\rm diag}(2,2,2,-3,-3) ~,
\end{equation}
where
\begin{equation}
F_\Sigma = V(A_{\lambda^\prime} - B_\Sigma)
+\frac{V}{2\mu_\Sigma}
\left[
B_\Sigma(A_{\lambda^\prime} -B_\Sigma)
-m_\Sigma^2
\right] + {\cal O}(M_{\rm SUSY}^3/M_{GUT})~.
\end{equation}
Using this result, we obtain the following matching conditions for the gaugino
masses \cite{Tobe:2003yj, Hisano:1993zu}:
\begin{align}
M_1 &= \frac{g_1^2}{g_5^2} M_5
-\frac{g_1^2}{16\pi^2}\left[10 M_5 +10(A_{\lambda^\prime} -B_\Sigma)
+\frac{2}{5}B_H\right]
+\frac{4cg_1^2V(A_{\lambda^\prime} -B_\Sigma)}{M_P} ~,
\label{eq:m1match}
\\[3pt]
M_2 &= \frac{g_2^2}{g_5^2} M_5
-\frac{g_2^2}{16\pi^2}\left[6 M_5 +6A_{\lambda^\prime} -4B_\Sigma
\right]
+\frac{12cg_2^2V(A_{\lambda^\prime} -B_\Sigma)}{M_P} ~,
\label{eq:m2match}
\\[3pt]
M_3 &= \frac{g_3^2}{g_5^2} M_5
-\frac{g_3^2}{16\pi^2}\left[4 M_5 +4A_{\lambda^\prime} -B_\Sigma
+B_H \right]
-\frac{8cg_3^2V(A_{\lambda^\prime} -B_\Sigma)}{M_P}
~.
\label{eq:m3match}
\end{align}
We again find that the contribution of the dimension-five operator
can be comparable to that of the one-loop threshold corrections.
The soft masses of the MSSM matter fields, as well as the $A$-terms of
the third-generation sfermions, are given by
\begin{align}
m^2_{Q} = m_{U}^2 = m^2_{E} = m^2_{{\bf 10}} ~,&
~~~~~~ m_{D}^2 = m_{L}^2 = m_{\overline{\bf 5}}^2 ~, \nonumber \\
m_{H_u}^2 = m_H^2 ~,& ~~~~~~ m_{H_d}^2 = m_{\overline{H}}^2 ~,
\nonumber \\
A_t = A_{\bf 10} ~,& ~~~~~~
A_b = A_\tau = A_{\overline{\bf 5}} ~.
\end{align}
Finally, for the $\mu$ and $B$ terms we have \cite{Borzumati:2009hu}
\begin{align}
\mu &= \mu_H - 3 \lambda V\left[
1+ \frac{A_{\lambda^\prime} -B_\Sigma}{2 \mu_\Sigma}
\right] ~,
\label{eq:matchingmu}
\\[3pt]
B &= B_H + \frac{3\lambda V \Delta}{\mu}
+ \frac{6 \lambda}{\lambda^\prime \mu} \left[
(A_{\lambda^\prime} -B_\Sigma) (2 B_\Sigma -A_{\lambda^\prime}
+\Delta) -m_\Sigma^2
\right]~,
\label{eq:matchingb}
\end{align}
with
\begin{equation}
\Delta \equiv A_{\lambda^\prime} - B_\Sigma - A_\lambda +B_H ~.
\label{eq:deltadef}
\end{equation}
These equations display the amount of fine-tuning required to obtain
values of $\mu$ and $B$ that are ${\cal O}(M_{\rm SUSY})$. Equation
\eqref{eq:matchingmu} shows that we need to tune $|\mu_H -3\lambda V|$
to be ${\cal O}(M_{\rm SUSY})$. On the other hand,
Eq.~\eqref{eq:matchingb} indicates that $V\Delta/\mu$ should be ${\cal
O}(M_{\rm SUSY})$, which requires $|\Delta| \leq {\cal O}(M_{\rm
SUSY}^2/M_{GUT})$. Therefore, we can neglect $\Delta$ in the following
calculations. Notice that the condition $\Delta = 0$ is stable against radiative corrections
as shown in Ref.~\cite{Kawamura:1994ys}.
The $\mu$ and $B$ parameters are determined by using the electroweak
vacuum conditions:
\begin{align}
\mu^2 &= \frac{m_{1}^2 -m_{2}^2 \tan^2\beta + \frac{1}{2} m_Z^2
(1-\tan^2\beta ) +\Delta_\mu^{(1)}}{\tan^2 \beta -1 +\Delta_\mu^{(2)}}
, \\
B\mu &= -\frac{1}{2}(m_1^2 + m_2^2 +2 \mu^2) \sin 2\beta +\Delta_B ~,
\end{align}
where $\Delta_B$ and $\Delta_\mu^{(1,2)}$ denote loop corrections
\cite{Barger:1993gh}.
We can determine the $B$ parameters in minimal SU(5)
by solving the conditions \eqref{eq:matchingb} and $\Delta = 0$
\footnote{We need to determine the $B$ parameters in order to obtain
the MSSM gaugino masses via Eqs.~(\ref{eq:m1match}--\ref{eq:m3match}).}.
However, we
find that there is an additional condition that must be satisfied in order for
these equations to be solvable. When eliminating $B_H$ from
Eq.~\eqref{eq:matchingb} using $\Delta = 0$, we obtain an
equation that is quadratic in $B_\Sigma$. This equation has a
real solution only if
\begin{equation}
A_{\lambda^\prime}^2 -\frac{\lambda^\prime \mu}{3\lambda}
\left(A_{\lambda^\prime} -4 A_\lambda + 4B\right)
+\left(\frac{\lambda^\prime \mu}{6\lambda}\right)^2
\geq 8 m_{\Sigma}^2 ~.
\label{alimit}
\end{equation}
This condition gives a non-trivial constraint on the input parameters,
especially on the trilinear coupling $A_0$. In particular, for
$\lambda^\prime \ll \lambda$, this constraint leads to
$A_{\lambda^\prime}^2 \simeq A_0^2 \geq 8 m_{\Sigma}^2 \simeq 8m_0^2$.
When we compute the proton lifetime, we need to evaluate the
color-triplet Higgs mass $M_{H_C}$. This can be done by using
Eqs.~\eqref{eq:matchmhc}, \eqref{eq:matchmgut}, and \eqref{eq:matchg5} together with
\begin{align}
M_{H_C} &= 5\lambda V ~,\label{eq:MHCV} \\
M_\Sigma &= \frac{5}{2} \lambda^\prime V ~, \\
M_X &= 5 g_5 V ~.\label{eq:MXV}
\end{align}
From these equations, we obtain
\begin{equation}
M_{H_C} = \lambda \left(\frac{2}{\lambda^\prime
g_5^2}\right)^{\frac{1}{3}}
\left(M_X^2 M_\Sigma\right)^{\frac{1}{3}} ~.
\label{eq:mhc}
\end{equation}
We can then determine $M_X^2 M_\Sigma$ using Eq.~\eqref{eq:matchmgut}. Eq. (\ref{eq:matchg5}) can be reduced to an equation with undetermined parameters $g_5$ and $M_{H_C}$ using Eq. (\ref{eq:MHCV}) and (\ref{eq:MXV}). Then once $\lambda$ and $\lambda^\prime$ are chosen, this equation plus Eq.~\eqref{eq:mhc} can be used
to determine $M_{H_C}$ and $g_5$. However, since $g_5$ is only logarithmical dependent on $M_{H_C}$, it will remain fairly constant for a broad range of $M_{H_C}$. As mentioned above, if we do not include the
contribution of the dimension-five operator, Eq.~\eqref{eq:matchmhc}
fixes $M_{H_C}$. In this case, $\lambda$ and $\lambda^\prime$ are restricted via
Eq.~\eqref{eq:mhc}, and thus we cannot regard both of them as free
parameters. The last term in Eq.~\eqref{eq:matchmhc} can relax this
restriction, and enables us to take $\lambda$ and $\lambda^\prime$ as
input parameters. In this case, $M_{H_C}$ is given by
Eq.~\eqref{eq:mhc}, and Eq.~\eqref{eq:matchmhc} determines the parameter
$c$. In the following analysis, we check that the coefficient $c$
has reasonable values, {\it i.e.}, $|c| < {\cal O} (1)$.
Using the above results, we see how the super-GUT CMSSM model is
specified by the following set of input parameters:
\begin{equation}
m_0,\ m_{1/2},\ A_0,\ M_{in},\ \lambda,\ \lambda',\ \tan \beta,\ {\rm
sign}(\mu) \, ,
\end{equation}
where the trilinear superpotential Higgs couplings, $ \lambda,\ \lambda'$,
are specified at $Q=M_{GUT}$.
\section{Proton Decay and GUT-Scale Phases}
\label{pdecay}
As is well known, in the minimal supersymmetric SU(5) GUT with weak-scale
supersymmetry breaking, the dominant decay channel
of proton is the $p\to K^+ \overline{\nu}$ mode \cite{Sakai:1981pk},
which is induced by the exchange of the color-triplet Higgs
multiplets, and the model is severely restricted by the proton decay
bound \cite{Goto:1998qg, mp}. The exchange of the GUT-scale gauge bosons can also induce
proton decay, but this contribution is usually subdominant because of the large
GUT scale in supersymmetric GUTs. The strong constraint from the
$p\to K^+ \overline{\nu}$ decay may, however, be evaded if the masses of
supersymmetric particles are well above the electroweak scale
\cite{Hisano:2013exa, McKeen:2013dma, Nagata:2013sba, evno,
Ellis:2015rya}. In addition, it turns out that the $p\to K^+
\overline{\nu}$ decay mode depends sensitively on the extra phases in the GUT
Yukawa couplings \cite{Ellis:1979hy}, which can suppress the
proton decay rate, as we discuss in this Section.
For more details of the proton decay calculation, see
Refs.~\cite{Hisano:2013exa, Nagata:2013sba, evno, Ellis:2015rya} and
the Appendix.
In supersymmetric models, the largest contribution to the decay rate of
the proton is determined by the dimension-five effective operators
generated by integrating out the colored Higgs multiplets
\cite{Sakai:1981pk},
\begin{equation}
{\cal L}_5^{\rm eff}= C^{ijkl}_{5L}{\cal O}^{5L}_{ijkl}
+C^{ijkl}_{5R}{\cal O}^{5R}_{ijkl}
+{\rm h.c.}~,
\label{eq:efflaggut}
\end{equation}
with ${\cal O}^{5L}_{ijkl}$ and ${\cal
O}^{5R}_{ijkl}$ defined by
\begin{align}
{\cal O}^{5L}_{ijkl}&\equiv\int d^2\theta~ \frac{1}{2}\epsilon_{abc}
(Q^a_i\cdot Q^b_j)(Q_k^c\cdot L_l)~,\nonumber \\
{\cal O}^{5R}_{ijkl}&\equiv\int d^2\theta~
\epsilon^{abc}\overline{u}_{ia}\overline{e}_j\overline{u}_{kb}
\overline{d}_{lc}~,
\end{align}
where $i,j,k,l$ are generation indices, $a,b,c$ are SU(3)$_C$ color
indices, and $\epsilon_{abc}$ is the totally antisymmetric three-index tensor.
The Wilson coefficients are given by
\begin{align}
C^{ijkl}_{5L}(M_{GUT})&
=\frac{2\sqrt{2}}{M_{H_C}}h_{{\bf 10}, i}
e^{i\phi_i}\delta^{ij}V^*_{kl}h_{\overline{\bf 5},l}~,\nonumber \\
C^{ijkl}_{5R}(M_{GUT})
&=\frac{2\sqrt{2}}{M_{H_C}}h_{{\bf 10}, i}V_{ij}V^*_{kl}h_{\overline{\bf 5},l}
e^{-i\phi_k}
~,
\label{eq:wilson5}
\end{align}
where $V_{ij}$ are the familiar CKM matrix elements, and the $\phi_i$ ($i =1,2,3$) are
the new CP- violating phases in the GUT Yukawa couplings. These are subject to the
constraint $\phi_1 + \phi_2 + \phi_3 = 0$, so there are two independent
degrees of freedom for these new CP-violating phases \cite{Ellis:1979hy}~\footnote{The
number of extra degrees of freedom in the GUT Yukawa couplings can be counted as
follows. Since $h_{\bf 10}$ is a $3\times 3$ symmetric complex matrix,
it has 12 real degrees of freedom, while $h_{\overline{\bf 5}}$ has
18. Field redefinitions of $\Psi_i$ and $\Phi_i$ span the ${\rm U}(3) \otimes
{\rm U}(3)$ transformation group, and thus 18 parameters are
unphysical. Hence, we have 12 physical parameters. Among them, 6 are specified by
quark masses, while 4 are for the CKM matrix elements. The remaining 2
are the extra CP phases, which we take to be $\phi_2$ and $\phi_3$.}.
We take $\phi_2$ and $\phi_3$ as free input parameters in the following
discussion. The coefficients in Eq.~\eqref{eq:wilson5} are then run
to the SUSY scale using the RGEs. At the SUSY scale, the sfermions associated with these Wilson
coefficients are integrated out through a loop containing either a wino
mass insertion or a Higgsino mass insertion, which are proportional to
$C_{5L}$ and $C_{5R}$, respectively. The wino contribution to the decay
amplitude for the $p\to K^+ \overline{\nu}_i$ mode is given by the sum
of the Wilson coefficients $C_{LL} (usd\nu_i)$ and $C_{LL} (uds\nu_i)$
multiplied by the corresponding matrix elements (see
Eq.~\eqref{eq:amplitudes}). These coefficients are approximated by
\begin{align}
C_{LL} (usd\nu_i) &= C_{LL} (uds\nu_i) \nonumber \\
&\simeq
\frac{2\alpha_2^2}{\sin 2\beta}\frac{m_t m_{d_i} M_2}{ m_W^2M_{H_C}M_{\rm
SUSY}^2} V_{ui}^*V_{td}V_{ts}e^{i\phi_3}\left(1 +
e^{i(\phi_2-\phi_3)}\frac{m_c V_{cd}V_{cs}}{m_tV_{td}V_{ts}}\right) ~,
\label{eq:cllaprox}
\end{align}
where $m_c$, $m_t$, $m_W$, and $m_{d_i}$ are the masses of the charm quark,
top quark, $W$ boson, and down-type quarks, respectively, and $\alpha_2 = g_2^2/4\pi$.
Since the ratio of Yukawa couplings and CKM matrix elements in the
parenthesis in Eq.~\eqref{eq:cllaprox} is ${\cal O}(1)$, this Wilson coefficient may be suppressed
for certain ranges of the phases.
On the other hand, the Higgsino exchange process contributes only to the
$p\to K^+ \overline{\nu}_\tau$ mode, and gives no contribution to
the $p\to K^+ \overline{\nu}_{e, \mu}$ modes. The relevant Wilson
coefficients for the $p\to K^+ \overline{\nu}_\tau$ mode are $C_{LL}
(usd\nu_\tau)$ and $C_{LL}
(uds\nu_\tau)$ in Eq.~\eqref{eq:cllaprox}, as well as $C_{RL}
(usd\nu_\tau)$ and $C_{RL} (uds \nu_\tau)$, which are approximately given by
\begin{align}
C_{RL} (usd\nu_\tau) &\simeq -
\frac{\alpha_2^2}{\sin^2 2\beta}\frac{m_t^2 m_s m_\tau \mu}{ m_W^4 M_{H_C}M_{\rm SUSY}^2}V_{tb}^*
V_{us} V_{td}e^{-i(\phi_2+\phi_3)} ~, \nonumber \\
C_{RL} (uds\nu_\tau) &\simeq -
\frac{\alpha_2^2}{\sin^2 2\beta}\frac{m_t^2 m_d m_\tau \mu}{m_W^4 M_{H_C} M_{\rm SUSY}^2}V_{tb}^*
V_{ud} V_{ts}e^{-i(\phi_2+\phi_3)} ~,
\label{eq:crlaprox}
\end{align}
where $m_d$, $m_s$, and $m_\tau$ are the masses of down quark, strange
quark, and tau lepton, respectively.
Contrary to the coefficients in Eq.~\eqref{eq:cllaprox}, the absolute
values of these coefficients do not change when the phases vary.
Equations \eqref{eq:cllaprox} and \eqref{eq:crlaprox} show that the
proton decay rate receives a $\tan \beta$ enhancement as well as
a suppression by the sfermion mass scale $M_{\rm SUSY}$. To evade
the proton decay bound, therefore, a small $\tan \beta$ and a high
supersymmetry-breaking scale are favored as shown in the subsequent
section. In addition, we note that the proton decay rate decreases as
$M_{H_C}$ is taken to be large. From Eq.~\eqref{eq:mhc}, we find
$M_{H_C} \propto \lambda/ (\lambda^\prime)^{\frac{1}{3}}$, and thus the
proton lifetime $\tau_p$ is proportional to
$\lambda^2/(\lambda^{\prime})^{\frac{2}{3}}$. This indicates that
larger $\lambda$ values and smaller $\lambda^\prime$ values help avoid
the proton decay bound.
\begin{figure}[t]
\begin{center}
\subfigure[${\cal A} (p\to K^+ \overline{\nu}_\tau)$]
{\includegraphics[clip, width = 0.46 \textwidth]{ampl.pdf}
\label{fig:ampl}}
\subfigure[$\tau (p \to K^+ \overline{\nu}_i)$]
{\includegraphics[clip, width = 0.51 \textwidth]{partlife.pdf}
\label{fig:partlife}}
\caption{\it (a): The absolute value of the contributions to the decay
amplitude of the $p\to K^+ \overline{\nu}_\tau$ channel as functions
of $\phi_2$. The red dashed, green dash-dotted and black solid lines
represent the absolute values of the wino, Higgsino, and total
contributions, respectively. (b): The phase dependences of the lifetimes
for the different $p\to K^+ \overline{\nu}$ decay modes. The green dash-dotted, blue dotted and red dashed
lines represent the first-, second-, and third-generation neutrino
decay modes, respectively, and the black solid line shows the total
lifetime. In both figures, we set $\phi_3 = 0$, and take the parameter
point indicated by the star $(\bigstar)$ in Fig.~\ref{fig:diskr1}. }
\label{fig:pknuphi2}
\end{center}
\end{figure}
To show the phase dependence of these contributions more clearly, we show in
Fig.~\ref{fig:ampl} each contribution to the decay amplitude of
the $p\to K^+ \overline{\nu}_\tau$ channel as a function of $\phi_2$
with $\phi_3$ fixed to be $\phi_3 = 0$. The red dashed, green
dash-dotted and black solid lines represent the absolute values of the
wino, Higgsino, and total contributions, respectively. We take the
parameter point indicated by the star $(\bigstar)$ in
Fig.~\ref{fig:diskr1} below. This figure shows that the wino contribution can
vary by almost an order of magnitude, while the size of the Higgsino
contribution remains constant. These contributions are comparable, and
thus a significant cancellation can occur. As a result, the total amplitude
varies by more than an order of magnitude. The wino contribution is
minimized at $\phi_2 \simeq 0.89\pi$, while the total amplitude is
minimized at $\phi_2 \simeq 0.44\pi$. This mismatch is due to the
Higgsino contribution.
In Fig.~\ref{fig:partlife} we show the phase dependence of the lifetime
of each $p\to K^+ \overline{\nu}$ decay mode with the same parameter set. The green dash-dotted,
blue dotted and red dashed lines represent the first-, second-, and
third-generation neutrino decay modes, respectively, while the black
solid line shows the total lifetime. We see that the lifetimes of the
$\overline{\nu}_e$ and $\overline{\nu}_\mu$ modes, which are induced
by wino exchange only, are maximized at $\phi_2 \simeq 0.89\pi$,
which deviates from the point where $\tau (p \to K^+
\overline{\nu}_\tau)$ is maximized. Due to this deviation, the phase dependence of
the total lifetime is much smaller than that of each partial lifetime,
but still it can change by an ${\cal O} (1)$ factor.
\begin{figure}[t]
\begin{center}
\subfigure[$\tau (p\to K^+ \overline{\nu})$]
{\includegraphics[clip, width = 0.46 \textwidth]{pknuphase.pdf}
\label{fig:pknuphase}}
\subfigure[Nucleon decay lifetimes]
{\includegraphics[clip, width = 0.51 \textwidth]{nucleon.pdf}
\label{fig:nucleon}}
\caption{\it (a): Contour plot for the proton decay lifetime in units of
$10^{35}$ years. The area within the 0.066 contour satisfies the current experimental bound.
The peak lifetime is denoted by $\spadesuit$. (b): Lifetimes of the nucleon decay modes as
functions of $\phi_2$. Calculated for for the reference point indicated by a star $(\bigstar)$
in Fig.~\protect\ref{fig:diskr1}.}
\label{fig:phasedep}
\end{center}
\end{figure}
In Fig.~\ref{fig:pknuphase}, we show a contour plot for the proton decay
lifetime in units of $10^{35}$ years in the $\phi_2$--$\phi_3$ plane, using
the same parameter set as in Fig.~\ref{fig:pknuphi2}. We find that
the proton lifetime exceeds the current experimental bound, $\tau (p\to
K^+ \overline{\nu}) > 6.6\times 10^{33}$~yrs \cite{Takhistov:2016eqm,
Abe:2014mwa}, in a significant area of the phase space shown by the contour labeled
0.066. The peak lifetime is marked in the upper part of the figure by a spade.
Although the $p\to K^+ \overline{\nu}$ modes may be suppressed for certain
values of the phases, other decay
modes that depend on the same phases are not suppressed in the same way. The other
decay modes that could restrict the parameter space are $p\to
\pi^+\overline{\nu}$ and $n\to \pi^0 \overline{\nu}$. The Wilson
coefficients for these proton decay modes are quite similar to those
that generate $p\to K^+\overline{\nu}$, and depend on exactly the same
combination of SUSY parameters. The differences in the calculations of their
lifetimes come from their different dependences on CKM matrix
elements. The $p\to \pi^+ \overline{\nu}$ and $n\to \pi^0
\overline{\nu}$ modes are suppressed relative to the $p\to K^+ \overline{\nu}$ modes
by off-diagonal components of the CKM matrix. Moreover, the experimental constraints on
these modes are weaker:
$\tau (p\to \pi^+ \overline{\nu}) > 3.9 \times 10^{32}$~yrs and $n\to \pi^0
\overline{\nu} > 1.1\times 10^{33}$~yrs \cite{Takhistov:2016eqm,
Abe:2013lua}, so these decay modes are less restrictive on the parameter
space. To ensure that these modes are not problematic, in
Fig.~\ref{fig:nucleon}, we show the lifetimes of these decay modes as
functions of $\phi_2$ for the same parameter set as in
Fig.~\ref{fig:pknuphi2}. We find that, although the $p\to \pi^+
\overline{\nu}$ mode can be dominant, it is still above the present
experimental limit. The $n\to \pi^0 \overline{\nu}$ is always
sub-dominant, and again exceeds the current bound. We also note that the
$p\to \pi^+ \overline{\nu}$ and $n\to \pi^0 \overline{\nu}$ modes
exhibit the same phase dependence, since they are related to each other
through isospin symmetry.
In the following analysis, we choose the CP-violating phases so as to maximize
the $p\to K^+ \overline{\nu}$ lifetime, thereby obtaining a conservative constraint on the
super-GUT model parameter space. Although not shown in the figures below, we have verified that each allowed point also meet the experimental constraint coming from $p\to \pi^+ \overline{\nu}$ and $n\to \pi^0 \overline{\nu}$.
\section{Results}
\label{sec:results}
To appreciate the effect of choosing $M_{in} > M_{GUT}$, we begin by
reviewing briefly some results for the CMSSM with $M_{in} = M_{GUT}$. We
note that we use here the {\tt FeynHiggs 2.11.3} code~\cite{FH} to
compute the Higgs mass. Previously we used {\tt FeynHiggs 2.10.0}, and
we note that due to a bug fix, the new version yields a significant
change in $m_h$ at large positive $A_0$~\footnote{Note that our sign
convention for $A_0$ is opposite that found in many public codes such as
{\tt SoftSusy}~\cite{softsusy}.}. A large value of $A_0/m_0$ is
necessary to obtain the correct relic density along the stop-coannihilation strip \cite{stop,eoz}, where the lighter stop and
neutralino LSP are nearly degenerate in mass. For $A_0/m_0 \gtrsim 2$,
we find that {\tt FeynHiggs 2.11.3} results in a $\simeq 1.5$~GeV drop in the
value of $m_h$ relative to the previous result, necessitating a
lower value of $A_0/m_0$. However, for $A_0/m_0 \lesssim 2$, the stop
strip is no longer present. On the other hand, the effect of updating
{\tt FeynHiggs} on $m_h$ at large negative $A_0/m_0$ is less
pronounced. We further note that our calculation of the proton lifetime
here is also updated with bug-fixes.
\subsection{CMSSM update}
\label{sec:cmssm}
In view of the proton lifetime constraint, which favours larger
sparticle masses, we consider here the possibilities that the correct
relic density of neutralino dark matter is obtained either in the
focus-point strip \cite{fp2,fp} or the stop-coannihilation strip
\cite{stop}, updating the results found in \cite{Ellis:2015rya}.
We use SSARD \cite{SSARD} to compute the particle mass spectrum, the
dark matter relic density, and proton lifetimes. The discussion of the
proton lifetime in Section~\ref{pdecay} motivates us to focus on
relatively small values of $\tan \beta$. For larger values of $\tan
\beta$, the proton lifetime becomes smaller than the current experimental bound,
and minimal supersymmetric SU(5) is not viable. For the CMSSM cases with
$M_{in} = M_{GUT}$, we have set $c=0$ and taken $M_{H_C}$ from
Eq.~\eqref{eq:matchmhc}.
In Fig.~\ref{fig:CMSSM}, we show four CMSSM $(m_{1/2},m_0)$ planes
displaying the focus-point (left) and stop-coannihilation (right) relic
density strips for the two choices of the sign of $\mu$. Higgs mass
contours are shown as red dot-dashed curves labeled by $m_h$ in GeV in
1~GeV intervals starting at 122~GeV. In the left panels, we choose $A_0
= 0$ \footnote{As we discussed in Section~\ref{sec:matchingcond}, if we
assume the minimal SU(5) GUT with the universality condition
\eqref{eq:inputcond}, then the $B$-term matching condition restricts
$A_0$ via Eq.~\eqref{alimit}. This constraint can, however, be evaded if
we relax the universality condition \eqref{eq:inputcond} (for $m_\Sigma$
in particular) or consider non-minimal Higgs content. With these
possibilities in mind, we do not take the condition \eqref{alimit} into
account in Section~\ref{sec:cmssm}, which allows the choice $A_0 = 0$.
} with $\mu > 0$ (top) and $\mu < 0$ (bottom). For this choice of $A_0$,
there is a relatively minor effect on $m_h$ due to the updated version
of {\tt FeynHiggs}. The light mauve shaded region in the parts of the
left panels with large $m_0/m_{1/2}$ are excluded because there are no
solutions to the EWSB conditions: along this boundary $\mu^2 = 0$. Just
below the regions where EWSB fails, there are narrow dark blue strips
where the relic density falls within the range determined by CMB and
other experiments~\cite{Planck15}~\footnote{Since the relic density of
dark matter is now determined quite accurately ($\Omega_\chi h^2 =
0.1193 \pm 0.0014$), for the purpose of visibility we display expanded
strips for which the relic density lies in the range $[0.06,
0.20]$.}. These strips are in the focus-point region~\cite{fp,fp2}. We
note also that the brown shaded regions in the portions of the panels
with low $m_0/m_{1/2}$ are excluded because there the LSP is the lighter
charged stau lepton. The planes also feature stau-coannihilation strips
\cite{stau} close to the boundaries of these brown shaded regions. They
extend to $m_{1/2} \simeq 1$~TeV, but are very difficult to see on the
scale of this plot, even with our enhancement of the relic density
range. There are also `thunderbolt'-shaped brown shaded bands at
intermediate $m_0/m_{1/2}$ where the chargino is the LSP. There are no
accompanying chargino-coannihilation strips, as at these multi-TeV
mass scales any such strip would lie within the shaded region and is
therefore excluded.
\begin{figure}[htb!]
\begin{minipage}{8in}
\includegraphics[height=3.in]{m0vM8aspnGr_5_0_0_f}
\hspace*{-0.17in}
\includegraphics[height=3.in]{m0vM8aspnGr_6_-42_0_1_fsef}
\hfill
\end{minipage}
\begin{minipage}{8in}
\includegraphics[height=3.in]{m0vM8aspnGr_5_0_1_f}
\hspace*{-0.17in}
\includegraphics[height=3.in]{m0vM8aspnGr_6_-42_1_1_fsef}
\end{minipage}
\caption{
{\it
Sample CMSSM $(m_{1/2}, m_0)$ planes showing the focus-point strip for
$\tan \beta = 5$ and $A_0 = 0$ (left) with $\mu > 0$ (upper) and $\mu <
0$ (lower), and the stop coannihilation strip with $\tan \beta = 6$ and
$A_0 = -4.2 \, m_0$ (right). In the light mauve shaded regions, it is
not possible to satisfy the electroweak symmetry breaking
conditions. In the brown shaded regions, the LSP is charged and/or
colored. The dark blue shaded strips show the areas where $0.06 <
\Omega_\chi h^2 < 0.2$ in the left panels and the further enlarged
range of $0.01 < \Omega_\chi h^2 < 2.0$ in the right panels. The red
dot-dashed contours indicate the Higgs mass, labeled in GeV, and the
solid black contours indicate the proton lifetime in units of
$10^{35}$~yrs. The bold solid black, blue, green, purple, and red lines
in each panel are current and future limits from the LHC at 8 TeV, $300$
and $3000$~fb$^{-1}$ at 14 TeV, 3000~fb$^{-1}$ with the HE-LHC at
33~TeV, and 3000~fb$^{-1}$ with the FCC-hh at 100~TeV, respectively,
taken from the analysis of \protect{\cite{Buchmueller:2015uqa}}.
}}
\label{fig:CMSSM}
\end{figure}
Contours of the proton lifetime calculated using down-type Yukawa couplings (see
the discussion in Section~\ref{sec:matchingcond}) are shown as solid
black curves that are labeled in units of $10^{35}$~yrs. The current
limit $\tau_p > 6.6 \times 10^{33}$~yrs \cite{Takhistov:2016eqm,
Abe:2014mwa} would exclude the entire area below the curve labeled
0.066. For the nominal value of $m_h = 125$~GeV, neglecting the
theoretical uncertainties in the calculation of $m_h$, we see that in
the upper left plane of Fig.~\ref{fig:CMSSM} the Higgs contour
intersects the focus-point region where $\tau_p \approx 5 \times
10^{33}$~yrs, very close to the experimental limit. Much of the
focus-point strip in this figure may be probed by future proton decay
experiments. Changing the sign of $\mu$ has almost no effect on the
proton lifetime, as seen in the lower left panel of
Fig.~\ref{fig:CMSSM}, but the calculated Higgs mass is smaller by $\sim
1$~GeV, which is less than the uncertainty in the {\tt FeynHiggs}
calculation of $m_h$.
In the right panels of Fig.~\ref{fig:CMSSM}, we have chosen large
negative $A_0/m_0 = -4.2$ and $\tan \beta = 6$, which allows a
sufficiently heavy Higgs and a viable stop strip. There are now brown
shaded regions in the upper left wedges of the planes where the stop is
the LSP (or tachyonic). Though it is barely visible, there is a stop
strip that tracks that boundary \footnote{In this case, and in the
super-GUT cases to follow, we have further extended the range on
$\Omega_\chi h^2$ to [0.01,2.0]. Otherwise the thickness of the strips
which are typically 10--50~GeV would be pixel thin for the range of
masses shown.}. Since we have taken an enhanced range for the relic
density the blue strip continues to the edge of the plot. In reality,
however, the stop strip ends~\cite{eoz} at the position marked by the
{\bf X} in the figure. We see that, for $\mu > 0$, the stop strip ends
when $m_h < 122$~GeV, whereas for $\mu < 0$ the strip ends when $m_h
\approx 123.5$~GeV, both of which are acceptable given the uncertainty
in the calculation of $m_h$. At the endpoint, which occurs at $(m_{1/2},
m_0) \simeq (5.2,8.8)$~TeV, the proton lifetime is approximately
$2\times 10^{34}$~yrs. Had we chosen a smaller value of $|A_0/m_0|$, the
stop strip would have extended to higher $m_h$. For example, for $\mu <
0$, the stop strip extends to 125~GeV for $A_0/m_0 = -3.5$ and the
endpoint is found at (5.1,11.3) TeV.
In all of the cases shown in Fig.~\ref{fig:CMSSM}, the favored parameter
regions predict the masses of supersymmetric particles to be in the
multi-TeV range. For example, as the gluino mass is $\simeq
2\times m_{1/2}$, it is expected to be as large as $\simeq 10$~TeV,
which is well above the LHC reach \cite{ATLAS20,
CMS20}. To see the current and future limits on the CMSSM parameter
space from the LHC and future hadron colliders such as the 33~TeV HE-LHC
option and the Future Circular Collider (FCC) \cite{Golling:2016gvc} which
aims at 100~TeV proton-proton collisions, we show the limits from LHC
at 8 TeV, and sensitivities with $300$ and $3000$~fb$^{-1}$ with the LHC at 14 TeV, 3000~fb$^{-1}$
with the HE-LHC at 33~TeV, and 3000~fb$^{-1}$ with the FCC-hh at 100~TeV as the
bold solid black, blue, green, purple, and red lines in each panel in
Fig.~\ref{fig:CMSSM}, respectively, following the analysis given in
Ref.~\cite{Buchmueller:2015uqa}. As we see, the parameter region in
which the proton decay bound is evaded is far beyond the reach of the
LHC, but may be probed at the 100 TeV collider. We further note that,
while the stop-coannihilation region shown may not be fully probed at 33 TeV,
the 100 TeV reach clearly extends beyond the stop endpoint marked by the {\bf X}.
On the other hand, the focus-point region is seen to extend beyond the 100 TeV reach.
\subsection{Super-GUT CMSSM}
As we discussed earlier, the super-GUT scenario introduces several new
parameters, making a complete analysis quite cumbersome. In addition to
the CMSSM parameters, we must specify the input universality scale
$M_{in}$ and the values of the two GUT couplings $\lambda$ and
$\lambda'$. In order to understand better the parameter space of the
super-GUT models, we begin by considering $(m_0,A_0/m_0)$ planes for
fixed $m_{1/2}, \tan \beta, \lambda$, and $\lambda'$ and several choices
of $M_{in}$, as shown in Fig.~\ref{fig:diskr1}.
\begin{figure}[htb!]
\begin{minipage}{8in}
\includegraphics[height=3.3in]{A0vm0pndiskr_6_4000_1_0_fsef}
\includegraphics[height=3.3in]{A0vm0pndiskr_6_4000_1_165_fsef}
\hfill
\end{minipage}
\begin{minipage}{8in}
\includegraphics[height=3.3in]{A0vm0pndiskr_6_4000_1_17_fsef}
\includegraphics[height=3.3in]{A0vm0pndiskr_6_4000_1_175_fsef}
\end{minipage}
\caption{
{\it
Super-GUT CMSSM $(m_{0}, A_0/m_0)$ planes for $\tan \beta = 6$ and $\mu
< 0$. The values of $M_{in}$ are $M_{GUT}$, $10^{16.5}$, $10^{17}$ and
$10^{17.5}$~GeV, as indicated. In each panel, we have fixed $m_{1/2} =
4$~TeV, $\lambda = 0.6$ and $\lambda^\prime = 0.0001$. In the light
mauve shaded regions, it is not possible to satisfy the matching
condition for $B$. In the brown shaded regions, the LSP is the
stop. The dark blue shaded regions show the areas where $0.01 <
\Omega_\chi h^2 < 2.0$. The red dot-dashed contours indicate the Higgs
mass, labeled in GeV, and the solid black contours indicate the proton
lifetime in units of $10^{35}$~yrs.
}}
\label{fig:diskr1}
\end{figure}
In Fig.~\ref{fig:diskr1}, we have fixed $\tan \beta = 6, m_{1/2} =
4$~TeV, $\lambda = 0.6$ and $\lambda^\prime = 10^{-4}$ with $\mu <
0$. We have chosen $M_{in} = M_{GUT}$, $10^{16.5}$, $10^{17}$ and
$10^{17.5}$~GeV in the upper left, upper right, lower left and lower
right panels, respectively. In each panel, the contours for $m_h$ and
$\tau_p$ are drawn using the same line styles as in the previous figure.
The brown shaded regions at large $m_0$ and $- A_0/m_0$ are excluded
because they contain a stop LSP (or tachyonic stop), and the stop relic
density strip tracks this boundary. Because $m_{1/2}$ is fixed, there is
no endpoint of the strip within the parameter ranges shown, and the
lightest neutralino is an acceptable LSP everywhere along the blue strip
(remembering that the thickness of the strip is exaggerated for clarity).
For $M_{in} > M_{GUT}$, there is a mauve shaded region at small $m_0$
and $- A_0/m_0$ that grows in size as $M_{in}$ is increased. In this
region, the $B$ matching condition \eqref{eq:matchingb} is violated, and
there is no solution to \eqref{alimit}~\footnote{For $M_{in} = M_{GUT}$,
the region excluded is $|A_0| \lesssim 2.8 m_0$, which is below the
range displayed in the Figure.}.
When $M_{in} = M_{GUT}$ with the parameters adopted in
Fig.~\ref{fig:diskr1}, the Higgs mass prefers smaller values of
$|A_0/m_0|$ and larger values of $m_0$. In the portion of the strip
where $m_h > 123$~GeV according to {\tt FeynHiggs} (which is consistent
with the experimental measurement), the proton lifetime is $> 10^{34}$
yrs. As $M_{in}$ is increased, we see that the stop LSP region moves to
larger $m_0$ and $|A_0/m_0|$, while low values of $|A_0/m_0|$ are
excluded because of the $B$ matching condition. For $m_h = 125$ GeV, the
allowed values of $m_0$ and $|A_0/m_0|$ increase as $M_{in}$ is
increased. For very large $M_{in}$, we see that the intersection of the
$m_h$ contour with the stop strip occurs at lower $\tau_p$ and for
$M_{in} = 10^{17.5}$~GeV, the intersection point occurs below the
current experimental bound. The star $(\bigstar)$ in the
lower left panel with $M_{in} = 10^{17}$ GeV, is a benchmark we used
in Section~\ref{pdecay} to discuss the choice of phases. At this point,
which is located at $m_0 = 11.6$ TeV and $A_0/m_0 = -3.7$, we must take
$c = -0.0095$ in Eq.~\eqref{eq:SigmaWW} in order to obtain $\lambda =
0.6$ with $\lambda^\prime = 10^{-4}$ and we find that the Higgs mass is
$m_h = 125.6$ GeV and $\tau_p \approx 10^{34}$ yrs. As shown in
Fig.~\ref{fig:pknuphase}, this lifetime requires phases $(\phi_2,
\phi_3) = (0.64, 1.96)\pi$. If the phases vanish, the lifetime drops by
a factor of about 5 to $\tau_p = 1.9 \times 10^{33}$ yrs. The mass
spectrum at this point is shown in Table~\ref{table}. As can be seen,
the gluino mass is $\simeq 2m_{1/2}\simeq 8$~TeV, which is within the
reach of the 100~TeV collider \cite{Golling:2016gvc}. On the other hand,
squark masses are $\gtrsim 10$~TeV, and thus it may be difficult to
discover squarks even at the 100~TeV collider.
\begin{table}[htb!]
\begin{center}
\begin{tabular}{cc|cc}
\hline
\hline
Particle & Mass [TeV] & Particle & Mass [TeV] \\
\hline
\hline
${\chi_1^0}$ & 1.75 & $\chi_2^0$ & 3.45 \\
${\chi_3^0}$ & 12.8 & $\chi_4^0$ &12.8 \\
${\chi^\pm_1}$ & 3.45 & $\chi_2^\pm$ & 12.8\\
$h$ & 0.1256 & $H$ & 14.9 \\
$A$ & 14.9 & $\tilde{g}$ & 7.97 \\
$\tilde{e}_L$ & 11.8 & $\tilde{e}_R$ & 12.0 \\
$\tilde{\nu}_e$ & 11.8 & & \\
$\tilde{\tau}_1$ & 8.29 & $\tilde{\tau}_2$ & 11.8 \\
$\tilde{\nu}_\tau$ & 11.8 & & \\
$\tilde{u}_L$ & 13.2 & $\tilde{u}_R$ & 12.9 \\
$\tilde{d}_L$ & 13.2 & $\tilde{d}_R$ & 13.0 \\
$\tilde{t}_1$ & 1.76 & $\tilde{\tau}_2$ & 7.48 \\
$\tilde{b}_1$ & 7.34 & $\tilde{b}_2$ & 12.9 \\
\hline
\hline
\end{tabular}
\caption{\it Particle Spectrum at the benchmark point indicated by a
star $(\bigstar)$ in Fig.~\ref{fig:diskr1}.}
\label{table}
\end{center}
\end{table}
The dependence of these results on $m_{1/2}$ can be gleaned from
Fig.~\ref{fig:CMSSM}. For smaller $m_{1/2}$, the Higgs mass and proton
lifetime both decrease. At higher $m_{1/2}$, we approach the endpoint of
the stop strip. For example, when $m_{1/2} = 6$~TeV, there would be no
blue strip alongside the red region (which would look similar to the
case displayed), as the relic density would exceed the Planck value even
for degenerate stops and neutralinos. The results scale as one might
expect with $\tan \beta$. At higher $\tan \beta$, the Higgs mass
increases while the proton lifetime decreases. For example, at $\tan
\beta = 7$, for the same value of $A_0/m_0$, the position of the star
when $M_{in} = 10^{17}$~GeV moves slightly to $m_0 = 11.5$ TeV, and the
Higgs mass increases to 126.1~GeV according to {\tt FeynHiggs}, but
$\tau_p$ decreases to $6.2 \times 10^{33}$ yrs.
From the discussion in section~\ref{pdecay}, we expect that there is a
strong dependence of $\tau_p$ on $\lambda^\prime$, while little else is
affected. For example, increasing (decreasing) $\lambda^\prime$ by an
order of magnitude moves the stop-coannihilation strip of the lower left
panel of Fig.~\ref{fig:diskr1} so that the star would be at 12.1~TeV
(11.2~TeV) for $A_0/m_0$ unchanged. The Higgs mass, $m_h$, for this
shifted point is almost unchanged, 125.8~GeV (125.5~GeV), while $\tau_p$
drops by a factor of 5 (increases by a factor of 4). The dependence on
$\lambda$ is discussed in more detail below. We also checked on the
effect of changing the sign of $\mu$ and the ratio of $m_\Sigma/m_0$ for
the case considered in the lower left panel of
Fig.~\ref{fig:diskr1}. For both changes, the stop strip and proton
lifetime are barely altered. For $\mu > 0$, the Higgs mass drops
significantly. At the position of the star, the Higgs mass is 117~GeV
for $\mu > 0$. For this reason we have largely focused on $\mu < 0$ in
this paper. For $m_\Sigma/m_0 = 0.1$ the only noticeable change in the
figure is the absence of the $B$ matching constraints which is greatly
relaxed when $m_\Sigma < m_0$. We note that, for $m_\Sigma^2 = 0$ or even
negative, we are able to recover solutions with $A_0 = 0$. However, when
$M_{in} > M_{GUT}$, one does not find a a focus-point region as
discussed previously \cite{emo}.
We next show two examples of $(m_{1/2},m_0)$ planes for $M_{in} =
10^{17}$~GeV, $\tan \beta = 6$ and $\mu < 0$, which can be compared with
the lower right panel of Fig.~\ref{fig:CMSSM}. In the left panel of
Fig.~\ref{fig:superG} we choose $A_0/m_0 = -4.2$ as in
Fig.~\ref{fig:CMSSM}. For this value of $M_{in}$, we see the appearance
of a mauve shaded region that is excluded because the $B$ matching
condition \eqref{alimit} cannot be satisfied. The $X$ located at (5.3,
12.0)~TeV again denotes the endpoint of the stop strip. This occurs when
$m_h = 125.5$~GeV and $\tau_p = 1.1 \times 10^{34}$~yrs. Thus only a
short segment of the stop strip is viable in this case. In the right
panel with $A_0/m_0 = -3.5$, we see that a larger fraction of the plane
is excluded by the failure to satisfy the $B$ matching condition. The
stop endpoint has moved to higher mass scales $(m_{1/2}, m_0) =
(5,16)$~TeV, where $m_h = 128.1$~GeV and $\tau_p = 2 \times
10^{34}$~yrs, and a larger portion of the strip is viable. In both
cases, the viable parameter points can be probed at future collider
experiments.
\begin{figure}[htb!]
\begin{minipage}{8in}
\includegraphics[height=3.3in]{m0vM8aspnGr_6_-42_1_17_1_fsef}
\includegraphics[height=3.3in]{m0vM8aspnGr_6_-35_1_17_1_fsef}
\hfill
\end{minipage}
\caption{
{\it
Super-GUT CMSSM $(m_{12}, m_0)$ planes for $M_{in} = 10^{17}$~GeV, $\tan \beta = 6$ and
$\mu < 0$, for $A_0/m_0 = -4.2$ (left) and $-3.5$ (right). In each panel,
we have fixed $\lambda = 0.6$ and $\lambda^\prime = 0.0001$.
Shadings and contours are as in Fig. \ref{fig:CMSSM}.
The mauve shaded regions are excluded because it is not possible to satisfy the matching condition for $B$.
The $X$ marks the endpoint of the stop coannihilation strip.}}
\label{fig:superG}
\end{figure}
Finally, we discuss the dependence on $\lambda$ and $\lambda'$ by considering the
$(\lambda, \tan \beta)$ plots shown in Fig. \ref{fig:diskr2}, which are for $m_{1/2} = 4$~TeV,
$m_0 = 10$~TeV and $\mu < 0$, with different values of $(M_{in}, A_0/m_0, \lambda^\prime)$.
The upper left panel is with the values $(10^{17}$~GeV$, -4.2, 0.0001)$, which serve as
references. We see that the dark matter strip is adjacent to the brown stop LSP region
at $\lambda \simeq 0.67$, growing only slightly with $\tan \beta$ in the range displayed.
Along this strip, the proton lifetime constraints is respected for $\tan \beta \lesssim 6.5$,
where $m_h \sim 125$~GeV according to {\tt FeynHiggs}.
Here, one sees very clearly the dependences of $m_h$ and $\tau_p$ on $\tan \beta$.
\begin{figure}[htb!]
\begin{minipage}{8in}
\includegraphics[height=3.3in]{A0vm0pndiskr1f_4000_10000_42_1_17_0001_1_fsef}
\includegraphics[height=3.3in]{A0vm0pndiskr1_4000_10000_42_1_175_0001_1_fsef}
\hfill
\end{minipage}
\begin{minipage}{8in}
\includegraphics[height=3.3in]{A0vm0pndiskr1_4000_10000_4_1_17_0001_1_fsef}
\includegraphics[height=3.3in]{A0vm0pndiskr1_4000_10000_42_1_17_1e-05_1_fsef}
\end{minipage}
\caption{
{\it
Super-GUT CMSSM $(\lambda,\tan \beta)$ planes with $m_{1/2} = 4$~TeV,
$m_0 = 10$~TeV, $\mu < 0$ and various values of $(M_{in}, A_0/m_0, \lambda^\prime)$
$ = (10^{17}$~GeV$, -4.2, 0.0001)$ (upper left), $ = (10^{17.5}$~GeV$, -4.2, 0.0001)$ (upper right),
$ = (10^{17}$~GeV$, -4.0, 0.0001)$ (lower left) and $ = (10^{17}$~GeV$, -4.2, 0.00001)$ (lower right).}
\label{fig:diskr2}}
\end{figure}
In the upper right panel of Fig. \ref{fig:diskr2},
$M_{in}$ is increased to $(10^{17.5})$~GeV, and we see that the dark matter-compatible
value of $\lambda$ decreases to $\sim 0.55$ and proton stability then enforces $\tan \beta \lesssim 5.2$,
with $m_h$ about a GeV smaller than before, but still compatible with the LHC measurement
when the {\tt FeynHiggs} uncertainties are taken into account.
Had we decreased $M_{in}$ to $10^{16.5}$ GeV, the coannihilation strip would
have moved to $\lambda \approx 0.90$,
and the proton stability constraint would have required $\tan \beta \lesssim 8.3$. At the limit, $m_h \simeq 127$~GeV
and is lower at lower $\tan \beta$.
In the lower left panel of Fig. \ref{fig:diskr2},
$- A_0/m_0$ is decreased to 4.0, with $M_{in}$ and $\lambda^\prime$ taking their
reference values. In this case, the dark matter constraint requires $\lambda \sim 0.6$ and
proton stability then imposes $\tan \beta \lesssim 5.5$, again compatible with $m_h$.
Increasing $- A_0/m_0$ to 4.4 would move the coannihilation strip to $\lambda \simeq 0.72$, and
the limit on $\tan \beta$ would become $\tan \beta \lesssim 6.6$ with $m_h$ close to 126 GeV.
Finally, we see in the lower right panel of Fig. \ref{fig:diskr2} that for $\lambda^\prime = 0.00001$ and the reference
values of $M_{in}$ and $A_0/m_0$ the dark matter density requires $\lambda \simeq 0.68$
and proton stability then allows $\tan \beta \lesssim 9.8$. Most of this part of the strip
is also compatible with $m_h$, given the uncertainty in the {\tt FeynHiggs} calculation.
A larger value of $\lambda^\prime = 0.001$ would require $\tan \beta \lesssim 3.6$, but for this value of
$\tan \beta$ the Higgs mass would be unacceptably small, around 120.4 GeV.
\section{Discussion}
\label{sec:discussion}
It is frequently stated that the minimal SU(5) GUT model is excluded by
the experimental lower limit on the proton lifetime. Taking into account
the cosmological constraint on the cold dark matter density, the LHC
measurement of $m_h$ and the unknown GUT-scale phases appearing in the
SU(5) GUT model, we have shown in this paper that this model is quite
compatible with the proton stability constraint.
We remind the reader that the amplitudes for the (normally) dominant
$p\to K^+ \overline{\nu}$ decay modes depend on two GUT-scale phases
that are beyond the CKM framework, and are not constrained by low-energy
physics. As we have discussed in detail, their effects on the $p\to K^+
\overline{\nu}_\tau$ decay amplitude are different from those on the
$p\to K^+ \overline{\nu}_{e, \mu}$ decay amplitudes. We take these
effects into account, and also consider their effects on the (normally)
subdominant $p\to \pi^+ \overline{\nu}$ and $n\to \pi^0 \overline{\nu}$
decays modes. In order to derive the most conservative bounds on the
model parameters, we choose the unknown GUT-scale phases so as to
maximize the $p\to K^+ \overline{\nu}$ lifetime.
The compatibility of the supersymmetric GUT model with the proton
stability constraint is already visible in the CMSSM with universality
of the soft supersymmetry-breaking scalar masses imposed at an input
scale $M_{in} = M_{GUT}$ and $\tan \beta \sim 5$. This is visible in
Fig.~\ref{fig:CMSSM} along the upper parts of the focus-point strips in
the left panels (with $A_0 = 0$) and of the stop-coannihilation strips
in the right panels (with $A_0 = - 4.2 m_0$). According to the latest
version of {\tt FeynHiggs}, large portions of these strips are also
compatible with the experimental measurement of $m_h$.
The super-GUT CMSSM with $M_{in} > M_{GUT}$ has more parameters, namely
the superpotential couplings $\lambda$ and $\lambda^\prime$ as well as
$M_{in}$. Correspondingly, the super-GUT CMSSM has greater scope for
compatibility with the proton stability and $m_h$ constraints. We had
previously noted \cite{emo} that, for $A_0 = 0$, the focus-point strip
move quickly to smaller $m_{1/2}$ and larger $m_0$ as $M_{in}$ is
increased. The stau LSP region also quickly recedes
\cite{superGUT,emo}. Here, we have added the matching condition for $B$,
previously neglected in other analyses. This led us to concentrate on
relatively large values of $|A_0/m_0|$. We have given some illustrative
examples of suitable parameter choices in Figs.~\ref{fig:diskr1},
\ref{fig:superG} and \ref{fig:diskr2}. Typical value of the model
parameters are $M_{in} = 10^{17}$~GeV, $m_{1/2} = 4$~TeV, $m_0 =
10$~TeV, $A_0/m_0 \sim - 4$, $\tan \beta \sim 5$, $\lambda \sim 0.6$ and
$\lambda^\prime \lesssim 0.0001$.
To evade the proton decay constraints, squarks are required to be as
heavy as $\gtrsim 10$~TeV, which are hard to probe even at the 100~TeV
collider; see~\cite{EZ}, however. On the other hand, the gluino mass can be $\lesssim 10$~TeV,
which can be probed at the 100~TeV collider \cite{Golling:2016gvc}.
Such heavy sparticle masses require fine-tuning at the
electroweak scale \cite{eenz}; at the expense of this, the simple models discussed
in this paper, the minimal SU(5) GUT with (super-GUT) CMSSM, are found
to be able to meet all the phenomenological requirements. Of course, by
extending the models and/or introducing more complicated mechanisms, we
may find a less fine-tuned sparticle spectrum with which the problems in
the minimal SU(5), such as the doublet-triplet splitting and the
dimension-five proton decay problems, can be evaded---this is beyond the
scope of the present work.
In view of the sensitivity of the proton lifetime to the unknown GUT-scale phases, it would
interesting to derive model predictions for them---another objective for theories of quark
and lepton mixing to bear in mind. Even more interesting would be to devise ways to
measure these phases experimentally. In principle, one way to do this would be to
measure the ratios of $p\to K^+ \overline{\nu}$, $p\to \pi^+ \overline{\nu}$ and
$n\to \pi^0 \overline{\nu}$ decay modes, as illustrated in Fig.~\ref{fig:nucleon}.
This may seem like a distant prospect, but let us remember that the
Hyper-Kamiokande project, in particular, has an estimated 90\% CL sensitivity to
$p\to K^+ \overline{\nu}$ at the level of $2.5 \times 10^{34}$~yrs~\cite{HyperK}. This covers the
range allowed in Fig.~\ref{fig:phasedep} for the reference point indicated by a star $(\bigstar)$
in Fig.~\ref{fig:diskr1}, and illustrates the capability of Hyper-Kamiokande to
probe the GUT-scale physics of proton decay. Let us be optimistic!
\section*{Acknowledgements}
The work of J.E. was supported in part by the UK STFC via the research grant ST/J002798/1.
The work of J.L.E., N.N. and K.A.O.
was supported in part by DOE grant DE-SC0011842 at the University of Minnesota.
\section*{Appendix}
|
\section*{Small-worlds, Watts and Strogatz}
Small-worldness now seems to be a ubiquitous characteristic of many complex systems; but its first, and still most familiar, appearance was in the form of social networks. We know that as individual agents (nodes) in a social network, we are connected by strong familial and friendship ties (edges) to a relatively few people who are likely also strongly connected to each other, forming a social clique, family or tribe. Yet we also know that we can travel far away from our tribal network, to physically remote cultures and places, and sometimes be surprised there to meet people -- often ``friends-of-friends'' -- who are quite closely connected to our home tribe: ``it's a small world'', we say. This common intuition was experimentally investigated by Milgram, who asked people in the mid-West of the US (Omaha, Nebraska) to forward a letter addressed to an unknown individual in Boston by posting it to the friend or acquaintance in their social network that they thought might know someone else who would know the addressee \cite{Milgram1967} (Fig.~\ref{Milgram}). It was discovered, on average over multiple trials of this procedure, that the letters successfully reaching Boston had been passed through 6 intermediate postings, which was considered much less than expected given the geographical distance between source and target addresses. In the language of graph theory, the characteristic path length of Milgram's social networks was short.
\begin{figure*}[h]
\begin{center}
\includegraphics[width=0.7\textwidth]{SWR_Milgram_Fig.pdf}
\caption{ \textbf{An illustration of the shortest path between Omaha and Boston in Milgram's social network experiment.} An image from Stanley Milgram's original experiment, published in {\em Psychology Today} in 1967. Here, the results of multiple experiments are represented as a composite shortest path between the source (a person in Omaha) and the target (a person in Boston). A letter addressed to the target was given to the source, who was asked to send it on (with the same instructions) to the friend or acquaintance that they thought was most likely to know the target, or someone else who might know the target personally. It was found that most letters that eventually reached the correct address in Boston passed through six intermediaries between source and target (denoted 1st remove, 2nd remove, etc), popularising the notion that each of us is separated by no more than ``six degrees of freedom'' from any other individual in a geographically distributed social network. Reproduced with permission from \cite{Milgram1967}. \label{Milgram}}
\end{center}
\end{figure*}
Famously, Watts \& Strogatz (1998) \cite{watts1998collective} combined this concept of path length (the minimum number of edges needed to make a connection between nodes) with a measure of topological clustering or cliquishness of edges between nodes (Fig.~\ref{cl_binary_weighted}). More formally, clustering measures the probability that the nodes $j$ and $k$, which are both directly connected to node $i$, are also directly connected to each other; this is equivalent to measuring the proportion of closed triangular 3-node motifs in a network \cite{sporns2004}. Watts \& Strogatz (WS) explored the behaviour of path length and clustering in a simple generative model (henceforth the WS model) (Fig.~\ref{WS_SW}). Starting with a binary lattice network of $N$ nodes each connected to the same number of nearest neighbors, by edges of identical weight (unity), the WS model iteratively re-wires the lattice by randomly deleting an existing edge, between nodes $i$ and $j$, and replacing it by a new edge between node $i$ and any node $k \neq j$. They found that as the probability of random rewiring was incrementally increased from zero, so that the original lattice was progressively randomised, sparsely rewired networks demonstrated both high clustering (like a lattice) and short path length (like a random graph). By analogy to social networks, these algorithmically generated graphs were called small-world networks.
\begin{figure*}[h]
\begin{center}
\includegraphics[width=0.5\textwidth]{cl_binary_weighted.pdf}
\caption{ \textbf{Diagrams of clustering and path length in binary and weighted networks.} \emph{(A)} In a binary network, all edges have the same weight, and that is a weight equal to unity. In this example of a binary graph, if one wishes to walk along the shortest path from the orange node to the green node, then one would choose to walk along the edges highlighted in red, rather than along the edges highlighted in blue. We also note that the clustering coefficient of the green node is equal to 1 (all neighbors are also connected to each other to form a closed triangular motif), while the clustering coefficient of the orange node is $<<1$ (only 3 out of 5 neighbors are also connected to each other). \emph{(B)} In a weighted graph, edges can have different weights. In this example, edges have weights of $3/3=1$, $2/3=0.66$, and $1/3=0.33$. If one wishes to traverse the graph from the orange node to the green node along the shortest path, one would choose to follow the path along the edges with weight equal to unity (stronger weights are equivalent to shorter topological distance). Note also that because the edges are now weighted, neither the orange nor the green nodes has a clustering coefficient equal to unity. \label{cl_binary_weighted}}
\end{center}
\end{figure*}
\begin{figure*}[h]
\begin{center}
\includegraphics[width=0.7\textwidth]{WS_Fig.pdf}
\caption{ \textbf{The Watts-Strogatz model and the generation of small-world networks.} The canonical model of a small-world network is that described by Duncan Watts and Steve Strogatz in their 1998 paper in Nature \cite{watts1998collective}. The model begins with a regular lattice network in which each node is placed along the circumference of a circle, and is connected to its $k$ nearest neighbors on that circle. Then, with probability $p$, edges are rewired uniformly at random such that (i) at $p=0$, the network is a lattice and (ii) at $p=1$, the network is random. Interestingly, at intermediate values of $p$, the network has so-called ``small-world'' characteristics with significant local clustering (from the lattice model) and short average path-length facilitated by the topological short-cuts created during the random rewiring procedure. Because this architecture can be defined mathematically, small-world graphs have proven fundamental in understanding game theory \cite{li2009largest} and even testing analytical results in subfields of mathematics \cite{konishi2011topology}. Yet, while this work provided a qualitative model of a small-world graph, it did not give a statistic to measure the degree of small-worldness in a particular data set. As a simple scalar measure of ``small-worldness'', Humphries and colleagues defined the \emph{small-world index}, $\sigma$, to be the ratio of the clustering coefficient (normalized by that expected in a random graph) to the average shortest path length (also normalized by that expected in a random graph) \cite{humphries2006brainstem}. The intuition here is that this index should be large (in particular, $\sigma>1$) when the clustering coefficient is much greater than expected in the random graph, and the average shortest path length is comparable to that expected in a random graph. Since this initial definition, other extensions have been proposed and utilized \cite{toppi2012how,telesford2011ubiquity}, building on the same general notions. \label{WS_SW}}
\end{center}
\end{figure*}
In addition to introducing this generative model, Watts \& Strogatz also showed how small-worldness could be estimated in naturally occurring networks \cite{watts1998collective}. The hybrid combination of high clustering and short path length that emerged in sparsely re-wired WS networks was proposed as a general quantitative measure of small-worldness (SW) in other networks. It was shown immediately that a nervous system was among the real-world networks that shared the SW pattern of topological organisation. Using data on the synaptic and gap junction connectivity between all $N=302$ neurons in the nervous system of {\em Caenorhabditis elegans} \cite{white1986structure}, a binary undirected graph was constructed representing each neuron as an identical node and each synapse ($\sim 5000$) or gap junction ($\sim 600$) as an identical, unweighted and undirected edge between nodes. This graph of about 5600 edges between 302 nodes was sparsely connected: only about 12\% of the maximum possible number of synaptic connections, $(N^2-N)/2 = 45,451$, actually existed. Compared to a random graph of $N$ nodes, {\em C elegans} had high clustering $\Gamma \sim 5.6$ and short path length $\Lambda \sim 1.18$. Thus the {\em C. elegans} connectome was small-world, in the same quantitative sense as the networks generated by the WS model at low re-wiring probabilities, less than 10\%. But note that does {\em not} necessarily mean that the {\em C. elegans} connectome was biologically generated by the WS algorithm of random rewiring of established connections (axonal projections) between neurons. To put it another way, the WS model can generate SW networks but not all SW networks were generated by a WS model. (And the WS model does not seem like a biologically plausible generative model for brain networks \cite{vertes2012simple,vertes2014generative,betzel2016generative}.)
\subsection*{Small-world brain graphs}
Following the small-world analysis of {\em C elegans}, pioneering topological studies of mammalian cortical networks used databases of tract-tracing experiments to demonstrate that the cat and macaque inter-areal anatomical networks shared similar small-world properties of short path length and high clustering \cite{sporns2004small,hilgetag2004clustered}. The first graph theoretical studies of neuroimaging data demonstrated that large-scale inter-areal networks of functional and structural connectivity in the human brain also had small-world properties \cite{salvador2005neurophysiological,bassett2006adaptive,vaessen2010effect}. These and other seminal discoveries were central to the emergence of connectomics as a major growth point of network neuroscience \cite{sporns2005human}.
About 10 years ago, we reviewed these and other data in support of the idea that the brain is a small world network \cite{bassett2006small}. Here, we aim to take another look at the concept of small-worldness, about one or two decades since it was first formulated quantitatively and applied to brain network analysis at microscopic and macroscopic scales of anatomical resolution. First, we review some of the key questions about small-worldness that have been a focus of work in the period 2006--2016; then we review the technical evidence for small-worldness in high resolution tract-tracing data from the macaque and the mouse; finally, we highlight some likely trends in the further evolution of small-worldness as part of a deeper understanding of the topology of weighted brain graphs.
\section*{What have we (not) learnt since 2006?}
We have learnt a lot about complex topological organisation of nervous systems since 2006, as evidenced by rapid growth in research articles, reviews and citations related to``brain graphs'' and ``connectomes" \cite{bullmore2009complex,bullmore2011brain,pessoa2014understanding}; by the publication of several textbooks \cite{sporns2011networks,fornito2016fundamentals}; and by the recent launch of new specialist journals for network neuroscience \cite{bassett2016network}. This emerging field of brain topology has grown much bigger than the foundational concept of small-worldness. But what have we learnt more specifically about brain small-worldness since 2006, and what do we still have to learn?
\subsection*{Universality}
There is no doubt that small-worldness -- the combination of non-random clustering with near-random path length -- has been very frequently reported across a wide range of neuroscience studies. Small-world topology has been highly replicated across multiple species and scales from structural and functional MRI studies of large scale brain networks in humans to multi-electrode array recordings of cellular networks in cultures \cite{bettencourt2007functional} and intact animals \cite{van2016comparative}. It seems reasonable to conclude that small-worldness is at least very common in network neuroscience; but is it a universal property of nervous systems? Universality is a strong claim and difficult to affirm conclusively. As Popper noted in his philosophy of science by hypothetical refutation \cite{popper1963}, the universal hypothesis that ``all swans are white'' can only be affirmed conclusively by a complete survey of every swan in the world. Whereas it can be immediately and decisively refuted by the observation of a single black swan. Similarly, the claim that {\em all} brains have small-world topology has not yet been (and never will be) affirmed by a complete connectomic mapping of every brain in the world. Some apparent counter-examples of brain networks that do {\em not} have small-world topology have been reported and deserve careful consideration as possible Popperian black swans (see below). However, we can provisionally conclude that enough evidence has amassed to judge that small-worldness is a nearly universal property of nervous systems. Indeed it seems likely that brains are only one of a large ``universality class'' of small-world networks comprising also many other non-neural or non-biological complex systems. Such near-universality of small-worldness, or any other brain network parameter, has a number of implications.
First, near-universality implies {\em self-similarity}. If the macro-scale inter-areal network of the human brain is small-world, as is the micro-scale inter-neuronal network of the worm or the fly, then we should expect also that the micro-scale inter-neuronal network of the human brain is small-world. Self-similarity of small-worldness would be indexed by scale invariance of network path length and clustering parameters as the anatomical resolution ``zooms in'' from macro- to micro-scales. Although there is abundant evidence for scaling, fractal or self-similar statistics in many aspects of brain network topology \cite{bullmore2009generic,bassett2010efficient,klimm2014resolving}, experimental data do not yet exist that could support a multi-scale, macro-to-micro analysis of small-worldness (and other network properties) in the same (human or mammalian) nervous system \cite{bassett2013multiscale}.
Second, near-universality suggests some very general selection pressures might be operative on the {\em evolution and development} of nervous systems across scales and species. This line of thinking has led to the formulation of generative models that can simulate brain networks by some probabilistic growth rule or genetic algorithm. It has been found that simple generative models, that add edges to a network based on the spatial distance and the topological relationships between nodes, can recapitulate small-worldness and many other properties of the connectome on the basis of two (spatial and topological) parameters \cite{vertes2012simple,vertes2014generative,betzel2016generative}. This serves as a reminder that the network phenotype of small-worldness can be generated by many different mechanisms and the biological mechanisms controlling formation of small-world properties in brain networks currently remain unknown.
Third, and from a somewhat more controversial perspective, universality might seem tantamount to {\em triviality}. If the brain is everywhere small-world, and so are almost all other complex systems in real-life \cite{bassett2006small,bullmore2009generic,gaiteri2014beyond,moslonka2011networks,sizemore2016classification} (for a few exceptions, see \cite{koschutzki2010structural}), then what is the small-worldness of the brain telling us that's of any interest specifically to neuroscience? There are two main answers to this important question, as we discuss in more detail below: (i) studies have recently succeeded in linking network topological metrics to biological concepts, like wiring cost \cite{bullmore2012economy,bassett2010efficient,bassett2011conserved,rubinov2015wiring}, and to biological phenotypes, like neuronal density \cite{acimovic2015effects,vandenheuvel2015bridging} or gene expression \cite{fulcher2016transcriptional,vertes2016philtrans,whitaker2016pnas}; and (ii) small-worldness is not the whole story of brain network organisation \cite{wang2016brain}.
\subsection*{Economical small-world networks}
At the risk of stating the obvious, small-worldness is a purely topological quantity that tells us nothing about the physical layout of the nodes or edges that constitute the graph \cite{bassett2010efficient,pessoa2014understanding}. However, it is equally obvious that brain networks are embedded in anatomical space \cite{klimm2014resolving,lohse2014resolving,bassett2011conserved,betzel2016modular}. Somehow the abstract, dimensionless topology of small-worldness must be reconciled to the anatomy of the brain. It turns out that the small-world topology of brain networks is (almost) always {\em economically} embedded in physical space \cite{bullmore2012economy,kaiser2006nonoptimal}.
For both clustering and path length, the two topological metrics combined in the hybrid small-world estimator, there is a strong relationship with brain anatomical space \cite{bassett2010efficient,bassett2011conserved,rubinov2015wiring}. The edges between clustered nodes tend to be shorter distance whereas the edges that mediate topological short cuts tend to traverse longer anatomical distances. Interpreting the Euclidean distance between brain regional nodes or neurons as a proxy for the wiring cost, i.e., the total biological cost of building a physical connection and maintaining communication between nodes, it has been argued that the brain is an economical small-world network \cite{latora2001efficient,bullmore2012economy}. Economical in this sense does not simply mean parsimonious or cheap; it is more closely related to the common-sense notion of ``value for money''. Topologically clustered nodes are anatomically co-located and thereby nearly minimise wiring cost. But small-world brain networks are not naturally lattices and if they are computationally rewired strictly to minimise wiring cost then brain networks are topologically penalised, losing integrative capacity indexed by increased characteristic path length and thus reduced small-worldness scalar $\sigma$.
The economical idea is that brain networks have been selected by the competition between a pressure to minimise biological cost \emph{versus} a pressure to maximise topological integration. More formally,
\begin{equation}
P_{i,j} \sim f(d_{i,j}) f(k_{i,j}),
\end{equation}
the probability of a connection between nodes $i$ and $j$, $P_{i,j}$, is a product of: a function of the physical distance in mm between nodes $d_{i,j}$ - often used as a proxy for wiring cost; and a function of the topological relationship between nodes - $k_{i,j}$.
Typically the functions of cost and topology are each parameterised by a single parameter, for example, simple exponential and power law functions. Several variants of this approach have been published, exploring a range of different topological relationships $k_{i,j}$ between nodes, for example, clustering and homophily \cite{vertes2012simple,vertes2014generative,betzel2016generative}. Economical models can generally reproduce the small world properties of brain networks quite realistically: clustering and path length are both increased as a function of the cost parameter \cite{avena2014using}. In other words, as the cost penalty becomes the dominant factor predicting the probability of a connection between nodes, economical models generate increasingly lattice-like networks, with strong spatial and topological clustering of connected nodes, approximating in the limit the minimal cost configuration of the network. The emergence of more integrative network features -- such as hubs mediating many inter-modular connections -- typically depends on some degree of relaxation of the cost penalty (reduced distance parameter) relative to the parameter controlling the importance of (integrative) topological relationships between nodes in predicting their connectivity. Thus small world networks can be generated by economical models for a certain range of the two parameters controlling the competitive factors of (wiring) cost and (topological) value.
\subsection*{Small-worldness is not the whole story}
Before getting further into the details of small-worldness, as we do below in relation to recent tract-tracing results, it is important to acknowledge that the specific metrics of path length $\Lambda$ and clustering $\Gamma$ introduced by Watts \& Strogatz \cite{watts1998collective}, and the small-worldness scalar derived from them $\sigma = {\Gamma}/{\Lambda}$ \cite{humphries2006brainstem}, are a few global topological metrics that have been of central importance to the growth of complex network science generally. But more than 15 years after the first discovery of small world properties in brain networks, the field of connectomics now extends into many other areas of topological analysis. There is much important recent work on topological properties like degree distribution and hubness \cite{achard2006resilient}, modularity \cite{simon1962architecture,meunier2009age,chen2008revealing,sporns2016modular,bassett2011dynamic,mattar2015functional,stoop2013beyond}, core/periphery organisation \cite{bassett2013task,senden2014rich,van2011rich}, controllability \cite{gu2015controllability,muldoon2016stimulation,betzel2016optimally} and navigability \cite{gulyas2015navigable} that are not simply related to small-worldness. Outside neuroscience there continues to be strong growth in the more general field of network science \cite{barabasi2016book}. It is nothing like a complete description of the brain to say it is small-world; we now turn to a more technical discussion of the evidence for small-worldness as a common property of nervous systems.
\section*{Challenges to small-worldness}
About 3--4 years ago, an important series of papers began to be published that could be regarded as ``black swans'' refuting the general importance of small-worldness in an understanding of brain networks \cite{markov2012cerebcortex,markov2013cortical,song2014spatial,knoblaugh2016}:
\begin{quote}
Previous studies of low density inter-areal graphs and apparent small-world properties are challenged by data that reveal high-density cortical graphs in which economy of connections is achieved by weight heterogeneity and distance-weight correlations. \cite{markov2013cortical}
\end{quote}
\begin{quote}
Recent connectomic tract tracing reveals that, contrary to what was previously thought, the cortical inter-areal network has high density. This finding leads to a necessary revision of the relevance of some of the graph theoretical notions, such as the small world property..., that have been claimed to characterise the inter-areal cortical network. \cite{knoblaugh2016}
\end{quote}
These remarks carried weight because they were based on sophisticated and highly sensitive measurements of mammalian cortical connectivity (Fig.~\ref{kennedy}). In each one of multiple carefully standardised experiments in the macaque monkey, a fluorescent tracer was injected into a (target) cortical region where it was taken up by synaptic terminals and actively transported to the cell bodies of neurons projecting to the target region. When the animal's brain was subsequently examined microscopically, the retrograde transport of the tracer from the injection site resulted in a fluorescent signal in the (source) regions of cortex that were directly connected to the target region. The basic technology of anatomical tract-tracing had been used by neuroanatomists since the late 20th century; but in the first decades of the 21st century it was possible to increase the scale and precision of the measurements dramatically, enabling the construction of connectivity matrices that summarised the strength or weight of axonal projections between a large number of cortical areas. These next-generation tract-tracing data thus represented a new standard of knowledge about mammalian cortical connectivity, that was more continuously quantified than the binary or ordinal rating of connectivity from traditional tract-tracing experiments \cite{stephan2001cocomac}, and much less ambiguously related to the cellular substrates of brain networks than the statistical measures of functional connectivity \cite{achard2006resilient,zhang2016choosing} and structural covariance \cite{bassett2008hierarchical,alexander2013imaging} used to build graphs from human neuroimaging data. It is clearly important to understand in some detail how the topology of brain networks can be modelled in contemporary tract-tracing data from the macaque (and subsequently the mouse \cite{oh2014mesoscale,rubinov2015wiring}) and what these results tell us about the small-worldness of brain networks.
\begin{figure*}[h]
\begin{center}
\includegraphics[width=0.9\textwidth]{SWR_KennedyScience.pdf}
\caption{\textbf{High density of the macaque cortical graph excludes sparse small world architecture} \emph{(A)} Comparison of the average shortest path length and density of the macaque cortical graph from \cite{markov2013cortical} with the graphs of previous studies. Sequential removal of weak connections causes an increase in the path length. Black triangle: macaque cortical graph from \cite{markov2013cortical}; gray area: 95\% confidence interval following random removal of connections from the macaque cortical graph from \cite{markov2013cortical}. Modha and Singh 2010: \cite{modha2010network}; Young 1993: \cite{young1993organization}; Honey et al., 2007: \cite{honey2007network}; Felleman and Van Essen 1991: \cite{felleman1991distributed}; Jouve et al., 1998: \cite{jouve1998mathematical}; Markov et al., 2012: \cite{markov2014weighted}. “Jouve et al., 1998 predicted” indicates values of the graph inferred using the published algorithm \cite{jouve1998mathematical}. \emph{(B)} Effect of density on Watts and Strogatz's formalization of a small world network. Clustering and path length variations generated by edge rewiring with probability range indicated on the $x$-axis applied to regular lattices of increasingly higher densities. The pie charts show graph density encoded via colors for path length ($L$) and clustering coefficient ($C$). The $y$-axis indicates the path length ratio ($L_{p}$/$L_{o}$) and clustering ratio ($C_{p}$/$C_{o}$) of the randomly rewired network, where $L_{o}$ and $C_{o}$ are the path length ($L_{o}$) and clustering ($C_{o}$) of the regular lattice, respectively. The variables $L_{p}$ and $C_{p}$ are the same quantities measured for the network rewired with probability $p$. Hence, for each density value indicated in the $L$ and $C$ pie charts, the corresponding $L_{p}$/$L_{o}$ and $C_{p}$/$C_{o}$ curves can be identified. Three diagrams below the $x$-axis indicate the lattice (\emph{left}), sparsely rewired (\emph{middle}), and the randomized (\emph{right}) networks. \emph{(C)} The small-world coefficient $\sigma$ \cite{humphries2006brainstem} corresponding to each lattice rewiring. Color code is the same as in panel \emph{(B)}. Dashed lines in \emph{(B)} and \emph{(C)} indicate 42\% and 48\% density levels. Reproduced with permission from \cite{markov2013cortical}. \label{kennedy}}
\end{center}
\end{figure*}
\subsection*{Binary graphs}
In general, a node represents a component of a system and an edge represents a connection or interaction between two nodes. Mathematically, we can capture these ideas with a graph $\mathcal G = (\mathcal V, \mathcal E)$ composed of a node set $\mathcal V$ and an edge set $\mathcal E$ \cite{bollobas1979graph,bollobas1985random}. We store this information in an association or weight matrix $\mathbf{W}$, whose $ij^{th}$ element indicates the strength or weight $w_{i,j}$ of the edge between node $i$ and node $j$. A simple way of building a graph from such an association matrix is to apply a threshold $\tau$ to each element of the matrix, such that if $w_{i,j} \geq \tau$ then an edge is drawn between the corresponding nodes, but if $w_{i,j} < \tau$ no edge is drawn \cite{achard2006resilient}. This thresholding operation thus binarizes the weight matrix and converts the continuously variable edge weights to either 1 (supra-threshold) or 0 (sub-threshold). It was on this basis that almost all brain graphs were constructed in the 15 years or so following the seminal small-world analysis of a binary graph representing the cellular connectome of {\em C. elegans} \cite{watts1998collective}. Most of the neuroimaging evidence for small-worldness in human brain networks, for example, is based on analysis of binary graphs constructed by thresholding a correlation coefficient or equivalent estimator of the weight of functional or structural connectivity or structural covariance between regions $i$ and $j$ \cite{wijk2010comparing}. It is well recognised that construction of binary graphs represents an extreme simplification of brain networks; indeed a binary undirected graph of homogenous nodes is as simple as it gets in graph theory \cite{bassett2012altered}. However, this approach has historically been preferred in neuroimaging because of limited signal-to-noise ratio in the data \cite{achard2006resilient}.
By varying the threshold $\tau$ used to construct a binary graph from a continuous weight matrix, the connection density of the network is made denser or sparser. If the threshold is low and many weak weights are added to the graph as edges then the connection density will increase; if the threshold is high and only the strongest weights are represented as edges, then the connection density will decrease. The connection density $D$ is quantified by the number of edges $E$ in the graph as a proportion of the total number of edges in a fully connected network of the same number of nodes $N$:
\begin{equation}
D = \frac{E}{N^2-N/2}
\end{equation} Often, this proportion is translated into a percentage. In many neuroimaging studies, the threshold is set to a large value to control for the high levels of noise in MRI data, resulting in connection densities in the range $5-30\%$ \cite{lynall2010functional}. In many of the first generation tract tracing studies, the connectivity data were collected on a binary or ordinal scale, and not all possible connections had been been experimentally measured, so these data were naturally modelled as binary graphs with connection densities $\sim 30\%$, a value that was constrained by the completeness and quality of the data \cite{bassett2006small}.
The small-world topology of a binary brain graph is defined by estimating two parameters in the data, path length $L$ and clustering $C$ (Fig.~\ref{cl_binary_weighted}A), and comparing each of these observed parameters to their distributions under a specified null model \cite{humphries2006brainstem}. More specifically,
\begin{equation}
L = \frac{1}{N} \sum l_{i,j}
\end{equation}
\noindent is the global or characteristic path length, where $l_{i,j}$ is the shortest path (geodesic) between nodes $i$ and $j$; and
\begin{equation}
C = \frac{1}{N} \sum c_{i,j}
\end{equation}
\noindent is the global clustering coefficient, where $c_{i,j}$ is the number of closed triangular motifs including node $i$. Each of these parameters is normalised by its value in a binary graph representing the null hypothesis. For example, if the null hypothesis is that clustering of brain networks $C_{brain}$ is no different from the clustering of a random graph, then it is reasonable to generate an Erd\"os-Reny\'i graph for $N$ nodes and $D$ connection density, measure the clustering coefficient in the random graph $C_{random}$, and use the ratio between brain and random graph clustering coefficients as a test statistic for non-random clustering. We note that there are many other possible ways in which a null model could be sampled, besides using the classical Erd\"os-Reny\'i model, and this is an active area of methodological research \cite{muldoon2016small,betzel2016modular}. However, in general one can define the normalized clustering coefficient as
\begin{equation}
\Gamma = \frac{C_{brain}}{C_{random}}.
\end{equation}
Likewise, the path length of the brain graph can be normalised by its value in a comparable random graph
\begin{equation}
\Lambda = \frac{L_{brain}}{L_{random}}.
\end{equation}
A small-worldness scalar can then be simply defined as
\begin{equation}
\sigma = \frac{\Gamma}{\Lambda}.
\end{equation}
With these definitions, small-world networks will have $\sigma > 1$, $\Gamma > 1$ and $\Lambda \sim 1$ \cite{humphries2006brainstem}.
\subsection*{Weighted graphs}
Although binary graph analysis has predominated to date in analysis of brain networks, this certainly does not represent the methodological limit of graph theory for connectomics. For example, provided the data are of sufficient quality, there is no need to threshold the weight matrix to estimate topological properties like clustering, path length and small-worldness. Indeed, while the binarization procedure was common in early applications of graph theory to neural data \cite{wijk2010comparing}, it remains fundamentally agnostic to architectural principles that may be encoded in edge weights \cite{rubinov2011weight}. This realization has more generally motivated the field to develop methods that remain sensitive to the patterns of weights on the edges \cite{ginestet2011brain}, and to the topologies present in weak \emph{versus} strong weights \cite{rubinov2011weight}. These efforts have included the development of alternative thresholding schemes \cite{bassett2012altered,lohse2014resolving} and fully weighted graph analysis \cite{rubinov2011weight,bassett2011dynamic}.
The mathematical tools exist to estimate and simulate the topological properties of weighted networks, and analysis of weighted networks is akin to studying the geometry of the graph, rather than simply its topology \cite{bassett2012influence,bassett2013task}. For example, weighted analogues of binary metrics of clustering, path length and small-worldness can be defined formally (Fig.~\ref{cl_binary_weighted}B). First, the weighted clustering coefficient of node $i$ can be defined as
\begin{equation}
C_{\mathrm{weighted}} = \frac{1}{k_i(k_i-1)}\sum\limits_{j,k}(\hat{w}_{ij}\hat{w}_{jk}\hat{w}_{ik})^{1/3},
\label{c_weighted}
\end{equation}
\noindent where $k_i$ is the number of edges connected to node $i$, or degree of node $i$ \cite{Onnela:2005de} (but see also \cite{Barrat:2004bk,Zhang:2005er} for other similar definitions). The weighted path length can be defined as
\begin{equation}
L_{\mathrm{weighted}}=\frac{1}{N(N-1)}\sum_{i\ne j}\delta_{ij},
\label{l_weighted}
\end{equation}
\noindent where the topological distance between two nodes is given by $\delta_{ij}=1/w_{ij}$ \cite{Newman:2001kc}. These two statistics can be combined to construct a weighted metric of small-worldness \cite{bolanos2013weighted}:
\begin{equation}
\sigma_{weighted}= \frac{\Gamma_{weighted}}{\Lambda_{weighted}}.
\end{equation}
With these definitions, small-world networks will have $\Gamma_{\mathrm{weighted}} > 1$, $\Lambda_{\mathrm{weighted}} \sim 1$ and $\sigma_{\mathrm{weighted}} > 1$, \cite{humphries2006brainstem}.
\subsection*{The small-world propensity}
\begin{figure*}[h]
\begin{center}
\includegraphics[width=0.6\textwidth]{figure_small_world_propensity.pdf}
\caption{ \textbf{Small-world propensity in weighted networks.} Here we illustrate an example of a generative small-world model, and its utility in estimating an empirical network's small-world propensity. \emph{(A)} We can extend the concept of a Watts-Strogatz model to weighted graphs by first building a lattice in which the edges are weighted by distance such that edges between spatially neighbouring nodes have more strongly weighted than edges between spatially distant nodes. These edge weights can then be rewired with a probability, $P$, to create a weighted small-world network. \emph{(B)} Weighted clustering coefficient and weighted path length can be estimated as a function of the rewiring parameter, $P$, and used to derive the small-world propensity of the graph compared to random and lattice benchmarks (Eq \ref{eq:SWP}). \emph{(C)} Weighted small-world propensity calculated for the same network as in panel \emph{(B)}. Error bars represent the standard error of the mean calculated over 50 simulations, and the shaded regions represent the range denoted as small-world. \emph{(D)} Weighted small-world propensity as a function of network density for a graph of 1000 nodes. Reproduced with permission from \cite{muldoon2016small}. \label{smallworldpropensity}}
\end{center}
\end{figure*}
There are several important limitations to the definitions of small-worldness described in the previous sections. First, the small-world scalar $\sigma$ (whether binary or weighted) can be greater than 1 even in cases when the normalized path-length is much greater than one; because it is defined as a ratio, if $\Gamma >> \Lambda > 1$, the scalar $\sigma > 1$. This means that a small-world network will always have $\sigma > 1$, but not all networks with $\sigma > 1$ will be small-world. Second, the measure is strongly driven by the density of the graph, and denser networks will more naturally have smaller values of $\sigma$ even if they are in fact generated from an identical small-world model. To address these and other limitations, Muldoon and colleagues recently developed a metric called the \emph{small-world propensity}. Specifically, the small-world propensity, $\phi$, reflects the deviation of a network's clustering coefficient, $C_{brain}$, and characteristic path length, $L_{brain}$, from both lattice ($C_{lattice}$, $L_{lattice}$) and random ($C_{random}$, $L_{random}$) networks constructed with the same number of nodes and the same degree distribution:
\begin{equation}
\label{eq:SWP}
\phi = 1-\sqrt{\frac{\Delta_{C}^{2}+\Delta_{L}^{2}}{2}},
\end{equation}
where
\begin{equation}
\Delta_{C}=\frac{C_{lattice}-C_{brain}}{C_{lattice}-C_{random}}
\end{equation}
and
\begin{equation}
\Delta_{L}=\frac{L_{brain}-L_{random}}{L_{lattice}-L_{random}}.
\end{equation}
\noindent The ratio $\Delta_{C/L}$ represents the fractional deviation of the metric ($C_{brain}$ or $L_{brain}$) from its respective null model (a lattice or random network). This quantity can be calculated for binary networks (using binary definitions of clustering and path length) or for weighted networks (using weighted definitions of clustering and path-length). Networks are considered small-world if they have small-world propensity $0.4 < \phi \leq 1$. However, this metric should be viewed as a continuous metric of small-worldness rather than a hard threshold \cite{muldoon2016small}.
Importantly, the small-world propensity overcomes several limitations of previous scalar definitions of small-worldness \cite{muldoon2016small}. First, it can incorporate weighted estimates of both the clustering coefficient and path-length, thus being generally applicable to any neural data that can be represented as a weighted network. Second, it is density-independent, meaning that it can be used to compare the relative small-worldness between two networks that have very different densities from one another. Third, the metric is informed by spatially-constrained null models \cite{expert2011uncovering,bassett2015extraction,papadopoulos2016evolution} in which nodes have physical locations and the edges that correspond to the smallest Euclidean distance between nodes are assigned the highest weights \cite{barthelemy2011spatial} (Fig \ref{smallworldpropensity}).
\section*{21st century tract-tracing}
The scale and quality of contemporary tract-tracing data, in both the macaque and the mouse, represents a quantitative change in terms of sensitivity in detecting anatomical connections, or axonal projections, between cortical areas. Using retrograde tracer experiments it has proven possible to demonstrate reliably that pairs of regions in the macaque brain may be connected by one or a few axonal projections. Likewise anterograde tracer experiments in the mouse have demonstrated that the minimal detectable weight of connectivity between cortical regions, that just exceeds the noise threshold, is equivalent to the projection of one or a few axons \cite{ypma2016}. This high sensitivity has led immediately to the recognition of a large number of weak and previously unreported axonal connections. In the macaque, it was estimated that 36\% of connections identified by contemporary tract tracing were so-called new found projections (NFP) that had not been described in the prior literature \cite{markov2014weighted}. The existence of so many weak connections is reflected in the log normal distributions of connectivity weight, ranging over 5--6 orders of magnitude, in both the macaque and the mouse \cite{oh2014mesoscale,ercsey2013predictive}. In short, tract-tracing can now resolve connections approximately equivalent to a single axonal projection and approximately a million times weaker than the strongest anatomical connections or white matter tracts.
How can we use graph theory to model the network organisation of such highly sensitive, highly variable data? Perhaps the simplest approach, borrowing from prior studies of less high quality datasets, is to apply a threshold and convert the log-normal weight matrix into a binary adjacency matrix. If the threshold is defined by the noise distribution of the measurements then it will be very close to zero for these sophisticated experiments, and correspondingly the connection density of the binary graph will be high. In the macaque, the connection density of a binary graph of 29 visual cortical areas was estimated to be 66\% \cite{markov2013cortical}, considerably higher than historical estimates in the range $25\%-45\%$ \cite{felleman1991distributed}. In the mouse, the connection density of a binary graph of 308 areas of the whole cortex was estimated to be 53\% \cite{rubinov2015wiring}.
In other words, the binary graphs generated from 21st century tract-tracing data are about twice as dense as the much sparser networks derived from human neuroimaging and 20th century tract-tracing. They are also considerably denser than brain networks constructed at a finer grained (ultimately cellular) resolution. For example, the connection density of the {\em C elegans} nervous system, which is still the only completely mapped synaptic connectome, is about 12\%. It is easy to see that the connection density of a binary network depends on the number of neurons comprising each node. In the limit, if the nervous system is parcellated into two large nodes the connection density will certainly be 100\%; as the same system is parcellated into a larger number of smaller nodes its connection density will monotonically decrease \cite{bassett2011conserved,zaleksy2010whole}. Thus the current interval estimate of mammalian cortical connection density $\sim 55-65\%$ is conditional both on the anatomical resolution of the parcellation scheme used to define the nodes and the sensitivity of the tract-tracing methods used to estimate the weights of the edges.
\begin{figure*}[h]
\begin{center}
\includegraphics[width=0.7\textwidth]{fig_sw_mouse_macaque_v5.pdf}
\caption{ \textbf{Binary and weighted small-worldness in mouse and macaque connectomes.} For the macaque connectome reported in \cite{markov2013cortical}, we show \emph{(A)} the binary network, a random graph of the same size and density, and the estimated small-world parameters $\Gamma$ (normalized clustering coefficient), $\Lambda$ (normalized path length), $\sigma$ (classical small-world scalar) and $\phi$ (small world propensity). In panel \emph{(B)} we show a weighted network analysis for the same data. For the mouse connectome reported in \cite{rubinov2015wiring}, we show \emph{(C)} the weighted network, a random graph of the same size and density, and the estimated small-world parameters $\Gamma$ (normalized clustering coefficient), $\Lambda$ (normalized path length), $\sigma$ (classical small-world scalar) and $\phi$ (small world propensity). In panel \emph{(D)} we show a binary network analysis for the same data. In the boxplots, the gray dotted line shows the threshold value of $\sigma = 1$, and the purple area shows the range of values of $0.4 < \phi \leq 1$ in which a network is considered small-world. \label{fig_sw_mouse_macaque}}
\end{center}
\end{figure*}
\subsection*{Small-worldness of binary tract-tracing networks}
Having constructed a high density binary graph from tract-tracing data on mammalian cortex, it is straightforward to estimate its clustering and path length, using the same metrics as for sparser binary graphs. However, simply because there is a larger number of connections in the denser network, its clustering will be considerably higher (there will be more closed triangular motifs) and its path length will be shorter (there will be more direct, pair-wise connections) than a sparser network. Indeed the clustering and path length of any binary graph at $60\%$ connection density will be close to the maximal clustering and minimal path length of a fully connected graph; and therefore the clustering and path length of a 60\% dense brain network will be very similar to the clustering and path length of a 60\% random network \cite{bassett2009cognitive}.
This means that when clustering and path length in brain networks are normalised by their corresponding values in equally dense random networks, the scaled metrics $\Gamma$ and $\Lambda$ will both be close to 1, and the small-world scalar $\sigma$ will be close to its critical value of 1 \cite{markov2013cortical}. For the macaque, at 66\% connection density, $\Gamma =1.21 \pm 0.014$, $\Lambda=1.00 \pm 0.000$, and $\sigma =1.21\pm0.014$; for the mouse, at 53\% connection density, $\Gamma =1.31 \pm 0.004$, $\Lambda=1.00 \pm 0.000$, and $\sigma =1.31 \pm 0.004$ (all given in mean $\pm$ standard deviation; Fig.~\ref{fig_sw_mouse_macaque}A,C; Table~\ref{tab1}). Since small-worldness has been traditionally defined as $\sigma > 1$, these results suggest that dense binary graphs constructed from tract tracing data are small-world, although the macaque is more similar to a random network than the mouse.
\begin{table*}[ht]
\hfill{}
\begin{tabular}{ |l|l|l|l|l|l|}
\hline
~& Macaque & ~ & ~ & Mouse & ~\\
\hline
~ & Binary & Weighted & ~ & Binary & Weighted \\
$\Gamma$ & $1.21 \pm 0.014$ & $1.59\pm0.007$ & ~ & $1.31 \pm 0.004$ & $1.76\pm0.009$ \\
$\Lambda$ & $1.00 \pm 0.000$ & $1.27 \pm 0.057$ & ~ & $1.00 \pm 0.000$ & $1.47\pm0.021$ \\
$\sigma$ & $1.21\pm0.014$ & $1.25 \pm 0.071$ & ~ & $1.31 \pm 0.004$ & $1.20\pm0.019$ \\
$\phi$ & N/A & $0.574\pm0.041$ & ~ & N/A & $0.800\pm0.002$ \\
\hline
\end{tabular}
\hfill{}
\caption{\textbf{Small-world metrics} For the macaque and mouse connectomes, we show the mean and standard deviation of the normalized clustering coefficient ($\Gamma$), the normalized path length ($\Lambda$), the small-world index $\sigma$, and the small-world propensity $\phi$ for both binary and weighted graphs. }
\label{tab1}
\end{table*}
These results do not look like a ``black swan'' that refutes universal claims that the brain always embodies small-world network topology. Nor do they undermine the credibility of previous studies demonstrating small-worldness in sparser brain graphs. However, our view is that binary graph models are very unlikely to be an optimal strategy for network analysis of tract-tracing data, because they fail to take account of the extraordinary range of connectivity weights, distributed log normally over 6 orders of magnitude, that has been discovered in mammalian cortical networks \cite{ercsey2013predictive}. The weakest connection between cortical areas is about a million times less weighted than the strongest connection: does it really make sense to set all these weights equivalently to 1 as edges in a binary graph? To ask the question is to answer it.
\subsection*{Small-worldness of weighted tract-tracing networks}
A weighted small-world analysis is easily done for these data (Fig.~\ref{fig_sw_mouse_macaque}B,D). The weighted clustering and weighted path length metrics (Eq.~\ref{c_weighted} and Eq.~\ref{l_weighted}) are estimated directly from the weight matrices, and the ratio of weighted clustering to weighted path length is the scalar summary of weighted small-worldness $\sigma_{\mathrm{weighted}} > 1$. In Fig.~\ref{fig_sw_mouse_macaque}, we directly compare binary and weighted graph theoretical results for the mouse \cite{oh2014mesoscale,rubinov2015wiring} and macaque \cite{markov2013cortical} connectomes. Compared to the results of binary graph analysis, both mouse and macaque networks have increased clustering for the weighted graph analysis, and $\sigma$ is increased for the macaque (see Table~\ref{tab1}).
The weighted graph of the mouse connectome is similarly small-world compared to the weighted macaque graph, as measured by $\sigma$, but is significantly more small-world as measured by the small-world propensity $\phi$. However, classical estimates of small-worldness may depend in a non-trivial way on the density of the graph. This relationship becomes obvious if we estimate the topology of both weighted graphs as a function of connection density (Fig.~\ref{fig_sw_mouse_macaque_density}). The classical small-world scalar $\sigma$ is greatest when it is estimated for a sparse graph comprising less than 20-30\% of the most strongly connected edges, and decreases progressively as the graph becomes denser. This might suggest that the macaque connectome seems less small-world than the mouse simply because it is denser. However, the small world propensity $\phi$ has the useful property that it is independent of network density and it is significantly greater, indicating more small-worldness, for the mouse than the macaque. This could be related to differences between the datasets in number of cortical areas and completeness of cortical coverage: the macaque dataset comprises fewer nodes of mostly visual cortex than the larger number of nodes across the whole mouse cortex.
\begin{figure*}[h]
\begin{center}
\includegraphics[width=1\textwidth]{fig_sw_mouse_macaque_density_v2.pdf}
\caption{ \textbf{Dependence of small-world characteristics on network density.} \emph{(A)} Macaque and \emph{(B)} mouse connectivity matrices in their natural state (\emph{left}), as well as after thresholding to retain the 5\% strongest (\emph{middle}) or 25\% strongest (\emph{right}) connections. Weighted small-world metrics including the normalized clustering coefficient ($\Gamma$), normalized path-length ($\Lambda$), small-world index ($\sigma$), and small-world propensity ($\phi$) as a function of network density for the \emph{(C)} macaque and \emph{(D)} mouse connectivity matrices. \label{fig_sw_mouse_macaque_density}}
\end{center}
\end{figure*}
\subsection*{Weighted small-worldness and the role of edge weights}
Why does a weighted graph analysis provide stronger evidence for non-random clustering than a binary graph analysis applied to the same tract-tracing data?
The most strongly weighted connections generally span the shortest physical distances between cortical areas \cite{klimm2014resolving,ercsey2013predictive,rubinov2015wiring}. This is not surprising based on what we know about the importance of cost constraints on brain organisation \cite{bassett2010efficient,bullmore2012economy,bassett2009cognitive,fornito2011genetic}. Strong connectivity weights indicate a large number of axonal projections, a big bandwidth bundle, perhaps macroscopically visible as a white matter tract. Building and resourcing a high bandwidth axonal signalling bundle is a significant biological cost that will increase as a function of connection distance: it is parsimonious to wire high bandwidth over short distances. Short distance connections are not only strongly weighted but also topologically clustered. So the strongest weights in both cortical networks define a topologically segregated and anatomically localised organisation. A map of the sub-network formed by the strongest weights shows spatial and topological clusters of regions (Fig.~\ref{WeakLinks}). In the mouse, the strongly weighted clusters each comprise functionally specialised areas of cortex (visual, motor, etc.) that are known to be densely inter-connected and anatomically localised \cite{rubinov2015wiring,ypma2016}. Thus it is not surprising that weighting the topological analysis of mammalian cortical networks will provide stronger evidence for non-random clustering than unweighted analysis of binary graphs.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=1\textwidth]{SWR_Fig2.pdf}
\caption{ \textbf{The existence of weak links and their topology in the mouse connectome.} Here we show the properties of the 5\% strongest and 5\% weakest edges of the mouse cortical network. \emph{(A,B)} Axial view of the mouse cortical network, red dots represent brain regions, blue lines represent the connections between them. Drawn are the \emph{(A)} 5\% weakest or \emph{(B)} 5\% strongest edges. Dot size corresponds to degree, the total number of incoming and outgoing edges connected to a node. In \emph{(B)}, the three nodes with highest degree have been labeled: VISp, primary visual area; MOp, primary motor area; SSs, supplemental somatosensory area. The strong connections are spatially organized, mainly connecting spatially adjacent or contralaterally homologous regions. The weak connections span longer distances and are topologically more random than the strongest connections. \emph{(C)} The distance distributions for (blue) the 5\% weakest edges, (red) the 5\% strongest edges, and (black) a random graph of the same size and connection density. \emph{(D)} The degree distributions for the weakest and strongest connections of the mouse connectome, and a comparable random graph, color-coded as
in panel \emph{(C)}. Reproduced with permission from \cite{ypma2016}. \label{WeakLinks}}
\end{center}
\end{figure*}
The most weakly weighted connections are an area of active, ongoing research (discussed in more detail below) and it is inevitable that there is still much to learn about a feature of network organisation -- replicable but very weak connections between large cortical areas -- that had not been measurable until recent advances in tract-tracing methodology. However, it is clear that weaker connections tend to subtend longer distances, and can be either more topologically random than \cite{ypma2016} or similarly topologically organized to \cite{bassett2012altered} strong connections.
We conclude that graph theoretical analysis of tract-tracing connectomes should respect the quality of the data and use weighted topological metrics to reflect the wide ranging variation in anatomical connectivity, from single fibres to major tracts, that is now measurable in the mammalian brain \cite{wang2016brain}. Weighted graph analysis demonstrates clearly that both the macaque and mouse connectomes are small-world networks, as are the human, cat, and nematode \cite{muldoon2016small}. Binary graph analysis has usefully measured high connection density, due to the existence of many new anatomical connections, but binarization of these data is not the best way to understand their complex topology and its economical embedding in anatomical space \cite{rubinov2011weight,bassett2011conserved,klimm2014resolving,bassett2012altered,rubinov2015wiring}. Future studies will likely also pay more attention to the fact that most tract-tracing markers are axonally transported only in one direction: anterograde or retrograde. This means that the weight matrix could be modelled more completely as a weighted and directed graph, representing a further evolution in the use of graph theoretical methods to capture a richer and biologically more meaningful model of brain network organisation than can be provided by binary graphs of unweighted and undirected edges.
\section*{The utility of weak connections}
At this juncture, one might naturally ask: ``From a neuroscientific perspective, do we need techniques that account for edge weights? Do these weights indeed capture information of relevance for cognition and behavior?'' Neuroanatomical data suggest that the weights of structural connections may be driven by developmental growth rules \cite{klimm2014resolving,lohse2014resolving,kaiser2006nonoptimal,ercsey2013predictive,markov2013cortical}, energetic and metabolic constraints \cite{bassett2010efficient}, and physical limitations on the volume of neural systems, particularly brains encapsulated by bone \cite{sherbondy2009think}. Yet the role of these edge weights in neural computations \cite{schneidman2006weak} and higher order cognition has been less well studied.
Recent studies have begun to elucidate the role of edge weights -- and particularly of weak connections -- in human cognition. In resting state fMRI data, weak functional connections from lateral prefrontal cortex to regions within and outside the frontoparietal network have been shown to display individual differences in strength that predict individual differences in fluid intelligence \cite{cole2012global}. The same general relationship was observed in a separate study in which individual differences in moderately weak, long-distance functional connections at rest were strongly correlated with full scale, verbal, and performance IQ \cite{santarnecchi2014efficiency}. Neither of these correlations were observed when considering strong connections. Indeed the utility of weak edges appears to extend to psychiatric illness, where the highly-organized topology of weak functional connections -- but not strong functional connections -- in resting state fMRI were able to classify people with schizophrenia from healthy controls with high accuracy and specificity \cite{bassett2012altered}. Interestingly, individual differences in these weak connections were significantly correlated with individual differences in cognitive scores and symptomatology. Together these results demonstrate that, indeed, methods that are sensitive to the strength (or weakness) of individual connections are imperative for progress to be made in understanding individual differences in cognitive abilities, and their alteration in psychiatric disease.
Importantly, the utility of weak connections is not only evident at the large scale in human brains, but also at the neuronal scale as measured in non-human species. In an influential paper published in 2006 with Bialek and colleagues, Schneidman demonstrated that weak pairwise correlations implied strongly correlated network states in a neural population, suggesting the presence of strong collective behaviour \cite{schneidman2006weak}. This result was initially counter-intuitive as one might expect that weak correlations would be associated with the lack of collective behavior. However, the original observation has withstood the test of time, and has been validated in several additional studies including work at the level of tract tracing in macaque monkeys \cite{goulas2015strength}. Intuitively, the juxtaposition of weak correlations and cohesive, collective behavior is thought to be driven by the underlying sparsity of neuronal interactions \cite{ganmor2011architecture}, which contain a few non-trivial higher-order interaction terms \cite{ganmor2011sparse}. Indeed, these higher-order interactions are the topic of some interest both from a computational neuroscience perspective \cite{giusti2016twos,sizemore2016classification}, and from the perspective of neural coding \cite{giusti2015clique}.
But perhaps the claim that weak connections are critically important for our understanding of neural systems should not be particularly surprising. Indeed, it is in fact an old story, first published at the inception of network science. In 1973, Granovetter wrote a seminal paper, titled ``The strength of weak ties'', which highlighted the critical importance of weakly connected components in global system dynamics \cite{granovetter1973strength}. Such weak connections are ubiquitous in many systems, from physician interactions \cite{bridewell2011social} to ecosystem webs \cite{ulanowicz2014limits} and atmospheric pathways \cite{lee2014tracking}. Looking forward, critical open questions lie in how these weak connections drive global dynamics, and how one can intervene in a system to manipulate those processes \cite{gu2015controllability,betzel2016optimally,muldoon2016stimulation}.
Acknowledging the role of weak connections, weighted small-world organization plays a critical role in system functions that are particularly relevant to neural systems: including coherence, computation, and control and robustness \cite{novkovic2016topological}. Perhaps the most commonly studied function afforded by small-world architecture is the ability to transmit information, a characteristic that is common in networks of coupled oscillators \cite{barahona2002synchronization,hong2002synchronization,nishikawa2003heterogeneity} (although see \cite{atay2006synchronization} for a few notable exceptions). This capability supports enhanced computational power \cite{lago2000fast}, via swift flow and transport \cite{hwang2010spectral}. In dynamic networks, oscillators coupled on small-world networks are much more sensitive to link changes than their random network counterparts \cite{kohar2014synchronization}, the time taken to reach synchronization is lowered, and the synchronized state is less stable over time, potentially enabling greater diversity of function. When such a system has both small-world topology and geometry, it directly impacts the network's ability to speed or slow spreading \cite{karsai2011small}, a potentially useful characteristic for resilience to dementia which is thought to be caused by the spread of prions \cite{raj2012network,raj2015network}.
The value of small-world architecture is not limited to its support of synchronization and information flow. Instead, it also supports a wide-range of computations in neural circuits. From early neural network studies, it is clear that the exact topology of connectivity patterns between network elements directly supports tradeoffs in the network's ability to learn new information \emph{versus} retain old information in memory \cite{hermundstad2011learning}. When these patterns are organized in a small-world manner, evidence suggests that local computations can be integrated across distributed cell assemblies to support functions as diverse as somatosensation \cite{zippo2013neuronal} and olfaction \cite{imam2012implementation}. The mechanism by which small-worlds support these computations may stem from the fact that their topological structure tends to contain both large cavities and high-dimensional cliques \cite{sizemore2016classification}, which when embedded in a physical space can strongly constrain the geometric properties of the computation \cite{giusti2015clique}.
While small-world structure can offer non-trivial advantages in terms of both communication and computation, it also directly informs the sorts of interventions that one could use to guide network dynamics and by extension system function. Indeed, computational studies have demonstrated that small-world network architecture requires specific control strategies if one wishes to stem the propagation of seizure activity \cite{ching2012distributed}, control the spread of viruses \cite{kleczkowski2012searching}, or enhance recovery following injury \cite{hubler2008mathematical}. To gain an intuition for how topology impacts control, we can consider the broad-scale degree distribution also characteristic of brain networks. Based on the Laplacian spectrum, one can observe that weakly connected nodes have the greatest potential to push the system into distant states, far away on an energy landscape \cite{pasqualetti2014controllability}; conversely, strongly connected hubs have the greatest potential to push the system into many local states, nearby on the energy landscape \cite{gu2015controllability}. Thus, control energy (such as that provided by brain stimulation) may be targeted to different locations in a small-world brain network to affect a specific change in brain dynamics \cite{muldoon2016stimulation}.
\section*{Conclusions}
Small-worldness remains an important and viable concept in network neuroscience. Nearly twenty years on from the first analysis of the complex topology of a binary graph representing the nervous system of {\em C. elegans}, it has been established that small-worldness is a nearly-universal and functionally valuable property of nervous systems economically embedded in anatomical space. Recent advances in tract tracing connectomics do not refute small-worldness; rather they considerably enrich and deepen our understanding of what it means in the brain. The extraordinary precision of contemporary tract tracing, and the important discovery that mammalian cortical networks are denser than expected, mandates the adoption of more sophisticated techniques for weighted graph theoretical modelling of inter-areal connectomes. On this basis, we expect the next ten years to yield further insights into the functional value of weak as well as strong connections in brain networks with weighted small-worldness.
\section*{Acknowledgements and disclosures}
We thank Rolf Ypma and Evelyn Tang for comments on an earlier draft of this manuscript, and for Jonathan Soffer for assistance. ETB is employed half-time by the University of Cambridge and half-time by the University of Cambridge; he holds stock in GSK. DSB acknowledges support from the John D. and Catherine T. MacArthur Foundation, the Alfred P. Sloan Foundation, the Army Research Laboratory and the Army Research Office through contract numbers W911NF-10-2-0022 and W911NF-14-1-0679,the National Institute of Mental Health (2-R01-DC-009209-11), the National Institute of Child Health and Human Development (1R01HD086888-01), the Office of Naval Research, and the National Science Foundation (BCS-1441502, BCS-1430087, PHY-1554488, and BCS-1631550).
\bibliographystyle{naturemag}
|
\section{Introduction}
A hadronic atom consists of a negatively-charged hadron
(e.g., $\pi^-$, $K^-$, $\bar{p}$, $\Sigma^-$, $\Xi^-$)
and electrons that are bound by a Coulomb field to an atomic nucleus.
Such a system can be used to probe the strong interaction
between hadrons and atomic nuclei in the low-energy limit.
The mass of the hadron,
significantly larger than the mass of the electron it replaces,
shifts the atomic transition energies in the hadronic atom to
significantly higher energies than those of the standard atom.
In addition, effects of the strong interaction
appear in the most tightly bound energy level.
These perturbations include an energy shift from that given only by the
electromagnetic interaction, and a lifetime broadening
due to absorption of the hadron by the nucleus.
The shift and width can be experimentally
extracted via x-ray-emission spectroscopy
of characteristic transitions into this lowest orbital.
In kaonic atoms,
the understanding of the low-energy $\overline{K} N$ interaction
has been substantially deepened
by the most recent kaonic-hydrogen-atom measurement \cite{SidKp}
and theoretical studies (e.g., \cite{Ikeda}).
However, the depth of the $\overline{K}$-nucleus potential
remains unknown
because of insufficient precision in the kaonic-atom data for $Z\ge2$;
this is one of the greatest present concerns
in strangeness nuclear physics \cite{Gal13}.
In recent years, Balmer-series x-rays of the $K^-$-He atom ($\sim$ 6 keV)
were measured at the K5 beamline of KEK-PS \cite{E570}
and the DA$\Phi$NE electron-positron collider
of the Laboratori Nazionali di Frascati \cite{SidHe4,SidHe3}.
These measurements followed discussions of the importance of the measurement of
the strong-interaction shift and width of the $2p$ level
in kaonic-helium atoms ($K^-$-He) over 30 years ago \cite{Bai83,Aka05}.
Both experiments employed Silicon Drift Detectors (SDDs)
whose FWHM energy resolution is typically $\sim$ 200 eV
at 6 keV \cite{E570,SidHe4,SidHe3}.
To observe spectral effects
as predicted by theoretical calculations \cite{Hir00,Bat90,Fri11},
e.g., a 0.2 eV shift and 2 eV line width,
a new approach with a significantly improved resolution is essential.
We are preparing a high-resolution x-ray measurement
of kaonic atoms at a kaon beamline of J-PARC \cite{K1.8BR}
using a novel superconducting transition-edge-sensor (TES)
microcalorimeter (J-PARC E62 \cite{E62}).
The FWHM energy resolution of this type of detector can be
as good as about 2 eV at 6 keV \cite{Smith12,Uhlig15}
which is about two orders of magnitude better than that of SDDs.
The spectrometer is a highly sensitive thermal sensor
that measures energy deposition
via the increase in the resistance
of a superconducting thin film that is biased
within the sharp phase transition
between the normal and superconducting phases.
The detailed working principles and the recent progress
of the TES system are described in Refs. \cite{Ens05,Ull15,Uhlig15}.
An alternate high-resolution x-ray spectroscopy technology
is based on diffraction from Bragg crystals.
However, they have not been used to study the strong interaction
in most hadronic atoms due to their low efficiency.
Only pion beams and the resulting x-ray emission are
intense enough to enable the study of pionic atom x-rays \cite{Mar76,Cha85,Sig95}
with crystal spectrometers.
On the other hand,
recent technological advances in multiplexed readout
of multi-pixel TES arrays
(more than 100 pixels; each having $\sim$0.1 mm$^2$ effective area)
allow the performance of a precision kaonic-atom measurement
in a realistic data acquisition time.
Until this experiment, a TES x-ray spectrometer has never been utilized
in a hadron-beam environment.
To study the in-beam performance of the TES spectrometer
at a hadron beamline and demonstrate the feasibility
of TES-based hadronic-atom x-ray spectroscopy,
we performed a pioneering experiment at a pion beamline
by measuring the x-rays from the $4f \to 3d$ transition of
pionic carbon ($\pi$-$^{12}$C).
This x-ray transition was chosen because the energy ($\sim$ 6.5 keV)
is similar to the $K^-$-He $3 \to 2$ x-ray energy,
while the strong-interaction effects are negligibly small.
\section{Experiment}
\begin{figure}[t]
\centering\includegraphics[width=1.\linewidth]{fig_setup.eps}
\caption{A top view (left) and birds-eye view (right)
of the experimental setup.}
\label{fig_setup}
\end{figure}
The experiment was carried out at the $\pi$M1 beamline \cite{piM1}
of the Paul Scherrer Institute (PSI) in October 2014.
Pions were created by a proton beam of energy 590 MeV at a current
of up to 2.3 mA passing through a graphite target located
in the main beamline of the PSI synchro-cyclotron,
and were transported through the $\pi$M1 beamline.
A schematic view of the experimental setup is shown in
Fig. \ref{fig_setup}.
With a 2.2 mA primary proton beam,
the $\pi^-$ beam rate with momentum 173 MeV/$c$
was 1.45 $\times$ 10$^6$ /sec
at BC1 (10$\times$10 cm$^2$ plastic scintillation counter)
located at a distance 30 cm upstream from the focal point.
The $\pi^-$/$e^-$ ratio was $\sim$0.7
due to a missing electrostatic separator in the beamline.
The incident $\pi^-$'s were degraded in moderators,
counted with beamline counters (BCs),
and stopped in a carbon graphite target
where $\pi$-$^{12}$C atoms were generated.
X-rays were emitted by the highly excited atoms
and detected by a TES x-ray spectrometer
through a 150$\mu$m-thick beryllium vacuum window
and three layers of 5-$\mu$m-thick aluminum IR-blocking filters
(one at each of the three temperature stages: 50K, 3K, 50 mK).
A 240-pixel TES array \cite{Ull14} was the x-ray spectrometer.
Each pixel had a 4-$\mu$m-thick bismuth absorber
(80\% absorption efficiency for 6.4 keV x-rays)
which converted the incident x-ray energy to heat;
the absorber was coupled to a sensitive thermal sensor, the TES,
composed of a superconducting Mo/Cu proximity bilayer film.
Each absorber had a collimated effective area
of 320 $\mu$m $\times$ 305 $\mu$m,
thus the total collecting area of the array was about 23 mm$^2$.
Each TES pixel was electrically biased
to the superconducting critical temperature ($T_\mathrm{C} \sim$ 107 mK)
and to about 20 - 30\% of their normal resistance of $\sim$10 m$\Omega$.
A pulse-tube-backed adiabatic
demagnetization refrigerator (ADR) \cite{HPD}
cooled the system to a regulated bath temperature
of 75 mK $\pm$ 7 $\mu$K (rms).
The regulated hold time was about 36 hours,
after which the ADR cycle
(magnetic field increased isothermally and decreased adiabatically) took 2 hours
before another 36 hours of operation at 75 mK.
SQUID current amplifiers were used to read out the current signal
from the low-resistance TESs.
For the 240-pixel readout,
a time division SQUID multiplexing (TDM) scheme \cite{TDM} was employed
to reduce the number of wires running to the low temperature stages
of the cryocooler.
The signal from each TES channel was coupled to a SQUID amplifier,
and the outputs from 30 individual channels were switched sequentially
and read out by a single amplifier.
The multiplexing frame time was 9.6 $\mu$s (0.32 $\mu$s/channel).
The sampling rate was therefore 104 kHz for each pixel.
The 240 pixel readout was realized by use of eight TDM columns in parallel.
TES data were continuously streamed into a PC server,
and were recorded to disk
only when a current ``pulse'' due to an x-ray event was triggered.
The system recorded 1024 samples (9.83 ms) for each event,
where the first 256 samples corresponded to a pre-triggered timing region.
The typical exponential rise and decay time constants of an x-ray pulse were
$\sim$200 $\mu$sec and $\sim$500 $\mu$sec respectively.
The x-ray energy corresponding to each pulse height was calculated
with an optimal-filter technique \cite{OptimalFilter}.
The data-acquisition system recorded also
a stopped-$\pi$ trigger timing
defined by beamline counters
as BC1 $\otimes$ BC2 $\otimes$ BC3 $\otimes$
$\overline{\mbox{BC4}}$,
where the pulse-height threshold of BC2 was carefully set
in order to select only stopped-pion events which deposited more energy
than an electron, a muon or an in-flight pion event.
The relative timing of this triggering system to the TES events
was reconstructed during the offline analysis.
Precise absolute energy calibration
was crucial for this measurement.
Because independent energy calibration was necessary for each of the 240 pixels,
it was essential to supply intense calibration x-rays
to achieve sufficient statistics for $in$-$situ$ calibration.
An x-ray tube source was installed
as shown in Fig. \ref{fig_setup}.
Characteristic x-rays were produced by shining
a controllable flux of x-rays,
which was generated by an electron gun with a Rh target,
on 99.999\% pure chromium and cobalt pieces.
These calibration x-rays traveled through the hollow,
conical carbon target to the TES spectrometers.
Figure \ref{fig_calib} shows
a summed x-ray energy spectrum of 209 working TES pixels
obtained with the x-ray generator without a pion beam.
The achieved energy resolution was 4.6 eV (FWHM) at 6.4 keV
with a count rate of 4.4 Hz/pixel.
The energy calibration was performed
with the four calibration x-ray lines,
Cr $K_\alpha$, Cr $K_\beta$, Co $K_\alpha$, and Co $K_\beta$,
with natural cubic-spline interpolation.
Energies and natural widths of the calibration x-rays
were fixed with the reference values \cite{Holzer1997}
in the spectral fitting.
Lower-yield x-rays, Fe $K_\alpha$ (6.4 keV) and Cu $K_\alpha$ (8.0 keV),
originated respectively from stainless steel vacuum fittings in the tube source
and Cu inside of the detector package. These were not used for calibration.
\begin{figure}[hbtp]
\centering\includegraphics[width=0.65\linewidth]{fig_calib.eps}
\caption{An x-ray energy spectrum obtained from a tube-source x-ray generator
shining on Cr and Co calibration pieces without a pion beam.
The Cr $K_\alpha$, Cr
$K_\beta$, Co $K_\alpha$, and Co $K_\beta$ lines are used for the
energy calibration.
Lower-yield x-rays, Fe $K_\alpha$ (6.4 keV) and Cu $K_\alpha$ (8.0 keV),
originate respectively from the stainless steel vacuum fittings
of the tube source and the metal structure of the 50 mK detector package.}
\label{fig_calib}
\end{figure}
\section{Results}
Time and energy distributions of $\pi$-$^{12}$C x-rays
measured with the 209 working TES pixels
are shown in Fig. \ref{fig_result}.
The data were accumulated for 13.5 hours
with a pion-beam intensity of 1.45 MHz
and a stopped-$\pi$ trigger rate of 34.5 kHz.
Figure \ref{fig_result} (b) shows a distribution of the time difference
between pion arrival and x-ray detection with the TES array
for stopped-$\pi^-$ trigger events.
A clear peak was observed with a time resolution of 1.2 $\mu$sec FWHM.
Figure \ref{fig_result} (a) shows a correlation plot of the time difference
versus the x-ray energy measured by TESs.
At around 6.43 keV, a clear time-energy correlation
corresponding to $\pi$-$^{12}$C $4f \to 3d$ x-ray was observed.
Figure \ref{fig_result} (c) shows an x-ray energy spectrum
of only those events recorded within the
stopped-$\pi^-$ timing gate ($\pm$1.5 $\mu$sec)
that is indicated in Fig. \ref{fig_result} (b).
A sharp peak from the $\pi$-$^{12}$C $4f \to 3d$ transition
is observed.
The TES spectrometer has sufficient energy resolution to observe
the parallel transition $4d \to 3p$.
Figure \ref{fig_result} (c) shows contributions from
those $4f \to 3d$ and $4d \to 3p$ lines whose centroids are only 7.4 eV apart.
The timing selection improved the peak-to-background ratio
of the $\pi$-$^{12}$C line from 2 to 10.
As a reference in this experiment,
a SDD is installed just beneath the carbon target
as shown in right panel of Fig. \ref{fig_setup}.
The FWHM energy resolution of the SDD of $\sim$ 165 eV
is not nearly enough to resolve any of
the important features of the spectrum (Fig \ref{fig_result} (d)).
\begin{figure}[hbtp]
\centering\includegraphics[width=0.85\linewidth]{fig_result.eps}
\caption{Results of the measured x-ray time
and energy distributions for stopped-$\pi^-$ trigger events.
(a) A correlation plot of
the time difference between pion arrival and x-ray detection vs
the x-ray energy measured by the TES array.
(b) The projection on the time axis
showing timing resolution of 1.2 $\mu$s (FWHM).
A time gate of $\pm$ 1.5 $\mu$s is used in the analysis.
(c) The projection on the energy axis
by selecting stopped-$\pi^-$ time gate indicated in (b),
where the fitted components for Fe $K_\alpha$ and
$\pi$-$^{12}$C x-rays
are shown as well.
(d) The same spectrum measured by the reference SDD
having a FWHM energy resolution of $\sim$165 eV.}
\label{fig_result}
\end{figure}
The tube source produces Cr and Co calibration x-rays constantly
during data acquisition with the pion beam.
The energy calibration with Cr and Co x-rays is recalculated
every 2 hours to mitigate gain drift.
A fit of the energy spectrum with the in-beam condition of $\pi^-$ beam
gives a FWHM energy resolution
of 6.8 eV at 6.4 keV
with a TES hit rate of 4.8 Hz/pixel, of which 0.4 Hz is
due to the pion beam.
The deterioration of the energy resolution
and the size of low-energy and high-energy tail components
in the non-Gaussian energy response of TESs
both increase as a function of pion beam intensity.
These beam-correlated effects are explained
by the production of thermal cross-talk pulses
which are due to direct charged-particle hits
in the silicon substrate of the TES array.
The detail of the detector response and the absolute energy calibration was
presented in a separate paper \cite{LTD16}.
The x-ray generator fluoresces materials
containing iron (e.g., the stainless steel vacuum fittings around the target cell);
therefore characteristic x-rays of iron
are observed that are uncorrelated with the pion-beam timing.
The Fe $K_{\alpha1}$ (6.404 keV)
and Fe $K_{\alpha2}$ (6.391 keV) lines
were used to evaluate the accuracy of energy calibration
that was then used to determine $\pi$-$^{12}$C x-ray energy.
A result of a spectral fit of Fe $K_\alpha$ and $\pi$-$^{12}$C x-rays
is shown in Fig. \ref{fig_result} (c).
The energy-calibration accuracy is assessed
by the fit to the measured Fe $K_{\alpha}$ line.
The measured energy of our Fe $K_{\alpha11}$ line is:
\begin{eqnarray*}
6404.07 \pm 0.10 \mbox{(stat.)} ^{+0.06}_{-0.04} \mbox{(syst.) eV}
\end{eqnarray*}
where the first error is statistical and the second is systematic.
The quoted systematic uncertainty is a quadratic summation
of the contributions from continuum background parameter
and asymmetry of the fit function.
A comparison with the reference value
of 6404.148(2) eV \cite{Holzer1997}
for pure, metallic iron shows good agreement within the errors.
To determine the energy of pionic carbon transition x-ray,
the critical issue is the uncertainty precisely at the lines of interest, 6430 eV.
It is difficult to assess the systematic uncertainty
introduced by our choice for the calibration curves in its full generality.
Fortunately, the Fe $K_{\alpha}$ line at 6404 eV is very close to the line of interest,
and this good agreement with the reference value validated the choice
of the present energy calibration curves.
Fits to our energy spectra determined
the energies of the
$\pi$-$^{12}$C $4f \to 3d$ and $4d \to 3p$ transition x-rays
and their yield ratios to be:
\begin{eqnarray*}
E (4f \to 3d) &=& 6428.39 \pm 0.13 \mbox{(stat.)} \pm 0.09 \mbox{(syst.) eV} \\
E (4d \to 3p) &=& 6435.76 \pm 0.30 \mbox{(stat.)} ^{+0.11}_{-0.07} \mbox{(syst.) eV} \\
I(4d \to 3p) / I(4f \to 3d) &=& 0.30 \pm 0.03 \mbox{(stat.)} \pm 0.02 \mbox{(syst.)}
\end{eqnarray*}
where the first error is statistical and the second is systematic.
The quoted systematic uncertainty is a quadratic summation
of the contributions from uncertainties of
the energy calibration,
the non-Gaussian response function,
and the timing window width.
The tail component of Fe $K_{\alpha}$ line affects the fit results.
We assessed it by varying the relative strength of the Fe x-ray tail
by changing the timing window. The Fe tail is the main source
of systematic errors both for x-ray energies and yield ratio.
The energy dependence of transmissions of these two $\pi$-$^{12}$C peaks
is negligible to determine the yield ratio,
since these energies are very close and no absorption edge structure
exists around those peaks.
\begin{table}[t]
\caption{Calculated values of pionic $^{12}$C electromagnetic
energies and strong-interaction energy shifts
via the Seki-Masutani potential \cite{SM83}.
For the $4f \to 3d$ and $4d \to 3p$ transitions,
electron-screening effects are assessed
with the cases of filling one $1s$ electron
and two $1s$ electrons in the K-shell.
The experimental results are shown as well.}
\label{table_calc}
\centering
\begin{tabular}{cccccccc}
\hline
\hline
State & K.G. &
\multicolumn{2}{c}{Vacuum polarization} &
Nuclear & Relativistic & Strong & Total \\
& energy &
$\alpha(Z\alpha)$ & $\alpha^2(Z\alpha)$ &
finite size & recoil effect & interaction & energy \\
& (eV) & (eV) & (eV) &
effect (eV) & (eV) & effect (eV) & (eV) \\
\hline
$3p$ & $-$14685.15 & $-$ 11.56 & $-$0.08 & $+$ 0.01 & $-$0.02 & $-$0.78 & $-$14697.58 \\
$3d$ & $-$14682.65 & $-$ 5.39 & $-$0.04 & $+$ 0.0005 & $-$0.02 & $<10^{-4}$ & $-$14688.10 \\
$4d$ & $-$8259.04 & $-$ 2.10 & $-$0.02 & $+$0.0003 & $-$0.01 & $<10^{-4}$ & $-$8261.17\\
$4f$ & $-$8258.59 & $-$ 0.72 & $-$0.004 & $+$0.0003 & $-$0.01 & $<10^{-4}$ & $-$8259.32\\
\hline
\hline
\end{tabular}
~\\
\vspace{5mm}
\centering
\begin{tabular}{ccccl}
\hline
\hline
Transitions & \multicolumn{3}{c}{Electron screening effect (eV)} & Transition \\
& Configuration & K-shell & L-shell & energy \\
& & contribution & contribution & (eV) \\
\hline
& no electron & - & - & 6428.78 \\
$4f \to 3d$ & $1s^1$ $2s^2$ $2p^1$ & $-$0.19 & $-$0.02 & 6428.57 \\
& $1s^2$ $2s^2$ $2p^1$ & $-$0.31 & $-$0.01 & 6428.46 \\
\cline{2-5}
& \multicolumn{3}{r}{Experimental result (this work) :} & 6428.39 $\pm$ 0.13 $\pm$ 0.09 \\
\hline
& no electron & - & - & 6436.41 \\
$4d \to 3p$ & $1s^1$ $2s^2$ $2p^1$ & $-$0.25 & $-$0.02 & 6436.14 \\
& $1s^2$ $2s^2$ $2p^1$ & $-$0.42 & $-$0.01 & 6435.98 \\
\cline{2-5}
& \multicolumn{3}{r}{Experimental result (this work) :} & 6435.76 $\pm$ 0.30 $^{+0.11}_{-0.07}$ \\
\hline
\hline
\end{tabular}
\end{table}
We have calculated the $3p$, $3d$, $4d$ and $4f$ energy levels
of the pionic $^{12}$C
using only the electromagnetic (EM) interaction
and tabulated the results in Table \ref{table_calc}.
These EM values are calculated from the Klein-Gordon equation
including vacuum polarization with higher-order correction,
the relativistic-recoil effect and the nuclear-finite-size effect.
The latest charged pion mass, 139.57018(35) MeV/$c^2$,
given by the particle data group \cite{PDG} is used.
Strong-interaction effects on these energy levels
were assessed via the Seki-Masutani potential \cite{SM83}.
The strong-interaction shift of the $3p$ level from its EM value
is calculated to be 0.78 eV,
while the shifts of the $4f$, $4d$, $3d$ levels are below $10^{-4}$ eV.
The effect of the electrons populating the K- and L-shells,
the so-called electron-screening effect, is not negligible
especially in the solid target.
The electron-screening correction to the x-ray energy is largely
determined by the number of $1s$ electrons in the atom, which depends on
the balance between Auger-electron emission and the electron-refilling process
from neighboring atoms.
We have estimated the electron screening effect
with the ``$Z-1$'' approximation, namely the captured pion
screens one unit of the nuclear charge seen by the electrons.
Thus, the maximum electron-screening correction is given
by the $1s^2$ $2s^2$ $2p^1$ electronic configuration.
In Table \ref{table_calc},
the EM values of the $4f \to 3d$ and $4d \to 3p$ transitions
are tabulated for the no electron,
$1s^1$ $2s^2$ $2p^1$ and $1s^2$ $2s^2$ $2p^1$ electronic configurations.
Here, the electron-density functions in each configuration are evaluated
using the hydrogenic wave function with effective nuclear charge
based on Hartree-Fock calculations.
The experimental results are consistent
with the $1s^2$ $2s^2$ $2p^1$ configuration within the errors.
\section{Conclusion}
We observed the $\pi$-$^{12}$C $4f \to 3d$ transition x-ray line
with a novel 240-pixel microcalorimetric x-ray detector
based on transition edge sensors.
The achieved averaged energy resolution is
6.8 eV FWHM at 6.4 keV under a high-rate pion beam intensity of 1.45 MHz.
The timing resolution is 1.2 $\mu$sec FWHM.
Absolute energy calibration is realized
by an x-ray generator shining on calibration metals during the data acquisition.
The resulting systematic uncertainty in the $\pi$-$^{12}$C $4f \to 3d$ transition x-ray energy
is less than 0.1 eV,
which meets our goal for a future measurement
of the kaonic-helium $3d \to 2p$ x-ray energy.
The TES spectrometer had sufficient energy resolution to observe
the parallel transition $\pi$-$^{12}$C $4d \to 3p$ x-ray for the first time.
The strong-interaction effect of the $3p$ level is not negligible
because it has smaller angular momentum than the $3d$ level.
The measured x-ray energy of the parallel transition $4d \to 3p$
was found to be consistent with the calculated strong-interaction effect
assessed via the Seki-Masutani potential \cite{SM83}.
Our data allow the determination of
the electron population status of the atoms.
Both the $4f \to 3d$ and $4d \to 3p$ transition energies obtained
favor two $1s$ electrons in the K-shell.
Moreover we have determined the yield ratio
between the $4f \to 3d$ and $4d \to 3p$ transitions.
Whether the observed x-ray yields can be consistently explained
by a cascade calculation with the electronic configuration
of two filled $1s$ electrons in the K-shell
remains to be solved in future theoretical work.
We successfully demonstrated the feasibility of hadronic-atom x-ray spectroscopy
with an absolute energy uncertainty of 0.1 eV in the 6 keV energy region.
This is an important milestone towards a more general use of high-resolution
microcalorimeter spectrometers at charged-particle beamlines.
\ack
The authors thank
K. Deiters and the PSI staff for beamline coordination and operation.
J. Uhlig thanks the Knut and Alice Wallenberg Foundation for their continued support.
This work was partly supported by RIKEN, NIST, KEK,
a Grants-in-Aid for Scientific Research
from MEXT and JSPS (Nos. 25105514, 26707014, 24105003,
15H05438 and 15H00785),
the strategic young researcher overseas visits program for
accelerating brain circulation by JSPS (No. R2509),
Incentive Research Grant from RIKEN,
Mitsubishi Foundation (26145),
and the NIST Innovations in Measurement Science Program.
|
\section{Introduction}
A key prediction of the concordance model of the universe is the
existence of structure in the dark matter distribution at nearly all
scales, limited only by the intrinsic temperature of the candidate
dark matter particle at the smallest scales and causality at the
largest scales \citep{2005PhR...405..279B}. At the galactic scale this
structure manifests as bound subhalos orbiting within the
gravitational potential of their host, with the largest mass
fraction in subhalos generally occuring near the virial radius of the
main halo \citep{Springel2008}. However it is also clear that many of
these subhalos, if they exist, must not host luminous galaxies, since
their number in cold-dark-matter simulations greatly
exceeds the number of satellite galaxies observed around the Milky Way
(MW) and other galaxies in nature, the so-called missing satellite
problem \citep[and Moore 1999]{1999ApJ...522...82K}. Solutions to this
discrepancy fall into two broad categories: those that modify the
properties of dark matter to raise the minimum size of substructures
to match observed satellites, and those in which the dark matter is cold (CDM)
and baryonic processes select some fraction of subhalos to
host galaxies. Solutions that adjust the baryons rather than the dark
matter thus also predict the existence of many ``dark subhalos''
devoid of baryonic matter, orbiting nearly invisibly in galactic
halos. Determining whether these dark subhalos exist is thus an
important test to discriminate between CDM and other models of dark
matter.
One possible observable signature of dark subhalos is the imprint they
leave on stellar tidal streams, the unbound remnants of accreted
satellite galaxies. These satellite galaxies initially inhabit the
subhalos that managed to retain their gas and form stars before
accreting onto their host galaxy. Many tidal streams have already
been found \citep{1999HelmiNat, 2001ApJ...547L.133I,
2001ApJ...554L..33V, 2003AJ....126.2385O, 2003ApJ...588..824Y,
2004ApJ...605..575Y, 2006ApJ...642L.137B, 2006ApJ...645L..37G,
2006ApJ...636L..97D, 2006ApJ...651L..33L, 2006ApJ...639L..17G,
2006ApJ...643L..17G, 2007ApJ...658..337B, 2009ApJ...698..865K,
2009ApJ...693.1118G, 2012ApJ...760L...6B, 2013ApJ...765L..39M} and
still more are thought to exist
\citep[e.g.][]{1999Helmi,Helmi2011,Gomez2013}. Tidal streams are
kinematically cold compared to the total phase-space distribution of
their host, since they come from less massive progenitors, and
their stars orbit as test particles in the host after being
unbound by tides from their progenitors. A tidal stream will thus
become even colder as it ages, since its stars conserve their total
phase-space volume and slight differences in their orbits cause their
positions to spread out over time \citep[e.g.][]{1999Helmi}. This
elongation of the volume in physical space makes streams sensitive probes of
the lumpiness of the mass distribution, since interactions with dark
subhalos will disturb the long (10s to 100s of kpc), thin (physical
width of 10s to 100s of pc, velocity dispersion of a few km
\unit{s}{-1}) stream of stars
\citep{Johnston2002,Ibata2002,Mayer2002,Penarrubia2006,Carlberg2009,Yoon2011}. The
small velocity spread and width of a stream mean that even small
disturbances can be detected, while its long length gives it a large
cross-section to interactions with the more isotropic population of
dark substructure. Additionally, the arrangement of the stream stars
along neighboring orbits will magnify alterations over time, since a
local encounter with a subhalo will affect a subset of stream stars
with similar orbital phases. Thus for example a ``gap'' in a stream
caused by scattering from a subhalo will widen over time as the
discrepancy between orbits at the edges of the gap increases
\citep{Carlberg2009,Yoon2011}.
Previous work on stream-subhalo interactions has fallen into two categories. \citet{Carlberg2009,2013ApJ...775...90C,2015ApJ...800..133C,2015ApJ...808...15C} and \citet{2015arXiv150705625E} considered the effects of these encounters on idealized streams on relatively circular orbits to determine the characteristics of their signatures and the mass range of subhalos likely to create observable gaps. Other work by \citet[][hereafter YJH]{Yoon2011}, \citet{Carlberg2012}, \citet{2014ApJ...788..181N}, and \citet{2015ApJ...803...75N} inserted simulated globular-cluster-like streams into dark matter halos based on cosmological simulations to estimate how often such interactions would occur in CDM halos. However, there has been no attempt to date to self-consistently measure the interaction rate between streams from accreted luminous substructures and nonluminous subhalos generated from a single cosmological simulation. That is the goal of this work.
To calculate the predicted frequency of both streams and subhalos, we use one of the Aquarius simulations of a Milky-Way-like dark matter halo \citep{Springel2008}, a portion of the Millenium II dark-matter-only cosmological simulation \citep{2009MNRAS.398.1150B} that has been resimulated at higher resolution. Specifically we use the simulation of Halo Aq-A-2 ($M_{200} = 1.842 \times 10^{12}\ M_{\odot}$) at resolution level 2 ($m_p = 1.37 \times 10^4\ M_{\odot}$). This simulation resolves dark substructures down to about $10^6\ M_{\odot}$ at size scales of about 0.3 kpc. To create stellar streams from some of the substructures, we tag some of the dark matter particles as ``stars" according to the procedure described in Section \ref{sec:retagging}. We then follow one such set of stars, which at the present day in the simulation forms a long thin stream, and track the close encounters between these stars and the many untagged dark subhalos in each snapshot of the simulation, from the infall of the stream's progenitor to the present day (Section \ref{sec:exampleStream}) and examine the rate of interactions and the evolution of the stream in phase space and constants of motion. Finally, we select 18 streams that still appear long and thin at the present day and track their interactions with subhalos in the same way to calculate an average interaction rate for a stream of a given age and present day length (Section \ref{sec:intxnStats}). In Section \ref{sec:discussion} we compare the rate to estimates from the literature and in Section \ref{sec:concl} we summarize our findings.
\begin{figure}
\centering
\includegraphics[width=0.43\textwidth]{size-vs-luminosity-v2-Aq-A-2-0-1-10-1}
\caption{Luminosity-velocity dispersion and luminosity-size relations of Aq-A-2 ``satellites'' still bound at the present day (green points, produced by tagging the 1\% most bound DM particles at infall) compared to known MW (red) and M31 (blue) satellites \citep{2015ApJ...799L..13C}. The green lines show multiples of the softening length used in the dark-matter-only N-body simulation. Compare with Figure 4 of \citet{2010MNRAS.406..744C}. The dark grey and light grey regions are surface brightness limits from SDSS and PandAs, respectively.}
\label{fig:lfunc}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{example_xy.png}
\caption{Time-evolution of stream C in both tagging schemes: the one used in Cooper et al. (black) and the 1-percent-at-infall scheme used in this work (green). The dots show the projected locations of the tagged particles relative to the center of the main halo, measured in kpc. The projected axes are the simulation $x$ and $y$ ($y$ is roughly aligned with the major axis of the halo; $x$ lies approximately along the minor axis). Time increases from left to right and then top to bottom, in pairs of panels labeled ``Cooper" and ``1\%" for the two tagging schemes. In the panels labeled ``Cooper," a random selection of 1241 out of the 2056 total tagged particles is plotted for each snapshot so that the same number of particles is plotted for each tagging scheme (the random selection is different in each snapshot). In the two panels on the top left, the circle shows the virial radius of the main Aq-A-2 halo (in the later snapshots the virial radius is larger than the frame).}
\label{fig:taggingComparisonXY}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{example_rvr.png}
\caption{As in Figure \ref{fig:taggingComparisonXY}, but showing the galactocentric radial velocity (in km \unit{s}{-1}) of the tagged particles in stream C as a function of galactocentric distance (in kpc).}
\label{fig:taggingComparisonRV}
\end{figure*}
\section{Streams in the tagged Aquarius simulations}
\label{sec:retagging}
Initially we used the scheme described in \citet{2010MNRAS.406..744C} to tag a subset of the dark-matter particles in the Aquarius subhalos as a stellar component, but we found that the assignment of stellar mass to dark matter particles in this scheme can be very lumpy; that is, one dark matter particle can end up with orders of magnitude more stellar mass assigned to it than its neighbor. This makes it difficult to interpret ``gaps'' in the surface brightness of the simulated streams. Because the tagging proceeds throughout the simulation, it is also unclear whether this method evenly samples the phase space of each subhalo. So we also tried a different, much simpler tagging method that assigns stellar mass more evenly to the DM particles.
\subsection{Tagging at infall}
In our new method we also start with the dark-matter-only Aq-A-2 simulation, but tag dark matter particles in subhalos with stellar mass by selecting the most bound 1\% of particles in each subhalo at the snapshot where it is first associated with the main halo in the friends-of-friends tree generated by the halo finder {\sc subfind} \citep{2001MNRAS.328..726S}. We refer to this snapshot as the ``infall'' snapshot. This strategy is similar to that followed by e.g. \citet{DeLucia2008} and \citet{2012ApJ...746..109L}. The amount of associated stellar mass is calculated using the semianalytic model of \citet{2013MNRAS.429..725S}, and is divided evenly between all selected DM particles. For tagged subhalos that are still bound at the present day in the simulation, this produces the luminosity--size relation shown in the lower panel of Figure \ref{fig:lfunc}. We tried several different tagging percentages up to 10\%, and found that a 1\% tagging fraction most closely matched the size-luminosity relation of the MW and M31 satellites down to the resolution limit of the simulation (where the half-light radius is comparable to a few times the softening length). The luminosity--velocity dispersion relation is also roughly matched with this tagging fraction.
\begin{figure*}
\begin{center}
\includegraphics[width=\textwidth]{streamO_gal.png}
\caption{View of stream O in ``galactic'' coordinates, for an arbitrary location 8 kpc from the center of the host halo. The size of the points is inversely proportional to distance (closer stars look bigger) while the line-of-sight velocity (negative = approaching) determines color.}
\label{fig:OrphanSky}
\end{center}
\end{figure*}
Tagging methods for representing the building blocks of the stellar halo are to some extent unsatisfactory, since they assume that the phase-space distribution of the stars and dark matter deep within the subhalo potential are identical. Recent work by \citet{2015arXiv150206371L} and \citet{2014ApJ...783...95B} points out discrepancies between the distributions of stars in particle-tagging schemes and those generated by SPH simulations. Some of these discrepancies are due to the way in which tagging proceeds, while others vary with the choice of feedback prescriptions in the subgrid SPH physics, which can either reconcile or exacerbate the differences. However, while neither representation of the streams resulting from tagged DM-only satellites is likely to be correct in detail, we do produce objects that agree with the luminosity-size relationship of the structures that are still bound at present day, which is the best available observational constraint.
We use a lower tagging percentage than \citet{2015arXiv150206371L},
who found that tagging the most-bound $\sim$5 percent of particles in
subhalos over the course of a dark matter (DM) only simulation
produced the best agreement with hydrodynamic (SPH) simulations. They
hypothesize that a larger tagging region is needed to properly account
for differences in the dark matter particle diffusion rate between the
hydro and DM-only cases (since, for example, the hydro simulations
form a disk that can shock satellites during mergers). However, their
agreement criteria differ from ours; we simply wish to properly
reproduce the observed luminosity-size and luminosity-velocity
relations of the bound MW satellites, while
\citeauthor{2015arXiv150206371L} focus mainly on the structure of the
unbound halo stars and the density profiles of the few largest simulated
satellites. Furthermore we tag at infall rather than over time, so
that the percentages being tagged are not directly
comparable. \citeauthor{2015arXiv150206371L} show that the difference
between tagging 1\% and 5\% of particles is most pronounced in the
massive satellite density profiles, and is much smaller in the unbound
halo structure. On the other hand, \citet{2011MNRAS.418..336L} focused on comparing the
radial profiles of observed stellar halos with those obtained through
tagging and found that tagged fractions of 1--3\% in the
dark-matter-only CLUES simulations could reproduce the observed
stellar halo profile depending on whether the MW, M31, or M33 was
considered, with the lower end of that range being appropriate for the
Milky Way. These authors also looked at CLUES runs including SPH and
found that in these cases a larger tagging fraction (3--5\%) was
required to match the observations, suggesting that matching SPH to
DM-only runs should give different results for the best tagging
fraction than matching either to observations. Additionally, the
choice of tagging fraction will primarily influence the length of the
resulting streams, which (as we will demonstrate in this work) spans a
very wide range, with far more variety than could be produced simply
by manipulating the tagging fraction within the range recommended by
all these studies. We therefore argue that our tagging choices are sufficient for our
purposes and consistent with the studies most similar to our
situation.
\begin{figure*}
\includegraphics[width=\textwidth]{xy-8057455-2.png}
\caption{Spatial view of the evolution of stream O with time. The first panel shows the progenitor of the stream at infall. The black points are the ``stars'' (tagged particles) making up the stream. The position axes are centered on the host galaxy center, marked with a blue cross. Frames outlined in green include an interaction with a subhalo at distance less than $r_{1/2}$; in these panels the $\mathcal{S}$ value defined by Equation \eqref{eq:strongEncRatio} is shown in green. }
\label{fig:OrphanXYsnaps}
\end{figure*}
\begin{figure*}
\includegraphics[width=\textwidth]{rvr-8057455-2.png}
\caption{Phase-space view of the evolution of stream O with time. The first panel shows the progenitor of the stream at infall. The black points are the ``stars'' (tagged particles) making up the stream. The radius and radial velocity axes are centered on the host galaxy. As in Figure \ref{fig:OrphanXYsnaps} the green-framed panels mark snapshots where a subhalo is within $r_{1/2}$. }
\label{fig:OrphanRVRsnaps}
\end{figure*}
\subsection{Comparison to Cooper et al. tagging scheme}
Both our simplified tagging scheme and the method used by \citeauthor{2010MNRAS.406..744C} match the properties of the present-day bound satellites reasonably well, but this agreement does not rule out possible differences in the distributions of tagged particles in the unbound structures. Differences are to be expected, since in \citeauthor{2010MNRAS.406..744C}'s scheme particles are tagged over time, and some dark matter particles could pass into and out of the most-bound region being tagged one or more times. This led us to expect that our single-snapshot tagging would lead to a consistent size of bound structures but without a more diffuse region of tagged particles that were in the subhalo's center at earlier times but not at infall. It was also unclear how much of the apparent lumpiness of the streams in the Cooper tagging scheme had to do with the tagging method, as opposed to structure either in the subhalos themselves or produced in the streams through encounters in the lumpy main halo.
Since we use a different semi-analytic model for our tagging scheme than \citeauthor{2010MNRAS.406..744C}, the particular subhalos chosen for tagging will generally not be the same between the two schemes, especially at lower masses where reionization and/or feedback can suppress star formation in some subhalos but not others. However, most of the highest-mass subhalos should be tagged in both cases and we can use these to compare the appearance of streams in our new scheme with those in \citeauthor{2010MNRAS.406..744C}. Figures \ref{fig:taggingComparisonXY} and \ref{fig:taggingComparisonRV} show one such stream that illustrates the differences and similarities between the tagging methods. We will refer to this as stream ``C,'' since we use it to {\it C}ompare the two tagging methods. The total stellar mass in this stream is $1.9\times 10^5\ M_{\odot}$ in the Cooper tagging scheme, and $4.6 \times 10^6\ M_{\odot}$ in the new tagging scheme. In both cases the progenitor crosses the virial radius as a very compact object (top left pair of panels) that only begins to tidally disrupt several Gyr after infall. Thus the stream is formed relatively recently even though its progenitor joins the main halo very early. In the figure, we plot the particles tagged using each of the two schemes.
The Cooper version of stream C includes about twice as many tagged
particles as the version in our new tagging scheme (2056 in Cooper
versus 1241 in the new scheme). To highlight differences in the
position and velocity of the tagged particles, we downsample the
Cooper stream by randomly choosing the same number of particles that
are tagged in our new scheme. Once this is done the stream is
virtually identical in the two tagging schemes. Figure
\ref{fig:taggingComparisonXY} shows that in the Cooper scheme there
are a few more outliers and slightly longer tails of tagged particles;
we think these represent particles that were only briefly deeply bound
within the subhalo in the past. However, the substructure within the
stream, visible in both position (Figure
\ref{fig:taggingComparisonXY}) and velocity (Figure
\ref{fig:taggingComparisonRV}), persists regardless of how the
particles are tagged, confirming that it is not an artifact of the
tagging process. As expected, our use of a different SAM also does not
affect our results, since it influences only our calibration of what
percentage of particles to tag as stars, which is set by comparing
with observed bound satellites. The only other way we use the SAM
information is to select streams by stellar mass for further analysis,
but since the simulated satellites are calibrated to match the luminosity function of the
Milky Way satellites \citep{Koposov2008,2013MNRAS.429..725S}, and because we use a wide
range of stellar masses ($10^4$--$10^6\ M_\odot$), we
expect our results not to be sensitive to details of the semi-analytic
model.
\begin{figure*}
\includegraphics[width=\textwidth]{infallDistribution.png}
\caption{Phase-space distribution of the progenitor of stream O at infall. Blue points show the locations of the tagged DM particles relative to the center of mass, in position and velocity. The histograms show the binned one-dimensional distributions of the stars in all 6 phase-space coordinates. All positions are given in kpc; all velocities are in km \unit{s}{-1}.}
\label{fig:InfallDistribution}
\end{figure*}
\FloatBarrier
\section{Subhalo interactions with an example stream}
\label{sec:exampleStream}
To see how subhalo interactions affect the streams we followed a single stream over time and tracked the number and mass of subhalos that approached the stream within several different threshold distances. We also looked at the evolution of the energy and angular momentum distribution of the stream ``stars'' (tagged DM particles) over time. The stream we tracked resembles the MW's Orphan Stream \citep{2006ApJ...645L..37G,2007ApJ...658..337B}; it has about the same stellar mass ($7.1\times 10^5$ \ensuremath{M_{\odot}}\ in Cooper, $9.7\times10^5$ \ensuremath{M_{\odot}}\ in our SAM), and its thickness (a few tens of pc) and angular span (about 1 full wrap) at the end of the simulation are similar. We will refer to it as stream ``O'' (for Orphan). Figure \ref{fig:OrphanSky} shows stream O ``on the sky", for an arbitrary choice of viewpoint 8 kpc from the center of the host halo. The color indicates the line-of-sight velocity, while the size is inversely proportional to the distance (so that closer stars look larger). A total of 857 DM particles in the progenitor subhalo were tagged with stellar mass using the Cooper et al. method, and 673 were tagged using our new method. The dark matter mass of the subhalo at the time of infall was $4.7\times10^8$ \ensuremath{M_{\odot}}.
\subsection{Initial distribution of the progenitor}
We followed stream O from the point at which its progenitor first becomes a subhalo ($t=1.23$ Gyr) until the end of the simulation ($t=13.58$ Gyr). The stream is well established by $t=9.4$ Gyr. It begins to form at about $t=4$ Gyr, after a few pericenter passages within 20 kpc of the galactic center. As the stream evolves it develops gaps and other density perturbations, features that are consistent with previous studies of the effects of encounters between the stream and dark substructure. In particular we see perturbations both in physical space (Figure \ref{fig:OrphanXYsnaps}) and projected phase space (Figure \ref{fig:OrphanRVRsnaps}) for stream O. However, it is important to determine whether these structures were indeed caused by such encounters, whether they arise from pre-existing structure in the stream progenitor, or whether they result from resolution effects. Thus in Figure \ref{fig:InfallDistribution} we examine the distribution of the tagged star particles in the first snapshot where the progenitor is a subhalo of the main (``Milky Way'') halo, hereafter referred to as the ``infall" snapshot. It appears from the figure that no prominent phase-space correlations exist in the progenitor; the distribution in every coordinate is smooth and roughly Gaussian. If this is so, then the smaller-scale structures apparent in the stream must have been produced after infall. However, some ``holes" are barely distinguishable in the distribution: one example is in the upper left-hand panel of Figure \ref{fig:InfallDistribution} at about (-0.05,0.05). This sort of feature may be due to shot noise in sampling the distribution, and so could also lead to an apparent gap if the noise is correlated in position and velocity. Therefore we wish to confirm that features on this level are consistent with the expected Poisson noise given the number of star particles in the stream.
To confirm quantitatively that the stream starts as a smooth distribution (down to the shot-noise limit) we calculated the mutual information of the distribution in phase space at the infall snapshot. The mutual information compares a distribution $p(\vect{w})$ to the product of its marginals $P(w_i)$:
\begin{equation}
\mathcal{M}[p(\vect{w})] \equiv \int p(w) \log \frac{p(\vect{w})}{\prod_{i=1}^{N_d}P(w_i)} d\vect{w},
\end{equation}
where $N_d$ is the number of dimensions of $\vect{w}$. In our case, $p(\vect{w})$ represents the phase-space density of the progenitor ($\vect{w}$ represents the six-dimensional phase-space position of a particle), normalized by the number of particles so that it integrates to unity. If the distribution is purely a smooth, multivariate Gaussian in all dimensions, with correlation matrix $\mathbf{\rho}$, then we expect $\mathcal{M}$ to be \citep[][Chap. 9]{Kullback59}
\begin{equation}
\sub{\mathcal{M}}{g} = -\frac{1}{2} \log |\mathbf{\rho}|.
\end{equation}
In the case of additional substructure within the distribution, $\mathcal{M}>\sub{\mathcal{M}}{g}$. To test whether additional substructure exists, we first calculated $\mathbf{\rho}$ for our distribution and obtained an analytic value for $\sub{\mathcal{M}}{g}$ of 0.15. Then we used the EnBID density estimator \citep{2006MNRAS.373.1293S} to calculate $\mathcal{M}$ for the infall distribution of the 673 star particles in stream O by Monte Carlo integration, obtaining a value of 0.31$\pm 0.02$, about a factor of 2 higher than the analytic estimate. However, previous tests have shown that at small $\mathcal{M}$ this numerical estimator tends to be biased high for samples with a low number of points, so we also constructed 100 random samples of 673 points from a multivariate Gaussian distribution with the same mean and $\mathbf{\rho}$ and estimated $\mathcal{M}$ from these samples, obtaining a value of 0.35$\pm 0.04$, statistically identical to the value for the actual phase-space distribution. Thus we can safely conclude that there is no more structure present in the infall distribution than would be expected for a smooth multivariate Gaussian.
\subsection{Frequency, velocity, and mass of interacting subhalos}
The effect of a subhalo encounter on a stream can be assessed, at zeroth order, by determining whether its distance of closest approach to the stream, $b$, brings it close enough that the potential energy of the nearest stream stars with respect to the subhalo (of mass \sub{M}{sub}) is comparable to the kinetic energy of the stream stars, which move at a velocity \sub{v}{rel} in the rest frame of the subhalo. Thus the condition for a ``strong'' encounter can be expressed as
\begin{equation}
\frac{\sub{GM}{sub}}{b} \geq \frac{\sub{v}{rel}^2}{2},
\label{eq:strongEncounterCondition}
\end{equation}
where $G$ is Newton's constant. One could therefore define a dimensionless ratio, the ``strength" $\mathcal{S}$ of the encounter, characterized by the scaling of Equation \ref{eq:strongEncounterCondition}:
\begin{equation}
\mathcal{S} = \frac{2G\sub{M}{sub}}{\sub{v}{rel}^2 b},
\end{equation}
so that strong encounters have $\mathcal{S} \gtrsim 1$. However, this criterion is quite conservative compared to what would actually be needed to induce a local gap in a stream: an $\mathcal{S} \sim 1$ interaction could for example reverse the direction of motion for stars in the stream, where what is actually needed to produce a gap in a coherent stream is to change the velocities of a small group of stream stars by an amount comparable to the stellar velocity dispersion, which is usually a few percent of $\sub{v}{rel}$ (i.e. a few km/s). This likewise corresponds to a few-percent change in the kinetic energy of the stars in the stream, so we choose to rescale $\mathcal{S}$ by the fractional velocity change $\eta$,
\begin{equation}
\mathcal{S}_{\eta} = \frac{2G\sub{M}{sub}}{\eta \sub{v}{rel}^2 b},
\end{equation}
and set $\eta=0.01$ so that an $\mathcal{S} \sim 1$ interaction changes the velocity of the interacting stream stars by one percent. Thus we define the quantity $\mathcal{S}_{1\%}$, using
\begin{equation}
\mathcal{S}_{1\%} = 100 \cdot \frac{2G\sub{M}{sub}}{ \sub{v}{rel}^2 b},
\label{eq:strongEncRatio}
\end{equation}
as a way to tell how likely an interaction is to open a gap in the stream.
By following stream O through the simulation, we can track the encounters it experiences and ask how many of them fit this basic criterion. To determine the type and rate of interactions between stream O and subhalos, we counted the number of subhalos in each snapshot for which the most-bound particle in the subhalo passed within $b=1$, 2, and 5 kpc of any star particle in the stream. We also counted how many subhalos passed within one or two times their half-mass radius, $r_{1/2}$, reflecting the region of influence in which we expect a subhalo to significantly alter the orbits of the stream stars. We tracked the identification numbers of the subhalos interacting with the stream to determine whether there were any repeat encounters with the same subhalo in different snapshots.
\begin{figure}
\includegraphics[width=0.5\textwidth]{fig-7-8057455.png}
\caption{Number of interactions between subhalos and any tagged star particle in stream O, as a function of time, for different interaction thresholds $b$. Only interactions after infall ($t=1.23$ Gyr; dotted vertical line) are shown. The time when the stream starts to form (identified as described in Section \ref{subsec:lengthAge}) is shown as a red dashed vertical line. In the top three panels, the dot-dashed line indicates the $t^2$ scaling posited by YJH, roughly fit by eye.}
\label{fig:OrphanIntxAll}
\end{figure}
\begin{figure}
\includegraphics[width=0.5\textwidth]{streamO-present-day-vectors.png}
\caption{Stars in stream O (black) and all subhalos within 2 kpc at present day (colored points with arrows). Each subhalo's velocity vector is shown. Subhalos are colored according to absolute relative velocity, and their size is proportional to the ratio given by Equation \eqref{eq:strongEncRatio} with $b$ the closest distance to any star in the stream. Values of this ratio closer to 1 imply stronger encounters. }
\label{fig:closeEncounters}
\end{figure}
The number of subhalos interacting with stream O in each snapshot, at the various distance thresholds, are plotted in Figure \ref{fig:OrphanIntxAll}. The $r_{1/2}$ threshold is not shown as there is only one interaction within this threshold over the course of the simulation, at 12 Gyr with a $10^7 M_\odot$ subhalo. This is not too surprising given that for most of the interacting subhalos, $r_{1/2}$ is much smaller than 1 kpc. As expected, the interaction rate increases with time as the stream increases in length, and roughly follows the $t^2$ scaling predicted by YJH (dot-dashed lines), beginning at the start of the formation of the stream at $\sim$6 Gyr. The number of encounters starts to fall beneath this scaling at late times, especially in the 2 and 5 kpc panels. We think this is due to particle resolution effects, which we discuss in more detail in Section \ref{subsec:resolution}. We do not see any pattern in the number of events with time for the $2r_{1/2}$ threshold (bottom panel), which is consistent with noise. Particle and time resolution probably combine to cause us to miss many of these encounters, so based on this result we decided to analyze encounters for the subhalo-dependent thresholds only in aggregate, and do not calculate rates for individual streams.
Figure \ref{fig:OrphanIntxAll} illustrates clearly that in any given snapshot after stream O is formed, it is interacting with many subhalos: up to 6 at a time within 1 kpc and up to 26 at a time within 2 kpc. Figure \ref{fig:closeEncounters} shows the stream with all subhalos within 2 kpc at present day. The velocities of the subhalos do not appear to be correlated with the velocity of the stream, implying that these subhalos did not simply fall in at the same time as the stream progenitor on similar orbits and follow the stream through space, but are passing though the stream with many different impact angles. This is also illustrated when we examine the interaction data for repeat encounters: one would expect these to occur frequently if the subhalos responsible for the interactions fell in with the stream along similar orbits. However, Figure \ref{fig:velmasstime} shows that all but one of the $\sim450$ interactions inside 2 kpc are one-time events.
We recorded the relative velocities of the interacting subhalos with respect to the stream stars within the threshold distance, in order to rank encounters by their strength and determine whether the field of perturbers is truly isotropic. For each subhalo interaction we identified the \sub{N}{close} stars within the threshold distance and calculated the average vector velocity $\vect{\bar{v}}$ of those stars component-by-component:
\begin{equation}
\vect{\bar{v}} = \frac{1}{\sub{N}{close}} \sum_{j=1}^{\sub{N}{close}} \vect{v}_{j}.
\end{equation}
Then we determined the relative velocity by subtracting this from the velocity of the subhalo:
\begin{equation}
\sub{\vect{v}}{rel} = \sub{\vect{v}}{sub} - \vect{\bar{v}}.
\end{equation}
The vector relative velocity was tabulated for every interaction in every snapshot.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{fig-9-8057455.png}
\caption{Encounters of stream O with subhalos as a function of time and speed relative to the stream, for the thresholds shown in Figure \ref{fig:speedDist}. Subhalos that have more than one encounter with the stream are shown in red. One-time encounters are shown in black; the encounter shown in blue in the bottom panel is within $r_{1/2}$. Only encounters after infall are shown. As in Figure \ref{fig:closeEncounters}, the size of the circle is proportional to the value of the ratio $\mathcal{S}_{1\%}$ in Equation \eqref{eq:strongEncRatio}. The infall time is shown as a vertical dotted line; the time when the stream starts to form (identified as described in Section \ref{subsec:lengthAge}) is shown as a red dashed vertical line.}
\label{fig:velmasstime}
\end{figure}
\begin{figure*}[!p]
\begin{center}
\includegraphics[width=\textwidth]{EnergyWithPhase-8057455-instantaneousL-2.png}
\caption{Energy versus phase angle for particles in stream O as a function of time. The red line marks the mean energy at infall (upper left panel). As in Figures \ref{fig:OrphanXYsnaps} and \ref{fig:OrphanRVRsnaps} the green frames mark snapshots with interactions at $b<r_{1/2}$.}
\label{fig:energyOrphan}
\end{center}
\end{figure*}
Figure \ref{fig:velmasstime} illustrates how, although many subhalo interactions do occur, most of them are likely too weak to substantially alter the stream. The size of the symbols in the figure correspond to the strength $\mathcal{S}_{1\%}$ of the encounter as defined in Equation \ref{eq:strongEncRatio}. The relative velocity \sub{v}{rel} is determined using particles within the impact parameter $b$, so $\mathcal{S}_{1\%}$ can vary slightly for different impact parameters. For the fixed distance thresholds, the mass distribution of subhalos in encounters reflects the steepness of the halo mass function in general, so one would expect most of the encounters to be very weak but also very frequent. Figure \ref{fig:velmasstime} shows that numerous relatively weak encounters are generally the rule over the lifetime of the stream. Only a handful of subhalos come within $2r_{1/2}$ of the stream over its lifetime, mostly more massive subhalos since these have larger half-mass radii. As shown in the bottom panel, selecting encounters with $b\leq 2r_{1/2}$ is similar (though not identical) to selecting encounters with large $\mathcal{S}_{1\%}$, which are those most likely to open a gap in the stream.
\subsection{Energy and angular momentum evolution}
\label{subsec:EandL}
The encounters between stream O and subhalos also leave an imprint on its energy and angular momentum distribution. Based on the results of YJH and the work of Carlberg and collaborators \citep{Carlberg2009,CarlbergConf2012,Carlberg2012,2013ApJ...775...90C,2014ApJ...788..181N,2015ApJ...800..133C,2015ApJ...803...75N,2015ApJ...808...15C}, we expect that the strongest subhalo encounters will rearrange part of the stream's energy-angular-momentum distribution. To investigate this, we followed the energy over time (including a computation of the potential energy directly from the N-body snapshot) as a function of the relative phase of the stream particles, as shown in Figure \ref{fig:energyOrphan}. To compute the phase angle $\phi$, we rotate the stream into the plane perpendicular to its mean angular momentum $L$, then compute the angle $\tan \phi \equiv y'/x'$, where $x'$, $y'$ are coordinates in the rotated frame. The angular momentum direction is not constant in time, as one would expect given that the potential is triaxial, and in fact it changes by of order 90 degrees over the course of the simulation. Since this change is large, we use the angular momentum of the stream in a given snapshot to compute the phase angle, so the orbital plane in each panel of the figure is different. The average energy at infall (red horizontal line) is roughly preserved as a function of time.
At $t$=10.2 Gyr long tidal tails have already formed around the remaining bound structure, and show signs of disturbance especially along the ends. There is also some structure that appears disconnected from the main energy distribution, around $E \sim -1.3\times 10^5\ (\textrm{km}\ \unit{s}{-1})^2$; this material was energetically detatched from the main body at around 8.2-8.3 Gyr. From about 11--11.5 Gyr the tails appear to rotate vertically in this view; Figure \ref{fig:OrphanXYsnaps} shows that this is because most of the stream is near apocenter.
The closest encounter (within $r_{1/2}$) is with a $10^5 M_\odot$ subhalo at 11.7 Gyr. The time-resolution of the snapshots prevents us from drawing a direct connection between either of these encounters and specific changes in the energy distribution of the stream stars, but the types of discontinuous features seen in the figure are reminiscent of those found by YJH in their simulations of a globular-cluster-like stream. Suggestively, the trailing tail in the 11.9 Gyr panel is sharply truncated. None of these so-called ``gaps'' are entirely devoid of star particles, but manifest rather as low-density regions. As time goes on the stream continues to lengthen, but displays significant variations in density that also widen with time. These gaps were likely created by subhalo encounters, but although we can identify snapshots where some of the closest encounters occur, the disturbances are so small in energy (recall $\mathcal{S}_{1\%}=1$ is a one-percent change in kinetic energy) that they are hard to identify in energy space in the same snapshot. Because the effects appear only after a delay while the gap grows in size, we cannot with certainty connect a particular subhalo encounter with a specific stream perturbation at the time-resolution available in this simulation.
\FloatBarrier
\begin{figure*}
\begin{center}
\includegraphics[width=\textwidth]{all_gal.png}
\caption{View of the set of thin streams in ``galactic" coordinates, for an arbitrary location 8 kpc from the center of the host halo. Each stream is a different color; some look thin from this vantage point and some do not.}
\label{fig:AllSky}
\end{center}
\end{figure*}
\section{Statistics of stream-subhalo interactions}
\label{sec:intxnStats}
We also considered a larger subset of streams in the Aq-A-2 halo that appear thin at the present day based on the retagging of star particles described in Section \ref{sec:retagging}. First we selected all streams with stellar mass between $10^4$ and $10^6\ \ensuremath{M_{\odot}}$. Then we visually inspected the present-day configuration of these streams and picked ones that were thin and not well-mixed in position and velocity space. These criteria resulted in a total sample of 18 streams, shown in sky projection in Figure \ref{fig:AllSky}. For convenience we labeled these streams with letters from A to R, with stream O assigned as before. Stream C, which was used in Section \ref{sec:retagging} to compare the different tagging schemes, is above the high end of this mass range and so is not included in the sample: we wanted a very well-resolved stream for the comparison of the methods, but the last panel of Figure \ref{fig:taggingComparisonXY} shows that this stream is fairly thick and disorganized at present day.
\begin{figure*}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=0.45\textwidth]{phi_dist_L-1153919.png} & \includegraphics[width=0.45\textwidth]{phi_dist_L-1153919_unwrapped.png}
\end{tabular}
\caption{Left: stream E from the sample, rotated into the plane perpendicular to its mean angular momentum vector. In both panels, the colors show the component of angular momentum (in the non-rotated frame) used to unwrap the stream. Right panel: the same stream unwrapped in the mean plane with outliers shown as open circled points. The black line is used to measure the physical length of the stream as described in the text.}
\label{fig:unwrapping}
\end{center}
\end{figure*}
\subsection{Measuring the lengths and ages of streams}
\label{subsec:lengthAge}
Since the streams have a wide variety of orbits and ages, we wished to normalize the interaction statistics by the length of each stream at the present day and by each stream's age. To measure stream length, we first calculate the phase angle $\phi$ as described in Section \ref{subsec:EandL}. An example is shown in the left panel of Figure \ref{fig:unwrapping}. The stream must then be ``unwrapped" in $\phi$, which we accomplish by adding and/or subtracting multiples of $2\pi$ for subsets of the stream particles so that the components of the angular momentum $\mathbf{L}$ (in the unrotated frame) vary continuously, since overlapping wraps tend to have very different values of at least one angular momentum component (right panel of Figure \ref{fig:unwrapping}). To track the continuous variation of $\vect{L}$ we pick the component in the unrotated frame which has the widest range of values; in the example shown in Figure \ref{fig:unwrapping} this is $L_x$ but it varies from stream to stream. Stream particles that are outliers in angular momentum and cannot be assigned unambiguously to a particular wrap are masked (light circled points in the right panel of Figure \ref{fig:unwrapping}). The angular span is derived by taking the range of the unwrapped angle without the masked particles, and the physical length is computed by binning small numbers of consecutive particles in the unwrapped $\phi$ coordinate, computing their average radius $\bar{r}_i$ and angular range $\delta\phi_i$, and calculating the Riemann approximation
\begin{equation}
\ell_s = \sum_{i=1}^{\sub{N}{bins}} \bar{r}_i \delta\phi_i.
\end{equation}
The number of particles per bin ranges from 1 to 10 depending on the number of particles in a particular stream, and is adjusted to give a relatively smooth approximation to the stream everywhere (thin black line in the right panel of Figure \ref{fig:unwrapping}).
As shown in the case of Stream O, there is usually a delay between the time that the parent halo becomes a satellite of the main galaxy and the time when a stream begins to form. We therefore determined for each stream in our sample the point when the stream started to form and used this to calculate the stream's age. To do so we compared the largest and smallest eigenvalues of the covariance matrix of the positions of the tagged star particles associated with each stream as a function of time. Before the stream starts to form, the stars are all still bound to the satellite so all three eigenvalues should be similar, while once the satellite begins to tidally disrupt, one eigenvalue (the one along the spreading direction) will become much larger. In practice, the ratio of largest to smallest eigenvalue is between about 1.5 and 4 for all our streams while in the satellite phase, and increases rapidly once the stream starts to form. As the stream wraps around the galaxy the ratio of eigenvalues begins to oscillate, and can decrease again as the stream becomes phase mixed, but the initial increase is diagnostic in every stream we examined. Based on this behavior we identified the snapshot where the ratio of eigenvalues is first larger than 5 as the start of stream formation, found the corresponding formation time $\sub{t}{form}$, and computed the age as
\begin{equation}
t_s = 13.6 \textrm{Gyr} - \sub{t}{form}.
\label{eq:age}
\end{equation}
Encounters before $\sub{t}{form}$ are very rare (less than one per stream) thanks to the small cross-section of the bound progenitor, so we consider the number of interactions since infall to be equivalent to the number since formation.
\begin{figure*}
\includegraphics[width=\textwidth]{velocity_histograms_new.png}
\caption{Distribution of average relative velocities between subhalos encountering any stream in the sample and the stars within the encounter radius, for three different choices of encounter radius (from left to right): 1 kpc, 2 kpc, and 5 kpc. The colors show different mass bins. Only encounters after the start of stream formation (identified as discussed in Section \ref{subsec:lengthAge}) are counted.}
\label{fig:speedDist}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=\textwidth]{fig15_scaled.png}
\caption{Total number of encounters per snapshot for all 18 streams tracked, for the five different thresholds considered. The encounters are stacked relative to the formation time of each stream $t_{\mathrm{form}}$ and timescales are normalized by the age of each stream $t_s$ (Section \ref{subsec:lengthAge}). The dotted gray lines show the expected time-evolution of the number of encounters within a fixed threshold postulated in YJH, $N_{\mathrm{enc}} \propto t^2$, fit to the scaled time range (0,0.25) for the three fixed distance thresholds.}
\label{fig:enc-with-time-all}
\end{center}
\end{figure*}
\subsection{Relative velocities of interactions}
\label{subsec:vrel}
Figure \ref{fig:speedDist} shows the distribution of encounter speeds for three bins in subhalo mass. We used a two-sample Kolmogorov-Smirnov (K-S) test to determine whether different-mass subhalos are drawn from the same speed distribution by comparing pairs of distributions in different mass bins for the same threshold. For the mass-independent thresholds shown in the figure, the K-S test generally detects a difference between the distributions in different mass bins where there are enough subhalos in the sample to do so, and the certainty with which one can conclude the distributions are different increases with the sample size. For the mass-dependent thresholds (not shown) the K-S test says all the distributions in different mass ranges are indistinguishable ($p\sim 0.86-0.92$), probably due to the small sample size.
To investigate differences in the relative speed distribution for different subhalo masses, we fit a Maxwell-Boltzmann (M-B) distribution to the CDF of relative speeds for each mass range shown in Figure \ref{fig:speedDist}. Although the M-B distribution does not describe the overall velocity distribution of dark matter in Aq-A-2 very well \citep{2009MNRAS.395..797V}, it is a fairly good fit to the relative speed distributions. The fits to different mass ranges result in velocity dispersions that systematically differ by about 10-15\%, with the lowest values (220--235 km \unit{s}{-1}) coming from the lowest mass range and the highest values (235--255 km \unit{s}{-1}) obtained for the highest mass range.
The broad speed distributions we find for the interacting subhalos are roughly in agreement with the reasoning outlined in \citet{Yoon2011} that the interaction speed distribution should be Maxwellian with characteristic width $\sqrt{2}\sigma$, where $\sigma$ is the velocity dispersion of the subhalos. Fitting a M-B distribution to the speed distribution of Aq-A-2 subhalos within 200 kpc (where all the stream particles examined are located) gives $\sigma=120$ km \unit{s}{-1}, predicting a width of 170 km \unit{s}{-1} for the interaction-speed distribution. We obtain slightly higher velocity dispersions than this for the interacting subhalos in our samples, but it is worth noting that the best-fit overall speed distribution is skewed substantially toward low velocities; if one instead matches the peak of the M-B distribution to that of the Aq-A-2 subhalos one obtains $\sigma=165$ km \unit{s}{-1} for the subhalo velocity dispersion and hence a predicted width of about 230 km \unit{s}{-1} for the interaction speed distribution, which is more in line with what we find. It is also worth noting that the recorded distributions of encounter speeds are biased towards low-speed encounters since these remain in the encounter volume longer, and thus are more likely to be recorded in a snapshot. For a fixed threshold distance this bias is roughly proportional to $1/\sub{v}{rel}$, and is also direction-dependent---subhalos moving parallel to the stream are more likely to be detected than those moving perpendicularly. We will discuss this bias, related to the time-resolution of the simulation, in more depth in Section \ref{subsec:resolution}. On the other hand, stream stars can reach speeds of up to $\sim 300$ km \unit{s}{-1} near pericenter (as seen in the right-hand panel of Figure \ref{fig:OrphanXYsnaps}), and the subhalo number density increases toward the center of the host halo, which may offset the time-resolution bias somewhat. Most importantly, the broad width and high mean speed of the subhalo distribution results in tails at both very high and very low encounter speeds. In general the lower the relative speed, the more disruptive the interaction will be; Figure \ref{fig:speedDist} confirms that the low-speed tail extends down to very low relative speeds (under 100 km \unit{s}{-1}) for even the lowest-mass subhalos considered.
\subsection{Frequency and mass spectrum of interactions}
\label{subsec:freqmass}
Figure \ref{fig:enc-with-time-all} shows the aggregate time-evolution of the encounter rate for all the streams in the sample. In order to stack the stream encounter histories, we time-shift each individual stream's encounters relative to its formation time, and then divide by its age, before adding all the encounters together and rebinning. We can then compare the aggregate encounter rate to the predicted $t^2$ behavior (dotted gray lines). In the case of the 5 kpc threshold, there appears to be some base rate of random encounters (roughly one per time-bin on average) before $\sub{t}{form}$ so we start the power-law from a constant instead of zero at $t=\sub{t}{form}$. As observed for stream O in Figure \ref{fig:OrphanIntxAll}, at late times the interaction rate appears to fall below this power-law behavior for the fixed-distance thresholds, which we attribute to a drop in particle resolution.
The mass-dependent thresholds have a much noisier signal, with the $r_{1/2}$ encounter rate consistent with noise and the $2r_{1/2}$ rate barely emerging from the noise level. Again, we attribute this to the low particle resolution of the tagged streams; the mass-dependent thresholds are generally well below 1 kpc so they are the most susceptible. For this reason, we decided to only consider aggregate results from the $2r_{1/2}$ threshold and not quote any results from the $r_{1/2}$ threshold since it is so undersampled.
\begin{figure*}
\includegraphics[width=\textwidth]{hist_submass_subvel_all_streams.png}
\caption{Distributions of relative velocity (left) and total mass (right) for all the encounters with the thin streams, for the three distance thresholds and the mass-dependent $2r_{1/2}$ threshold, and for all encounters within 5 kpc with $\mathcal{S}_{1\%}>1$. }
\label{fig:mass-and-vrel}
\end{figure*}
\begin{figure*}
\includegraphics[width=\textwidth]{encs_per_length_per_time.png}
\caption{Rate of subhalo encounters (Equation \eqref{eq:rate}) within the three fixed distance thresholds for the 18 streams in our sample. Horizontal error bars are 10\% of the estimated stream length; vertical error bars include the length uncertainty and Poisson uncertainty on the number of encounters, added in quadrature.}
\label{fig:encounterRate}
\end{figure*}
Figure \ref{fig:mass-and-vrel} shows the distributions of total mass and relative speed for all the encounters as a function of threshold. We also include the distributions of all encounters with $\mathcal{S}_{1\%}>1$ for comparison. The relative velocity distribution is similar for the fixed-distance and mass-dependent thresholds, while the strong velocity-dependence in the definition of $\mathcal{S}_{1\%}$ is apparent here. As expected, the three fixed-distance thresholds sample the overall mass distribution while the mass-dependent threshold shows the competition between the subhalo mass function and the scaling of subhalo size with mass. The distribution of encounters with $\mathcal{S}_{1\%}>1$ is similar to the mass-dependent threshold, which justifies our use of the half-mass radius to represent the ``sphere of influence" of a subhalo. We can conclude from these results that while the typical halo interacting with a stream will have a low mass, the masses of subhalos that come close enough to significantly affect the stream are fairly uniformly distributed over several decades in mass: a $10^6\ M_\odot$ subhalo is roughly as likely to have a strong interaction with a stream as a $10^8\ M_\odot$ subhalo.
Finally, using the measured lengths $\ell_s$ and ages $t_s$ of the streams, determined as described in Section \ref{subsec:lengthAge}, we computed the length-normalized interaction rate,
\begin{equation}
\label{eq:rate}
\eta \equiv \frac{\sub{N}{enc}}{\ell_s t_s},
\end{equation}
for each stream in the sample and for each of the constant distance thresholds. As discussed in Section \ref{subsec:freqmass}, we do not computer individual rates for the mass-dependent thresholds since the frequency of interactions is consistent with noise. Figure \ref{fig:encounterRate} shows the results. Although the streams vary widely in length, rates for the large group of streams with lengths of a few hundred kpc do not show length- or age-dependence in their encounter rate (though there is a large scatter). The large range of stream lengths (nearly a factor of 100) can apparently account for much of the equally wide range in the number of encounters shown in Figure \ref{fig:encounterRate}. Even after normalization, the highest encounter rates tend to be for the shortest streams and the lowest encounter rates for the longest streams. This effect is most likely an indication of our limited numerical resolution and our approximate method for measuring stream lengths, which can be greatly affected by the inclusion/exclusion of single particles in the limit of low resolution. As an example of how these effects can combine, the streams with the two highest rates in the 5 and 2 kpc panels are the two shortest streams in the sample: both are near apocenter at present day (i.e. at the shortest length of any orbital phase). However, both streams also have a fairly large number of particles in a diffuse region, surrounding a more concentrated nucleus, that were excluded as outliers during the length determination. The phase-dependence of the length and the uncertainty in the length measurement itself, which in this case is probably skewed to the short side, combine to produce very high rates for these two streams.
Taking the median of our sample of 18 streams, we find that a stream that is 10 Gyr old and 10 kpc long at present day should have experienced $61.8^{+211}_{-40.6}$ subhalo encounters within 5 kpc, $9.1^{+17.5}_{-7.1}$ encounters within 2 kpc, and $1.5^{+3.0}_{-1.1}$ encounters within 1 kpc. The error bounds given here are the extrema of the range of encounter rates over the streams we tracked.
\begin{figure}
\includegraphics[width=0.45\textwidth]{encounter_rate_scatter_2kpc.png}
\caption{Number of encounters within 2 kpc versus number of particles in stream. Size and color denote the stellar mass of the stream progenitor and age of the stream (defined as in Section \ref{subsec:lengthAge}), respectively. The point circled in black is stream O. }
\label{fig:resolution-effects}
\end{figure}
\subsection{Resolution effects}
\label{subsec:resolution}
Encounters with the stream are undersampled because of the low time-resolution of the snapshots: for $t\gtrsim 1$ Gyr, the time between snapshots is 0.1546 Gyr, so a subhalo with a typical velocity of ~130 km/s moves about 20 kpc between snapshots. The undersampling is less severe for larger threshold distances, and is not strictly proportional to the single-particle threshold distance since we look for subhalos within that distance of \emph{any} particle in the stream. As the stream gets longer, this effect is mitigated somewhat as the stream length approaches the average distance traveled by a subhalo between two snapshots. Thus the slope at which the interaction rate increases at middle times in Figure \ref{fig:OrphanIntxAll} may be slightly steeper than if we had infinitely good time-resolution.
The decrease in the slope of the interaction rate toward the end of the simulation (evident for Stream O in Figure \ref{fig:OrphanIntxAll} and still marginally present in the aggregated history of Figure \ref{fig:enc-with-time-all}) is probably an effect of particle resolution. Toward the end of the simulation when the tagged stream particles are very spread out, a lack of resolution can cause us to miss subhalos that would have interacted with the stream. Figure \ref{fig:resolution-effects} illustrates this point for our entire sample: the number of encounters is primarily correlated with the number of particles in the stream, whether one considers a fixed threshold (left) or a mass-dependent one (right).
Time- and particle resolution both affect our results in the same way, so the results we present in this work should therefore underestimate the number of interactions between a stream and dark substructures. We think that particle resolution is probably the more seriously limiting of the two issues based on the scaling of the median interaction rates for the 18 streams. If we had perfect time-resolution, the median rates for the fixed-distance threshold should scale as the cube of the threshold length: $1:8:125$ for the 1:2:5 kpc thresholds. Our measured rates scale as $1:6.1^{+11.6}_{-4.8}:41.2^{+141}_{-27.1}$, so the medians are a bit on the low side but the proper scaling is within the range we obtain for different streams.
The mass resolution of the simulation mainly affects our ability to detect repeat encounters, which would likely be interactions with subhalos that accompanied the progenitor of the stream during infall. The dark-matter halos of the stream progenitors are not very massive, however, and if a stream progenitor is the largest halo in an infall group, then its most massive companion is likely to be only about 1/100th as massive, and hence will be resolved with about as many particles as are tagged for the stellar stream. For a progenitor like that of our example O stream, the most massive companion will only be resolved with about 700 particles. It takes about 100 particles to satisfactorily resolve a subhalo, so only the few most massive companion subhalos to a stream progenitor, that might give rise to multiple encounters, will be resolved in the simulation. Thus the frequency of repeat encounters is likely also underestimated by this work.
\section{Discussion}
\label{sec:discussion}
The encounter rate that we measure can be compared to analytic
expressions in the literature. Figure \ref{fig:comparison} compares our results to the estimate derived by \citet{Yoon2011} for the total number of encounters in the mass range $10^6-10^7\ M_{\odot}$ within a constant impact parameter $\sub{b}{max}$, based on the subhalo density in the Via Lactea II simulation (their Equation 16):
{\small
\begin{eqnarray}
\sub{N}{enc}(10^6<\sub{M}{sub}<10^7\ M_{\odot}) = 20 \left(\frac{\sub{R}{circ}}{13.7 \textrm{ kpc}}\right)\left(\frac{\sub{b}{max}}{0.58\ \textrm{kpc}}\right) \qquad \nonumber \\
\times \left(\frac{\sigma}{120 \textrm{ km}\unit{s}{-1}} \right) \left(\frac{t}{8.44\ \textrm{Gyr}}\right)^2 \left(\frac{\sub{n}{sub}}{0.0006\ \unit{kpc}{-3}}\right) \left(\frac{\Delta\Psi}{1^\circ}\right)\left(\frac{0.55\ \textrm{Gyr}}{T_{\Psi}}\right).
\label{eq:yjh}
\end{eqnarray}
}
Given that the VLII halo is quite similar to Aq-A-2 in mass, we use the fiducial values for the stream lengthening rate per orbit $\Delta\Psi$, and the orbital period $T_\Psi$. We also take the fiducial value of the subhalo density $\sub{n}{sub}$; although this value is calculated at the low end of our streams' distance range (13.7 kpc), the Aq-A-2 halo has slightly more substructure than VLII so the subhalo density at 10-60 kpc is actually about the same as VLII at 13.7 kpc \citep{Springel2008}. We use a subhalo velocity dispersion $\sigma=139.3$ km \unit{s}{-1}, which is the best-fit value from a fit to a Maxwell distribution: as we discuss in Section \ref{subsec:vrel}, this functional form is not a very good fit to the real subhalo velocity distribution, but the model of YJH presumes this form so we use it for consistency. The predictions depend only linearly on the velocity dispersion, and more strongly on other things like the age of each stream, which varies among the different streams, whereas $\sigma$ is the same for all of them and so will only shift the predictions up or down by some constant. Additionally the best-fit value does approximate the width of the distribution (perhaps slightly over-estimated thanks to skewness that cannot be accommodated by the model). For these reasons we do not expect the poor fit to have a strong effect on the result. Plugging in each stream's median radius for $\sub{R}{circ}$, each stream's age $t_s$ as $t$, and using the different constant distance thresholds as \sub{b}{max} produces the predictions shown as red points, which we compare to the measured number of encounters within the same thresholds (black circles). At small \sub{b}{max} our rates are lower by a factor of about 10, while at large \sub{b}{max} the predictions agree; this is partially expected behavior given the resolution limits since wider thresholds are less sensitive both to the time between snapshots and increasing distance between particles. On the other hand, the orbital period and lengthening rate for our streams is likely longer than the fiducial value since many are on orbits at larger distances in the halo (VLII is also a bit more massive than Aq-A-2), which may also contribute to the larger rates calculated by \citeauthor{Yoon2011}.
\begin{figure}
\includegraphics[width=0.45\textwidth]{comparison_yjh.png}
\caption{Comparison between our measured numbers of encounters within different distance thresholds (black circles) and analytic estimates based on \citet[][Equation \ref{eq:yjh}]{Yoon2011} (red points). See text for the details of how the predictions were calculated.}
\label{fig:comparison}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{comparison_carlberg.png}
\caption{Comparison between our measured numbers of encounters within different distance thresholds and mass ranges (blue circles) and analytic predictions based on \citet{Carlberg2012}. In the left-hand column, all interactions within $2r_{1/2}$ with $M>10^6M_\odot$ are compared with the predicted rate of Equation \eqref{eq:carlberg} presuming that any interaction within this distance will open a discernible gap. In the right-hand column, interactions with $b<1$ kpc and $M>10^{7.28}M_\odot$ are compared with the rate predicted by Equation \eqref{eq:carlberg}, based on applying Carlberg's empirical criterion for opening a gap as described in the text.}
\label{fig:comparison_ray}
\end{center}
\end{figure}
\citet{Carlberg2012} also derives an expression for the rate of subhalo interactions with a slightly different approach. Using constrained simulations they determine the maximum impact parameter required to open a gap in a stream of a fixed density contrast relative to the unperturbed stream, as a function of the time since the interaction, the mass of the perturbing subhalo, and the orbital radius $r$. Because our mass-dependent thresholds are heavily subject to undercounting, we sought instead to try to identify the set of subhalos within a fixed encounter threshold that satisfy this criterion, which we computed for each stream in our sample using its mean orbital distance for $r$ and its age $t_s$ for the time since interaction. This is an upper limit on the required encounter distance, since interactions that happened later on would need to come closer to produce a gap of the same depth in a shorter time. We find that based on this reasoning, our 1 kpc fixed threshold corresponds to the required encounter distance for subhalos in the mass range $10^{7.28} - 10^{8}\ M_{\odot}$. There are 22 encounters within 1 kpc within this mass range, over a total of 18 streams; the number of qualifying encounters per stream ranges from zero to 8.
To compare with Carlberg's prediction we first tried simply applying Equation 5 of \citet{Carlberg2012},
{\small
\begin{equation}
\frac{\mathcal{R}_{\cup}(\hat{M}, r)}{\unit{kpc}{-1} \unit{Gyr}{-1}} = 0.0066\left(\frac{r}{100 \textrm{ kpc}}\right)^{0.23}\left( \frac{n(r)/n_0}{6} \right) \left(\frac{\sigma}{120 \textrm{ km}\unit{s}{-1}}\right) \left(\frac{\hat{M}}{10^8\ M_{\odot}}\right)^{-0.35},
\label{eq:carlberg}
\end{equation}
}
which uses fiducial values taken from the subhalo distribution in the Aquarius
halos. We used the median distance of star particles in each stream as
the galactocentric distance $r$ and set the subhalo velocity dispersion $\sigma$ to 139.3 km \unit{s}{-1} as with the YJH estimate. The scaled subhalo number density
$n(r)/n_0$ is obtained for Aq-A-2 from Figure 11 of
\cite{Springel2008}, and is about 15 for the range explored by our
sample streams (10-60 kpc). Based on our application of the gap-opening criterion, we calculated upper and lower limits on $\mathcal{R}_{\cup}$ for each stream by taking $\hat{M} = 10^{7.28}$ and $\hat{M} = 10^{8}$ as the range of possible minimum masses to open a gap given 1 kpc maximum impact parameter. We then multipiled Equation \eqref{eq:carlberg} by each
stream's measured length $\ell_s$ and age $t_s$ to get an estimate of the expected number of encounters in each stream, to compare with the 22 encounters in this same mass range over the 18 streams. This comparison is shown in the right-hand column of Figure \ref{fig:comparison_ray}, with the range of the predicted number of encounters denoted by the green boxes. This approach presumes that each stream stays at its present-day length over its entire formation time, whereas in reality they grow roughly linearly with time. Furthermore it presumes a circular orbit whereas the actual stream orbits are fairly eccentric and spend most of their time at larger distances where the subhalo density is lower. Added to the fact that for this distance threshold we are likely missing some events due to particle and time resolution, Equation \eqref{eq:carlberg} should produce an over-estimate relative to what we observe and we find that this is so; in fact the over-estimate is something like a factor of 10 to 50. However, the number of scattering events we detect in this mass range is very small, so it is difficult to quantify by how much this approach over-estimates the number of encounters.
One could also construe Equation \eqref{eq:carlberg} in terms of the mass-dependent criteria we use to count events within $r_{1/2}$ or $2r_{1/2}$, since this criterion is another way to select interactions that approach close enough to open a gap. Carlberg finds from his simulations that for a discernible gap to open from an interaction 7 Gyr in the past (the fiducial age assumed for all his predictions) the perturbing subhalo must approach within 1-3 scale radii presuming an NFW profile. Satellite galaxies (and simulated subhalos) are tidally truncated beyond a few scale radii; for subhalos orbiting at a few times the scale radius of the parent halo (where our streams are located) this truncation radius is also about 1-3 times the scale radius \citep{Hayashi2003} so it is roughly consistent to use $2r_{1/2}$ as the gap opening criterion for the approach distance. We compare all encounters within $2r_{1/2}$ to the prediction of Equation \eqref{eq:carlberg} for a minimum mass of $\hat{M}=10^6\ M_{\odot}$: in this case all the events we record are gap-opening, so the minimum mass to open a gap is equal to the lowest mass in our sample, which we cut off at $10^6\ M_{\odot}$. This comparison is shown in the left-hand column of Figure \ref{fig:comparison_ray}. This set of interactions, although still undersampled for the reasons described above, has slightly more events in it than the mass-limited, 1 kpc version in the right-hand column, but the rates we measure are still much lower than predicted by Carlberg's formula.
The typical stream in our sample has had only a handful of close interactions with subhalos in its lifetime (less than 10 within 2 kpc, and a 20\% chance of one within $2r_{1/2}$). Given this fairly low rate of encounters, it is at first glance surprising how disturbed the morphology of a stream can be: our example stream starts out as a completely smooth distribution but has some significant lumps and holes in it (both in position and energy) by the end of the simulation, yet we count only one interaction with a subhalo within $r_{1/2}$ over its entire lifetime. \citet{2015ApJ...803...75N} found similarly disorganized-looking streams in their simulations when starting from a smooth globular-cluster-like phase space distribution and using the Via Lactea simulation. In our case, the interaction rate is undercounted for a number of reasons, most notably the fairly large time between snapshots and the low resolution of the streams themselves, but also because we did select streams that still appear fairly coherent at present day, which automatically picks out ones that have had fewer disruptive encounters. The stream lengths on the other hand are probably over-estimated compared to what one would be able to identify as part of the stream observationally, since the ends of our simulated streams are often very diffuse. Additionally, gaps in streams will grow faster in a gravitational potential like Aq-A-2's, which is triaxial, than in a spherically-symmetric potential \citep{1999Helmi}, especially for streams on more eccentric orbits \citep{2015ApJ...808...15C}. Streams in less symmetric potentials are also intrinsically more complex-looking, even without the addition of substructure \citep[e.g.][]{2015ApJ...799...28P}, since these admit a wider range of orbit families than streams in spherical potentials. Gap size also varies significantly with orbital phase \citep{2016arXiv160608782H}, leading to a wider variety of effects. Given the range of possible gap growth rates, the variety of orbits and phases our streams explore, and the fact that we are missing some encounters due to time-resolution, it is perhaps not as surprising that the streams look more ``messed-up,'' and the gaps in them harder to identify and disentangle, than one might expect from our measured encounter rate.
\section{Summary and conclusions}
\label{sec:concl}
In this work we used a new, simplified tagging scheme to produce an accreted stellar halo from the Aquarius A-2 dark-matter-only simulation, and studied interactions between dark subhalos and thin stellar streams orbiting the main halo. We recorded encounters where a subhalo (identified by its most-bound particle) came within several threshold distances of any tagged star particle in a stream. We used three constant distance thresholds (1, 2, and 5 kpc) and two mass-dependent thresholds based on the half-mass radius of each subhalo ($\sub{r}{1/2}$ and $2\sub{r}{1/2}$).
First we tracked a single stellar stream, observing the rate at which
interactions occurred at different thresholds, the evolution of the
physical and phase-space shape of the stream over time, and the mass
and velocity spectra of the encounters. The initial phase-space
distribution of the tagged particles in the stream's progenitor is
smooth at the time of infall, yet we see significant structure,
including discontinuities in position and energy, arise as the stream
evolves. As the stream gets longer the number of encounters increases
and by late times, multiple subhalos are passing close to the stream
in every snapshot, and up to 26 within 2 kpc at a given point in time. Only some of these passes bring the subhalo closer than its own half-mass radius: in the case of our example stream, only 12 subhalos pass within $2r_{1/2}$ and 2
within $r_{1/2}$ over its lifetime. Additionally, we find only one
case out of many where a subhalo experienced multiple encounters with
the stream, and the velocities of the interacting subhalos are not
significantly correlated with that of the stream.
Then we identified a sample of 18 stellar streams that still appeared
streamlike at the present day and repeated the tracking of encounters
for all streams in the sample. The growth of the number of encounters
with time over the entire sample confirms what we observed for the
single stream, although at late times low resolution likely causes us
to under-count interactions. The relative speeds of the interacting
subhalos are representative of the general subhalo population, as are
their masses when a fixed threshold distance is used. For the
mass-dependent threshold distances, subhalos in the range $10^6$ to
$10^8\ M_{\odot}$ (and perhaps up to $10^{8.5}$) are equally likely to have an interaction
with a stream within $2r_{1/2}$.
Comparing the rate of interactions we measure to analytic predictions, we find that our measured rate is lower (as one would expect given the resolution limitations) and closer to these predictions for larger maximum impact parameter (as one would also expect for a resolution-limited rate). Our median measured rates for encounters, per 10 Gyr per 10 kpc, are: 62 within 5 kpc, 9 within 2 kpc, 1.5 within 1 kpc, and 0.2 within $2r_{1/2}$.
\begin{acknowledgements}
RES is supported by an NSF Astronomy and Astrophysics Postdoctoral Fellowship under award AST-1400989. RES and AH gratefully acknowledge support from the European Research Council under ERC-Starting Grant GALACTICA-240271. AH is also supported by a VICI grant from NWO.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
Evolutionary game theory is used on different levels of biological systems, ranging from the genetic level to ecological systems. The language of game theory allows to address basic questions of ecology, related to the emergence of cooperation and biodiversity, as well as the formation of coalitions with applications to social systems. Darwinian dynamics, in particular frequency-dependent selection, can be formulated in terms of game-theoretic arguments \cite{nowak}. The formation of dynamical patterns is considered as one of the most important promoters of biodiversity \cite{may,levin,durret,hassel}. Here we consider games of competition, where the competition is realized as predation among $N$ species, where each species preys on $r$ others in a cyclic way. A subclass of these $(N,r)$-games are cyclic games, that is $(N,1)$ with $(3,1)$ being the famous rock-paper-scissors game. An extensive overview on cyclic games is given in \cite{szabo,perc}. The $(3,1)$-game has been studied in various extensions (spatial, reproduction and deletion, swapping or diffusion, mutation). One of the first studies of a $(3,1)$-game without spatial assignment, but in a deterministic and stochastic realization revealed that fluctuations due to a finite number of agents can drastically alter the mean-field predictions, including an estimate of the extinction probabilities at a given time \cite{reichen1}. This model was extended to include a spatial grid in \cite{reichen2}, where the role of stochastic fluctuations and spatial diffusion was analyzed both numerically and analytically. The influence of species mobility on species diversity was studied in \cite{reichen3}, pattern formation close to a bifurcation point was the topic of \cite{reichen4}, see also \cite{reichen5} and the impact of asymmetric interactions was considered in \cite{reichen6}.
An extension to four species, first without spatial assignment, shows interesting new features as compared to $(3,1)$: Already in the deterministic limit the trajectories show a variety of possible orbits, and from a certain conserved quantity the late-time behavior can be extrapolated \cite{durney}. The four species can form alliance pairs similarly to the Game of Bridge \cite{case}. Under stochastic evolution various extinction scenarios and the competition of the surviving set of species can be analyzed \cite{28}. Domains and their separating interfaces were studied in \cite{arXiv:1205.4914}. $(4,1)$ cyclic games on a spatial grid were the topic in \cite{szabosznaider,luetz}. A phase transition as a function of the concentration of vacant sites is identified between a phase of four coexisting species and a phase with two neutral species that protect each other and extend their domain over the grid. For an extension of this model to long-range selection see \cite{hua}.
In this paper we focus on the $(6,3)$-game, including both spiral formation inside domains and domain formation. It is a special case of $(N,r)$-games, which were considered for $N\ge3$ and $r\ge1$ by \cite{m1,m2} and more recently by \cite{m3}. The authors of \cite{m1,m2} were the first to notice that for certain combinations of N and r one observes the coexistence of both spiral formation and domain formation. However, it should be noticed that our set of reactions, even if we specialize $(N,r)$ to the $(3,1)$-game, is similar, but not identical with the versions, considered in \cite{m1,m2,m3} or in \cite{reichen1}-\cite{reichen6}. The seemingly minor difference refers to the implementation of an upper threshold to the occupation number of single sites (set to 1 or a finite fixed number), while we use a ``bosonic" version. We introduce a dynamical threshold, realized via deletion reactions, so that we need not explicitly restrict the occupation number per site. Due to this difference, the bifurcation structure of the mean-field equations is changed.
The reason why we are interested in the particular combination of $N=6$ and $r=3$ is primarily motivated by two theoretical aspects rather than by concrete applications. As to the first aspect, this game is one of the simplest examples of ``games within games" in the sense that the domains effectively play a $(2,1)$-game as transient dynamics on a coarse scale (the scale of the domain diameter), while the actors inside the domains play a $(3,1)$-game on the grid scale. Finally, one of the domains gets extinct along with all its actors. As such, this game provides a simple, yet non-trivial example for a mechanism that may be relevant for evolution: In our case, due to the spatial segregation of species, the structural complexity of the system increases in the form of patterns of who is chasing whom, appearing as long-living transients, along with a seemingly change of the rules of the game that is played between the competing domains on the coarse scale, while the rules, which individuals use on the elementary grid sites, are not changed at all. As outlined by Goldenfeld and Woese \cite{goldenfeldwoese}, it is typical for processes in evolution, in particular in ecology, that ``the governing rules are themselves changed", as the system evolves in time and the rules depend on the state. In our example it is spatial segregation, which allows for a change of rules from a coarse perspective, as we shall see.
As to the second aspect, an interesting feature of such an arrangement is the multitude of time and spatial scales that are dynamically generated. Concretely in the $(6,3)$-game the largest of the reaction/diffusion rates sets the basic time unit. When the species segregate and form domains, the next scale is generated: it is the time it takes the two domains to form until both cover the two-dimensional grid or the one-dimensional chain. The domains are not static, but play the $(2,1)$-game that has a winner in the end. So the extinction time of one of the domains sets the third scale. A single domain then survives, including the moving spirals from the remaining $(3,1)$-game inside the domain. The transients can last very long, depending on the interaction rates and the system size. In the very end, however, in a stochastic realization as well as due to the finite accuracy in the numerical solutions even in the mean-field description, only one out of the three species will survive, and the extinction of the other two species sets the fourth scale. Along with these events, spatial scales emerge, ranging from the basic lattice constant to the radii of spirals and the extension of the domains.
One of the challenges is to explore which of the observed features in the Gillespie simulations can be predicted analytically. We shall study the predictions on the mean-field level, which is rather conclusive in our ultralocal implementation of reactions and reproduces the results of the Gillespie simulations quite well, since fluctuations turn out to play a minor role for pattern formation. The deterministic equations are derived as the lowest order of a van Kampen expansion. The eigenvalues of the Jacobian are conclusive for the number of surviving species in a stable state, the composition of the domains, and transient behavior, which is observed in the Gillespie simulations. The mean-field equations, including the diffusion term, will be integrated numerically and compared to the results of the Gillespie simulations.
The paper is organized as follows. In section~\ref{sec_reactions} we present the model in terms of basic reactions and the corresponding master equation. For generic $(N,r)$ games we summarize in section~\ref{sec_vankampen} the derivation of the mean-field equations from a van Kampen expansion, followed by a stability analysis via the Jacobian with and without spatial dependence for the specific $(6,3)$ game, and a derivation of the numerical solutions of the mean-field equations in section~\ref{sec_Jacobian}. In section~\ref{sec_numerical} we present our results from the Gillespie simulations in comparison to the mean-field results. Section~\ref{sec_conclusions} summarizes our conclusions and gives an outlook to further challenges related to this class of games. For comparison, the supplementary material contains a detailed stability analysis for the $(3,1)$-game with spiral formation and the $(3,2)$-game with domain formation, as well as the numerical solutions of the mean-field equations and the corresponding Gillespie simulations.
\section{Reactions and Master Equation}\label{sec_reactions}
We start with the simplest set of reactions that represent predation between individuals of different species, followed by reproduction, deletion
\begin{eqnarray}\label{eq:rec_sys}
X_{\alpha,i}\, +\, X_{\beta,i} & \overset{k_{\alpha\beta}/V}{\longrightarrow} & X_{\alpha,i} \label{pred}\\
X_{\alpha,i}\, & \overset{r_{
\alpha,i}}{\longrightarrow} & 2 X_{\alpha,i} \label{repr} \\
2X_{\alpha,i}\,& \overset{p_{\alpha}/V}{\longrightarrow} & X_{\alpha,i} \label{anih}
\end{eqnarray}
and finally diffusion
\begin{eqnarray}\label{diff}
X_{\alpha,i}\, & \overset{D_{\alpha}/h^2}{\longrightarrow} & X_{\alpha,j}.
\end{eqnarray}
$X_{\alpha,i}$ represents an individual of species $\alpha$ at lattice site $i$, while the total number of individuals of species $\alpha$ at site $i$ will be denoted with $n_{\alpha,i}$. (We use small characters $n$ for convenience, although the meaning of $n$ is not a density, but the actual occupation number of a certain species at a certain site.) In view of applications to ecological systems, each lattice site stands for a patch housing a subpopulation of a metapopulation, where the patch is not further spatially resolved. Eq.~(\ref{pred}) represents the predation of species $\alpha$ on species $\beta$ with rate $k_{\alpha\beta}/V$, where the parameter $V$ does not stand for the physical volume, but parameterizes the distance from the deterministic limit in the following way: According to our set of reactions, larger values of $V$ lead to higher occupation numbers $n_{\alpha,i}$ of species $\alpha$ at sites $i$, since predation and deletion events are rescaled with a factor $1/V$, and therefore to a larger total rate. The fluctuations in occupation numbers, realized via the Gillespie algorithm, are independent of $V$ or the occupation numbers of sites, since only relative rates enter the probabilities for a certain reaction to happen. Therefore the size of the fluctuations relative to the absolute occupation numbers or to the overall $V$ gets reduced for large V, that is, in the deterministic limit. Predation is schematically described in figure~\ref{(6,3)}.
Eq.~(\ref{repr}) represents reproduction events with rate $r_{\alpha}$, and Eq.~(\ref{anih}) stands for death processes of species $\alpha$ with rate $p_{\alpha}/V$. Death processes are needed to compensate for the reproduction events, since we do not impose any restriction on the number of individuals that can occupy lattice sites. Here we should remark why we implement death processes in the form of Eq.~\ref{anih} rather than simpler as $X_{\alpha,i} \overset{p_{\alpha}}{\longrightarrow} \oslash$. The latter choice could be absorbed in a term $(\rho-\gamma)\phi_i\equiv \tilde{\rho}\phi_i$ in the mean-field equation (\ref{eq:pde}) below with uniform couplings $\rho$ and $\gamma$. This choice would not lead to a stable coexistence-fixed point \cite{josef} and therefore not to the desired feature of games within games\footnote{For the $(6,3)$-game we would have 40 fixed points, the sign of the eigenvalues would then only depend on the sign of the parameter $\tilde{\rho}$. At $\tilde{\rho}=0$ all fixed points collide and exchange stability through a multiple transcritical bifurcation. For $\tilde{\rho}>0$ (the only case of interest), the system has no stable fixed points, and the numerical integration of the differential equations diverges. (Similarly for the (3,1)-game, for $\tilde{\rho}>0$, the trivial fixed point with zero species is always an unstable node, while the coexistence fixed point is always a saddle.)}.
The species diffuse within a two-dimensional lattice, which we reduce to one dimension for simplicity if we analyze the behavior in more detail. We assume that there can be more than one individual of one or more species at each lattice site. Individuals perform a random walk on the lattice with rate $D_{\alpha}/h^d$, where $D_{\alpha}$ is the diffusion constant, $h$ the lattice constant and $d$ the dimension of the grid. Diffusion is described by Eq.~(\ref{diff}), where $i$ represents the site from which an individual hops, and $j$ is one of the neighboring sites to which it hops. It should be noticed that diffusion is the only place, which leads to a spatial dependence of the results, since apart from diffusion, species interact on-site, that is, within their patch.
In summary, the main differences to other related work such as references \cite{reichen1,reichen2,reichen3,reichen4,reichen5,reichen6,durney,case, 28,m1,m2,m3} are the ultralocal implementation of prey and predation, no swapping, no mutations as considered in \cite{mobilia1,mobilia2}, and a bosonic version with a dynamically ensured finite occupation number of sites. Even if qualitatively similar patterns like spirals or domains are generated in all these versions, the bifurcation diagram, that is, the stability properties and the mode of transition from one to another regime depend on the specific implementation.
We can now write a master equation for the probability of finding $\{n\}$ particles at time $t$ in the system for reaction and diffusion processes, where $\{n\}$ stands for $(n_{1,1},...,n_{N,L^d})$ and $N$ is the number of species, $L^d$ the number of sites.
\begin{eqnarray}\label{eq:me_reac}
\frac{\partial P^{reac} \left( \left\{ n \right\};t \right)}{\partial t} &=&
\underset{i}{\sum} \left\{
%
\underset{\alpha,\beta}{\sum} \frac{k_{\alpha\beta}}{V} \left[
n_{\alpha,i}\left(n_{\beta,i}+1\right)P\left(n_{\alpha,i},n_{\beta,i}+1,...;t \right)
- n_{\alpha,i}n_{\beta,i}P\left(\{n\};t \right) \right] \right.\nonumber \\
&+&\left. \underset{\alpha}{\sum} \frac{p_{\alpha}}{V} \left[
\left( n_{\alpha,i}+1 \right) n_{\alpha,i} P\left( ...,n_{\alpha,i}+1,...;t \right)
- n_{\alpha,i}\left( n_{\alpha,i}-1 \right) P(\{n\};t) \right] \right. \nonumber \\
%
&+& \left. \underset{\alpha}{\sum} r_{\alpha} \left[
(n_{\alpha,i}-1)P(n_{\alpha,i}-1,...;t) - n_{\alpha,i}P(\{n\};t) \right]
\right\}
\end{eqnarray}
\begin{eqnarray}\label{eq:me_diff}
\frac{\partial P^{diff} \left( \left\{ n \right\};t \right)}{\partial t} &=& \underset{\alpha}{\sum} \frac{D_\alpha}{h^2} \underset{\left\langle i,j \right\rangle}{\sum} \left[ (n_{\alpha,i}+1)P(...,n_{\alpha,i}+1,n_{\alpha,j}-1,...;t)-n_{\alpha,i}P(\{n\};t)\right. \nonumber\\
&+&\left. (n_{\alpha,j}+1)P(...,n_{\alpha,i}-1,n_{\alpha,j}+1,...;t)-n_{\alpha,j}P(\{n\};t)\right]
\end{eqnarray}
with $n_{\alpha,i}\ge 1$ for all $\alpha,i$, and
\begin{equation}\label{eq7}
\partial_tP = \partial_tP^{reac}+\partial_tP^{diff}.
\end{equation}
As uniform (with respect to the grid) random initial conditions we assume a Poissonian distribution on each site $i$
\begin{equation}
P \left( \{n\} ;0 \right)=\underset{\alpha,i}{\prod}\left( \frac{\overline{n}^{n_{\alpha,i}}_{\alpha,0}}{n_{\alpha,i}!} e^{-\overline{n}_{\alpha,0}} \right),
\end{equation}
where $\overline{n}_{\alpha,0}$ is the mean initial number of individuals of species $\alpha$ per site.
\section{Derivation of the mean-field equations}\label{sec_vankampen}
The master equation is continuous in time and discrete in space. The diffusion term is included as a random walk. Next one takes the continuum limit in space, in which the random walk part leads to the usual diffusion term in the partial differential equation (pde) for the concentrations $\varphi(\vec{x},t)\equiv n_\alpha(\vec{x},t)/V$. The mean-field equations can then be derived by calculating the equations of motion for the first moments $\langle n_\alpha(\vec{x},t)\rangle$ from the master equation, where the average is defined as $\langle n_\alpha(\vec{x},t)\rangle=\sum n_\alpha(\vec{x},t)P(\{n_\alpha(\vec{x},t)\})$ with $P(\{n_\alpha(\vec{x},t)\})$ being a solution of the master equation, and factorizing higher moments in terms of first-order moments.
Alternatively, we insert the ansatz for the van Kampen expansion according to $n_\alpha=V\varphi_\alpha + \sqrt{V}\eta_\alpha$ in the reaction part. To leading order in $V$ we obtain the deterministic pde for the concentrations of the reaction part. Combined with the diffusion part this leads to the full pde that is given as Eq.~\ref{eq:pde} in the next section. While this leading order then corresponds to the mean-field level, the next-to-leading order leads a Fokker-Planck equation with associated Langevin equation, from which one can determine the power spectrum of fluctuations. In our realization, the visible patterns are not fluctuation-induced, differently from noise-induced fluctuations as considered in \cite{goldenbutler}. Therefore our power spectrum of fluctuations is buried under the dominating spectrum that corresponds to patterns from the mean-field level. Therefore we do not further pursue the van Kampen expansion here\footnote{For details of a possible derivation of the mean-field equations we refer to \cite{darkathesis}; however, there we derived the mean-field equations via a longer detour towards a field theoretic formulation, where we read off the mean-field equations as leading order of a van Kampen expansion, not applied to the master equation, but to a Lagrangian that appears in the path integral derived from the master equations, in analogy to the derivation in \cite{goldenbutler}.}.
\section{Stability analysis of the mean field equations and their solutions}\label{sec_Jacobian}
We perform a linear stability analysis of the mean field equations by finding the fixed points of the system of partial differential equations
\begin{equation}\label{eq:pde}
\partial_t\varphi_\alpha =
D_\alpha\nabla^2\varphi_\alpha+
r_\alpha\varphi_\alpha-
p_\alpha\varphi_\alpha^2 -
\underset{\beta}{\sum}k_{\beta\alpha}\varphi_\alpha\varphi_\beta,
\end{equation}
with $\phi_\alpha$ the concentration of species $\alpha$,
by setting $\partial_t\varphi_\alpha =D_\alpha\nabla^2\varphi_\alpha=0$. We will focus on the system with homogeneous parameters $r_\alpha=\rho$, $k_{\alpha\beta}=\kappa$ if species $\alpha$ preys on $\beta$ and 0 otherwise, $p_\alpha=\gamma$, and $D_\alpha=\delta$, $\forall\alpha,\beta\in\{1,...,N\}$ and consider the special case of the (6,3)-game. After finding the fixed points, we look at the eigenvalues of the Jacobian $J$ of the system~(\ref{eq:pde}) to determine the stability of the fixed points. We then extend our analysis to a spatial component by analyzing a linearized system in Fourier space, with Jacobian~\cite{cianci}
\begin{equation}
J^{SP}=J+\underline{D}\tilde{\Delta},
\end{equation}
where $\tilde{\Delta}=-k^2$ is the Fourier transform of the Laplacian and $\underline{D}$ is the diffusion matrix evaluated at a given fixed point. In our case, the diffusion matrix is a diagonal matrix $\underline{D}=\delta\mathbb{1}$. This leads to a dependence of the stability of the fixed points on diffusion. In the following we focus on the special case to be considered.
\subsection{Stability analysis and numerical integration for the (6,3)-game}
{\bf Stability analysis of the (6,3)-game.}
The (6,3)-game is given by the system of mean field equations:
\begin{eqnarray}\label{eq:MF(6,3)}
\frac{\partial\varphi_1}{\partial t} & = & \delta\nabla^2\varphi_1 + \rho\varphi_1 - \gamma\varphi_1^2 - \kappa\varphi_1(\varphi_4+\varphi_5+\varphi_6) \nonumber \\
\frac{\partial\varphi_2}{\partial t} & = & \delta\nabla^2\varphi_2 + \rho\varphi_2 - \gamma\varphi_2^2 - \kappa\varphi_2(\varphi_5+\varphi_6+\varphi_1) \nonumber \\
\frac{\partial\varphi_3}{\partial t} & = & \delta\nabla^2\varphi_3 + \rho\varphi_3 - \gamma\varphi_3^2 - \kappa\varphi_3(\varphi_6+\varphi_1+\varphi_2) \nonumber\\
\frac{\partial\varphi_4}{\partial t} & = & \delta\nabla^2\varphi_4 + \rho\varphi_4 - \gamma\varphi_4^2 - \kappa\varphi_4(\varphi_1+\varphi_2+\varphi_3) \nonumber\\
\frac{\partial\varphi_5}{\partial t} & = & \delta\nabla^2\varphi_5 + \rho\varphi_5 - \gamma\varphi_5^2 - \kappa\varphi_5(\varphi_2+\varphi_3+\varphi_4) \nonumber\\
\frac{\partial\varphi_6}{\partial t} & = & \delta\nabla^2\varphi_6 + \rho\varphi_6 - \gamma\varphi_6^2 - \kappa\varphi_6(\varphi_3+\varphi_4+\varphi_5). \nonumber\\
\end{eqnarray}
In total there are 64 different fixed points $FP_1$ to $FP_{64}$, of which some have the same set of eigenvalues and differ only by a permutation of the fixed-point coordinates of the eigenvalues,
so that we can sort all fixed points in 12 groups $FP^1$-$FP^{12}$: for example, the fixed points $(\rho/(\gamma+\kappa),0,\rho/(\gamma+\kappa),0,\rho/(\gamma+\kappa),0)$ and $(0,\rho/(\gamma+\kappa),0,\rho/(\gamma+\kappa),0,\rho/(\gamma+\kappa))$ are in the same group $FP^3$. We will refer to all fixed points by the number of the group they belong to, that is to $FP^1$ to $FP^{12}$, instead of $FP_1$ to $FP_{64}$.\\
The zero-fixed point $FP^1$, where all components are equal to zero, with all eigenvalues equal to $\rho$ for $\delta=0$, and equal to $\rho-\delta k^2$ for $\delta\neq0$, is unstable for a system without spatial assignment, while it can become stable for a spatial system if $\rho<\delta k^2$, as in the cases of the (3,1) and (3,2) games, which are discussed in detail in the supplementary material. In the coexistence-fixed point $FP^2$, all components are equal to $\rho/(\gamma+3 \kappa)$. It is stable for $\kappa/\gamma<1$, three of the eigenvalues are always negative, the first one being $-\rho$ and an the second and third one equal to $-\rho\gamma/(\gamma+\kappa)$, two are complex conjugates $-\rho(\gamma-\kappa\pm i \sqrt{3}\kappa)/(\gamma+3\kappa)$, and the last one is real $-\rho(\gamma-\kappa)/(\gamma+3\kappa)$. At $\kappa/\gamma=1$, $FP^2$ becomes a saddle, three of the six eigenvalues change sign, complex conjugates change sign of their real part, so a Hopf bifurcation occurs, and the direction corresponding to the last eigenvalue becomes unstable.\\
Other fixed points include the survival of one species ($FP^4$), two species (for both $FP^5$ and $FP^6$), three (for $FP^7$ and $FP^8$), four ( for $FP^9$, $FP^{10}$ and $FP^{11}$), and five species (for $FP^{12}$). All fixed points $FP^4$ to $FP^{12}$ are always saddles in the case of $\delta=0$.\\
For $\delta\neq0$ all eigenvalues get a $(-\delta k^2)$-term, which can extend the stability regime in the parameter space, as long as $k\neq0$, and lead to the coexistence of stable fixed points, which cannot be found for $\delta=0$.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=5cm]{fig1.pdf}
\end{center}
\caption{Diagram of a (6,3)-game. Colors represent species, each preys on three other species in clockwise direction, shown only for the red species by black arrows. Red lines connect species, which form one domain (red, green, and blue), the other three species (cyan, magenta, and yellow) form the second domain. Each species (like the red one) preys on only one species from its own domain (green), and on two species from the other domain (cyan and magenta), this way eliminating all predators of the third species (blue) from the domains. These rules are characteristic for all games, in which a domain forms of three species playing the (3,1)-game. Colors in this scheme will be used throughout the paper to represent species one (red) to six (yellow). }\label{(6,3)}
\end{figure}
In view of pattern formation we shall distinguish three regimes. Before we go into detail, let us first give an overview of the sequence of events, if we vary the bifurcation parameters $\kappa$ and $\gamma$ as $\kappa/\gamma$. These are events, which we see both in the Gillespie simulations and the numerical integration of the mean-field equations in space, described as a finite grid upon integration.
\begin{itemize}
\item $\kappa/\gamma<1$: the first regime with $\kappa/\gamma$ smaller than its value at the first Hopf bifurcation at $\kappa/\gamma=1$, where the 6-species coexistence fixed point becomes unstable. As long as this fixed point is stable, we see no patterns, as the system converges at each site of the grid to the 6-species fixed point without dominance of any species, so that the uniform color is gray.
\item $1<\kappa/\gamma<2$: the second regime with $\kappa/\gamma$ chosen between the first and second Hopf bifurcations, where the second one happens at $\kappa/\gamma=2$ for the $FP^3$ fixed points. When $\kappa/\gamma=1$ is approached from below, that is from $\kappa/\gamma<1$, two fixed points, belonging to the $FP^3$-group, become stable through a transcritical bifurcation until $\kappa/\gamma=2$, where they become unstable through the second Hopf bifurcation. Each of the two predicts the survival of three species, the ones, which are found inside the domains. Each of these fixed points is, of course, a single-site fixed point, so in principle a subset of the nodes of the grid can individually approach one of the two fixed points, while the complementary set of the nodes would approach the other fixed point. However, as a transient we see two well separated domains with either even or odd species. At the interfaces between them all six species are present and oscillate with small amplitude oscillations, caused by the first Hopf bifurcation of the six-species coexistence fixed-point, where it became a saddle. Which one of the domains wins the effective (2,1)-game in the end, where a single domain with all its three species survives, depends on the initial conditions and on the fact that diffusion is included; the mere stability analysis only suggests that six species at a site destabilize the interface between domains with either even or odd species. In fact, the numerical integration and the Gillespie simulations both show that one domain gets extinct if the lattice size is small enough and/or the diffusion fast enough. As long as the two fixed points are stable, the (3,1)-game is played at each site of a domain in the sense of coexisting three species, which are not chasing each other, related to the neighboring sites only via diffusion, without forming any patterns. Patterns are only visible at the interface of the domains as a remnant of the unstable six-species coexistence fixed point.
\item $\kappa/\gamma>2$: the third regime, which is of most interest for pattern formation.
Starting from random initial conditions, the species segregate first into two domains, each consisting of three species, one with species 1,3, and 5, the second one with species 2, 4, and 6, and inside both domains the three species play a rock-paper-scissors game, chasing each other, since the two fixed points of the $FP^3$ group became unstable at the second Hopf bifurcation. Due to the interactions according to an effective $(2,1)$-game at the interfaces of the domains (here with either two or four species coexisting), one of the domains will also here get extinct, including the involved three species, while the remaining three survive. Which domain survives depends also here on the initial conditions. As we shall see, the temporal trajectories of the concentrations of the three species in the surviving domain show that they still explore the vicinity of the second Hopf bifurcation from time to time, while they otherwise are attracted by the heteroclinic cycle. The three species in the surviving domain live the longer, the larger the grid size is, in which the species continue playing (3,1). In contrast to the second regime, however, two of the three species in the surviving domain will get extinct as well, and a single one remains in the end. This extinction is caused by fluctuations in the finite population in the stochastic simulation or by the numerical integration on a spatial grid with finite numerical accuracy, respectively.
\end{itemize}
So the linear stability analysis indicates options for when we can expect oscillatory trajectories: it is the Hopf bifurcations in the (6,3)-game for the $FP^2$ and $FP^3$ fixed points that induce the creation of limit cycles, which here lead to the {\it formation of spirals} in space in the third regime and only temporary patterns at the interfaces in the second regime, before the system converges to one of the $FP^3$ fixed points.
Moreover, it is the two $FP^3$ fixed points in the (6,3)-game that correspond to the {\it formation of two domains}. In both the (3,2) and the (6,3)-games, one of these fixed points will be approached as a collective fixed point (shared by all sites of the grid), while the domain corresponding to the other one gets extinct, and patterns are seen if this fixed point is unstable. So in the (6,3)-game the existence of domains including their very composition is due to two stable (second regime) or unstable (third regime) fixed points. Their coexistence is in both regimes transient. In the second regime three species will survive in the end, because the three-species coexistence-fixed point is stable, and it would need a large fluctuation to kick it towards a 1-species unstable fixed point. In contrast, only one species will survive in the third regime, where the same fixed point is unstable. Obviously here it does not need a rare, large fluctuation to kick the system towards the 1-species unstable fixed point, as we always observed a single species to survive in the end, both in the Gillespie simulations and the numerical integration in a relatively short time.
We should mention, however, that from our Gillespie simulations we cannot exclude that after all, a large fluctuation would also kick the system in the second regime from its metastable state towards one of the unstable 1-species fixed points as well as in the first regime to either one of the two three-species fixed points, or to one of the six 1-species fixed points, when the six-species fixed point is stable in the deterministic limit. So far we have not searched for these rare events, in which two, three or five species would get extinct, respectively.
\vskip5pt
\textbf{Numerical solutions of the (6,3)-game.}
\begin{figure}[tp]
\begin{center}
\includegraphics[scale=1]{fig2.pdf}
\end{center}
\caption{Evolution of the (6,3)-game in the second regime in one dimension. The parameters are $ \gamma = \kappa = 1 $, $ \delta/dx = 0.1 $, and $ \kappa = 1.5 $. The right and middle columns show the species of each domain separately.
For further explanations see the text.}\label{(6,3)_array_DOM}
\end{figure}
In the following we show evolutions of species concentrations in space and time for parameters, chosen from the second and third regime of the (6,3)-game. These solutions are obtained from the numerical integration of Eq.~\ref{eq:MF(6,3)}.
For the representation on a lattice we will use the following procedure to visualize site occupation: Odd species are represented by the rgb-(red, green, blue) color scheme, while even species are represented by cmy-colors (cyan, magenta, yellow). The three numbers of species $(r,g,b)$, or $(c,m,y)$, divided by the total sum of all species at the site, give a color in the rgb-, or cmy-spectrum that results from a weighted superposition of individual colors, where the weights (color intensities) depend only on the ratios of occupation numbers, rather than on absolute numbers. Moreover, we display the rgb-color scheme if odd species make up the majority at a site and the cmy-scheme otherwise. We should note that a well mixed occupation of odd (even) species leads to a dark (light) gray color in these color schemes.
Figure~\ref{(6,3)_array_DOM} shows coexisting domains with oscillations at the interfaces in the second regime. To justify the visualization of data according to the ``majority rule", we show even and odd species also separately in the two right panels. This way we can see the transitions at the interfaces of the domains between even and odd species more clearly. The light (dark) gray domain corresponds to a well mixed occupancy with even (odd) species, respectively. On the boundaries of the domains all six species are present, and if we zoom into the boundary, we can see small amplitude oscillations caused by the Hopf bifurcation of the 6-species coexistence-fixed point, see figure~\ref{(6,3)_tx_DOM}. Figures~\ref{(6,3)_array_DOM} (a)-(c) show the evolution of the system at the first 100 t.u. at which time the domains are already starting to form. In panel (a) it is seen how the transient patterns, generated by transient small domains, shortly after the domains disappear also fade away, so that the transient patterns are generated by oscillations at the interfaces. The figure also reminds to the early time evolution of condensate formation in a zero-range process, where initially many small condensates form, which finally get absorbed in a condensate that is located at a single site with macroscopic occupation in the thermodynamic limit. Here initially many small and short-lived domains form, which get first absorbed into four domains as seen in the figure, but later end up in a single domain with three surviving species. So we see a ``condensation" in species space, where three out of six species get macroscopically occupied as a result of the interaction, diffusion and an unstable interface, while the remaining three species get extinct, so that the symmetry between the species in the cyclic interactions with identical rates gets dynamically broken.
\begin{figure}[tp]
\begin{center}
\includegraphics[scale=1]{fig3.pdf}\\
\end{center}
\caption{Trajectories of all six species corresponding to figure~\ref{(6,3)_array_DOM} at an interface. Red (1), green (3), and blue (5) represent odd species, cyan (2), magenta (4), and yellow (6) even species, from smaller to larger labels, respectively. (a) and (b) show temporal and spatial trajectories, respectively, at the beginning of the integration, corresponding to (a)-(c) in figure~\ref{(6,3)_array_DOM}, while (c) and (d) refer to late times. For further explanations see the text.}\label{(6,3)_tx_DOM}
\end{figure}
Panels (d)-(f) show the evolution from 10000-10100 time units (t.u.). The displayed domains were checked to coexist numerically stable up to $10^6$ t.u., while for smaller lattices and faster diffusion one domain gets extinct. Figures~\ref{(6,3)_array_DOM} (a) and (d) should be compared with figures~\ref{(6,3)1D_2} (a) and (b) of the Gillespie simulations, respectively.
Figure~\ref{(6,3)_tx_DOM} shows the corresponding oscillating concentration trajectories at early (a) and late (c) times at a site of an interface (x=124), where all six species oscillate around the coexistence-saddle fixed point, as indicated by the horizontal black line in (c), while the spatial dependence at (b) (early) and (d) (late) times displays the domain formation due to two stable fixed points, corresponding to figures~\ref{(6,3)_array_DOM} (a) and (d), respectively, so that the oscillations are restricted to the interfaces.
\begin{figure}[tp]
\begin{center}
\includegraphics[scale=1]{fig4.pdf}
\end{center}
\caption{Evolution of the (6,3)-game in the third regime in one-dimension. The parameters are $ \gamma = \kappa = 1 $, $ \delta/dx = 0.1 $, and $ \kappa = 4.0 $. The middle and right columns show the species of each domain separately. For further explanations see the text. }\label{(6,3)_array_OSC}
%
\end{figure}
The evolution of the (6,3)-game in the third, oscillatory regime in one dimension is shown in figure~\ref{(6,3)_array_OSC}. Species are represented in the same way as in figure~\ref{(6,3)_array_DOM}. Panel (a)-(c) show the evolution of the system in the first 100 t.u., (d)-(f) in the first 10000 t.u. The two stable fixed points from the second regime became unstable (saddles) through the second Hopf bifurcation. As in the second regime, at the beginning of the integration there is a separation of odd and even species, but at the same time they start to chase each other, resulting in oscillatory behavior in space and time. Here we see no longer traces of the limit cycle around the six-species coexistence fixed point as in the second regime, since no sites have six species coexisting, even not for a short period of time. At the interfaces between even and odd species usually three species coexist, either two odd and one even, or vice versa, two even and one odd, but these mixtures are not stable, as these 3-species coexistence-fixed points in the deterministic limit are saddles. It also happens that just two or four species coexist at the interface, but also their coexistence-fixed points are saddles. Therefore also here the coexistence of domains is not stable, only one of them survives, and which one depends on the initial conditions, resulting in the extinction of three either odd or even species. In view of Gillespie simulations, figures~\ref{(6,3)_array_OSC} (a) (early times) and (b) (late times) should be compared with figures~\ref{(6,3)1D_1} (a) (early) and (b) (late), respectively.
Figure~\ref{(6,3)_tx_OSC} (a) shows the evolution in time at late times, when only one domain survives. All three species oscillate between zero and one, corresponding to the heteroclinic cycle. From time to time the trajectories are also attracted by the saddle-limit cycle, which is created by the second Hopf bifurcation of the three species-fixed point (black line) as indicated by the small amplitude oscillations. Apart from the amplitude, the heteroclinic and saddle-limit cycles differ in their frequency: the saddle-limit cycle has a higher frequency than the heteroclinic cycle. Panel (b) shows the spatial trajectories at the beginning of the integration when both domains still coexist. Yet we see no mixing of all six species at a single site, the 6-species coexistence-fixed point is no longer felt in this regime.
\begin{figure}[tp]
\begin{center}
\includegraphics[scale=1]{fig5.pdf}\\
\end{center}
\caption{In correspondence to figure~\ref{(6,3)_array_OSC} (a) temporal trajectories of the surviving domain at late times in the interval 10200-10600 t.u. when only even species exist, and (b) spatial trajectories at early times when still both domains exist. For further explanations see the text.}\label{(6,3)_tx_OSC}
%
\end{figure}
As we see from figure~\ref{(6,3)_array_OSCend}, the numerical integration evolves to one of the saddles after having spent a finite time on the heteroclinic cycle and not according to the analytical prediction, where it were only in the infinite-time limit that the trajectory would get stuck in one of the saddles, which are connected by the heteroclinic cycle. According to figure~\ref{(6,3)_array_OSCend}(a) all trajectories get absorbed in one (the pink one) of the 1-species saddles already at finite time as a result of the finite accuracy of the numerical integration. Yet figure~\ref{(6,3)_array_OSCend} (b) shows the characteristics of a heteroclinic cycle at finite time: The dwell time of the trajectory in the vicinity of the 1-species saddles gets longer and longer in each cycle, before it fast moves towards the next saddle in the cycle. This escape stops after a finite number of cycles, when the concentration of two of the three species are zero within the numerical accuracy, and therefore no ``resurrection" is possible.
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=1]{fig6.pdf}
\end{center}
\caption{Extinction of all but one species upon approaching the heteroclinic cycle, (a) (x,t)-diagram, (b) species concentrations as a function of time. For further explanations see the text.}\label{(6,3)_array_OSCend}
%
\end{figure}
\section{Numerical Methods and Results}\label{sec_numerical}
Going back to the set of reactions, in this section we describe their Gillespie simulations.
We solve the system (\ref{eq:rec_sys})
by stochastic simulations on a regular square lattice as well as on a one-dimensional ring, using the Gillespie algorithm~\cite{gillespie}, combined with the so-called next-subvolume method~\cite{elf}. This method is one option to generalize Gillespie simulations to spatial grids. We choose periodic boundary conditions on a square $L\times L$ lattice or on a ring with $L$ nodes. In our case nodes, or synonymously sites, represent subvolumes. All reactions except of the diffusion happen between individuals on the same site (in the same subvolume), and a diffusion reaction is a jump of one individual to a neighboring site. One event can change the state of the system of only one (if a reaction happens) or two neighboring (if a diffusion event happens) subvolumes. At each site the initial number of individuals of each species is chosen from a Poisson distribution $P \left( \{n\} ;0 \right)=\underset{\alpha,i}{\prod}\left( \frac{\overline{n}^{n_{\alpha,i}}_{\alpha,0}}{n_{\alpha,i}!} e^{-\overline{n}_{\alpha,0}} \right)$, with a mean $\overline{n}_{\alpha,0}$, which is randomly chosen for each species $\alpha$.
In the next-subvolume method we assign the random times of the Gillespie algorithm to subvolumes rather than to a specific reaction. To each subvolume, or site, we assign a time $\tau$, at which one of the possible events, in our case reactions (predation, birth or death), or diffusion, will happen. The time $\tau$ is calculated as $\tau = -\ln(rn)/r_{total}$, where $rn$ is a random number generated from a uniform distribution between 0 and 1. The total rate $r_{total}$ depends on the reaction rates and the number of individuals which participate in the event. Events happen at sites in the order of the assigned times $\tau$. Once it is known at which subvolume the next event happens, the event (reaction or diffusion) is chosen randomly according to the specified reaction rates.
We start the simulations with initial conditions from a Poisson distribution such that each site of the entire lattice is well mixed with all species. We want to study the dynamics of the system in a parameter regime, where we expect pattern formation. From the linear stability analysis of the mean-field system we expect stable patterns in the regime without stable fixed points in the (6,3)-game, as long as three species are alive, and transient patterns in the (6,3)-game for coexisting stable fixed points.
We use the same color scheme as we used before to visualize the numerical solutions of the mean-field equations.
As one Gillespie step (GS) we count one integration step here.
Our results confirm the predictions from the mean-field analysis: there are on-site oscillations in time in a limit-cycle regime. There are also oscillations in space, which form spirals on two-dimensional lattices. If the evolution approaches the stochastic counterpart of a stable fixed point in the deterministic limit, we call it shortly ``noisy fixed point", where the trajectories fluctuate around a value that is the mean field-fixed point value multiplied by the parameter $V$ as defined before. Of particular interest is the influence of the diffusion in relation to the ratio $\kappa/\gamma$ on the patterns. It was the ratio of $\kappa/\gamma$ that determines the stability of the fixed points. As mentioned earlier, the value of $\delta k^2$, which enters the stability analysis, can extend the stability regime. So in the Gillespie simulations it is intrinsically hard to disentangle the following two reasons for the absence of patterns in the case of fast diffusion: either the stability regime of a fixed point with only one surviving species is extended, or the diffusion is so fast, that the extension of visible patterns is larger than the system size, so that a uniform color may just reflect the homogeneous part within a large pattern.
All the mean-field-fixed points are proportional to the value of the parameter $\rho$. If this value is much larger than $\kappa$ and $\gamma$, the fixed-point value is very large. This leads to a large occupation on the sites, which slows down the formation of patterns. The reason is that the number of reactions, which are needed for the system to evolve to stable trajectories, either to oscillations, or to fixed points, increases with the number of individuals in the system.
\vskip5pt
We study the stochastic dynamics of a (6,3)-game in regimes, for which we expect pattern formation, i.e. for $\gamma<\kappa$. When the coexistence-fixed point $FP^2$ becomes unstable at $\gamma=\kappa$, we find the formation of two domains, each consisting of three species, one domain containing odd species, in the figures represented by shared colors red, green, and blue in the rgb-color scheme. The other domain consists of even species, represented by shared colors of cyan, magenta, and yellow in the cmy-color representation, see figure~\ref{(6,3)2D}. Inside the domains the three species play the (3,1)-game and form spiral patterns. We have checked that the domains in figure~\ref{(6,3)2D} are not an artefact of the visualization, and determined, for example, the occupancy on a middle column of the lattice (not displayed here).
On sites with oscillations of species from one domain, there is a very small or no occupation of species of the second domain, confirming the very existence of the domains.
The time evolution of the six species on two sites, chosen, for example, from the middle column of the lattice confirm that the species' trajectories oscillate in time, reflecting the stable limit cycles in the deterministic limit.
Here a remark is in order as to whether radii, propagation velocity or other features of the observed spiral patterns can be predicted analytically. While spiral patterns in spatial rock-paper-scissors games were very well predicted via a multi-scale expansion in the work of \cite{mobilia1,mobilia2}, we performed a multi-scale expansion (see, for example \cite{bookkuramoto}) to derive amplitude equations for the time evolution of deviations from the two unstable fixed points, which lose their stability at the two Hopf bifurcations. However, the resulting amplitude equations differ from Ginzburg-Landau equations by a missing imaginary part, which can be traced back to the absence of an explicit constraint to the occupation numbers on sites and the absence of a conserved total number of individuals. As a result, the amplitude equations only predict the transient evolution as long as the trajectory is in the very vicinity of the unstable fixed point, but cannot capture the long-time behavior, which here is determined by an attraction towards the heteroclinic cycle that is responsible for the spiral patterns in our case. So it seems to be this non-local feature in phase space that the multi-scale expansion about the Hopf bifurcation misses.
For a further discussion of how the patterns depend on the choice of parameters we shall focus on the results on a one-dimensional lattice, since the simulation times are much longer for two dimensions. (In two dimensions, the period of oscillations is as long as about one fifth of the $2^{30}$ Gillespie steps.)
\begin{figure}[tp]
\begin{center}
\includegraphics[scale=0.9]{fig7.pdf}
\end{center}
\caption{Pattern formation for a (6,3)-game on a two-dimensional $(64\times 64)$-lattice for weak diffusion and far from the bifurcation point. Snap shots are taken at $1000\cdot2^{15}$ (a), $10000\cdot2^{15}$ (b), and $32000\cdot2^{15}$ (c) GS. Two domains are formed, each containing three species, indicated by the different color groups. These species play a (3,1)-game inside the domains and evolve spiral patterns. The parameters are $\rho=1$, $\kappa=1$, $\delta=1$, and $\gamma=0.2$.
}\label{(6,3)2D}
\end{figure}
\begin{figure}[tp]
\begin{center}
\includegraphics[scale=1]{fig8.pdf}
\end{center}
\caption{Pattern formation in the (6,3)-game on a one-dimensional lattice of 64 sites for $\kappa/2<\gamma<\kappa$, that is in the second regime, where the coexistence-fixed points $FP^3$ are stable, for weak diffusion $\delta=0.1$ ((a) and (b)), and strong diffusion $\delta=1$ ((c) and (d)). For both strengths of the diffusion domains form. In the case of weak diffusion no extinction of domains is observed within the simulation time of $2^{30}$ GS. For strong diffusion, one domain goes extinct after $800\cdot2^{15}$ GS. Initially, oscillatory patterns appear as remnants of many interfaces between small domains, where within the interfaces six species oscillate due to the unstable 6-species coexistence-fixed point, which fade away with time. This confirms the analytical results that in this parameter regime the $FP^3$-fixed points with $2\times 3$ coexisting species are both stable, leading to the black color in (c) and (d) for the one surviving domain. Panel (a) shows the time evolution on a lattice for the time interval $(0-1000)\cdot2^{15}$, (b) for $(30000-31000)\cdot2^{15}$, (c) for $(0-1000)\cdot2^{15}$, and (d) for $(10000-11000)\cdot2^{15}$ GS. The parameters are $\gamma=\rho=0.5$, $\kappa=0.6$.}\label{(6,3)1D_2}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[scale=1]{fig9.pdf}
\end{center}
\caption{Pattern formation in the (6,3)-game on a one-dimensional lattice of size 64, shown on a space-time grid, for $\gamma<\kappa/2$, that is, in the third regime, where the coexistence-fixed points $FP^3$ are unstable, weak diffusion $\delta=0.1$ ((a) and (b)), and strong diffusion $\delta=0.1$ ((c) and (d)). The parameters are $\gamma=\rho=0.5$ and $\kappa=1.2$. Two domains form, of which one goes extinct after $600\cdot2^{15}$ GS in the case of weak diffusion and $150\cdot2^{15}$ GS in the case of strong diffusion. The surviving domain keeps playing the (3,1)-game. For weak diffusion no further extinction is observed for the simulation time of $2^{30}$ GS, while for strong diffusion, an extinction of all but one species, here the red one, happens after $1470\cdot2^{15}$ GS. Panel (a) shows patterns in a time interval of $(0-700)\cdot2^{15}$, (b) for $(30000-30100)\cdot2^{15}$, (c) for $(0-200)\cdot2^{15}$, and (d) for $(0-1470)\cdot2^{15}$ GS.}\label{(6,3)1D_1}
\end{figure}
As to diffusion, for stronger diffusion patterns are more homogeneous, extinction events happen faster, sometimes they happen only for sufficiently strong diffusion, see figure~ \ref{(6,3)1D_2}. The extinction time depends also on the $\kappa/\gamma$-ratio, i.e. if the ratio is from the interval (i) $1<\kappa/\gamma<2$, where $FP^3$ is stable and $FP^2$ is unstable, or, (ii) from $\kappa/\gamma>2$, where both fixed points are unstable.
For case (i), the $FP^3$-fixed points are stable, yet at the beginning of the simulations the dynamics shows oscillatory behavior, caused by the interfaces between the small domains, where six species feel the unstable coexistence-fixed point at $\gamma=\kappa$; but after about $500\cdot2^{15}$ Gillespie steps for weak diffusion, and $1000\cdot2^{15}$ Gillespie steps for strong diffusion, and for our choice of parameters, the patterns fade away and the system evolves to a homogeneous state in both domains as long as they coexist, see figures~\ref{(6,3)1D_2} (b) and \ref{(6,3)1D_2} (d). The closer the system is to the bifurcation point $\gamma=\kappa$, the longer live the oscillatory patterns, the stronger feels the system the unstable 6-species fixed point.
Figure~\ref{(6,3)1D_2} (a) should be compared with the corresponding mean-field solution of figure~\ref{(6,3)_array_DOM} (a) at early times and figure~\ref{(6,3)1D_2} (b) with figure~\ref{(6,3)_array_DOM} (d) at late times, from which we see that the mean-field solutions reproduce the qualitative features including the transient patterns.
In case (ii), the third regime, domains get faster extinct, both domains play a (3,1)-game inside the domains. After one domain gets extinct, the surviving one keeps on playing the (3,1)-game, until only one species survives. Here one should compare figure~\ref{(6,3)1D_1} (a) with the corresponding mean-field solution of figure~\ref{(6,3)_array_OSC} (a) at early times and figure~\ref{(6,3)1D_1} (b) with figure~\ref{(6,3)_array_OSC} (d) at later times, where one is left with one domain and three species chasing each other; the final extinction of two further species is not visible in this figure due to the weak diffusion and the larger extinction time.
\section{Conclusions and Outlook}\label{sec_conclusions}
Beyond the emerging dynamically generated time-and spatial scales, the most interesting feature of the (6,3)-game is the fact that the rules of the game, specified initially as (6,3), dynamically change to effectively (2,1) and (3,1) as a result of spatial segregation. In view of evolution, here the rules of the game change while being played. They change as a function of the state of the system if the state corresponds to the spatial distribution of coexisting species over the grid.
In preliminary studies, we investigated the $(27,17)$-game with the following set of coexisting games in a transient period: From a random start we observe segregation towards nine domains playing $(9,7)$ with each other and inside the domains again the $(3,1)$ game. From the superficial visualization of Gillespie simulations, this system looks like a fluid with whirling ``vortices", where the (3,1)-game is played inside the domains. We expect a rich variety of games with new, emerging, effective rules on a coarser scale, if we not only increase the number $N$ of species, or release the restriction to cyclic predation, but allow for different time scales, defined via the interaction rates. So far we chose the same rates for all interactions and always started from random initial conditions.
We performed a detailed linear stability analysis, which together with the numerical integration of the mean-field equations reproduced all qualitative features of the Gillespie simulations, even extinction events. That the mean-field analysis worked so well is due to the ultralocal implementation of the interactions, so that the spatial dependence enters only via diffusion. The stability analysis revealed already a rather rich structure with 12 groups of in total 64 fixed points for the (6,3)-game. We focussed on coexistence-fixed points of six, three or one species.
Along with the fixed points' repulsion or attraction properties we observed three types of extinction, whose microscopic realization is different and deserves further studies:
(i) In the second regime of the (6,3)-game, both domains with either even or odd species are in principle stable, as long as they are not forced to coexist. We have seen a spatial segregation towards a domain with only even and one with only odd species, occupying the sites. At the interface between both domains, six species cannot escape from playing the (6,3)-game. Since the 6-species coexistence-fixed point is unstable, the unstable interface seems to be the driving force to initiate the extinction of one of the two domains including its three species, since interface areas should be reduced to a minimal size. From the coarse-grained perspective, one domain preys on the other domain, which is a (2,1)-game.
(ii) In the third regime of the (6,3)-game, the domain structure in odd and even domains is kept, but in the interior of the domains the species follow heteroclinic cycles, which explain the patterns of three species, chasing each other, inside each domain.
At the interface between the domains, two to four species coexist at a site, but for small enough diffusion, coexistence-fixed points of the respective species are always saddles, so also here the instability of the interfaces seems to induce their avoidance, leading
again to the extinction of one of the two domains. So from the coarse-grained perspective, again a (2,1)-game is played between the domains.\\
It should be noticed that in contrast to systems, where the fate of interfaces between domains is explained in terms of the competition between free energy and interface tension, here the growth of domains and the reduction of interfaces are traced back to the linear stability analysis of the system in the deterministic limit, which is conclusive for the dynamics.
(iii) The third type of extinction event was the extinction of two species, when the individual trajectories move either in the vicinity of, or along a heteroclinic cycle, and either a fluctuation from the Gillespie simulations, or the finite numerical accuracy on the grid (used for integration) captured the trajectory in one of the 1-species saddles.
We have not studied rare large fluctuations, which could induce other extinction events and kick the system out of the basin of attraction from the 6- or 3-species stable coexistence-fixed points when stochastic fluctuations are included. Neither have we measured any scaling of the extinction times with the system size or of the domain growth with the system size. This is left for future work.\\
Furthermore, for future work it would be challenging to derive and predict the domain formation on the coarse scale from the underlying $(6,3)$-game on the basic lattice scale in the spirit of the renormalization group, here, however, applied to differential equations rather than to an action.
\section{Acknowledgments}
One of us (D.L.) is grateful to the German Research Foundation (DFG)(ME-1332/25-1) for financial support during this project. We are also indebted to the German Exchange Foundation (DAAD)(ID 57129624)for financial support of our visit at Virginia Tech Blacksburg University, where we would like to thank Michel Pleimling for valuable discussions. We are also indebted to Michael Zaks (Potsdam University) for useful discussions.\\
\section{References}
|
\section{Introduction}
As in other galaxy groups, also the dwarf galaxies inhabiting the Local Group
can be broadly divided in two categories, late- and early-types,
on the basis of the presence or absence of both gas and current star formation,
respectively. It is still a question whether these morphological
types share similar ancestors and owe their distinct properties at redshift
zero to a different evolutionary path, or whether their ``destiny'' as
a late- or early-type system was somewhat imprinted in the conditions at
their formation. The fact that Local Group
late- and early-type dwarfs share similar
scaling relations (see \cite[Tolstoy, Hill \& Tosi 2009]{2009ARA&A..47..371T} for a recent review),
appear to follow the same stellar mass-metallicity relation
(\cite[Kirby et al. 2013]{2013ApJ...779..102K}) but inhabit clearly different environments -
with late-types found in isolation and early-types in general satellites
of the large Local Group spirals - hints to environmental effects being
relevant for the evolution of these galaxies.
On the basis of lifetime star formation histories from very deep
colour-magnitude-diagrams, \cite[Gallart et al.(2015)]{2015ApJ...811L..18G} propose the existence of two
main types of dwarf galaxies: {\it slow dwarfs}, i.e.
those that have formed stars at a relatively low rate for the whole
life of the system, and {\it fast dwarfs}, i.e. those that have experienced
most of their star formation activity in the first Gyrs of evolution.
Late-type systems map into {\it slow dwarfs}, but this is not always the
case for early-type and {\it fast-} dwarfs. Gallart et al. propose that
the two evolutionary paths are imprinted from the start and related to the
environment in which the dwarfs were born.
The spectroscopic study of the evolved stellar component of dwarf galaxies
offers a complementary perspective, as for Local Group galaxies
it is possible to determine metallicities and line-of-sight velocities
for individual stars, and hence study their internal kinematics and
metallicity properties in detail and out to large radii.
Red giant branch stars are particularly interesting targets,
because they probe a very large range of ages,
from $\sim$1.5Gyr to the early star formation episodes. Being very luminous,
they are within reach of current spectrographs on 8m-10m telescopes for galaxies
to within $\sim$1 Mpc.
\section{The data-set}
Here we present preliminary results from our VLT/FORS2 MXU
spectroscopic survey of RGB stars in the Phoenix transition type (dT). This
system is as luminous as a typical ``classical'' spheroidal, but
found at a distance of $\sim$400kpc from the Milky Way, hence it can provide an
interesting view of the properties of low mass dwarf galaxies that have mostly
evolved in isolation.
We obtained VLT/FORS2 MXU spectroscopic data for the Phoenix dT in service mode (Programme 083.B-0252; PI: Battaglia).
We used the 1028z grism in combination with the OG590+32 order separation filter (wavelength coverage 7730 - 9480\,\AA\,);
the slits were set to have a width=1'' for a final resolving power of R$\sim$2500. We obtained 11 pointings
in which we allocated slits to a total of $254$ pre-selected red giant branch (RGB; I\,$<21.5$\,mag) stars
from \cite[Battaglia et al.(2012)]{2012MNRAS.424.1113B}
that cover a total area of roughly $10' \times 20' $ across the field of the galaxy.
In addition to these data,
we use VLT/FORS2 MXU spectra for ~30 stars in a central pointing (Programme 71.B-0516; PI: Cole).
The data reduction and extraction of the spectra was performed with standard
IRAF procedures, together with custom made IDL scripts. We extracted line-of-sight velocities (v$_{\rm l.o.s.}$) and
equivalent widths (EWs) from the Ca~II triplet (CaT) nIR lines. Line-of-sight velocities were obtained
by cross-correlating the continuum normalized spectra with a synthetic spectrum derived
from a stellar atmospheric model with similar parameters as expected for the Phx targets. The EWs were
determined by fitting a Voigt profile and transformed into [Fe/H] using the calibration from
\cite[Starkenburg et al.(2010)]{2010A&A...513A..34S}. The CaT [Fe/H] was validated
against two calibrating globular clusters; these were also used to test the accuracy of the v$_{\rm l.o.s.}$ determination,
together to Phx stars overlapping among different masks. The median error in v$_{\rm l.o.s.}$ is 8 km s$^{-1}$.
The sample of probable Phoenix members among the two observing programmes
consists of $\sim$190 stars, as indicated by the stars' position on the colour-magnitude-diagram and
radial velocity.
\section{Results}
We obtain a systemic l.o.s. velocity equal to -21.2$\pm$1.0 km s$^{-1}$ and a dispersion of 9.3$\pm$0.7 km s$^{-1}$.
The systemic velocity is consistent with the determination of -13$\pm$9 km s$^{-1}$ obtained by \cite[Irwin \& Tolstoy (2002)]{2002MNRAS.336..643I} from a
much smaller sample of 7 stars and in excellent agreement with the velocity of the HI cloud (-23 km s$^{-1}$) proposed by
\cite[St-Germain et al.(1999)]{1999AJ....118.1235S} as physically associated to Phoenix.
It is well-known that dwarf galaxies inhabiting dense environments
such as within the virial radii of the Milky Way or M31
are devoid of gas, except for relatively massive systems such as e.g. for the Magellanic Clouds or IC~10.
The fact that such a small system as Phoenix has been able to
form stars until almost present day (\cite[Hidalgo et al. 2009]{hidalgo+2009})
and still contains gas, together with its large distance and velocity direction approaching the Milky Way
(v$_{\rm GSR}= -108.6 \pm 1.0$km s$^{-1}$),
strongly suggests that Phoenix is still to enter the Milky Way virial radius for the first time and has evolved
undisturbed from interactions with large galaxies.
Interestingly, the metallicity properties of this system appear alike to those of similarly luminous
dwarf spheroidal galaxies satellites of the Milky Way, which are devoid of gas and are found in a higher
density environment. In particular we detect a clear negative, approximately linear [Fe/H] gradient (see Fig.~\ref{fig1}),
whose slope compares well to those of Milky Way dwarf spheroidal galaxies (e.g. \cite[Leaman et al. 2013]{2013ApJ...767..131L}).
This finding leads us to speculate that
metallicity gradients in these small dwarf galaxies are likely to be driven by internal mechanisms rather than by environmental ones,
although this conclusion needs to be put on a more secure foot by
similar spectroscopic surveys of larger samples of Local Group dwarf galaxies inhabiting different environments.
The detailed metallicity and kinematic properties of Phoenix from the VLT/FORS2 MXU sample
here presented will be discussed in Kacharov et al. (in prep.)
\begin{figure}[b]
\begin{center}
\includegraphics[width=3.4in]{Phx_metallicity_gradient.png}
\caption{Metallicity versus elliptical radius for the Phoenix probable member stars. The red lines
show the running average and the linear fit to the metallicity properties.}
\label{fig1}
\end{center}
\end{figure}
\section*{Acknowledgements}
We are grateful to A.Cole for useful discussions and for providing the reduced
VLT/FORS2 data from program 71.B-0516.
|
\section{Appendix}
We include pseudocode for generating questions (Alg.~\ref{alg:question}) and multiple choice answer sets (Alg.~\ref{alg:passage}).
\input{./persuade_code.tex}
\end{document}
\section*{Acknowledgments}
\section{Introduction}
Researchers distinguish the problem of general knowledge question answering from that of reading comprehension~\cite{anonymized,Goldilocks}.
Reading comprehension is more difficult than
knowledge-based or IR-based question answering in two ways.
First,
reading comprehension systems must infer answers from a given unstructured passage
rather than structured knowledge sources such as Freebase~\cite{bollacker2008freebase} or the Google Knowledge Graph~\cite{knowedgegraph}.
Second, machine comprehension systems cannot exploit the large level of
redundancy present on the web to find statements that provide a strong syntactic match to the question~\cite{yang2015wikiqa}.
In contrast, a machine comprehension system must use the single phrasing in the given passage, which
may be
a poor syntactic match to the question.
\ignore{\mbcomment{After motivating machine comprehension above, we should probably start this next para with first a brief discussion of why current MC datasets are insufficient and hence our new dataset is much needed...}}
In this paper, we describe the construction of a new reading comprehension dataset that we refer to as ``Who-did-What''.
Two typical examples are shown in Table~\ref{table:sample}.\footnote{The passages here only show certain salient portions of the passage. In the actual dataset, the entire article is given. The correct answers are (3) and (2).}
The process of forming a problem starts with the selection of a question article from the English Gigaword corpus.
The question is formed by deleting a person named entity from the first sentence of the question article.
An information retrieval system is then used to select a passage with high overlap with the first sentence of the question article,
and an answer choice list is generated from the person named entities in the passage.
Our dataset differs from the CNN and Daily Mail comprehension tasks~\cite{anonymized} in that it forms questions from two distinct articles rather than summary points.
This allows problems to be derived from document collections that do not contain manually-written summaries.
This also reduces the syntactic similarity between the question and the relevant sentences in the passage, increasing the need for deeper semantic analysis.
To make the dataset more challenging we selectively remove problems so as to suppress four simple baselines --- selecting the most mentioned person, the first mentioned person, and two language model baselines.
This is also intended to produce problems requiring deeper semantic analysis.
The resulting dataset yields a larger gap between human and machine
performance than existing ones. Humans can answer questions in
our dataset with an 84\% success rate compared to the estimates of
75\% for CNN \cite{DBLP:journals/corr/ChenBM16a} and 82\% for the CBT
named entities task \cite{Goldilocks}. In spite of this higher level of human performance, various existing readers
perform significantly worse on our dataset than they do on the CNN
dataset. For example, the Attentive Reader \cite{anonymized} achieves 63\% on CNN but only 55\% on Who-did-What
and the Attention Sum Reader
\cite{AttentionSum} achieves 70\% on CNN but only 59\% on Who-did-What.
In summary, we believe that our Who-did-What dataset is more challenging, and requires deeper
semantic analysis, than existing datasets.
\section{Dataset Construction}
\label{sec:details}
We now describe the construction of our Who-did-What dataset in more detail. We sketch the procedure below and provide more specific details in the appendix.
To generate a problem we first generate the question by selecting a random article --- the ``question article'' --- from the Gigaword corpus
and taking the first sentence of that article --- the ``question sentence'' --- as the source of the cloze question.
The hope is that the first sentence of an article contains prominent people and events which are likely to be discussed in other
independent articles.
To convert the question sentence to a cloze question, we first extract named entities using the
Stanford NER system~\cite{finkel2005incorporating} and parse the sentence using the Stanford PCFG parser~\cite{Klein:2003:AUP:1075096.1075150}.
The person named entities are candidates for deletion to create a cloze problem.
For each person named entity we then identify a noun phrase in the automatic parse that is headed by that person. For example,
if the question sentence is ``President Obama met yesterday with Apple Founder Steve Jobs'' we identify the two person noun phrases ``President Obama''
and ``Apple Founder Steve Jobs''. When a person named entity is selected for deletion, the entire noun phrase is deleted. For example, when deleting the second named entity, we get
``President Obama met yesterday with XXX'' rather than ``President Obama met yesterday with Apple founder XXX''. This increases the difficulty of the problems because systems cannot rely on descriptors and other local contextual cues.
About 700,000 question sentences are generated from Gigaword articles (8\% of the total number of articles).
Once a cloze question has been formed we select an appropriate article as a passage. The article should be independent of the question article but should discuss the
people and events mentioned in the question sentence.
To find a passage we search the Gigaword dataset using the Apache Lucene information retrieval system~\cite{McCandless:2010:LAS:1893016}, using the question sentence as the query.
The named entity to be deleted is included in the query and required to be included in the returned article. We also restrict the search to
articles published within two weeks of the date of the question article.
Articles containing sentences too similar to the question in word overlap and phrase matching near the blanked phrase are removed.
We select the best matching article satisfying our constraints. If no such article can be found, we abort the process and move on to a new question. See the appendix for details.
Given a question and a passage we next form the list of choices. We collect all person named entities in the passage except unblanked person named entities in the question.
Choices that are subsets of longer choices are eliminated. For example the choice ``Obama'' would be eliminated if the list also contains ``Barack Obama''.
We also discard ambiguous cases where a part of a blanked NE
appears in multiple candidate answers, e.g., if a passage has ``Bill Clinton'' and ``Hillary Clinton'' and the blanked phrase is ``Clinton''. We found this simple coreference rule to work well in practice since news articles usually employ full names for initial mentions of persons.
If the resulting choice list contains fewer than two or more than five choices, the process is aborted and we move on to a new question.\footnote{The maximum of five helps to avoid sports articles containing structured lists of results.}
After forming an initial set of problems we then remove ``duplicated'' problems. Duplication arises because Gigaword contains many copies of the same article or
articles where one is clearly an edited version of another. Our duplication-removal process ensures that no two problems have very similar questions. Here, similarity is
defined as
the ratio of the size of the bag of words intersection to the size of the smaller bag.
In order to focus our dataset on the most interesting problems, we remove some problems to suppress the performance of the following simple baselines:
\begin{itemizesquish}
\item First person in passage: Select the person that appears first in the passage.
\item Most frequent person: Select the most frequent person in the passage.
\item $n$-gram: Select the most likely answer to fill the blank under a 5-gram language model trained on Gigaword minus articles which are too similar to one of the questions in word overlap and phrase matching.
\item Unigram: Select the most frequent last name using the unigram counts from the 5-gram model.
\end{itemizesquish}
To minimize the number of questions removed we
solve an optimization problem defined by limiting the performance of each
baseline to a specified target value while removing as few problems as possible, i.e.,
\begin{equation}
\max_{\alpha(C)} \sum_{C \in \{0,1\}^{|b|}} \alpha(C) | T(C) |
\end{equation}
subject to
\begin{eqnarray}
&\forall i& \ \sum_{ C : C_i = 1} \frac{\alpha(C) | T(C)| }{N} \le k \nonumber \\
&N& = \sum_{C \in \{0,1\}^{|b|}} \alpha(C) | T(C) |
\end{eqnarray}
where $T(C)$ is the subset of the questions solved by the subset $C$ of the suppressed baselines, and $\alpha(C)$ is a keeping rate for question set
$T(C)$. $C_i=1$ indicates $i$-th baseline is in the subset and $|b|$ is a number of baselines. Then $N$ is a total number of questions and $k$ is an upper bound for the baselins after suppression. $k$ is set to the random performance.
The performance of these baselines before and after suppression are shown in Table~\ref{table:beforeKnockOff}.
The suppression removed 49.9\% of the questions.
Table~\ref{table:statistics} shows statistics of our dataset after suppression. We split the final dataset into train, validation, and test by taking the validation and
test to be a random split of the most recent 20,000 problems as measured by question article date. In this way there is very little overlap in semantic subject matter between the training set and either validation or test.
We also provide a larger ``relaxed'' training set formed by applying less baseline suppression (a larger value of $k$ in the optimization). The relaxed training set then has a slightly different distribution from the train, validation, and test sets which are all fully suppressed.
\ignore{
For initial problem set $Q$, we classifies questions $q \in Q$ by its difficulty, e.g. the set of baseline systems which can solve the question, and defined subsets of questions as $T \subset Q$ for each size $m$ of the multiple choice answer set:
\begin{equation}
T(C) = \{ q \in Q | \mathbb{I}(\text{ $b_i$ solves $q$}) = C_i \}
\end{equation}
where $C \in \{0,1\}^{|b|}$ and $b$ is the above list of baseline systems. Each subsets corresponds to the difficulty.
We removed questions in each subset $T(C)$ with probabilities $\alpha(C)$ which is a solution of the following optimization problem for each size of $m$:
\begin{equation}
\max_{\alpha(C), C\in \{0,1\}^{|b|}} \sum_{C \in \{0,1\}^{|b|}} \alpha(C) | T(C) |
\end{equation}
subject to
\begin{eqnarray}
&\forall i& \ \sum_{ C : C_i = 1} \frac{\alpha(C) | T(C)| }{N} \le k \nonumber \\
&N& = \sum_{C \in \{0,1\}^{|b|}} \alpha(C) | T(C) |
\end{eqnarray}
where $\sum_{ C : C_i = 1} \frac{\alpha(C) | T(C)|}{N}$ is a new accuracy of a baseline $b_i$ after removing trivial questions and $k \in [0,1]$ is the target accuracy which all new baseline accuracies are intended to be. We employed the random chance to the target accuracy and relaxed it with 12\% for data size, e.g. $k=1/m + 0.12$.
}
\begin{table}[t]
\centering
\begin{tabular}{@{}l|ll@{}}
\toprule
& \multicolumn{2}{c}{Accuracy} \\
\multicolumn{1}{c|}{Baseline} & \multicolumn{1}{c}{Before} & \multicolumn{1}{c}{After} \\ \hline
First person in passage & 0.60 & 0.32 \\
Most frequent person & 0.61 & 0.33 \\
$n$-gram & 0.53 & 0.33 \\
Unigram & 0.43 & 0.32 \\
Random$^\ast$ & 0.32 & 0.32 \\
\bottomrule
\end{tabular}
\caption{Performance of suppressed baselines. $^\ast$Random performance is computed as a deterministic function of the number of times each choice set size appears. Many questions have only two choices and there are about three choices on average.
}
\label{table:beforeKnockOff}
\end{table}
\begin{table}
\centering
\begin{tabular}{@{}lrrrr@{}}
\toprule
& relaxed & train & valid & test \\
& train & & & \\ \midrule
\# queries & 185,978 & 127,786 & 10,000 & 10,000 \\
Avg \# choices & 3.5 & 3.5 & 3.4 & 3.4 \\
Avg \# tokens & 378 & 365 & 325 & 326 \\
Vocab size & 347,406 & \multicolumn{3}{c}{308,602} \\
\bottomrule
\end{tabular}
\caption{Dataset statistics.}
\label{table:statistics}
\end{table}
\section{Performance Benchmarks
\label{sec:baselines}
We report the performance of several systems to characterize our dataset:
\begin{itemizesquish}
\item Word overlap: Select the choice $c$ inserted to the question $q$ which is the most similar to any sentence $s$ in the passage, i.e., ${\rm CosSim}({\rm bag}(c+q), {\rm bag}(s))$.
\item Sliding window and Distance baselines (and their combination) from \newcite{richardson-burges-renshaw:2013:EMNLP}.
\item Semantic features: NLP feature based system from \newcite{wang-EtAl:2015:ACL-IJCNLP2}.
\item Attentive Reader: LSTM with attention mechanism \cite{anonymized}.
\item Stanford Reader: An attentive reader modified with a bilinear term \cite{DBLP:journals/corr/ChenBM16a}.
\item Attention Sum (AS) Reader: GRU with a point-attention mechanism \cite{AttentionSum}.
\item Gated-Attention (GA) Reader: Attention Sum Reader with gated layers \cite{DBLP:journals/corr/DhingraLCS16}.
\end{itemizesquish}
\ignore{
\item N-gram : Select the most likely candidate answer with n-gram trained on "???" on the training set.
\item Uni-gram : Select the most likely candidate answer with uni-gram trained on "???" on the training set.
}
\ignore{
{\bf Word overlap: } Select the choice $c$ maximizing the maximum over sentence $s$ in the passage of $\mathrm{overlap}(q(c),s)$
where $q(c)$ is the question with answer $c$ inserted into the blank.
\begin{equation}
\hat{a} = \operatornamewithlimits{argmax}_{c \in A} \max_{s \in P} \;\mathrm{overlap}(q(c),s))
\end{equation}
where $q(c)$ is a question whose blank is filled by a candidate choice $c$.
}
\noindent Table~\ref{table:baseline} shows the performance of each system on the test data. For the Attention and Stanford Readers, we anonymized the Who-did-What data by replacing named entities with entity IDs as in the CNN and Daily Mail datasets.
We see consistent reductions in accuracy when moving from CNN to our dataset. The Attentive and Stanford Reader drop by up to 10\% and the AS and GA reader drop by up to 17\%. The ranking of the systems also changes.
In contrast to the Attentive/Stanford readers, the AS/GA readers explicitly leverage the frequency of the answer in the passage, a heuristic which appears beneficial for the CNN and Daily Mail tasks. Our suppression of the most-frequent-person baseline appears to more strongly affect the performance of these latter systems.
\begin{table}
\centering
\begin{tabular}{@{}lrr@{}}
\toprule
System & WDW & CNN\\ \midrule
Word overlap & 0.47 & -- \\
Sliding window & 0.48 & -- \\
Distance & 0.46 & -- \\
Sliding window + Distance & 0.51 & -- \\
Semantic features & 0.52 & -- \\ \hdashline
Attentive Reader & 0.53 & $0.63^{I}$ \\
Attentive Reader (relaxed train) & 0.55 & \\
Stanford Reader & 0.64 & $0.73^{I\hspace{-.1em}I}$ \\
Stanford Reader (relaxed train) & 0.65 & \\
AS Reader & 0.57 & $0.70^{I\hspace{-.1em}I\hspace{-.1em}I}$ \\
AS Reader (relaxed train) & 0.59 & \\
GA Reader & 0.57 & $0.74^{I\hspace{-.1em}V}$ \\
GA Reader (relaxed train) & 0.60 & \\ \hline
Human Performance & $84/100$ & $0.75+^{I\hspace{-.1em}I}$ \\ \bottomrule
\end{tabular}
\caption{System performance on test set. Human performance was computed by two annotators on a sample of 100 questions. Result marked ${I}$ is from (Hermann et al., 2015), results marked ${I\hspace{-.1em}I}$ are from (Chen et al., 2016), result marked $I\hspace{-.1em}I\hspace{-.1em}I$ is from (Kadlec et al., 2016), and result marked $I\hspace{-.1em}V$ is from (Dhingra et al., 2016). }
\label{table:baseline}
\end{table}
\ignore{
\begin{table}
\centering
\caption{Performance of unsuppressed baselines on test set.}
\label{table:baseline}
\begin{tabular}{@{}lr@{}}
\toprule
Baselines & Accuracy \\ \midrule
Word overlap, monogram & 0.522 \\
Word overlap, bigram & 0.499 \\
Sliding window & 0.520 \\
Richardson Distance & 0.503 \\
Richardson SW+D & 0.540 \\
Attentive Reader & 0.581 \\
Human readers & $89/100$ \\ \bottomrule
\end{tabular}
\end{table}
}
\ignore{
\subsection{Performance of human readers}
Two native speakers of American English who are working in NLP research took a exam of our dataset. They took exams with 50 question for each and also annotated a difficulties of each question by 3 categories, "Confident", "Reasonably Confident", and "Guess". The scores for each human reader are 92\% and yy\%, zz\% in average. More than 90\% of questions are reasonably solvable by human readers. And xx\% of questions are solved by only reading, yy\% requires common knowledge and less than 10\% requires knowledge about a specific domain.
\begin{table*}
\centering
\caption{Difficulty and Correctness in 100 annotated questions}
\label{table:marginalAnnotation}
\begin{tabular}{@{}lccc@{}}
\toprule
& Confident & Reasonably Confident & Guess \\ \midrule
Correct & 65 & 22 & 2 \\
Incorrect & 3 & 4 & 4 \\ \bottomrule
\end{tabular}
\end{table*}
\begin{table*}[]
\centering
\caption{Required knowledge}
\label{table:reqKnowledge}
\begin{tabular}{@{}lccc@{}}
\toprule
& English & Common knowledge & Specific Knowledge in this domain \\ \midrule
\# quesitons & 74 & 22 & 4 \\ \bottomrule
\end{tabular}
\end{table*}
}
\ignore{
Two native speakers of American English who are working in NLP research took a exam of our dataset. We examined 99 questions and each readers annotated 66 questions. 33 of 66 questions are annotated both of 2 readers for a correlation. Readers also annotated how confident with his answer before he sees the correct answer.
We obtained 132 annotations from 99 questions as table(\ref{table:marginalAnnotation}). More than 85\% of questions are answered with "confident" or "reasonably confident" and the performance of human readers is more than 90\%. We found some questions which are answered with confident or reasonably confident but reader's answer is incorrect. They include "human error" and "misleading parse error".
The agreement of difficulties is on table (\ref{table:DiffCorrelation}). 12 questions of 33 are disagreed but the half of this disagreement happens between "confident" and "reasonably confident". Some questions are disagreed between "confident" and "guess" since a reader knows the news as a general knowledge and another does not.
We obtained more than 85\% agreement in correctness.
\begin{table*}
\centering
\caption{Performance of baseline systems on 88 reasonable questions (error : $\pm$ 0.06)}
\label{table:baseline}
\begin{tabular}{@{}ll@{}}
\toprule
Baselines & Accuracy \\ \midrule
Random & TBC \\
Sliding window & \\
Richardson SW+D & \\
Ave. human readers & \\ \bottomrule
\end{tabular}
\end{table*}
\begin{table}[]
\centering
\caption{Correlation of difficulty between readers A and B in 33 questions}
\label{table:DiffCorrelation}
\begin{tabular}{|rr|ccc|}
\hline
& & & \multicolumn{1}{c}{A} & \\
& & confident & \begin{tabular}[c]{@{}l@{}}reasonably\\ \quad confident\end{tabular} & guess \\ \hline
& confident & 17 & 1 & 1 \\
B & \begin{tabular}[c]{@{}l@{}}reasonably\\ \quad confident \end{tabular} & 5 & 2 & 1 \\
& guess & 3 & 1 & 2 \\ \hline
\end{tabular}
\end{table}
\begin{table}[]
\centering
\caption{Correlation of correctness between readers A and B in 33 questions}
\label{table:CrrCorrelation}
\begin{tabular}{|lr|cc|}
\hline
& & \multicolumn{2}{c|}{A} \\
& & correct & incorrect \\ \hline
\multirow{2}{*}{B} & correct & 29 & 2 \\
& incorrect & 1 & 1 \\ \hline
\end{tabular}
\end{table}
}
\section{Conclusion}
We presented a large-scale person-centered cloze dataset whose
scalability and flexibility is suitable for neural methods. This dataset is different in a variety of ways from existing large-scale cloze datasets
and provides a significant extension to the training and test data for machine comprehension.
\ignore{
Our dataset is answerable by human readers with xx\% and requires reading comprehension of text which is challenging for machines.
We published our dataset (www....) to help compare each systems of reading comprehension.
}
|
\section{Introduction}
\label{intro}
It is important to study the characteristics of plasma in presence of neutrinos, since such systems are important in understanding various physical phenomena during the evolution of early Universe as well as the systems like core core-collapsing supernovae and magnetars (see for e.g. \cite{volpe_14} for a brief overview). The presence of the cosmic neutrino background can influence cosmic microwave anisotropy and matter clustering \cite{komatsu_10,bashinsky_04} and it can also influence dynamics of the primordial magnetic field \cite{kandu_98,jedamzik_98,shaw_10}. There exists several studies in literature where the neutrino plasma interaction has been analysed in a variety of physical situations. Non-linear coupling of intense neutrino flux with collective plasma oscillations is studied in the Ref. \cite{Bingham_94}. Authors have shown that a neutrino flux as intense as that in supernovae core can cause parametric instabilities in the surrounding plasma. Effect of a neutrino medium in the evolution of lepton plasma had been studied invoking ponderomotive description \cite{Shukla_1997,Bethe_96}. In these cases it was shown that the ponderomotive force is proportional to the gradient of neutrino density and the electrons are repelled from the regions where neutrino density is large. Interaction of very large number of neutrinos with collective plasma and oscillation and the excitation of plasma turbulence is considered in the Ref. \cite{Tsytovich_98}. Different kinds of plasma neutrino interactions using the ponderomotive force description and the effect on collective plasma properties can be found in the references, \cite{Silva_99,Bento_99,Esposito_99,Brizard_00,Bento_01,Serbeto_02,Serbeto_fluid_02}. In the fore mentioned ponderomotive force description, it was assumed that the neutrino field satisfy the naive Klein-Gordon equation with appropriate interaction terms. Thus in this formalism the information about the chiral structure of the weak interaction is absent. Here we note that by Silva {et. al} in Refs. \cite{Silva_1999} and \cite{Silva_2006} the problem of neutrino driven streaming instability, which in turn can generate an significant energy transfer from neutrino to the plasma, was considered in the kinetic theory formalism.
Formulation to study the plasma interaction with intense neutrino beam using the field theory techniques is developed in \cite{Bento_99_recent}. Photon polarization tensor in a medium consistent with gauge and Lorentz invariance can be found in \cite{Palash_89}. In this work it is shown that, in presence of a medium, the photon polarization tensor can have anti-symmetric part indicating $P$ and $CP$ violations. Further studies of such effect in presence of neutrinos for different physical scenarios are explored in [\cite{Nieves_00},\cite{Nieves_05}].
In the context of early Universe, it has been shown by Shukla {\it at. el.} \cite{Shukla_1997}that, the ponderomotive force of non-uniform intense neutrino beam can be responsible for large scale quasi-stationary magnetic field. In fact, later was the first one to suggest magnetic field generation in plasma due to plasma-neutrino interactions. Further, large-scale magnetic field generation at the time of neutrino decoupling due to the evolution of plasma in presence of asymmetric neutrino background is studied in \cite{Boyarsky_12} and \cite{Doglov_02}. This field can act as a seed for generation of the galactic magnetic field via the galactic dynamo mechanism (see e.g \cite{Vainshtein_72} to read about galactic dynamo mechanism). It is to be noted that at finite lepton/baryon density the loop corrections to the photon polarization tensor are non-vanishing.
With these corrections the photon polarization tensor acquires a non-zero parity odd contribution $\Pi_2(\bm k)$ where, $\bm k$ is the wave vector.
A finite and non-zero values of $\Pi_2(\bm k)$ in the photon polarization tensor means that there can be single field derivative terms in the effective Lagrangian and free energy, which dominates the kinetic energy part of the free energy which is having double derivative term. For e.g. the free energy for a static gauge field can be written as $\mathcal{F}[A]=\int d^3pA_i(k)\Pi_{ij}(k)A_j(-k)$ and with parity violating interactions $\Pi_{ij}(k)$ can have a contribution $i\Pi_2(p^2)\epsilon_{ijl}k_l$. Thus a non-zero value of $\Pi_2(0)$ means a term $\Pi_2(0)\bm A\cdot\nabla\times\bm A$ in the expression for free energy.
This in turn means that there can be a generation of a large scale ($k\rightarrow 0$) magnetic field by an instability arising due to non-zero values of parity-odd contributions $\Pi_2(0)$ to the polarization tensor \cite{Boyarsky_12}. In the ref. \cite{Dornikov_15}, thermal field theory calculations were carried out to study the corrections to the photon polarization in presence of a background neutrino which is asymmetric in left-right number densities. Authors have shown that the axial part $\Pi_{2}$ is proportional to the neutrino asymmetry parameter and argued that the contribution to $\Pi_{2}$ due to the plasma which is interacting with the neutrino gas is $\sim 10^1$ times larger than the contribution to $\Pi_2$ through the correction due to the virtual process. In ref. \cite{Bhatt_16} using a kinetic theory approach
it was shown that the photon polarization tensor can have the parity-odd contribution
$\Pi_2(\bm k)$ due to the asymmetric neutrino background in both the collision less and
collision dominated regime. In the collision dominated regime the result for $\Pi_2(\bm k)$
using the kinetic approach agrees with that in ref. \cite{Dornikov_15}.
In a recent work \cite{Diaz_16} authors have calculated the effective potential or refractive index for the cosmic neutrino background (CNB) and future experimental implications have been discussed.
Further, recent theoretical calculations showed that the asymmetry in the neutrino density can be transmuted to the fluid helicity for sufficiently large electron neutrino interaction\cite{Yamamotto_16}. This neutrino induced vorticity can act as axial chemical potential for the chiral electrons. This phenomenon can induce {\it helical plasma instability} that generate strong magnetic field \cite{Yamamotto_16}. In this work the plasma particles
are considered to be massless and chirally polarized. Moreover, it was assumed, in this work, that the neutrino mean free path $l_\nu$ is much smaller than the system dynamics
at the length scale $L$ i.e. $L\gg l_\nu$. This allows one to write the equations for the neutrino
hydrodynamics \cite{Yamamotto_16}.
Though this assumption is justifiable for a core collapsing supernova, it is hard to be satisfied in other scenarios like the early Universe. Electroweak plasma in a rotating matter is studied in \cite{Dornikov_15}. In this work it is shown that electric current can be induced in the direction of rotation axis due to the parity violating nature of the interaction. This phenomenon is called \textit{galvano-rotational effect} (GRE). In a recent work \cite{Suvorov_16}, spin paramagnetic deformation of neutron star has been studied and authors have calculated the ellipticity of a strongly magnetized neutron star using the spin magneto hydrodynamic equations developed in \cite{Brodin_07}.
In the present work we are interested in developing magnetohydrodynamic description of the plasma
in presence of the left-right asymmetric neutrino background. The expression of the interaction Lagrangian of a charged lepton field and the asymmetric neutrinos suggests that the neutrino can couple with spin of the electron [\cite{Giunti_07},\cite{Dornikov_15}]. It is interesting to note here that there exist a lot of literatures in the usual electron-ion plasmas where the dynamics of spin
degree can play a significant role. For example it was suggested that a spin polarized
plasma in a fusion reactor can yield higher nuclear reaction cross section \cite{Kulsrud_82} and the spin depolarization process in the plasma can remain small \cite{Cowley_86}. Effect of spin dynamics using single particle description, valid for a dilute gas, is studied in the context of laser plasma interaction in the Ref. \cite{Walser_02}.
The collective effects within the framework of spin-magnetohydrodynamics has been studied
in Refs.\cite{Brodin_07,Mahajan_11} ( for general discussion see \cite{Haas_11}). These works can have applications in studying environments with a strong external magnetic field like pulsar and magnetars. In the present work we generalize the spin-magnetohydrodynamics considered in \cite{Brodin_07}
to incorporate the effect of asymmetric neutrino background.
The report is organised in the following way. In section \ref{lag_non_rel} we consider the low energy Lagrangian for our system and the equations of motion and spin evolution equations are derived invoking the non-relativistic approximations. MHD equations are considered in section \ref{hydro_eqns}. Velocity perturbations and electromagnetic perturbations in a magnetized plasma interacting with neutrino background is considered in section \ref{mode_analysis}.In section \ref{pul_kick} we apply our theory to neutron star to calculate the kick and section \ref{discussion} is about summary and conclusions. We provide a brief summery of relativistic generalization of the theory in appendix \ref{app_A}.
\section{The Lagrangian and non-relativistic approximation}
\label{lag_non_rel}
Lagrangian density for lepton field interacting with background neutrino is given by,
\begin{equation}
\mathcal{L}=\bar\psi[i\gamma^\mu\partial_\mu\psi-\gamma_\mu(f^\mu_LP_L+f^\mu_RP_R)-m]\psi
\label{lagrangian}
\end{equation}
where, m is mass of the lepton, $\gamma^\mu=(\gamma^0, \bm{\gamma})$ are the Dirac matrices and $P_{L,R}=\frac{1\mp\gamma^5}{2}$ are the
chiral projection operator with $\gamma^5=i\gamma^0\gamma^1\gamma^2\gamma^3 $. $f^\mu_{L,R}=(f^0_{L,R}, \bm{f}_{L,R})$
are the neutrino currents and they are regarded as an external macroscopic quantities.
An explicit form of $f^\mu_{L,R}$ can be calculated from effective Lagrangian (\cite{Giunti_07}, \cite{Dornikov_15})
\begin{equation}
\mathcal{L}_{eff}=[-\sqrt{2} G_F\sum_\alpha\bar\nu_\alpha\gamma^\mu\frac{(1-\gamma^5)}{2}\nu_\alpha][\bar\psi\gamma_\mu(a_L^\alpha P_L+a_R^\alpha P_R)\psi]
\label{Leff}
\end{equation}
where, label $\alpha$ denotes neutrino species $\alpha=e,\,\mu,\,\tau$ and $G_F=1.17\times10^{-11}MeV^{-2}$ is the Fermi constant. The coefficients $a_L^\alpha$ \& $a_R^\alpha$
are given by
\begin{equation}
a_L^\alpha=\delta_{\alpha,e}+sin^2{\theta_W}-1/2 , a_R^\alpha=
sin^2{\theta_W}
\label{aLR}
\end{equation},
with $\theta_W$ being the Weinberg angle. Next, we assume that $\nu \bar\nu$ form an isotropic background gas.
This in turn means that in averaging over the
neutrino ensemble, only non-zero quantity will be\, $<\bar\nu_\alpha\gamma^0(1-\gamma^5)\nu>=2(n_{\nu_\alpha}-n_{\bar\nu_\alpha})$. Number densities
of neutrinos and anti-neutrinos can be calculated using corresponding Fermi-Dirac distribution function
\begin{equation}
n_{\nu_\alpha , \bar\nu_\alpha}=\int \frac{d^3p}{(2\pi)^3}\frac{1}{e^{\beta_{\nu_\alpha}(|\bm p|\mp\mu_{\nu_\alpha} )}+1}
\label{FDdistribution}
\end{equation}
where $\beta$ is the inverse temperature. Using Eqns. (\ref{lagrangian}-\ref{FDdistribution}) one obtains
\begin{align}
f^0_L=&2\sqrt 2 G_F[\Delta n_{\nu_e}+(sin^2\theta_W-1/2)\sum_\alpha\Delta n_{\nu_\alpha})] ,\\ f^0_R=&2\sqrt 2 G_F sin^2
\theta_W\sum_\alpha\Delta n_{\nu_\alpha}.
\end{align}
Thus the equation of motion obtained from Eqn. (\ref{lagrangian}) can be written as
\begin{equation}
i\frac{\partial\psi}{\partial t}=[\bm\alpha\cdot\bm{\hat p}\psi+\beta m-(f^0_LP_L+f^0_RP_R)]\psi.
\label{eompsi}
\end{equation}
Writing $\psi=\left(\begin {array}{c} \phi \\ \chi \end {array}\right ) $ in the Eqn.(\ref{eompsi}) and following the standard procedure \cite{Greiner_reqm}, Hamiltonian
for the large component of the spinor can be obtained as,
\begin{equation}
\mathcal{ H}=\frac{1}{2m}\bm{(\sigma\cdot p)(\sigma\cdot p)}+\frac{\Delta f^0}{2m}\bm{(\sigma\cdot p)}+\frac{f^0}{2}+O(f_{L,R}^2)
\nonumber
\end{equation}
where, $f^0=f^0_L+f^0_R$ and $\Delta f^0=f^0_L-f^0_R$. In the above equation, we have neglected terms proportional to $G_F^2$.
In the presence of external electromagnetic field, momentum $\bm p$ has to be
replaced by $\bm{p-eA}$. Thus the Hamiltonian for charged fermion in interacting with an external electromagnetic
field and background neutrino is given by,
\begin{equation}
\mathcal {H}=
\frac{(\bm{p-eA})^2}{2m}-\bm{\mu\cdot B}+eA^0+\frac{\Delta f^0}{2m}\bm\sigma\cdot(\bm p-e\bm A)+\frac{f^0}{2}
\label{H}
\end{equation}
\noindent
where, $\bm\mu=\frac{eg}{4m}\bm\sigma$ is the electron magnetic moment and $g$ is the Land\`{e} g-factor. The first three terms on the right hand
side are well known and very well studied in the literature. The fourth and fifth terms are due to the neutrino background. The last term
might contribute
to the energy of the system, but it will not enter into the equations of motion as the neutrino background considered to be constant.
If the neutrino background vary with space and time, this term would modify force
equation as $\bm F\propto \bm\nabla f=\bm\nabla\psi_\nu^*\psi_\nu$. This force is called ponderomotive force. Such a scenario was studied
in Ref. \cite{Bethe_96}, however in their formalism
the fourth term was not considered.
In order to find the equation of motion
for a charged particle in an electromagnetic
field and the neutrino background, one can
use Eqn.(\ref{H}) and the
Heisenberg equation $\dot{\hat O}= i[\hat {\mathcal{H}},\hat O]$ and write:
\begin{equation}
\bm v=\dfrac{\bm p-e\bm A}{m}+\frac{\Delta f^0}{2m}\bm\sigma
\label{dotx}
\end{equation}
where we wrote $\dot{\bm x}=\bm v$.
\begin{equation}
\dot{\bm p}=\frac{e}{m}(\bm p-e\bm A)_k\bm\nabla\bm A_k+\frac{eg}{m}\bm\nabla\bm{(s\cdot B)}-e\bm{\nabla}A^0
\label{dotp}
\end{equation}
where we have defined $\bm s=\bm\sigma /2$ and
\begin{equation}
\dot{\bm s}=\mu_B(\bm s\times\bm B)-\Delta f^0(\bm s\times\bm v)
\label{spin_dyn}
\end{equation}
From the equations \ref{dotx}-\ref{spin_dyn} we get
\begin{equation}
\ddot{\bm x}=\frac{e}{m}[\bm E+\bm v\times\bm B]+\frac{e\Delta f^0}{2m^2}(\bm s\times \bm B)+\frac{eg}{2m^2}\bm\nabla(\bm s\cdot\bm B)
\label{ddotx}
\end{equation}
\section{The hydrodynamic equations}
\label{hydro_eqns}
In this section we follow the methods developed in Ref. \cite{Brodin_07} to derive the hydrodynamic equations from the quantum Lagrangian for spin half particles. We consider a system of electrons and ions in presence of homogeneous neutrino background. The neutrino background assumed to have a left-right asymmetry. Furthermore we treat electrons as quantum particles and ions as classical particles so that we can neglect the spin dynamics and other quantum effects for ions.
For simplicity, first let us Consider the Eqn. (\ref{H}) without the neutrino interaction term. We can decompose the wave function as $\psi_\alpha =\sqrt{n_\alpha} e^{iS_\alpha}\chi_\alpha$, where $n_\alpha$ is the density, $S_\alpha$ is the phase and $\chi_\alpha$ is a two component spinor in which the spin-1/2 information is contained. Inserting this decomposition and considering the real and imaginary parts of the resulting equation we get the continuity and momentum conservation equation for the "species $\alpha$" as,
\begin{equation}
\frac{\partial n_\alpha}{\partial t}+\nabla \cdot (n_\alpha \bm{\upsilon}_\alpha)=0
\label{cont_spe}
\end{equation}
and
\begin{align}
&\nonumber m_\alpha(\frac{\partial}{\partial t}+\bm\upsilon_\alpha\cdot\bm{\nabla})\bm\upsilon_\alpha =
q_\alpha(\bm E+\bm\upsilon_\alpha\times\bm B) +\\ & 2\mu(\nabla\otimes\bm B)\cdot\bm s_\alpha-\nabla Q_\alpha-\frac{1}{m_\alpha n_\alpha}\nabla\cdot (n_\alpha
\bm\Sigma_\alpha)
\label{mom_cons_spe}
\end{align}
Velocity is defined via $m_\alpha\bm\upsilon_\alpha=\bm j_\alpha/\psi^\dagger\psi$, from which we obtain
\begin{equation}
m_\alpha\bm\upsilon_\alpha=(\nabla S-i\chi_\alpha^\dagger\nabla\chi_\alpha)-q_\alpha\bm A
\label{vel_spe_def}
\end{equation}
and,
\begin{equation}
\bm s_\alpha=\frac{1}{2}\chi_\alpha^\dagger\bm\sigma\chi_\alpha
\end{equation}
The quantity $Q_\alpha$ is known as the quantum potential(Bohm potential) defined as
\begin{equation}
Q_\alpha=-\frac{1}{2m_\alpha\sqrt n_\alpha}\nabla^2\sqrt n_\alpha
\end{equation}
and $\bm \Sigma_\alpha$ is the symmetric spin gradient tensor.
\begin{equation}
\bm \Sigma_\alpha=\nabla\bm s_{(\alpha)a}\otimes\nabla\bm s_{(\alpha)}^a
\end{equation}
where $a=1,2,3$. By contracting the Pauli equation with $\psi^\dagger\bm\sigma$, one can obtain the spin evolution equation
as
\begin{equation}
(\frac{\partial}{\partial t}+\bm\upsilon_\alpha\cdot\bm{\nabla})\bm s_\alpha=2\mu (\bm s_\alpha\times\bm B)+\frac{\bm s_\alpha\times[\partial_a(n_\alpha\partial^a\bm s_\alpha)]}{m_\alpha n_\alpha}
\label{spin_evo_spe}
\end{equation}
In presence of neutrino background, the continuity equation remain unchanged. But both momentum conservation and spin evolution equations are modified in the following way.
\begin{align}
&\nonumber m_\alpha(\frac{\partial}{\partial t}+\bm\upsilon_\alpha\cdot\bm{\nabla})\bm\upsilon_\alpha =
q_\alpha(\bm E+\bm\upsilon_\alpha\times\bm B) +\\ & 2\mu(\nabla\otimes\bm B)\cdot\bm s_\alpha-\nabla Q_\alpha-\frac{1}{m_\alpha n_\alpha}\nabla\cdot (n_\alpha\bm\Sigma_\alpha)+\frac{\Delta f^0}{2m}\bm s_\alpha\times\bm B
\label{mom_cons_spe_nu}
\end{align}
and
\begin{equation}
(\frac{\partial}{\partial t}+\bm\upsilon_\alpha\cdot\bm{\nabla})\bm s_\alpha=2\mu (\bm s_\alpha\times\bm B)-\frac{\Delta f^0}{2}\bm s_\alpha\times\bm\upsilon_\alpha+\frac{\bm s_\alpha\times[\partial_a(n_\alpha\partial^a\bm s_\alpha)]}{m_\alpha n_\alpha}
\label{spin_evo_spe_nu}
\end{equation}
Now, in order to define hydrodynamic quantities, we need to specify how to calculate the expectation values. Suppose that we have N wave function with same kind of particles with magnetic moment $\mu$ charge $q$ and mass $m$ so that the wave function for the entire system can be factorised as $\psi=\psi_{(1)}\psi{(2)}...\psi_{(N)}$. Then we can define the total particle density for charge $q$ as, $n_q=\sum_{\alpha}n_\alpha$ and the expectation value of any quantity $f$ as $<f>=\sum_\alpha\frac{n_\alpha}{n_q}f$.
Using these arguments we define the total fluid velocity $\bm V_q=<\bm\upsilon_\alpha>$ and $\bm S_q=<\bm s_\alpha>$. In order to simplify further calculations, we redefine these quantities such that $\bm w_\alpha=\bm \upsilon_\alpha-\bm V_q$ and $\bm{\mathcal S}_\alpha=\bm s_\alpha-\bm S_q$, satisfying $<\bm w_\alpha>=0$ and $<\bm{\mathcal S}_\alpha>=0$. Now taking the ensemble average of the equations (\ref{cont_spe}), (\ref{mom_cons_spe_nu}) and (\ref{spin_evo_spe_nu}) we get the following expressions.
\begin{equation}
\frac{\partial n_q}{\partial t}+\nabla\cdot(n_q\bm V_q)=0
\end{equation}
\begin{equation}
m_q n_q\Big(\frac{\partial}{\partial t}+\bm V_q\cdot\nabla\Big)\bm V_q=qn_q\Big(\bm E+\bm V_q\times\bm B\Big)-\nabla\cdot\bm\Pi-\nabla P+\bm{\mathcal{C}}_{qi}+\bm F_Q+\bm F_{\nu e}
\label{Hydro_eqn_V_q}
\end{equation}
and
\begin{equation}
n_q\Big(\frac{\partial}{\partial t}+\bm V_q\cdot \nabla\Big)\bm S_q=2\mu_B n_q\bm S_q\times\bm B-\frac{\Delta f^0}{2}\bm S_q\times \bm V_q+\bm\Omega_s-\nabla\cdot \bm {K}_q+\bm K_{\nu e}
\label{Hydro_eqn_S_q}
\end{equation}
Where, $\bm\Pi$ is the traceless anisotropic part of the pressure tensor and $P$ is the homogeneous part. $\mathcal{C}_{qi}$ represents the collision between particle with charge $q$ and ion denoted using the letter $i$ and the quantum force density $\bm F_Q$ and force due to the interaction with the neutrino back ground $\bm F_{\nu e}$ has the definitions,
\begin{align}
\bm F_Q&=2\mu_B n_q\Big(\nabla\otimes\bm B\Big)\cdot\bm{S}_q-n_q\Big<\nabla Q_\alpha \Big>-\frac{1}{m}\nabla\cdot\Big(n_q\bm\Sigma\Big)-\frac{1}{m}\nabla\cdot\Big(n_q\tilde\Sigma\Big) \label{F_Q}
\\
\nonumber &-\frac{1}{m}\nabla\cdot\Big[\Big(\nabla\bm S_a\Big)\otimes\Big<\nabla\mathcal{S}^{a}_\alpha\Big>+n_q\Big<\nabla\mathcal{S}^{a}_\alpha\Big>\otimes\Big(\nabla\bm S_a\Big)\Big]
\end{align}
and
\begin{equation}
\bm F_{\nu e}=n_ee\frac{\Delta f^0}{2m}\bf S_q\times \bm B
\end{equation}
The quantities $\bm \Omega_s, \bm\Sigma$ and $\bm{\tilde\Sigma}$ depends on the spin of the particles and their precise definitions can be found in the Ref. \cite{Brodin_07}. $\bm K_q=\Big<\mathcal{\bm S}_{\alpha }\otimes \bm w_{\alpha }\Big>$ is the spin thermal coupling and $\bm K_{\nu e}=\epsilon_{ijk}\frac{\delta f^0}{2}\Big<\mathcal{\bm S}_{\alpha j}\bm w_{\alpha k}\Big >$ is the thermal spin coupling induced by neutrino interaction.
In the following sections, we will replace the subscript $q$ with $e$ and $i$ for electrons and ions respectively. Since we are considering ions as classical particle, we can neglect the contributions from spin and other quantum effects for ions. Thus, the fluid equations for ions read,
\begin{equation}
\frac{\partial n_i}{\partial t}+\nabla\cdot(n_i\bm V_i)=0
\label{cont_ion}
\end{equation}
\begin{equation}
m_i n_i\Big(\frac{\partial}{\partial t}+\bm V_i\cdot\nabla\Big)\bm V_i=q_in_i\Big(\bm E+\bm V_i\times\bm B\Big)-\nabla\cdot\bm\Pi_i-\nabla P_i+\bm{\mathcal{C}}_{iq}
\label{mom_ion}
\end{equation}
Note that there is no spin evolution equation for ions. Therefore whatever spin contributions governing the dynamics of the system are only due to the spin of the electrons. Now we can construct the single fluid equations from the above equations for electrons and ions. In order to do that we define the total mass density, $\rho=(m_en_e+m_in_i)$, the centre of mass velocity of the fluid $\rho\bm V=(m_en_e \bm V_e+m_in_i\bm V_i)$ and the current density $\bm j=(-en_e\bm V_e+Zen_i\bm V_i)$ and assuming quasi-neutrality $n_e=Zn_i$, one can immediately obtain the continuity equation,
\begin{equation}
\frac{\partial\rho}{\partial t}+\nabla\cdot(\rho\bm V)=0
\label{cont_flu}
\end{equation}
and momentum conservation equation
\begin{equation}
\rho\Big(\frac{\partial}{\partial t}+\bm V\cdot\nabla\Big)\bm V=\bm j\times\bm B-\nabla\cdot\bm\Pi-\nabla P+\bm{F}_Q+\bm{F}_{\nu e}
\end{equation}
Note that, with the assumption of quasi-neutrality we can write $n_e=\rho /(me_e+m_i)$ and $\bm V_e=\bm V-m_i\bm j/Ze\rho$. There fore we can express the quantum terms in terms of the total density and , centre of mass velocity of the fluid and current. Thus the spin evolution equation becomes,
\begin{align}
\rho\Big(\frac{\partial}{\partial t}+\bm V\cdot\nabla\Big)\bm S&=\frac{m_i}{Ze\rho}\bm j\cdot\nabla \bm S+2\mu\rho\bm S\times\bm B-\Big(m_e+m_i/Z\Big)\nabla\cdot\bm K_e+\Big(m_e+m_i/Z\Big)\bm{\Omega}_s\\\nonumber &-\frac{e\Delta f^0\rho}{2m_e} \bm S\times\Big(\bm V-\frac{m_i\bm j}{Ze\rho}\Big)-\frac{e\Delta f^0}{2m_e}\Big<\bm w_e\times\mathcal{\bm S}_e\Big>
\end{align}
in general, for a magnetized medium with magnetization density $\bm M$ we can write the free current density $\bm j=\frac{\nabla\times\bm B}{\mu_0}-\bm j_M$, where $\bm {j}_M=\nabla\times \bm M$ is the magnetization current density. Note that, here we have discarded the displacement current term $\frac{\partial\bm E}{\partial t}$.
In order to simplify the further calculation, we consider only the transverse waves
in that case the Bohm potential i.e. $<Q_\alpha>$ term in Eqn.(\ref{F_Q}) can be dropped \cite{Haas_11}. Further all the other terms in Eqn.(\ref{F_Q}) are
are second order in the spin variable and of order $\hbar^2$. We neglect these terms.
However, $\bm F_{\nu e}$ term in Eqn. (\ref{Hydro_eqn_V_q}) and the spin dynamic Eqn. (\ref{Hydro_eqn_S_q}) are order $\hbar$. These
terms are retained in the calculation. In such a situation can write the total force density exerted on the fluid element as,
\begin{equation}
F^i=-\partial^i\Big(\frac{B_0^2}{2}-\bm B\bm\cdot M\Big)+\bm B\cdot\nabla H^i-\partial^i P-\partial^j \Pi^{ij}
\end{equation}
For an isotropic plasma, the trace-free part of the pressure tensor $\Pi^{ij}$ is zero. It is worth noting that the spacial part of the stress tensor take the form $T^{ij}=-H^iB^j+\big(B^2/2-\bm B\cdot \bm M\big)\delta^{ij}$ \cite{Robinson_00}, apart from the pressure terms. Thus the total force density on a magnetized fluid element can be written as, $F^i=-\partial^j T^{ij}$. Therefore the momentum conservation equation takes the form,
\begin{equation}
\rho\Big(\frac{\partial}{\partial t}+\bm V\cdot\nabla\Big)\bm V=-\nabla\Big(\frac{B^2}{2}-\bm B\bm\cdot \bm M\Big)+\Big(\bm B\cdot\nabla\Big)\bm H-\nabla P-\nabla\cdot\bm\Pi
\label{mom_approx_fluid}
\end{equation}
Following the procedure in Ref. \cite{Boyd_69} we can write,
\begin{equation}
\bm j\sim\frac{\sigma m_i}{\rho e}\nabla P+\sigma\Big(
\bm E+\bm V\times\bm B\Big)+\frac{\sigma m_i}{\rho e}\bm j\times\bm B+\frac{\sigma m_i}{\rho e}\bm {F}_Q-n_0\frac{\sigma m_i}{\rho}\Big(\frac{\Delta f^0}{2m_e}\Big)\bm S\times\bm B
\label{j_fluid}
\end{equation}
Taking $\rho\sim n_0 m_i$ the expression for total current can be written as,
\begin{equation}
\bm j\sim\sigma \Big(\bm E+\bm V\times \bm B\Big)-\sigma\Big(\frac{\Delta f^0}{2m_e}\Big)\bm S\times\bm B+\bm{j}_M
\end{equation}
For the above expression for the hydrodynamic current, the time evolution for the magnetic field $\bm B$ is given by
\begin{equation}
\frac{\partial\bm B}{\partial t}=-\eta\nabla\times\Big(\nabla\times \bm B\Big)+\nabla\times\Big(\bm V\times\bm B\Big)-\Big(\frac{\Delta f^0}{2m_e}\Big)\nabla\times\Big(\bm S\times\bm B\Big)+\eta\nabla\times\bm{j}_M
\label{B_with _finite_eta}
\end{equation}
Where, $\eta=1/\sigma$ is the resistivity.
\section{Neutrino induced vorticity, Alfv\'en wave and normal modes }
\label{mode_analysis}
In this section, we consider a very simple scenario. A background magnetic field $\bm B_0=B_0 \bm{\hat z}$ is applied to the plasma. As a result, there is a non-zero constant magnetization in the system even in the absence of any perturbations, which also imply that $\bm S\times \bm B=0$ for the plasma at equilibrium. In this case the spin of the electrons align anti-parallel to the magnetic field to reduce the energy and there fore we can assume the equilibrium magnetization density $\bm M_0$ to take the form $\bm M_0=-\mu_B n_e \bm S_0=\mu_Bn_e\xi\big(\frac{\mu_B B_0}{T_e}\big)\bm{\hat z}$, where $\xi (x)=tanh(x)$ is the Brillouin function. For the following discussions we make the approximation $\xi\big(\frac{\mu_B B_0}{T_e}\big)\sim \big(\frac{\mu_B B_0}{T_e}\big)$ so that $\bm S_0\sim -\frac{1}{2}\frac{\mu_BB_0}{T}\bm{\hat z}$. Furthermore we assume that there are no electromagnetic perturbations in the system and the fluid velocity enters into the governing equations as perturbation. That is, $\bm E=0, \bm B= B_0\bm{\hat z}, \bm V=\delta\bm V$ and $\bm S=\bm S_0+\bm \delta \bm S$. With these assumptions, up to linear order in perturbations, we use the hydrodynamic equations in the following form.
\begin{equation}
\frac{\partial\delta\bm S}{\partial t}=2\mu_B\delta\bm S\times\bm B_0-\Big(\frac{\Delta f^0}{2m_e}\Big)\bm S_0\times\delta\bm V
\label{spin_evo_no_em}
\end{equation}
\begin{equation}
\rho_0\frac{\partial\delta\bm V}{\partial t}=-\frac{\mu_Bn_e}{T}\Delta f^0\nabla(\bm B_0\cdot\delta\bm V)+
\frac{\mu_Bn_e}{T}\Delta f^0(\bm B_0\cdot\nabla)\delta \bm V
\label{mom_con_no_em}
\end{equation}
From the equation (\ref{spin_evo_no_em}) we get $\frac{\partial\delta\bm S\cdot\bm B_0}{\partial t}=0$. In order to satisfy this conditions, we choose $\delta\bm S\cdot\bm B_0=0$. We also take the space-time dependence of the perturbations to be of the following form,
\begin{equation}
\delta\bm V(t,\bm x)=\delta\bm V_{\omega,\bm k}e^{-i\big(\omega t-\bm k\cdot\bm x\big)},~~~~ \delta\bm S(t,\bm x)=\delta\bm S_{\omega,\bm k}e^{-i\big(\omega t-\bm k\cdot\bm x\big)}
\end{equation}
With these assumptions we get,
\begin{equation}
\delta\bm S_{\omega,\bm k}=-\Big(\frac{\Delta f^0}{2\Omega_e}\Big)|\bm S_0|\delta\bm V_{\omega,\bm k}
\label{spin_no_em}
\end{equation}
Where, $\Omega_e=\frac{eB_0}{m_e}$. To obtain the above expression we have assumed that $\frac{\omega^2}{\Omega_e^2}\ll 1$. Thus, from the equations (\ref{mom_con_no_em}), and (\ref{spin_no_em}), we get
\begin{equation}
-i\omega \delta\bm V_{\omega,\bm k}=\Big(\frac{\mu_Bn_e}{T}\Delta f^0\Big)\bm\Omega_{\omega,\bm k}\times\bm{B}_0
\label{neu_ind_vor_no_em}
\end{equation}
Where, $\bm\Omega_{\omega,\bm k}=i\bm k\times\delta\bm V_{\omega,\bm k}$ is the vorticity in the Fourier space. Note that, in the above expressions, we have kept the terms only up to linear in $\Delta f^0$. From the above expression we can see that vorticity term will not contribute to the fluid dynamics if $\Delta f^0=0$. Therefore we conclude that $\bm\Omega$ is induced via the electron neutrino interaction . From the Eqn. (\ref{neu_ind_vor_no_em}) we can obtain the dispersion relations. For the case $\bm k ||\bm B_0$,
\begin{equation}
\omega=-\Big(\frac{\mu_B n_e B_0}{\rho_0}\Big)\Big(\frac{\Delta f^0}{T}\Big)k
\label{neu_ind_vor_no_em_w_para}
\end{equation}
Group velocity of this new mode is given by,
\begin{align}
v_g=&\bigg|\frac{d\omega}{dk}\bigg|
=\bigg |\Big(\frac{\mu_Bn_e B_0}{\rho_0}\Big)\Big(\frac{\Delta f^0}{T}\Big)\bigg| \\
\sim & 2\sqrt 2\Big(\frac{\mu_Bn_e B_0}{\rho_0}\Big)\Big(\frac{G_F}{T}\Big)\bigg | n_{\nu e}-n_{\bar\nu e}\bigg |
\label{new_mod_vg}
\end{align}
The Eqn. (\ref{neu_ind_vor_no_em_w_para}) corresponds to a new type of transverse mode propagating in the direction parallel to the background magnetic field, \textit{induced} by asymmetry in the neutrino background. The wave velocity not just depend on the strength of magnetic field but also on the neutrino asymmetry. This new mode
is similar to the one found in very high energy plasma with the chiral-anomaly \cite{Yamamotto_16}. In contrast to Ref. \cite{Yamamotto_16}, in our work
the electrons are not considered to be chirally polarized. However, the parity violating
interaction in our work arises due to neutrino-electron interaction. Further, the effect
of dissipation can easily be introduced by incorporating contribution of the finite shear viscosity $-ik^2\eta_{vis}$ and the resistivity $-i\sigma_1 B^2_0$ into the dispersion relation (\ref{neu_ind_vor_no_em_w_para}), where $\eta_{vis}$ is the kinematic viscosity and
$\sigma_1=\sigma/\rho_0$ with $\sigma$ being the resistivity.
Next, we consider the effect of electro-magnetic perturbations. That is we take the perturbations in the following form.
\begin{equation}
\bm V=\delta\bm V,~~\bm B=B_0\bm{\hat z}+\delta\bm B,~~ \bm E=\delta\bm E.
\end{equation}
For this case, linearized hydrodynamic equations, Eqn.(\ref{cont_flu}) and (\ref{mom_approx_fluid}) takes the form,
\begin{equation}
\frac{\partial\rho}{\partial t}+\rho_0\nabla\cdot\delta\bm V=0
\label{cont_with_em}
\end{equation}
\begin{equation}
\rho_0\frac{\partial\delta\bm V}{\partial t}=-\nabla\Big(\bm {B}_0\cdot\delta\bm B-\bm {M}_0\cdot\delta
\bm B-\delta\bm M\cdot\bm {B}_0\Big)+\bm{B}_0\cdot\nabla\delta\bm H-\nabla P
\label{mom_with_em}
\end{equation}
And the spin evolution equation becomes,
\begin{equation}
\Big(\frac{\partial\bm S}{\partial t}\Big)=2\mu_B\bm S\times\bm B-\frac{\Delta f^0}{2}\bm S\times\delta\bm V
\end{equation}
Where, $\bm S=\bm{S}_0+\delta\bm S$. For a perfectly conducting medium $(\eta\rightarrow 0)$, the equation (\ref{B_with _finite_eta}) becomes,
\begin{equation}
\frac{\partial\delta\bm B}{\partial t}=\nabla\times\Big(\delta\bm V\times\bm {B}_0\Big)-\Big(\frac{\Delta f}{2m_e}\Big)\nabla\times\Big(\delta\bm S\times\bm {B}_0+\bm {S}_0\times\delta\bm B\Big)
\label{evo_delta_b_with_em}
\end{equation}
Following the same procedure in the last section with same assumptions, we get the expression for $\delta\bm S$ as,
\begin{equation}
\delta\bm S_{\omega,\bm k}=\frac{\mu_B}{T}\delta\bm B_{\omega,\bm k}-\frac{\Delta f^0}{T}\delta\bm V_{\omega,\bm k}
\label{spin_with_em}
\end{equation}
Where, $\omega_p^2$ is the plasma frequency. Using the equations (\ref{cont_with_em}), (\ref{mom_with_em}) and (\ref{spin_with_em}) and using the approximation $\bm M_0=-\mu_B n_e \bm S_0=\mu_Bn_e\eta\big(\frac{\mu_B B_0}{T_e}\big)\bm{\hat z}$ we get,
\begin{equation}
-i\omega\rho_0\delta\bm V_{\omega, \bm k}=i\Big[\frac{\omega_p^2}{m_eT}\Big]\bm {B}_0\times\Big(\bm k\times\delta\bm B_{\bm k,\omega}\Big)-\Big[\frac{\mu_Bn_e}{T}\Delta f^0\Big]\bm {B}_0\times\bm\Omega_{\bm k,\omega}
\label{evo_delta_V_with_em}
\end{equation}
Where, $\bm\Omega_{\bm k,\omega}=i\bm k\times\delta\bm V$ is the vorticity in the Fourier space. We can see that the last term in the Eqn.(\ref{evo_delta_V_with_em} ) is proportional to the neutrino asymmetry of the background. expression for velocity in the Fourier space as,
\begin{equation}
\delta\bm V_{\bm k,\omega}=\Bigg(\frac{\bm B_0\cdot\bm k}{\rho_0\omega^2}\Big[\frac{\omega_p^2}{m_eT}-1 \Big]+\frac{1}{\omega}\frac{\mu_B n_e}{\rho_0 T}\Delta f^0\Bigg)\Bigg((\bm B_0\cdot\delta\bm V_{\omega,\bm k})\bm k-(\bm B_0\bm\cdot \bm k)\delta\bm V_{\omega,\bm k} \Bigg)
\end{equation}
Note that, we have neglected the contributions from the pressure terms in the above expression. Taking $\bm k$ in the direction of background magnetic field and assuming $\bm {B}_0\cdot \delta\bm V=0$, we get the following dispersion relation,
\begin{equation}
\omega=-\frac{\tilde v_A}{\sqrt{\rho_0\alpha}}\frac{\mu_B n_e}{2T}\Delta f^0 k\pm \tilde v_A k
\label{omega_no_eta_with_em}
\end{equation}
where, $\alpha=(1-\frac{\omega_p^2}{m_eT})$ and $\tilde v_A=v_A\alpha^{1/2}$ is the spin-modified Alfv\'en velocity \cite{Brodin_07}. Here we note that the quantity $\alpha$ describes the spin corrections and in the absence of spin dynamics $\alpha=1$. It is clear from the Eqn. (\ref{omega_no_eta_with_em}) that the group velocity $v_g$ can have two values given by,
\begin{equation}
v_g^{\pm}=\tilde v_A\bigg|1\pm\frac{1}{\sqrt{\rho_0\alpha}}\frac{\mu_B n_e}{2T}\Delta f^0 \bigg|
\label{group_v}
\end{equation}
which is absent in the absence of any neutrino asymmetry ($\Delta f^0=0)$. Thus we can have two different group velocities for the Alfv\'en waves propagating parallel or anti-parallel to $\bm B_0$.
For finite value of conductivity, we have to take into account of the first and last terms of the Eqn.(\ref{B_with _finite_eta}) and the dispersion relation can be obtained from,
\begin{equation}
\omega^2+\omega\Big[i\alpha\eta k^2+\frac{\mu_Bn_e\Delta f^0}{T\sqrt{\rho_0\alpha}}\tilde v_Ak\Big]-\tilde v_A^2k^2=0
\end{equation}
Solving for $\omega$ we get,
\begin{equation}
\omega=-\frac{1}{2}\bigg[i\alpha\eta k^2+\frac{\mu_B n_e \tilde v_A}{T\sqrt{\rho_0\alpha}}\Delta f^0 k\bigg]\pm\frac{1}{2}\sqrt{\bigg[i\alpha\eta k^2+\frac{\mu_B n_e \tilde v_A}{T\sqrt{\rho_0\alpha}}\Delta f^0 k\bigg]^2+4\tilde v_A^2k^2}
\label{omega_eqn_with_eta}
\end{equation}
We can see that in the absence of any neutrino asymmetry and $\eta$, the Eqn.(\ref{omega_eqn_with_eta}) reduces to $\omega^2=\tilde v_A^2k^2$, which is the same in magneto hydrodynamics with spin corrections as obtained in \cite{Brodin_07}.
\section{Neutrino asymmetry and the pulsar kick}
\label{pul_kick}
We use our formalism for a qualitative calculation of observed pulsar kick \cite{Minkowski_1970,Lyne_1982,Hansen_1997}. There are several attempts to explain the reason for the kick, for eg. see the references \cite{Gott_1970,Iben_1996,Kusenko_1999,Kusenko_2004}. Recently there have been attempts to explain the pulsar kick using anomalous hydrodynamic theories (for eg. see Ref.\cite{kaminski2016anomalous}), but the exact reason for the pulsar kick is not yet resolved.
We note that, the energy flux associated with the wave is equal to the energy density in the wave times the group velocity \cite{Stix_1992}, which is the Poynting vector $\bm P=\bm E\times\bm B$ in our case \cite{Stix_1992,Cramer_2011}. The Poynting vector can be expressed in the form,
\begin{equation}
\bm P=(\omega A^2)\bm k
\label{Poynting_defi}
\end{equation}
where $A$ is the magnitude of the vector potential $\bm A_{\omega,\bm k}$. Using the Eqn. (\ref{omega_no_eta_with_em}) we write,
\begin{align}
|\bm P|&=k^2 A^2\tilde v_A\Big(1\pm\frac{1}{\sqrt{\rho_0\alpha}}\frac{\mu_B n_e}{2T}\Delta f^0 \Big)\label{mod_flux}\\
&=\Big(k^2A^2\Big)v_g\label{E_den_times_vg}
\end{align}
From Eqn. (\ref{E_den_times_vg}), we infer that the energy density associated with the wave is $k^2A^2$. Further we note from the Eqn. (\ref{mod_flux}) that the energy transported in the direction of the background field $\bm B_0$ and opposite to $\bm B_0$ are different due to the parity violation within the system. An excess amount of energy is transported in the direction of magnetic field. This excess amount of energy transported per unit area per unit time is given by,
\begin{equation}
\Delta P=\Big(k^2A^2\tilde v_A\Big)\Big(\frac{\Delta f^0}{T}\frac{\mu_Bn_e}{\sqrt{\rho_0\alpha}}\Big)
\label{Delta_P}
\end{equation}
Which is essentially the momentum carried by the \textit{excess} photons leaving the pulsar per unit area per unit time. Therefore the change in velocity experienced by the pulsar can be expressed as,
\begin{equation}
\Delta V_{NS}=\frac{\Delta P}{M_{NS}}\times \Delta t\times (area)
\label{Delta V_NS_defi}
\end{equation}
Where $M_{NS}\sim 10^{30}~kg$ is the mass of neutron star and $\Delta t$ is the time span we assume for the kick to last, which is approximately $10$ seconds. The radius of the neutron star $R_{NS}$ is approximately $10~km$. Taking $k\sim A\sim T$, $\Delta n_{\nu e}\sim 1.6\times 10^{8}~(MeV)^3$, $T\sim 10^{12}~K $ and $B_0\sim(10^{15}-10^{16})~Gauss$
\cite{Andreas_Reisenegger}, we get $\Delta V_{NS}\sim (10^2-10^3)~km/s$, which is within the order of magnitude of observed pulsar kicks.
\section{Discussion and conclusion}
\label{discussion}
In conclusion we have developed spin magnetohydrodynamic equations in the presence of asymmetric background neutrinos and analysed the normal modes of the plasma
in presence of a constant magnetic field. We have shown that a
new kind of wave (Alfv\'en) is generated Eqn.\ref{neu_ind_vor_no_em_w_para} can exist
whose velocity depends on the neutrino asymmetry. Such a wave can be generated in
a dense astrophysical plasma such as magnetar. For example
for $B_0^{15}$ Gauss, $T\sim 10 MeV$ and $\Delta n_{\nu e}\sim 1.6\times 10^{8}$(MeV)$^3$, one can estimate the velocity
of the wave (in units of speed of light) around $10^{-5}$. Similarly for the Alfv\'en
waves (Eqn. (\ref{group_v}) ) can have two different speeds.
We have shown that the background neutrino asymmetry can change the wave-velocity in directions parallel and anti-parallel to the external magnetic field (as shown in Eqn. (\ref{group_v})). We have used our formalism to calculate the kick received by pulsar during its birth. An order of magnitude calculation matches with the observations $\Delta V_{NS}\simeq(10^2-10^3)~km/s$. In the appendix we have
derived the relativistic hydrodynamical equation for the electrons using Dirac equation. For the case when the electrons
are relativistic the estimate given here for the Alfv\'en velocity can be suppressed by a factor $1/\sqrt{2}$.
\begin{appendix}
\section{}
\label{app_A}
In the many astrophysical situations it is necessary to consider the system temperature to be greater
than its rest-mass and therefore we discuss relativistic generalization of the electron fluid.
For such a generalization in the context of quantum plasma one needs to start with the Dirac equation.
Works by Pauli \cite{WP_36}, Harish Chandra \cite{HC_45} and T. Takabayasi\cite{TT_57} have shown that Dirac equation can be cast into hydrodynamical form.
Here we use the methodology similar to that given in \cite{TT_57}(see also \cite{FA_11})
to describe the fluid equations for relativistic electrons in presence of the asymmetric neutrino
background. In the standard MHD-approximation electron contributes in defining the current whereas
the ion provides the inertia and therefore significantly contributes to the fluid velocity
[\cite{Boyd_69},\cite{Brodin_07}]. In this appendix we first derive electron fluid equations from
the Dirac equation and then carry-out the MHD approximations with the non-relativistic ion fluid
and obtain an expression for the relativistic corrections to the MHD current and finally
discuss the changes this brings about the our (non-relativistic) results on the Alfv\'en waves.
The subsequent derivation is rather lengthy and involved we would like to refer the readers to Ref.\cite{TT_57} for further details.
Following \cite{TT_57} we start with writing the bilinear covariants with hydrodynamic variables and establishing the relations among them using properties of the gamma matrices.
And also establishing their evolution equations from moments of the corresponding Dirac equation. We choose following bilinear covariants.
\begin{align}
\Omega=&\bar\psi\psi \\
\bar\Omega=&i\bar\psi\gamma^5\psi\\
S^\mu=&\bar\psi\gamma^\mu\psi \\
\bar S^\mu=&\bar\psi\gamma^5\gamma^\mu\psi\\
M^{\mu\nu}=&\bar\psi\sigma^{\mu\nu}\psi \\%
\bar M^{\mu\nu}=&i\bar\psi\gamma^5\sigma^{\mu\nu}\psi
\end{align}
Where, $\sigma^{\mu\nu}=(i/2)[\gamma^\mu, \gamma^\nu]$.
One can obtain the equations of motion for $\psi$ and $\bar\psi$ from Eqn.(\ref{lagrangian}) and
using these equations of motion one can write following two generic equations involving the dynamics of the above bilinear forms:
\begin{equation}
i\bigg(\bar\psi\gamma^A\gamma^\mu\partial_\mu\psi+\partial_\mu\bar\psi\gamma^\mu\gamma^A\psi\bigg)-eA_\mu\bar\psi[\gamma^A, \gamma^\mu]\psi-\frac{\Delta f_\mu}{2}\bar\psi[\gamma^A, \gamma^\mu\gamma^5]\psi=0
\label{D_Eq+}
\end{equation}
\begin{equation}
i\bigg(\bar\psi\gamma^A\gamma^\mu\partial_\mu\psi-\partial_\mu\bar\psi\gamma^\mu\gamma^A\psi\bigg)-eA_\mu\bar\psi\{\gamma^A, \gamma^\mu\}\psi-\frac{\Delta f_\mu}{2}\bar\psi\{\gamma^A, \gamma^\mu\gamma^5\}\psi-2m\bar\psi\gamma^A\psi=0
\label{D_Eq-}
\end{equation}
Using the definition for covariant differential operator $\delta^*_\mu(\bar\psi\gamma^A\psi)=i(\bar\psi\gamma^A\partial_\mu\psi-\partial_\mu\bar\psi\gamma^A\psi)-2eA_\mu\bar\psi\gamma^A\psi $ we define,
\begin{align}
j_\mu=&(1/2m)\delta^*_\mu\Omega \label{Def_jmu}\\
\bar j_\mu=&(1/2m)\delta^*_\mu\bar\Omega \\
T_\mu^\nu=&(1/2m)\delta^*_\mu S^\nu \\
\bar T_\mu^\nu=&(1/2m)\delta^*_\mu \bar S^\nu \\
N^{\mu\nu}_\alpha=&(1/2m)\delta^*_\alpha M^{\mu\nu}\\
\bar N^{\mu\nu}_\alpha=&(1/2m)\delta^*_\alpha \bar M^{\mu\nu}
\end{align}
The quantities $M^{\mu\nu}$ and $\bar M^{\mu\nu}$ can be expressed as $\rho^2 M^{\mu\nu}=-\bar\Omega( S^\mu\bar S^\nu-S^\nu\bar S^\mu)+\Omega\epsilon^{\mu\nu\kappa\lambda}S_\kappa\bar S_\lambda$ and $\bar M^{\mu\nu}=(1/2)\epsilon^{\mu\nu\kappa\lambda} M_{\kappa\lambda}$, where $\rho=\sqrt{\Omega^2+\bar\Omega^2}$ has the interpretation of density.
From Eqns.(\ref{D_Eq+}) and (\ref{D_Eq-}) we obtain the evolution equation of the above defined quantities.
\begin{align}
\partial_\mu S^\mu=&0 \label{evo_1}\\
\partial_\mu\bar S^\mu =&-2m\bar\Omega \label{evo_2}\\
(1/2m)\partial_\nu M^{\mu\nu} +j^\mu-S^\mu+&\frac{\Delta f_\nu}{2m}\bar M^{\mu\nu}=0 \label{evo_3}\\
(1/2m)\partial_\nu \bar M^{\mu\nu} +\bar j^\mu-&\frac{\Delta f_\nu}{2m} M^{\mu\nu}=0 \label{evo_3}
\end{align}
Next, one defines\cite{FA_11} four-velocity $v_\mu=S_\mu/\rho$ and four-spin $w_\mu={\bar S}_\mu/\rho$ in such a way
that it satisfies the following constraints: $v^\mu v_\mu=1$, $w^\mu w_\mu=-1$ and $v^\mu W_\mu=0$.
From the last constraint, it is clear that $w_0={\mathbf v}\cdot {\mathbf w}/u^0$ and thus in the
rest-frame zeroth component of the four spin $w_0=0$.
By taking the divergence of ${\bar T}^{\mu \nu}$ one obtains the following equation
\begin{equation}
\partial_\nu{\bar T}^{\mu\nu}=\frac{e}{m}w_\nu F^{\mu\nu}-{\bar j}^\mu_{st}+
\delta f_\nu\left[-\rho sin\theta\,\left(v^\mu w^\nu-v^\nu w^\mu\right)+
\rho cos\theta\,\epsilon^{\mu\nu\kappa\lambda}v_\kappa w_\lambda \right]
\label{spin1}
\end{equation}
where we have used $M^{\mu\nu}=\left[-\rho sin\theta\,\left(v^\mu w^\nu-v^\nu w^\mu\right)+
\rho cos\theta\,\epsilon^{\mu\nu\kappa\lambda}v_\kappa w_\lambda \right]$ following \cite{TT_57} and ${\bar j}^{\mu}_{st}$ has the same standard expression as given in Refs.[\cite{TT_57} or \cite{FA_11}]. Besides we have used
new definitions: $cos\theta=\Omega/\rho$ and $sin\theta={{\bar\Omega}/\rho}$. Here we note that $f^{\mu}$ term for the neutrino current does not appear in the above equation. Equation (\ref{spin1}) is at the single body particle-antiparticle state level and one is required to take the fluid average for a collection of $N$ such states. This $N$ particle spinor must be written as a $4^N\times 4^N $ Slatter determinant
of $N$ one-particle states this procedure had been developed in Ref.\cite{FA_11} and we follow it up for our calculation. We find following equation,
in thermal equilibrium, for the spatial part of the spin dynamic:
\begin{equation}
\gamma\left(\partial_t+{\mathbf v}\cdot\nabla\right) {\mathbf S}=\frac{e/m}{<cos\theta>}
\left(W^0{\mathbf E}/2+ {\mathbf S}\times{\mathbf B}\right)-
\gamma\Delta f^0{\bf S}\times {\mathbf v}+
\gamma\Delta f^0\frac{<sin\theta>}{<cos\theta>}{\mathbf S}
\label{rel_spin_dyn}
\end{equation}
where, $\gamma=\frac{1}{\sqrt{1-v^2}}$ and $W^0={\mathbf S}\cdot {\mathbf v}$.
Here we note that as we have assumed before we have dropped the spin-thermal coupling and the non-linear spin terms.
The last two terms on the right hand side gives an additional contribution to the spin dynamics of the electron-fluid dynamics given in Ref.\cite{FA_11}. This additional term solely depends on the neutrino background as it should be the case. Following electron relativistic hydrodynamical model in Ref.\cite{FA_11}, we we regard $\theta$ as a constant parameter which is zero for the non-relativistic quantum case and for an extreme relativistic quantum case $\theta=\pi/4$. For the non-relativistic spin dynamics (Eqn.(\ref{Hydro_eqn_S_q})) can be reproduced when we take $\theta=0$. When the electrons are at relativistic temperature one can
replace $m n$ by $(\epsilon+p)$ i.e. by enthalpy density \cite{LLB}.
Now one can define the total mass density $\rho=(\epsilon+p)+m_i n_i$ where, $(\epsilon+p)$ represents
the enthalpy density of the electrons. Since in the magnetohydrodynamic equations the inertia of
the fluid is dominated by ions, the momentum of the fluid is dominated by the ion momenta \cite{Boyd_69}
(see also \cite{Brodin_07}). This remains true for the relativistic electron case also provided $\rho\sim m_i n_i$ and Eqn.(\ref{mom_approx_fluid}) remains valid for us.
Next we derive the analogue of Eqn.(\ref{neu_ind_vor_no_em_w_para}) when the electrons are relativistic. For this consider that there
is an external magnetic field $B_0$ in $z-$direction and there is no streaming of fluid ${\mathbf v}_0=0$.
Also there is no electromagnetic perturbations i.e. $\delta {\mathbf E} $, $\delta {\mathbf B}$=0. The background spin vector is anti-parallel to the external magnetic field and given by
$S_0=-\mu_B B_0/2T$ as considered before. Next one can eliminate the electron velocity in the spin
equation ${\mathbf v}_e={\mathbf v}-m_i{\mathbf j}/(Ze\rho)$. Since there is no electromagnetic perturbation
for this case ${\mathbf j}=0$ and one can obtain using Eqs.(\ref{mom_approx_fluid},\ref{rel_spin_dyn}) one obtains the following dispersion
relation:
\begin{equation}
\omega=-\left( \frac{\mu_B n_e}{\sqrt{\rho_0}} \right)\left( \frac{\Delta f^0}{2T}<cos\theta>\right)k v_A
\label{omega_new_rel_alf}
\end{equation}
where, $v_A=\frac{B_0}{\sqrt{\rho_0}}$. Here we we note here that when
then the last term in Eqn.(\ref{rel_spin_dyn}) does not contribute to significantly to the dispersion relation.
Similarly for the electromagnetic perturbation for the standard Alfv\'en waves one obtains
the following dispersion relation:
\begin{equation}
\omega=-\frac{\tilde{v}_A}{\sqrt{\rho_0\alpha}}\frac{\mu_B n_e}{2T}\Delta f^0<cos\theta>k\pm k\tilde{v}_A
\label{alf_corre_neu+rel}
\end{equation}
Here we note that both the new Alfv\'en waves(Eqn.(\ref{omega_new_rel_alf})) and the regular Alfv\'en waves (Eqn.(\ref{alf_corre_neu+rel})), in the ideal MHD limit, gets
correction due to the relativistic effect which is characterized by
the $<cos\theta>$ factors. Now for an ultra relativistic limit
if one takes $\theta=\pi/4$ following Ref.\cite{FA_11}, one
gets $1/\sqrt{2}$ factor suppression in the speed of the new Alfv\'en wave compared to the non-relativistic case (with
$\theta=0$) case.
\end{appendix}
|
\section{Introduction}
In Paper I in this series,
we have presented the \textsc{Hi} observations for eight edge-on late-type galaxies.
One of our main conclusions was that the \textsc{Hi} showed clear signs of
self-absorption, which created the risk of a rather drastic underestimation
of the baryonic content of the galaxy.
We developed a new \textsc{Hi} modelling and fitting system called \textsc{Galactus}
in Paper II to address this problem.
We showed that the \textsc{Hi} self-absorption indeed has a drastic impact on the
observed maximum surface brightness temperature.
Self-absorption needs to be taken into account when modelling the \textsc{Hi} in
edge-on galaxies.
The main goal of our project throughout this series of papers is to
measure the hydrostatics at the central plane of edge-on galaxies.
To get to this, we will need to derive the rotation curve $v_\textrm{rot}(R)$,
face-on surface density $A_\textsc{Hi} (R)$, thickness of the HI layer $z_0(R)$ and
its velocity
dispersion $\sigma(R)$ accurately.
Secondary parameters, such as the exact kinematic position in RA, DEC and the
systemic velocity $v_\textrm{sys}$, will also need to be derived.
While warps are of scientific interest, they represent a disturbance of the
central plane of the galaxy.
The physics behind warps is not well understood \citep{kf11}.
Using the hydrostatics from the warped region would be a dangerous and
ultimately futile exercise.
We therefore refrain from modelling the warps.
The channel maps of the observations, with the model superimposed on it,
will be used to estimate the position at which the warp sets in.
Beyond this position, we will not calculate the hydrostatics.
Fitting the \textsc{Hi} structure and kinematics of edge-on galaxies by modelling the
position-velocity (XV) diagram has a long-standing record of accomplishment.
Various methods have been devised, initially only to derive the rotation curve
and the
radial distribution of HI surface brightness, and in some cases also the
flaring (the increasing
thickness as a function of galactocentric radius) of the \textsc{Hi} layer
\citep[e.g.][and references therein]{Sancisi1979A, vdk81c,
Garcia-Ruiz2002A, Takamiya2002A,
Uson2003A, kk04, Kregel2004B}.
More recently, \citet{Olling1996A, Olling1996B} and
\citet{OBrien2010A,OBrien2010B,OBrien2010C,OBrien2010D} have measured the
\textsc{Hi} velocity dispersion.
The paper by \citet{OBrien2010B} provides a detailed description of the
various methods and a discussion on the relative merits and pitfalls.
In this paper, we develop a new approach to modelling the \textsc{Hi} properties
of an edge-on galaxy, which consists of fitting channel maps near the terminal
velocity of the galaxy.
In Section \ref{sec:HIfittingstrategy}, we explain this fitting strategy in
more detail.
Section \ref{sec:HIfittingtestmodels} then applies this strategy on a series
of test models and demonstrates the validity of the method.
The kinematics and structure of six edge-on galaxies is subsequently presented
in Section \ref{sec:HIfittingresults}.
\section{Fitting Strategy for Edge-on Galaxies}\label{sec:HIfittingstrategy}
In Paper II, we demonstrated that the face-on galaxy NGC\,2403 has at least 10\%
more \textsc{Hi} mass than an optically thin model would have measured.
Rotating NGC\,2403 to an edge-on geometry and assuming
a spin temperature of 100\,K,
we found that 22\% of the total \textsc{Hi} mass would have been hidden by
self-absorption.
A lower median spin temperature of 80\,K increases this fraction to 30\%
(see Section 7 of Paper II).
Any attempt to model an edge-on galaxy as optically thin will thus suffer
from a drastic underestimation of total \textsc{Hi} mass.
While this is troubling enough by itself, other questions regarding the
effect of the self-absorption can also be raised:
\begin{itemize}
\item What is the effect of self-absorption on the rotation-curve estimate?
\item Is the thickness of the disc still measured properly?
\item What is the impact on the measured velocity dispersion?
\end{itemize}
A common practice for measuring the kinematics of edge-on galaxies is
through the outer envelope of the (integrated) XV diagram, either adopting
the peak flux as the rotation at the line of nodes, or by fitting a
half-Gaussian to the terminal velocities
\citep[e.g.][and references therein]{Sancisi1979A,Olling1996A,kk04, OBrien2010B, OBrien2010C}.
Both the velocity dispersion and thickness of the disc can
change the mid-plane surface brightness in a particular channel.
This creates a degeneracy between the two parameters.
As such, there is no practical way to correct for the
self-absorption using only the mid-plane or integrated XV-diagrams
of a galaxy.
We therefore opt for a different strategy and model
the channel-maps of the \textsc{Hi} data cube directly.
An advantage to this strategy is that we can fit all
four main parameters (i.e. the rotation curve, face-on
surface density, velocity dispersion and the thickness)
simultaneously and self-consistently.
After much trial-and-error, we have devised the following
three-pass strategy for fitting edge-on galaxies, using
the \textsc{Galactus} fitting code developed in Paper II.
In the first pass, we fix the velocity dispersion at a
constant value of 10 km/s and fit the rotation curve,
face-on surface density and thickness of the disc.
The central position ($x$, $y$, $v_\textrm{sys}$) along
with the position angle $PA$ are also fitted during this pass.
Tests have shown that measuring the exact inclination is
difficult, so we fix the inclination at $90^\circ$.
We fit the entire \textsc{Hi} data cube. The galaxy has already been projected
such that the major axis aligns with the horizontal axis as already discussed
in Paper I.
The results are subsequently inspected and corrections made where necessary.
We next fix the position and position angle.
The galaxy is split in two halves, which will be fit separately.
Based on the results from the initial fit, we create a mask, such that
the terminal velocity is still visible at all radii, but the radii
at lower rotational velocities are masked.
The line-of-nodes velocity is expected near the terminal velocity.
We do this because warps and asymmetries, together with the
self-absorption, lead to degeneracies in the solution and
thus can confuse the fitting algorithm.
The masking is done by hand. Results from subsequent fits
are inspected to ensure the line of nodes rotation is well
beyond the masked positions.
This is effectively creating an outer-envelope mask. Rather
than just mask the XV diagram, we are now masking the \textsc{Hi} data cube itself.
After both sides have their \textsc{Hi} data cubes masked, we fit the
rotation curve, thickness, face-on surface density and velocity dispersion.
The results are again inspected, corrected and refit where necessary.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Petersetal-IIIfig1.jpg}
\caption[Moment-zero maps of warps]{Moment-zero maps of a model without warp (top), face-on warp (middle) and a side-on warp (bottom).}
\label{fig:maskingtheenvelope1}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{Petersetal-IIIfig2a.jpg}
\includegraphics[width=0.45\textwidth]{Petersetal-IIIfig2b.jpg}
\caption[Integrated XV diagram and the effects due to warps]{Left: Integrated XV-diagrams for a model without a warp (top), a line-of-sight warp (middle) and a side-on warp (bottom). Right: Integrated XV diagram of the difference \textsc{Hi} data cube between the non-warped and face-on warp model (top), and the non-warped and side-on warp model (middle). The blue contours show the outline of the full XV diagram. The bottom panel reproduces the middle panel, but has the envelope mask superimposed on top of it. Note that the fitting occurs on the channel maps rather than the XV-diagram.}
\label{fig:maskingtheenvelope2}
\end{figure*}
To illustrate the need for an outer-envelope mask, we use
\textsc{Galactus} to model various types of warps.
We begin with the parameters of the model from Section
7 of Paper II. This model is based on the \textsc{Hi} in NGC\,2403, as
determined by us using a fitting of the data taking self-absorption
into account. The models will be simulated
as perfectly edge-on.
We generate three models: The first will have no warp, while
the second will have a strong line-of-sight warp (the
maximum deviation from the mid-plane occurs along the line
of sight) and the third one a strong side-on warp (the maximum
deviation from the mid-plane occurs
perpendicular to the line of sight).
The models are run in self-absorption mode at a spin
temperature of 100\,K.
Both warps begin at a radius of 13.6\,kpc from the centre of the galaxy.
The offset of the warp above the plane the increases
linearly with one kpc height for every one kpc radius.
It peaks at a radius of 20.4\,kpc at a height of 6.8\,kpc.
In Figure \ref{fig:maskingtheenvelope1}, we demonstrate
the zeroth-moment maps for the three models.
The presence of a line-of-sight warp is very hard to
detect from this image, as is typical for these types
of warps \citep[e.g.][and references therein]{Gentile2003A}.
In contrast, the side-on warp is clearly visible.
Shown in Figure \ref{fig:maskingtheenvelope2} are
the height-integrated XV diagrams of the three models.
While some variation is present between the three
images, the differences are minor.
To illustrate the effect of the warp, we have created
two difference XV-maps in Figure \ref{fig:maskingtheenvelope2}.
Here we have subtracted the non-warped \textsc{Hi} data cube from both warped
\textsc{Hi} data cubes and
integrated the absolute values of these differences along the minor axis.
Both warps create a different signature in these difference maps.
The line-of-sight warp creates a bar spread evenly over all lower
velocity channels, while the side-on warp is far more pronounced at larger
positions of $x$.
Unsurprisingly, beyond $x>13.6$\,kpc the warps affect the
terminal-velocity channels as well and will thus affect the parameters
extracted from this region.
However, as demonstrated by the superimposed mask in the last panel,
the outer envelope of the \textsc{Hi} data cube at radii less than
13.6 kpc is not affected by the warp.
Using an envelope mask of the \textsc{Hi} data cube, we can thus measure the parameters
of an edge-on galaxy inside the warped region, regardless of the
presence of either a line-of-sight or side-on warp.
Had we not masked the lower velocities, then the effect of the
warps would have `confused' the fitting algorithm.
In the final pass, the results are sampled with the Monte-Carlo Markov-Chain
(MCMC) code \textsc{emcee}, which is used to sample the likelihood
distribution in each parameter (see Section 2.8 of Paper II
for more details).
The results, auto-correlation and traces are inspected after a sufficient
number of samples (i.e. 100.000+) have been drawn.
When the MCMC has settled into a stable distribution, the final 10.000
samples are used to calculate the parameter distributions.
We visualize these parameter distribution in subsequent figures based
on the central 68.3, 95.5 and 99.7\% fractions.
We fit each galaxy in two ways: using both an optically thin model and
a self-absorption model with a spin temperature of 100\,K.
As the physics of warps is not well understood\footnote{see \citet{kf11}
and references therein}, we do not concern ourselves with warps.
We fit the entire galaxy and determine by eye at what radius the warp
starts to affect the flaring measurement.
Beyond that radius, the data is considered unreliable.
\section{Testing the Strategy}\label{sec:HIfittingtestmodels}
\begin{figure*}
\centering
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig3a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig3b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig3c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig3d.jpg}
\caption[Optically thin test model fit (0.1\,K)]{Results from a fit to both sides of an optically thin test model with low noize (0.1\,K). Top-left panel shows the rotation curve $v_\textrm{rot}$, top-right panel the face-on surface density $A_\textsc{Hi} $. The lower-left panel shows the FWHM thickness of the disc. The velocity dispersion $\sigma$ is shown in the lower-right panel. Green colours denote the left side of the galaxy, blue colours show the right side. The grey band is the combined result. Dotted lines show the true parameters of the model.}\label{fig:testmodel00}
\end{figure*}
\subsection{High S/N optically thin fit}
We first test our strategy on a series of models.
We begin with an optically thin model, with a very low noise at
$\sigma=0.1$\,K. The model resembles UGC\,7321, except that the
linear scale is twice as small.
The data is modelled at a distance of 10\,Mpc, with pixels of
6.8'', a channel width of 3.3\,km/s and a FWHM beam of 13.8''.
At only 48\,K, the maximum surface brightness in the model is
still low, but the maximum signal to noise is very high at 480.
We present the results of the fit to this model in Figure
\ref{fig:testmodel00}.
As is clear from the image, the fit matches the initial near
perfectly. Only at very large radii, where the \textsc{Hi} surface density is
lower than about $10^{20}$ atoms/cm$^{2}$, do we find a sudden increase in FWHM
and decrease in $\sigma$ that are not present in the model. For such levels,
where nearly no \textsc{Hi} is present the fits become unreliable.
We do find such behaviour in noisier models and in actual fits on our data
below, usually also at levels below $10^{20}$ atoms/cm$^{2}$. Obviously
such results should not and will not be used in any further modelling.
\subsection{Optically Thin Fits}
So how does the fitting deteriorate when the noise becomes higher?
We double the size of the model from the previous test, such that the
radial distance of each value of the parameters is twice as far out.
The cell size is also doubled to 13.8''. The model now has the same
linear size as UGC\,7321, on which it is based.
This has the effect of raising the maximum surface brightness
temperature to a more realistic value $\sim90$\,K.
The models are run with noise levels of 1, 3 and 5\,K.
The maximum signal to noise then becomes 93, 30 and 20.
The parameters found in the fits on either sides are smoothed.
To create the combined parameters, the unsmoothed fits to either side are
averaged together and the combined results smoothed.
For the rotation curve, flaring and face-on surface density with a
kernel of [1/4, 1/2, 1/2] and the velocity dispersion with a kernel of
[1/3, 1/3, 1/3], as these parameters are found to be more sensitive to noise.
For the combined result, both sides of the galaxy are averaged together
from the unsmoothed fits, and only then is the combined data smoothed.
We show the results for this fit in Figure \ref{fig:testmodels-thin}.
The rotation curve and the face-on surface density are recovered well
in all cases.
Measuring the thickness of the disc works reasonably well for the 1
and 3\,K models -- at least out to 10-12 kpc --,
but the 5\,K result starts to deviate more.
The velocity dispersion is the hardest parameter to fit.
Only for the 1\,K model is the input model recovered sufficiently
well out to 10 kpc, the 3\,K result is marginal and the 5\,K result is
doubtful even at 8 kpc.
\begin{figure*}
\centering
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig4a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig4b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig4c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig4d.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig4e.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig4f.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig4g.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig4h.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig4i.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig4j.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig4k.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig4l.jpg}
\caption[Optically thin test model fits (1\,K, 3\,K and 5\,K)]{Optically thin test models with noises of 1\,K, 3\,K and 5\,K (left-to-right). Top panels shows the rotation curve $v_\textrm{rot}$, second row shows the face-on surface density $A_\textsc{Hi} $. Third row shows the thickness of the model. Fourth row shows the fit to the velocity dispersion. Green colours denote the left side of the galaxy, blue colours show the right side. The grey band is the combined result. Dashed lines show the true parameters of the model.}\label{fig:testmodels-thin}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig5a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig5b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig5c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig5d.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig5e.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig5f.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig5g.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig5h.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig5i.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig5j.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig5k.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig5l.jpg}
\caption[Self-absorption test model fits (1\,K, 3\,K and 5\,K)]{Self-absorption ($T_\textrm{spin}=100$\,K) test models with noises of 1\,K, 3\,K and 5\,K (left-to-right). Top panels shows the rotation curve $v_\textrm{rot}$, second row shows the face-on surface density $A_\textsc{Hi} $. Third row shows the thickness of the model. Fourth row shows the fit to the velocity dispersion. Green colours denote the left side of the galaxy, blue colours show the right side. The grey band is the combined result. Dashed lines show the true parameters of the model.}\label{fig:testmodels-thick}
\end{figure*}
\subsection{Self-absorption Fits}
We perform a similar test using these models, but include
self-absorption at a median spin temperature of 100\,K.
The face-on surface density is increased to compensate for
the self-absorption and now peaks around $1.5\times10^{20}$\,atoms/cm$^2$.
We show the results in Figure \ref{fig:testmodels-thick}.
We find that the rotation curve is recovered well in all
three models.
Both the measurements of face-on surface density and those of
the velocity dispersion become
increasingly difficult at lower signal-to-noise ratio.
The velocity dispersion is only recovered well in the 1\,K model.
\begin{figure*}
\centering
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig6a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig6b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig6c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig6d.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig6e.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig6f.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig6g.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig6h.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig6i.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig6j.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig6k.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig6l.jpg}
\caption[Optically thin fits to self-absorption models]{Optically thin fits to self-absorption ($T_\textrm{spin}=100$\,K) test models with noises of 1\,K, 3\,K and 5\,K (left-to-right). Top panels shows the rotation curve $v_\textrm{rot}$, second row shows the face-on surface density $A_\textsc{Hi} $. Third row shows the thickness of the model. Fourth row shows the fit to the velocity dispersion. Green colours denote the left side of the galaxy, blue colours show the right side. The grey band is the combined result. Dashed lines show the true parameters of the model.}\label{fig:HIwhenmodelsgobad}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig7a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig7b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig7c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig7d.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig7e.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig7f.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig7g.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig7h.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig7i.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig7j.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig7k.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig7l.jpg}
\caption[Optically thin test model fits at $i=88.8^\circ$ (1\,K, 3\,K and 5\,K)]{Optically thin test models with noises of 1\,K, 3\,K and 5\,K (left-to-right). The reference images have an inclination of $88.8^\circ$. Top panels shows the rotation curve $v_\textrm{rot}$, second row shows the face-on surface density $A_\textsc{Hi} $. Third row shows the thickness of the model. Fourth row shows the fit to the velocity dispersion. Green colours denote the left side of the galaxy, blue colours show the right side. The grey band is the combined result. Dashed lines show the true parameters of the model.}\label{fig:testmodels-wronginclination}
\end{figure*}
\subsection{A Self-absorption Model Fit as Optically Thin}
If in reality the \textsc{Hi} is self-absorbing, how wrong will models be
that assume an
optically-thin medium?
To test this, we fit the self-absorption models from the previous test with
optically-thin models.
The results are shown in Figure
\ref{fig:HIwhenmodelsgobad}.\label{sec:whenmodelgobad}
Again the rotation curve is recovered well out to 12 kpc or so.
Clearly, the rotation curve does not suffer much from self-absorption.
As expected, the face-on surface density is not
recovered well at all, with the highest column densities estimated at
only $\sim70$\% of their true values.
The disc is consistently estimated to be somewhat
thicker than it in reality is.
The velocity dispersion is not recovered well at all.
Performing optically thin fits on self-absorbed distributions produces
results significantly in error, even at relatively small radii.
\subsection{A More Complicated Model}\label{sec:notedgeon}
In the previous tests we have modelled and fitted the galaxies at
a perfect edge-on inclination of $90^\circ$. \label{sec:whenHIgoesbad}
Our fitting strategy always assumes this perfectly edge-on inclination.
As a final test, we examine the quality of the fit if the actual
galaxy is not perfectly edge-on, but is at an inclination of $88.8^\circ$.
We also include a modest side-on warp that begins at $R=8$\,kpc.
We again run the fits with intrinsic noises of 1, 3 and 5\,K
and use optically thin models.
The results are shown in Figure \ref{fig:testmodels-wronginclination}.
The true systemic velocity is recovered to within 0.6\,km/s,
while the channel width was 3.3\,km/s.
The position angle is accurate to within $0.9^\circ$.
The central position is found to within 0.7''.
The measurement of the thickness of the disc is wrong beyond
a radius of 9\,kpc, which is due to the onset of the warp.
Our conclusions remain the same as the previous tests: All
parameters are recovered well at 1\,K, but the measurements of the
velocity dispersion become increasingly difficult at 3\,K and for
all parameters at 5\,K.
\subsection{Concluding Remarks}
So what can we conclude?
We have shown that our fitting strategy can recover the parameters
in these edge-on galaxies.
The rotation curve and the face-on surface density are recovered
well in all cases.
The biggest problem in fitting these galaxies comes from the velocity
dispersion, which can only be fit in very low noise \textsc{Hi} data cubes.
We will thus adopt a constant velocity dispersion for galaxies where
the noise is too high.
The flaring will only be problematic in high noise \textsc{Hi} data cubes.
We note that our strategy produces these results in idealized circumstances.
In reality, the galaxies can (and will) have asymmetries, spiral arms,
warps, lagging haloes, etc.; which will make our measurements less accurate
then the presented test models.
We have also shown that fitting a self-absorption reference model with an
optically thin model will lead to bad results.
In particular, the face-on surface density will be too low, while the
thickness measured will be too high and the velocity dispersion will be wrong.
Only the rotation curve will be estimated to a reasonable degree.
As we argued before, galaxies are in reality expected to have self-absorbing
\textsc{Hi} .
This result thus implies that the results found in the
literature based on the assumption
of zero optical thickness are underestimating the true \textsc{Hi} content of
galaxies.
\section{Fits to actual galaxies}\label{sec:HIfittingresults}
In Paper I, we have presented the \textsc{Hi} observations
for eight edge-on galaxies.
In this section, we will focus on modelling these galaxies in more detail.
We aim to fit the rotation curve, velocity dispersion, thickness
and face-on surface density of the galaxies, as we will use these
in Paper V to model the hydrostatics of the gas.
Based on the results of the previous section, it is of vital
importance to have the lowest noise levels we possibly can.
To this end, we re-image the data such that we can still resolve
the vertical structure of the disc, but have the lowest noise possible.
For more details on the \textsc{Hi} reduction, see Section 4.1 of Paper I.
In Table 3 of Paper I, we denote the original \textsc{Hi} data cubes.
IC\,5052 has been re-imaged with a FWHM beam of 25'' and a noise
$\sigma$ of 1.7\,K.
ESO\,115-G021 now has a FWHM beam of 25'' and a noise $\sigma$ of 1.27\,K.
For ESO\,274-G001, we adopt the 30'' beam from Table
3 of Paper I, which has a noise $\sigma$ of 1.2\,K.
UGC\,7321 has been re-imaged with a FWHM beam of 25'' and a
noise $\sigma$ of 1.7\,K.
Based on these noise levels and the results of the previous section,
we have decided to model UGC\,7321, ESO\,115-G021 and ESO\,274-G001
with the velocity dispersion as a free parameter that can vary with radius.
The other galaxies will be modelled using an assumed constant velocity
dispersion of 10\,km/s.
We do not include the results for IC\,2531 and ESO\,146-G014.
In Section 5.1 of Paper I, we commented on the remarkable thin
disc that IC\,2531 appeared to have.
This disc, combined with the high noise, made it impossible to
measure the thickness of the \textsc{Hi} disc accurately.
We were unable to make a reliable fit to ESO\,146-G014.
As we discussed in Section 5.6 of Paper I, the galaxy probably
has a warp and is not sufficiently edge-on for our measurements.
\subsection{IC\,5052}
Galaxy IC\,5052 has proven hard to model.
We show both the optically thin model and the self-absorption model in
Figure \ref{fig:IC5052-velocitydispersion-thin}.
As we noted in Section 5.2 of Paper I, the galaxy has a very strong warp.
Looking back at Figure 3 in Paper I, we can see that
the onset of the warp occurs just beyond that same radius (180'').
This warp is responsible for the very large thickness that is found
beyond 5\,kpc.
The declining rotation curve beyond 5\,kpc is also due to this component.
The self-absorption models have a total mass of
$9.5\pm0.9\times10^8$\,M$_\odot$, while the optically thin model recovers
only $7.4\pm0.7\times10^8$\,M$_\odot$.
This is less than our initial estimate of $8.9\times10^8$\,M$_\odot$
in Table 4 of Paper I.
This trend is also visible in the other galaxies. It is because the
fit can only model a smooth medium. Local bright regions in the
observations therefore cannot be accurately recovered.
Any warp will also harbor mass, which is not recovered here.
Comparing the two models (and ignoring the right-hand self-absorption
model), the optically thin model is missing about a quarter of the \textsc{Hi} .
\citet{OBrien2010C} also had trouble with modelling IC\,5052.
They did not resolve the warp, but found clear evidence for asymmetry.
Comparing the rotation curves to our work, the results on both sides
look similar up to 6\,kpc, but subsequent downturn is not detected as
strongly in their work. The thickness estimates are similar.
Their face-on column density is beyond 2\,kpc similar to our optically
thin result, although due to our smoothing the data is less bumpy.
The inner parts are different, but are not well resolved in our models.
The results we find for the flaring beyond a radius of 5 kpc are doubtful.
We have chosen in Paper V in this series where we analyze our data further
not to consider this galaxy.
\begin{figure*}
\centering
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig8a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig8b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig8c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig8d.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig8e.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig8f.jpg}
\caption[\textsc{Hi} decomposition of IC\,5052]{Decomposition results for IC\,5052 assuming an optically thin mode (left column) and self-absorbing \textsc{Hi} medium (right column). Colour ranges have been chosen such that they contain 68\%, 95\% and 99.7\% of the distribution. Assuming a normal distribution, this represents the 1, 2 and 3$\sigma$ dispersion from the mean. Green lines are for the left side of the galaxy, cyan lines for the right side.}\label{fig:IC5052-velocitydispersion-thin}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig9a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig9b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig9c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig9d.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig9e.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig9f.jpg}
\caption[\textsc{Hi} decomposition of IC\,5249 (optically thin)]{Decomposition results for IC\,5249 assuming an optically thin mode (left column) and self-absorbing \textsc{Hi} medium (right column). Colour ranges have been chosen such that they contain 68\%, 95\% and 99.7\% of the distribution. Assuming a normal distribution, this represents the 1, 2 and 3$\sigma$ dispersion from the mean. Green lines are for the left side of the galaxy, cyan lines for the right side. The grey band is the combined result.} \label{fig:IC5249-velocitydispersion-thin}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig10a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig10b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig10c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig10d.jpg}
\caption[\textsc{Hi} decomposition of ESO\,115-G001 (optically thin)]{Decomposition results for ESO\,115-G021 assuming an optically thin model. Colour ranges have been chosen such that they contain 68\%, 95\% and 99.7\% of the distribution. Assuming a normal distribution, this represents the 1, 2 and 3$\sigma$ dispersion from the mean. Green lines are for the left side of the galaxy, cyan lines for the right side. The grey band is the combined result.}
\label{fig:ESO115-G021-velocitydispersion-thin}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig11a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig11b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig11c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig11d.jpg}
\caption[\textsc{Hi} decomposition of ESO\,115-G021 (self-absorbing)]{Decomposition results for ESO\,115-G021 assuming a self-absorption model. Colour ranges have been chosen such that they contain 68\%, 95\% and 99.7\% of the distribution. Assuming a normal distribution, this represents the 1, 2 and 3$\sigma$ dispersion from the mean. Green lines are for the left side of the galaxy, cyan lines for the right side. The grey band is the combined result.}
\label{fig:ESO115-G021-velocitydispersion-thick}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig12a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig12b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig12c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig12d.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig12e.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig12f.jpg}
\caption[\textsc{Hi} decomposition of ESO\,138-G014 (optically thin)]{Decomposition results for ESO\,138-G014 assuming an optically thin mode (left column) and self-absorbing \textsc{Hi} medium (right column). Colour ranges have been chosen such that they contain 68\%, 95\% and 99.7\% of the distribution. Assuming a normal distribution, this represents the 1, 2 and 3$\sigma$ dispersion from the mean. Green lines are for the left side of the galaxy, cyan lines for the right side. The grey band is the combined result.}\label{fig:ESO138-G014-velocitydispersion-thin}
\end{figure*}
\subsection{IC\,5249}
The \textsc{Hi} structure and kinematics of galaxy IC\,5249 have been previously
analyzed by \citet{Abe1999}, \citet{vanderKruit2001A} and \citet{OBrien2010C}.
\citet{Abe1999} report a linearly rising rotation curve to about 100\,km/s
at 17\,kpc.
The data was re-analyzed by \citet{vanderKruit2001A}, who reports a steeper
inner rotation curve that flattens out to 105 km/s at 10\,kpc and remains
flat throughout the remaining disc.
This rotation curve was confirmed by \citet{OBrien2010C}.
Our fits with both the optically thin model and the self-absorption model
confirms their results (Figure \ref{fig:IC5249-velocitydispersion-thin}).
We however note that rotation curves differs between both sides, so the
galaxy might be asymmetric.
Comparing the face-on surface densities, our optically thin fit looks very
similar to the observed face-on density by \citet{vanderKruit2001A}. The
peak surface density occurs at a radius of 17-18\,kpc at a value of
$5.8\times10^{20}$ atoms/cm$^2$. More inward, the profile has a lower
density. We also reproduce their dip in the face-on profile at $\sim11$\,kpc.
The face-on surface density of \citet{OBrien2010C} looks similar to
our left-side profile.
Their joined profile does recover the peak at 17-18\,kpc, but does
not have the lower central densities.
Our self-absorption fit finds more \textsc{Hi} . The peak at 17-18\,kpc
now becomes a plateau near at $8\times10^{20}$\,atoms/cm$^2$ (30\% higher).
The profile also shows the dip near 11\,kpc, although the
average inner density remains far larger than the optically thin result.
The inner parts of the optically thin model are `saturated'
by the \textsc{Hi} from outer radii, and the model compensates for
this by lowering the face-on column densities in the inner parts.
We find a strongly flaring disc in both fits, with an inner
thickness of only 500\,pc, but increasing linearly with radius
to nearly 2.6\,kpc at 25\,kpc. In Paper V we will only use the flaring
out to 14 kpc for the fits to determine the axis ratio of the dark matter
halo.
This is different than \citet{OBrien2010C}, who finds the same
linearly increasing thickness, but only going out to 1.5\,kpc
thickness at a radius of 20\,kpc.
An average thickness of $1.1\pm0.3$\,kpc was reported by
\citet{vanderKruit2001A}, although their Figure 6 can be
seen to increase linearly to a thickness of 3.3\,kpc at 14\,kpc.
Given their error bars, this result is consistent with our findings.
Comparing the total masses of both types of fits, we find
that the optically thin model has a total \textsc{Hi} mass of
$4.8\pm0.2\times10^9$\,M$_\odot$.
The self-absorption model finds a total of $7.8\pm0.8\times10^9$\,M$_\odot$.
This would put the hidden \textsc{Hi} mass fraction at about a third.
The optically thin mass is however significantly lower than our
initial estimate of $5.6 \times 10^{9}$\,M$_\odot$
(Table 4 of Paper I).
\subsection{ESO\,115-G021}
Galaxy ESO\,115-G021 is a symmetric, slow rotating
galaxy which we have managed to fit well (see Figures
\ref{fig:ESO115-G021-velocitydispersion-thin}
and \ref{fig:ESO115-G021-velocitydispersion-thick}).
The rotation curve rises from an initial 30\,km/s
in the inner parts almost linearly to 65\,km/s around 8\,km/s.
The only other analysis of the rotation curve of
this galaxy is by \citet{OBrien2010C}, who report a very similar rotation curve.
The face-on surface density of our optically thin
model is also similar to theirs: A plateau between
$3-5\times10^{20}$\,atoms/cm$^2$ and then rapidly dropping off.
The self-absorption fit to the surface density yields
a similar profile, although the inner part has a bit more \textsc{Hi} .
The thickness of this galaxy is remarkably well behaved
and consistent between both fits.
Starting near 700\,pc in the inner parts, it rises
linearly to 1.5\,kpc at a radius of 8\,kpc.
This is different to \citet{OBrien2010C}, who report
a stronger increase to over 3\,kpc beyond a radius
of 6\,kpc, although they note they had insufficient available
positions in their data for a proper fit.
The velocity dispersion fit to the optically thin model is
surprising; the galaxy appears to have an increasing velocity
dispersion, which is 7\,km/s in the inner parts and 10\,km/s at 6\,kpc.
The self-absorption velocity dispersion is more stable, although
it is still varying a bit.
This behavior is likely due to noise in the data.
Comparing the total masses, the optically thin model yields
$5.6\pm0.2\times10^9$\,M$_\odot$ and $7.2\pm0.1\times10^9$\,M$_\odot$
for the self-absorption model.
Thus about a quarter of the \textsc{Hi} is hidden in the model.
\subsection{ESO\,138-G014}\label{sec:HIdecom138}
Galaxy ESO\,138-G014 has been well resolved.
The optically thin model and the self-absorption
model are shown in Figure \ref{fig:ESO138-G014-velocitydispersion-thin}.
The rotation curve rises steeply to 70\,km/s in the
inner parts, and gradually increase to a maximum of 105\,km/s near 8\,kpc.
Beyond this radius, it levels off.
The curve is very similar to that measured by \citet{OBrien2010C}.
The disc is reasonably thick, starting at 1.1\,kpc
near 2.5\,kpc radius and increasing linearly towards
1.8\,kpc at 9\,kpc (for the optically thin model).
In the self-absorption model, the disc is less thick,
starting around 1\,kpc in the inner parts and increasing
only to 1.5\,kpc at 9\,kpc.
In both models, around 12-13\,kpc the thickness suddenly
increases drastically, almost certainly due to the warp
seen in the total \textsc{Hi} map (Fig. 7 in Paper I).
Note that the increase towards a thickness of 2\,kpc in
the centre is unreliable.
\citet{OBrien2010C} do not trust their thickness
measurements and so we cannot compare their results with ours.
In Paper V we will only use the flaring
out to 6 kpc for the fits to determine the axis ratio of the dark matter
halo.
In the optically thin model, the face-on surface
density shows a clear plateau between 5 and $6\times10^{20}$\,atoms/cm$^2$
up to a radius of 12-13\,kpc, and dropping off rapidly beyond that.
The self-absorption model results in more \textsc{Hi} , with the plateau at a
higher $6-9\times10^{20}$\,atoms/cm$^2$.
The (optically thin) model of \citet{OBrien2010C} also finds a plateau
near $6\times10^{20}$\,atoms/cm$^2$, although it is higher near the inner parts.
Comparing the two mass estimates, our optically thin model has a total
mass of $3.4\pm0.1 \times 10^9$\,M$_\odot$, while our self-absorption model
finds $4.6\pm 0.2 \times 10^9$\,M$_\odot$.
The total hidden \textsc{Hi} fraction is thus $28\pm1$\% of the total \textsc{Hi} .
\begin{figure*}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig13a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig13b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig13c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig13d.jpg}
\caption[\textsc{Hi} decomposition of ESO\,274-G001 (optically thin)]{Decomposition results for ESO\,274-G001 assuming an optically thin model. Colour ranges have been chosen such that they contain 68\%, 95\% and 99.7\% of the distribution. Assuming a normal distribution, this represents the 1, 2 and 3$\sigma$ dispersion from the mean. Green lines are for the left side of the galaxy, cyan lines for the right side. The grey band is the combined result.}
\label{fig:ESO274-G001-velocitydispersion-thin}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig14a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig14b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig14c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig14d.jpg}
\caption[\textsc{Hi} decomposition of ESO\,274-G001 (self-absorbing)]{Decomposition results for ESO\,274-G001 assuming a self-absorption model. Colour ranges have been chosen such that they contain 68\%, 95\% and 99.7\% of the distribution. Assuming a normal distribution, this represents the 1, 2 and 3$\sigma$ dispersion from the mean. Green lines are for the left side of the galaxy, cyan lines for the right side. The grey band is the combined result.}
\label{fig:ESO274-G001-velocitydispersion-thick}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig15a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig15b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig15c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig15d.jpg}
\caption[\textsc{Hi} decomposition of UGC\,7321 (optically thin)]{Decomposition results for UGC\,7321 assuming an optically thin model. Colour ranges have been chosen such that they contain 68\%, 95\% and 99.7\% of the distribution. Assuming a normal distribution, this represents the 1, 2 and 3$\sigma$ dispersion from the mean. Green lines are for the left side of the galaxy, cyan lines for the right side. The grey band is the combined result.}
\label{fig:UGC7321-velocitydispersion-thin}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig16a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig16b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig16c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig16d.jpg}
\caption[\textsc{Hi} decomposition of UGC\,7321 (self-absorbing)]{Decomposition results for UGC\,7321 assuming a self-absorption model. Colour ranges have been chosen such that they contain 68\%, 95\% and 99.7\% of the distribution. Assuming a normal distribution, this represents the 1, 2 and 3$\sigma$ dispersion from the mean. Green lines are for the left side of the galaxy, cyan lines for the right side. The grey band is the combined result.}
\label{fig:UGC7321-velocitydispersion-thick}
\end{figure*}
\subsection{ESO\,274-G001}\label{sec:HIdecomposition_ESO274}
As we discussed in Section 5.7 in Paper I, galaxy ESO\,274-G001
has a strong continuum source at its central position.
This causes strong \textsc{Hi} absorption, leading to our inability to fit the
central radii of the galaxy accurately.
We trust our results beyond 1.5\,kpc.
We show our optically thin and self-absorption fits in Figures
\ref{fig:ESO274-G001-velocitydispersion-thin} and
\ref{fig:ESO274-G001-velocitydispersion-thick}.
The rotation curve of the galaxy starts near 30\,km/s in the inner
parts and rises continuously towards a maximum of 80\,km/s at 5\,kpc.
A similar result was found by \citet{OBrien2010C}.
The face-on surface density profile of the optically thin model
hovers between 3 and $5\times10^{20}$ atoms/cm$^2$, with an average
around $3.7\times10^{20}$ atoms/cm$^2$.
The face-on surface density profile of the self-absorption model
has more \textsc{Hi} , in a plateau around $5\times10^{20}$ atoms/cm$^2$ up to 5\,kpc.
In both models, the \textsc{Hi} density drops rapidly towards zero at 8\,kpc.
The (optically thin) face-on surface density model of \citet{OBrien2010C}
yielded a similar result as our optically thin model.
With the exception of the inner part, the thickness of
the optically thin galaxy starts at 600\,pc at 2\,kpc radius,
and increases to 750\,pc at a radius of 6\,kpc.
Similar to previous galaxies, the self-absorption yields a less
thick disc, starting at 550\,pc at 2\,kpc, but only increasing to
600\,pc at 6\,kpc radius.
Beyond 6\,kpc, both models yield a very large thickness, although
the \textsc{Hi} has dropped to almost zero at those radii.
The flaring is very different from that measured by \citet{OBrien2010C},
whom found 100\,pc near the inner parts and flaring out linearly towards
1.8\,kpc in the outer parts.
Both our models find a velocity dispersion that is dropping with radius.
The optically thin model starts near 13\,km/s and drops towards 7\,km/s
at 6\,kpc.
The self-absorption model drops towards 6\,km/s at that radius.
Beyond this radius, the velocity dispersion is unreliable.
The velocity dispersion measured by \citet{OBrien2010C} shows a very
different behavior, staying constant around 6.5\,km/s up to 6\,kpc.
The optically thin model has a total \textsc{Hi} mass of
$3.1\pm0.02\times10^8$\,M$_\odot$.
The self-absorption model has a total mass of $4.1\pm0.1\times10^8$\,M$_\odot$.
A total of $25\pm1$\% of the \textsc{Hi} is thus hidden by self-absorption.
\subsection{UGC\,7321}
The \textsc{Hi} of galaxy UGC\,7321 has been previously modelled by
\citet{Uson2003A} and \citet{OBrien2010A}.
\citet{Matthews2003A} investigated the high-latitude \textsc{Hi} of
the galaxy is detail.
The galaxy is the only in our sample which has been observed
using the VLA.
Together with ESO\,274-G001, the galaxy has the highest signal
to noise in our sample.
We show the results for the fits in Figures
\ref{fig:UGC7321-velocitydispersion-thin} and
\ref{fig:UGC7321-velocitydispersion-thick}.
The rotation curve rises steeply in the inner
parts, only to level of around 9\,kpc at 110\,km/s.
A similar behavior was reported by \citet{Uson2003A} and \citet{OBrien2010A}.
The thickness of the disc starts, in the optically
thin model, near 750\,pc and increases linearly
towards 1.5\,kpc at a radius of 10\,kpc.
Beyond this radius, the thickness increases rapidly,
due to the presence of a warp.
\citet{Uson2003A} estimate the onset of the warp at
150'' (7.2\,kpc), although it becomes stronger beyond this radius.
Our self-absorption fit shows a different profile than
the optically thin model.
Between zero and five kpc, the thickness is nearly constant
at 600-700\,pc, and only then flares out to 1.5\,kpc at 10\,kpc radius.
In Paper V we will only use the flaring
out to 7.5 kpc for the fits to determine the axis ratio of the dark matter
halo.
Both the optically thin model and the self-absorption
model find a plateau in the face-on surface density profile.
This plateau remains constant out to roughly 7\,kpc,
between 4 and $5\times10^{20}$\,atoms/cm$^2$ in the optically
thin model, and 6 and $7\times10^{20}$\,atoms/cm$^2$ in the
self-absorption model.
The optically thin results are consistent with both \citet{Uson2003A} and
\citet{OBrien2010A}.
The high quality of the data has also allowed us to fit the velocity
dispersion.
The optically thin model hovers between 9 and 10\,km/s up to 10\,kpc radius.
The self-absorption model yields a decrease from 10 to 8\,km/s in the
first 3\,kpc, but remains flat at 8\,km/s up to 10\,kpc radius.
The optically thin model has a total mass of $8.9\pm0.4\times10^8$\,M$_\odot$,
while the self-absorption model has a total mass of
$1.17\pm0.02\times10^9$\,M$_\odot$.
A total of $24\pm1$\% of the \textsc{Hi} is thus hidden by self-absorption.
\begin{figure*}[t]
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig17a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIIfig17b.jpg}
\caption[Rotation curve and velocity dispersion with height]{Fits to the rotation curve and velocity dispersion as function of height. Measurement in ESO\,274-G001. The grey curve is the fit to the total galaxy, the cyan is the fit to the central plane of the galaxy and the green curve is the fit to the data that excludes the central plane.}\label{fig:sigmatest}
\end{figure*}
\subsection{Variations of the Velocity Dispersion with Height}\label{sec:velocitydispersionwithheight}
One of the (many) assumptions required for modelling the
hydrostatics in a galaxy, is that the \textsc{Hi} velocity dispersion
is constant with height (isothermal) and only varies with radius.
This has been assumed for a very long time, as for example by
\citet{vdk81c}, based on the Galactic work done by \citet{Celnik1979A},
who found little evidence for variations.
To conclude this paper, we test this assumption.
We use the highly resolved observation of ESO\,274-G001.
We begin with the self-absorption results from Section
\ref{sec:HIdecomposition_ESO274} and fix the flaring at the measured values.
A mask was created on the central plane of the galaxy with a
thickness of 290\,pc.
Two models were created, one in which only this central plane is
visible, and one in which this plane is masked.
The results are shown in Figure \ref{fig:sigmatest}.
The rotation curve of the high latitude fit can be seen to lag
compared to the combined and main profiles.
The velocity dispersion of the high latitude fit is 1\,km/s
larger than the central plane fit, with the combined fit falling
in between the other two curves.
Note that this test is stretching the data to its absolute limits.
The error-bars are large; the only consolation is that we observe
this behavior in the fits to the individual sides.
These results do not rule out that the \textsc{Hi} in galaxies is not isothermal.
It is beyond the scope of this project and the available data to
test this in more detail in this series of papers.
It will be very interesting to explore this using the next
generation of radio synthesis telescopes.
\section{Discussion \& Conclusions}
In the previous section, we have measured the structure and kinematics of
six edge-on galaxies.
For each of these galaxies, we presented a fit that assumed an optically
thin medium, as well as fit of a self-absorption medium with a spin
temperature of 100\,K.
As already predicted in Section \ref{sec:whenmodelgobad}, not
accounting for self-absorption leads to very different results
than the self-absorption models give.
The most striking of this is the face-on column density, which
is often measured 25\% lower in the inner parts of the optically
thin models.
Other problems that occur when assuming an optically thin medium
are an overestimation of the thickness of the \textsc{Hi} disc and a
completely wrong velocity dispersion.
In our test models (Section \ref{sec:whenmodelgobad}), the
velocity dispersion tends to be too high in the inner parts and
too low in the outer part, when fitting the optically thin results
to a self-absorption model, although this exact behaviour might not
be universal.
The implication of this work is that the \textsc{Hi} mass, thickness and
velocity dispersion of galaxies are incorrect in the literature.
Comparing the optically thin results to those published in the literature,
we find that our optically thin result match well.
Only the velocity dispersion of the optically thin model appear very
different compared to the work by \citet{OBrien2010C}.
This difference is not fully understood, but is most likely due to
the $z$-integrated PV diagram fitting by \citet{OBrien2010C}.
Self-absorption has the effect of flattening Gaussian profile of
the \textsc{Hi} at a particular radius.
As the data is integrated with height, it will lead to a different
observed slope at the outer edges, than would appear in a normal
Gaussian profile.
\begin{table*}
\centering
\begin{tabular}{l|rr}
Name & Optically thin mass [M$_\odot$] & Self-absorbing mass [M$_\odot$]\\\hline\hline
IC\,5052&$7.4\pm0.7\times10^8$&$9.5\pm0.9\times10^8$\\
IC\,5249&$4.8\pm0.2\times10^9$& $7.8\pm0.8\times10^9$\\
ESO\,115-G021&$5.6\pm0.2\times10^9$&$7.2\pm0.1\times10^9$\\
ESO\,138-G014&$3.4\pm0.1\times10^9$&$4.6\pm0.2\times10^9$\\
ESO\,274-G001&$3.1\pm0.02\times10^8$&$4.1\pm0.1\times10^8$\\
UGC\,7321&$8.9\pm0.4\times10^8$&$1.17\pm0.02\times10^9$\\\hline\hline
\end{tabular}
\caption[Total \textsc{Hi} mass of galaxies]{The total \textsc{Hi} mass of each galaxy as measured with our outer envelope strategy assuming either an optically thin or a self-absorbing mass.}
\label{table:HImassmeasured}
\end{table*}
In Table \ref{table:HImassmeasured}, we show the list of all \textsc{Hi}
masses as measured using our outer envelope strategy.
On average we find that $27\pm6$\% of the \textsc{Hi} mass is hidden by
self-absorption.
In Section 7 of Paper II, we predicted that
self-absorption would hide 30\% of the mass if galaxy NGC\,2403
had been viewed edge-on.
These results are thus compatible.
However, we stress that our analysis is based on a couple of
important assumption.
The first is the constant spin temperature of 100\,K.
While this appears to work well for the data, its value is
chosen only for convenience.
In reality, the spin temperature is not constant, but is
dependent on location inside the galaxy and the state of the \textsc{Hi} gas.
Assuming a lower average spin temperature would have
resulted in a much higher total \textsc{Hi} mass.
The second assumption is our treatment of the gas as a
uniform medium.
As we already discussed in Section 2.3 of Paper II,
this is not realistic.
In our Galaxy, many lines of sight are known to be optically
thick within a couple of hundred parsec, with the highest
concentrations of \textsc{Hi} forming into distinct cloud structures
\citep{Taylor2003,Allen2012A}.
The simulation of the Galactic plane by \citet{Douglas2010}
shows that the opacity $\tau_\nu$ can go well above 25.
Due to beam smearing and the large distances, we cannot resolve those
cloud structures in our galaxies and as such, a uniform medium is justified.
We stress however that because of the cloud structure of the \textsc{Hi} , it
could well hiding even more \textsc{Hi} in the densest parts of these clouds.
An analysis of the cloud structure in the nearby \emph{face-on}
galaxy M\,31 by \citet{Braun2012A}, showed that $34\%$ of the \textsc{Hi} was
already hidden.
The \emph{edge-on} galaxies analyzed here could thus hide far more \textsc{Hi}
than was measured here.
With these assumptions in mind, we conclude that our new fitting
strategy, along with the presented self-absorption models, are a
more accurate representation of the neutral hydrogen content of
edge-on galaxies, than the optically-thin decomposition strategies
presented here and by other authors.
\section*{Acknowledgments}
SPCP is grateful to the Space Telescope Science Institute, Baltimore, USA, the
Research School for Astronomy and Astrophysics, Australian National University,
Canberra, Australia, and the Instituto de Astrofisica de Canarias, La Laguna,
Tenerife, Spain, for hospitality and support during short and extended
working visits in the course of his PhD thesis research. He thanks
Roelof de Jong and Ron Allen for help and support during an earlier
period as visiting student at Johns Hopkins University and
the Physics and Astronomy Department, Krieger School of Arts and Sciences
for this appointment.
PCK thanks the directors of these same institutions and his local hosts
Ron Allen, Ken Freeman and Johan Knapen for hospitality and support
during many work visits over the years, of which most were
directly or indirectly related to the research presented in this series op
papers.
Work visits by SPCP and PCK have been supported by an annual grant
from the Faculty of Mathematics and Natural Sciences of
the University of Groningen to PCK accompanying of his distinguished Jacobus
C. Kapteyn professorhip and by the Leids Kerkhoven-Bosscha Fonds. PCK's work
visits were also supported by an annual grant from the Area of Exact
Sciences of the Netherlands Organisation for Scientific Research (NWO) in
compensation for his membership of its Board.
|
\section{Introduction}
Hard diffractive events in hadron-hadron collisions were first observed at the Tevatron \cite{Abachi:1994hb,Abe:1994de} more than 20 years ago, nevertheless the QCD dynamics at play has yet to be fully understood. In spite of the large transfer of transverse momentum involved in such processes, a satisfactory weak-coupling description remains elusive, and one has to settle for phenomenological models.
To estimate cross-sections of hard processes in single diffractive dissociation (when one hadron escapes the collision intact) and central diffractive dissociation (when both hadrons escape the collision intact), a modern version of the resolved-Pomeron model \cite{Ingelman:1984ns} is being widely used.
This model describes hard diffractive scattering in the following way: hadrons scatter through the exchange of a colorless objet called the Pomeron, which carries a longitudinal momentum fraction denoted $\xi$ and a four-momentum squared denoted $t$. Then, imitating what happens in collinear factorization, a long-distance/short-distance separation of the Pomeron-induced subprocesses is assumed, into perturbative partonic cross-sections and non-perturbative parton distribution functions (pdfs) of the Pomeron, that depend on the parton fractional longitudinal momentum $\beta$, and on the hard scale of the problem $\mu^2$.
The motivation for this model comes from the fact that, in electron-hadron collisions, the diffractive part of the deep inelastic scattering (DIS) cross-section does obey collinear factorization \cite{Collins:1997sr}. The further factorization of the diffractive parton densities $f_{a/h}^D$ into a Pomeron flux $\Phi_{\mathbb{P}/h}(\xi,t)$ and Pomeron parton distributions $f_{a/\mathbb{P}}(\beta,\mu^2)$ is an assumption, called Regge factorization, which is accurate and routinely used in QCD fits of diffractive DIS data, from which $f_{q/\mathbb{P}}$, $f_{g/\mathbb{P}}$ and $\Phi_{\mathbb{P}/h}$ are extracted.
When imported to hadron-hadron collisions, such factorization does not apply for diffractive processes, even at very large momentum scales, as shown by comparisons to Tevatron data \cite{Affolder:2000vb}. The presence of additional soft interactions between the colliding hadrons, which may fill the rapidity gap(s), is the standard interpretation of this factorization breaking, and there are empirical indications that it can be compensated by an overall factor, called the gap survival probability, roughly independent of the details of the hard process. In fact, the phenomenology of this factor is still a topic of intense debate~\cite{sgap1,sgap2,sgap3,sgap4} and it represents the last ingredient of the resolved-Pomeron model.
At the LHC, a whole new set of experimental studies has started, in order to provide answers to a number of unsolved questions. Is the gap survival probability only a function of the collision energy, as often assumed? Does one need a different factor for single diffraction and central diffraction? Is the quark and gluon composition of the Pomeron extracted from HERA data compatible with LHC measurements? In this letter, we would like to study a different aspect that has not been investigated extensively: the possibility that the diffractive scattering happens through the exchange of a Reggeon, as opposed to a Pomeron. As a matter of fact, quality fits to diffractive DIS data do require that both Pomerons and Reggeons contribute to the diffractive pdfs: $f_{a/h}^D\!=\!\Phi_{\mathbb{P}/h}f_{a/\mathbb{P}}\!+\!\Phi_{\mathbb{R}/h}f_{a/\mathbb{R}}$.
The differences between the two contributions resides in the $\xi$ and $t$ dependence of their fluxes. Reggeon exchange matters mostly at high $\xi$, notably for $\xi>0.1$ and the shape of the $t$ distribution is also different, showing a less steep decrease than in the Pomeron case.
At the LHC, when large diffractive masses are considered - which is the case in a number of studies (see for instance \cite{N.Cartiglia:2015gve}) - such large values are easily reached and one may therefore wonder about the importance of the Reggeon contribution.
Previous studies have shown that the Reggeon contribution is not always negligible~\cite{Luszczak:2014mta,Luszczak:2014cxa,Luszczak:2016csq} but this is not always taken into account by standard codes. Our goal in this work is to illustrate, within a very simple model where the parton content of the Reggeon is obtained from the pion structure function, that indeed the Reggeon contribution cannot be safely neglected, and that for processes where both protons escape the collision intact, a double-Reggeon exchange can even dominate over a double-Pomeron exchange.
The plan of the letter is as follows. In Section 2, we present more details about the resolved-Pomeron model for hard diffraction in hadron-hadron collisions, we explain its implementation into the Forward Physics Monte Carlo (FPMC) program \cite{fpmc} that we shall utilize, and we outline our subsequent analysis of diffractive di-jet production, the process we have chosen to consider. In Section 3, we present our results, when the Reggeon contribution is included, for both single and central diffractive di-jets. Section 4 is devoted to conclusions and outlook.
\section{Hard diffractive processes with Reggeon exchanges}
\subsection{Resolved Pomeron model supplemented with Reggeons}
\begin{figure}
\includegraphics[scale=0.37]{Fig1a.pdf}
\hspace{2.5cm}
\includegraphics[scale=0.37]{Fig1b.pdf}
\caption{Leading-order diagrams for di-jet production in single-diffractive events (left) and central-diffractive events (right) in proton-proton collisions. Intact protons can scatter through the exchange of either a Pomeron ($\mathbb{P}$) or a Reggeon ($\mathbb{R}$).}
\label{fig0}
\end{figure}
The resolved-Pomeron model is a long-distance/short-distance collinear factorization framework commonly used to calculate hard single-diffractive (SD) and central-diffractive (CD) processes. In this work we focus on di-jet production at the LHC. Typical leading-order (LO) diagrams for this process are pictured in Fig.~\ref{fig0}, and the cross-section in the resolved-Pomeron model reads:
\begin{eqnarray}
d\sigma^{pp\to pJJX}={\cal S}_{SD}\ &&\sum_{a,b} \int f^D_{a/p}(\xi_1,t_1,\beta_1,\mu^2)f_{b/p}(x_2,\mu^2)\
\otimes d\hat\sigma^{ab\to JJX}\label{colfact1}
\end{eqnarray}
\begin{eqnarray}
d\sigma^{pp\to pJJXp}={\cal S}_{CD}\ &&\sum_{a,b} \int f^D_{a/p}(\xi_1,t_1,\beta_1,\mu^2)f^D_{b/p}(\xi_2,t_2,\beta_2,\mu^2)\
\otimes d\hat\sigma^{ab\to JJX}\label{colfact2}
\end{eqnarray}
where $d\hat\sigma$ is the short-distance partonic cross-section, which can be computed
order by order in perturbation theory (provided the transverse momentum of the jets is
sufficiently large). $f_{a/p}$ denotes the standard proton parton distributions functions (pdfs) while $f^D_{a/p}$ denotes the diffractive ones.
These are non-perturbative objects, however their evolution with the factorization scale $\mu$ is perturbative and given by the
Dokshitzer-Gribov-Lipatov-Altarelli-Parisi~\cite{dglap} evolution equations (in the following $\mu$ is set to the transverse momentum of the leading jet). In Eq.~\eqref{colfact1} and~\eqref{colfact2} and in Fig.~\ref{fig0}, the variables $\xi_i$ and $t_i$ for the intact protons denote their fractional energy loss and the four-momentum squared transferred in the collision, respectively. The convolution is done over the longitudinal momentum fractions of the partons $a$ and $b$ with respect to the incoming protons, namely $x_1$ and $x_2$, respectively. In the case of intact protons, it is common to use instead $\beta_i \equiv x_i/\xi_i$, the longitudinal momentum fraction of the parton with respect to the exchanged Pomeron or Reggeon.
Formulae \eqref{colfact1} and \eqref{colfact2} are reminiscent of the collinear factorization obeyed for inclusive processes. However, it is known that hard diffractive cross-sections in hadronic collisions do not factorize in such a way, due to possible secondary soft interactions between the colliding hadrons which can fill the rapidity gaps. In the resolved-Pomeron model, the so-called gap survival probabilities ${\cal S}_{SD}$ and ${\cal S}_{CD}$ act as corrections to collinear factorization in order to account for the effects of the soft interactions. Since those happen on much longer time scales compared to the hard process, they are modeled by an overall factor, function of the collision energy only~\cite{sgap1,sgap2}.
\subsection{Pomeron and Reggeon parton-content}
In our computations, we shall use the diffractive pdfs $f^D_{a/p}$ extracted from HERA data \cite{Aktas:2006hy} for diffractive DIS, a process for which collinear factorization does hold. They are obtained by means of a next-to-leading-order (NLO) QCD fit, in which they are decomposed into Pomeron and Reggeon fluxes $\Phi_{\mathbb{P},\mathbb{R}/p}$ and their corresponding parton distribution functions $f_{a/\mathbb{P},\mathbb{R}}$ which depict the partonic structure of the exchanged color singlet objects:
\begin{equation}
f^D_{a/p}(\xi,t,\beta,\mu^2) = \Phi_{\mathbb{P}/p}(\xi,t) f_{a/\mathbb{P}}(\beta,\mu^2) + n_{\mathbb{R}} \Phi_{\mathbb{R}/p}(\xi,t) f_{a/\mathbb{R}}(\beta,\mu^2) \quad \mbox{with}
\quad \Phi_{\mathbb{P},\mathbb{R}/p}(\xi,t) = A_{\mathbb{P},\mathbb{R}}\ \frac{e^{B_{\mathbb{P},\mathbb{R}}t}}{\xi^{2\alpha_{\mathbb{P},\mathbb{R}}(t)-1}}\ .
\label{dpdfs}
\end{equation}
The diffractive slopes $B_{\mathbb{P},\mathbb{R}}$, and the Regge trajectories $\alpha_{\mathbb{P},\mathbb{R}}(t)=\alpha_{\mathbb{P},\mathbb{R}}(0)+t\alpha'_{\mathbb{P},\mathbb{R}}$ are given in Table \ref{fitvalues} for two different fits (known as $A$ and $B$).
The flux normalizations $A_{\mathbb{P},\mathbb{R}}$ are chosen such that $\xi\times\int_{t_{\rm{min}}}^{t_{\rm{max}}}dt\ \Phi_{\mathbb{P},\mathbb{R}/p}(\xi,t)=1$ at $\xi=0.003$, with $t_{\rm{min}}=-1$ GeV$^2$ and $t_{\rm{max}}=-m_p^2\xi^2/(1\!-\!\xi)$ ($m_p$ denotes the proton mass).
The factor $n_{\mathbb{R}}$ is an extra normalization to the Reggeon contribution.
The Pomeron pdf $f_{a/\mathbb{P}}$ is obtained from fits to H1 data~\cite{Aktas:2006hy}.
The Pomeron structure is well constrained by those fits, which clearly show that its parton content is gluon dominated.
By contrast, the HERA data do not constrain $f_{a/\mathbb{R}}$. The Reggeon contribution is however needed in order to obtain a quantitative description of the high-$\xi$ measurements.
Following the description in Ref.~\cite{Aktas:2006hy} we treat the Reggeon contribution as an exchange of a quark-antiquark pair and take $f_{a/\mathbb{R}}$ as the pion structure function.
\begin{table}[b]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\ \textbf{Fits}\, &$\ \Phi_{\mathbb{P},\mathbb{R}/p}\, $&$\alpha(0)$ & $\alpha'$ & $n_{\mathbb{R}}$ & $B$ \\ \hline &&&&&\\
&$\mathbb{P}$ & $1.118\pm 0.008$ & $0.06_{-0.06}^{+0.19}$ GeV$^{-2}$ & - & $5.5_{-2.0}^{+0.7}$ GeV$^{-2}$ \\ A &&&&& \\
&$\mathbb{R}$ & $0.50\pm 0.10$ & $0.3^{+0.6}_{-0.3}$ GeV$^{-2}$ & $(1.7\pm 0.4) \times 10^{-3}$ & $1.6_{-1.6}^{+0.4}$ GeV$^{-2}$ \\ &&&&& \\ \hline &&&&& \\
&$\mathbb{P}$ & $1.111\pm 0.007$ & $0.06_{-0.06}^{+0.19}$ GeV$^{-2}$ & - & $5.5_{-2.0}^{+0.7}$ GeV$^{-2}$ \\ B &&&&& \\
&$\mathbb{R}$ & $0.50\pm 0.10$ & $0.3^{+0.6}_{-0.3}$ GeV$^{-2}$ & $(1.4\pm 0.4) \times 10^{-3}$ & $1.6_{-1.6}^{+0.4}$ GeV$^{-2}$ \\ &&&&& \\ \hline
\end{tabular}
\end{center}
\caption{
Parameters of the Pomeron and Reggeon fluxes as described in the text, from Ref.~\cite{Aktas:2006hy}.
}
\label{fitvalues}
\end{table}
The Reggeon contribution is often not implemented in hard diffraction studies in hadron-hadron collisions, and measurements at the LHC will allow to test the validity of this assumption.
The Reggeon contribution to the diffractive pdfs $f^D_{a/p}$ is important only at large values of $\xi$~\cite{Aktas:2006hy,Aktas:2006hx}.
It is routinely disregarded, and subsequently the theoretical description from \eqref{colfact1} and \eqref{colfact2} was dubbed the resolved-Pomeron model. But of course it can be supplemented with resolved Reggeons, and it should be as we will argue below.
Similarly, this is the reason why what we call central diffractive events in this work is usually referred to as double-Pomeron-exchange events in the literature. Since double-Reggeon exchanges or mixtures of Pomeron and Reggeon exchanges are also possible, we choose to use the terminology \enquote{central diffractive}.
The central diffractive di-jet final states considered in this work are not exclusive since they contain the so-called Pomeron or Reggeon remnants $X$, that are made of soft particles accompanying the production of the hard di-jet system. They reduce the rapidity gaps when compared to the exclusive case, but they do not fill them entirely.
\subsection{Hard diffractive di-jet analysis with FPMC}
The above theoretical description of hard diffractive processes in hadron-hadron collisions, in which one or both hadrons remain intact, is implemented by the FPMC generator that we shall employ to perform our analysis. The parton-level matrix elements are imported from HERWIG \cite{herwig} routines and calculated at LO, while the NLO fit B is adopted for the diffractive pdfs~\footnote{We decided to use LO partonic cross-sections convoluted with NLO pdfs due to the lack of an updated LO diffractive pdf.}. For proton tagging at the LHC, a region of $\xi<0.17$ for both protons is chosen for a center of mass energy of 13 TeV, as well as a lower cut of $0.15$ GeV for their transverse momenta \cite{Trzebinski:2014vha}.
In general, the lower boundary for the $\xi$ values depends on the minimum mass of the diffractive system and is thus related to the jet transverse momenta. For the acceptance we have chosen the lower $\xi$ boundary is approximately $10^{-5}$.
For the di-jet system, we apply at parton level a transverse momentum cut on $p_T\!>\!5$ GeV, and a pseudo-rapidity cut of $|\eta|\!<\!5$. The jets are reconstructed using the FastJet~\cite{fastjet:2012} package and the anti-kt algorithm, with a value of 0.4 for the jet radius, and a 10 GeV threshold for the transverse momenta. Then, the selection criteria requires at least two jets with $p_T$ larger than 20 GeV, and the two highest transverse momentum jets tagged with $p_{T}(j_{1})\!>\!p_{T}(j_{2})$. The di-jet mass fraction is defined as $R_{JJ}\!=\!m_{JJ}/M$, i.e. the ratio of the invariant mass of the di-jet system to the invariant mass of the whole diffractive
final state, $M\!=\!\sqrt{\xi s}$ and $M\!=\!\sqrt{\xi_{1}\xi_{2}s}$ for single and central diffraction, respectively.
Experimentally, the di-jet mass fraction is a good variable for identifying, and for our purpose excluding, possible exclusive di-jet events. In such events, the di-jet mass is essentially equal to the mass of the central system
because no Pomeron (or) Reggeon remnants are present, and if the jet definition is such that little is left outside the cones, then the presence of an exclusive event would manifest itself as an excess towards $R_{JJ} \approx 1$.
This observation of exclusive events does not depend on the overall normalization of the event distribution, which might be strongly dependent on the detector simulation and acceptance of the roman pot detectors \cite{Kepka:2007nr}.
Finally, the histograms are normalized according to the relation $(\sigma \times \mathcal{L})/N_{gen}$, and our predictions are presented for an integrated luminosity of $1$~pb$^{-1}$ which represents the expected data to be collected in high-$\beta^*$ low pile-up runs at the LHC. Note that for the gap survival probabilities, we have assumed a constant value for Pomerons and Reggeons $S_{SD}=S_{CD} \!\simeq\! 0.03$. There have been several attempts to estimate those probabilities~\cite{Khoze:2000cy,Khoze:2008cx,Kaidalov:2001iz,Bartels:2006ea,Luna:2006qp,Frankfurt:2006jp,Gotsman:2007ac,Gotsman:2011xc,Achilli:2007pn,sgap1,sgap2,sgap3,sgap4}, but the actual values are rather uncertain.
The chosen value of the gap survival factor can be considered a lower limit given the recent available experimental results~\cite{Chatrchyan:2012vc,Aad:2015xis}.
\begin{figure}[t]
\includegraphics[width=0.47\textwidth]{sd_dijetspt20xileq017_ptproton_13tev.pdf}
\includegraphics[width=0.47\textwidth]{sd_dijetspt50xileq017_ptproton_13tev.pdf}
\caption{Number of single diffractive di-jet events as a function of $\xi$ for p$_{T}$(proton)$>$ 0.15 GeV assuming either a Pomeron exchange (solid lines) or a Reggeon exchange (dashed lines), for $p_{T}(j_{1,2})>20$ GeV (left plot) or $p_{T}(j_{1,2})>50$ GeV (right plot).}
\label{xi_sd}
\end{figure}
\section{Numerical results for the LHC}
\subsection{Single diffractive di-jets}
The total cross-sections predicted by FPMC at 13 TeV for single diffractive di-jets assuming either a Pomeron exchange ($\mathbb{P}+p\to jjX$) or a Reggeon exchange ($\mathbb{R}+p\to jjX $) are $1.51 \times 10^{8}$ pb and $2.3\times 10^{7}$ pb, respectively. These values assume an acceptance of $\xi_1\equiv\xi\leq 0.17$ for the final state intact proton. The $\xi$ distributions are plotted in Figure~\ref{xi_sd} for two different values of minimum jet $p_T$, 20 or 50 GeV. One clearly sees the dominance of the Pomeron exchange at small $\xi$ (the Reggeon contribution can be neglected for $\xi\lesssim0.07$), but also the fact that the Reggeon contribution becomes comparable to it for $\xi\gtrsim0.1$, depending slightly on the jet $p_T$ cut.
\begin{figure}[t]
\includegraphics[width=0.47\textwidth]{sd_dijetsratiopt20_xi017_ptproton_13tev.pdf}
\includegraphics[width=0.47\textwidth]{sd_dijetsratiopt20_xi010_ptproton015_13tev.pdf}
\caption{Di-jet mass fraction distribution in single diffraction assuming either a Pomeron exchange (solid lines) or a Reggeon exchange (dashed lines), for
$p_{T}(j_{1,2})>20$ GeV, p$_{T}$(proton) $>$ 0.15 GeV and for $\xi<0.17$ (left) or $0.1<\xi<0.17$ (right).}
\label{dmfsdpt04}
\end{figure}
This is confirmed by Table~\ref{neventssd_ptcuts} where the number of events are displayed for three different $\xi$ ranges: no minimum $\xi$ cut, $\xi>0.015$ and $\xi>0.1$. In the latter case, the number of events for the Pomeron and Reggeon contributions have the same order of magnitude. Note that the events for the Pomeron process take into account the uncertainty of the QCD fits at high $\beta$: the gluon density $f_{g/\mathbb{P}}(\beta,Q^2)$ is multiplied by an uncertainty factor $(1-\beta)^{\nu}$, with $\nu = -0.5$, $0$, or $0.5$ (the default value in FPMC is $\nu=0$). The uncertainty range of the Reggeon contribution is not known. Finally the di-jet mass fraction distributions are displayed in Figure \ref{dmfsdpt04}, for $\xi<0.17$ and $0.1<\xi<0.17$. While the Reggeon exchange can be as important as the Pomeron exchange, there is no kinematical window where it clearly dominates which would allow to experimentally isolate it.
Those findings confirm the expectations, that for single diffractive processes sensitive to $\xi>0.1$, the Reggeon contribution should play a non-negligible role. Therefore, the LHC capabilities should be utilized in order to constrain it better, and improve the theoretical predictions of the various high-mass diffractive studies. We will demonstrate below that central diffractive di-jet production can also be used to constrain the Reggeon contribution.
\begin{table}[h]
\renewcommand{\arraystretch}{1.5}%
\adjustbox{max height=\dimexpr\textheight-1.0cm\relax,max width=\textwidth}{
\begin{tabular}{|p{0.18\linewidth}|p{0.30\linewidth}|p{0.10\linewidth}|p{0.30\linewidth}|p{0.10\linewidth}|}
\hline
\textbf{Process} & $\mathbb{P}\ p\to jjX$ & $\mathbb{R}\ p\to jjX$ & $\mathbb{P}\ p\to jjX$ & $\mathbb{R}\ p\to jjX$ \\
\hline
\textbf{Acceptance} & \multicolumn{2}{c|}{$p_{T}(j_{1},j_{2})>20$ GeV } & \multicolumn{2}{c|}{$p_{T}(j_{1},j_{2})>50$ GeV } \\
\hline
$\xi_{1,2} < 0.17$ & $6.06 \times 10^{5}$ [$5.85\times 10^{5}$, $6.74\times 10^{5} $] & $1.38 \times 10^{4} $ & $ 2.51 \times 10^{4}$ [$ 2.26 \times 10^{4} $, $2.86 \times 10^{4} $ ] & $5450 $ \\
$0.015 < \xi_{1,2} < 0.17$ & $4.58 \times 10^{5}$ [$4.53\times 10^{5}$, $5.03\times 10^{5} $] & $1.37 \times 10^{4} $ & $ 1.99 \times 10^{4}$ [$ 1.81\times 10^{4} $, $2.32 \times 10^{4} $ ] & $5419 $ \\
$0.10 < \xi_{1,2} < 0.17$ & $1.49 \times 10^{5}$ [$1.46\times 10^{5}$, $1.62\times 10^{5} $] & $8.77 \times 10^{4} $ & $6561 $ [$6341$, $8827 $] &$3521$ \\
\hline
\end{tabular}
}
\caption{Number of single diffractive di-jet events for an integrated luminosity of $1$ pb$^{-1}$, and different kinematical windows. For the Pomeron process, the left values inside the brackets stand for $\nu=0.5$, whereas the right values stand for $\nu=-0.5$.}
\label{neventssd_ptcuts}
\end{table}
\subsection{Central diffractive di-jets}
In central diffraction, besides the double-Pomeron ($\mathbb{P}\pom\to jjX$) and double-Reggeon ($\mathbb{R}\reg\to jjX$) exchanges, there are also cross terms
($\mathbb{P}\mathbb{R},\mathbb{R}\mathbb{P}\to jjX$) which makes the approximation of disregarding the Reggeon contributions even more questionable. The total cross-sections predicted by FPMC for proton-proton collisions at 13 TeV for those distinct channels are $ 1.7\times 10^{7}$ pb ($\mathbb{P}\pom$), $9.1 \times 10^{6}$ pb ($\mathbb{P}\mathbb{R}\!+\!\mathbb{R}\mathbb{P}$), and $9.03 \times 10^{5}$ pb ($\mathbb{R}\reg$), again for an acceptance of $\xi_{1,2}\leq 0.17$ for the final state intact protons. These values are in agreement with the prediction that single diffractive cross-sections should be approximately 10 times greater than in the central diffractive case~\cite{Aad:2012pw}.
The $\xi$ distributions are plotted in Figure~\ref{centraldiffrative_xi} for a minimum jet $p_T$ of 20 GeV. The Pomeron exchange is still dominant at small values of $\xi_{1,2}$, albeit by a lesser margin than in the single diffractive case, but now the Reggeon contributions dominate for large values of proton momentum loss. They become comparable to the Pomeron one for $\xi_1\sim 0.14$ when $\xi_2$ is integrated in the whole acceptance (Fig.~\ref{centraldiffrative_xi} - right panel).
\begin{figure}[t]
\includegraphics[width=0.47\textwidth]{pomreg_dijetsxi12sumsquare_ptproton_13tev.pdf}
\includegraphics[width=0.47\textwidth]{pomreg_dijetsxi1_ptproton_13tev.pdf}
\caption{Number of central diffractive di-jet events as function of $\sqrt{\xi_1\xi_2}$ (left plot) and $\xi_1$ (right plot) for $p_{T}(j_{1,2}) > 20$ GeV and p$_{T}$(proton) $>$ 0.15 GeV. The solid line stands for the double-Pomeron exchange while the dashed line represents the total Reggeon contribution $(\mathbb{R}\ \mathbb{R}$ and $\mathbb{P}\ \mathbb{R}+ \mathbb{R}\ \mathbb{P})$.}
\label{centraldiffrative_xi}
\end{figure}
\begin{figure}[t]
\includegraphics[width=0.49\textwidth]{pomreg_dijetsratioxi017_13tev.pdf}
\includegraphics[width=0.47\textwidth]{pomreg_dijetsratioxi010_pt015_13tev.pdf}
\caption{Di-jet mass fraction distribution in central diffraction for $p_{T}(j_{1,2})>20$ GeV, p$_{T}$(proton) $>$ 0.15 GeV and for $\xi_{1,2}<0.17$ (left plot) or $0.1<\xi_{1,2}<0.17$ (right plot). The solid line stands for the double-Pomeron exchange while the dashed line represents the total Reggeon contribution from the double-Reggeon and the Pomeron-Reggeon exchanges.
}
\label{dmfcentraldiffrative_xi}
\end{figure}
This is confirmed by Table~\ref{neventsptcut1} where the number of events is displayed for the same three different $\xi_{1,2}$ ranges and two different jet $p_T$ cuts considered before. As expected, using a larger minimum jet $p_T$ further enhances the importance of the Reggeon exchange, however in central diffraction the number of events quickly becomes too small and it may not be efficient to use a value greater than 20 GeV. Finally, the di-jet mass fraction distributions are displayed in Figure \ref{dmfcentraldiffrative_xi}, for $\xi_{1,2}\!<\!0.17$ and $0.1\!<\!\xi_{1,2}\!<\!0.17$.
\begin{table}[h]
\renewcommand{\arraystretch}{1.5}%
\adjustbox{max height=\dimexpr\textheight-1.0cm\relax,max width=\textwidth}{
\begin{tabular}{|p{0.18\linewidth}|p{0.30\linewidth}|p{0.30\linewidth}|p{0.10\linewidth}|p{0.14\linewidth}|p{0.15\linewidth}|p{0.10\linewidth}|p{0.10\linewidth}|p{0.10\linewidth}|p{0.10\linewidth}|}
\hline
\textbf{Process } & $\mathbb{P}\ \mathbb{P}\to jjX$ & $\mathbb{P}\ \mathbb{R}\!+\!\mathbb{R}\ \mathbb{P}\to jjX$ & $\mathbb{R}\ \mathbb{R}\to jjX$ & $\mathbb{P}\ \mathbb{P}\to jjX$ & $\mathbb{P}\ \mathbb{R}\!+\!\mathbb{R}\ \mathbb{P}\to jjX$ & $\mathbb{R}\ \mathbb{R}\to jjX$\\
\hline
\textbf{Acceptance} & \multicolumn{3}{c|}{$p_{T}(j_{1},j_{2})>20$ GeV } & \multicolumn{3}{c|}{$p_{T}(j_{1},j_{2})>50$ GeV } \\
\hline
$ \xi_{1,2} < 0.17$ & $3.34\times 10^{4} $ [$2.88\times 10^{4} $, $4.42\times 10^{4} $] & $1.56\times 10^{4}$ [$1.41\times 10^{4}$, $1.68\times 10^{4} $] & 1610 & 1489 [1198,1829] & 697[594,771] & 72 \\
$0.015 < \xi_{1,2} < 0.17$ & $2.19\times 10^{4} $ [$1.99 \times 10^{4} $, $2.79\times 10^{4} $] & $1.26\times 10^{4} $ [$1.16\times 10^{4}$, $1.35\times 10^{4}$] & 1590 &1030 [876,1269] &576 [536,644] & 70 \\
$0.10 < \xi_{1,2} < 0.17$ & 2530 [2319,3193] & 2802 [2627,2850] & 680 & 120[113,135] &148 [140,174] & 35\\
$0.10 < \xi_{1,2} < 0.17\ [\star]$ & 544 [499,736] & 865 [813,877] & 312 &20.5 [10,23] &42 [36,52] & 30 \\
\hline
\end{tabular}
}
\caption{Number of central diffractive di-jet events for an integrated luminosity of $1$ pb$^{-1}$, and different kinematical windows. For the Pomeron process, the left values inside the brackets stand for $\nu=0.5$, whereas the right values stand for $\nu=-0.5$. The last line $[\star]$ is for $p_{T}$(proton) $>$ 0.4 GeV, instead of the default value 0.15 GeV.}
\label{neventsptcut1}
\end{table}
Finally, in Figure~\ref{newprotonptcut} we study the sensitivity of our results with respect to the cut on the proton transverse momentum. By choosing an alternative cut of $0.4$ GeV, we are able to increase the sensitivity to the Reggeon contribution, such that near the edge of the proton detector acceptance, it becomes clearly dominant.
By measuring these distributions, we should be able to study the Reggeon contribution in the LHC data.
Our results show that di-jets in central diffractive events (formerly known as double-Pomeron-exchange events) at the LHC could be used to study the Reggeon contribution to hard diffractive processes, since a kinematic window of dominance has been identified which could be used experimentally to isolate and constrain it.
\begin{figure}[t]
\includegraphics[width=0.49\textwidth]{pomreg_dijetsxi1_ptproton04_13tev.pdf}
\includegraphics[width=0.47\textwidth]{pomreg_dijetsratio_01xi017_13tev.pdf}
\caption{
Number of central diffractive di-jet events as function of $\xi_1$ for $\xi_2<0.17$ (left plot) and di-jet mass fraction distribution for $0.1<\xi_{1,2}<0.17$ (right plot), with $p_{T}(j_{1,2}) > 20$ GeV and p$_{T}$(proton) $>$ 0.4 GeV. Increasing the last cut enhances the sensitivity to the Reggeon contribution.}
\label{newprotonptcut}
\end{figure}
\section{Conclusions}
In this letter, we studied hard diffractive processes in hadron-hadron collisions using the resolved-Pomeron model \eqref{colfact1} and \eqref{colfact2}, supplemented with a Reggeon term according to formula \eqref{dpdfs}. Our goal is to verify whether or not this contribution can be safely neglected at LHC energies. For the moment, it is often ignored when estimating hard diffractive cross-sections in hadron-hadron collisions, even though it is needed for a quantitative description of the diffractive DIS HERA data. The Pomeron structure used in the resolved-Pomeron model is extracted from DIS, therefore a consistency check is in order at the LHC.
To do this, we chose to analyze the diffractive di-jet process at the LHC, assuming an integrated luminosity of $1$~pb$^{-1}$. We have assumed a simple model in which the parton content of the Reggeon $f_{a/\mathbb{R}}(\beta,\mu^2)$ is given by the pion structure function, but it should be pointed out that the related uncertainties are large since the Reggeon structure at low $\beta$ and high transverse momentum scales is essentially unknown and unconstrained experimentally.
Our calculations have been performed using the Forward Physics Monte Carlo program. In the case of single diffractive di-jet production, our results confirm the expectation that the Reggeon contribution is comparable to the Pomeron contribution only for $\xi \gtrsim 0.1$. Since the acceptance of the LHC forward proton detectors can go up to $\xi \sim 0.15 - 0.2$, it must be carefully taken into account when the total diffractive mass becomes large, which is the case for a number of final states considered in the literature, e.g.~\cite{N.Cartiglia:2015gve}.
In the case of central diffractive di-jet production, we find that Reggeon exchanges contribute much more, and can almost never be completely ignored, at least in our model.
For large values of $\xi_{1,2}$ but still within the detector acceptances, processes involving Reggeons can even dominate over the double-Pomeron exchange. This should allow relatively clean experimental studies in order to better constrain the Reggeon parton content and correct the model. Subsequently, many phenomenological studies of double-Pomeron-exchange events at the LHC, such as \cite{Marquet:2012ra,Marquet:2013rja,Kohara:2015nda}, will have to be corrected in order to take into account the possibility to exchange Reggeons as well.
\begin{acknowledgments}
D.E. Martins, M. Rangel and A.V. Pereira acknowledge the financial support from the funding agencies CNPq, CAPES and FAPERJ (Brazil). D.E. Martins thanks the Centre de Physique Th\'eorique of \'Ecole Polytechnique for hospitality. He also warmly thanks Gregory Soyez for explanations about the FastJet implementation and the related technical aspects on this analysis.
\end{acknowledgments}
|
\section{Introduction}
The discovery of the Higgs boson established the existence of spin-0
particles in nature and this opened up the new era in looking for
extensions of the Standard Model (SM) at accelerators. It is now expected
that at higher energies, the SM may be embedded in larger
gauge structures, whose gauge symmetries would have been broken by the
new Higgs scalars. So, we can expect signals of the new gauge bosons,
additional Higgs scalars as well as the extra fermions required to
realize the higher symmetries.
One of the extensions of the SM with the gauge group $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ provides
strong promise of new physics that can be observed at the LHC or the
next generation accelerators
\cite{Singer:1980sw,Valle:1983dk}. Recently there has been a renewed
interest in this model as it can provide novel ways to understand
neutrino masses \cite{Boucenna:2014ela,Boucenna:2014dia}.
The $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model proposed by Singer, Valle and Schechter (SVS)
\cite{Singer:1980sw} has the special feature that it is not anomaly
free in each generation of fermions, but only when all the three
generations of fermions are included the theory becomes anomaly
free. As a result, different multiplets of the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ group appear with
different multiplicity and as a result it becomes difficult to unify
the model within usual grand unified theories. For this reason string
completions have been suggested~\cite{Addazi:2016xuh}. In this article
we study how such a theory can be unified in a larger SU(6) gauge
theory that can emerge from a E(6) Grand Unified Theory
(GUT)~\cite{Gursey:1975ki}. We find that the anomaly free
representations of the SVS 331 model can all be embedded in a
combination of anomaly free representations of SU(6), which in turn
can be potentially embedded in the fundamental and adjoint
representations of the group E(6) motivated by F-theory GUTs with
matter and bulk exotics obtained from the flux breaking
mechanism~\cite{1126-6708-2009-01-059,King:2010mq,Callaghan:2011jj,Callaghan:2013kaa}.
Interestingly,
the SVS 331 model can also be refurbished in an anomaly free multiplet
structure which can be right away embedded in a minimal anomaly free
combination of representations of SU(6) as an E(6) subgroup. We refer
to this new 331 model as the sequential 331 model. This scheme is
particularly interesting since its embedding in SU(6) does not require
any bulk exotics to account for the chiral families; and in that sense
it provides a truly minimal unification scenario in the same spirit
akin to the minimal SU(5) construction~\cite{Georgi:1974sy}.
The article is organized as follows. In Section~\ref{sec2} we discuss
the basic structure of the SVS $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model whereas Section~\ref{sec3}
describes the sequential $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model. In Section~\ref{sec4} we then
analyze the resulting renormalization group running of the gauge
couplings in the SVS model with and without additional octet states,
and discuss necessary conditions for gauge unification. In
Section~\ref{sec5} we then embed the different variants of the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$
model in an SU(6) unification group and demonstrate successful
unification scenarios. Section~\ref{sec6} concerns the experimental
constraints from achieving the correct electroweak mixing angle and
satisfying proton decay limits. We conclude in
Section~\ref{sec:discussion-outlook}.
\section{The SVS $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ Model}{\label{sec2}}
The $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ extension of the SM was originally proposed to
justify the existence of three generations of fermions, as the model
is anomaly free only when three generations are present.
Such a \emph{non-sequential} model, which is generically referred to as the
331 model, breaks down to the SM at some higher energies,
usually expected to be in the TeV range, making the model testable in
the near future. The symmetry breaking: $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ $\to$ $\mathrm{ SU(3)_c \otimes SU(2)_L \otimes U(1)_Y}$ allows us to
identify the generators of the 321 model in terms of the generators of
the 331 model. Writing the generators of the $\mathrm{SU(3)_L}$ group
as
\begin{eqnarray}{\label{2.1}}
T_3 = \frac{1}{2} I_3 = \begin{pmatrix} \frac{1}{2}&0&0 \\ 0 & - \frac{1}{2} & 0 \\
0 & 0 & 0\end{pmatrix} \quad &{\rm and }& \quad T_8 = \frac{1}{2 \sqrt{3}} I_8 = \frac{1}{2 \sqrt{3}}
\begin{pmatrix}1&0&0 \\ 0&1&0 \\ 0&0&-2 \end{pmatrix} \nonumber \\
{\rm with}~~ I_3 = {\rm diag} [ 1,~-1, ~ 0 ] ~ &{\rm and}&~
I_8 = {\rm diag} [ 1,~1, ~ -2 ],
\end{eqnarray}
we can readily identify the SM hypercharge and the electric charge as
\begin{equation}{\label{2.2}}
Y= \frac{1}{\sqrt{3}} T_8 + X = \frac{1}{6} I_8 + X ~~~ {\rm and} ~~~
Q = T_3 + Y = \frac{1}{2} I_3 + \frac{1}{6} I_8 + X \,.
\end{equation}
This allows us to write down the fermions and the representations
in which they belong as
\begin{eqnarray}{\label{2.3}}
Q_{iL} &=& \begin{pmatrix} u_{iL} \\ d_{iL} \\ D_{iL} \end{pmatrix} \equiv
[3,3,0], \quad
Q_{3L} = \begin{pmatrix} b_L \\ t_L \\ T_{L} \end{pmatrix} \equiv
[3,{3}^\ast,1/3], \nonumber \\[.15in]
u_{iR} &\equiv& [3,1,2/3], ~~ d_{iR} \equiv [3,1,-1/3], ~~ D_{iR}
\equiv [3,1,-1/3] \nonumber \\[.15in]
b_R &\equiv& [3,1,-1/3], ~~ t_{R} \equiv [3,1,2/3], ~~ T_{R}
\equiv [3,1,2/3], \nonumber \\[.15in]
\psi_{aL} &=& \begin{pmatrix} e_{aL} \\ \nu_{aL} \\ N_{aL} \end{pmatrix} \equiv [ 1,{3^\ast},-1/3], \quad
e_{aR} \equiv [1,1,-1].
\end{eqnarray}
The generation index $i=1,2$ corresponds to the first two generations
with the quarks $u_{L,R},d_{L,R},D_{L,R}$ and
$c_{L,R},s_{L,R},S_{L,R}$. For the leptons, the generation index is
$a=1,2,3$.
There are several variants of the model that allow slightly different
choices of fermions as well as their baryon and lepton number
assignments. Here we shall restrict ourselves to the one which
contains only the quarks with electric charge $2/3$ and $1/3$ and no
lepton number ($L$). In this scenario all quarks (usual ones and the
exotic ones) carry baryon number ($B=1/3$) and no lepton number
($L=0$), while all leptons carry lepton number ($L=1$) and no baryon
number ($B=0$). Notice that in Ref. \cite{Boucenna:2014dia} the lepton
number is defined as $L=4/\sqrt{3}~ T_8+ {\cal L}} \def\B{{\cal B}$, where $U(1)_{{\cal L}} \def\B{{\cal B}}$ is a
global symmetry and a $Z_{2}$ symmetry is introduced to forbid a
coupling like $ \psi_L \psi_L \phi_0$, in connection with neutrino
masses. Since the charge equation given in Eq.~(\ref{2.2}) remains the
same for this assignment, the following discussion regarding
Renormalization Group Equations (RGE) in the SVS model remains valid
for this assignment as well.
For the symmetry breaking and the charged fermion masses, the
following Higgs scalars and their vacuum expectation values (vevs) are
assumed,
$$
\phi_0 \equiv [1,3^\ast, 2/3] ~~~ {\rm and} ~~~ \phi_{1,2}
\equiv [1,3^\ast,-1/3],
$$
\begin{equation}{\label{2.4}}
\langle \phi_0 \rangle = \begin{pmatrix} k_0 \\ 0 \\ 0 \end{pmatrix} , ~~~~
\langle \phi_{1} \rangle = \begin{pmatrix} 0 \\ k_1 \\ n_1\end{pmatrix} , ~~~~
\langle \phi_2 \rangle = \begin{pmatrix} 0 \\ k_2 \\ n_2\end{pmatrix} \,.
\end{equation}
Here we assume $k_{0,1,2} \sim m_W$ to be of the order of the
electroweak symmetry breaking scale and $n_{1,2} \sim M_{331}$ to be
the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale. We shall not discuss here the details of
fermion masses and mixing, which can be found in Refs. \cite{Boucenna:2014ela,Boucenna:2014dia}.
\section{The sequential $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ Model}{\label{sec3}}
In this model the fields are assigned in a way such that the anomalies
are cancelled for each generation separately. The multiplet structure
is given by
\begin{eqnarray}{\label{3.1}}
Q_{aL} &=& \begin{pmatrix} u_{aL} \\ d_{aL} \\ D_{aL} \end{pmatrix} \equiv
[3,3,0], ~
u_{aR} \equiv [3,1,2/3], ~ d_{aR} \equiv [3,1,-1/3], ~
D_{aR}
\equiv [3,1,-1/3], \nonumber \\[.15in]
\psi_{aL} &=& \begin{pmatrix} e^{-}_{aL} \\ \nu_{aL} \\ N^{1}_{aL} \end{pmatrix} \equiv [ 1,{3^\ast},-1/3], ~
\xi_{aL} = \begin{pmatrix} E^{-}_{aL} \\ N^{2}_{aL} \\ N^{3}_{aL} \end{pmatrix} \equiv [ 1,{3^\ast},-1/3], ~
\chi_{aL} = \begin{pmatrix} N^{4}_{aL} \\ E^{+}_{aL} \\ e^{+}_{aL} \end{pmatrix} \equiv [ 1,{3^\ast},2/3].\nonumber\\
\end{eqnarray}
It is straightforward to check that each family is anomaly free. In
order to drive symmetry breaking and generate the charged fermion
masses, we assume a Higgs sector and vevs similar to the SVS 331 model
\footnote{A model with similar fermion content and with
$k_{1}=n_{2}=0$ in the scalar sector was discussed in
Ref. \cite{Sanchez:2001ua} using the trinification group
$\mathrm{SU(3)_c \times SU(3)_L \times SU(3)_R}$.}.
The Yukawa Lagrangian for the quark sector can be written as
\begin{eqnarray}{\label{3.2}}
\mathcal{L}_{\rm{quarks}}=y_{u_a} \overline{Q_{aL}} u_{aR}\phi_{0}^{\ast}+y_{d_a}^{i} \overline{Q_{aL}} d_{aR}\phi_{i}^{\ast}+y_{D_a}^{i} \overline{Q_{aL}} D_{aR}\phi_{i}^{\ast} + {\rm{h.c.}} \quad,
\end{eqnarray}
with $i=1,2$ and where we neglect any flavour mixing. After the chain of spontaneous symmetry breaking the
up-type quarks obtain a mass term
\[
m_{u_a}=y_{u_a} k_{0},
\]
while the down--type and vectorlike down--type quarks form a mass matrix
in the $(d,D)$ basis given by
\begin{eqnarray}{\label{3.3}}
m_{dD}^{a}=\begin{pmatrix} y_{d_a}^{1} k_1 +y_{d_a}^{2} k_2 & y_{D_a}^{1} k_1 +y_{D_a}^{2} k_2\\
y_{d_a}^{1} n_1 +y_{d_a}^{2} n_2 & y_{D_a}^{1} n_1 +y_{D_a}^{2} n_2\end{pmatrix} .
\end{eqnarray}
Note that in the cases
$y_{d_a}^{1}=y_{d_a}^{2}=y_{D_a}^{1}=y_{D_a}^{2}\equiv y_d$; or
$k_1=k_2=k$ and $n_1=n_2=n$ the determinant of the above Yukawa
matrix vanishes giving $m_{d_a}=0$ and
$m_{D_a}=y_{d_a}^{1} k_1 +y_{d_a}^{2} k_2 +y_{D_a}^{1} n_1
+y_{D_a}^{2} n_2$. However, in the absence of any symmetries forcing
the above conditions, the down quarks obtain mass as a result of the
mixing with the vector--like quarks. One can determine it
perturbatively by expanding the Yukawa contributions in terms of
$k_i/n_i \ll 1$ so as to obtain
$$m_{d_a}=\left(y_{d_a}^{1} k_1 +y_{d_a}^{2} k_2 \right) -\left( y_{D_a}^{1} k_1 +y_{D_a}^{2} k_2 \right) \frac{y_{d_a}^{1} n_1 +y_{d_a}^{2} n_2}{y_{D_a}^{1} n_1 +y_{D_a}^{2} n_2}+\cdots ,$$
$$m_{D_a}=\left(y_{d_a}^{1} k_1 +y_{d_a}^{2} k_2 \right)+\left(y_{D_a}^{1} n_1 +y_{D_a}^{2} n_2\right)-\left| M_{dD}^{a} \right|/ \left(y_{D_a}^{1} n_1 +y_{D_a}^{2} n_2\right) +\cdots ,$$
where
$$\left| M_{dD}^{a} \right|=\left(y_{d_a}^{1} k_1 +y_{d_a}^{2} k_2 \right)\left(y_{D_a}^{1} n_1 +y_{D_a}^{2} n_2\right)-\left( y_{D_a}^{1} k_1 +y_{D_a}^{2} k_2 \right)\left( y_{d_a}^{1} n_1 +y_{d_a}^{2} n_2 \right).$$
This structure can be used to account for the SM quark masses and CKM
mixing, as well as the heavier vector--like quark mass limits from
the LHC.
Turning now to the lepton sector, the relevant Yukawa terms are given by
\begin{eqnarray}{\label{3.4}}
\mathcal{L}_{\rm{leptons}}=\epsilon_{\alpha\beta\gamma}\left[ \psi_{\alpha L}^{T} C^{-1}\left(y_{1} \xi_{\beta L}\phi_{0\gamma}+y_{2}^{i}\chi_{\beta L}\phi_{i \gamma}\right)+\xi_{\alpha L}^{T} C^{-1} y_{3}^{i} \chi_{\beta L}\phi_{i \gamma} \right] + {\rm{h.c.}} \quad,
\end{eqnarray}
where $\alpha, \beta, \gamma$ are the $\mathrm{SU(3)_{L}}$ tensor
indices ensuring antisymmetric Dirac mass terms, $C$ is the charge
conjugation matrix, and $i=1,2$. After the symmetry breaking, these
Yukawa terms give rise to the mass matrices for charged and neutral
leptons. In the basis $(e,E)$ the mass matrix is given by
\begin{eqnarray}{\label{3.5}}
m_{eE}=\begin{pmatrix} -\left(y_{2}^{1} k_1 +y_{2}^{2} k_2\right) & \left(y_{2}^{1} n_1 +y_{2}^{2} n_2\right)\\
-\left(y_{3}^{1} k_1 +y_{3}^{2} k_2 \right)& \left( y_{3}^{1} n_1 +y_{3}^{2} n_2 \right)\end{pmatrix} ,
\end{eqnarray}
with the eigenvalues given by
$$m_{e}=-\left(y_{2}^{1} k_1 +y_{2}^{2} k_2\right) +\left(y_{3}^{1} k_1 +y_{3}^{2} k_2 \right) \frac{y_{2}^{1} n_1 +y_{2}^{2} n_2}{y_{3}^{1} n_1 +y_{3}^{2} n_2}+\cdots ,$$
$$m_{E}=\left( y_{3}^{1} n_1 +y_{3}^{2} n_2 \right)-\left(y_{2}^{1} k_1 +y_{2}^{2} k_2\right)-\left| M_{eE}^{a} \right|/ \left( y_{3}^{1} n_1 +y_{3}^{2} n_2 \right) +\cdots ,$$
where
$$\left| M^{a}_{eE}\right|=\left(y_{2}^{1} n_1 +y_{2}^{2}
n_2\right)\left(y_{3}^{1} k_1 +y_{3}^{2} k_2 \right)-\left(y_{2}^{1}
k_1 +y_{2}^{2} k_2\right)\left( y_{3}^{1} n_1 +y_{3}^{2} n_2
\right).$$
For the case of neutral leptons the mass matrix can be written as:
\begin{eqnarray}{\label{3.6}}
m_{\nu N}=\begin{pmatrix}
0 & 0 & y_1 k_0 & 0 & -\left(y_{2}^{1} n_1 +y_{2}^{2} n_2\right)\\
0 & 0 & 0 & -y_1 k_0 & \left(y_{2}^{1} k_1 +y_{2}^{2} k_2\right) \\
y_1 k_0 & 0 & 0 & 0 & \left(y_{3}^{1} k_1 +y_{3}^{2} k_2 \right) \\
0 & -y_1 k_0 & 0 & 0 & -\left(y_{3}^{1} n_1 +y_{3}^{2} n_2 \right) \\
-\left(y_{2}^{1} n_1 +y_{2}^{2} n_2\right) & \left(y_{2}^{1} k_1 +y_{2}^{2} k_2\right) & \left(y_{3}^{1} k_1 +y_{3}^{2} k_2 \right) &-\left(y_{3}^{1} n_1 +y_{3}^{2} n_2 \right) & 0
\end{pmatrix} ,\nonumber\\
\end{eqnarray}
in the basis $(\nu, N_1, N_3, N_2, N_4)$, where $N_1, N_3$ are
SU(2)$_{\rm L}$ isosinglets and $\nu, N_2, N_4$ are components of doublets. Next,
we rotate the above mass matrix by an orthogonal transformation
$m_{\nu N}^{\prime}=R^{T} m_{\nu N} R$, where
\begin{eqnarray}{\label{3.6.1}}
R=\begin{pmatrix}
0 &- \frac{1}{2} & \frac{1}{2} & \frac{1}{2} & \frac{1}{2}\\
- \frac{1}{\sqrt{2}} & \frac{1}{2} &\frac{1}{2} & 0 & 0 \\
\frac{1}{\sqrt{2}} & \frac{1}{2} & \frac{1}{2} & 0 & 0 \\
0 & \frac{1}{2} & -\frac{1}{2}& \frac{1}{2} & \frac{1}{2} \\
0 & 0 & 0 & \frac{1}{\sqrt{2}} & -\frac{1}{\sqrt{2}}
\end{pmatrix} ,\nonumber\\
\end{eqnarray}
This yields the rotated mass matrix $m^{\prime}_{\nu N}$ given by
\begin{eqnarray}{\label{3.7}}
m^{\prime}_{\nu N}=\begin{pmatrix}
0 & 0 & 0 & \frac{u}{\sqrt{2}}-\frac{x-z} {2}& \frac{u}{\sqrt{2}}+\frac{x-z} {2}\\
0 & -u & 0 & \frac{(X-Z)+(x+z)} {2\sqrt{2}} & -\frac{(X-Z)+(x+z)} {2\sqrt{2}} \\
0 & 0 & u & -\frac{(X-Z)-(x+z)} {2\sqrt{2}} & \frac{(X-Z)-(x+z)} {2\sqrt{2}} \\
\frac{u}{\sqrt{2}}-\frac{x-z} {2} & \frac{(X-Z)+(x+z)} {2\sqrt{2}} & -\frac{(X-Z)-(x+z)} {2\sqrt{2}} & -\frac{X+Z}{\sqrt{2}} & 0 \\
\frac{u}{\sqrt{2}}+\frac{x-z} {2} & -\frac{(X-Z)+(x+z)} {2\sqrt{2}} & \frac{(X-Z)-(x+z)} {2\sqrt{2}} & 0 & \frac{X+Z}{\sqrt{2}}
\end{pmatrix} ,\nonumber\\
\end{eqnarray}
where
$$u=y_1 k_0,~ X= \left(y_{2}^{1} n_1 +y_{2}^{2} n_2\right),
~ x= \left(y_{2}^{1} k_1 +y_{2}^{2} k_2\right), ~
z=\left(y_{3}^{1} k_1 +y_{3}^{2} k_2 \right), ~ Z=\left(y_{3}^{1}
n_1 +y_{3}^{2} n_2 \right). $$
Now we recall that
$k_{0,1,2} \sim m_W$ is of the order of the electroweak symmetry
breaking scale, while and $n_{1,2} \sim M_{331}$ is of the order of
the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale, and hence one expects that
$X,Z\gg u, x, z$. If we further assume $X+Z\gg X-Z$, then we can
identify the 44 and 55 entries as the heaviest in the mass matrix
given in Eq. (\ref{3.7}) and these rotated isodoublet states form a
pair of heavy quasi Dirac neutrinos with mass of the order of the
$\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale. We can now readily use perturbation
theory to obtain the masses for the three remaining lighter
states. Up to second order in perturbation theory we obtain two Dirac
states with mass of the order of the electroweak symmetry breaking
scale $\pm u=\pm y_1 k_0$ and a light seesaw Majorana neutrino with
mass $2u(z-x)/(X+Z)$. With this we see that the model has enough
flexibility to account for the observed pattern of fermion masses.
It is not our purpose here to present a detailed study of the
structure of the fermion mass spectrum, but only to check its
consistency in broad terms.
\section{Renormalization Group Equations and
Gauge Coupling Unification} {\label{sec4}}
In this section we study the SVS model RGEs to explore if unification of the three gauge
couplings~\cite{PhysRevLett.33.451} can be obtained in the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ theory
at a certain scale $M_U$, without any presumptions about the nature
of the underlying group of grand
unification~\cite{Boucenna:2014dia}. Using the RGEs we express the
hypercharge (and X) normalization and the unification scale as a
function of $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ breaking scale. Next we study the allowed range of
$\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ breaking scale such that one can obtain a guaranteed unification
of the gauge couplings. First we discuss the SVS model discussed in
section \ref{sec2}. Then, we study the impact of adding three
generations of leptonic octet representations $[1,8,0]$ that can give
gauge coupling unification for a TeV scale $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ breaking while driving
an interesting radiative model for neutrino mass
generation~\cite{Boucenna:2014dia}.
The evolution for running coupling constants at one loop level is
governed by the RGEs
\begin{equation}{\label{4.1}}
\mu\,\frac{\partial g_{i}}{\partial \mu}=\frac{b_i}{16 \pi^2} g^{3}_{i},
\end{equation}
which can be written in the form
\begin{equation}{\label{4.2}}
\frac{1}{\alpha_{i}(\mu_{2})}=\frac{1}{\alpha_{i}(\mu_{1})}-\frac{b_{i}}{2\pi} \ln \left( \frac{\mu_2}{\mu_1}\right),
\end{equation}
where $\alpha_{i}=g_{i}^{2}/4\pi$ is the fine structure constant for
$i$--th gauge group, $\mu_1, \mu_2$ are the energy scales with
$\mu_2 > \mu_1$. The beta-coefficients $b_i$ determining the evolution
of gauge couplings at one-loop order are given by
\begin{eqnarray}{\label{4.3}}
&&b_i= - \frac{11}{3} \mathcal{C}_{2}(G)
+ \frac{2}{3} \,\sum_{R_f} T(R_f) \prod_{j \neq i} d_j(R_f)
+ \frac{1}{3} \sum_{R_s} T(R_s) \prod_{j \neq i} d_j(R_s).
\label{oneloop_bi}
\end{eqnarray}
Here, $\mathcal{C}_2(G)$ is the quadratic Casimir operator for the
gauge bosons in their adjoint representation,
\begin{equation}{\label{4.4}}
\mathcal{C}_2(G) \equiv \left\{
\begin{matrix}
N & \text{if } SU(N), \\
0 & \text{if } U(1).
\end{matrix}\right.
\end{equation}
On the other hand, $T(R_f)$ and $T(R_s)$ are the Dynkin indices of the
irreducible representation $R_{f,s}$ for a given fermion and scalar,
respectively,
\begin{equation}{\label{4.5}}
T(R_{f,s}) \equiv \left\{
\begin{matrix}
1/2 & \text{if } R_{f,s} \text{ is fundamental}, \\
N & \text{if } R_{f,s} \text{ is adjoint}, \\
0 & \text{if } R_{f,s} \text{ is singlet},
\end{matrix}\right.
\end{equation}
and $d(R_{f,s})$ is the dimension of a given representation $R_{f,s}$
under all gauge groups except the $i$-th~gauge group under
consideration. An additional factor of $1/2$ is multiplied in the case
of a real Higgs representation.
The electromagnetic charge operator is given by
\begin{equation}{\label{4.6}}
Q = T_3 + Y = T_3 +\frac{1}{\sqrt{3}} T_8 + X,
\end{equation}
where the generators (Gell-Mann matrices) are normalized as
$\text{Tr} (T_{i}T_{j})=\frac{1}{2}\delta_{ij}$. We define the
normalized hypercharge operator $Y_{N}$ and $X_{N}$ as
\begin{equation}{\label{4.7}}
Y=n_{Y} Y_{N},\quad X=n_{X}X_{N},
\end{equation}
such that we have
\begin{equation}{\label{4.8}}
n_{Y}^{2}=\frac{1}{3}+n_{X}^{2},
\end{equation}
and the normalized couplings are related by
\begin{equation}{\label{4.9}}
n_{Y}^{2} {\left(\alpha^{N}_{Y}\right)}^{-1}=\frac{1}{3}\alpha_{3L}^{-1}+\left(n_{Y}^{2}-\frac{1}{3}\right){\left(\alpha^{N}_{X}\right)}^{-1},
\end{equation}
where
\begin{equation}{\label{4.10}}
\alpha^{N}_{Y} =n_{Y}^{2} \alpha_{Y}, \quad \alpha^{N}_{X}=\left(n_{Y}^{2}-\frac{1}{3}\right)\alpha_{X}, \quad \alpha_{3L}=\alpha_{2L}.
\end{equation}
Now using Eqs. (\ref{4.1}, \ref{4.9}, \ref{4.10}) we obtain
\begin{eqnarray}{\label{4.11}}
\alpha^{-1}_{U} &=& \frac{1}{n_{Y}^{2}-\frac{1}{3}}
\left\{
\alpha_{\text{em}}^{-1}(M_{Z})\cos^{2}\theta_{w} (M_{Z})
-\frac{1}{3}\alpha^{-1}_{2L}(M_Z)
-\frac{b^{\text{UN}}_{Y}-\frac{1}{3}b_{2L}}{2\pi}
\ln\left(\frac{M_X}{M_Z}\right)
-\frac{b^{\text{UN}}_{X}}{2\pi}
\ln\left(\frac{M_U}{M_X}\right)
\right\},\nonumber\\
\alpha^{-1}_{U}&=&\alpha^{-1}_{2L}(M_Z)-\frac{b_{2L}}{2\pi}\ln\left(\frac{M_X}{M_Z}\right)-\frac{b_{3L}}{2\pi}\ln\left(\frac{M_U}{M_X}\right),\nonumber\\
\alpha^{-1}_{U}&=&\alpha^{-1}_{3C}(M_Z)-\frac{b_{3C}}{2\pi}\ln\left(\frac{M_X}{M_Z}\right)-\frac{b^{X}_{3C}}{2\pi}\ln\left(\frac{M_U}{M_X}\right).
\end{eqnarray}
Here, the SM running is described by the the SU(3)$_C$ coefficient $b_{3C}$, the SU(2)$_L$ coefficient $b_{2L}$ and the U(1)$_Y$ unnormalized coefficient $b^{\text{UN}}_{Y}$. Likewise, in the unbroken $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ phase, the gauge running coefficients for the SU(3)$_C$, SU(3)$_L$ and unnormalized U(1)$_X$ components are $b^{X}_{3C}$, $b_{3L}$ and $b^{\text{UN}}_{X}$, respectively. The scale $M_{Z}$ corresponds to the $Z$
boson-pole, the 331 symmetry breaking scale is denoted by $M_{X}$ and
$M_{U}$ is the scale of unification for the normalized gauge
couplings. From the above set of equations the unification scale $M_U$
can be obtained as a function of $M_{X}$,
\begin{equation}{\label{4.12}}
M_U=M_{X}\left(\frac{M_X}{M_Z}\right)^{-\frac{b_{3C}-b_{2L}}{b^{X}_{3C}-b_{3L}}} \exp \left[2\pi \frac{\alpha^{-1}_{3C}(M_Z)-\alpha^{-1}_{2L}(M_Z)}{b^{X}_{3C}-b_{3L}}\right].
\end{equation}
Similarly, $n_{Y}^{2}$ can be expressed as a function of $M_{X}$,
\begin{eqnarray}{\label{4.13}}
n_{Y}^{2}&=&\frac{1}{3}+\left[\alpha_{\text{em}}^{-1}(M_{Z})\cos^{2}\theta_{w} (M_{Z})-\frac{1}{3}\alpha^{-1}_{2L}(M_Z)-\frac{b^{\text{UN}}_{Y}-\frac{1}{3}b_{2L}}{2\pi}\ln\left(\frac{M_X}{M_Z}\right)\right.\nonumber\\
&+&\left. b^{\text{UN}}_{X}\left\{ \frac{1}{2\pi}\frac{b_{3C}-b_{2L}}{b^{X}_{3C}-b_{3L}}\ln\left(\frac{M_X}{M_Z}\right)- \frac{\alpha^{-1}_{3C}(M_Z)-\alpha^{-1}_{2L}(M_Z)}{b^{X}_{3C}-b_{3L}}\right\}\right]\nonumber\\
&\times& \left[\alpha^{-1}_{2L}(M_Z)-\frac{b_{2L}}{2\pi}\ln\left(\frac{M_X}{M_Z}\right)+b_{3L}\left\{\frac{1}{2\pi}\frac{b_{3C}-b_{2L}}{b^{X}_{3C}-b_{3L}}\ln\left(\frac{M_X}{M_Z}\right)-\frac{\alpha^{-1}_{3C}(M_Z)-\alpha^{-1}_{2L}(M_Z)}{b^{X}_{3C}-b_{3L}}\right\} \right]^{-1}.\nonumber\\
\end{eqnarray}
The above two relations are valid provided $b^{X}_{3C}\neq b_{3L}$ and
$(b^{X}_{3C}-b_{3L})\neq (b_{3C}-b_{2L})$, which are satisfied in the
cases that we shall discuss below. Furthermore, we take $M_{X}\leq
M_{U}\leq 10^{17}$~GeV and assume that 331 is the only gauge group (in
other words $M_{X}$ is the only intermediate scale) between $M_{Z}$
and the unification scale $M_U$.
\subsection{The minimal SVS Model}
The first case of interest is the minimal scenario described in
section \ref{sec2}. The relevant gauge quantum numbers are given in
Eqs. (\ref{2.3},\ref{2.4}). The Higgs sector involves three
$\mathrm{SU(3)_{L}}$ triplets, namely the minimal set necessary for
adequate symmetry breaking and generation of fermion masses. First we
notice that the model described in Ref. \cite{Boucenna:2014dia} has
the same RGE evolution, since the extra gauge singlets added to the
fermion spectrum to generate neutrino masses do not enter the
RGEs. For the SM the one-loop beta-coefficients are given by
$b_{2L} = -19/6$, $b^{\text{UN}}_{Y} = 41/6$, $b_{3C} = -7$, while in
the $SU(3)_c \times SU(3)_L \times U(1)_X$ phase they are given by
$b_{3L} = -13/2$, $b^{\text{UN}}_{X} = 26/3$, $b^{X}_{3C} = -5$.
\begin{figure}[t!]
\includegraphics[width=0.49\linewidth]{p4.pdf}
\includegraphics[width=0.47\linewidth]{p5.pdf}
\caption{(Left) Allowed range for $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale
$M_{X}$ for guaranteed unification. The solid line represents
$M_{U}$ as a function of $M_{X}$. The dashed and dot-dashed lines
correspond to $M_{U}=10^{17}$~GeV and $M_{U}=M_{X}$
respectively. (Right) The hypercharge normalization factor
$n_{Y}^{2}$ as a function of the 331 symmetry breaking scale
$M_{X}$. The shaded region represents the allowed region $n_{Y}^{2}\geq
\frac{1}{3}$ with the dashed line corresponding to the lower limit
$n_{Y}^{2}= \frac{1}{3}$. The solid line gives $n_{Y}^{2}$ as a
function of $M_X$ and the red dot-dashed line shows the standard $SU(5)$ normalization $n_Y^2=
\frac{5}{3}$.}
\label{fig1}
\end{figure}
In Fig.~\ref{fig1} (left) we plot the allowed range for $M_X$. The intersection of the line corresponding to $M_U$
evaluated as a function of $M_{X}$ in Eq. (\ref{4.12}) with the lines
for $M_{U}=M_{X}$ and $M_{U}=10^{17}$~GeV gives the lower and upper
bound on $M_{X}$ respectively such that there is a guaranteed
unification. In this scenario, the scale $M_X$ of $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ breaking is therefore always high and very close to the unification scale $M_U$.
Next, in Fig. \ref{fig1} (right) we plot the hypercharge normalization
factor $n_{Y}^{2}$ as a function of $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale
$M_{X}$. The dashed horizontal line represents the lower limit
$n_{Y}^{2}= \frac{1}{3}$ of the allowed value for $n_{Y}^{2}$. As can
be seen from the figure, for the allowed $M_X$ range from the
condition $M_{X}\leq M_{U}\leq 10^{17}$~GeV the hypercharge
normalization $n_{Y}^{2}$ is almost constant $\approx 1.3$ and well above the allowed lower limit.
Finally, in Fig.~\ref{fig3} we give an example of gauge coupling
running with respect to the 331 symmetry breaking scale
$M_{X}=7.9\times 10^{16}$~GeV. It demonstrates that successful gauge
unification at the scale $M_{U} = 8.05 \times 10^{16}$~GeV with
$n_{Y}^{2}=1.3$ can be achieved, albeit this requires a very high scale of $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ breaking very near to the unification scale.
\begin{figure}[t!]
\includegraphics[width=\linewidth]{p6.pdf}
\caption{Gauge coupling running in the SVS Model with 331 symmetry
breaking scale $M_{X}=7.9\times 10^{16}$~GeV, demonstrating
successful gauge unification at the scale
$M_{U} = 8.05 \times10^{16}$~GeV with $n_{Y}^{2}=1.3$ . The right
plot shows the magnified view of gauge coupling running around
$M_{X}$.}
\label{fig3}
\end{figure}
\subsection{The SVS Model with fermionic octets}
In this model, in addition to the field content of model I, we include
three generations of fermion octets $\Omega$ with the assignments
under the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ group given by
\begin{equation}{\label{4.15}}
\Omega\equiv [1, 8^{\ast}, 0].
\end{equation}
The Higgs sector involves the same three $\mathrm{SU(3)_{L}}$
triplets as before. Although this model has the same content as the one
considered in Ref.~\cite{Boucenna:2014dia}, here we take a completely
different approach to unification. Indeed, we do not consider the
usual SU(5) normalization for the hypercharge and the octet mass scale
is the same as the 331 symmetry breaking scale. In this model, the
neutrinos are massless at tree level, however at one-loop level the
exchange of gauge bosons give rise to dimension-nine operator which
generates neutrino masses after 331 symmetry
breaking~\cite{Boucenna:2014dia}. For the SM the one-loop
beta-coefficients remain the same as Model I, while in the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ phase
they are given by $b_{3L} = -1/2$, $b^{\text{UN}}_{X} = 26/3$,
$b^{X}_{3C} = -5$.
In Fig. \ref{fig7} (left) we plot the allowed range for $M_X$ for
which unification is guaranteed at a scale
$M_{X}\leq M_{U}\leq 10^{17}$GeV. Interestingly, in this model we find
that for a $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale $M_{X}$ as low as TeV it is
possible to achieve unification. Note that in contrast to
Ref. \cite{Boucenna:2014dia}, here we do not assume another
intermediate scale corresponding to the fermion octet mass scale in
addition to $M_{X}$. Formally, unification can thus be achieved for
any scale $M_X$ between $M_Z$ and $M_U$, however,
$M_{U}\lesssim 10^{15.5}$~GeV is disfavored by the current
experimental limits on the lifetime of the proton decay
\cite{Agashe:2014kda}. This consequently puts a lower limit of
the order of $M_X \gtrsim 10^5$~GeV on the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ breaking scale,
although we should emphasize that we here do not specify the GUT
group and thus cannot predict the proton decay rate accurately.
\begin{figure}[t!]
\includegraphics[width=0.49\linewidth]{p1.pdf}
\includegraphics[width=0.49\linewidth]{p2.pdf}
\caption{Same as Fig.~\ref{fig1}, but for the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ SVS model with three additional fermionic octets. The purple dotted line corresponds to the lower limit of $M_U$ allowed by the current experimental limits on the lifetime of the proton decay.}
\label{fig7}
\end{figure}
In Fig. \ref{fig7} (right) we plot the hypercharge normalization
factor $n_{Y}^{2}$ as a function of $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale
$M_{X}$. In this case as well, for the allowed $M_X$ range from the
condition $M_{X}\leq M_{U}\leq 10^{17}$~GeV the hypercharge
normalization $n_{Y}^{2}$ is well above its allowed lower limit.
In Fig.~\ref{fig9} we show an example gauge coupling running with $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$
symmetry breaking scale $M_{X}= 3000$~GeV, demonstrating successful
gauge coupling unification at a scale $M_{U} = 10^{14.9}$~GeV with
$n_{Y}^{2}=1.8$. Thus, from the perspective of a low $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry
breaking scale within the reach of accelerator experiments like the
LHC $\sim{\mathcal{O}}$(TeV)) this model is the most interesting
candidate leading to a successful gauge coupling unification. In
addition to the new gauge bosons, the model can harbor a plethora of
new states associated to the new exotic fermions as well as extra
Higgs bosons.
\begin{figure}[t!]
\includegraphics[width=0.7\linewidth]{p3.pdf}
\caption{Gauge coupling running in the SVS Model adding three
generations of leptonic octets with $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale at
$M_{X}= 3000$~GeV, demonstrating successful gauge unification at the
scale $M_{U} = 10^{14.9}$~GeV with $n_{Y}^{2}=1.8$.}
\label{fig9}
\end{figure}
\section{SU(6) Grand Unification}{\label{sec5}}
We consider the possibility of grand unification of the
$\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model in an SU(6) gauge unification group. Two possibilities
arise. Unlike other conventional grand unified theories, in SU(6) one
can have different components of the 331 subgroup with different
multiplicity. Such a scenario may emerge from the flux breaking of the
unified group in an E(6) F-theory GUT. This provides new ways of
achieving gauge coupling unification in 331 models.
Alternatively, a sequential variant of $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model can have a minimal
SU(6) grand unification, which in turn can be a natural E(6)
subgroup. This minimal SU(6) embedding does not require any bulk exotics to
account for the chiral families and allows for a TeV scale $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$
model.
We now demonstrate how the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model fermions can be embedded
in an SU(6) grand unified gauge group. Our main consideration is to
explore whether the combinations of the SU(6) gauge group
representations form an anomaly free set, which can contain all the
required fermions. In the subsequent subsections we discuss how
different multiplicities of the SVS version of the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model can be
explained when this SU(6) grand unified model is embedded in an E(6)
F-theory and how the sequential $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model can be embedded in a
minimal anomaly free combination of representations of SU(6) as an
E(6) subgroup. For the minimal SVS version of the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model, gauge
coupling unification can be obtained by including both the matter
multiplets
in the 27-dimensional fundamental representations of E(6) as
well as the bulk exotics from the 78-dimensional adjoint
representations of E(6). In particular the octet of
$\mathrm{SU(3)_L}$ coming from the bulk plays a crucial role in
allowing the unification of the gauge couplings with a low 331
symmetry breaking scale. On the other hand, the embedding of
sequential 331 model in SU(6) does not require any bulk exotics to
account for the chiral multiplets and imply, by adding three
generations of $\mathrm{U(1)_X}$ neutral fermionic octets, one can
obtain SU(6) unification with a TeV scale $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ breaking scale.
We shall first write down some of the product decompositions of the
group SU(6):
\begin{eqnarray}{\label{5.1}}
6 \times 6 &=& 15_a + 21_s, \nonumber \\
6 \times \bar{6} &=& 1 + 35, \nonumber \\
6 \times 15 &=& 20 + 70, \nonumber \\
6 \times 21 &=& 56 + 70.
\end{eqnarray}
The SU(6) has $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ as a maximal subgroup with the same rank. For
convenience we write down some of the representations of SU(6) under
this maximal subgroup $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$:
\begin{eqnarray}{\label{5.2}}
6 &=& [3,~1,~-1/3] ~+ ~[ 1,~3,~1/3], \nonumber \\
15 &=& [\bar{3},~ 1,~-2/3] ~+~ [1,~\bar{3},~ 2/3] ~+~
[3,~3,~0], \nonumber\\
20 &=& [1,~1,~-1] ~+~ [1,~1,~1] ~+~ [3,~\bar{3},~1/3]~+~ [\bar{3},~3,~-1/3],
\nonumber\\
21 &=& [3,~3,~0] ~+~ [6,~ 1,~-2/3] ~+~ [1,~6,~ 2/3], \nonumber \\
35 &=& [1,~1,~0] ~+~ [8,~1,~0] ~+~ [1,~8,~0] ~+~
[3,~\bar{3},~-2/3]~+~ [\bar{3},~3,~2/3],
\nonumber \\
56 &=& [10,~1,~-1] ~+~ [1,~10,~1] ~+~ [6,~ 3,~-1/3] ~+~ [3,~6,~ 1/3],
\nonumber \\
70 &=& [6, 3,-1/3] + [3,6, -1/3] + [3,\bar{3},1/3]+
[\bar{3},3,-1/3] + [8,1,-1] + [1,8,1].
\end{eqnarray}
The anomaly for the various representations of the group SU(6) are
\begin{eqnarray}{\label{5.3}}
{\cal A}[6] = 1, ~ {\cal A}[15] = 2, ~ {\cal A}[20] = 0, ~
{\cal A}[21] = 10, ~ {\cal A}[35] = 0, ~ {\cal A}[56] = 54, ~
{\cal A}[70] = 27.
\end{eqnarray}
We now turn to two concrete model constructions.
\subsection{SU(6) Grand Unification of the SVS Model}
It can be easily verified that all fermions of the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model proposed
by SVS (discussed in section \ref{sec2}) can be included in the
anomaly free combination of representations under SU(6):
$$ \bar{6} ~+~ \bar{6} ~+~ 15 ~+~ 20.$$
There will be some extra fermions and the multiplicity of the
different representations are now different. It is to be noted that
these states can be naturally embedded in an E(6) theory. We start
with the maximal $\mathrm{SU(2) \times SU(6)}$ subgroup of E(6), and
write down the decomposition:
\begin{eqnarray}{\label{5.4}}
27 & = & [2,~\bar{6}] ~+~ [1,~ 15] \nonumber \\[.1in]
78 &=& [1, ~35] ~+~ [2,~20] ~+~ [3,~1] \nonumber
\end{eqnarray}
Thus the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ anomaly free representations of the SVS model can all be
embedded in a combination of anomaly free representations of SU(6),
which in turn can be embedded in the fundamental and adjoint
representations of the group E(6). The next question is how to match
the multiplicity of the different representations of the SVS 331
model, which is nontrivial. At this stage we resort to the symmetry
breaking at the GUT scale induced by flux breaking through the
Hosotani mechanism~\cite{HOSOTANI1989233}. Assigning particular
geometry to the flux breaking, we identify the different states with
the different algebraic varieties, and then the intersection numbers
would give us the multiplicities of the different representations. A
detailed study of such E(6) F-theory
GUTs~\cite{1126-6708-2009-01-059,King:2010mq,Callaghan:2011jj,Callaghan:2013kaa}
is beyond the scope of this article and we shall rather take a
phenomenological approach to the problem. We consider the required
representations to match the low energy phenomenological
requirements. The first step is to keep the known fermions light and
also to have $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale as low as TeV, while at the
same time requiring for gauge coupling unification.
\begin{figure}[t!]
\includegraphics[width=0.7\linewidth]{svs_su6_unif.pdf}
\caption{Gauge coupling running in the SVS Model with three generations of
leptonic octets with 331 symmetry breaking scale $M_{X}= 3000$~GeV
and octet mass scale $M_{8}= 9000$~GeV, demonstrating successful
gauge unification at the scale $M_{U} = 10^{15.5}$~GeV with
$n_{Y}=\sqrt{5/3}$ and $n_X=2/\sqrt{3}$.}
\label{svs_unif}
\end{figure}
Considering the $\bar{6}$ representation of SU(6), which contains the
down antiquarks $d^{c}_{L}$ with hypercharge $Y=1/3$, isospin lepton
doublet containing $e_{L}$ and $\nu_{L}$ with $Y=-1/2$, and $N_{L}$
with $Y=0$; we can get the normalization for the hypercharge from
$Tr(Y^2)=5/6 n^{-2}_{Y}$ in the notation of Eq. (\ref{4.7}). The
$\mathrm{U(1)_Y}$ normalization defined in Eq. (\ref{4.7}) is given by
$n_Y=\sqrt{5/3}$ and using Eq. (\ref{4.8}) we obtain the
$\mathrm{U(1)_X}$ normalization given by $n_X=2/\sqrt{3}$, which is
below the normalizations required for a guaranteed unification in
Fig. \ref{fig1} and Fig. \ref{fig7}. However, it is still possible to
obtain gauge coupling unification following the prescription in
Ref. \cite{Boucenna:2014dia}, where the octet scale is decoupled from
the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale and is assumed to lie between the
$\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale and unification scale. Here, the octets
belong to 35 of SU(6), which belongs to the bulk exotics coming from
the 78-dimensional adjoint representations of E(6).
Using Eq.~(\ref{4.3}) the the one-loop beta-coefficients $b_i$ can be
calculated for the different phases. For the phase between the
electroweak symmetry breaking scale and the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking
scale ($M_Z$ to $M_X$) the one-loop beta-coefficients are given by
$b_{2L} = -19/6$, $b_{Y} =41/10$, $b_{3C} = -7$. For the phase between
the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale and the octet mass scale ($M_X$ to
$M_8$) the one-loop beta-coefficients are given by $b_{3L} = -13/2$,
$b_{X} =13/2$, $b^{331}_{3C} = -5$. Finally, for the phase between the
octet mass scale to the unification scale ($M_8$ to $M_U$) the
one-loop beta-coefficients are given by $b^{8}_{3L} =2n-13/2 $, where
$n$ is the number of generations of the fermionic octets
($ \Omega\equiv [1, 8^{\ast},0]$), $b^{8}_{X} =13/2$,
$b^{8}_{3C} = -5$.
In Fig.~\ref{svs_unif} we plot the gauge coupling
running of SVS $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model with the field content given in
Eqs.~(\ref{2.3},\ref{2.4}) and three generations of fermionic octets with $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale
$M_{X}= 3000$~GeV and octet mass scale $M_{8}= 9000$~GeV,
demonstrating successful gauge unification at the scale
$M_{U} = 10^{15.5}$~GeV with $n_{Y}=\sqrt{5/3}$ and $n_X=2/\sqrt{3}$. A relative modest variation of the octet mass scale from the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ scale, $M_8 / M_X = 3$, therefore lifts the scale of successful unification from the value $M_{U} = 10^{14.9}$~GeV found in Fig.~\ref{fig9} and thus relaxes the tension with proton decay limits, cf. Section~\ref{sec6}.
\subsection{SU(6) Grand Unification of the sequential $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ Model}
It is easy to verify from Eq. (\ref{5.1}) that each generation of the
fermionic multiplets of the sequential 331 model written in
Eq. (\ref{3.1}) fits perfectly in the anomaly free combination of
SU(6) representations: $ \bar{6} ~+~ \bar{6^{\prime}} ~+~ 15 $, where
$\bar{6}$ contains $d^{c}_{L}\equiv [3,1,-1/3]$ and
$\psi_{L}\equiv [1,3^{\ast},-1/3]$; $\bar{6^{\prime}}$ contains
$D^{c}_{L}\equiv [3,1,-1/3]$ and $\xi_{L}\equiv [1,3^{\ast},-1/3]$;
and $15$ contains $u^{c}_{L}\equiv [3^{\ast},1,-2/3]$,
$\chi_{L}\equiv [1,3^{\ast},2/3]$ and $Q_{L}\equiv [3,3,0]$. Now the
fundamental $27$ of E(6) branches under the maximal
$\mathrm{SU(2) \otimes SU(6)}$ subgroup as
$27 = [2,~\bar{6}] ~+~ [1,~ 15]$. Thus three $27$s of E(6) contain
three sets of $ \bar{6} ~+~ \bar{6^{\prime}} ~+~ 15 $ accommodating
the three generations of the fermionic multiplets of the sequential
331 model. However the minimal content of the sequential 331 model
does not have a low scale unification. However, by adding three
generations of fermionic octets
again leads to a successful gauge coupling unification.
\begin{figure}[t!]
\includegraphics[width=0.7\linewidth]{seq_su6_unif.pdf}
\caption{Gauge coupling running in the sequential 331 Model with three
generations of fermionic octets with $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale
$M_{X}= 3000$~GeV and octet mass scale $M_{8}= 8\times 10^7$~GeV,
demonstrating successful gauge unification at the scale $M_{U} =
10^{15.5}$~GeV with $n_{Y}=\sqrt{5/3}$ and $n_X=2/\sqrt{3}$.}
\label{seq_unif}
\end{figure}
For the
phase between the electroweak symmetry breaking scale and the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$
symmetry breaking scale ($M_Z$ to $M_X$) the one-loop
beta-coefficients are given by $b_{2L} = -19/6$, $b_{Y} =41/10$,
$b_{3C} = -7$. For the phase between the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale
and the octet mass scale ($M_X$ to $M_8$) the one-loop
beta-coefficients are given by $b_{3L} = -9/2$, $b_{X} =13/2$,
$b^{331}_{3C} = -5$. Finally, for the phase between the octet mass
scale to the unification scale ($M_8$ to $M_U$) the one-loop
beta-coefficients are given by $b^{8}_{3L} =2n-9/2 $, where $n$ is the
number of generations of the fermionic octets
($ \Omega\equiv [1, 8^{\ast},0]$), $b^{8}_{X} =13/2$,
$b^{8}_{3C} = -5$.
In Fig. \ref{seq_unif} we plot the gauge coupling
running of the sequential $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model with the field content given in
Eqs. (\ref{2.4},\ref{3.1}) and three generations of fermionic octets with $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale
$M_{X}= 3000$~GeV and octet mass scale $M_{8}= 8\times 10^7$~GeV,
demonstrating successful gauge unification at the scale
$M_{U} = 10^{15.5}$~GeV with $n_{Y}=\sqrt{5/3}$ and $n_X=2/\sqrt{3}$. In this scenario, the octet mass scale has to be detached rather strongly from the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ scale in order to achieve successful unification.
\section{ $\sin^{2}\theta_{w}$ and proton decay in SU(6) grand unification}
\label{sec6}
Using the RGEs and the relations among the coupling constants
corresponding to different gauge groups one can express
$\sin^{2}\theta_{w} (M_Z)$ in terms of the different scales associated
with the SU(6) grand unified theory. Noting that for SU(6) grand
unification we have $n_{Y}=\sqrt{5/3}$ and $n_X=2/\sqrt{3}$, the
relation between normalized couplings at the scales $M_Z$ and $M_X$
are given by
\begin{eqnarray}{\label{sin:1.1}}
\alpha_{2L}^{-1}(M_Z) &=& \alpha_{\rm{em}}^{-1}(M_Z)-\frac{5}{3} {\alpha^{N}_{Y}}^{-1}(M_Z), \\
{\alpha^{N}_{Y} }^{-1} (M_X) &=& \frac{1}{5}\alpha_{3L}^{-1} (M_X) +\frac{4}{5}{\alpha^{N}_{X}}^{-1} (M_X).
{\label{sin:1.2}}
\end{eqnarray}
Using Eq.~(\ref{sin:1.1}) it is straightforward to obtain
\begin{equation}{\label{sin:1.3}}
\sin^{2} \theta_w (M_Z) \equiv \frac{\alpha_{\rm{em}}(M_Z)}{\alpha_{2L}(M_Z)}=\frac{3}{8}+\frac{5}{8} \alpha_{\rm{em}}(M_Z) \left[ \alpha_{2L}^{-1} (M_Z) - {\alpha^{N}_{Y} }^{-1} (M_Z)\right] .
\end{equation}
Finally, using Eqs. (\ref{4.2}), (\ref{sin:1.2}) the above equation can be written in the form
\begin{eqnarray}{\label{sin:1.3}}
\sin^{2} \theta_w (M_Z) = &&\frac{3}{8}+\frac{5}{8} \alpha_{\rm{em}}(M_Z) \left[ \frac{4}{5} \left\{ \frac{b_{3L}}{2\pi} \ln \left( \frac{M_8}{M_X}\right) +\frac{b_{3L}^8}{2\pi} \ln \left( \frac{M_U}{M_8}\right) \right\} \right.\nonumber\\
&& \left. + \frac{(b_{2L}-b_Y)}{2\pi} \ln \left( \frac{M_X}{M_Z}\right) -\frac{4}{5} \frac{b_{X}}{2\pi} \ln \left( \frac{M_U}{M_X}\right) \right] ,
\end{eqnarray}
which can be readily used to obtain the prediction for
$\sin^{2}\theta_{w} (M_Z)$. For example, in the sequential 331 model
taking $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ symmetry breaking scale $M_{X}= 3000$~GeV, octet mass
scale $M_{8}= 8\times 10^7$~GeV, and unification scale
$M_{U} =10^{15.5}$~GeV we obtain
$\sin^{2}\theta_{w} (M_Z) \simeq 0.231$, which is consistent with the
electroweak precision data \cite{Agashe:2014kda}.
Turning to the prediction for proton decay, we note that being a
non-supersymmetric scenario the gauge $d=6$ contributions for proton
decay are most important here. An analysis of all $\mathrm{ SU(3)_c \otimes SU(2)_L \otimes U(1)_Y}$ invariant
operators
\cite{Weinberg:1979sa,Weinberg:1980bf,Wilczek:1979hc,Weldon:1980gi}
that can induce proton decay in SU(6) is beyond the scope of this
article and will be addressed in a separate communication. Here we
will consider the decay mode $p\rightarrow e^{+}\pi^{0}$, which is
constrained by experimental searches to have a life time
$\tau^{\rm{expt}}_{p}\geq 1 \times 10^{34}$ \cite{Agashe:2014kda}. The
relevant effective operators in the physical basis are given by
\cite{FileviezPerez:2004hn,Nath:2006ut}
\begin{eqnarray}{\label{pd:6.1}}
\mathcal{O}(e^{c}_{\alpha},d_{\beta}) &=& c (e^{c}_{\alpha},d_{\beta}) \epsilon_{ijk} \overline{u^{c}_{i}}_{L} \gamma^{\mu} u_{jL} \overline{e^{c}_{\alpha}}_{L} \gamma_{\mu} {d_{k\beta}}_{L} ,
\nonumber\\
\mathcal{O}(e_{\alpha},d^{c}_{\beta}) &=& c (e_{\alpha},d^{c}_{\beta}) \epsilon_{ijk} \overline{u^{c}_{i}}_{L} \gamma^{\mu} u_{jL} \overline{d^{c}_{k\beta}}_{L} \gamma_{\mu} {e_{\alpha}}_{L} ;
\end{eqnarray}
where
\begin{eqnarray}{\label{pd:6.2}}
c (e^{c}_{\alpha},d_{\beta}) &=& k_1^2 \left[ V^{11}_{1} V^{\alpha \beta}_{2} +(V_1 V_{UD})^{1\beta} (V_2 V_{UD}^{\dagger})^{\alpha 1} \right] ,
\nonumber\\
c (e_{\alpha},d^{c}_{\beta}) &=& k_1^2 V^{11}_{1} V^{ \beta \alpha}_{3}+k^2_2 (V_4 V_{UD}^{\dagger})^{\beta 1}+(V_1 V_{UD} V_4^{\dagger} V_3)^{1\alpha} ;
\nonumber\\
\alpha &=& \beta \neq 2.
\end{eqnarray}
Here $i,j,k=1,2,3$ are the color indices and $\alpha,\beta =1,2$;
$V_{1,2,3,4}$ and $V_{UD}$ are the mixing matrices
$V_1=U^{\dagger}_{C} U$, $V_2=E^{\dagger}_{C}D$,
$V_3=D^{\dagger}_{C}E$, $V_{UD}=U^{\dagger}D$; where $U,D,E$ are the
unitary matrices diagonalizing the Yukawa couplings e,g.
$U^{T}_{C}Y_U U=Y_{U}^{\rm{diag}}$. $k_1=g_{\rm{GUT}}/M_{(X,Y)}$ and
$k_2=g_{\rm{GUT}}/M_{(X^{\prime},Y^{\prime})}$, where
$M_{(X,Y)}, M_{(X^{\prime},Y^{\prime})}\sim M_{\rm{GUT}}$ are the
masses of the superheavy gauge bosons and $g_{\rm{GUT}}$ is the
coupling constant at the GUT scale. The decay rate for
$p\rightarrow e^{+}\pi^{0}$ mode is given by
\begin{eqnarray}{\label{pd:6.3}}
\Gamma (p\rightarrow e^{+}\pi^{0})=\frac{m_p}{16\pi f_{\pi}^{2}}A_{L}^{2} \vert \alpha_H \vert^{2} (1+D+F)^{2} \left[ \vert c (e,d^{c})\vert^{2} + \vert c (e^{c},d) \vert^{2} \right],
\end{eqnarray}
where $m_p=938.3$ MeV is the proton mass, $f_{\pi}=139$ MeV is the
pion decay constant, $A_{L}$ is the long distance renormalization
factor; $D,F$ and $\alpha_H$ are parameters of the chiral
Lagrangian. For a rough estimate, taking
$\alpha_H(1+D+F)\sim 0.012 ~{\rm{GeV}}^3$
\cite{Aoki:2006ib,Aoki:2008ku}; $A_R=A_L A^{\rm{SD}}_R \sim 3$, where
$A^{\rm{SD}}_R$ is the short distance renormalization factor; the
parameter depending on the mixing matrices $F_{q}(V)\sim 5$, we obtain
\begin{eqnarray}{\label{pd:6.4}}
\Gamma^{-1} (p\rightarrow e^{+}\pi^{0}) \sim 10^{36}\text{ yrs}
\left(\frac{\alpha^{-1}_\text{GUT}}{35}\right)^2
\left(\frac{M_U}{10^{16}\text{ GeV}}\right)^4.
\end{eqnarray}
Now noting that $M_{U} =10^{15.5}$~GeV and
$\alpha^{-1}_{\rm{GUT}}\sim35$ in the SVS and sequential 331 models,
the lifetime of the proton decay mode $p\rightarrow e^{+}\pi^{0}$
comes out to be \footnote{In fact for a more careful estimation, one
should also take into account the GUT threshold corrections which
might improve on this limit, however, given the uncertainties in the
hadronic parameters here we do not worry about such effects.}
$\sim 10^{34}$~yrs, which is consistent with the current
experimental limit \cite{Agashe:2014kda}.
\section{Discussion and outlook}
\label{sec:discussion-outlook}
In this paper we have considered the possibility of conventional
non-supersymmetric grand unification of extended electroweak models
based upon the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ gauge framework within an SU(6) gauge unification
group.
In contrast to other conventional grand unified theories, in SU(6) one
can have different components of the $\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ subgroup with different
multiplicity.
Such scenarios may emerge from the flux breaking of the unified group
in an E(6) F-theory GUT framework. While it allows for successful
unification, the required 331 scale is typically very close to
unification.
However, the sequential addition of a leptonic octet provides a way of
achieving gauge coupling unification at 331 scales accessible at
collider experiments.
Alternatively, we have also considered a sequential variant of the
$\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model that can have a minimal SU(6) grand unification, which in
turn can be a natural E(6) subgroup.
Such minimal SU(6) embedding does not require any bulk exotics in
order to account for the chiral families and allows for a TeV scale
$\mathrm{ SU(3)_c \otimes SU(3)_L \otimes U(1)_X}$ model as well as seesaw-induced neutrino masses.
In both cases the gauge coupling unification is associated to
the presence of sequential a leptonic octet at some intermediate
scale between the 331 scale, which lies in the TeV range, and the
unification scale.
It is important to stress that the presence of the octet plays a key
role in the mechanism of neutrino mass generation. In other words,
the same physics that drives unification is responsible for the
radiative origin of neutrino masses~\cite{Boucenna:2014dia}.
\section*{Acknowledgements}
This work is supported by the Spanish grants FPA2014-58183-P,
Multidark CSD2009-00064, SEV-2014-0398 (MINECO) and
PROMETEOII/2014/084 (GVA). CH would like to thank the organizers of
Planck 2016, Valencia, for their warm hospitality and IFIC's AHEP
Group, Institut de Fisica Corpuscular -- C.S.I.C./Universitat de
Valencia, Spain, where part of this work was carried out. The work of SP is partly supported by DST, India under the financial grant SB/S2/HEP-011/2013. The work of US is supported partly by the JC Bose National Fellowship grant under DST, India.
|
\section{Introduction}
\label{sec:introduction}
Slice-to-volume registration has received increasing attention during the last decades within the medical imaging community. Given a tomographic 2D slice and a 3D volume as input, this challenging problem consists in finding the slice (extracted from the input volume and specified by an arbitrary rigid transformation) that best matches the 2D input image. We stress the fact that we are working with 2D slices (e.g. ultrasound (US)) as opposed to projective 2D images (e.g. X-ray images). This is important since both problems are usually refereed as 2D/3D registration, even if they are intrinsically different. In slice-to-volume registration, every pixel from the 2D image corresponds to a single voxel in 3D space. However, in a projective 2D image every pixel represents a projection of voxels from a given viewpoint.
One can formulate different versions of slice-to-volume registration, depending on several aspects of the problem such as the \textit{matching criterion} used to determine the similarity between the images, the \textit{transformation model} we aim at estimating, the \textit{optimization strategy} used to infer the optimal transformation model (continuous or discrete) and the \textit{number of slices} given as input. In this work, we propose an iconic method (where the matching criterion is defined as a function of the image intensity values) to infer rigid transformation models (specified using 6-DOF). The input consists of a single slice and a single volume, and we formulate it as a discrete optimization problem.
Discrete methods, where the tasks are usually formulated as a discrete labeling problem on a graph, have become a popular strategy to model vision problems \cite{Wang2013} (and particularly, biomedical vision problems \cite{Paragios2016}) thanks to their modularity, efficiency, robustness and theoretical simplicity. This paper presents a graph-based formulation (inspired by the work of \cite{Zikic2010a, Zikic2010b}) to solve rigid (only) slice-to-volume registration using discrete methods. As we will discuss in section \ref{sec:previousWork}, other works have tackled similar problems. However, to date, no work has shown the potential of discrete methods to deal with rigid slice-to-volume registration. Our main contribution is to put a new spin on graph-based registration theory, by demonstrating that discrete methods and graphical models are suitable to estimate rigid transformations mapping slice-to-volume. We validate our approach using a dataset of magnetic resonance images (MRI) of the heart, and we compare its performance with a state-of-the-art approach based on continuous optimization using simplex method. Moreover, in the spirit of encouraging reproducible research, we make the source code of the application publicly available at the following website: \texttt{
https://gitlab.com/franco.stramana1/slice-to-volume}.
\subsection{Motivation}
\label{sec:motivation}
In the extensive literature of medical image registration, it is possible to identify two main problems which motivated the development of slice-to-volume registration methods during the last decades. The first one is the fusion of pre-operative high-definition volumetric images with intra-operative tomographic slices to perform diagnostic and minimally invasive surgeries. In this case, slice-to-volume registration is one of the enabling technologies for computer-aided image guidance, bringing high-resolution pre-operative data into the operating room to provide more realistic information about the patient's anatomy \cite{Liao2013}. This technique has been used when dealing with several scenarios such as liver surgery \cite{Bao2005}, radio-frequency thermal ablation of prostate cancer \cite{Fei2003}, minimally invasive cardiac procedures \cite{Huang2009}, among many others.
The second problem is the correction of slicewise motion in volumetric acquisitions. In a variety of situations, inter slice motion may appear when capturing a volumetric image. For example, in case of fetal brain imaging (essential to understand neurodevelopmental disabilities in childhood and infancy) \cite{Rousseau2005}, fetus motion generates inconsistencies due to the slice acquisition time. Another case is related to functional magnetic resonance images (fMRI), usually acquired as time series of multislice single-shot echoplanar images (EPI). Patient head motion during the experiments may introduce artifacts on activation signal analysis. Slice-to-volume registration can be used to alleviate this problem by registering every slice to an anatomical volumetric reference following the well-know map-slice-to-volume (MSV) method \cite{Kim1999}.
\subsection{Previous Work}
\label{sec:previousWork}
Graph-based image registration, where the task is conceived as a discrete labeling problem on a graph, constitutes one of the most efficient and accurate state-of-the-art methods for image registration \cite{Sotiras2013}. Even if they have shown to be particularly suitable to estimate deformable non-linear transformations \cite{Glocker2011, Heinrich2013a}, they were also adapted to the linear case \cite{Zikic2010a}. Most of the publications on the field focus on registering images which are in dimensional correspondence (2D/2D or 3D/3D). In case of projective 2D/3D image registration, only linear transformations were estimated using discrete methods by \cite{Zikic2010b, Zikic2010a}. More recently, several graph-based approaches to perform deformable slice-to-volume registration were introduced in \cite{Ferrante2013, Ferrante2015, Ferrante2015a}. In these works, rigid transformations were computed as a by-product of the deformable parameters, leading to unnecessary computational burden (since rigid transformations are by far lower dimensional than deformable ones). To the best of our knowledge, rigid (only) slice-to-volume registration has not been formulated within this powerful framework. To date, all the methods focusing on this challenging problem are based on continuous (e.g. simplex~\cite{Fei2003}, gradient descent~\cite{Rousseau2005}, Powell's method~\cite{Gholipour2010}, etc.) or heuristic approaches (evolutionary algorithms~\cite{Tadayyon2011}, simulated annealing~\cite{Birkfellner2007}), missing the aforementioned advantages offered by discrete optimization. Based on the work of \cite{Zikic2010b}, we propose a discrete Markov Random Field (MRF) formulation of this problem, delivering more precise results than the state-of-the-art continuous approaches.
Moreover, inspired by the work of \cite{Lempitsky2008} in the context of vector flow estimation, we discuss how continuous state-of-the-art approaches can be used to further refine the rigid transformations obtained through discrete optimization, resulting in more accurate solutions.
\section{Rigid Slice-to-Volume Registration through Markov Random Fields}
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{images/interpretationTransformation2.png}
\caption{Interpretation of the components of Eq \ref{eq:sliceToVolume}. (a) Image $I$ corresponds to the input 2D image, which is moved by different rigid transformations $\pi$. (b) Image $J$ corresponds to the 3D image. Given a rigid transformation $\pi$, a slice $\pi[J]$ is extracted (using trilinear interpolation). In that way, it is possible to explore the space of solutions by sampling several rigid transformations $\pi$. (c) The matching criterion $\mathcal{M}$ quantifies the dissimilarity of both 2D images, $I$ and $\pi[J]$. Higher values indicate dissimilar images while lower values indicate better alignment.}
\label{fig:interpretationTransformation}
\end{figure}
We formulate rigid slice-to-volume registration as an optimization problem. Given a 2D image $I: \Omega_{I} \in \Re^2 \rightarrow \Re$, and a 3D image $J: \Omega_J \in \Re^3 \rightarrow \Re$, we aim at recovering the rigid transformation specified by $\pi = (r_x, r_y, r_z, t_x, t_y, t_z)$ that better aligns both images, by solving:
\begin{equation}
\hat{\pi} = \operatornamewithlimits{argmin}_{\pi} \mathcal{M}(I, \pi[J]),
\label{eq:sliceToVolume}
\end{equation}
where $\pi[J]$ corresponds to the slice extracted from image $J$ (using trilinear interpolation) and specified by the rigid transformation $\pi$ (as explained in Figure \ref{fig:interpretationTransformation}). $\mathcal{M}$ is the so-called matching criteria, that quantifies the dissimilarity between the 2D image $I$ and the slice $\pi[J]$. Alternative matching criteria can be adopted depending on the type of images we are registering. For example, in monomodal cases where intensities tend to be linearly correlated in both images, simple functions such as sum of absolute differences (SAD) or sum of squared differences (SSD) may make the job. However, for more complicated cases like multimodal registration (where the relation between intensity values in both images is usually non-linear), more elaborated functions like mutual information (MI) are applied.
This optimization problem is commonly solved through continuous (gradient or non-gradient based) methods, which are considerably sensible to initialization and may be stuck in local minima. As discussed in Section \ref{sec:previousWork}, in this work we model rigid slice-to-volume registration as a discrete labeling problem following the discretization strategy proposed by \cite{Zikic2010a}.
\subsection{Rigid Slice-to-Volume Registration as a Discrete Labeling Problem}
Rigid slice-to-volume registration, as well as many other problems in computer vision, can be formulated as a discrete labeling problem on a pairwise Markov Random Field (MRF) \cite{Wang2013}. Formally, a discrete pairwise MRF is an undirected graph $\mathcal{G} = \langle \mathcal{V}, \mathcal{E} \rangle$, where each node $v_i \in \mathcal{V}, i=1 ... |\mathcal{V}|$ represents a discrete variable. Any two variables $v_i, v_j$ depend on each other if there is an edge $(v_i, v_j) \in \mathcal{E}$ linking the corresponding nodes. The range of values that can be assigned to a discrete variable is determined by the label space $L$. A discrete labeling problem on a pairwise MRF consists on assigning a label $l_i \in L$ to every $v_i \in \mathcal{V}$, such that the following energy is minimized:
\begin{equation}
\mathcal{P}(\mathbf{x}; G, F) = \sum_{v_i \in V} g_i(l_i) + \sum_{(v_i,v_j) \in \mathcal{E}} f_{ij}(l_i, l_j),
\label{eq:mrfEnergy}
\end{equation}
\noindent where $\mathbf{x}=\{l_1, ..., l_n\}$ is a labeling assigning a label $l_i$ to every $v_i \in \mathcal{V}$, $G = \{g_i(\cdot)\}$ are the unary potentials associated to $v_i \in \mathcal{V}$ and $F = \{f_{ij}(\cdot, \cdot)\}$ are the pairwise potentials associated to edges $(v_i,v_j) \in \mathcal{E}$. These functions return scalar values when labels $l_i$ are assigned to variables $v_i$. Since we pose the optimization as a minimization problem, potentials must associate lower values to labelings representing good solutions, and higher values otherwise.
In the formulation presented in Eq~\ref{eq:sliceToVolume}, one would like to explore the space of parameters $\pi$ and chose the values giving the best matching. Since we are modeling it as a discrete problem, we must adopt a discretization strategy for the (naturally) continuous space of rigid transformations $\pi$. In \cite{Zikic2010a}, authors proved that it is possible to estimate linear (an particularly, rigid body) transformations by solving a discrete and approximated version of this formulation. Following their proposal, we model rigid slice-to-volume registration through a graph $\mathcal{G} = \langle \mathcal{V}, \mathcal{E} \rangle$, associating every parameter of the rigid transformation $\pi = (r_x, r_y, r_z, t_x, t_y, t_z)$ to a variable $v_i \in \mathcal{V}$, giving a total of 6 variables (nodes in the graph). $\mathcal{G}$ is a fully connected pairwise graph where $\mathcal{E}=\{(v_i, v_j)\}, \forall i \neq j$, meaning that all variables (parameters) depend on each other. Note that this pairwise model is clearly an approximation, since the real dependency between the parameters is not pairwise but high-order. However, as stated in \cite{Zikic2010a}, similar approximations have shown to be good enough to estimate linear transformations, while making the problem tractable.
In our discrete strategy, every parameter $v_i$ is updated through a discrete variation $d_{l_i}$ associated to the label $l_i$. Given an initial transformation $\pi_0 = (r^0_x, r^0_y, r^0_z, t^0_x, t^0_y, t^0_z)$, we explore the space of solutions by sampling discrete variations of $\pi_0$, and choosing the one that generates the slice $\pi[J]$ best matching image $I$. Therefore, for a maximum size $\omega_i$ and a quantization factor $\kappa_i$, we consider the following variations to the initial estimate of $v_i$: $\{0, \pm \frac{\omega_i}{\kappa_i}, \pm \frac{2\omega_i}{\kappa_i}, \pm \frac{3\omega_i}{\kappa_i}, ..., \pm \frac{\kappa_i \omega_i}{\kappa_i} \}$. The total number of labels results $|L| = 2 \kappa + 1$. Note that 0 is always included since we give the possibility of keeping the current parameter estimate. For example, in case that $v_0$ corresponds to $r_x$, $\omega_0=0.2$ and $\kappa_0=2$, the label space of $v_0$ will correspond to $\{r^0_x, r^0_x \pm 0.1, r^0_x \pm 0.2\}$.
Ideally, we would like to explore the complete search space around $\pi_0$ given by all possible combinations of labels. Since we have an exponential number of potential solutions, we adopt a pairwise approximation where only variations for pairs of variables are considered. This variations are encoded in the pairwise terms of the energy defined in Eq~\ref{eq:mrfEnergy} as $f_{ij}(l_i, l_j) = \mathcal{M}(I, \pi_{l_i, l_j}[J])$. Here $\pi_{l_i, l_j}$ denotes the updated version of $\pi_0$, where only $v_i$ and $v_j$ were modified according to the labels $l_i$ and $l_j$, while the rest of the parameters remained fixed. Unary potentials $g_i$ are not considered since we are only interested in the interaction between variables. Therefore, the discrete version of the optimization problem introduced in Eq~\ref{eq:sliceToVolume} becomes:
\begin{equation}
\hat{\mathbf{x}} = \operatornamewithlimits{argmin}_{\mathbf{x}} \mathcal{P}(\mathbf{x}; F) = \operatornamewithlimits{argmin}_{\mathbf{x}} \sum_{(v_i,v_j) \in \mathcal{E}} \mathcal{M}(I, \pi_{l_i, l_j}[J]),
\label{eq:sliceToVolumeMRF}
\end{equation}
where the optimal labeling $\hat{\mathbf{x}}$ represents the final rigid transformation $\hat{\pi}$ used to extract the solution slice $\hat{\pi}[J]$.
\subsection{Discrete Optimization}
We solve the discrete multi-labeling problem from Eq~\ref{eq:sliceToVolumeMRF} using FastPD. FastPD is a discrete optimization algorithm based on principles from linear programming and primal dual strategies, which at the same time generalizes the well known $\alpha$-expansion \cite{Komodakis2008}. One of the main advantages of FastPD is its modularity/scalability, since it deals with a much wider class of problems than $\alpha$-expansion, being an order of magnitude faster while providing the same optimality guarantees when performing metric labeling \cite{Komodakis2007a}. Our problem does not fulfill the conditions to be considered a metric labeling problem (we refer the reader to \cite{Boykov2001} for a complete discussion about metric labeling); however, FastPD has shown promising results for similar formulations \cite{Zikic2010a}.
FastPD solves a series of max-flow min-cut \cite{Boykov2004} problems on a graph. In that sense, it is similar to $\alpha$-expansion which also performs discrete inference on multi-label problems by solving successive binary max-flow min-cut problems. The main difference between these approaches is the construction of the graph where max-flow min-cut algorithm is applied. $\alpha$-expansion constructs the binary problem by restricting the label space, so that the only options for a given variable are to remain in its current assignment, or to take a label $\alpha$ (which varies in every iteration). Instead, FastPD constructs these binary problems by performing a Linear Programming Relaxation (LPR) of the integer program that represents the discrete MRF formulation.
\subsection{Incremental Approach}
Discrete approximations of continuous spaces usually suffer from low accuracy (since it is bounded by the quality of the discretization). Thus, we adopt an incremental approach to explore the space of solutions in a finer way. The idea is to successively solve the problem from Eq~\ref{eq:sliceToVolumeMRF}, using the solution from time $t$ as initialization for time $t+1$, keeping a fixed number of labels but decreasing the maximum sizes $\omega_i$ in a factor $\alpha_i$. Moreover, we also adopt a pyramidal approach, where we generate a Gaussian pyramid for both input images $I$ and $J$, and we run the complete incremental approach on every level of the pyramid. In that way, we increase the capture range of our method.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{images/visualResults.png}
\caption{Visual results for two slices of the individual tests. The first column corresponds to the input 2D slice. The second column shows the difference between the input 2D slice and the initial slice. The other columns show the difference between the input and the one resulting slices applying simplex, discrete and refined approaches. Grey values indicate no difference, while white and black value indicate inconsistencies. As it can be observed, the solution given by the refined approach is outperforming the others.}
\label{fig:visualResults}
\end{figure}
\subsection{Simplex Refinement Step}
Let us advance one of the conclusions of this work, so that we can motivate the last step of our approach. In Section \ref{sec:resultsIndividualTests}, we compare the performance of our discrete approach with a continuous method based on simplex \cite{Nelder1965} algorithm. As we will see, when the initialization $\pi_0$ is good enough, both continuous and discrete approaches perform well. In fact, in some cases, simplex is delivering more accurate solutions than discrete. However, as we move away from the initialization, discrete optimization gives more and more significant improvements, thanks to its wider capture range. In order to improve the accuracy of our proposal, and inspired by similar conclusions discussed by \cite{Lempitsky2008} in the context of vector flow estimation, we refine the results obtained with our approach by optimizing Eq \ref{eq:sliceToVolume} through simplex, using the discrete solution as initialization.
\section{Experiments and Results}
\label{sec:results}
In this section, we present the results obtained using the proposed method (considering also the refined version), and we compare them with a state-of-the-art approach based on continuous optimization trough downhill simplex \cite{Nelder1965} (also known as Nelder-Mead or amoeba method). Simplex is one of the most popular optimization algorithms used to deal with rigid slice-to-volume registration (some examples are \cite{Xu2014b, Kim1999, Birkfellner2007, Park2004}). It is a continuous and derivative-free method, which relies on the notion of simplex (which is a special polytope of $n+1$ vertices living in a $n$-dimensional space) to explore the space of solutions in a systematic way. We used a dataset composed of MRI images of a beating heart. Given an initial sequence of 3D images $M_i, i=0 .. 19$ of a beating heart (with a resolution of $192 \times 192 \times 11$ voxels and a voxel size of $1.25 \times 1.25 \times 8 mm$), we generated slices which were used for two different experiments.
\subsection{Implementation Details and Parameters Setting}
We implemented the three versions of the algorithms discussed in this paper (simplex, discrete and refined) mainly using Python and ITK for image manipulation \footnote{The source code can be downloaded from \\ \texttt{
https://gitlab.com/franco.stramana1/slice-to-volume}}. For simplex optimization we used the method implemented in \textit{scipy.optimize} package, while discrete optimization was performed using a Python wrapped version of the standard C++ implementation of FastPD. In all the experiments, we used a pyramidal approach with 4 Gaussian levels (3D images where not downsampled in z axis because of the low resolution of the images in this direction). The matching criterion adopted in all the experiments was the sum of squared differences, since pixel intensities are equivalent in both 2D and 3D images. The matching criterion $\mathcal{M}$ based on SSD is simply defined as:
\begin{equation}
\mathcal{M}(I_1, I_2) = \sum_{i \in \Omega} (I_1(i) - I_2(i))^2,
\label{eq:matchingCriterion}
\end{equation}
For the discrete case, at every pyramid level we decreased the maximum label size for both, rotation ($\omega_{rot} = [$0.02, 0.015, 0.0125, 0.01$]$rad) and translation ($\omega_{trans} = [$7, 6.5, 6, 5$]$mm) parameters. Starting from these maximum sizes, we solved Eq~\ref{eq:sliceToVolumeMRF} running FastPD several times per level until no improvement is produced or a maximum number of iterations is achieved ([200, 100, 150, 600]), using different label space decreasing factors at every pyramid level ($\alpha$=[0.08, 0.07, 0.05, 0.03]). The total number of labels was fixed to 5 ($\kappa = 2$) for all the experiments. For the continuous case (where Eq~\ref{eq:sliceToVolume} was optimized using simplex), we used again a 4-levels pyramidal approach, with simplex running until convergence in every level. Finally, for the refined experiment, we just ran the simplex experiment initialized with the solution estimated with the discrete method. For every registration case, continuous approach took around 30secs while the discrete version took 9mins, running on a laptop with an Intel i7-4720HQ and 16GB of RAM.
\label{sec:resultsIndividualTests}
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{images/individualsSAD.pdf}
\caption{Individual tests where 100 2D slices (extracted at locations specified using random rigid transformations) are considered as independent registration cases. Every point form the scatter plot represents the mean of absolute differences (MAD) between the input 2D image and the slice extracted at the initial position (X axis) vs the estimated position (Y axis). We also include a linear trend estimation (fitted using least squares method) to compare the robustness of the method to bad initializations.}
\label{fig:individualSAD}
\end{figure}
\subsection{Experiments}
We performed two different type of experiments, considering individual registration cases as well as image series. For validation, we measured three different indicators: the distance in terms of translation and rotation parameters between the estimated and ground truth transformations, together with the mean of absolute differences (MAD) between the input 2D image and the slice specified by the estimated rigid transformation.\\
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{images/boxplots.png}
\caption{Boxplot corresponding to the estimated error (in terms of rotation and translation parameters) for the 300 individual tests. As it can be observed, discrete and refined approaches are reducing both the mean error (shown as a dotted line in every box) and the dispersion.}
\label{fig:individualBoxplot}
\end{figure}
\begin{table}[t]
\centering
\label{tab:individualTests}
\caption{Average error estimated in terms of rotation (expressed in radians), translation (expressed in millimeters) and MAD for the three alternative approaches discussed in this paper. As it can be observed, the discrete and refined methods outperform the standard continuous approach optimized through simplex.}
\begin{tabular*}{\columnwidth}{@{\extracolsep{\stretch{1}}}*{8}{c}@{}}
\hline
\textbf{Method} & $R_x$ & $R_y$ & $R_z$ & $T_x$ & $T_y$ & $T_z$ & \textit{MAD} \\ \hline
\textit{Simplex} & 0,14 & 0,13 & 0,12 & 8,46 & 9,62 & 10,69 & 51,88 \\ \hline
\textit{Discrete} & 0,12 & 0,08 & 0,09 & 5,87 & 6,72 & 6,18 & 42,36 \\ \hline
\textit{Refined} & \textbf{0,10} & \textbf{0,07} & \textbf{0,08} & \textbf{5,09} & \textbf{5,96} & \textbf{4,92} & \textbf{36,45} \\ \hline
\end{tabular*}
\end{table}
\noindent \textbf{Individual tests.} The first set of experiments measures the accuracy of the three approaches using individual tests, where 100 random slices extracted from the 20 volumes are considered as single images (independently of the series), and registered to the first volume $M_0$. We run the same experiment for every slice using three different initializations (resulting in 300 registration cases), where ground truth parameters were randomly perturbed in three different ranges ([5, 12), [12, 18), [18, 25) millimeters for translation and [0.1 , 0.2), [0.2, 0.3), [0.3, 0.4) radians for rotation parameters) to guarantee that both, good and bad initializations, are considered for every slice. Quantitative results are reported in Figures \ref{fig:individualSAD}, \ref{fig:individualBoxplot} and summarized in Table \ref{tab:individualTests}. Visual results for qualitative evaluation are reported in Figure \ref{fig:visualResults}.
Results in the scatter plot from Figure \ref{fig:individualSAD} indicate that, as we go farther away from the initialization (in this case, it is quantified by the MAD between the input 2D image and the slice corresponding to the initialization), discrete and refined methods tend to be more robust. This robustness is clearly reflected by the slope of the trend lines: the refined method presents the trend line with the lower slope, meaning that even for bad initializations it converges to better solutions. The boxplot from Figure \ref{fig:individualBoxplot} and the numerical results from Table \ref{tab:individualTests} confirm that discrete and refined methods perform better not only in terms of MAD, but also with respect to the distance between the rotation/translation estimated and ground truth parameters.\\
\noindent \textbf{Temporal series test.} The idea behind the second experiment is to simulate an image guided surgery (IGS) scenario, where a fixed pre-operative volume must be fused with consecutive intra-operative 2D images suffering deformations (in this case, due to heart beating). Given the temporal sequence of 20 volumetric MRI images $M_i, i=0 .. 19$, we generated a sequence of 20 2D slices to validate our method. It was extracted as in \cite{Ferrante2013}: starting from a random initial translation $T_0=(T_{x_0}, T_{y_0}, T_{z_0})$ and rotation $R_0=(R_{x_0}, R_{y_0}, R_{z_0})$, we extracted a 2D slice $I_0$ from the initial volume $M_0$. Gaussian noise was added to every parameter in order to generate the position used to extract the next slice from the next volume. We used $\sigma_r=3^\circ$ for the rotation and $\sigma_t=5\mathit{mm}$ for the translation parameters. All the slices were registered to the first volume $M_0$. The solution of the registration problem for slice $I_i$ was used as initialization for the slice $I_{i+1}$. The first experiment was initialized randomly perturbing its ground truth transformation with the same noise parameters. As it can be observed in Figure \ref{fig:temporalSeries}, discrete and refined approaches manage to keep a good estimation error while simplex can not. Note that different strategies could be used in real scenarios to obtain good initializations for the first slice. For example, in IGS, physicians could start from a plane showing always the same anatomical structure, or identify landmark correspondences in the first slice and the 3D image useful to estimate an initial transformation.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{images/series-chart.png}
\caption{Results for the temporal series experiment. In this case, the transformation estimated for the slice $i$ was used as initialization for the next slice of the series. We reported results in terms of MAD and rotation/translation error for the estimated transformations using the three approaches.}
\label{fig:temporalSeries}
\end{figure}
\section{Discussion, Conclusions and Future Works}
In this paper we presented a strategy to solve rigid slice-to-volume registration as a discrete graph labeling problem, following the discretization strategy introduced by \cite{Zikic2010a}. We validated our proposal using a MRI dataset of a beating heart, where arbitrary 2D slices are fused with a 3D volume. The experimental results showed that our discrete approach produces more accurate and robust estimates for the rigid transformations than a continuous method based on simplex. Moreover, they also reflected that results obtained using such a method can be further refined using a continuous approach like simplex, leading to even more accurate estimations. This is coherent with the conclusions presented by \cite{Lempitsky2008} for the case of optical flow estimation.
An interesting discussion about the limitations of our approach, emerges when we observe the results obtained in previous work by \cite{Ferrante2013, Ferrante2015, Ferrante2015a} using similar images. In these works, both rigid and local deformable parameters are estimated in a one-shot discrete optimization procedure, delivering results which are considerably better than ours, even for the refined approach. Since we are dealing with 2D images which are deformed with respect to the static 3D volume (due to heart beating), estimating both rigid and deformable parameters at the same time seems to be the correct solution since there is a clear mutual dependence between them. However, if we look at the results corresponding to the first slices of the temporal series in Figure \ref{fig:temporalSeries} (where there is almost no deformation, and even null deformation for the 1st slice), we can see that the quality of the solution is significantly better than in the other cases. In fact, the error is almost 0. It suggests that when the 2D image is not deformed with respect to the input volume, our method is enough to capture slice-to-volume mapping. This limitation is somehow inherent to the model we are using: rigid transformations can not deal with local deformations. To improve the results in these cases, we plan to extend our approach to linear transformations where also anisotropic scaling and shearing can be considered. Following the strategy by \cite{Zikic2010a}, it will result straightforward.
Finally, a future line of research has to do with applying discrete rigid (or linear) slice-to-volume registration to other problems. As discussed in Section \ref{sec:motivation} motion correction for volume reconstruction is another problem requiring mapping slice-to-volume. It would be interesting to explore how our approaches performs in this case.
\bibliographystyle{splncs}
|
\section{Introduction}
The quantum physics of black holes is full of various paradoxes.
One of these problems is related to the understanding of the particle production
process induced by black holes. The created particles $\{|\tilde{\psi}_i\rangle\}$ possess a thermal profile and
turn out to appear in empty space, i.e. outside of the matter that has gravitationally collapsed into a black hole.
The fact is that such a process
is in apparent conflict with the unitary evolution in quantum theory~\cite{Hawking}. In other words,
the $S$-matrix between initial vacuum state and final thermal state cannot be unitary. If one
demands the $S$-matrix be unitary, one comes to another problem - the firewall
paradox~\cite{Almheiri&Marolf&Polchinski&Sully}, - i.e. the near-horizon region is populated by the
high-energy particles of the type $\{|\tilde{\psi}_i\rangle\}$. If this was correct, then this would imply
the equivalence principle, being at the heart of general relativity, would not in general hold in quantum
theory.
There exists a logically non-excludable interpretation of the black-hole evaporation (non-vanishing
positive outgoing energy flux) which seems to be free of the mentioned inconsistencies.
The quantum state remains empty outside of the collapsed matter, i.e. there are no physical particles in
the locally Minkowski
vacuum (we denote this vacuum as $|\Omega\rangle$ in the following). In other words, the initial
Hilbert space representation $\mathcal{H}$ of the algebra $\mathcal{A}$ of all field operators known
in the Standard Model is still a physical representation even after the black hole has formed.
This implies in particular that the spectrum of particles $\{|\psi_i\rangle\}$ as being elements of
$\mathcal{H}$ (assigned far from the event horizon to the irreducible representations of the
Poincar\'{e} group $\mathcal{P}_+^\uparrow$; below we show how this can be done near the
event horizon) from which the collapsing matter were composed are still physical excitations after
the gravitational collapse.
The crucial question is then whether thermally distributed particles $\{|\tilde{\psi}_i\rangle\}$
are elements of $\mathcal{H}$. To put it differently, the question is whether there exist projection operators
$\hat{P}_i \in \mathcal{A}$ which map $\mathcal{H}$ on a one-dimensional subspace associated with
each of $|\tilde{\psi}_i\rangle$. The fact, however, is that these projections do not exist in
$\mathcal{A}$ and, hence, the particles $\{|\tilde{\psi}_i\rangle\}$ are elements of another Hilbert
space $\tilde{\mathcal{H}} \not\subset \mathcal{H}$~\cite{Emelyanov-2015-2}. This can be shown by
employing the argument that the splitting of the algebra $\mathcal{A}$ into factor subalgebras does not lead
to the factorization of the Hilbert space $\mathcal{H}$ in quantum field theory (in contrast to quantum
mechanics with finite dimensional Hilbert space representations, e.g., of a finite qubit system, which are all
unitarily equivalent according to the Stone-von Neumann theorem). Thus, there does not exist a
unitary $S$-matrix mapping elements of $\mathcal{H}$ into elements of $\tilde{\mathcal{H}}$ within the
framework of local quantum field theory. Therefore, for the states $\{|\tilde{\psi}_i\rangle\}$
to be physically realisable as
quanta of the Standard Model fields, there must occur a phase transition
$\mathcal{H} \rightarrow \tilde{\mathcal{H}}$
whenever a black hole forms. This scenario is physically unacceptable.
How can the black-hole evaporation then be understood? The evaporation process can still be accounted
for within local quantum field theory in a consistent way without referring to the states
$\{|\tilde{\psi}_i\rangle\}$. The
non-trivial value of the renormalised stress tensor $\langle \hat{T}_\nu^\mu\rangle$ of a certain quantum
field is assigned to the vacuum $|\Omega\rangle$. In other words, the vacuum is gravitationally ``active"
after the event horizon has formed (as $\langle \hat{T}_\nu^\mu\rangle$ enters the Einstein equations).
This occurs, because the field operators are sensitive (through their field equations) to
the geometry of
space-time and, hence, are modified not too far away from the horizon. The modification of the field
operators results in the change of the vacuum stress tensor $\langle \hat{T}_\nu^\mu\rangle$ (this
explains the featurelessness of the outgoing energy flux as the state $|\Omega\rangle$ is full of the
featureless quantum fluctuations only). This tensor decreases as $1/r^2$ far from the
hole~\cite{Christensen&Fulling,Candelas} and is thus practically zero sufficiently far away from
the horizon.
An analogous effect occurs, e.g., in the Casimir set-up, where the (Minkowski) vacuum $|\Omega\rangle$
also possesses a non-vanishing energy-momentum tensor. This is due to the modification of the
electromagnetic operators $\hat{E}_i$ and $\hat{B}_i$ between the conducting plates. This leads to
$\langle\hat{E}_i\hat{E}_j\rangle \neq 0$ and $\langle\hat{B}_i\hat{B}_j\rangle \neq 0$, although these
vanish in the absence of the plates.
This analogy seems also to hold when one studies electromagnetic properties of the vacuum within quantum
electrodynamics. The electromagnetic properties of the vacuum $|\Omega\rangle$ are characterised by
the electric permittivity $\epsilon$ and the magnetic permeability $\mu$ entering the Maxwell equations.
It has been found within the Casimir set-up that the dispersion relation of low-energy
photons is modified in-between the plates~\cite{Scharnhorst}. We have recently shown that a similar effect
exists in the black-hole background~\cite{Emelyanov-2016-1}. Moreover, we have shown that photons
acquire an effective mass decreasing as $1/r$ and a point-like electric charge can be partially screened due
to black holes~\cite{Emelyanov-2016-2}.
In this paper we study properties of the vacuum $|\Omega\rangle$ in the framework of the
massless $\lambda\Phi^4$-theory. The main purpose is to find out any physical inconsistencies related to
the idea of having medium-like characteristics of the vacuum in the background of black holes. Specifically,
we consider a self-interacting scalar model in the far-from- and near-horizon
region of a large black hole. The
one-loop correction to the self-energy and coupling constant are computed in the near-horizon
region. As expected, we find that these corrections are suppressed as $(\lambda_\mathbf{p}/r_H)^2$ near the
event horizon, where $\lambda_\mathbf{p}$ is the de Broglie wavelength of the scalar particle and $r_H$ is
a size of the black-hole horizon.
The outline of this paper is as follows. In Sec.~\ref{sec:physical-particles}, we introduce the principle of
how the physical particles should be defined in curved space-times. The basic idea is to employ a local
Minkowski frame to identify a particle with a localised state as one has been successfully
doing that in
Minkowski space. The equivalence principle plays a crucial role in extending the particle notion to any
space-time domain in which gravity is not too strong. This tacitly implies that the notion of particle
is \emph{not} observer-dependent, i.e. covariant, as opposed to the common belief.
In Sec.~\ref{sec:scalar-model},
the self-interacting scalar field is considered in the background of an evaporating and
spherically symmetric black hole of astrophysical mass, i.e. $M \geq M_\odot$, where $M_\odot$ is the
solar mass. For these black holes, the local inertial frame might be of the size
$l_M \gtrsim 300(M/M_\odot)\,\text{m}$. Thus, a particle detector of the size $l_D \ll 30(M/M_\odot)\,\text{m}$
could be employed to study how particle scattering reactions quantitatively differ from the
same reactions
in the asymptotically flat region, i.e. in the region, where the influence of black holes can be ignored.
In Sec.~\ref{sec:lpp}, we provide an argument why the firewall paradox does not exist.
The non-existence of the unitary $S$-matrix relating the states $\{|\psi_i\rangle\}$ with
$\{|\tilde{\psi}_i\rangle\}$ is on the contrary consistent with local quantum field theory, but does not imply
the unitarity in the gravitational collapse is broken. The outstanding problem is how to take into account
the backreaction of the quantum fields on the quantum state of the collapsed matter which is under the
horizon.
In Sec.~\ref{sec:concluding remarks}, the main results are summarised.
Throughout this paper the fundamental constants are set to $c = G = k_\text{B} = \hbar = 1$, unless
stated otherwise.
\section{Local notion of particle in curved space-time}
\label{sec:physical-particles}
We want to study certain scattering processes in the vicinity of the horizon of a large black hole
formed through the gravitational collapse and compare these with observations in the far-from-horizon
region. For this purpose it is first necessary to define a physical notion of particle in curved
space-time. Bearing in mind the remarkable success of particle physics based on QFT,
the notion of particle should be related to the pole structure of the Feynman propagator when computed
in the local Minkowski frame. Therefore, the guiding principle should be based on reproducing the standard
results of particle physics at any given point in the universe.
Particle physics formulated in Minkowski space and based on the Standard Model
has successfully passed so far all tests in the particle colliders up to the energy scale $1\,\text{TeV}$.
Certainly, there is physics beyond the Standard Model which is assumed to be associated with
the theory itself (e.g., the neutrino oscillation implies that
at least two among of three neutrino flavors are massive), rather than the modifications of
the basic QFT principles.
However, the universe is globally non-flat. The observable part of the universe looks
at cosmological scales
($100\,\text{Mpc} \lesssim l \lesssim 3000\,\text{Mpc}$) as de Sitter space due to dark energy (see,
e.g.,~\cite{Mukhanov}). The universe becomes inhomogeneous and anisotropic at smaller scales due
to dark matter, clusters of galaxies, galaxies and so on. At much smaller, but still macroscopic scales,
the universe definitely looks as being nearly Minkowski space.
Moreover, earth is also a source of the non-trivial local curvature.
If one introduces the normal Riemannian coordinates $y$, then the local geometry becomes flat:
\begin{eqnarray}
g_{\mu\nu}(y) &=& \eta_{\mu\nu} + \frac{1}{3}R_{\mu\lambda\rho\nu}(y)\,y^{\lambda}y^{\rho} +
\text{O}\big(\nabla{R}\,y^3\big) \quad \text{with} \quad |Ry^2| \;\ll\; 1\,.
\end{eqnarray}
One usually employs the Minkowski-space approximation in order to describe various
scattering processes in the particle colliders. Therefore, it is \emph{a posteriori}
legitimate to use the Fourier-transform
technique (with integration over the whole space-time) whenever one restricts oneself to space-time
regions with a size $l \ll r(r/r_\oplus)^{\frac{1}{2}}$ ($r_\oplus \approx 8.7\,\text{mm}$ is earth's
gravitational radius) for the reason explained below. Specifically, the size $l_\text{LHC}$ of the
LHC is about $27\,\text{km}$ and we find
$R_\oplus(R_\oplus/r_\oplus)^{\frac{1}{2}} \approx 1.7{\times}10^{8}\,\text{km} \gg R_\oplus \gg l_\text{LHC}$,
where $R_\oplus \approx 6.4{\times}10^{3}\,\text{km}$ is earth's raduis. In other words, the gravitational
influence of earth on scattering processes in the LHC can be safely ignored.
The integration over the whole space-time (as if it is infinitely large Minkowski space) in vertices in
non-linear field theories does not entail any sort of inconsistencies,
because particles are described by localized states in
quantum field theory~\cite{Haag}. The particle state looks like the vacuum $|\Omega\rangle$
for measurements performed outside of its support. This is in turn characterised by the size of particle's wave
packet. For instance, electron has a size which is about its Compton wavelength
$\lambda_e = \frac{h}{m_ec} \approx 2.4{\times}10^{-12}\,\text{m}$. Thus, it makes a physical sense
to speak about the electron (at rest) as a localised object, whenever one restricts oneself to space-time regions
of the size $\lambda_e \ll l \ll l_c$, where $l_c$ is a characteristic size of the curvature
(e.g., $l_c \sim 10^{8}\,\text{km}$ for earth and a hypothetical particle of the rest mass less than
$10^{-17}\,\text{eV}$ could not be understood as being localised).
In Minkowski space-time, the wave packet $h_e(x)$ characterising the electron is a positive energy
solution of the Dirac field equation. It possesses a non-vanishing support in the spatial
region of the extent $\lambda_e$. The notion of energy is defined with respect to the
Minkowski time
translation operator $G = \partial_\tau$ whose integral curves are geodesics, i.e. it satisfies the
geodesic equation $\nabla_GG = 0$. Thus, one has
\begin{eqnarray}\label{eq:wave-packet}
h_e(x) &=& \frac{1}{\left(2\pi\right)^{3/2}}\int d^4p\;
\theta\big(p_0\big)\delta\big(p^2 - m_e^2\big)e^{-ipx}h_e(p)\,,
\end{eqnarray}
where $h_e(p)$ is a wave packet in momentum space and $x^0 = \tau$. At scales $l \ll l_c$ in curved
space-time, the vector $G$ approximately satisfies the Killing equation, i.e.
$l_c\nabla_{(\mu}G_{\nu)} \approx 0$, so that it is one of the generators of the \emph{local} Poincar\'{e}
group. It turns out to be $l_c\nabla_{(\mu}G_{\nu)} \sim \text{O}(1)$ at larger scales $l \gtrsim l_c$, i.e.
$G$ is a local Killing vector.
Since the positive-energy packet $h(x)$ is chosen with respect to $G$ in particle physics and this choice
is consistent with the observations performed so far on earth (freely-moving, rather than moving along
any global Killing vector), we come to the following principle: \\[2mm]
\emph{A physical particle corresponds to a covariant wave packet $h(x)$ being a
positive energy solution of the field equation with respect to a geodesic vector $G$ determining the
dynamics in the local Minkowski frame.}\vspace{1.5mm}
The quantisation procedure of, e.g., a scalar non-interacting field $\hat{\Phi}(x)$ is performed by expanding
the field over the positive- and negative-frequency modes defined with respect to a certain
time-like Killing vector $K$:
\begin{eqnarray}\label{eq:field-splitting}
\hat{\Phi}(x) &=& \int{}d\mu(\mathbf{p})\Big(\phi_\mathbf{p}(x)\,\hat{a}_\mathbf{p} +
\phi_\mathbf{p}^*(x)\,\hat{a}_\mathbf{p}^\dagger\Big) \;=\; \hat{a}(x) + \hat{a}^\dagger(x)\,,
\end{eqnarray}
where $K\phi_\mathbf{p} = - i\omega\phi_\mathbf{p}$ and
$K\phi_\mathbf{p}^\dagger = + i\omega\phi_\mathbf{p}^\dagger$ and $d\mu(\mathbf{p})$ is some positive
measure of integration/summation. Due to the linearity of the field equation, there are infinitely many ways
of choosing the modes $\phi_\mathbf{p}(x)$. According to our principle, there is a preferred
choice (unique up the local Lorentz transformation). This depends on a
point in curved space-time, such that $K = G$ and, hence, only locally satisfies the Killing equation.
In this case, $\hat{a}^\dagger(h)$ is a creation operator of the physical particle from the vacuum
$|\Omega\rangle$ around of a support $\sigma$ of the wave packet $h(x)$. However, one has
\begin{eqnarray}\label{eq:particle-creation-operator}
\hat{a}^\dagger(h) &=& i\int_\Sigma d\Sigma_\mu\sqrt{-g(x)}\,g^{\mu\nu}(x)
\Big(\hat{\Phi}^\dagger(x)\nabla_\nu h(x) - h(x)\nabla_\nu\hat{\Phi}^\dagger(x)\Big)
\end{eqnarray}
for an arbitrary choice of the mode functions, where $\Sigma$ is a Cauchy surface. Taking now into
account the finite support of the wave packet $h(x)$ whose size is supposed to be much smaller than the
characteristic curvature scale $l_c$, we obtain
\begin{eqnarray}\nonumber
\hat{a}^\dagger(h) &=& i\int_\sigma d^3\mathbf{y}
\Big(\hat{\Phi}^\dagger(y)\,\partial_{y^0}h(y) - h(y)\,\partial_{y^0}\hat{\Phi}^\dagger(y)\Big) \quad
\text{with} \quad y_0 \;=\; \tau\,.
\end{eqnarray}
This is consistent with the Minkowski-space approximation one has been successfully employing in
particle physics. The physical particle is thus given by
\begin{eqnarray}
|\psi\rangle &=& \hat{a}^\dagger(h)|\Omega\rangle\,,
\end{eqnarray}
which is localised over the support of the wave packet $h(x)$ and, hence, is normalisable.
Thus, the equivalence principle plays an essential role in our proposal for defining the physical notion
of particle at any space-time point of the universe, where the local curvature length is
much larger than the particle size. Since these particles when considered in the local
Minkowski frame are identical to Wigner's ones, we can employ the standard methods of the Feynman
rules and diagrams to describe their scattering reactions.
\section{Self-interacting scalar field in Schwarzschild space}
\label{sec:scalar-model}
We shall consider a massless scalar field with the conformal coupling to gravity and the quartic
self-interacting term in the background of an evaporating and spherically symmetric black hole.
Specifically, the Lagrangian $\mathcal{L}$ is given by
\begin{eqnarray}\label{eq:lofsf1}
\mathcal{L} &=& -\frac{1}{2}\,\Phi{\Box}\Phi + \frac{1}{12}R\,\Phi^2 - \frac{\lambda}{4!}\,\Phi^4\,,
\end{eqnarray}
where $R$ is the Ricci scalar. The Ricci scalar vanishes in the Schwarzschild geometry which is described
by the line element
\begin{eqnarray}\label{eq:line-element}
ds^2 &=& g_{\mu\nu}dx^\mu dx^\nu \;=\; f(r)dt^2 - \frac{dr^2}{f(r)} - r^2d\Omega^2\,, \quad
\text{where} \quad f(r) \;=\; 1 - \frac{r_H}{r}\,,
\end{eqnarray}
where $r_H = 2M$ is a size of the black-hole horizon of mass $M$. In the following we shall use
the surface gravity $\kappa$ on the horizon which is defined as $\kappa \equiv
\frac{1}{2}f'(r_H) = 1/2r_H$.
A freely-falling frame in Schwarzschild space is characterised by an affine parameter $\tau$. This
parameter corresponds to the Painlev\'{e}-Gullstrand time (see, e.g.,~\cite{Martel&Poisson} for a
brief review). The line element~\eqref{eq:line-element} in the Painlev\'{e}-Gullstrand coordinates reads
\begin{eqnarray}
ds^2 &=& f(r)d\tau^2 - 2\big(1 - f (r)\big)^{\frac{1}{2}}d\tau dr - dr^2 - r^2d\Omega^2\,.
\end{eqnarray}
The time coordinate $\tau$ reduces to the standard Minkowski (M) time $t_M$ for $r \gg r_H$. That is also the case
for the Schwarzschild (S) time $t_{S}$ for $r \rightarrow \infty$.\footnote{Note that a
stationary observer, i.e. the observer moving along the Killing vector $\partial_{t_S}$, possesses
a non-trivial acceleration, which asymptotically vanishes in the spatial infinity.} However, the
Painlev\'{e}-Gullstrand and Schwarzschild time considerably differ from each other in the near-horizon
region. We shall establish the relation between these times at $r \sim r_H$ in what follows.
\subsection{Wightman function}
We found in~\cite{Emelyanov-2016-2} the Wightman two-point function
of the non-interacting ($\lambda = 0$), massless scalar field for the Unruh (U) state~\cite{Unruh}.
In this paper, we want to approximate the local Minkowski vacuum $|\Omega\rangle$
by the Unruh state in the far- and near-horizon region. This approximation is sufficiently accurate
at $r \gg r_H$. In case of $r \sim r_H$, the approximation is still adequate up to some corrections (we
shall come back to this issue below).
The 2-point function in the Unruh state reads
\begin{eqnarray}\label{eq:unruh-2-point-function}
W_U(x,x') &=& \vec{W}_{\beta}(x,x') + \cev{W}(x,x')
\\[1mm]\nonumber
&=&
\int\limits_0^{+\infty}d\omega
\left(\frac{\cos\left(\omega\Delta{t} +
i\frac{\omega\beta}{2}\right)}{4\pi\omega\sinh\left(\frac{\beta\omega}{2}\right)}\,\vec{K}_\omega({\bf{x}},{{\bf{x}}}') +
\frac{\exp\left(-i\omega\Delta{t}\right)}{4\pi\omega}\,\cev{K}_\omega({\bf{x}},{{\bf{x}}}')\right),
\end{eqnarray}
where $\beta \equiv 1/T_H$ is the inverse Hawking temperature. The right arrow refers to the ``outgoing"
modes, whereas the left one to the ``ingoing" modes. The functions
$\vec{K}_\omega({\bf{x}},{{\bf{x}}}')$ and $\cev{K}_\omega({\bf{x}},{{\bf{x}}}')$ are given by
\begin{eqnarray}\label{eq:kvec}
\vec{K}_\omega({\bf{x}},{{\bf{x}}}') &\approx&
\frac{\bar{\Delta}^{\frac{1}{2}}(\rho)\sin(\omega\rho)}{4\pi\omega\rho(f(r)f(r'))^{\frac{1}{2}}}
\left\{
\begin{array}{ll}
4\omega^2 - \frac{(f(r)f(r'))^{\frac{1}{2}}}{rr'}\,\Gamma_\omega\,, & r \sim r_H\,, \\[3mm]
\frac{(f(r)f(r'))^{\frac{1}{2}}}{rr'}\,\Gamma_\omega\,, & r \gg r_H\,,
\end{array}
\right.
\end{eqnarray}
and
\begin{eqnarray}\label{eq:kcev}
\cev{K}_\omega({\bf{x}},{{\bf{x}}}') &\approx&
\frac{\bar{\Delta}^{\frac{1}{2}}(\rho)\sin(\omega\rho)}{4\pi\omega\rho(f(r)f(r'))^{\frac{1}{2}}}
\left\{
\begin{array}{ll}
\frac{(f(r)f(r'))^{\frac{1}{2}}}{rr'}\,\Gamma_\omega\,, & r \sim r_H\,, \\[3mm]
4\omega^2 - \frac{(f(r)f(r'))^{\frac{1}{2}}}{rr'}\,\Gamma_\omega\,, & r \gg r_H\,,
\end{array}
\right.
\end{eqnarray}
where $\rho \equiv (2\bar{\sigma}({\bf{x}},{{\bf{x}}}'))^{\frac{1}{2}}$, $\bar{\sigma}({\bf{x}},{{\bf{x}}}')$
is the three-dimensional geodetic interval for the ultra-static or optical metric
$\bar{g}_{\mu\nu} = g_{\mu\nu}/f(r)$, $\bar{\Delta}(x,x')$ is the Van Vleck-Morette
determinant and
\begin{eqnarray}\label{eq:gamma}
\Gamma_\omega &\equiv& \sum_{l = 0}^{+\infty}(2l + 1)|B_{\omega l}|^2 \;\approx\; 27\omega^2M^2
\end{eqnarray}
in the DeWitt approximation~\cite{DeWitt}.
The origin of the ``ingoing" and ``outgoing" part of the Wightman function $W_U(x,x')$ can be understood
as follows. The scalar field operator $\hat{\Phi}(x)$ is represented through a sum of the non-Hermitian
operators, namely $\hat{\Phi}(x) = \hat{a}(x) + \hat{a}^\dagger(x)$, where $\hat{a}(x)$ annihilates the vacuum,
i.e. $\hat{a}(x)|\Omega\rangle = 0$ (cf. Eq.~\eqref{eq:field-splitting}). This operator can in turn be
represented as
\begin{eqnarray}
\hat{a}(x) &=& \hat{a}_>(x) + \hat{a}_<(x)\,,
\end{eqnarray}
such that $\hat{a}_<(x)|\Omega\rangle = \hat{a}_>(x)|\Omega\rangle = 0$, but $\hat{a}_<(x)$
has a non-vanishing support only for the advanced Finkelstein-Eddington time $v < v_H$, while
$\hat{a}_>(x)$ possesses a non-vanishing support for $v > v_H$, where $v_H$ corresponds to
the moment when the event horizon has formed (see, e.g.,~\cite{DeWitt}).\footnote{In
Minkowski space this splitting
can be done, e.g., with respect to the origin of the reference frame, such that $\hat{a}_>(x)$ and
$\hat{a}_<(x)$ have a support only on $\Sigma_>$ and $\Sigma_<$, respectively, where
$\Sigma_>$ is a part of the Cauchy surface for $x >0$, whereas $\Sigma_<$ for $x < 0$.
The total Cauchy surface $\Sigma$ in space is thus given by $\Sigma_<{\cup}\Sigma_>$.}
Since the operators $\hat{a}_<(x)$ and $\hat{a}_>(x)$ have non-intersecting supports, it generally
holds
\begin{eqnarray}
[\hat{a}_<(x),\hat{a}_>(x')] &=& [\hat{a}_<(x),\hat{a}_>^\dagger(x')] \;=\; 0\,.
\end{eqnarray}
The operator $\hat{a}_<(x) + \text{H.c.}$ is further split as follows~\cite{Hawking}:
\begin{eqnarray}
\hat{a}_<(x) + \text{H.c.} &=& \hat{b}(x) + \hat{c}(x) + \text{H.c.}\,,
\end{eqnarray}
where $\hat{b}(x)$ and $\hat{c}(x)$ have a non-vanishing support above and under the
horizon, respectively, and $\hat{b}(x)|\tilde{\Omega}\rangle = \hat{c}(x)|\tilde{\Omega}\rangle = 0$,
where $|\tilde{\Omega}\rangle$ is the Boulware vacuum~\cite{Boulware}. These operators commute
with each other as possessing non-intersecting supports:
\begin{eqnarray}
[\hat{b}(x),\hat{c}(x')] &=& [\hat{b}(x),\hat{c}^\dagger(x')] \;=\; 0\,.
\end{eqnarray}
The scalar field operator expressed through these operators becomes
\begin{eqnarray}\label{eq:scala-operator-splitting}
\hat{\Phi}(x) &=& \hat{\Phi}_>(x) + \hat{\Phi}_<(x) \;=\;
\hat{\Phi}_>(x) + \hat{\Phi}_b(x) + \hat{\Phi}_c(x)\,,
\end{eqnarray}
and, hence, the Wightman function $W_U(x,x')$ above the horizon ($\hat{\Phi}_c(x) = 0$) is
\begin{eqnarray}\nonumber
\langle\hat{\Phi}(x)\hat{\Phi}(x')\rangle &=&
\langle\hat{\Phi}_>(x)\hat{\Phi}_>(x')\rangle +
\langle\hat{\Phi}_b(x)\hat{\Phi}_b(x')\rangle + \langle\hat{\Phi}_c(x)\hat{\Phi}_c(x')\rangle
\\[2mm]
&=& \langle\hat{\Phi}_>(x)\hat{\Phi}_>(x')\rangle +
\langle\hat{\Phi}_b(x)\hat{\Phi}_b(x')\rangle \;=\; \cev{W}(x,x') + \vec{W}_{\beta}(x,x')\,.
\end{eqnarray}
Therefore, the ``ingoing" part of the 2-point function is due to the operator $\hat{\Phi}_>(x)$, while the ``outgoing"
one originates from the operator $\hat{\Phi}_b(x)$. Note that the vacuum $|\Omega\rangle$ responds
to the action of the operator $\hat{\Phi}_b(x)$ as a thermal state at the Hawking temperature $T_H$ defined
with respect to the Schwarzschild time $t_S$, while as an empty state when probed by $\hat{\Phi}_>(x)$.
In general, any polynomial composed of the operator $\hat{\Phi}_>(x)$ probes $|\Omega\rangle$ as
an empty state, while composed of $\hat{\Phi}_b(x)$ as if $|\Omega\rangle$ is a mixed state. It is worth
emphasising that the latter effect is due to the field operator $\hat{\Phi}_b(x)$, rather than the vacuum
state (see for an earlier version of this point~\cite{Emelyanov-2014-2,Emelyanov-2015-1}).\footnote{This
can be elucidated as follows. Consider a quantum-mechanical system of two
non-interacting harmonic oscillators $\{\hat{a},\hat{a}^\dagger\}$ and $\{\hat{b},\hat{b}^\dagger\}$ of the
same frequency $\omega$ and with a ground state $|0\rangle = |0_a\rangle{\otimes}|0_b\rangle$, such
that $\langle 0|\hat{a}^\dagger\hat{a}|0\rangle = \langle 0|\hat{b}^\dagger\hat{b}|0\rangle = 0$. Perform
now a double squeezed transformation, i.e. $\hat{\alpha} = \cosh\theta\,\hat{a} - \sinh\theta\,\hat{b}^\dagger$
and $\hat{\beta} = \cosh\theta\,\hat{b} - \sinh\theta\,\hat{a}^\dagger$. It is straightforward to show that
$[\hat{\alpha},\hat{\beta}] = [\hat{\alpha},\hat{\beta}^\dagger] = 0$, i.e. the oscillators
$\{\hat{\alpha},\hat{\alpha}^\dagger\}$ and $\{\hat{\beta},\hat{\beta}^\dagger\}$ are independent.
If one chooses the parameter $\theta$ of the transformation to satisfy
$\tanh\theta = \exp(-\omega\beta/2)$, then one has, e.g.,
$\langle 0|\hat{\alpha}^\dagger\hat{\alpha}|0\rangle = 1/(\exp(\beta\omega) - 1)
= \text{tr}(\hat{\rho}_\beta\hat{\alpha}^\dagger\hat{\alpha})$, where $\hat{\rho}_\beta$ is a density matrix
of inverse temperature $\beta$. Thus, one may say that the ground state $|0\rangle$ is
a mixed state when probed by operators of the type $\hat{\alpha}$. Thus, a pure state
can sometimes respond as a mixed state due to the non-triviality of quantum operators. Note
that a ground state $|\tilde{0}\rangle$ annihilated by both $\hat{\alpha}$ and $\hat{\beta}$
(but not by $\hat{a}$ or $\hat{b}$) can be mapped to $|0\rangle$ by a unitary operator. It is an easy
exercise to show that and is actually guaranteed by the Stone-von Neumann theorem. This is not
anymore the case in quantum field theory (QFT), where this map is \emph{not} unitarily implementable.
This fact allows in particular to describe phase transitions (e.g. normal phase $\leftrightarrow$ superconductive
phase) in the framework of
QFT, which is impossible in quantum mechanics. Note that, for that reason, toy models in the background
of black holes based on qubits presuppose a physical realisation of the Hawking excitations. Therefore,
this kind of the models cannot anyhow provide a resolution of the information loss problem, but can and does
cause further confusions.}
It is tempting to conclude that particles $|\tilde{\psi}\rangle$ defined as
\begin{eqnarray}
|\tilde\psi\rangle &=& \hat{b}^\dagger(\tilde{h})|\tilde{\Omega}\rangle
\end{eqnarray}
are thermally populated in the vacuum $|\Omega\rangle$, where $\tilde{h}(x)$ is a wave packet being
a positive frequency solution of the field equation with respect to $t_S$ with definite
values of the orbital and magnetic numbers. However,
one can show that $\langle\tilde{\psi}|\Omega\rangle$ is identically zero, i.e.
$\langle\tilde{\psi}|\Omega\rangle = 0$. Moreover, $\langle\tilde{\Omega}|\Omega\rangle = 0$ and,
hence, the vacua $|\Omega\rangle$ and $|\tilde{\Omega}\rangle$ give unitarily inequivalent Hilbert
space representations ($\mathcal{H}$ and $\tilde{\mathcal{H}}$, respectively) of the same operator
algebra $\mathcal{A}$~\cite{Emelyanov-2015-2}. This means that the physical interpretation of the
state $|\Omega\rangle$ as a thermal state of the particles $|\tilde\psi\rangle$ at the Hawking
temperature $T_H$ is not self-consistent, although most of the researchers take the contrary for
granted. We shall come back to this issue below.
\subsection{One-loop correction to self-energy}
In general, it is hardly possible to obtain the full propagator $G(x,x')$ in the non-linear theories.
This exact propagator in the massless $\lambda\Phi^4$-theory up to the 1-loop order satisfies
\begin{eqnarray}
\Big(\Box + m_\Phi^2 + \text{O}\big(\lambda^2\big)\Big)G(x,x') &=& \frac{-i}{(-g(x))^\frac{1}{2}}\,\delta\big(x-x'\big)\,,
\end{eqnarray}
Expanding the propagator $G(x,x')$ through the free propagator $G_U(x,x')$ by employing the standard
Feynman rules for this theory, one can obtain the effective mass of the scalar field at one-loop approximation.
This can be pictorially expressed as follows
\begin{eqnarray}
m_\Phi^2\,G_U(x,x') &=& -\Box_x\left(\mathbf{\imineq{1-loop-self-phi4.eps}{5.5}}
\right).
\end{eqnarray}
To compute the effective mass of the scalar field, one thus needs to establish the Feynman propagator
$G_U(x,x')$ in the Unruh state that approximates the local Minkowski one $|\Omega\rangle$.
This vacuum state is supposed to be physically (unitarily) equivalent to the vacuum state before the black
hole has formed. In other words, we do not expect the change of quantum physics (the change of the
Hilbert space representation of the field operator algebra or the phase transition) as a result of the
black-hole formation.
\subsubsection{Far-horizon region: $R \gg r_H$}
\label{subsubsec:weak}
Far away from the black hole, i.e. $R \gg r_H$, where $R$ is the distance to the centre of the black hole, one
can approximate geometry by Minkowski space. It implies that the geodetic interval for the optical and
physical metric approximately coincide and read
\begin{eqnarray}\label{eq:tadpole-far}
2\sigma(x,x') &\approx&
2\bar{\sigma}(x,x') \;\approx\; (t-t')^2 - (\mathbf{x}-\mathbf{x}')^2\,,
\end{eqnarray}
where $\mathbf{x} = (r\cos\theta\cos\phi,r\cos\theta\sin\phi,r\sin\theta)$. It should be noted
that the Schwarzschild time $t_S$ approaches the Painlev\'{e}-Gullstrand time $\tau$ in the asymptotically flat
region as
\begin{eqnarray}
t_S &=& \tau\,\big(1 + \text{O}(r_H/R)\big)\,,
\end{eqnarray}
such that $t_S \rightarrow \tau$ for $R \rightarrow \infty$.
The 2-point function in the asymptotically flat region is
\begin{eqnarray}\label{eq:correlator-far}
W_U(x,x') &\approx& \Big(1 - \frac{27r_H^2}{16R^2}\Big) W_M(x,x') + \frac{27r_H^2}{16R^2}\,W_M^\beta(x,x')\,
\end{eqnarray}
for $|x-x'| \ll R$, where $W_M^\beta(x,x')$ is functionally given by the Minkowski two-point function at the
inverse temperature $\beta = 1/T_H$, which becomes $W_M(x,x')$ for $T_H \rightarrow 0$. In the limit
$R \rightarrow \infty$, $W_U(x,x')$ reduces to the standard Minkowski correlator $W_M(x,x')$ and, hence,
the influence of the black hole on local physics can be fully neglected.\footnote{It is not the case for eternal
black holes. This does not serve a problem, because black holes of this type are not realistic. However, if
a tiny black hole of this type could appear as a result of a quantum space-time fluctuation, then its influence
on local physics cannot be neglected even at spatial infinity. For instance, photons acquire
an effective thermal mass of the order of $\alpha^\frac{1}{2}T_H$ for $T_H \gg m_e$ in
QED, assuming such a black hole exists for a sufficiently large time interval~\cite{Emelyanov-2016-2}.
These are certainly ruled out if the vacuum state is given by the Hartle-Hawking one.}
The Feynman propagator $G_U(x,x')$ can be expressed through the commutator function and reads
\begin{eqnarray}\label{eq:propagator-far}
G_U(k,k') &\approx& \left(\frac{i}{k^2 + i\varepsilon}
+ 2\pi\left(\frac{27r_H^2}{16R^2}\right)\frac{\delta\big(k^2\big)}{e^{|k_0|/T_H} - 1}\right)\delta\big(k-k'\big)
\end{eqnarray}
in the momentum representation~\cite{Emelyanov-2016-2}. The frequency $k_0$ here is defined
with respect to the Painlev\'{e}-Gullstrand time $\tau$. The second term on the right-hand side
of~\eqref{eq:propagator-far} depends on the distance to the black hole and vanishes in the limit
$R \rightarrow \infty$. Again, it means that local physics does not change in the asymptotic region,
where the field propagator is insensitive to the black-hole properties.
Thus, one obtains
\begin{eqnarray}\label{eq:smfar}
\Box_x\left(\mathbf{\imineq{1-loop-self-phi4.eps}{5.5}}\right) &=&
-\lambda\int dx_1(-g(x_1))^{\frac{1}{2}}\;\Box_xG_U(x,x_1)G_U(x_1,x_1)G_U(x_1,x')
\\[2mm]\nonumber
&=& -\lambda G_U(x,x)G_U(x,x')
\;\approx\; -\lambda\int\frac{d^3\mathbf{k}}{(2\pi)^3}\frac{1}{|\mathbf{k}|}
\left(\frac{1}{2} + \frac{27r_H^2}{16R^2}\frac{1}{e^{|\mathbf{k}|/T_H} - 1}\right)G_U(x,x')\,.
\end{eqnarray}
The integral over $\mathbf{k}$ in~\eqref{eq:smfar} diverges unless one subtracts the first term in
the parenthesis. In Minkowski space with a hot physical plasma, one can renormalise this by adding
a mass counter-term to the Lagrangian density. This essentially implies that one
removes this UV divergence by subtracting all terms which do
not vanish in the limit of the vanishing plasma temperature. In the black-hole background, this can
be accounted for the divergent Boulware contribution to the scalar mass.
Having got rid of the ultraviolet divergence in~\eqref{eq:smfar}, we obtain
\begin{eqnarray}\label{eq:smfar-reg}
\Box_x\left(\mathbf{\imineq{1-loop-self-phi4.eps}{5.5}}\right) &\approx&
-\frac{\lambda\,\xi}{16\pi^2R^2}\,G(x,x') \quad
\text{with} \quad \xi \;\equiv\; \frac{9}{128}\,.
\end{eqnarray}
The effective mass of the scalar field is thus finite and reads
\begin{eqnarray}
m_\Phi^2 &\approx& \frac{\lambda\,\xi}{16\pi^2R^2}\,.
\end{eqnarray}
The scalar-field mass $m_\Phi$ turns out not to depend on the black-hole mass $M$ in the
leading order of the approximation. It appears to be the case, because \eqref{eq:smfar-reg}
equals $T_H(r_H/R) \propto 1/R$ up to a numerical factor. We found a similar property of the one-loop
correction to the photon self-energy in the case of evaporating black holes of mass $M \ll 10^{16}\,\text{g}$
in~\cite{Emelyanov-2016-2}.
\subsubsection{Near-horizon region: $R \sim r_H$}
\label{sec:olctse-nhr}
The Wightman function $W_U(x,x')$ has a geometrical prefactor $1/f(r)$. Therefore, it is
tempting to conclude that the
loop diagrams are divergent on the horizon. For instance, the 1-loop correction to the self-energy of the
scalar field seems to increase as $1/f(r)$ in the near-horizon region, while as $1/f^2(r)$ at 2-loop level.
However, this does not happen to be the case if one studies this carefully in the freely-falling frame.
To analyse near-horizon physics, we introduce a local Minkowski frame at $R \sim r_H$ of a
black hole of mass $M \gtrsim M_\odot$. The size of the local Minkowski frame is then about
\begin{eqnarray}
l_M &\gtrsim& 300{\times}(M/M_\odot)\,\text{m}\,.
\end{eqnarray}
The local dynamics in this frame is set by the geodesic vector $G$, such that $G_\mu = \{1,0,0,0\}$
in the Painlev\'{e}-Gullstrand coordinates with the time coordinate $\tau$. This vector $G$ is one of the
generators of the local Poincar\'{e} group whose irreducible representations correspond to the
particle states.
The light-cone Kruskal-Szekeres coordinates $U$ and $V$ near the horizon behave as
\begin{subequations}\label{eq:kruskal-coordinates}
\begin{eqnarray}
U &=& \alpha\Big(\tau - \tau_0 - 2\kappa(\tau - \tau_0)^2 + \text{O}\big((\tau - \tau_0)^3\big)\Big)\,,
\\[1mm]
V &=& (e/\alpha)\Big(2/\kappa + \tau - \tau_0 + \text{O}\big((\tau - \tau_0)^3\big)\Big)\,,
\end{eqnarray}
\end{subequations}
where $e$ is the Euler number and $0 < \alpha \leq e^\frac{1}{2}$. The event horizon corresponds to
$U_H = 0$ or $\tau = \tau_0 > 0$, while $V_H = 2e/(\alpha\kappa)$ holds at $\tau_0$. This reveals the
geometrical meaning of the constant $\alpha$ in~\eqref{eq:kruskal-coordinates}.
Introducing $\tau' = (V+U)/2e^{\frac{1}{2}}$ and $x' = (V-U)/2e^{\frac{1}{2}}$, we obtain
\begin{eqnarray}
ds^2 &\approx& d\tau'{}^2 - dx'{}^2 - dy^2 - dz^2\,
\end{eqnarray}
near the horizon, where $y^2+z^2 = 4r_H^2\tan^2(\theta/2)$, $z/y = \tan\phi$ and $y^2 + z^2 \ll r_H^2$.
Employing the local Lorentz transformation with $v/c = (e-\alpha^2)/(e + \alpha^2)$, the time coordinate
$\tau'$ can be transformed to the Painlev\'{e}-Gullstrand time $\tau$ (up to a translation).
On the other hand, the Schwarzschild metric in the near-horizon region can be approximated by the
Rindler metric. Specifically, it holds
\begin{eqnarray}
ds^2 &\approx& \kappa^2\rho^2dt^2 - d\rho^2 - dy^2-dz^2\,,
\end{eqnarray}
where $\rho = \int dr/f^{\frac{1}{2}}(r) \approx (4r_H(r-r_H))^{\frac{1}{2}}$. This line element
can be further transformed into Minkowski one via the diffeomorphism
$\tau = \rho\sinh(\kappa t)$ and $x = \rho\cosh(\kappa t)$:
\begin{eqnarray}
ds^2 &\approx& d\tau^2 - dx^2 - dy^2 - dz^2\,.
\end{eqnarray}
Note that we have tacitly introduced a new time $\tau$ which is up to the Lorentz transformation
coincides with the Painlev\'{e}-Gullstrand time. Therefore, we have denoted both by the same symbol.
We now want to study local physics in this coordinate system. For this
purpose we need to derive the Feynman propagator $G_U(x,x')$ in the freely-falling frame. For the
reasons which become clear later on, we study first the vacuum expectation value of the operator
$\hat{\Phi}^2(x)$.
\subparagraph{Wick squared operator $\hat{\Phi}^2(x)$ near horizon}
Using the background-field method to compute the 1-loop contribution to the scalar field equation,
we find
\begin{eqnarray}
\Big(\Box + \frac{\lambda}{2}\,\langle\hat{\Phi}^2(x)\rangle\Big)\Phi(x)
&=& 0\,,
\end{eqnarray}
where $\langle\hat{\Phi}^2(x)\rangle$ has been appropriately renormalised (see below). Thus, the effective
scalar mass at 1-loop level is given by
\begin{eqnarray}\label{eq:smws}
m_\Phi^2 &=& \frac{\lambda}{2}\,\langle\hat{\Phi}^2(x)\rangle\,.
\end{eqnarray}
The renormalised value of the Wick squared operator $\hat{\Phi}^2(x)$ in the Unruh vacuum
was computed in~\cite{Candelas}. This quantity turns out to be finite on the horizon and decreases as $1/R^2$
at the spatial infinity. Specifically, it holds that
\begin{eqnarray}\label{eq:wick-squared-operator}
\langle \hat{\Phi}^2(x)\rangle &\approx&
\left\{
\begin{array}{ll}
\big(\frac{1}{3} - 2\xi\big)T_H^2\,, & R \sim r_H\,, \\[3mm]
\xi/8\pi^2R^2\,, & R \gg r_H\,.
\end{array}
\right.
\end{eqnarray}
This is in full agreement with our result obtained above for $R \gg r_H$ using the diagrammatic approach.
The finiteness of $\langle \hat{\Phi}^2(x)\rangle$ on the black-hole horizon seems \emph{a priori} not to be
guaranteed. Indeed, the Wick squared operator is defined in general as
\begin{eqnarray}
\hat{\Phi}^2(x) &=& \lim\limits_{x' \rightarrow x}\left(\hat{\Phi}(x)\hat{\Phi}(x') - H(x,x')\hat{1}\right),
\end{eqnarray}
where $H(x,x')$ is the Hadamard parametrix. It is a geometrical (state-independent) object and designed
to subtract the ultraviolet divergences in the 2-point function only. In our case, the Hadamard parametrix
is of the form
\begin{eqnarray}
H(x,x') &=& -\frac{1}{8\pi^2\sigma(x,x')} + \textrm{O}\big(\sigma\ln\sigma\big)\,,
\end{eqnarray}
where $\sigma(x,x')$ is the geodetic interval for the physical metric. It is worth mentioning that the term
$\sigma\ln\sigma$ in $H(x,x')$ is fully responsible for the trace aka conformal
anomaly~\cite{Moretti,Decanini&Folacci}.
The parametrix can be expressed via the geodetic interval in the optical metric:
\begin{eqnarray}\label{eq:hp}
H(x,x') &=& -\frac{\left(f(r)f(r')\right)^{-\frac{1}{2}}}{8\pi^2\bar{\sigma}(x,x')}
+ \frac{M^2f^{-1}(r)}{48\pi^2r^4} + \textrm{O}\big(\bar{\sigma}\ln\bar{\sigma}\big)\,.
\end{eqnarray}
The first term in~\eqref{eq:hp} coincides with the 2-point function $W_B(x,x')$ in the
Boulware vacuum for $x \sim x'$, whereas the second term in~\eqref{eq:hp} gives a non-vanishing
value of $-\langle\tilde{\Omega}| \hat{\Phi}^2(x)|\tilde{\Omega}\rangle$ far away
from the event horizon which disappears in the limit $M \rightarrow 0$~\cite{Candelas}.
In the far-from-horizon region, the geodetic intervals $\sigma(x,x')$ and $\bar{\sigma}(x,x')$
go over to the Minkowski geodetic distance. These significantly differ from each other in the near-horizon
region. We obtain that
\begin{eqnarray}\label{eq:gdnh}
\sigma(x,x') &=& \sigma_0(x,x')
\\[1mm]\nonumber
&& + \;\frac{1}{24} \left(\frac{f'(r)^2-4\kappa^2}{f(r)}\frac{f'(r')^2-4\kappa^2}{f(r')}\right)^{\frac{1}{2}}\sigma_0^2(x,x')
+ \textrm{O}\big(\sigma_0^3(x,x')\big)\,,
\end{eqnarray}
where we have taken into account that $f'(r_H) = 1/r_H = 2\kappa$ and
\begin{eqnarray}
2\sigma_0(x,x') &\approx& \Delta\tau^2 - \Delta{x}^2 - \Delta{y}^2 - \Delta{z}^2
\end{eqnarray}
close to the horizon, while $\bar{\sigma}(x,x')$ can be found in~\cite{Emelyanov-2015-1} and approaches
in the limit $R \rightarrow r_H$ to the geodetic interval of the static space with the hyperbolic spatial
section. Substituting $\sigma(x,x')$ into the Hadamard parametrix, we find that
\begin{eqnarray}\label{eq:hadamard-parametrix-lif-near-horizon}
H(x,x') &=& -\frac{1}{8\pi^2\sigma_0(x,x')} - \frac{\kappa^2}{12\pi^2}
+ \textrm{O}\big(\sigma_0\ln\sigma_0,\kappa^2f(r)\big)\,.
\end{eqnarray}
This result and \eqref{eq:kvec} with \eqref{eq:kcev} at $R \sim r_H$ in the local Minkowski
frame in turn allow us to compute the Wightman function in the near-horizon region:
\begin{eqnarray}\label{eq:2-point-function-neat-horizon}
W_U(x,x') &\approx& -\frac{1}{8\pi^2\sigma_0(x,x')} - \frac{\xi\kappa^2}{2\pi^2}\,.
\end{eqnarray}
Thus, we reproduce the result \eqref{eq:wick-squared-operator} obtained in~\cite{Candelas} by subtracting
$H(x,x)$ from $W_U(x,x)$.\footnote{It is worth noticing that the finite terms in $W_U(x,x')$ and $H(x,x')$ in
the coincidence limit $x' \rightarrow x$ are separately regular on the event horizon $r = r_H$ in the freely-falling
frame. In the Schwarzschild frame, these parts of the Wightman function and the Hadamard
parametrix increase as $1/f(r)$ for $r \rightarrow r_H$, but their difference turns out to be non-singular at
$r = r_H$ as found in~\cite{Candelas}.}
\subparagraph{Feynman propagator near horizon}
The Feynman propagator can be expressed through the commutator function $C(x,x')$ which is equal to
$W_U(x,x') - W_U(x',x)$. The commutator $C(x,x')$ is insensitive to the time-independent term in the
correlation function. There are at least two possibilities to deal with the second term in $W_U(x,x')$:
Either one needs to
introduce an imaginary ``temperature" $\Theta \sim i\kappa$ or a discrete frequency spectrum
$\omega_n \sim \kappa n$. Since we expect merely a slight change of physics in the local inertial frame
even near the horizon of a large black hole, we do not consider the possibility of the
change of the continuous
spectrum into the discrete one.\footnote{It is worth mentioning that the finite term in the coincidence limit
$x' \rightarrow x$ of the thermal two-point function of temperature $T$ is given by $+T^2/12$. However,
the Wightman function in the near-horizon region given in~\eqref{eq:2-point-function-neat-horizon} has
a negative correction to the term $-1/\sigma_0(x,x')$. This effectively corresponds to the imaginary
``temperature".}
The time-independent term in \eqref{eq:2-point-function-neat-horizon} is negligibly small, because we have
been working in the regime $\sigma_0\kappa^2 \ll 1$. To reproduce the value of
$\langle \hat{\Phi}^2(x)\rangle$ through the tadpole diagram, we thus define
\begin{eqnarray}\label{eq:propagator-near}
G_U(p,p') &\approx& \left(\frac{i}{p^2 + i\varepsilon} + \frac{2\pi\delta(p^2)}{e^{|p_0|/\Theta_\varepsilon} - 1}\right)
\delta\big(p-p'\big)\,,
\end{eqnarray}
where we have introduced an \emph{effective} imaginary ``temperature":
\begin{eqnarray}\label{eq:imaginary-temperature}
\Theta_\varepsilon &\equiv& \varepsilon + \frac{i\kappa}{\pi}\,(6\xi)^{\frac{1}{2}} \quad \text{with} \quad
\varepsilon \;\rightarrow\; +0\,.
\end{eqnarray}
Since $\langle \hat{\Phi}^2(x)\rangle \propto G_U(x,x)$ from the tadpole diagram
(see Eq.~\eqref{eq:smfar}), one needs to renormalise it by subtracting the ``Hadamard propagator" at $x=x'$.
We define it as follows
\begin{eqnarray}\label{eq:propagator-hadamard}
G_H(p,p') &\approx& \left(\frac{i}{p^2 + i\varepsilon} + \frac{2\pi\delta(p^2)}{e^{|p_0|/\theta_\varepsilon} - 1}\right)
\delta\big(p-p'\big)\,, \quad \textrm{where} \quad
\theta_\varepsilon \;\equiv\; \varepsilon + \frac{i\kappa}{\pi}\,.
\end{eqnarray}
We can now reproduce the result of~\cite{Candelas} for the Wick squared or the effective scalar mass
if we renormalise the ultraviolet divergence of the tadpole diagram by subtracting $G_H(x,x')$ from $G_U(x,x')$.
Specifically, we have
\begin{eqnarray}\label{eq:effective-mass-diagram}
m_\Phi^2 &=& \frac{\lambda}{2}\Big(G_U(x,x) - G_H(x,x)\Big)
\\[2mm]\nonumber
&\approx& \frac{\lambda}{2}\int\frac{d^3\mathbf{k}}{(2\pi)^3}\frac{1}{|\mathbf{k}|}
\left(\frac{1}{e^{|\mathbf{k}|/\Theta_\varepsilon} - 1} - \frac{1}{e^{|\mathbf{k}|/\theta_\varepsilon} - 1}\right)
\;=\; \frac{\lambda}{24}\Big(\Theta^2 - \theta^2\Big)
\;=\; \frac{\lambda}{2}\Big(\frac{1}{3} - 2\xi\Big)T_H^2\,.
\end{eqnarray}
Note that both $\Theta^2$ and $\theta^2$ are negative, but the effective scalar mass $m_\Phi^2$
is positive, because the absolute value of $\theta$ is larger than that of $\Theta$.
It should be noted that the first term of the right-hand side in
Eq.~\eqref{eq:effective-mass-diagram} is due to $\theta$ that comes in turn from the Hadamard
parametrix. It is a geometrical object and its contribution to $m_\Phi^2$ is a result of the coordinate
transformation from the local Rindler geometry into the local Minkowski geometry near the event horizon.
\subsection{Local renormalisation scheme}
Above we have found that the effective scalar mass is finite and coincides with~\eqref{eq:smws} if we add
the minus ``Hadamard loop" to the Feynman loop. One may represent this subtraction pictorially
as follows
\begin{eqnarray}\label{eq:effective-mass-lrs}
m_\Phi^2\,G(x,x') &=& -\Box_x\left(\mathbf{\imineq{1-loop-self-phi4.eps}{5.5}} +
\mathbf{\imineq{1-loop-self-m-phi4.eps}{5.5}} + \text{O}\big(\lambda^2\big)\right),
\end{eqnarray}
where the dashed line corresponds to the minus Hadamard propagator, i.e. $-G_H(x,x')$.
We want to propose a local renormalisation scheme in curved space-times. Specifically, we
introduce a fictitious scalar field $\phi(x)$ with a \emph{negative} norm (a wrong sign in front of the propagator)
with the propagator being constructed from the Hadamard parametrix. We need also to introduce extra
vertices in the Lagrangian $\mathcal{L}$, namely
\begin{eqnarray}
\frac{\lambda}{4!}\,\phi^4\,,\quad \frac{4\lambda'}{4!}\,\phi^3\Phi\,,\quad
\frac{6\lambda}{4!}\,\phi^2\Phi^2\quad \text{and}\quad \frac{4\lambda'}{4!}\,\phi\Phi^3\,,
\end{eqnarray}
such those
\begin{subequations}
\begin{eqnarray}\label{eq:extra-vertices-1}
\mathbf{\imineq{0-loop-lambda-phi4-2.eps}{4.5}} &=&
\mathbf{\imineq{0-loop-lambda-phi4-4.eps}{4.5}} \;=\;-i\lambda\,,
\\[2mm]\label{eq:extra-vertices-2}
\mathbf{\imineq{0-loop-lambda-phi4-1.eps}{4.5}} &=&
\mathbf{\imineq{0-loop-lambda-phi4-3.eps}{4.5}} \;=\;-i\lambda'
\end{eqnarray}
\end{subequations}
in the momentum representation. The ratio $\lambda'/\lambda$ will be fixed below in order to recover
the standard result for the running coupling constant in the asymptotically flat region.
\subsection{One-loop correction to coupling constant}
We now study how the coupling constant $\lambda$ changes in the near-horizon and asymptotically
flat region at 1-loop level to further investigate the imprints of evaporating black holes in scattering
processes. The four-point
vertex function can be computed by functionally differentiating the path integral
over the external current which is linearly coupled to the scalar field. The result of this method is by now
standard and reads
\begin{eqnarray}\label{eq:vertex}
\Gamma^{(4)}(x_i) &=& \mathbf{\imineq{0-loop-lambda-phi4.eps}{4.5}}
+ \frac{3}{2}\left(\mathbf{\imineq{1-loop-lambda-phi4.eps}{5.5}}
+2\mathbf{\imineq{1-loop-lambda-phi4-1.eps}{5.5}}
+\mathbf{\imineq{1-loop-lambda-phi4-2.eps}{5.5}}\right) + \text{O}\big(\lambda^3\big)\,,
\end{eqnarray}
where $i$ runs from $1$ to $4$.
\subsubsection{Near-horizon region: $R \sim r_H$}
The first diagram in the momentum representation is given by $-i\lambda$, whereas the first
1-loop diagram in Eq.~\eqref{eq:vertex} is
\begin{eqnarray}\label{eq:2-loop-coupling}
\mathbf{\imineq{1-loop-lambda-phi4.eps}{5.5}} &=& -i\lambda^2
\int\frac{d^4p}{(2\pi)^4}\left(\frac{i}{p^2 + i\varepsilon}\frac{1}{(q-p)^2 + i\varepsilon}
+\frac{4\pi\delta\big(p^2\big)}{e^{|p_0|/\Theta_\varepsilon}-1}\frac{1}{(q-p)^2 + i\varepsilon}\right),
\end{eqnarray}
where we have chosen the external momenta be non-exceptional
($k_0^i = 0$, $\mathbf{k}^i\mathbf{k}^j = q^2(\delta^{ij} - 1/4)$ for the four external legs, i.e. $i,j = 1,2,3,4$
and $q = k_1 + k_2 =- k_3-k_4$)\footnote{Note that the Mandelshtam variables $s = (k_1+k_2)^2$,
$t = (k_1+k_3)^2$ and $u = (k_1+k_4)^2$ are all equal to $-q^2$ for this choice of the external momenta.
This explains the factor of $3$ in Eq.~\eqref{eq:vertex}.}
not to have extra IR divergence of the logarithmic type that is due to the zero mass of the scalar
field~\cite{Kleinert&Frohlinde}. The origin of this divergence is the same as in quantum electrodynamics.
Specifically, the zero photon mass leads to the non-negligible mutual influence of two charged
particles even when these are at infinite distance from each other. This in turn entails the IR divergence
of the $S$-matrix constructed from the asymptotic particle states.
The first integral in~\eqref{eq:2-loop-coupling} can be evaluated by employing the standard technique
of the dimensional regularisation~\cite{Kleinert&Frohlinde}. The second integral can be simplified after
the integration over the solid angle. The result reads
\begin{eqnarray}
\mathbf{\imineq{1-loop-lambda-phi4.eps}{5.5}} &=& \frac{i\lambda^2}{(4\pi)^2}\left(
\frac{2}{\epsilon} - \gamma + 2 - \ln\Big(\frac{q^2}{4\pi \mu^2}\Big) +
2\int\limits_0^{+\infty}dx\ln\Big|\frac{1+x}{1-x}\Big| \frac{1}{e^{qx/2\Theta_\varepsilon} -1}\right),
\end{eqnarray}
where $\epsilon \rightarrow 0$ in our case, $\gamma$ is the Euler constant and $\mu$ is an arbitrary
mass scale inherent to the dimensional regularisation. The same result holds for the third 1-loop
diagram in~\eqref{eq:vertex} after the substitution $\Theta \rightarrow \theta$.
The second 1-loop diagram in~\eqref{eq:vertex} equals a quarter of the sum of the first and third
1-loop diagram if we set
\begin{eqnarray}
\lambda' &=& \lambda/2^{\frac{1}{2}}\,.
\end{eqnarray}
Then, after the $\overline{\text{MS}}$ renormalisation, we obtain
\begin{eqnarray}\nonumber\label{eq:lambda-far}
\lambda(q,r_H) &=& \lambda + \frac{3\lambda^2}{32\pi^2}\left(\ln\Big(\frac{q^2}{4\pi \mu^2}\Big) -
\int\limits_0^{+\infty}dx\ln\Big|\frac{1+x}{1-x}\Big| \Big(\frac{1}{e^{qx/2\Theta_\varepsilon} -1} +
\frac{1}{e^{qx/2\theta_\varepsilon} -1}\Big)\right) + \text{O}\big(\lambda^3\big)
\\[2mm]
&=&\lambda + \frac{3\lambda^2}{32\pi^2}\left(\ln\Big(\frac{q^2}{4\pi \mu^2}\Big)
- \frac{\pi^2}{3q^2}\big(\Theta^2 + \theta^2\big)\right) + \text{O}\big(\lambda^3,\lambda^2\kappa^4/q^4\big)
\end{eqnarray}
in the regime $q^2 \gg \kappa^2$ which is consistent with our approximation.
Note that, whenever some loop correction depends on the external momenta, the UV divergence is not
cancelled by the fictitious field. We could rearrange the local renormalisation scheme in a manner that
the UV divergence of the
diagram \eqref{eq:2-loop-coupling} is absent, but then the entire 1-loop correction to the coupling
constant $\lambda$ would not depend on the external momenta in the limit $\kappa \rightarrow 0$. This
turns out to be in disagreement
with the well-known result in particle physics, namely $\lambda$ depends on the energy scale at
which one is measuring the coupling constant. For this reason, one had to employ the
dimensional regularisation (or any other standard regularisation) and the
$\overline{\text{MS}}$ renormalisation as well.
Substituting $\Theta$ and $\theta$ in~Eq.~\eqref{eq:lambda-far}, we find
\begin{eqnarray}\label{eq:1lc-cc-nh}
\lambda(q,r_H) &\approx& \lambda + \frac{3\lambda^2}{32\pi^2}\left(\ln\Big(\frac{q^2}{4\pi \mu^2}\Big)
+ \Big(\frac{1}{3} + 2\xi\Big)\frac{1}{4r_H^2q^2}\right).
\end{eqnarray}
Thus, the coupling constant $\lambda$ becomes stronger at 1-loop level in the near-horizon region.
It should be noted that the correction due to the Schwarzschild black hole is given by
the second term in the parenthesis of Eq.~\eqref{eq:1lc-cc-nh}. This is in turn composed of two
contributions. One of these originates from the Hadamard subtraction, while another (that is
proportional to $\xi$) comes from the correction to the propagator that is related to the black-hole
evaporation.
\subsubsection{Far-horizon region: $R \gg r_H$}
Far away from the horizon, $G_U(x,x')$ is given by~\eqref{eq:propagator-far}. The
Hadamard parametrix $H(x,x')$ takes the form at the leading order of the approximation
as if there is no black hole, namely
\begin{eqnarray}
H(x,x') &=& -\frac{1}{8\pi^2\sigma_0(x,x')} + \text{O}\big(\sigma_0\ln\sigma_0,M^2/R^4\big)\,.
\end{eqnarray}
In this case, we find
\begin{eqnarray}\nonumber
\lambda(q,R) &\approx& \lambda + \frac{3\lambda^2}{32\pi^2}\left(\ln\Big(\frac{q^2}{4\pi \mu^2}\Big) -
\frac{27r_H^2}{16R^2}\int\limits_0^{+\infty}dx\ln\Big|\frac{1+x}{1-x}\Big|\frac{1}{e^{qx/2T_H} -1}\right)
+ \text{O}\big(\lambda^3\big)
\\[2mm]
&=&\lambda + \frac{3\lambda^2}{32\pi^2}\left(\ln\Big(\frac{q^2}{4\pi \mu^2}\Big)
- \frac{\xi}{2R^2q^2}\right) + \text{O}\big(\lambda^3,\lambda^2\kappa^2/q^4R^2\big)
\end{eqnarray}
Thus, we reproduce the standard result known in particle physics far away ($R \gg r_H$) from
an evaporating black hole. For a large, but fixed $R$, the vacuum polarisation
induced by the black hole slightly suppresses the coupling constant at one-loop approximation.
\section{Local particle physics}
\label{sec:lpp}
In Sec.~\ref{sec:scalar-model}, we have derived corrections to the self-energy and coupling
constant at 1-loop level near to and far away from the event horizon. In this section, we study their physics.
\subsubsection{Particle physics: Near-horizon region}
We observe no physical particles in the locally Minkowski vacuum $|\Omega\rangle$
near the event-horizon region.
Indeed, one can speak about a \emph{massless} particle as a localised object in the near-horizon region
when its de Broglie wavelength $\lambda_\mathbf{p}$ is much smaller than the size of the event horizon
$r_H$, i.e. $\lambda_\mathbf{p} \ll r_H$. More precisely, $\lambda_\mathbf{p}$ must actually be much
smaller than the size of a particle detector $l_D$ which is in turn much smaller than the horizon size. In
this case, the Wightman function $W_U(x,x')$ approximately coincides with the two-point
function as if there is no black hole
plus a small correction of the order of $(\lambda_\mathbf{p}/r_H)^2 \ll (\lambda_p/l_D)^2 \ll 1$.\footnote{To
our knowledge, this correction
which is due to the ``ingoing" part of the correlation function has not been discussed in the literature. In
the case of eternal black holes, this correction does not appear, because the decreasing (with the distance)
parts of the ``ingoing" and ``outgoing" modes cancel each other.} Therefore, the state $|\Omega\rangle$
is not populated by the real particles. The same holds in the stationary frame near the
event horizon as the notion of particle is covariant.
There has been recently argued that an in-falling observer should discover a firewall (a sort of cloud of
the high-energy (blue-shifted) Hawking particles $|\tilde{\psi}\rangle$) in the near-horizon
region~\cite{Almheiri&Marolf&Polchinski&Sully}. If this is a real phenomenon, the equivalence principle
does not hold, because the event horizon of evaporating black holes would then physically be a
distinguishable set of space-time points. If so, the whole framework of general relativity which led to the
notion of black hole would be not reliable. Importantly, no evidences have been found so far that the
principle of equivalence does not hold.
The equivalence principle was sacrificed in favour of the unitarity~\cite{Almheiri&Marolf&Polchinski&Sully}.
By the unitarity one should here understand the existence of the unitary $S$-matrix between the in-state
and the thermal out-state. The in-state in our notations is identified with the vacuum $|\Omega\rangle$,
while the out-state corresponds to the Boulware vacuum $|\tilde{\Omega}\rangle$. However, the mathematical
subtlety is that the in-state can only \emph{formally} be represented as a thermally populated state of the
Hawking particles defined with respect to the out-state. The relation between the in-state and out-state is
formal, because these do not define unitarily equivalent representations of the field operator
algebra~\cite{Emelyanov-2015-2}. Thus, this means that if one demands that there exists a unitary operator
$\hat{S}$ that relates the in-state and the thermally populated out-state, then one comes to the idea of
having the firewall near the event horizon. The problem is that the existence of the unitary operator $\hat{S}$
is not consistent with the principles of local quantum field theory as pointed out in~\cite{Emelyanov-2015-2}.
We have found above that the Feynman propagator $G_U(x,x')$ in the near-horizon region is given by a
thermal-like propagator with the imaginary ``temperature" $\Theta$ given in~\eqref{eq:imaginary-temperature}.
We interpret this small correction as being due to the modification of the field operator in the presence of
black holes. A similar effect occurs in the Casimir set-up. Indeed, if one considers a wave packet of a photon
of the de Broglie wavelength $\lambda_\mathbf{p} \ll d$, where $d$ is a distance between the conducting plates,
then the photon propagator turns out to be as a thermal-like one (for the modes in the perpendicular direction
with respect to the plates) with an imaginary ``temperature" $T_C = i/2d$ when localised far from the
plates~\cite{Latorre&Pascual&Tarrach}. The ``thermal" term in this propagator is understood as being due to
the boundary conditions satisfied by the electromagnetic operators or the vacuum fluctuations
which are present in the Minkowski vacuum.\footnote{The very existence of $T_C$ in the Casimir effect
can be envisaged from the Tomita-Takesaki theorem~\cite{Haag}. Indeed, according to the theorem
the operator algebra composed of the electromagnetic field operators in-between the conducting plates must
satisfy the Kubo-Martin-Schwinger condition in the Minkowski vacuum with respect to a certain one
parameter group of automorphism of the quantum operators. This group can in general be of a geometrical
as well as non-geometrical origin. In the present case, it is a symmetry related with the periodicity of the
operators in the spatial direction which is transferred to the periodicity in the Minkowski time with a real
period, which corresponds to the imaginary ``temperature". In a private discussion with Bernard Kay I got
to know about a representation of the Minkowski vacuum as an ``imaginary-temperature state" in-between
the conducting plates, which was found in~\cite{Kay}.}
We want now to discuss the vacuum expectation value of the particle number operator in $|\Omega\rangle$.
This operator is defined as
\begin{eqnarray}\label{eq:number-operator-physical}
\hat{N}(h) &=& \hat{a}^\dagger(h)\hat{a}(h)\,,
\end{eqnarray}
where the wave packet $h(x)$ is the same as~\eqref{eq:wave-packet}, but for the scalar particle
and $h(p)$ having a maximum near the momentum $q$. Due to the thermal-like term of the imaginary
``temperature" $\Theta$ in $G_U(x,x')$, the (imaginary) quantity $\langle \hat{N}(h) \rangle$ does
\emph{not} vanish. However, it is in general true that
the localised operators have a non-vanishing vacuum expectation value in the Minkowski vacuum
due to its Reeh-Schlieder property~\cite{Haag}.\footnote{As a consequence of this property, a
sufficiently sensitive thermometer should measure a non-zero ``temperature" of the vacuum as being
a local operator. For the same reason, the particle detector is excited all the time by ``particles"
in the vacuum. The ``temperature" and ``particles" of the vacuum are merely a quantum noise.}
The physical interpretation of this mathematical theorem is given in terms
of the quantum fluctuations. One needs thus to subtract these or calibrate the particle
detector~\cite{Yngvason}. If we do the same in the near-horizon region, we obtain
$\langle\hat{N}(h)\rangle = 0$.\footnote{We do not consider the imaginary
``temperature" $\Theta$ as being of any fundamental meaning,
rather than a footprint of our approach. We further study the near-horizon physics by employing
quantum kinetic theory in~\cite{Emelyanov-2017a}, wherein we derive the 2-point function
in the local inertial frame at $R \sim r_H$ without any reference to $\Theta$.}
\subsubsection{Particle physics: Far-horizon region}
The Feynman propagator $G_U(x,x')$ far away from the black hole ($R \gg r_H$) is on the contrary
given by the propagator with a thermal-like term at the Hawking temperature $T_H$. This extra term
vanishes as $1/R^2$ at $R \rightarrow \infty$ and can be assigned to the modification of the field
operator in the presence of evaporating black holes.
Utilising the number operator introduced in Eq.~\eqref{eq:number-operator-physical}, one can define a
number density operator. Its vacuum expectation value in $|\Omega\rangle$ for the ``outgoing"
plane-wave modes within the frequency range from $\omega$ to $\omega +d\omega$ is given by
\begin{eqnarray}\
dn_\omega &=& \langle \hat{n}_\omega \rangle\,d\mu_\omega\,, \quad \text{where} \quad
\mu_\omega \;=\; \frac{\omega^2d\omega}{2\pi^2}
\end{eqnarray}
is the standard measure of integration, and the distribution of the modes in the frequency interval
$(\omega,\omega + d\omega)$ is given by
\begin{eqnarray}\label{eq:number-local}
\langle \hat{n}_\omega \rangle &\approx& \frac{1}{4\omega^2R^2}\,\frac{\Gamma_\omega}{e^{\omega/T_H} - 1}\,,
\end{eqnarray}
where $\Gamma_\omega$ has been given in Eq.~\eqref{eq:gamma}.
It is worth pointing out that the right-hand side of~\eqref{eq:number-local} is fully due to
$\vec{W}_\beta(x,x')$ or the field operator $\hat{\Phi}_b(x) = \hat{b}(x) + \hat{b}^\dagger(x)$.
The prefactor $1/4\omega^2R^2$ also appears in the \emph{effective} covariant Wigner function
$\mathcal{W}_\text{eff}(x,p)$ playing a role of the phase-space distribution function in relativistic
kinetic theory (see, e.g.,~\cite{deGroot&vanLeeuwen&vanWeert}), namely
\begin{eqnarray}
\mathcal{W}_\text{eff}(x,k) &\approx& \frac{1}{(2\pi)^3}\frac{\Gamma_\omega}{4\omega^3R^2}\frac{\delta(\omega - k)}
{e^{\omega/T_H}-1} \;=\; f_\text{eff}(x,k)\,\frac{\delta(\omega - k)}{\omega}\,,
\end{eqnarray}
where $\omega = k_0 = k = |\mathbf{k}|$ with $\mathbf{k}^i = (k,0,0)$ and the index $i$ runs over
$\{r,\theta,\phi\}$. The effective function $f_\text{eff}(x,k)$ is known as the one-particle Wigner distribution. We
reproduce the result~\eqref{eq:number-local} by employing the standard formula known in kinetic
theory:
\begin{eqnarray}
n &=& \int d^3p \,f_\text{eff}(x,p) \;=\; \int dn_\omega \;=\; \int d\mu_\omega \,\langle \hat{n}_\omega \rangle\,.
\end{eqnarray}
Moreover, we can compute the outward positive energy flux as found in~\cite{Christensen&Fulling,Candelas}
as the second moment (with respect to the momentum) of the distribution
function~\cite{deGroot&vanLeeuwen&vanWeert}, namely
\begin{eqnarray}
\int d^3p\,p f_\text{eff}(x,p) &=& \frac{L}{4\pi R^2}\,, \quad
\text{where} \quad L \;=\; \frac{1}{2\pi}\int d\omega\,\frac{\omega\,\Gamma_\omega}{e^{\omega/T_H} - 1}
\end{eqnarray}
is the luminosity. We further study local quantum physics near the event
horizon by employing quantum kinetic theory in~\cite{Emelyanov-2017a}.
The equation~\eqref{eq:number-local}, however, differs from the distribution of the Hawking
modes~\cite{Hawking}:
\begin{eqnarray}\label{eq:number-global}
\langle \hat{n}_\omega \rangle_H &=& \frac{\Gamma_\omega}{e^{\omega/T_H} - 1}\,.
\end{eqnarray}
By now the equation~\eqref{eq:number-global} is a widely-accepted result in black-hole physics.
This has been derived by computing the Bogolyubov coefficients relating $\hat{b}(x)$ with $\hat{a}_<(x)$
and $\hat{a}_<^\dagger(x)$~\cite{Hawking-1}. The reason of the discrepancy is that
the formula~\eqref{eq:number-local} is
\emph{local} and expected to be valid only in a volume of the size being much smaller than $R \gg r_H$,
whereas~\eqref{eq:number-global} is a global result. Thus, the physical meaning of~\eqref{eq:number-local}
and~\eqref{eq:number-global} is different. Specifically, the number of the locally plane-wave modes in a
spherical shell of volume $dV = 4\pi R^2dR$ for the fixed distance $R \gg r_H$ from the black hole is given
by
\begin{eqnarray}
dN &=& n dV \;=\; 4\pi n R^2dR \;=\; 4\pi n R^2(dR/d\tau)d\tau \;=\; 4\pi n R^2d\tau\,,
\end{eqnarray}
where we have set $dR/d\tau = c = 1$. Therefore, we obtain the flux of these modes:
\begin{eqnarray}
\dot{N} &=& \frac{1}{2\pi}\int d\omega\,\langle \hat{n}_\omega \rangle_H\,
\end{eqnarray}
(or the number of the modes per the radial distance $dR$, i.e. $dN/dR$). It is worth noticing that the
number of these modes in a detector of volume $l_D{\times}l_D{\times}l_D$ drops out with the
distance as $(l_D/R)^2$ for $R \gg r_H$. Analogously, we rederive a well-known result
\begin{eqnarray}
\dot{E} &=& \frac{1}{2\pi}\int d\omega\,\omega\,\langle \hat{n}_\omega \rangle_H
\end{eqnarray}
from the local distribution~\eqref{eq:number-local} by computing the energy in the spherical
shell of the volume $dV = 4\pi R^2dR$ or $dV = 4\pi R^2d\tau$. This implies that~\eqref{eq:correlator-far}
and~\eqref{eq:propagator-far} are good approximations to the exact Wightman function and Feynman
propagator whenever the conditions $|r-r'| \ll R$ and $r_H \ll R$ are satisfied.
All local observables are composed of the fundamental field operator $\hat{\Phi}(x)$. For instance, the
particle creation operator is given by Eq.~\eqref{eq:particle-creation-operator} and this is employed
to construct the particle number operator. The field operator $\hat{\Phi}(x)$ can in turn be
represented as the sum $\hat{\Phi}_>(x) + \hat{\Phi}_b(x)$ (see Eq.~\eqref{eq:scala-operator-splitting}).
As emphasised above, the Wigner distribution $\mathcal{W}(x,k)$ is due to the part $\hat{\Phi}_b(x)$
of the field operator $\hat{\Phi}(x)$. In the asymptotically flat region, local physics is oblivious to the
presence of black holes, because the Wigner distribution drops out as $(r_H/R)^2$ for
$R \rightarrow \infty$. This means that if we choose the standard wave packet $h(x)$ in the spatial
infinity ($R \rightarrow \infty$) as we have been doing that on earth when we study scattering processes
in particle physics, we then find
\begin{eqnarray}
\hat{a}^\dagger(h) &=& \hat{a}_>^\dagger(h)\,,
\end{eqnarray}
where $\hat{a}_>^\dagger(x)$ is the standard creation operator of the scalar particle in Minkowski space-time
and the vacuum $|\Omega\rangle$ in the spatial infinity ``reduces" to the Minkowski
vacuum (in the sense that the local operators probe $|\Omega\rangle$ at $R \rightarrow \infty$
as the Minkowski vacuum in Minkowski space-time), such that $\hat{a}_>(h)|\Omega\rangle = 0$. This
implies that the \emph{effective} density matrix\footnote{Note that the description of the vacuum
$|\Omega\rangle$ in terms of the thermal density matrix comes from the assumption that the
probes of this vacuum in the spatial infinity ($R \rightarrow \infty$) are performed \emph{only}
by the operators composed of $\hat{b}(x)$ and $\hat{b}^\dagger(x)$~\cite{Hawking}. However, this turns
out not to be the case, because $\hat{\Phi}(x)$ reduces to $\hat{\Phi}_>(x)$, rather than $\hat{\Phi}_b(x)$ in
the spatial infinity. The effective density matrix characterises the quantum
fluctuations of the $\hat{\Phi}_b$-part of the field $\hat{\Phi}(x)$, which is locally irrelevant at
$R \gg r_H$.} introduced in black-hole physics must
actually decrease with the distance as $(r_H/R)^2$, such that the vacuum $|\Omega\rangle$ is probed
at $R \rightarrow \infty$ as being pure, but as if it is mixed at finite $R \gg r_H$. It implies that the formal
representation of the vacuum $|\Omega\rangle$ as the thermally populated state of the particles defined
with respect to $|\tilde{\Omega}\rangle$ in the spatial infinity is \emph{not} a self-consistent picture of
the black-hole evaporation.
Indeed, if we consider a wave packet $h(\mathbf{p}_n)$ (one of the elements of the countable set of
orthonormalised functions) with a definite value of the momentum $\mathbf{p}_n$
localised in a ball of volume $\sigma \sim (\lambda_\mathbf{p})^3$ at the distance $R \gg r_H$ from
an evaporating black hole, then the vacuum expectation value of the operator $\hat{N}(h)$
defined in Eq.~\eqref{eq:number-operator-physical} is given by
\begin{eqnarray}
\langle \hat{N}_{\mathbf{p}_n}\rangle &=& \frac{27r_H^2}{16R^2}\,\frac{1}{e^{\omega_n/T_H} - 1}
\;\rightarrow\; 0 \quad \text{for} \quad R \;\rightarrow\; \infty\,,
\end{eqnarray}
where $\omega_n = |\mathbf{p}_n|$. However, if we take a wave packet $h(\omega_n| lm)$
which corresponds to a spherical wave of the frequency $\omega_n$,
the orbital number $l$ and the magnetic number $m$ localised around radial distance $R \gg r_H$,
we obtain
\begin{eqnarray}
\langle \hat{N}_{\omega_n lm}\rangle &=& \frac{\Gamma_{\omega_n}}{e^{\omega_n/T_H} - 1}
\;\approx\; \frac{1}{4}\,\frac{27r_H^2\,\omega_n^2}{e^{\omega_n/T_H} - 1}\,.
\end{eqnarray}
This does \emph{not} depend on the radial distance $R$, because the support of the spherical shell
scales as $R^2$. Therefore, the Hawking particles are associated with the spherical waves. The
spherical waves are not localised in the angular directions. For this
reason, one cannot understand these as localised excitations, which one can put, for instance, in
a box of the one-cubic-meter size.
As shown above, the quantum operator $\hat{\Phi}(x)$ can be represented as $\hat{\Phi}_>(x) + \hat{\Phi}_b(x)$
above the horizon ($R > r_H$). The operators $\hat{\Phi}_>(x)$ and $\hat{\Phi}_b(x')$ commute with
each other for any points $x$ and $x'$, i.e.
\begin{eqnarray}
[\hat{\Phi}_>(x),\hat{\Phi}_b(x')] &=& 0\, \quad \text{for} \quad \forall\;\; x,x'\,.
\end{eqnarray}
In the asymptotically flat region, the operator $\hat{\Phi}_b(x')$ drops out to zero as $r_H/R$. It seems
this means that the matter outside of the hole is composed only of the operator $\hat{\Phi}_>(x)$
at $R \rightarrow \infty$ and \emph{cannot} be used to \emph{directly} discover the Hawking modes
which are due to $\hat{\Phi}_b(x)$. It is still possible to discover these indirectly through its gravitational
influence, because $\langle \hat{T}_{\nu}^{\mu}\rangle \neq 0$. Thus, these could be a sort of ``the dark
radiation". Another argument in favour of this idea is the following: If we prepare a thermal gas in a small
box at $R \rightarrow \infty$ and let it fall towards the horizon, then the energy-momentum tensor will be
finite at $R = r_H$ within the box. This is \emph{not} the case for the Hawking gas, because the
energy-momentum tensor will diverge for any temperature $T \neq T_H$ on the horizon and the whole
consideration becomes self-inconsistent (unless one starts to treat this gedankenexperiment at the level
of non-perturbative quantum gravity). Therefore, one might conclude that ``the Hawking
matter" is decoupled from
the normal matter. However, this statement makes sense only if one assumes that the
``outgoing" modes correspond to real particles which can be literally used to prepare the Hawking
gas heated up to any temperature $T$. This decoupling does not make any sense, if one understands
these as virtual particles/vacuum fluctuations, because the fundamental field is $\hat{\Phi}(x)$, rather than
$\hat{\Phi}_>(x)$ or $\hat{\Phi}_b(x)$ separately.
If we calibrate the particle detector (as made above at $R \sim r_H$) or subtract the vacuum contribution to
the Wigner function at large, but fixed $R$, then we obtain $|\Omega\rangle$ is empty. It does
not imply the vanishing energy-momentum tensor, i.e. $\langle \hat{T}_\nu^\mu \rangle \neq 0$.
For the same reason, the Fourier transform of the Wightman function with respect to the time (this
represents the Unruh-DeWitt detector) is also non-vanishing (although $|\Omega\rangle$ does not
contain the physical particles), because this yields the frequency spectrum of the vacuum fluctuations
(according to the Wiener-Khinchin theorem) as pointed out in~\cite{Candelas}.
Since we have been trying to interpret the black-hole evaporation as the vacuum would possess
(inhomogeneous, but isotropic) medium-like properties (because this interpretation seems to be
self-consistent and does not suffer from the absence of the unitary $S$-matrix as understood by
many researchers as well as the firewall problem), it is of interest to be aware of any other physical
examples when this kind of the viewpoint is fruitful. There exists at least one to our
knowledge. Specifically, the propagation of photons in the Minkowski vacuum with a super-strong
magnetic field $B$ ($\,\gtrsim \pi m_e^2/\alpha e$, where $\alpha$ is the fine structure constant
and $e$ the elementary charge) occurs as if the photons move through a magnetised physical
plasma, i.e. in the plasma held at the external magnetic field~\cite{Melrose&Stoneham,Dittrich&Gies}.
\section{Concluding remarks}
\label{sec:concluding remarks}
\subsection{Particles in black-hole geometry}
We have proposed a new, covariant definition of the notion of particle in curved space-time
which is observer-independent. This definition is mostly motivated by the success of particle physics we have
been testing on earth and is consistent with the particle creation effect in expanding universe~\cite{Parker}
(see for a recent short review~\cite{Parker-1}).
We have found that the term in the 2-point function that is (partially) responsible for the black-hole
evaporation is of the sub-leading
order with respect to the term providing the correct pole structure of the Feynman propagator. Namely,
this hierarchy of the terms is regulated by the ratio $(\lambda_\mathbf{p}/r_H)^2$ near the horizon, where
$\lambda_\mathbf{p}$ is a de Broglie wavelength of the scalar particle of momentum $\mathbf{p}$. The
modes leading to this correction should thus be understood as vacuum fluctuations. In the far-horizon region,
this suppression is even stronger: $(\lambda_\mathbf{p}/R)^2$ for $R \gg r_H$. This implies that the black-hole
evaporation originating in the near-horizon region cannot be understood as a local effect that agrees
with~\cite{Bardeen}.
The flux of the energy density changes its direction at the distance $R \sim 3M$ outside of the
event horizon~\cite{Unruh-1,Giddings}. Hence, it might imply that the sub-leading term in the propagator
is associated with the flux of the negative energy density inside the black-hole horizon
(see~\cite{Emelyanov-2017a}). Thus, we come to a conclusion that the Hawking's partner
mode is a vacuum fluctuation. The same observation based on a different argument has been
recently made in~\cite{Hotta&Schuetzhold&Unruh}.
\subsection{One-loop correction to self-energy}
The tadpole diagram yields the effective mass of the self-interacting scalar field $\Phi(x)$. Employing
the effective action approach for computing the 1-loop correction to the scalar field equation, one can
obtain the well-known result $m_\Phi^2 = \frac{\lambda}{2}\langle\hat{\Phi}^2(x)\rangle$. The Wick squared
operator $\hat{\Phi}^2(x)$ in the Unruh vacuum was derived in~\cite{Candelas}.
Hence, the value of the effective scalar mass at one-loop approximation is a straightforwardly computable
quantity.
Perhaps, the less trivial computation is to reproduce the result for $m_\Phi^2$ by applying Feynman's
method in the freely-falling frame. The non-trivial part of this computation is how to ``properly" renormalise
the ultraviolet divergence of the tadpole diagram. As a guideline, we have used the above mentioned
findings for the Wick squared operator.
This can be considered as an intermediate step to compute other scattering reactions
with the radiative corrections.
\subsection{One-loop correction to coupling constant}
A presence of black holes entails the modification of quantum field propagators. This
in turn leads to the non-trivial corrections to the self-energy and coupling constant in the loop
expansion. The latter gets a loop contribution depending on the external momenta $q \gg \kappa$.
We have found that the 1-loop correction to $\lambda(q)$ in the near-horizon region
$R \sim r_H$ is larger than that in the absence of the black hole, while this reduces to the standard
result found in the Minkowski-space approximation in the asymptotically flat region
$R \gg r_H$.\footnote{We have derived in~\cite{Emelyanov-2017a} the higher-order corrections
to the 2-point function in $\Delta\mathbf{x}$ up to the second order in both the far-horizon and
near-horizon region. These corrections give a contribution to the running coupling constant
$\lambda(q)$ of the order of $(1+3\cos^2\gamma)\,T_H^4/q^4$, where $\gamma$ is an angle between
$\mathbf{q}$ and $\mathbf{n} = \mathbf{R}/R$. This is much smaller than the
leading order term found above in the regime $q \gg T_H$.}
\subsection{One-loop correction to vacuum energy}
The energy-momentum tensor of the scalar field gets a correction due to the self-coupling.
The one-loop contribution to the vacuum energy is given by the standard method of the
perturbation theory, namely
\begin{eqnarray}\label{eq:loc-emt}
\langle\Delta\hat{T}_{\mu\nu}\rangle &=& \frac{\lambda}{4!}\,\eta_{\mu\nu}\langle\hat{\Phi}^4\rangle \;=\;
\frac{\lambda}{8}\,\eta_{\mu\nu}\langle\hat{\Phi}^2\rangle^2\,.
\end{eqnarray}
The same result can be reproduced through taking into account vacuum bubbles at two-loop level,
namely
\begin{eqnarray}\nonumber\label{eq:emt-one-loop-correction}
\langle\Delta\hat{T}_{\mu\nu}\rangle &=& \frac{i}{8V_4}\,\eta_{\mu\nu}\left(
\mathbf{\imineq{2-loop-vacuum-diagram-1.eps}{6.0}} +
2\mathbf{\imineq{2-loop-vacuum-diagram-2.eps}{6.0}} +
\mathbf{\imineq{2-loop-vacuum-diagram-3.eps}{6.0}}\right)
\\[2mm]
&=& \frac{\lambda}{8}\,\eta_{\mu\nu}\Big(G_U(x,x) - G_H(x,x)\Big)^2 \;=\;
\frac{\lambda}{8}\,\eta_{\mu\nu}\langle\hat{\Phi}^2\rangle^2\
\end{eqnarray}
according to Eqs. \eqref{eq:smws} and \eqref{eq:effective-mass-diagram},
where $V_4 = \int d^4x$ is a four-dimensional volume of
a local Minkowski frame as if it is infinitely large.
Thus, we find that the 2-loop correction to the vacuum energy after having been renormalised is finite
at $R \sim r_H$ in the freely-falling frame. In the asymptotically flat region, the 2-loop correction to the
vacuum energy is of the order of $1/R^4$. This is much smaller than the 1-loop contribution that
vanishes as $1/r_H^2R^2$ in the limit $R \rightarrow \infty$~\cite{Christensen&Fulling,Candelas}.
\subsection{Local renormalisation scheme}
A sort of ambiguity is inherent to the local renormalisation scheme we have proposed. One could choose
the fictitious field $\phi(x)$ be anticommuting (instead of commuting) and with a \emph{positive} norm
(with the correct sign in front of the propagator). This would change only the 1-loop correction to the
coupling constant which depends on the external momenta, while \eqref{eq:effective-mass-lrs} and
\eqref{eq:emt-one-loop-correction} are insensitive to these modifications. Our choice is, however,
symmetric with respect to how the propagators $G_U(x,x')$ and $G_H(x,x')$ contribute to the coupling
constant $\lambda$ (at least) at one-loop level.
In addition, the coupling constant $\lambda$ in the thermal state of temperature $T$ without introducing
the fictitious field $\phi(x)$ would be
\begin{eqnarray}
\lambda(q,T) &=& \lambda + \frac{3\lambda^2}{32\pi^2}\left(\ln\Big(\frac{q^2}{4\pi \mu^2}\Big)
- \frac{2\pi^2}{3q^2}\,T^2\right) + \text{O}\big(\lambda^3,\lambda^2T^4/q^4\big).
\end{eqnarray}
Thus, the temperature-dependent correction is double of what one finds if the fictitious field is taken
into account. However, it will be the same result if we also heat the fictitious field up to the temperature $T$.
In the case of a massive $\lambda\Phi^4$-theory, the one-loop correction to the scalar self-energy
after having been regularised by one of the standard methods (dimensional or Pauli-Villars regularisation)
is non-vanishing after the subtraction of the UV-divergent term. However, the extra UV-divergent term
($\,\propto m^2\ln\sigma_0(x,x')$) of $\langle\hat{\Phi}(x)\hat{\Phi}(x')\rangle$ in the limit $x' \rightarrow x$ is
precisely cancelled by the same term in the Hadamard parametrix. Thus, the mass of the scalar field
does not get a quantum correction at one-loop level in this renormalisation approach.
The UV-divergent part of the stress tensor in Minkowski space is associated with the
vacuum bubbles. One usually ignores this in particle physics as these do not show up for scattering
processes (where one measures the energy differences only). This is illegitimate when one takes
gravity into account. One of the methods to compute the renormalised stress tensor is to subtract
the Hadamard parametrix from the Wightman function (see, e.g.,~\cite{Decanini&Folacci} for a brief
review).
It is a well-known problem in standard electroweak theory, that the quantum/loop contributions to
the self-energy of the Higgs field are (polynomially) divergent. That is the origin of the hierarchy
problem. In this case, the one-loop self-energy term depends on the external momentum of the
Higgs particle. If we do not introduce the fictitious field in the loop diagrams containing the external
momenta, then we recover the standard results.
To summarise, we should either introduce the fictitious field only for the purpose to renormalise
expectation values of local quantum operators in Hadamard states or demand that the fictitious field
also appears in the loop diagrams depending on the external momenta. The consequences of the
latter should however be investigated in detail to draw any decisive conclusions.
\subsection{Locally Minkowski and Unruh vacuum}
These vacua are indistinguishable in the far-from-horizon region at the leading order of the
approximation. The deviation of the Unruh vacuum from the locally Minkowski one $|\Omega\rangle$
can be, however, established near the event horizon, taking into account that these vacua are
characterised by the 2-point functions $W_U(x,x')$ and $W(x,x')$, respectively.
The singular part of $W_U(x,x')$ at $r \sim r_H$ is given by $\vec{W}_{\beta}(x,x')$ (see
Eq.~\eqref{eq:unruh-2-point-function}). It is inversely proportional to
$\cosh(\kappa\Delta{t}_S) - \cosh(\kappa(2\bar{\sigma}(\mathbf{x},\mathbf{x}'))^\frac{1}{2})$, where
$\kappa = \frac{1}{2}f'(r_H)$. This can be transformed to the Minkowski form $\sigma_0(x,x')$ following
the procedure outlined in Sec.~\ref{sec:olctse-nhr}, where, e.g., the new time coordinate is approximately
equal to $f^\frac{1}{2}(r)\sinh(\kappa t_S)/\kappa$.
The singular part of $W(x,x')$ is approximately given by $-1/(8\pi^2\sigma(x,x'))$, where $\sigma(x,x')$
is a geodetic distance between the space-time points $x$ and $x'$. This can be expressed in terms of
$\sigma_0(x,x') \propto \cosh(\kappa\Delta{t}_S) - \cosh(\kappa(2\bar{\sigma}(\mathbf{x},\mathbf{x}'))^\frac{1}{2})$
as found in Eq.~\eqref{eq:gdnh}. However, in terms of the Riemann normal coordinates,
$\sigma(x,x')$ acquires a Minkowski form as well, but at a fixed spatial point at $r \sim r_H$, the Riemann
normal time reads
\begin{eqnarray}
\Delta{y}^0 &\approx& \frac{f^\frac{1}{2}(r)}{\frac{1}{2}f'(r)}\,\sinh\left(\frac{1}{2}f'(r)\Delta{t}_S\right).
\end{eqnarray}
Thus, the Unruh vacuum differs from the locally Minkowski vacuum in the near-horizon region,
because $\kappa = \frac{1}{2}f'(r_H)$ and $\frac{1}{2}f'(r)$ coincide in the limit $r \rightarrow r_H$
(implying $f(r) \rightarrow 0$) only. We shall determine $W(x,x')$ at any radial distance from the event horizon
elsewhere.
\section*
ACKNOWLEDGMENTS}
I am thankful to Frans Klinkhamer, Jos\'{e} Queiruga and Frasher Loshaj for discussions and their
comments on an early version of this paper. It is also a pleasure to thank Eduardo Grossi for the
reference~\cite{deGroot&vanLeeuwen&vanWeert}.
|
\section{Introduction}\label{sec:introduction}
The response of electromagnetic materials to applied fields is spatially nonlocal, i.e.\ the response at any point depends on the value of the fields throughout a neighboring region\cite{LLcm,rukhadze1961}. This nonlocal response, or spatial dispersion, can be neglected in many regimes of interest, even when frequency dispersion is significant. Nevertheless, there are cases where nonlocal response is a significant factor and interest in this topic has increased in recent years. Metallic nano-structures have been developed whose properties can be accurately predicted only by including their nonlocal response\cite{raz11,wie12,pen12,tos15,sch16}. Near-field radiative heat transfer between materials is also modified by spatial dispersion\cite{sin15a,sin15b}. More generally, thermal and zero-point electromagnetic energy in the presence of materials is significantly affected by spatial dispersion\cite{henkel2006,nar10,hor14}. This has implications for spontaneous emission rates of emitters inside materials or placed close to surfaces\cite{pur46,dre68,bar98,Novotny,hor14}, and also for thermal and zero-point forces on curved boundaries\cite{hor14}. There are thus many interesting questions, some of them quite basic, that require a proper account of spatial dispersion.
In this paper we employ the macroscopic Maxwell equations to explore boundary effects in nonlocal media. We extend previous results on reflection and transmission at planar boundaries to the most general isotropic spatially dispersive dielectric. The key extra ingredient here is to allow for both transverse and longitudinal susceptibilities that have different values\cite{LLcm,rukhadze1961} (see below). We also show how the spectral energy density of thermal and zero-point radiation depends on the material susceptibilities in this general case.
The electric susceptibility $\chi$ is frequently described by a damped-oscillator model, in which nonlocal response may be incorporated by a simple wave-vector dependence\cite{hopfieldthomas} :
\begin{align}
\chi({\bm k},\omega)
=
\chi_0
+
\frac{\omega_p^2}{\omega_T^2+\sigma^2k^2-\omega^2-i\gamma\omega}.
\label{eq:susceptibility_basic}
\end{align}
Here $\omega_T$ is the resonant frequency, $\gamma$ quantifies the absorption, $\omega_p$ is the oscillator strength, and $\sigma$ is a spatial-dispersion parameter. The term $\chi_0$ collects contributions from other resonances and acts as a background susceptibility. A more complete model would dispense with $\chi_0$ and include additional resonance terms. A justification of (\ref{eq:susceptibility_basic}) based on properties of semi-conductors was given by Hopfield and Thomas~\cite{hopfieldthomas} but it can also be derived from a simple classical model~\cite{chu16}. In the local case $\sigma=0$ the usual Maxwell boundary conditions, namely the continuity of the tangential components of ${\bm E}$ and ${\bm H}$ and the normal components of ${\bm D}$ and ${\bm B}$, are sufficient to calculate the reflection coefficient at a sharp boundary. However, the introduction of a nonlocal term as in (\ref{eq:susceptibility_basic}) leads to the presence of two transverse and one longitudinal wave inside the medium\cite{rukhadze1961}.
The usual Maxwell boundary conditions are now insufficient to solve for the four unknown amplitudes of the transmitted and reflected waves. Additional information is required about the relationship between the amplitudes. Many authors have expressed this as additional boundary conditions (ABC's) on the polarization ${\bm P}$ of the medium associated with the spatially dispersive resonance\cite{agarwal1971a,agarwal1971b,agarwal1972,birman1972,agarwal1973,maradudin1973,birman1974,mills1974,foley1975,bishop1976,ting1975,kliewer1968,kliewer1971,kliewer1975,ruppin1981,rimbey1974,rimbey1975,rimbey1976,rimbey1977,rimbey1978,pekar1958a,pekar1958b,pekar1958c,pekar1959}
(i.e.\ without the background term $\chi_0$).
Each ABC is motivated by the type of medium considered, as we describe in more detail in Sec. \ref{sec:p-polarization}. Several of these authors note that ABCs are equivalent to introducing a phenomenological scattering term to the susceptibility in the presence of a boundary, in order to describe the behavior of the medium at the surface. The denominator of the second term in (\ref{eq:susceptibility_basic}) has zeros that correspond to the dispersion relation for waves of polarization $\bm{P}$ in the material. These polarization waves are reflected at the surface and complex parameters are introduced to serve as the corresponding reflection coefficients. Halevi and Fuchs\cite{Halevi} incorporated all previous examples of these extra parameters into values of a set $U_i$ ($i\in\{x,y,z\}$) of reflection coefficients for the polarization waves (see below). They then derived a general expression for the electromagnetic reflection coefficients at the boundary in terms of arbitrary complex $U_i$, in the case of a dielectric whose bulk susceptibility is the scalar $\chi({\bm k},\omega)$ of the form (\ref{eq:susceptibility_basic}).
However, in the presence of spatial dispersion the susceptibility is a tensor, as the wave-vector generates a distinctive direction\cite{LLcm}. In a homogeneous, isotropic, non-gyroscopic medium:
\begin{align}
\chi_{ij}({\bm k},\omega)
=
\delta_{ij}
\chi_{\perp}({\bm k},\omega)
+
\frac{k_ik_j}{k^2}
\left[
\chi_{\parallel}({\bm k},\omega)
-
\chi_{\perp}({\bm k},\omega)
\right],
\label{eq:susc}
\end{align}
where $\perp$ and $\parallel$ denote transverse and longitudinal terms, respectively. If one assumes $\chi_{\parallel}=\chi_{\perp}$ then the susceptibility (\ref{eq:susc}) is essentially still a scalar, but the most general isotropic susceptibility is a tensor in the nonlocal case.
In comparison to the simplified scalar case, the tensor nature of the susceptibility is generally overlooked.
While Rimbey and Mahan include this in their calculation for a specific ABC\cite{rimbey1974}, their choice of $U_i$ leads to the absence of a longitudinal wave in the medium.
Garcia-Moliner and Flores\cite{garcia1977} derive the reflection coefficient in the tensor case, but they restrict themselves to a scalar $U$ and obtain a result in an integral form.
The first aim of this paper is to extend Halevi and Fuchs' derivation\cite{Halevi} to the tensor susceptibility (\ref{eq:susc}), with $\chi_{\parallel}\neq\chi_{\perp}$, where $\chi_{\parallel}$ and $\chi_{\perp}$ each have the form (\ref{eq:susceptibility_basic}). We derive a general expression for reflection and transmission coefficients at a planar boundary, allowing for arbitrary complex reflection coefficients of the polarization waves at the surface.
Our second goal is to use the general electromagnetic reflection coefficients derived in the first part of the paper to calculate the spectral energy density of thermal and zero-point radiation outside the boundary of the spatially dispersive medium. It is well known that the result for a local medium is proportional to $1/z^3$ close to the surface\cite{candelas1982,henkel2000,joulain2005}. Note that the divergence of the spectral energy on the boundary also occurs for purely thermal radiation (dropping the zero-point part) so it is not due to the divergence of (total) vacuum energy. This unphysical divergence can be removed by introducing a cutoff wave-vector \cite{candelas1982} based on the interatomic or lattice spacing. But a more accurate picture is obtained by the inclusion of spatial dispersion, which must naturally remove the divergence without the need for additional modifications to the calculation. This has been shown\cite{henkel2006} to work for the Lindhard susceptibility of a plasma, which can be used to model the response of the conduction electrons in a metal. The susceptibility due to the core electrons, however, will still lead to a divergence if it is taken to be local. We will show that the general reflection coefficients derived here give a finite thermal and zero-point spectral energy density at a planar boundary. Moreover, we find that the difference $\chi_{\parallel}-\chi_{\perp}$ between the transverse and longitudinal susceptibilities has a large effect on the spectral energy close to the boundary. The influence of a metal boundary on the spectral thermal energy density has been measured using near-field microscopy\cite{wilde06}.
As our treatment is based on macroscopic electromagnetism, we do not include quantum mechanical features, such as ``electron spill-out''. that are not directly encoded in the bulk susceptibility. This means that our results for quantities close to a sharp boundary will lose accuracy below a few nanometers. In practice however, it has been found that some quantum features of the surface can be incorporated through a spatially dispersive susceptibility\cite{wie12}. Our work also assumes a smooth boundary, but surface roughness can potentially be incorporated in a similar fashion to that employed for a local medium\cite{halevi93,biehs11}. Boundary layers containing slits\cite{ren15} or other nontrivial structures\cite{xiong13} would require additional considerations of the field behavior in the interface layer. Our model can be used to find the reflection and transmission coefficients for spatially dispersive metamaterials when the wavelength is such that an effective medium description can be used. Finally, for materials such as thin films or nanospheres, a different approach to that used here is required because of more complicated possibilities for the behavior of polarization waves (e.g.\ multiple reflections from closely separated boundaries).
The paper is organized as follows.
In Sec. \ref{sec:model} we present the spatially dispersive susceptibility model for a half-infinite dielectric with a tensor permittivity and derive the field equations.
In Secs. \ref{sec:p-polarization} and \ref{sec:transmission} we derive the general expressions for the reflection and transmission coefficients and present the results for a variety of ABC's. In Sec. \ref{sec:energy} we calculate the zero-point and thermal spectral energy density and show in detail how the nonlocal response removes the divergence in this quantity that is present in a local model.
\section{Dielectric Model}\label{sec:model}
We first consider an infinite, homogeneous, spatially-dispersive dielectric with the susceptibility (\ref{eq:susc}).
The electric field ${\bm E}$ and polarization field ${\bm P}$ satisfy the wave equation:
\begin{align}
{\bm \nabla}
\times
{\bm \nabla}
\times
{\bm E}({\bm r},\omega)
-
\frac{\omega^2}{c^2}
{\bm E}({\bm r},\omega)
=
\frac{\omega^2}{c^2}
{\bm P}({\bm r},\omega),
\label{eq:wave_equation}
\end{align}
where the polarization field is:
\begin{align}
P_i({\bm r},\omega)=
\int d^3{\bm r}^\prime
\sum_j
\chi_{ij}({\bm r}-{\bm r}^\prime,\omega)
E_j({\bm r^\prime},\omega).
\label{eq:infinite_polarization}
\end{align}
Using the Fourier transformation:
\begin{align}
P_i({\bm r},\omega)
=
\frac{1}{(2\pi)^3}
\int
d^3{\bm k}
P_i({\bm k},\omega)
e^{i {\bm k} \cdot {\bm r}}
\end{align}
we have:
\begin{align}
P_i({\bm k},\omega)
=
\sum_j
\chi_{ij}({\bm k},\omega)
E_j({\bm k},\omega).
\end{align}
The wave equation (\ref{eq:wave_equation}) has solutions for ${\bm E}$ when the frequency and wave vector satisfy the dispersion relation\cite{rukhadze1961}:
\begin{align}
(\omega/c)^2\left[1+\chi_\perp({\bm k},\omega)\right]=k^2,
\label{eq:transverse_disp_rel}
\end{align}
for transverse waves with ${\bm E}\cdot{\bm k}=0$ or:
\begin{align}
1+\chi_\parallel({\bm k},\omega)=0,
\label{eq:longitudinal_disp_rel}
\end{align}
for longitudinal waves with ${\bm E}\times{\bm k}=0$.
As the electric field is parallel to the wave vector for the longitudinal wave, this wave has no magnetic field.
With an $\textrm{exp}(ik_zz)$ field dependence we restrict ourselves to wave vectors with $\textrm{Im}[k_z]>0$. There are two solutions to (\ref{eq:transverse_disp_rel}) which we denote ${\bm k_1}$, ${\bm k_2}$ and one solution to (\ref{eq:longitudinal_disp_rel}) which we denote ${\bm k_3}$.
\begin{figure}[!htb]\centering
{\includegraphics[width=\linewidth]{Figure01.pdf}}
\caption{
Schematic of the model.
The $z<0$ vacuum half-space contains the incident (${\bm k_0}$) and reflected (${\bm k_r}$) wave.
The angle of incidence is $\theta_i$.
The $z>0$ spatially dispersive dielectric half-space contains two transverse (${\bm k_1}$, ${\bm k_2}$) and one longitudinal (${\bm k_3}$) transmitted waves.
The coordinate system is chosen such that the $xz$-plane coincides with the plane of incidence and $k_y=0$.
}
\label{fig:model}
\end{figure}
We now consider a half-infinite dielectric that occupies the $z>0$ region as shown in Fig. \ref{fig:model}.
In the vacuum region we have the incident (${\bm E_0}$) and reflected wave (${\bm E_r}$) with wave-vectors $\bm{k}_0$ and $\bm{k}_r$ ($k_0=k_r=\omega/c$), while in the dielectric we have the one longitudinal wave (${\bm E^{(3)}}$) and two transverse waves (${\bm E^{(1)}},{\bm E^{(2)}}$) previously derived.
We choose our coordinate system such that the $xz$-plane coincides with the plane of incidence, with $k_y=0$, $k_x=K$.
The various wave-vectors differ only in the value of $k_z$ .
The angle of incidence is given by $\cos{\theta_i}=\sqrt{k_0^2-K^2}/k_0$. The usual Maxwell boundary conditions are now insufficient to solve for the unknown amplitudes of the three transmitted and one reflected wave.
Additional relationships between the amplitudes are required.
These are usually expressed as additional boundary conditions on the polarization, denoted $\mathcal{P}_i$, at $z=0^+$ due to the second term in the susceptibility (\ref{eq:susceptibility_basic}):
\begin{align}
\alpha_j\mathcal{P}_j(0^+)
+\beta_j
\partial_z\mathcal{P}_j(0^+)
=0,
\label{eq:robinABC}
\end{align}
for some parameters $\alpha_j$ and $\beta_j$.
A generalized approach was developed by Halevi and Fuchs\cite{Halevi} for a scalar susceptibility, equivalent to (\ref{eq:susc}) with $\chi_\perp=\chi_\parallel$.
We will now modify their derivation to the tensor case $\chi_{ij}$.
Due to the presence of the boundary, the polarization field now depends on a position-dependent susceptibility $\chi^\prime$:
\begin{align}
P_i({\bm r})=
\int d^3{\bm r}^\prime
\sum_j
\chi^\prime_{ij}({\bm r},{\bm r}^\prime)
E_j({\bm r^\prime}),
\end{align}
where we have omitted the $\omega$ dependence for notational simplicity.
After a Fourier transform in the $xy$-plane:
\begin{align}
\tilde{P}_i(k_x,k_y,z)=
\int_{0}^{\infty} dz^\prime
\sum_j
\tilde{\chi}^\prime_{ij}(k_x,k_y,z,z^\prime)
\tilde{E}_j(k_x,k_y,z^\prime).
\label{eq:polarization_half_infinite}
\end{align}
We assume that the overall susceptibility $\chi^\prime$ of the half-infinite dielectric can be expressed in terms of the bulk susceptibility:
\begin{equation}
\tilde{\chi}^\prime_{ij}(k_x,k_y,z,z^\prime)=
\left\{
\begin{aligned}
&\tilde{\chi}_{ij}(k_x,k_y,z-z^\prime)+\\
&U_{ij}\tilde{\chi}_{ij}(k_x,k_y,z+z^\prime),&\text{if } z,z^\prime > 0 \\
&0&\text{otherwise},
\end{aligned}
\right.
\label{eq:susc_U}
\end{equation}
where we have Fourier transformed the bulk susceptibility in (\ref{eq:susc}) to real space in the $z$-direction.
The first term in (\ref{eq:susc_U}) for $z,z^\prime > 0$ is position independent and gives the non-local bulk response.
The second term depends on the distance from the boundary and describes a polarization wave propagating from $z^\prime$ to the surface, reflecting with a (complex in general) amplitude coefficient $U_{ij}$ and continuing to $z$, with $|U_{ij}|=1$ implying elastic reflection. A similar expression to (\ref{eq:susc_U}) had been used previously\cite{garcia1977}, with a scalar $U$ used as a phenomenological description of the dielectric surface response.
Halevi and Fuchs\cite{Halevi} considered a scalar $\chi$, leading to a vector $U_i$ with general values, and showed this to be equivalent to using the ABC's in (\ref{eq:robinABC}).
After substituting the half-infinite susceptibility (\ref{eq:susc_U}) into (\ref{eq:polarization_half_infinite}), the polarization field takes the form:
\begin{align}
\tilde{P}_i(z)=&
\frac{1}{2\pi}
\int_{-\infty}^{\infty} dq
\int_{0}^{\infty} dz^\prime
\bigg[
e^{iq(z-z^\prime)}
\sum_j
\chi_{ij}(q)
\tilde{E}_j(z^\prime)
\nonumber\\&
+
e^{iq(z+z^\prime)}
\sum_j
U_{ij}
\chi_{ij}(q)
\tilde{E}_j(z^\prime)
\bigg],
\quad
z>0,
\label{eq:polarization_integral}
\end{align}
where $q=k_z$ and we have omitted the $k_x$ and $k_y$ values as they are the same in all arguments.
We now substitute the tensor (\ref{eq:susc}) with expressions for $\chi_\perp$ and $\chi_\parallel$ of the form (\ref{eq:susceptibility_basic}).
The susceptibility (\ref{eq:susceptibility_basic}) can be rewritten as:
\begin{align}
\chi(k_x,k_y,q,\omega)
=
\chi_0
+
\frac{\omega_p^2/\sigma^2}{q^2-\Gamma^2},
\label{eq:susceptibility_simplified}
\end{align}
\begin{align}
\Gamma^2
=
\frac{\omega^2-\omega_T^2+i\gamma\omega-\sigma^2(k_x^2+k_y^2)}{\sigma^2}.
\end{align}
The transverse (longitudinal) susceptibility takes the form in (\ref{eq:susceptibility_simplified}), but with $\sigma$ and $\Gamma$ replaced by $\sigma_\perp$ ($\sigma_\parallel$) and the corresponding $\Gamma_\perp$ ( $\Gamma_\parallel$).
We define the relationship $\sigma_\parallel^2=(1+\delta)\sigma_\perp^2$, so that the susceptibility tensor reduces to $\delta_{ij}\chi_\perp(k,\omega)$ in the $\delta\to0$ limit.
At this point we introduce an ansatz for the ${\bm E}$ field inside the medium - a linear combination of three plane waves\cite{Halevi}:
\begin{align}
\tilde{E}_j(z)
=
\sum_{n=1}^3
\tilde{E}_j^{(n)}
e^{iq_nz},
\label{eq:E_ansatz}
\end{align}
where $n=1,2$ are the transverse waves and $n=3$ is the longitudinal wave.
Substituting (\ref{eq:E_ansatz}) into (\ref{eq:polarization_integral}) and evaluating the integrals gives:
\begin{align}
\tilde{P}_i(z)=&
\sum_n
\sum_j
\chi_{ij}(q_n)
\tilde{E}^{(n)}_j
e^{iq_nz}
\nonumber\\&
+
\sum_n
\sum_j
\phi_{ij}^{(n)}
\tilde{E}^{(n)}_j
e^{i\Gamma_\perp z}
\nonumber\\&
+
\sum_n
\sum_j
\psi_{ij}^{(n)}
\tilde{E}^{(n)}_j
e^{i\Gamma_\parallel z},
\end{align}
where:
\begin{align}
\phi_{ij}^{(n)}
=&
-
\left[q_n(1+U_{ij})+\Gamma_\perp(1-U_{ij})\right]\chi_\perp(q_n)
\nonumber\\&\times
\frac{1}{2\Gamma_\perp}
\left(\delta_{ij}-\frac{k^{(\perp)}_ik^{(\perp)}_j}{\Gamma_\perp^2+K^2}\right),
\end{align}
\begin{align}
\psi_{ij}^{(n)}
=&
-
\left[q_n(1+U_{ij})+\Gamma_\parallel(1-U_{ij})\right]\chi_\parallel(q_n)
\nonumber\\&\times
\frac{1}{2\Gamma_\parallel}
\left(\frac{k^{(\parallel)}_ik^{(\parallel)}_j}{\Gamma_\parallel^2+K^2}\right),
\end{align}
and ${\bm k}^{(\perp/\parallel)}=\left(K,0,\Gamma_{\perp/\parallel}\right)$.
This must be substituted into the right-hand side (RHS) of the wave equation (\ref{eq:wave_equation}).
All left-hand side terms are proportional to $\textrm{exp}(iq_nz)$, so
for the wave equation to hold for all $z$ values, we require the two RHS sums proportional to $\textrm{exp}(i\Gamma_\perp z)$ and $\textrm{exp}(i\Gamma_\parallel z)$ to equal zero:
\begin{align}
\sum_n
\sum_j
\phi_{ij}^{(n)}
\tilde{E}^{(n)}_j
=
0,
\qquad
\sum_n
\sum_j
\psi_{ij}^{(n)}
\tilde{E}^{(n)}_j
=
0.
\label{eq:abc}
\end{align}
These two equations act as the additional boundary conditions for the system, once the $U_{ij}$ in (\ref{eq:susc_U}) are specified.
\section{$p$-Polarization}\label{sec:p-polarization}
The wave can be decomposed to components with $\bm{E}$ perpendicular to ($s$-polarized) or in ($p$-polarized) the plane of incidence.
For $s$-polarization there is no longitudinal wave and the second term of the susceptibility (\ref{eq:susc}) does not contribute, effectively reducing the susceptibility tensor to the diagonal form $\delta_{ij}\chi_\perp({\bm k},\omega)$ used by Halevi and Fuchs \cite{Halevi}. The derivation in this case is identical to their work and will not be repeated here.
In contrast, the $p$-polarization includes the longitudinal wave and the second term in (\ref{eq:susc}) contributes. We proceed to analyze this case.
\subsection{Field Equations}
For $p$-polarized light $E_y=0$, $E_x\ne0$ and $E_z\ne0$.
After a Fourier transform in the $xy$-plane, we write the $x$ and $z$ components of the wave equation (\ref{eq:wave_equation}) inside the material ($z>0$). Using equation (\ref{eq:infinite_polarization}) and the ansatz (\ref{eq:E_ansatz}) these components are:
\begin{align}
\sum_{n=1}^3
\bigg\{&
\left[
k_0^2(1+\chi_{xx}(q_n))-q_n^2
\right]\tilde{E}_x^{(n)}
\nonumber\\&
+
\left[
Kq_n+k_0^2\chi_{xz}(q_n)
\right]\tilde{E}_z^{(n)}
\bigg\}
e^{iq_nz}
=0,
\end{align}
\begin{align}
\sum_{n=1}^3
\bigg\{&
\left[
Kq_n+k_0^2\chi_{zx}(q_n)
\right]\tilde{E}_x^{(n)}
\nonumber\\&
+
\left[
k_0^2(1+\chi_{zz}(q_n))-K^2
\right]\tilde{E}_z^{(n)}
\bigg\}
e^{iq_nz}
=0.
\end{align}
These must hold for all values of $z$, giving:
\begin{align}
&
\left[
k_0^2(1+\chi_{xx}(q_n))-q_n^2
\right]\tilde{E}_x^{(n)}
\nonumber\\&+
\left[
Kq_n+k_0^2\chi_{xz}(q_n)
\right]\tilde{E}_z^{(n)}
=0,
\label{eq:Ex}
\end{align}
\begin{align}
&
\left[
Kq_n+k_0^2\chi_{zx}(q_n)
\right]\tilde{E}_x^{(n)}
\nonumber\\&+
\left[
k_0^2(1+\chi_{zz}(q_n))-K^2
\right]\tilde{E}_z^{(n)}
=0.
\label{eq:Ez}
\end{align}
If $\tilde{E}_x^{(n)}$ and $\tilde{E}_z^{(n)}$ are non-zero, the determinant of these equations must vanish for all three values of $n$. This requirement leads to the dispersion relation (\ref{eq:transverse_disp_rel}) for $n=1,2$ and (\ref{eq:longitudinal_disp_rel}) for $n=3$. For the form of $\chi_\perp$ and $\chi_\parallel$ used here these dispersion relations are:
\begin{gather}
\left[
\left(1+\chi_0\right)k_0^2
-K^2
-
q_n^2
\right]
\left[
\Gamma_\perp^2
-
q_n^2
\right]
=
k_0^2
\frac{\omega_p^2}{\sigma_\perp^2},
\label{eq:q1q2} \\
\left[
\Gamma_\parallel^2
-
q_3^2
\right]
=
\frac{\omega_p^2}{\left(1+\chi_0\right) \sigma_\parallel^2}.
\label{eq:q3}
\end{gather}
Rearranging (\ref{eq:Ex}) and (\ref{eq:Ez}) gives relations between the components of ${\bm \tilde{E}}$ :
\begin{align}
\tilde{E}_z^{(n)}=\eta^{(n)}\tilde{E}_x^{(n)},
\label{eq:Ex_Ez}
\end{align}
where $\eta^{(n)}$ takes the role of $\gamma^{(n)}$ in the Halevi \& Fuchs derivation\cite{Halevi}, and $\eta^{(1)}=-K/q_1$, $\eta^{(2)}=-K/q_2$ and $\eta^{(3)}=q_3/K$.
\subsection{Surface Impedance}
The reflection coefficient will be calculated below from the surface impedance, which for $p$-polarized light is given by:
\begin{align}
Z_p
=
\frac{
E_x(0^+)
}{
H_y(0^+)
}.
\end{align}
(Here ${\bm H}=\mu_0{\bm B}$.) The magnetic field $B_y$ can be expressed in terms of the electric field using $k_0{\bm B}={\bm k}\times{\bm E}$ and (\ref{eq:E_ansatz}):
\begin{align}
B_y^{(n)}&
=
\frac{1}{k_0}
\left[
q_nE_x^{(n)}
-
KE_z^{(n)}
\right]
e^{iq_nz}
\nonumber\\&
=
\left[
\frac{
q_n
-
K\eta^{(n)}
}{k_0}
\right]
E_x^{(n)}
e^{iq_nz}
\nonumber\\&
=
\tau^{(n)}
E_x^{(n)}
e^{iq_nz}.
\end{align}
Here we have substituted for $E_z$ using (\ref{eq:Ex_Ez}) and defined $\tau^{(n)}$ by:
\begin{align}
\tau^{(n)}=
\frac{q_n^2+K^2}{q_n k_0}
=
\frac{k_0}{q_n}
\left(
1
+
\chi_0
+
\frac{\omega_p^2/\sigma_\perp^2}{q_n^2-\Gamma_\perp^2}
\right)
\end{align}
for $n=1,2$ and $\tau^{(3)}=0$ for the longitudinal wave.
The surface impedance can now be expressed in terms of field amplitude ratios:
\begin{align}
Z_p
=&
\frac{1}{\mu_0}
\frac{
E_x^{(1)}
+
E_x^{(2)}
+
E_x^{(3)}
}{
\tau^{(1)}
E_x^{(1)}
+
\tau^{(2)}
E_x^{(2)}
}
=
\frac{1}{\mu_0}
\frac{1
+
\frac{E_x^{(2)}}{E_x^{(1)}}
+
\frac{E_x^{(3)}}{E_x^{(1)}}
}{
\frac{q_1^2+K^2}{q_1 k_0}
+
\frac{q_2^2+K^2}{q_2 k_0}
\frac{E_x^{(2)}}{E_x^{(1)}}
}.
\label{eq:zp_ratios}
\end{align}
\subsection{Additional Boundary Conditions}
At this point we require the field amplitude ratios of the transmitted waves to find the surface impedance (\ref{eq:zp_ratios}).
By using the relation in (\ref{eq:Ex_Ez}), we rewrite the additional boundary conditions in (\ref{eq:abc}) solely in terms of $E_x$:
\begin{align}
\sum_n
\sum_j
\phi_{ij}^{(n)}
E^{(n)}_j&
=
\sum_n
\left[
\phi_{ix}^{(n)}
E^{(n)}_x
+
\phi_{iz}^{(n)}
E^{(n)}_z
\right]
\nonumber\\&
=
\sum_n
\left[
\phi_{ix}^{(n)}
+
\phi_{iz}^{(n)}
\eta^{(n)}
\right]
E^{(n)}_x
=
0.
\end{align}
We collect together the terms in square brackets to new variables $a_n$ and $b_n$ for $i=x$ and $z$ respectively:
\begin{align}
\sum_n
\left[
\phi_{xx}^{(n)}
+
\phi_{xz}^{(n)}
\eta^{(n)}
\right]
E^{(n)}_x
=&
\sum_n
a_n
E^{(n)}_x
=
0,
\nonumber\\
\sum_n
\left[
\phi_{zx}^{(n)}
+
\phi_{zz}^{(n)}
\eta^{(n)}
\right]
E^{(n)}_x
=&
\sum_n
b_n
E^{(n)}_x
=
0.
\label{eq:abc1}
\end{align}
The $\psi$ terms in (\ref{eq:abc}) are collected in a similar fashion to define $c_n$ and $d_n$:
\begin{align}
\sum_n
\left[
\psi_{xx}^{(n)}
+
\psi_{xz}^{(n)}
\eta^{(n)}
\right]
E^{(n)}_x
=&
\sum_n
c_n
E^{(n)}_x
=
0,
\nonumber\\
\sum_n
\left[
\psi_{zx}^{(n)}
+
\psi_{zz}^{(n)}
\eta^{(n)}
\right]
E^{(n)}_x
=&
\sum_n
d_n
E^{(n)}_x
=
0.
\label{eq:abc2}
\end{align}
We now have four ABC equations, compared to the two in the Halevi and Fuchs derivation\cite{Halevi}, which must all be satisfied.
With some manipulation we obtain:
\begin{align}
\frac{E_x^{(2)}}{E_x^{(1)}}
=\frac{
(3,1)_{\mu\nu}
}{
(2,3)_{\mu\nu}
},
\qquad
\frac{E_x^{(3)}}{E_x^{(1)}}
=\frac{
(1,2)_{\mu\nu}
}{
(2,3)_{\mu\nu}
},
\label{eq:amp_ratios}
\end{align}
where we define the symbol $(i,j)_{\mu\nu}=\mu_i\nu_j-\mu_j\nu_i$ with $\mu,\nu \in \{a,b,c,d\}$ and $\mu\neq\nu$. The field amplitude ratios in (\ref{eq:amp_ratios}) must give the same value for any combination of $\mu$ and $\nu$ ($\mu\neq\nu$).
Given the fact that $a$ and $c$ contain only $U_{xx}$ and $U_{xz}$ while $b$ and $d$ contain only $U_{zx}$ and $U_{zz}$, there must be some restrictions on the values that $U_{ij}$ can take. We find that (\ref{eq:amp_ratios}) can be satisfied for all $\mu,\nu$ combinations with $U_{xx}=U_{zx},U_{xz}=U_{zz}$, so that $b_n=(-k/\Gamma_\perp)a_n$ and $d_n=(\Gamma_\parallel/k)c_n$, reducing (\ref{eq:abc1}) and (\ref{eq:abc2}) to two equations.
Under these conditions we can make clear comparisons to the choice of ABC's presented by Halevi and Fuchs\cite{Halevi}, by associating their $U_x$ and $U_z$ with $U_{xx}$ and $U_{zz}$ as in Table \ref{tab:abc}.
The choice of ABC is typically dependent on the type of material, with various authors making arguments based on the microscopic behavior of the system.
Both the Pekar\cite{pekar1958a,pekar1958b,pekar1958c,pekar1959} and
Rimbey-Mahan\cite{rimbey1974,rimbey1975,rimbey1976,rimbey1977,rimbey1978} ABC were developed for Frenkel (tight-binding) excitons systems such as molecular crystals, although the second excluded the coupling of light to longitudinal modes of the medium.
Ting \emph{et al.}\cite{ting1975} looked at a crystal model with Wannier-Mott (weak-binding) excitons, typically found in semiconductors.
The Fuchs-Kliewer\cite{ting1975,kliewer1968,kliewer1971,kliewer1975,ruppin1981} ABC considered a metal with specular reflection of electrons at the inner surface.
The Agarwal \emph{et al.}\cite{agarwal1971a,agarwal1971b,agarwal1972,birman1972,agarwal1973,maradudin1973,birman1974,mills1974,foley1975,bishop1976}
ABC was not for a specific type of material, but derived under the assumption that changes in the susceptibility arising from the presence of the boundary can be neglected when considering bulk effects such as reflection and refraction. Henneberger\cite{hen98} considered a thin surface layer on the boundary as a source of radiation and found the ABC of Ting \emph{et al.}\cite{ting1975} in a simple case.
\begin{table}[!htb]\centering
\caption{\label{tab:abc}List of ABC's}
\begin{ruledtabular}
\begin{tabular}{llll}
& $U_{xx}$ & $U_{yy}$ & $U_{zz}$\\
\hline
Agarwal \emph{et al}.\cite{agarwal1971a,agarwal1971b,agarwal1972,birman1972,agarwal1973,maradudin1973,birman1974,mills1974,foley1975,bishop1976}
& \phantom{-}0 & \phantom{-}0 & \phantom{-}0\\
Ting \emph{et al}.\cite{ting1975}& \phantom{-}1 & \phantom{-}1 & \phantom{-}1\\
Fuchs-Kliewer\cite{ting1975,kliewer1968,kliewer1971,kliewer1975,ruppin1981} & \phantom{-}1 & \phantom{-}1 & -1\\
Rimbey-Mahan\cite{rimbey1974,rimbey1975,rimbey1976,rimbey1977,rimbey1978} & -1 & -1 & \phantom{-}1\\
Pekar\cite{pekar1958a,pekar1958b,pekar1958c,pekar1959} & -1 & -1 & -1\\
\end{tabular}
\end{ruledtabular}
\end{table}
Using (\ref{eq:amp_ratios}), the surface impedance (\ref{eq:zp_ratios}) can be written:
\begin{align}
Z_p
=
\frac{1}{\mu_0}
\frac{
(2,3)_{ac}
+
(3,1)_{ac}
+
(1,2)_{ac}
}{
\frac{k_1^2}{q_1 k_0}
(2,3)_{ac}
+
\frac{k_2^2}{q_2 k_0}
(3,1)_{ac}
},
\label{eq:surface_imp}
\end{align}
where $k_n^2=K^2+q_n^2$. From now on we will only use the combination $\mu=a,\nu=c$ in the symbol $(i,j)_{\mu\nu}$, as in (\ref{eq:surface_imp}), so hereafter we omit the subscript $ac$ for notational simplicity.
In the simplest ABC case $U_{ij}=0$ and we find:
\begin{widetext}
\begin{align}
Z_p
=
-
\frac{k_0}{\mu_0}
\Bigg\{
\Gamma_\perp
-
\frac{
\left[
K^4
+K^2(q_1^2+q_1q_2+q_2^2)
+(q_1+q_2)q_1q_2q_3
\right]
-
\frac{\left(\Gamma_\parallel-\Gamma_\perp\right)(K^2+q_3^2)}{\left(\Gamma_\perp^2+K^2\right)+\left(\Gamma_\parallel-\Gamma_\perp\right)\left(\Gamma_\perp+q_3\right)}
\left[
q_1q_2(q_1+q_2)
\right]
}{
\left[
K^2(q_1+q_2-q_3)
+q_1q_2q_3
\right]
+
\frac{\left(\Gamma_\parallel-\Gamma_\perp\right)(K^2+q_3^2)}{\left(\Gamma_\perp^2+K^2\right)+\left(\Gamma_\parallel-\Gamma_\perp\right)\left(\Gamma_\perp+q_3\right)}
\left[
K^2-q_1q_2
\right]
}
\Bigg\}^{-1}
.
\end{align}
\end{widetext}
In the general case where $U_{xx}=U_{zx}$ and $U_{xz}=U_{zz}$, (\ref{eq:surface_imp}) reduces to the Halevi and Fuchs result in the $\delta\to0$ limit where $\chi_{ij}=\delta_{ij}\chi_\perp$.
\subsection{Reflection Coefficient}
\label{sec:reflection}
Using the vacuum surface impedance $Z_p^{(0)}=\sqrt{k_0^2-K^2}/\mu_0k_0$ and (\ref{eq:surface_imp}), we can construct the $p$-polarization reflection coefficient\cite{kliewer1968}:
\begin{align}
r_p=\frac{
Z_p^{(0)}-Z_p
}{
Z_p^{(0)}+Z_p
},
\label{eq:rp}
\end{align}
where:
\begin{align}
r_p
=\frac{{ E_r}}{{ E_0}}.
\label{eq:rp_definition}
\end{align}
\begin{figure}[!htb]\centering
{\includegraphics[width=\linewidth]{Figure02.pdf}}
\caption{
Absolute value of the reflection coefficient of ZnSe at $\omega=\omega_T$ as a function of incident angle $\theta_i$ for propagating waves with $\delta=0$ (solid lines), $\delta=0.5$ (dashed) and $\delta=-0.5$ (dotted).
Includes Agarwal \emph{et al}. (Red), Ting \emph{et al}. (Brown), Fuchs-Kliewer (Green), Rimbey-Mahan (Blue) and Pekar (Purple) ABC's. The black curve has the spatial dispersion removed ($\sigma_\parallel=\sigma_\perp=0$).
}
\label{fig:rp_prop_angle}
\end{figure}
\begin{figure}[!htb]\centering
{\includegraphics[width=\linewidth]{Figure03.pdf}}
\caption{
Absolute value of the reflection coefficient of ZnSe at $\theta_i=\pi/4$ as a function of $\omega$.
Plot styles follow the conventions in Fig \ref{fig:rp_prop_angle}.
The parameter $\delta$ has the greatest effect near the reflection minimum.
}
\label{fig:rp_freq}
\end{figure}
\begin{figure}[!htb]\centering
{\includegraphics[width=\linewidth]{Figure04.pdf}}
\caption{
Detail of Fig. \ref{fig:rp_freq} near the reflection minimum, where $\delta$ has the greatest effect.
}
\label{fig:rp_freq_detail}
\end{figure}
As an example, we consider parameters for ZnSe, the same material used by Halevi and Fuchs\cite{Halevi,ZnSe},
with $\chi_0=8.1$, $\omega_p=3.25\times10^{14}$ rad s$^{-1}$, resonant frequency $\omega_T=4.25\times10^{15}$ rad s$^{-1}$ and damping $\gamma=4.25\times10^{10}$ rad s$^{-1}$.
We define the non-local term $\sigma^2=\hbar\omega_T/(m_e+m_h)$ in the same manner as Halevi and Fuchs\cite{Halevi}, where $m_e$ and $m_h$ are the electron and hole mass.
For ZnSe, $\sigma_\perp=7.45\times10^5$m s$^{-1}$.
The exact value for $\delta$ is unknown, so we will present our results over the range $\delta=-0.5$ to $0.5$.
Figure \ref{fig:rp_prop_angle} shows the absolute value of the reflection coefficient for a range of ABC's and $\delta$ values at the resonant frequency $\omega_T$.
The choice of $U_{ij}$ values, specifically $U_{xx}$, has the greatest effect on $r_p$ near this frequency.
The $\delta$ parameter modifies the result to a much smaller extent, with the greatest change near the reflection minimum in the Agarwal \emph{et al}. ABC\cite{agarwal1971a,agarwal1971b,agarwal1972,birman1972,agarwal1973,maradudin1973,birman1974,mills1974,foley1975,bishop1976} , followed by Fuchs-Kleiwer\cite{ting1975,kliewer1968,kliewer1971,kliewer1975,ruppin1981}, Pekar\cite{pekar1958a,pekar1958b,pekar1958c,pekar1959} and Ting \emph{et al}\cite{ting1975}. The Rimbey-Mahan\cite{rimbey1974,rimbey1975,rimbey1976,rimbey1977,rimbey1978} result remains unchanged by $\delta$. This ABC was chosen so that no longitudinal wave could be generated, so in this case $\chi_\parallel$ and $\delta$ have no effect on $r_p$.
Figures \ref{fig:rp_freq} and \ref{fig:rp_freq_detail} show the $\omega$ dependence of $|r_p|$ for a fixed angle $\theta_i=\pi/4$.
It can be seen that $\delta$ has the greatest effect at frequencies slightly larger than $\omega_T$ near the reflection minima.
Agarwal \emph{et al}., Fuchs-Kliewer and Pekar are the most affected by $\delta$, while the change in Ting \emph{et al}. is significantly smaller.
\section{$p$-polarization transmission coefficients}
\label{sec:transmission}
We can find the transmission coefficients for the three transmitted waves by imposing the continuity of the tangential ${\bm E}$ field across the boundary.
Our choice of coordinate system means we simply equate the $E_x$ components on each side:
\begin{align}
\left[E_0-E_r\right]\cos{\theta_i}
=
\left[E^{(1)}_x+E^{(2)}_x+E^{(3)}_x\right].
\end{align}
By using (\ref{eq:amp_ratios}), (\ref{eq:rp_definition}) and $\cos{\theta_i}=\sqrt{k_0^2-K^2}/k_0$, this can be rewritten in terms of a single wave amplitude on the right:
\begin{align}
\frac{\sqrt{k_0^2-K^2}}{k_0}\left[1-r_p\right]E_0
=
\frac{\left[(2,3)+(3,1)+(1,2)\right]}{(2,3)}E^{(1)}_x.
\end{align}
Similar expressions can be found for $n=2,3$.
By using:
\begin{align}
E^{(n)}
=
\sqrt{
\left[E^{(n)}_x\right]^2
+
\left[E^{(n)}_z\right]^2
}
\end{align}
and (\ref{eq:Ex_Ez}), we derive the three transmission coefficients:
\begin{align}
t^{(1)}_p
=&
(2,3)\frac{k_1}{q_1}
\frac{\sqrt{k_0^2-K^2}}{k_0}\frac{\left[1-r_p\right]}{\left[(2,3)+(3,1)+(1,2)\right]}
,
\nonumber\\
t^{(2)}_p
=&
(3,1)\frac{k_2}{q_2}
\frac{\sqrt{k_0^2-K^2}}{k_0}\frac{\left[1-r_p\right]}{\left[(2,3)+(3,1)+(1,2)\right]}
,
\nonumber\\
t^{(3)}_p
=&
(1,2)\frac{k_3}{K}
\frac{\sqrt{k_0^2-K^2}}{k_0}\frac{\left[1-r_p\right]}{\left[(2,3)+(3,1)+(1,2)\right]}
,
\label{eq:tp}
\end{align}
where:
\begin{align}
t^{(n)}_p
= \frac{{ E^{(n)}}}{{E_0}}.
\label{eq:tp_definition}
\end{align}
\begin{figure}[!htb]\centering
{\includegraphics[width=\linewidth]{Figure05.pdf}}
\caption{
Absolute values of the transmission coefficient for the three waves in ZnSe at $\omega=\omega_T$ as a function of incidence angle $\theta_i$.
Plot styles follow the conventions in Fig \ref{fig:rp_prop_angle}.
}
\label{fig:tp_prop_angle}
\end{figure}
Figure \ref{fig:tp_prop_angle} shows the absolute value of the transmission coefficients at $\omega=\omega_T$ as a function of incident angle for ZnSe.
For the two transverse waves, the $U_{ij}$ values, specifically $U_{xx}$, have the greatest effect.
Pekar and Rimbey-Mahan give near-identical results while Ting \emph{et al}. and Fuchs-Kleiwer also have similar values.
The effect of $\delta$ is negligible for $n=1$, which corresponds to the wave present when the material has local response, but its effect is larger for the second transverse wave introduced by the nonlocal dependence.
The longitudinal wave shows different behavior.
At normal incidence $t_p^{(3)}=0$, since ${\bm E_0}$ is polarized parallel to the surface and there is no $z$-component to excite the longitudinal wave.
The results show a large spread with $\delta$. Fuchs-Kleiwer is generally the largest, while Rimbey-Mahan is always zero due to the absence of a longitudinal wave in that case.
Beyond these features, the behavior of $t_p^{(3)}$ for the various ABC's, and also the effect of $\delta$, is strongly dependent on the material parameters and frequency.
For example, while Agarwal \emph{et al.} and Ting \emph{et al.} give similar results in Fig. \ref{fig:tp_prop_angle} while Fuchs-Kleiwer is not affected by $\delta$, this is not true in general.
We have already highlighted the Rimbey-Mahan ABC, where the values of $U_{ij}$ leads to $c_1=c_2=0$. As a result $(1,2)=0$ and there is no longitudinal wave ($E^{(3)}=0$). We now check the possibility of choosing $U_{ij}$ to give no \emph{transverse} waves in the medium. With no transverse waves the ${\bm B}$-field in the medium is zero, leading to $Z_p=\infty$ and perfect reflection with $r_p=-1$.
This requires $a_3=0$ and $c_3=0$, leading to the following values of $U_{xx}$ and $U_{zz}$:
\begin{align}
U_{xx}
=
1+
\frac{
2q_3(\Gamma_\parallel^2+K^2)
}{
(\Gamma_\perp\Gamma_\parallel+K^2)(\Gamma_\parallel-q_3)
}
\nonumber\\
U_{zz}
=
-1+
\frac{
2\Gamma_\perp(\Gamma_\parallel^2+K^2)
}{
(\Gamma_\perp\Gamma_\parallel+K^2)(\Gamma_\parallel-q_3)
}.
\label{eq:long_only_1}
\end{align}
In the $\delta=0$ case, this reduces to:
\begin{align}
U_{zz}
=
U_{xx}
=
\frac{\Gamma+q_3}{\Gamma-q_3}.
\label{eq:long_only_2}
\end{align}
Using the $q_3$ definition in (\ref{eq:q3}) we find from (\ref{eq:long_only_2}) the required $U$ to get perfect reflection for given $K$ and $\omega$ values in the $\delta=0$ case:
\begin{align}
U_{xx}
=U_{zz}
=
\frac{
1+\sqrt{
1+\frac{1}{1+\chi_0}\frac{\omega_p^2}{\omega_T^2+\sigma^2K^2-\omega^2-i\gamma\omega}
}
}{
1-\sqrt{
1+\frac{1}{1+\chi_0}\frac{\omega_p^2}{\omega_T^2+\sigma^2K^2-\omega^2-i\gamma\omega}
}
}.
\label{eq:longU}
\end{align}
The relation gives $\left|U_{xx}\right|>1$, conflicting with the definition of $U$ as the reflection coefficient of the polarization waves at the surface. Note that (\ref{eq:longU}) does not allow $\left|U_{xx}\right|=1$ because the square-root quantity does not vanish for any real $K$ and $\omega$. Thus we must have transverse transmitted waves in the material.
\section{Spectral Energy density}\label{sec:energy}
\begin{figure*}[!htb]
{\includegraphics[width=178mm]{Figure06.pdf}}
\caption{
Logarithmic plot of Im$[r_p]$ used in the $u_{tot}$ integration (\ref{eq:utot}) at $\omega=0.999\omega_T$ (top left) and $\omega=1.01\omega_T$ (all others) as a function of $K$ for evanescent waves.
Plot styles follow the conventions in Fig \ref{fig:rp_prop_angle}.
}
\label{fig:rp_evan_k}
\end{figure*}
We now apply the previous results to the problem of electromagnetic zero-point and thermal radiation near material boundaries. It is well known that the neglect of spatial dispersion leads to an unphysical divergence in the energy density of thermal radiation at a planar boundary\cite{henkel2000,joulain2005}. The divergence occurs at the level of the spectral energy density (i.e.\ the energy density per unit frequency) and this same divergence is present for the zero-point spectral energy density. Although the total zero-point energy density will always diverge if it is not regularized, the \emph{spectral} energy density of zero-point radiation should be finite without regularization\cite{hor14}. When the nonlocal response of materials is taken into account, all these spurious divergences must disappear and thus spatial dispersion is the key property that determines the spectral energy density of zero-point and thermal radiation near material boundaries. It has already been shown\cite{henkel2006} that a plasma described by the nonlocal Lindhard susceptibility gives a finite spectral energy density at a planar boundary. A similar model removes an unphysical divergence in spectral zero-point and thermal correlations inside a homogeneous material\cite{nar10,hor14}. Here we show that the quite general dielectric model used here is free of the aforementioned divergences at a planar boundary, and we also show that differences between the transverse and longitudinal susceptibilities can have a large effect.
The average energy density of zero-point and thermal radiation in the vacuum region outside the semi-infinite dielectric of Fig.~\ref{fig:model} is given by:
\begin{align}
\langle U\rangle
=&
\frac{\varepsilon_0}{2}\langle \left|{\bm E}\left({\bm r},t\right) \right|^2\rangle
+
\frac{\mu_0}{2}\langle \left|{\bm B}\left({\bm r},t\right) \right|^2\rangle
\nonumber\\
=&
\int_0^\infty
d\omega \, u_{tot}\left(z,\omega\right),
\end{align}
where $u_{tot}(z,\omega)$ is the spectral energy density that depends on $z$. We assume that the semi-infinite dielectric is in thermal equilibrium with the surroundings and we will include the zero-point contribution. The expression for $u_{tot}(z,\omega)$ can be written in terms of the reflection coefficients for $s$- and $p$-polarized light at the planar boundary (see for example Ref.~\onlinecite{kit05}):
\begin{align}
& \!\!\! u_{tot}(z,\omega)
=
\nonumber\\
& \!\!\! \frac{u_0}{k_0}
\int_0^{k_0} \!\!\!
\frac{KdK}{\sqrt{k_0^2-K^2}}
\left[
1+\frac{K^2\textrm{Re} \left[ (r_s+r_p) e^{-2i\sqrt{K^2-k_0^2}z} \right] }{2k_0^2}
\right]
\nonumber\\
& +
\frac{u_0}{2 k_0^3} \int_{k_0}^\infty
\frac{K^3dK}{\sqrt{K^2-k_0^2}}
\textrm{Im}[r_s + r_p]
e^{2\sqrt{K^2-k_0^2}z}
,
\label{eq:utot}
\end{align}
where $u_0$ is the spectral energy density in the absence of the material, given by:
\begin{align}
u_0&
=
\frac{\Theta(\omega,T)\omega^2}{\pi^2c^3}, \\
\Theta(\omega,T)&
=
\hbar\omega
\left(
\frac{1}{2}
+\frac{1}{e^{\hbar\omega/k_BT}-1}
\right).
\end{align}
The quantity $\Theta(\omega,T)$ is the mean energy of a harmonic oscillator in thermal equilibrium, the first term of which gives rise to the electromagnetic zero-point energy. The first term in (\ref{eq:utot}) is the contribution of propagating waves while the second term comes from evanescent waves.
If spatial dispersion is ignored then as $K\to\infty$ the reflection coefficients have the limits $r_s\to0$, $r_p\to\chi(\omega)/(2+\chi(\omega))$, where $\chi(\omega)$ is the local susceptibility of the isotropic medium.
For large $K$ the $r_p$-term in the second integral in (\ref{eq:utot}) is then proportional to $K^2$, dominating the final result for $u_{tot}(z,\omega)$ when $z$ is much smaller than the wavelength. This leads to a simple approximate expression\cite{joulain2005} for the second integral in (\ref{eq:utot}) as $z\to0$:
\begin{align}
\frac{1}{4z^3}
\frac{\textrm{Im}[\chi(\omega)]}{|2+\chi(\omega)|^2},
\label{eq:utot_nondispersive}
\end{align}
which diverges as $z\to0$ for complex $\chi(\omega)$. This unphysical divergence is removed when the bulk susceptibility has a dependence on $k$ of the form of the second term in (\ref{eq:susceptibility_basic}), which is the form we used for the transverse and longitudinal susceptibilities. But the background term $\chi_0$ in (\ref{eq:susceptibility_basic}) will still lead to a divergence in the spectral energy density if it is complex. As already noted, the background term $\chi_0$ should be replaced by additional resonance terms in a more general susceptibility $\chi(k,\omega)$. For our parameters for ZnSe, however, $\chi_0$ is real and so it causes no difficulties in the numerical calculations below.
We now consider the spectral energy density (\ref{eq:utot}) for our model with bulk tensor susceptibility (\ref{eq:susc}). The $p$-polarization reflection coefficient $r_p$ was found in Sec.~\ref{sec:p-polarization} (see (\ref{eq:rp})) and $r_s$ is given by Halevi and Fuchs\cite{Halevi}:
\begin{align}
r_s=\frac{
Z_s^{(0)}-Z_s
}{
Z_s^{(0)}+Z_s
},
\label{eq:rs}
\end{align}
where $Z_s^{(0)}=k_0/\mu_0\sqrt{k_0^2-K^2}$ and:
\begin{widetext}
\begin{align}
Z_s=
\frac{(1+U_{yy})k_0(q_1q_2+\Gamma_\perp^2)
+(1-U_{yy})k_0\Gamma_\perp(q_1+q_2)
}{(1+U_{yy})q_1q_2(q_1+q_2)
+(1-U_{yy})\Gamma_\perp(q_1^2+q_1q_2+q_2^2-\Gamma_\perp^2)}.
\label{eq:Zs}
\end{align}
\end{widetext}
We will now substitute these refection coefficients into (\ref{eq:utot}) and perform the integration over $K$, with the same material parameters as used previously. As we have seen, it is the behavior of the $r_p$ term in the second (evanescent wave) integral in (\ref{eq:utot}) that determines the spectral energy density near the boundary.
Figure \ref{fig:rp_evan_k} shows the behavior of $\textrm{Im}[r_p]$ at $\omega=0.999\omega_T$ and $\omega=1.01\omega_T$ for evanescent waves ($K>k_0$). In contrast to the $r_p$ results for propagating waves in Sec \ref{sec:p-polarization}, $\delta$ has a significant effect on $\textrm{Im}[r_p]$ for evanescent waves.
For $K<\sqrt{(1+\chi_0)k_0^2}$, the reflection coefficient closely matches the local result, whereas for $K>\sqrt{(1+\chi_0)k_0^2}$ the plots show how $\textrm{Im}[r_p]$ has a very different behavior from the local model. Spatial dispersion causes $\textrm{Im}[r_p]$ to fall off as $1/K^4$ for large $K$, but its behavior for smaller $K$ differs significantly for $\omega<\omega_T$ compared to $\omega>\omega_T$. For $\omega=0.999\omega_T$, Agarwal \emph{et al}., Fuchs-Kleiwer and Rimbey-Mahan are nearly identical in the $\delta=0$ limit, while Ting \emph{et al}. is larger and Pekar is smaller.
The $\delta$ parameter has the greatest effect on Fuchs-Kleiwer and a smaller effect on Agarwal \emph{et al}., Ting \emph{et al}. and Pekar. Rimbey-Mahan remains unchanged with $\delta$ due to the absence of the longitudinal wave. For $\omega=1.01\omega_T$ there is a peak in $\textrm{Im}[r_p]$, followed by a sharp drop, at the value of $K$ where $\textrm{Re}[\Gamma_\perp^2]$ changes sign to a negative value, with a similar peak at the value of $K$ where $\textrm{Re}[\Gamma_\parallel^2]$ changes sign. For $\delta=0$ we have $\Gamma_\perp=\Gamma_\parallel$ and there is only one such peak. The exceptions to this behavior are Rimbey-Mahan, which always displays only the $\Gamma_\perp$ peak and Fuchs-Kleiwer, which displays only the $\Gamma_\parallel$ peak.
The large $K$ behavior of $\textrm{Im}[r_p]$ at all frequencies means the function in the evanescent integral of (\ref{eq:utot}) without the exponential is proportional to $1/K^2$ as $K\to\infty$. As a result, the integral over evanescent waves converges to a finite value even in the $z=0$ case.
\begin{figure}[!htb]\centering
{\includegraphics[width=\linewidth]{Figure07.pdf}}
\caption{
Rescaled spectral energy density at $\omega=0.999\omega_T$ as a function of distance from the surface $z$.
Plot styles follow the conventions in Fig \ref{fig:rp_prop_angle}.
}
\label{fig:utotsmall}
\end{figure}
\begin{figure}[!htb]\centering
{\includegraphics[width=\linewidth]{Figure08.pdf}}
\caption{
Rescaled spectral energy density at $\omega=1.01\omega_T$ as a function of distance from the surface $z$.
Plot styles follow the conventions in Fig \ref{fig:rp_prop_angle}.
Agarwal \emph{et al}., Fuchs-Kleiwer and Rimbey-Mahan results are almost identical in the $\delta=0$ case.
}
\label{fig:utotlarge}
\end{figure}
Figures \ref{fig:utotsmall} and \ref{fig:utotlarge} show the spectral energy density $u_{tot}(z,\omega)$ divided by $u_0$ as function of distance from the boundary. Results for the various ABC's are shown together with the local result. Figure~\ref{fig:utotsmall} is for $\omega<\omega_T$ while figure~\ref{fig:utotlarge} is for $\omega>\omega_T$.
For the smaller frequency $\omega=0.999\omega_T$ (Fig. \ref{fig:utotsmall}), the integral is dominated by the smaller values of $K$ for which $\textrm{Im}[r_p]$ is very similar for all the ABC's. As a result the spectral energy density shows small differences between the ABC's while differing significantly from the local (diverging) result as $|z|\to0$. The effect of $\delta$ is negligible as it only affects large-$K$ values of $\textrm{Im}[r_p]$ that are already very small. The choice of $U_{ij}$ and $\delta$ is more significant at the larger frequency $\omega=1.01\omega_T$ (Fig. \ref{fig:utotlarge}).
In the $\delta\to0$ limit, the Agarwal \emph{et al}., Fuchs-Kleiwer and Rimbey-Mahan results are almost identical, while Ting \emph{et al}. is larger and Pekar is smaller.
Fuchs-Kleiwer is affected the most by $\delta$, followed by Ting \emph{et al}., Agarwal \emph{et al}. and finally Pekar, while Rimbey-Mahan remains unchanged.
Below 20nm for $0.999\omega_T$ and 8nm for $1.01\omega_T$, the spatially dispersive result begins to differ from the local medium.
These values of $|z|$ match the condition:
\begin{align}
\sigma^2\left(\frac{2\pi}{z}\right)^2=|\omega_T^2-\omega^2-i\gamma\omega|, \label{scale1}
\end{align}
since this distance corresponds to the wavelength of the polarization waves. The nonlocal $u_{tot}$ begins to saturate to a finite value below 1nm, removing the divergent $1/z^3$ behavior of the local medium. This distance is given by:
\begin{align}
\sigma^2\left(\frac{2\pi}{z}\right)^2=\omega_T^2, \label{scale2}
\end{align}
corresponding to wavelengths of the polarization waves below which the nonlocal term starts to dominate the resonance term ($\omega_T^2$). Both of the distance scales (\ref{scale1}) and (\ref{scale2}) depend on the relevant material parameters. For very small distances, certainly below 1nm, this model is no longer valid as other effects need to be included, e.g.\ higher-order terms of $k$ in the denominator of $\chi$, surface roughness, and quantum properties of the surface.
\begin{figure}[!htb]\centering
{\includegraphics[width=\linewidth]{Figure09.pdf}}
\caption{
Rescaled spectral energy density as a function of $\omega$ at a distance of 3nm (top) and 10nm (bottom) from the dielectric surface. Note the difference in scales.
Plot styles follow the conventions in Fig \ref{fig:rp_prop_angle}.
The Agarwal \emph{et al}., Fuchs-Kleiwer and Rimbey-Mahan results are almost identical in the $\delta=0$ case at $z$=3nm.
}
\label{fig:utotfreq}
\end{figure}
Figure \ref{fig:utotfreq} shows $u_{tot}$ as a function of frequency at fixed distances of 3nm and 10nm from the surface for the different ABC's.
Similar behavior can be observed in both cases, with two key features present.
The first is the small peak at $\omega_T$, which is a feature of the $\textrm{Im}[r_s]$ integral and as a result is unaffected by $\delta$. The second, larger peak at higher frequencies occurs when the sign of $\textrm{Re}[\Gamma_\perp^2]$ and $\textrm{Re}[\Gamma_\parallel^2]$ can change with $K$ leading to peaks in Im$[r_p]$ as a function of $K$.
This larger peak in $u_{tot}$ increases and broadens as the surface is approached.
In the $\delta\to0$ limit, Ting \emph{et al}. gives the largest value, Pekar is the smallest and
Agarwal \emph{et al}., Fuchs-Kleiwer and Rimbey-Mahan all take very similar intermediate values. The peak is strongly dependent on the value of $\delta$; for example, the greatest effect is in the Fuchs-Kleiwer peak, which varies by almost a factor of 3 over the range $-0.5<\delta<0.5$ at 3nm.
The effect of $\delta$ decreases in the order Ting \emph{et al}., Agarwal \emph{et al}. and finally Pekar.
This strongly contrasts with the almost negligible effect $\delta$ has on the reflection coefficient for propagating waves.
From these results it is clear that the tensor nature of the susceptibility and the difference between $\chi_\perp$ and $\chi_\parallel$ must be taken into consideration when considering the spectral energy density of zero-point and thermal radiation close to material boundaries.
We note that the thermal energy density near metal surfaces has been probed using near-field microscopy\cite{wilde06}. The zero-point spectral energy density can be probed by measuring spontaneous emission rates close to a boundary\cite{pur46,dre68,bar98,Novotny}. In addition, curved boundaries experience a deforming force (Casimir ``self-force") due to the local zero-point and thermal radiation\cite{boy68,mil80,candelas1982,hor14}, although this effect will be difficult to measure experimentally in any direct manner.
\section{Conclusion}\label{sec:conclusion}
We have derived exact expressions for reflection and transmission coefficients at a boundary of an isotropic spatially dispersive dielectric, taking into account that such a material has a tensor susceptibility.
Surface effects have been included by introducing phenomenological reflection coefficients $U_{ij}$ for polarization waves at the boundary. We have compared the effect of specific values of $U_{ij}$ corresponding to different ABC sets in the literature and also the effect of the inequality between the transverse and longitudinal susceptibilities ($\chi_\perp$ and $\chi_\parallel$ ). As noted by Halevi and Fuchs \cite{Halevi}, the coefficients $U_{ij}$ will in reality depend on frequency, in contrast to the simple constant values assumed in the ABC sets.
The reflection coefficient for $s$-polarization has already been derived by Halevi and Fuchs \cite{Halevi} so here we looked in detail at the $p$-polarization reflection coefficient and transmission coefficients. For propagating waves, differences between $\chi_\perp$ and $\chi_\parallel$ have the greatest effect on $r_p$ near the reflection minima, but it is the choice of ABC that has a far more significant effect on $r_p$.
We also considered in detail the zero-point and thermal spectral energy density $u_{tot}(z,\omega)$ outside the dielectric. The inclusion of spatial dispersion naturally removes the $1/z^3$ divergence of the local-medium result so that $u_{tot}(z,\omega)$ attains a constant value at distances of the order of 1nm, depending on material parameters. The inequality between $\chi_\perp$ and $\chi_\parallel$ was found to have a very significant effect on the maximum of $u_{tot}(z,\omega)$ as a function of $\omega$, even at a distance of 10nm from the surface. These results demonstrate that divergences in the (regularized) zero-point energy density and stress at planar boundaries\cite{sop02,bar05} are due to the neglect of spatial dispersion\cite{hor14}. Similar divergences in zero-point and thermal radiation at curved boundaries should also be removed by including non-local response\cite{hor14} and this will enable proper estimates of Casimir self-forces on objects like the dielectric ball and spherical shell\cite{boy68,mil80,candelas1982}.
The ABCs investigated here arose from consideration of different materials and models (as described in Sec.~\ref{sec:p-polarization}). The question of which ABC is appropriate for a given dielectric is difficult to assess in practice, given the necessarily simplified analysis which leads to the ABCs. In the case of the conduction electrons of a metal, far more is known both from theory and experiment\cite{raz11,wie12,pen12,tos15,sch16}. Comparison of further experimental results with the predictions of different ABCs (such as those calculated here) may shed light on this interesting question.
\acknowledgments
We thank S.A.R. Horsley for helpful discussions.
|
\section{Introduction}\label{sec:intro}
The problem of an unknown common change in means of the panels is studied here, where the panel data consist of $N$ panels and each panel contains $T$ observations over time. Various values of the change are possible for each panel at some unknown common time $\tau=1,\ldots,N$. The panels are considered to be independent, but this restriction can be weakened. In spite of that, observations within the panel are usually not independent. It is supposed that a~common unknown dependence structure is present over the panels.
Tests for change point detection in the panel data have been proposed only in case when the panel size $T$ is sufficiently large, i.e., $T$ increases over all limits from an asymptotic point of view, cf.~\cite{CHH2013} or~\cite{HH2012}. However, the change point estimation has already been studied for finite $T$ not depending on the number of panels $N$, see~\cite{Bai2010} or~\cite{PP2016}. The remaining task is to develop testing procedures to decide whether a~common change point is present or not in the panels, while taking into account that the length $T$ of each observation regime is fixed and can be relatively small.
\section{Panel Change Point Model}\label{sec:model}
Let us consider the panel change point model
\begin{equation}\label{model}
Y_{i,t}=\mu_i+\delta_i\mathcal{I}\{t>\tau\}+\sigma\varepsilon_{i,t},\quad 1\leq i\leq N,\, 1\leq t\leq T;
\end{equation}
where $\sigma>0$ is an unknown variance-scaling parameter and $T$ is fixed, not depending on $N$. The possible \emph{common change point time} is denoted by $\tau\in\{1,\ldots,T\}$. A~situation where $\tau=T$ corresponds to \emph{no change} in means of the panels. The means $\mu_i$ are panel-individual. The amount of the break in mean, which can also differ for every panel, is denoted by $\delta_i$. Furthermore, it is assumed that the sequences of panel disturbances $\{\varepsilon_{i,t}\}_t$ are independent and within each panel the errors form a~weakly stationary sequence with a~common correlation structure. This can be formalized in the following assumption.
\begin{assump}\label{ass:A1}
\normalfont The vectors $[\varepsilon_{i,1},\ldots,\varepsilon_{i,T}]^{\top}$ existing on a~probability space $(\Omega,\mathcal{F},\mathbbm{P})$ are $iid$ for $i=1,\ldots,N$ with $\mathsf{E}\varepsilon_{i,t}=0$ and $\mathsf{Var}\,\varepsilon_{i,t}=1$, having the autocorrelation function
\[
\rho_t=\mathsf{Corr}\,\left(\varepsilon_{i,s},\varepsilon_{i,s+t}\right)=\mathsf{Cov}\,\left(\varepsilon_{i,s},\varepsilon_{i,s+t}\right),\quad\forall s\in\{1,\ldots,T-t\},
\]
which is independent of the lag $s$, the cumulative autocorrelation function
\[
r(t)=\mathsf{Var}\,\sum_{s=1}^t \varepsilon_{i,s}=\sum_{|s|<t}(t-|s|)\rho_s,
\]
and the shifted cumulative correlation function
\[
R(t,v)=\mathsf{Cov}\,\left(\sum_{s=1}^t\varepsilon_{i,s},\sum_{u=t+1}^v\varepsilon_{i,u}\right)=\sum_{s=1}^t\sum_{u=t+1}^v\rho_{u-s},\quad t<v
\]
for all $i=1,\ldots,N$ and $t,v=1,\ldots,T$.
\end{assump}
The sequence $\{\varepsilon_{i,t}\}_{t=1}^T$ can be viewed as a~part of a~\emph{weakly stationary} process. Note that the dependent errors within each panel do not necessarily need to be linear processes. For example, GARCH processes as error sequences are allowed as well. The assumption of independent panels can indeed be relaxed, but it would make the setup much more complex. Consequently, probabilistic tools for dependent data need to be used (e.g., suitable versions of the central limit theorem). Nevertheless, assuming, that the claim amounts for different insurance companies are independent, is reasonable. Moreover, the assumption of a~common homoscedastic variance parameter $\sigma$ can be generalized by introducing weights $w_{i,t}$, which are supposed to be known. Being particular in actuarial practice, it would mean to normalize the total claim amount by the premium received, since bigger insurance companies are expected to have higher variability in total claim amounts paid.
It is required to test the \emph{null hypothesis} of no change in the means
\[
H_0:\,\tau=T
\]
against the~\emph{alternative} that at least one panel has a~change in mean
\[
H_1:\,\tau<T\quad\mbox{and}\quad\exists i\in\{1,\ldots,N\}:\,\delta_i\neq 0.
\]
\section{Test Statistic and Asymptotic Results}\label{sec:results}
We propose a~\emph{ratio type statistic} to test $H_0$ against $H_1$, because this type of statistic does not require estimation of the nuisance parameter for the variance. Generally, this is due to the fact that the variance parameter simply cancels out from the nominator and denominator of the statistic.
For surveys on ratio type test statistics, we refer to \cite{CH1997}, \cite{Barborka2011}, and~\cite{HHH2008}. Our particular panel change point CUSUM test statistic is
\[
\mathcal{C}_N(T)=\frac{1}{\sqrt{N}}\max_{t=1,\ldots,T-1}\left|\sum_{i=1}^N\sum_{s=1}^t\left(Y_{i,s}-\widebar{Y}_{i,T}\right)\right|,
\]
which is going to be compared with the ratio test statistics
\[
\mathcal{R}_N(T)=\max_{t=2,\ldots,T-2}\frac{\max_{s=1,\ldots,t}\left|\sum_{i=1}^N\sum_{r=1}^s\left(Y_{i,r}-\widebar{Y}_{i,t}\right)\right|}{\max_{s=t,\ldots,T-1}\left|\sum_{i=1}^N\sum_{r=s+1}^T\left(Y_{i,r}-\widetilde{Y}_{i,t}\right)\right|},
\]
where $\widebar{Y}_{i,t}$ is the average of the first $t$ observations in panel $i$ and $\widetilde{Y}_{i,t}$ is the average of the last $T-t$ observations in panel $i$, i.e.,
\[
\widebar{Y}_{i,t}=\frac{1}{t}\sum_{s=1}^t Y_{i,s}\quad\mbox{and}\quad\widetilde{Y}_{i,t}=\frac{1}{T-t}\sum_{s=t+1}^T Y_{i,s}.
\]
elaborated in~\cite{PP2015}. It will be demonstrated by simulations that $\mathcal{R}_N(T)$ keeps the theoretical significance level, however, $\mathcal{C}_N(T)$ does not.
Firstly, we derive the behavior of the test statistics under the null hypothesis.
\begin{theorem}[Under Null]\label{underNull}
Under hypothesis $H_0$ and Assumption~\ref{ass:A1}
\[
\mathcal{C}_N(T)\xrightarrow[N\to\infty]{\mathscr{D}}\sigma\max_{t=1,\ldots,T-1}\left|X_t-\frac{t}{T}X_T\right|
\]
and
\[
\mathcal{R}_N(T)\xrightarrow[N\to\infty]{\mathscr{D}}\max_{t=2,\ldots,T-2}\frac{\max_{s=1,\ldots,t}\left|X_s-\frac{s}{t}X_t\right|}{\max_{s=t,\ldots,T-1}\left|Z_s-\frac{T-s}{T-t}Z_t\right|},
\]
where $Z_t:=X_T-X_t$ and $[X_1,\ldots,X_T]^{\top}$ is a~multivariate normal random vector with zero mean and covariance matrix $\bm{\Lambda}=\{\lambda_{t,v}\}_{t,v=1}^{T,T}$ such that
\[
\lambda_{t,t}=r(t)\quad\mbox{and}\quad\lambda_{t,v}=r(t)+R(t,v),\,\, t<v.
\]
\end{theorem}
The limiting distribution does not depend on the variance nuisance parameter $\sigma$, but it depends on the unknown correlation structure of the panel change point model, which has to be estimated for testing purposes. The way of its estimation is shown in Section~\ref{sec:covest}. Note that in case of independent observations within the panel, the correlation structure and, hence, the covariance matrix $\bm{\Lambda}$ is simplified such that $r(t)=t$ and $R(t,v)=0$.
Next, we show how the test statistic behaves under the alternative
\begin{assump}\label{alternativeDelta}
\normalfont $\lim_{N\to\infty}\frac{1}{\sqrt{N}}\left|\sum_{i=1}^N\delta_i\right|=\infty$.
\end{assump}
\begin{theorem}[Under Alternative]\label{underAlternative}
If $\tau\leq T-3$, then under Assumptions~\ref{ass:A1}, \ref{alternativeDelta} and alternative $H_1$
\begin{equation}
\mathcal{C}_{N}(T)\xrightarrow[N\to\infty]{\mathbbm{P}}\infty\quad\mbox{and}\quad\mathcal{R}_{N}(T)\xrightarrow[N\to\infty]{\mathbbm{P}}\infty.
\end{equation}
\end{theorem}
Assumption~\ref{alternativeDelta} is satisfied, for instance, if $0<\delta\leq\delta_i\,\forall i$ (a~common lower change point threshold) and $\delta\sqrt{N}\to\infty,\, N\to\infty$. Another suitable example of $\delta_i$s for the condition in Assumption~\ref{alternativeDelta}, can be $0<\delta_i=KN^{-1/2+\eta}$ for some $K>0$ and $\eta>0$. Or $\delta_i=Ci^{\alpha-1}\sqrt{N}$ may be used as well, where $\alpha\geq 0$ and $C>0$. The assumption $\tau\leq T-3$ means that there are at least three observations in the panel after the change point. It is also possible to redefine the test statistic by interchanging the nominator and the denominator of $\mathcal{S}_{N}(T)$. Afterwards, Theorem~\ref{underAlternative} for the modified test statistic would require three observations before the change point, i.e., $\tau\geq 3$.
Theorem~\ref{underAlternative} says that in presence of a~structural change in the panel means, the test statistic explodes above all bounds. Hence, the procedure is consistent and the asymptotic distribution from Theorem~\ref{underNull} can be used to construct the test.
\section{Estimation of the Covariance Structure}\label{sec:covest}
The estimation of the covariance matrix $\bm{\Lambda}$ from Theorem~\ref{underNull} requires panels as vectors with elements having common mean (i.e., without a~jump). Therefore, it is necessary to construct an~estimate for a~possible change point. A~\emph{consistent estimate} of the change point $\tau$ in the panel data is proposed in~\cite{PP2016} as
\begin{equation}\label{tauhat}
\widehat{\tau}_N:=\arg\min_{t=2,\ldots,T}\frac{1}{w(t)}\sum_{i=1}^N\sum_{s=1}^t(Y_{i,s}-\widebar{Y}_{i,t})^2,
\end{equation}
where $\{w(t)\}_{t=2}^T$ is a~sequence of weights specified in~\cite{PP2016}.
Since the panels are considered to be independent and the number of panels may be sufficiently large, one can estimate the correlation structure of the errors $[\varepsilon_{1,1},\ldots,\varepsilon_{1,T}]^{\top}$ empirically. We base the errors' estimates on \emph{residuals}
\begin{equation}\label{widehate}
\widehat{e}_{i,t}:=\left\{
\begin{array}{ll}
Y_{i,t}-\widebar{Y}_{i,\widehat{\tau}_N},& t\leq\widehat{\tau}_N,\\
Y_{i,t}-\widetilde{Y}_{i,\widehat{\tau}_N},& t>\widehat{\tau}_N.
\end{array}
\right.
\end{equation}
Then, the empirical version of the autocorrelation function is
\[
\widehat{\rho}_t:=\frac{1}{\widehat{\sigma}^2 NT}\sum_{i=1}^N\sum_{s=1}^{T-t}\widehat{e}_{i,s}\widehat{e}_{i,s+t}.
\]
Consequently, the kernel estimation of the cumulative autocorrelation function and shifted cumulative correlation function is adopted in lines with~\cite{Andrews1991}:
\begin{align*}
\widehat{r}(t)&=\sum_{|s|<t}(t-|s|)\kappa\left(\frac{s}{h}\right)\widehat{\rho}_s,\\
\widehat{R}(t,v)&=\sum_{s=1}^t\sum_{u=t+1}^v\kappa\left(\frac{u-s}{h}\right)\widehat{\rho}_{u-s},\quad t<v;
\end{align*}
where $h>0$ stands for the window size and $\kappa$ belongs to a~class of kernel
\begin{multline*}
\Big\{\kappa(\cdot):\,\mathbbm{R}\to[-1,1]\,\big|\,\kappa(0)=1,\,\kappa(x)=\kappa(-x),\,\forall x,\,\int_{-\infty}^{+\infty}\kappa^2(x)\mbox{d} x<\infty,\Big.\\
\Big.\kappa(\cdot)\,\mbox{ is continuos at $0$ and at all but a finite number of other points}\Big\}.
\end{multline*}
Since the variance parameter $\sigma$ simply cancels out from the limiting distribution of Theorem~\ref{underNull}, it neither has to be estimated nor known. Nevertheless, one can use $\widehat{\sigma}^2:=\frac{1}{NT}\sum_{i=1}^{N}\sum_{s=1}^{T}\widehat{e}_{i,s}^2$.
\section{Simulations}\label{sec:simul}
A~simulation experiment was performed to study the \emph{finite sample} properties of the test statistics for a~common change in panel means. In particular, the interest lies in the empirical \emph{sizes} of the proposed tests (i.e., based on $\mathcal{R}_{N}(T)$ and $\mathcal{S}_{N}(T)$) under the null hypothesis and in the empirical \emph{rejection} rate (power) under the alternative. Random samples of panel data ($5000$ each time) are generated from the panel change point model~\eqref{model}. The panel size is set to $T=10$ and $T=25$ in order to demonstrate the performance of the testing approaches in case of small and intermediate panel length. The number of panels considered is $N=50$ and $N=200$.
The correlation structure within each panel is modeled via random vectors generated from iid, AR(1), and GARCH(1,1) sequences. The considered AR(1) process has coefficient $\phi=0.3$. In case of GARCH(1,1) process, we use coefficients $\alpha_0=1$, $\alpha_1=0.1$, and $\beta_1=0.2$, which according to \citet[Example~1]{Lindner2009} gives a~strictly stationary process. In all three sequences, the innovations are obtained as iid random variables from a~standard normal $\mathsf{N}(0,1)$ or Student $t_5$ distribution. Simulation scenarios are produced as all possible combinations of the above mentioned settings.
When using the asymptotic distribution from Theorem~\ref{underNull}, the covariance matrix is estimated as proposed in Section~\ref{sec:covest} using the Parzen kernel
\[
\kappa_{P}(x)=\left\{\begin{array}{ll}
1-6x^2+6|x|^3, & 0\leq|x|\leq 1/2;\\
2(1-|x|)^3, & 1/2\leq|x|\leq 1;\\
0, & \mbox{otherwise}.
\end{array}
\right.
\]
Bartlett window as well
Several values of the smoothing window width $h$ are tried from the interval $[2,5]$ and all of them work fine providing comparable results. To simulate the asymptotic distribution of the test statistics, $2000$ multivariate random vectors are generated using the pre-estimated covariance matrix. To access the theoretical results under $H_0$ numerically, Table~\ref{tab:H0} provides the empirical specificity (one minus size) of the asymptotic tests based on $\mathcal{R}_{N}(T)$ and $\mathcal{S}_{N}(T)$, where the significance level is $\alpha=5\%$.
\begin{table}[!ht]
\caption{Empirical specificity ($1-$size) of the test under $H_0$ for test statistics \colorbox{sedaseda}{$\mathcal{R}_{N}(T)$} and \colorbox{seda}{$\mathcal{C}_{N}(T)$} using the asymptotic critical values, considering a~significance level of $5\%$, weight function $w(t)=t^2$, and smoothing window width $h=2$}
\label{tab:H0}
\begin{center}
\begin{tabular}{cccx{.85cm}x{.85cm}x{.85cm}x{.85cm}x{.85cm}x{.85cm}}
\toprule
$T$ & $N$ & innovations & \multicolumn{2}{c}{IID} & \multicolumn{2}{c}{AR(1)} & \multicolumn{2}{c}{GARCH(1,1)} \tabularnewline
\midrule
$10$ & $50$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.948$} & \cellcolor[gray]{0.9}{$.934$} & \cellcolor[gray]{0.8}{$.933$} & \cellcolor[gray]{0.9}{$.823$} & \cellcolor[gray]{0.8}{$.946$} & \cellcolor[gray]{0.9}{$.935$} \tabularnewline
& & $t_5$ & \cellcolor[gray]{0.8}{$.951$} & \cellcolor[gray]{0.9}{$.925$} & \cellcolor[gray]{0.8}{$.932$} & \cellcolor[gray]{0.9}{$.822$} & \cellcolor[gray]{0.8}{$.946$} & \cellcolor[gray]{0.9}{$.929$} \tabularnewline
& $200$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.950$} & \cellcolor[gray]{0.9}{$.933$} & \cellcolor[gray]{0.8}{$.939$} & \cellcolor[gray]{0.9}{$.825$} & \cellcolor[gray]{0.8}{$.950$} & \cellcolor[gray]{0.9}{$.938$} \tabularnewline
& & $t_5$ & \cellcolor[gray]{0.8}{$.948$} & \cellcolor[gray]{0.9}{$.927$} & \cellcolor[gray]{0.8}{$.935$} & \cellcolor[gray]{0.9}{$.821$} & \cellcolor[gray]{0.8}{$.948$} & \cellcolor[gray]{0.9}{$.937$} \tabularnewline
\cmidrule(){1-9}
$25$ & $50$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.946$} & \cellcolor[gray]{0.9}{$.940$} & \cellcolor[gray]{0.8}{$.932$} & \cellcolor[gray]{0.9}{$.780$} & \cellcolor[gray]{0.8}{$.947$} & \cellcolor[gray]{0.9}{$.945$} \tabularnewline
& & $t_5$ & \cellcolor[gray]{0.8}{$.948$} & \cellcolor[gray]{0.9}{$.945$} & \cellcolor[gray]{0.8}{$.932$} & \cellcolor[gray]{0.9}{$.790$} & \cellcolor[gray]{0.8}{$.946$} & \cellcolor[gray]{0.9}{$.943$} \tabularnewline
& $200$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.949$} & \cellcolor[gray]{0.9}{$.939$} & \cellcolor[gray]{0.8}{$.930$} & \cellcolor[gray]{0.9}{$.801$} & \cellcolor[gray]{0.8}{$.951$} & \cellcolor[gray]{0.9}{$.939$} \tabularnewline
& & $t_5$ & \cellcolor[gray]{0.8}{$.953$} & \cellcolor[gray]{0.9}{$.941$} & \cellcolor[gray]{0.8}{$.931$} & \cellcolor[gray]{0.9}{$.813$} & \cellcolor[gray]{0.8}{$.952$} & \cellcolor[gray]{0.9}{$.946$} \tabularnewline
\bottomrule
\end{tabular}
\end{center}
\end{table}
It may be seen that both approaches are close to the theoretical value of specificity $.95$. As expected, the best results are achieved in case of independence within the panel, because there is no information overlap between two consecutive observations. The precision of not rejecting the null is increasing as the number of panels is getting higher and the panel is getting longer as well.
The performance of both testing procedures under $H_1$ in terms of the empirical rejection rates is shown in Table~\ref{tab:H1}, where the change point is set to $\tau=\lfloor T/2 \rfloor$ and the change sizes $\delta_i$ are independently uniform on $[1,3]$ in $33\%$, $66\%$ or in all panels.
\begin{table}[!ht]
\caption{Empirical sensitivity (power) of the test under $H_1$ for test statistics \colorbox{sedaseda}{$\mathcal{R}_{N}(T)$} and \colorbox{seda}{$\mathcal{C}_{N}(T)$} using the asymptotic critical values, considering a~significance level of $5\%$, weight function $w(t)=t^2$, and smoothing window width $h=2$}
\label{tab:H1}
\begin{center}
\begin{tabular}{ccccx{.95cm}x{.95cm}x{.95cm}x{.95cm}x{.95cm}x{.95cm}}
\toprule
$H_1$ & $T$ & $N$ & innovations & \multicolumn{2}{c}{IID} & \multicolumn{2}{c}{AR(1)} & \multicolumn{2}{c}{GARCH(1,1)} \tabularnewline
\midrule
$33\%$ & $10$ & $50$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.235$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.256$} & \cellcolor[gray]{0.9}{$.999$} & \cellcolor[gray]{0.8}{$.193$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & & $t_5$ & \cellcolor[gray]{0.8}{$.174$} & \cellcolor[gray]{0.9}{$.999$} & \cellcolor[gray]{0.8}{$.202$} & \cellcolor[gray]{0.9}{$.996$} & \cellcolor[gray]{0.8}{$.201$} & \cellcolor[gray]{0.9}{$.999$} \tabularnewline
& & $200$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.453$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.486$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.387$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & & $t_5$ & \cellcolor[gray]{0.8}{$.360$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.393$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.389$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
\cmidrule(l){2-10}
& $25$ & $50$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.376$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.394$} & \cellcolor[gray]{0.9}{$.992$} & \cellcolor[gray]{0.8}{$.312$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & & $t_5$ & \cellcolor[gray]{0.8}{$.294$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.301$} & \cellcolor[gray]{0.9}{$.993$} & \cellcolor[gray]{0.8}{$.312$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & $200$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.685$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.699$} & \cellcolor[gray]{0.9}{$.995$} & \cellcolor[gray]{0.8}{$.584$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & & $t_5$ & \cellcolor[gray]{0.8}{$.561$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.565$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.590$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
\cmidrule(){1-10}
$66\%$ & $10$ & $50$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.450$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.491$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.386$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & & $t_5$ & \cellcolor[gray]{0.8}{$.360$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.377$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.390$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & $200$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.774$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.807$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.677$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & & $t_5$ & \cellcolor[gray]{0.8}{$.642$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.692$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.688$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
\cmidrule(l){2-10}
& $25$ & $50$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.688$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.694$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.581$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & & $t_5$ & \cellcolor[gray]{0.8}{$.558$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.570$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.594$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & $200$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.951$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.959$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.905$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & & $t_5$ & \cellcolor[gray]{0.8}{$.874$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.888$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.906$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
\cmidrule(){1-10}
$100\%$ & $10$ & $50$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.641$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.667$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.563$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & & $t_5$ & \cellcolor[gray]{0.8}{$.519$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.547$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.546$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & $200$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.928$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.945$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.868$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & & $t_5$ & \cellcolor[gray]{0.8}{$.844$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.869$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.872$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
\cmidrule(l){2-10}
& $25$ & $50$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.873$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.884$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.792$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & & $t_5$ & \cellcolor[gray]{0.8}{$.760$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.771$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.789$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & $200$ & $\mathsf{N}(0,1)$ & \cellcolor[gray]{0.8}{$.997$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.997$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.985$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & & $t_5$ & \cellcolor[gray]{0.8}{$.977$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.982$} & \cellcolor[gray]{0.9}{$1.000$} & \cellcolor[gray]{0.8}{$.986$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
\bottomrule
\end{tabular}
\end{center}
\end{table}
One can conclude that the power of both tests increases as the panel size and the number of panels increase, which is straightforward and expected. Moreover, higher power is obtained when a~larger portion of panels is subject to have a~change in mean. The test power drops when switching from independent observations within the panel to dependent ones. Innovations with heavier tails (i.e., $t_5$) yield smaller power than innovations with lighter tails. Generally, the newly defined test statistic $\mathcal{S}_{N}(T)$ \emph{outperforms} $\mathcal{R}_{N}(T)$ in all scenarios with respect to the power.
Finally, an~early change point is discussed very briefly. We stay with standard normal innovations, iid observations within the panel, the size of changes $\delta_i$ being independently uniform on $[1,3]$ in all panels, and the change point is $\tau=3$ in case of $T=10$ and $\tau=5$ for $T=25$. The empirical sensitivities of both tests for small values of $\tau$ are shown in Table~\ref{tab:tau}.
\begin{table}[!ht]
\caption{Empirical sensitivity of the test for small values of $\tau$ under $H_1$ for test statistics \colorbox{sedaseda}{$\mathcal{R}_{N}(T)$} and \colorbox{seda}{$\mathcal{C}_{N}(T)$} using the asymptotic critical values, considering a~significance level of $5\%$, weight function $w(t)=t^2$, and smoothing window width $h=2$}
\label{tab:tau}
\begin{center}
\begin{tabular}{x{.55cm}x{.55cm}x{.65cm}x{.95cm}x{.95cm}x{.55cm}x{.55cm}x{.65cm}x{.95cm}x{.95cm}}
\toprule
$T$ & $\tau$ & $N$ & \multicolumn{2}{c}{$H_1$, iid, $\mathsf{N}(0,1)$} & $T$ & $\tau$ & $N$ & \multicolumn{2}{c}{$H_1$, iid, $\mathsf{N}(0,1)$} \tabularnewline
\midrule
$10$ & $3$ & $50$ & \cellcolor[gray]{0.8}{$.551$} & \cellcolor[gray]{0.9}{$1.000$} & $25$ & $5$ & $50$ & \cellcolor[gray]{0.8}{$.629$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
& & $200$ & \cellcolor[gray]{0.8}{$.867$} & \cellcolor[gray]{0.9}{$1.000$} & & & $200$ & \cellcolor[gray]{0.8}{$.927$} & \cellcolor[gray]{0.9}{$1.000$} \tabularnewline
\bottomrule
\end{tabular}
\end{center}
\end{table}
When the change point is not in the middle of the panel, the power of the test generally falls down. The source of such decrease is that the left or right part of the panel possesses less observations with constant mean, which leads to a~decrease of precision in the correlation estimation. Nevertheless, $\mathcal{S}_{N}(T)$ again outperforms $\mathcal{R}_{N}(T)$ even for early or late change points (the late change points are not numerically demonstrated here).
\section{Conclusions}\label{sec:concl}
In this paper, we consider the change point problem in panel data with fixed panel size. Occurrence of common breaks in panel means is tested. We introduce a~ratio type test statistic and derive its asymptotic properties. Under the null hypothesis of no change, the test statistic weakly converges to a~functional of the multivariate normal random vector with zero mean and covariance structure depending on the intra-panel covariances. As shown in the paper, these covariances can be estimated and, consequently, used for testing whether a~change in means occurred or not. This is indeed feasible, because the test statistic under the alternative converges to infinity in probability. Furthermore, the whole stochastic theory behind requires relatively simple assumptions, which are not too restrictive.
A~simulation study illustrates that even for small panel size, both investigated approaches---the newly derived one based on $\mathcal{S}_{N}(T)$ and the older one proposed in~\cite{PP2015}---work fine. One may judge that both methods keep the significance level under the null, while various simulation scenarios are considered. Besides that, the power of the test is higher in case of $\mathcal{S}_{N}(T)$ compared to $\mathcal{R}_{N}(T)$. Finally, the proposed method is applied to insurance data, for which the change point analysis in panel data provides an~appealing approach.
\subsection{Discussion}\label{subsec:disc}
Our setup can be modified by considering large panel size, i.e., $T\to\infty$. Consequently, the whole theory leads to convergences to functionals of Gaussian processes with a~covariance structure derived in a~very similar fashion as for fixed $T$. However, our motivation is to develop techniques for fixed and relatively small panel size.
Dependent panels may be taken into account and the presented work might be generalized for some kind of asymptotic independence over the panels or prescribed dependence among the panels. Nevertheless, our incentive is determined by a~problem from non-life insurance, where the association of insurance companies consists of a~relatively high number of insurance companies. Thus, the portfolio of yearly claims is so diversified, that the panels corresponding to insurance companies' yearly claims may be viewed as independent and neither natural ordering nor clustering has to be assumed.
|
\section{Introduction}
Frustration in magnetic systems occurs by competing exchange
interactions and leads frequently to disordered spin-liquid states
\cite{Nor09,Bal10,Lucile}. Recent progress in understanding transition
metal oxides with orbital degrees of freedom demonstrated many unusual
properties of systems with active $t_{2g}$ degrees of freedom --- they
are characterized by anisotropic hopping
\cite{Kha00,Har03,Dag08,Nic11,Wro10} which generates Ising-like orbital
interactions \cite{Jac04,Kha05,Jac07,Jac08,Kru09,Che09,Ryn10,Tro12,Che13},
similar to the
orbital superexchange in $e_g$ systems \cite{Dag04,Rei05}.
Particularly challenging are $4d$ and $5d$ transition metal oxides,
where the interplay between strong electron correlations and spin-orbit
interaction leads to several novel phases \cite{Wit14,Brz15}. In
iridates the spin-orbit interaction is so strong that spins and orbital
operators combine to new $S=1/2$ pseudospins at each site \cite{Jac09},
and interactions between these pseudospins decide about the magnetic
order in the ground state.
The $A_2$IrO$_3$ ($A$=Na, Li) family of honeycomb iridates has
attracted a lot of attention as these compounds have $t_{2g}$ orbital
degree of freedom and lie close to the exactly solvable $S=1/2$ Kitaev
model \cite{Kit06}. This model has a number of remarkable features,
including the absence of any symmetry breaking in its quantum Kitaev
spin-liquid (KSL) ground state, with gapless Majorana fermions
\cite{Kit06} and extremely short-ranged spin correlations \cite{Bas07}.
We emphasize that below we call a KSL also disordered spin-liquid
states which arise near the Kitaev points in presence of perturbing
Heisenberg interactions $\propto J$.
By analyzing possible couplings between the Kramers doublets it was
proposed that the microscopic model adequate to describe the honeycomb
iridates includes Kitaev interactions accompanied by Heisenberg exchange
in form of the Kitaev-Heisenberg (KH) model \cite{Jac09,Cha10}. Soon
after the experimental evidence was presented that several features of
the observed zigzag order are indeed captured by the KH model
\cite{Sin10,Liu11,Sin12,Cho12,Ye12,Comin,Gre13,Tro13,Cha13}. Its
parameters for $A_2$IrO$_3$ compounds
are still under debate at present \cite{Kat14,Val16}. One
finds also a rather unique crossover from the quasiparticle states to
a non-Fermi liquid behavior by varying the frustrated interactions
\cite{Tro14}.
Unfortunately, however, it was recently realized that this model does
not explain the observed direction of magnetic moments in Na$_2$IrO$_3$
and its extension is indeed necessary to describe the magnetic order
in real materials \cite{Chu15,Cha15}.
For example, bond-anisotropic interactions associated with the trigonal
distortions have to play a role to explain the differences between
Na$_2$IrO$_3$ and Li$_2$IrO$_3$ \cite{Rau15}, the two compounds with
quite different behavior reminiscent of the unsolved problem of
NaNiO$_2$ and LiNiO$_2$ in spin-orbital physics \cite{Rei05}.
On the other hand, the KH model
might be applicable in another honeycomb magnet $\alpha$-RuCl$_3$, see
e.g. a recent study of its spin excitation spectrum \cite{Ban16}.
Understanding the consequences of frustrated Heisenberg interactions on
the honeycomb lattice is very challenging and has stimulated several studies
\cite{Alb11,Cab11,Son16}. The KH model itself is highly nontrivial and
poses an even more interesting problem in the theory
\cite{Cha10,Cha13,Rau14,Oit15}: Kitaev term alone has
intrinsic frustration due to directional Ising-like interactions between
the spin components selected by the bond direction \cite{Kit06}.
In addition, these interactions are disturbed by nearest neighbor
Heisenberg exchange which triggers long-range order (LRO) sufficiently
far from the Kitaev points \cite{Cha10,Cha13,Rau14,Oit15}.
In general, ferromagnetic (FM) and antiferromagnetic (AF) interactions
coexist and the
phase diagram of the KH model is quite rich as shown in several previous studies \cite{Cha10,Cha13,Rau14,Oit15,Tre11,Ire14}.
Finally, the KH
model has also a very interesting phase diagram on the triangular
lattice \cite{Li15,Bec15,Jac15,Rou16}. These studies motivate better
understanding of quantum effects in the KH model on the honeycomb
lattice in the full range of its competing interactions.
The first purpose of this paper is to revisit the phase diagram of the KH
model and to investigate it further by combining exact diagonalization
(ED) result \cite{Cha13} with the self-consistent cluster mean field
theory (CMFT),
supplemented by the insights from the linear spin-wave (LSW) theory
and the second--order perturbation theory (SOPT).
The main advantage of CMFT is that it goes beyond
a single site mean field classical theory and gives not only the
symmetry-broken states with LRO, but also includes partly quantum
fluctuations, namely the ones within the considered clusters
\cite{Alb11,Brz12}. In this way the treatment is more balanced and may
allow for disordered states in cases when frustration of interactions
dominates. We present below a complete CMFT treatment of the phase
diagram which includes also the Kitaev term in MF part of the
Hamiltonian and covers the entire parameter space (in contrast to the
earlier prototype version of CMFT calculation on a single hexagon for
the KH model \cite{Got15}). Note that the CMFT complements the ED which
is unable to get symmetry breaking for a finite system, but nevertheless
can be employed to investigate the phase transitions in the present KH
model by evaluating the second derivative of the ground state energy to
identify phase transitions by its characteristic maxima \cite{Cha10,Cha13}.
ED result can be also used to recognize the type of magnetic order by
transforming to reciprocal space and computing spin-structure factor.
The second purpose is to investigate further the difference between
quantum KSL regions around both Kitaev points mentioned
in Ref. \cite{Cha13} and LRO/KSL boundaries.
The paper is organized as follows: In Sec. \ref{sec:KH} we introduce
the KH model and define its parameters. In Sec. \ref{sec:methods} we
present three methods of choice:
(i) the exact diagonalization in Sec. \ref{sec:ED},
(ii) the self-consistent CMFT in Sec. \ref{sec:CMFT}, and
(iii) linear spin wave theory in Sec. \ref{sec:lsw}.
An efficient method of solving the self-consistence problem obtained
within the CMFT is introduced in Sec. \ref{sec:lin}.
The numerical
results are presented and discussed in Sec. \ref{sec:res}:
(i) the phase transitions and the phase diagram are introduced in Sec.
\ref{sec:phd}, and
(ii) the phase boundaries, the values of the ground state energies
and the magnetic moments obtained by different methods are
presented and discussed in Secs. \ref{sec:qcen} and \ref{sec:qcom}, and
(iii) we discuss the compatibility of the Kitaev interaction with
different spin ordered states in Sec. \ref{sec:qcom}.
Spin correlations obtained for various phases are presented in
Sec. \ref{sec:ss}.
The dynamical spin susceptibility and spin structure factor are
presented for different phases in Sec. \ref{sec:chi}. Finally,
in Sec. \ref{sec:summa} we present the main conclusions and short
summary. The paper is supplemented with Appendix where we explain
the advantages of the linearization procedure implemented on the CMFT
on the example of a single hexagon.
\section{Kitaev-Heisenberg Model}
\label{sec:KH}
We start from the KH Hamiltonian with nearest neighbor interactions on
the honeycomb lattice in a form,
\begin{eqnarray}
{\cal H}&\equiv&
K\sum_{\langle ij\rangle\parallel \gamma} S_{i}^{\gamma} S_{j}^{\gamma}
+ J\sum_{\langle ij\rangle} \vc{S}_{i}\cdot \vc{S}_{j},
\label{ham_in}
\end{eqnarray}
where $\gamma=x$,$y$,$z$ labels the bond direction. The Kitaev term
$\propto K$ favors local bond correlations of the spin component
interacting on the particular bond. The superexchange $J$
is of Heisenberg form and alone would generate a LRO state,
antiferromagnetic or ferromagnetic, depending on whether $J>0$ or $J<0$.
We fix the overall energy scale, $J^{2}+K^{2}=1$, and choose angular
parametrization
\begin{eqnarray}
\label{k}
K&=&
\sin\varphi,\\
\label{j}
J&=&
\cos\varphi,
\end{eqnarray}
varying $\varphi$ within the interval $\varphi\in[0,2\pi]$. This
parametrization exhausts all the possibilities for nearest neighbor
interactions in the KH model.
While zigzag AF order was observed in Na$_2$IrO$_3$
\cite{Sin12,Cho12,Ye12,Comin,Gre13}, its microscopic explanation has been
under debate for a long time.
The \textit{ab initio} studies \cite{Foy13,Kat15} give motivation to
investigate a broad regime of parameters $K$ (\ref{k}) and $J$ (\ref{j}).
Further motivation comes from the honeycomb magnet $\alpha$-RuCl$_3$
\cite{Ban16}. Note that we do not intend to identify the parameter sets
representative for each individual experimental system, but shall
concentrate instead on the phase diagram of the model Eq. (\ref{ham_in})
with nearest neighbor interactions only.
\section{Calculation methods}
\label{sec:methods}
\subsection{Exact diagonalization}
\label{sec:ED}
We perform Lanczos diagonalization for $N=24$-site cluster with
periodic boundary conditions (PBC). This cluster respects all the
symmetries of the model, including hidden ones. Among the accessible
clusters it is expected to have the minimal finite-size effects.
\subsection{Cluster mean field theory}
\label{sec:CMFT}
A method which combines ED with an explicit breaking of Hamiltonian's
symmetries is the so-called self-consistent CMFT. It has
been applied to several models with frustrated interactions, including
Kugel-Khomskii model \cite{Brz12}. The method was also extensively used
by Albuquerque \textit{et al.} \cite{Alb11} as one of the means to
establish the full phase diagram of Heisenberg-$J_{2}$-$J_{3}$ model on
the honeycomb lattice.
Within CMFT the internal bonds of the cluster [connecting the circles
in Fig. \ref{1}(a)] are treated exactly. The corresponding part
$H_{\rm IN}$ of the Hamiltonian is the nearest neighbor KH Hamiltonian,
Eq. \eqref{ham_in}.
The external bonds that connect the boundary sites ($\bullet$) with
the corresponding boundary sites of periodic copies of the cluster
($\Box$) are described by the MF part of the Hamiltonian,
\begin{eqnarray}
H_{\rm MF}&\equiv&
K\sum_{[ij]\parallel z} \langle S_{i}^{z}\rangle S_{j}^{z}
+ J\sum_{[ ij]} \langle S_{i}^{z}\rangle S^{z}_{j},
\label{ham_mf}
\end{eqnarray}
where $[ij]$ marks the external bonds. Since the ordered moments in KH
model align always along one of the cubic axes $x$, $y$, $z$ (see e.g.
Ref.~\cite{Cha10}) we have put
\begin{equation}
\langle\vec{S}_i\rangle\cdot\vec{S}_j\equiv\langle S^z_i\rangle S^z_j
\label{SzSz}
\end{equation}
in $H_{\rm MF}$ to simplify the calculations.
\begin{figure}[t!]
\includegraphics[width=7cm]{fig1.pdf}
\caption{(a) 24-site cluster and the introduction of the mean fields.
Gray (black) circles indicate internal (boundary) sites.
In CMFT the internal bonds of the cluster are
treated exactly while the external bonds crossing the cluster boundary
(dashed) are treated on the MF level.
The sites marked by $\Box$ generate an effective magnetic fields
on the boundary sites $\bullet$.
Labels $x$, $y$ and $z$ stand for three inequivalent bond directions
determining the active products $S^{\gamma}_{i}S^{\gamma}_{j}$ in
Kitaev part of the Hamiltonian (\ref{ham_in}), e.g. bonds of $x$
direction contribute with the $S^x_iS^x_j$ product to the Hamiltonian.
The pseudospin axes used here are parallel to the cubic axes indicated
in the top view of a single octahedron.
(b) Unit cells: for honeycomb lattice (coinciding with single hexagon
of that lattice), for triangular lattice (inner dotted hexagon) and
zigzag magnetic unit cell (dashed rectangle). Black and white circles
indicate one of three equivalent zigzag patterns.
(c) Corresponding Brillouin zones and special $\vc{q}$ points for the
lattice constant $a=1$.
}
\label{1}
\end{figure}
The averages $\langle S_i^z\rangle$ generate effective magnetic fields
acting on the boundary sites of the cluster. The total Hamiltonian
\begin{eqnarray}
{\cal H}&\equiv&
H_{\rm IN}+H_{\rm MF},
\label{ham_num}
\end{eqnarray}
is diagonalized in a self-consistent manner, taking slightly different
approach than the one presented in Ref. \cite{Alb11}:
instead of starting with random wave function our algorithm begins
with expectation values $\langle S_i^z\rangle_\mathrm{ini}$
on each boundary site $i$ of the cluster. These can represent a certain
pattern (zigzag, stripy, N\'eel, FM) or be set randomly to have a
``neutral" starting point.
After diagonalizing the Hamiltonian
(\ref{ham_num}) (again by the ED Lanczos method) the ground state of
the system is obtained and we recalculate the expectation values
$\langle S^{z}_i\rangle$
to be used in the second iteration.
The procedure is repeated until self-consistency is reached.
\subsection{Linearized cluster mean field theory}
\label{sec:lin}
A single iteration of the self-consistent MF calculation may be viewed
as a nonlinear mapping of the set of initial averages
$\{\langle S_i^z\rangle_\mathrm{in}\}$ to the resulting averages
$\{\langle S_i^z\rangle_\mathrm{fin}\}$. The self-consistent solution is
then a stable stationary point of such a mapping. To find the leading
instability, we may consider the case of small initial averages in the
CMFT calculation and identify the pattern characterized by the fastest
growth during the iterations. To this end we linearize the above
mapping.
\begin{figure}[t!]
\includegraphics[width=8.2cm]{fig2.pdf}
\caption{
(a) The values of $\lambda$ obtained by the linearization
of CMFT for an embedded cluster of $N=24$ sites with fixed magnetic
order patterns: FM, AF, stripy, and zigzag.
Leading $\lambda >1$ indicates the order that sets in.
The disordered KSL phase is indicated by the red color.
(b) Second derivative of the ground state energy,
$-\mathrm{d}^2 E_0(\varphi)/\mathrm{d}\varphi^2$,
obtained by ED. Adopted from Ref.~\cite{Cha13}.
}
\label{la24}
\end{figure}
In the lowest order the mapping corresponds to the multiplication
of the vector of the averages $\{\langle S_i^z\rangle_\mathrm{in}\}$
by the matrix,
\begin{equation}
F_{ij} = \frac{\partial \langle S_i^z\rangle_\mathrm{fin}}
{\partial \langle S_j^z\rangle_\mathrm{in}},
\end{equation}
where $i$ and $j$ run through the cluster boundary sites. During iterations,
the patterns corresponding to the individual eigenvectors of the matrix
$F$ grow as $\lambda^n$, where $\lambda$ is a particular eigenvalue and
$n$ is the number of iterations. The ordering pattern obtained by CMFT
is then given by the eigenvector with largest $\lambda_{max}>1$. In the
quantum KSL regimes, all the eigenvalues are less than $1$
and no magnetic order emerges. An example of linearized CMFT applied to
a single hexagon with PBC can be found in the Appendix.
A modified version of this method, used to obtain Fig. \ref{la24}(a),
assumes a particular ordered pattern (N\'eel, zigzag, FM, or stripy
phase) and uses a single spin average $\langle S^z\rangle_\mathrm{in}$
distributed along the boundary sites outside the cluster, with the
signs consistent with this pattern. The resulting values,
$\langle S^z_i\rangle_\mathrm{fin}$, are then averaged correspondingly.
In this case the matrix $F$
is reduced to a single value $\lambda$ plotted in Fig. \ref{la24}(a).
We observe that the largest eigenvalue either drops below 1 when the
disordered KSL state takes over, or interchanges with
another eigenvalue corresponding to a different ordered phase.
\subsection{Linear spin-wave theory}
\label{sec:lsw}
The LSW method is a basic tool to determine
spin excitations and quantum corrections in systems with long-range
order \cite{Wal63}. For systems with coexisting AF and FM bonds quantum
corrections are smaller than for the N\'eel phase but are still
substantial for $S=1/2$ spins \cite{Rac02}. For the KH model the LSW
theory \cite{Cha10,Cha13,Cho12} has to be implemented separately for
each of the four ordered ground states: N\'eel (N),
zigzag (ZZ), FM, or stripy (ST). Then for a particular ground state
the Hamiltonian is rewritten in terms of the Holstein-Primakoff
bosons~\cite{Cho12,Mak15} and only quadratic terms
in bosonic operators are kept. The spectrum of such quadratic
Hamiltonian is finally obtained using the successive Fourier and
Bogoliubov transformations.
While the spin wave dispersion relations are usually of prime
interest~\cite{Cho12,Mak15,Cha10,Cha13}, there are also two other quantities
which can easily be calculated using LSW and which will be important
in the discussion that follows:
(i) the value of the total ordered moment $\langle M\rangle$ per site, and
(ii) the total energy per site $\langle E\rangle$.
These observables are calculated in a standard way~\cite{Wal63,Rac02}
and expressed in terms of the eigenvalues, i.e., spin-wave
energies $\omega_{{\bf k}\alpha}$, and the eigenvector components ($v_{{\bf k}
\alpha \lambda}$) of the bosonic
Hamiltonian {\it before} the Bogoliubov transformation:
\begin{align}
\langle M \rangle =S- \frac{1}{L V} \sum_{\alpha, \lambda = 1, ..., L}
\int_{{\bf k} \in BZ} |v_{{\bf k} \alpha, \lambda}|^2\ d^2{\bf k},
\label{m}
\end{align}
and
\begin{align}
\langle E \rangle =& E_{\rm class}\ [S^2 \rightarrow S(S+1)] \nonumber \\
&+ \frac{S}{2 L V} \sum_{\alpha = 1, ..., L} \int_{{\bf k} \in BZ}
{\omega_{{\bf k} \alpha}}\ d^2{\bf k},
\label{e}
\end{align}
where the choice of the sign of the eigenvalues and the normalization
of their eigenvectors is described in Ref.~\cite{Wal63}. Here
$E_\mathrm{class}$ is the classical ground state energy per site, e.g.
\begin{equation}
E_{\rm class} = -JzS^2/2,
\end{equation}
with $z=3$ for the N\'eel phase at $K=0$ and $S=1/2$ is the value of
spin quantum number. $L$ in Eqs. (\ref{m})-(\ref{e}) is the number of
the eigenvalues of the problem (spin-wave modes) and $\alpha$ enumerates
these modes. For all cases except for the zigzag order~\cite{Cha10},
the integrals go over the two-sublattice ($L=2$) rectangular Brillouin
zone (BZ)~\cite{Wei91} with its volume $V=8\pi^2 / 3\sqrt{3}$ and
$ -\pi / \sqrt{3} \le k_x \le \pi / \sqrt{3}$,
$-2 \pi /3 \le k_y \le 2 \pi / 3 $
(as already mentioned we assume the lattice constant $a=1$).
For the zigzag state $L=4$ and the rectangular BZ can be chosen as:
$ -\pi/\sqrt{3} \le k_x \le \pi/\sqrt{3}$ and
$- \pi / 3 \le k_y \le \pi / 3$
and its volume is $V=4 \pi^2 / 3\sqrt{3}$.
\section{Quantum phase transitions}
\label{sec:res}
\subsection{Phase diagram}
\label{sec:phd}
Here we supplement the ED--based phase diagram for the KH model established
in Ref.~\cite{Cha13} with the one obtained within CMFT.
Figure \ref{phase_diag} displays the phase boundaries obtained with ED
\cite{Cha13}, within CMFT, as well as classical (Luttinger-Tisza) phase
boundaries. The latter are included for completeness and
to highlight the fact that the quantum fluctuations stabilize the KSL
phases beyond single points, see below. To examine them in
more detail it is instructive to analyze the data in
Fig.~\ref{la24}(a) for the boundaries obtained from linearized CMFT and
Fig.~\ref{la24}(b) for the peaks in the second derivative of energy,
$-\mathrm{d}^2 E_0(\varphi)/\mathrm{d}\varphi^2$,
giving phase boundaries in ED~\cite{Cha13}.
\begin{figure}[t!]
\includegraphics[width=8.2cm]{fig3.pdf}
\caption{
T$=0$ phase diagram for KH model. The outer ring is composed
from ED data for the 24-site cluster, reproducing the result from
Ref. \cite{Cha13} in the new parametrization, the middle ring shows
CMFT results also for 24-site cluster and the inner black circle
represents the classical result. The convention used
for the angular parameter $\varphi$ which determines
coupling constants [see Eqs. (\ref{k}) and (\ref{j})] is shown in the
center of the inner circle. The colors represent particular phases,
shown also as mini-drawings next to suitable regions of the phase
diagram. Starting from $\varphi=0$ green colored region corresponds to
N\'eel order, red --- KSL, yellow --- zigzag order,
dark blue --- FM, red --- KSL, light blue --- stripy
phase and again green --- N\'eel phase.
}
\label{phase_diag}
\end{figure}
It is clearly visible that all the methods that include quantum
fluctuations give quantum versions of the four classically established
magnetic phases: N\'eel, zigzag, FM and stripy. As the most important
effect we note that when quantum fluctuations are included within a
classical phase, the energy is generally lowered and that the emerging
phase is expected to expand beyond the classical boundaries, but only
in case if a phase which competes with it has weaker quantum
fluctuations. This implies that phases of AF nature will expand at the expense
of the FM ones as the latter phases have lower energy gains by quantum
fluctuations (which even vanish exactly for the FM order at $K=0$ and
$J<0$).
We summarize the phase boundaries obtained within different methods
in Table \ref{table1}. One finds substantial corrections to the quantum
phase transitions which follow from quantum fluctuations. These
corrections are quite substantial in both KSLs at the Kitaev
points ($K=+1$, $\varphi=\frac{1}{2}\pi$ and $K=-1$,
$\varphi=\frac{3}{2}\pi$, first column of Table~\ref{table1}). Indeed,
in the classical approach massively degenerate ground states exist just
at isolated points but they are replaced by disordered spin-liquid states
that extend to finite intervals of $\varphi$ when
quantum fluctuations are included, see the second, third and fourth
column in Table \ref{table1}. The expansion of N\'eel and zigzag phases
beyond classical boundaries is given by particularly large corrections
and is well visible.
The most prominent feature in the phase diagram described above is
however the difference in size between two KSL regions, already
addressed before using ED \cite{Cha13} and also visible now in the CMFT
data. Therefore, the CMFT result supports the claim from Ref.
\cite{Cha13} that the stability of KSL perturbed by
relatively small Heisenberg interaction depends on the nature of the
phases surrounding the spin liquid and the amount of quantum
fluctuations that they carry.
In the following we discuss the above issues more thoroughly, examining:
(i) ground state energy curves emerging from ED, CMFT, SOPT within the linked
cluster expansion and LSW, (ii) the ordered moment given by various methods,
(iii) the spin--spin correlation functions, and (iv) the spin structure factor
as well as the dynamical spin susceptibility in the vicinity of the Kitaev
points.
\subsection{Quantum corrections: energetics}
\label{sec:qcen}
We start the discussion of quantum corrections to the energy of the ordered
phases by noting that, even though it properly captures finite order
parameters, the CMFT looses quantum energy on the external bonds and does not
therefore provide a reliable estimate of the ground-state energy. Instead, the
energy obtained using the ED calculations [see Fig. \ref{sz24}(a)] will be
treated as a reference value. This is supported by the fact that the ED phase
boundaries were roughly confirmed by tensor networks (iPEPS) \cite{Ire14} and
DMRG results \cite{Tre11}:
While the iPEPS phase
boundaries agree with ED for AF KSL/LRO transitions and
the boundaries between different LRO phases differ only slightly from
those found in ED (iPEPS: zigzag/FM -- $0.808\pi$, stripy/N\'eel --
$1.708\pi$), for FM KSL/LRO transition however the iPEPS result
KSL/stripy -- $1.528\pi$).
On the other hand, DMRG boundaries agree perfectly with ED and due to
four--sublattice dual transformation~\cite{Kha05,Cha10} one can
reproduce the FM/zigzag as well as FM/KSL boundaries. Only the
extent of the AF spin-liquid phase cannot be extracted from this result,
but that is already confirmed by iPEPS.
\begin{table}[t!]
\caption{Phase boundaries for KH model, parameterized by the angle
$\varphi$ (in units of $\pi$), see Eqs. (\ref{k}) and (\ref{j}).
Columns: classical Luttinger-Tisza approximation,
second--order perturbation theory,
exact diagonalization, and
self--consistent cluster mean field theory. }
\begin{ruledtabular}
\begin{tabular}{ccccc}
boundary & classical & SOPT & ED & CMFT \cr
\hline
N\'eel/KSL &$0.5$&$0.492$&$0.494 $&$0.493 $ \cr
KSL/zigzag&$0.5$&$0.507$&$0.506 $&$0.505$\cr
zigzag/FM&$ 0.75$&$0.813$&$0.814$ &$0.825$\cr
FM/KSL&$1.5$ &$1.463$&$1.448$&$1.481$\cr
KSL/stripy& $1.5$&$1.530$&$1.539$&$1.517$\cr
stripy/N\'eel&$1.75$&$1.705$&$1.704$&$1.699$\cr
\end{tabular}
\end{ruledtabular}
\label{table1}
\end{table}
Fig. \ref{sz24}(a) shows a quite remarkable agreement between the energy
values and critical values of $\varphi$ obtained by the simplest SOPT
\cite{Cha10} and our reference ED results.
This suggests that this analytical method can be utilized to get better
insight to the quantum contributions to the ground state energy.
For the four phases with LRO, the energy per site ${\cal E}$, written as a sum of
the classical energy $E_\mathrm{class}$ and the quantum fluctuation
contribution $\Delta E$, is obtained as:
\begin{eqnarray}
\label{enen}
{\cal E}_{\rm N} &=&-\frac18 (K+3J) -\frac{1}{16}(K+3J), \\
\label{enez}
{\cal E}_{\rm ZZ}&=&-\frac18 (K-J) -\frac{1}{16}(K-J) , \\
\label{enef}
{\cal E}_{\rm FM}&=&+\frac18 (K+3J) +\frac{1}{16} \frac{K^2}{K+2J}, \\
\label{enes}
{\cal E}_{\rm ST}&=&+\frac18 (K-J) +\frac{1}{16}\frac{(K+2J)^2}{K}.
\end{eqnarray}
In addition, to get the LRO/KSL phase boundary points in Table \ref{table1},
we estimate the energy of the KSL phase as
\begin{equation}
{\cal E}_{\rm KSL}\simeq
\frac{3}{2}(K+J)\langle S^\gamma S^\gamma\rangle_\mathrm{Kitaev},
\end{equation}
using the analytical result for the Kitaev points \cite{Bas07},
$\langle S^\gamma S^\gamma\rangle_\mathrm{Kitaev}\approx\pm 0.131$.
\begin{figure}[t!]
\includegraphics[width=8.4cm]{fig4.pdf}
\caption{
(a) Comparison between ground state energies per site obtained
using various methods:
classical Luttinger-Tisza approximation (dashed black),
SOPT (solid red),
LSW approximation (dashed red),
ED for 24-site cluster (solid blue, see \cite{Cha13} for this result
in a different parametrization),
and CMFT (solid green).
(b) Ordered moment value obtained from CMFT (green line) and LSW
(dashed red line). For CMFT the values were obtained by averaging
$\pm\langle S_i^z\rangle$ over the boundary sites, with the signs
determined by the particular type of magnetic order.
}
\label{sz24}
\end{figure}
The two spin-liquid phases in the phase diagram of KH model differ
strongly in their extent, despite the formal equivalence of the FM
and AF Kitaev points provided by an exact mapping of the Hamiltonian
\cite{Kit06}. As mentioned earlier, this is due to the fact that the
two KSLs compete with LRO phases of a distinct nature. Here we
give a simple interpretation based on the strength of the quantum corrections
of the LRO phases estimated using \eqref{enen}--\eqref{enes}.
Later, in Secs. \ref{sec:ss} and \ref{sec:chi} we illustrate the different
nature of the transitions between FM and AF KSL and the surrounding it
LRO phases in terms of spin correlations and spin dynamics.
Let us now compare the quantum fluctuation contribution and the
classical one. For the LRO phases surrounding the AF spin liquid ---
N\'eel and zigzag --- we always have
$\Delta E/E_\mathrm{class}=\frac{1}{2}$ as deduced from Eqs.
(\ref{enen}) and (\ref{enez}), i.e., only $\frac23 {\cal E}_{\rm N}$
and $\frac23 {\cal E}_{\rm ZZ}$ are found in the classical approach.
This guarantees that the quantum phase transition between these two
types of order occurs at the same value of $\varphi=\pi/2$ in SOPT and
in the classical approach that do not capture the spin-liquid phase in
between these ordered states, see Fig.~\ref{sz24}(a).
In contrast, the phases neighboring to the FM
spin liquid --- FM and stripy --- would reach the value of
$\Delta E/E_\mathrm{class}=\frac12$ only at the FM Kitaev point with
$J=0$ and away from this point the contribution of quantum fluctuations
decreases rapidly allowing for large extent of the FM spin-liquid phase.
Note, that both these latter phases contain a point which is exactly
fluctuation free --- for FM phase when frustration is absent ($K=0$),
and for stripy phase it is related to the FM one by the interaction
transformation \cite{Cha15} at $K=-2J$.
Moving to the CMFT energy analysis (green line in Fig.~\ref{sz24}(a))
one should also keep in mind that within the CMFT method the external
bonds between $\langle S^z_i\rangle$ and $S^z_j$ do not include quantum
fluctuations fully. This implies worse estimate of the energy for
regions of the phase space that allow quantum fluctuations.
As a consequence the region of stability of FM spin-liquid phase is
smaller than that obtained in the ED. Finally, the estimates obtained
from LSW, which represents a harmonic approximation to the quantum
fluctuations, are typically better than those from the CMFT but not as
good as those from SOPT, see dashed red lines in Fig. ~\ref{sz24}(a).
As expected, the LSW energy fits well with ED curve for FM and stripy
phases with less quantum fluctuations and starts to diverge when
beyond quantum phase transitions within N\'eel and zigzag phases.
\subsection{Quantum corrections: ordered moment}
\label{sec:qcom}
As usual,
getting the correct value of the ordered moment turns out to
be a more difficult task than estimating the ground state energy. This
is primarily due to the fact that the ED does not capture the
symmetry-broken states and the ordered moment can only be indirectly
extracted from the $m^2$; moreover, the SOPT may not be reliable here.
Hence, we are mostly left with the results obtained with CMFT and LSW.
We discuss the corresponding data [shown in Fig.~\ref{sz24}(b)]
together with the several values given already in the literature.
Let us begin with the Heisenberg AF point $\varphi=0$: here it is
expected that the ordered moment should be
strongly reduced by quantum fluctuations.
LSW approximation estimates the ordered moment value at $0.248$ \cite{Wei91}.
Similar values were extracted from $m^2$ in quantum Monte Carlo
($0.268$ \cite{Reg89,Cas06,Low09}) and ED calculations ($0.270$ \cite{Alb11}).
In the last case however the authors admit that the set of clusters
for finite size scaling was chosen so as to make the best agreement
with quantum Monte Carlo.
Another method --- series expansion (high order perturbation theory)
\cite{Oit15} sets ordered moment value at a somewhat higher value of
$0.307$. While all the above results seem roughly consistent, CMFT
value seems to stand out ($0.374$ for $\varphi=0$). Nevertheless,
one should note that the ordered moment estimated from $m^2$ for
24--site cluster ED equals $0.45$ \cite{Alb11} which is above the CMFT value.
This suggests that at this point the finite size scaling is important.
Before transferring to the frustrated regime we briefly mention that
the the trivial ordered moment value at $\varphi=\pi$ is here correctly
reproduced by both CMFT and LSW. Besides, for the regions around the
fluctuation--free FM (and stripy) point the
ordered moments predicted by CMFT and LSW also match. Following the
ground state energy analysis, LSW gives the correct result because
quantum fluctuations contribution is small compared to the classical
state. The further we move towards the Kitaev points, however, the more
incorrect the LSW approximation should be because of the strong
reduction of the ordered moment due to the growing frustration.
In contrast, the lack of quantum fluctuations on the external bonds
makes CMFT steadily biased except for FM and stripy phases.
However, since for the internal part of the cluster the fluctuations
are still fully included, the frustration should be well handled and
CMFT should give more predictable results than LSW in frustrated parts
of the phase diagram.
Here it is also important to stress, that the series expansion captures
correctly the fluctuation--free point at $\varphi=\pi$ (FM) and
$\varphi=-\arctan 2$ (stripy) and predicts a broader region of FM
KSL phase \cite{Oit15}. The order parameter is also
qualitatively correctly estimated and is reduced more to $m\simeq 0.3$
for both N\'eel and zigzag phases \cite{Oit15}. However, while the
ordered moment values obtained by CMFT are consistent with the
four--sublattice dual transformation, the ordered moment data from the
high--order perturbation theory \cite{Oit15} are not, as the ordered
moment values differ at the points connected by the mapping.
Unfortunately the largest difference appears near the FM LRO/KSL
boundaries. This observation uncovers certain limits of the
high--order perturbation theory.
\subsection{Quantum corrections: naive interpretation}
\label{sec:naive}
Let us conclude the discussion of the quantum corrections with the
following more general observation:
Developing the argumentation presented by Iregui, Corboz, and Troyer
\cite{Ire14}, the dependence of the quantum correction to the energy
and to the ordered moment on the angle $\varphi$ suggests that the
Kitaev interaction is less ``compatible'' with the FM/stripy ground
states than with the N\'eel/zigzag ones. This can be understood in
the simple picture of the KH model on a 4-site segment of the honeycomb
lattice consisting of three bonds attached to a selected lattice site,
as presented below.
Starting with $\varphi=\pi$ (FM ground state. e.g. along the $z$
quantization axis) and increasing $\varphi$ leads to an addition of the
FM Kitaev term, which favors FM-aligned spins along the $x$, $y$, and
$z$ quantization axes for the $x$, $y$, and $z$ directional bonds,
respectively. It can easily be seen that, e.g. for the $x$ bond, the
eigenstate of the FM Kitaev-only Hamiltonian on that bond
($|\uparrow_x \uparrow_x \rangle$) has a $25\%$ overlap with the FM
ground state,
$|\langle \uparrow_z\uparrow_z|\uparrow_x\uparrow_x\rangle|^2=\frac14$.
In contrast, while a similar situation happens for the $y$ bond,
for the $z$ bond there is a $100\%$ overlap between such states.
Next, we perform a similar analysis for $\varphi=0$ and firstly assume
that we have a {\it classical} ground state. In this case for the
``unsatisfied'' bonds from the point of view of the increasing AF
Kitaev interaction we also obtain that the eigenstate of the AF
Kitaev-only Hamiltonian ($|\uparrow_x\downarrow_x\rangle$) on that bond
has a $25\%$ overlap with the classical N\'eel ground state --- e.g.:
$|\langle\uparrow_z\downarrow_z|\uparrow_x\downarrow_x\rangle|^2=
\frac{1}{4}$. However, this situation changes once we consider that
the spin quantum fluctuations dress the classical N\'eel ground state.
This can be best understood if we assumed the unrealistic but
insightful case of very strong quantum fluctuations destroying the
classical N\'eel ground state: then for the $x$ bond a singlet could be
stabilized and the overlap between such a state and the state
``favored'' by the Kitaev term increases to $50\%$:
$|\langle 0 | \uparrow_x \downarrow_x \rangle |^2=\frac{1}{2}$.
This suggests that the N\'eel ground state, which {\it contains}
quantum spin fluctuations, is more ``compatible'' with the states
``favored'' by the Kitaev terms than the FM ground state, resulting in
more stable values of ordered moment for N\'eel phase.
It seems that the above difference is visible in CMFT data but not in
LSW ones. We shall discuss this issue further by analyzing spin
correlations below.
\section{Spin correlations}
\label{sec:ss}
\begin{figure}[t!]
\includegraphics[width=8.2cm]{fig5.pdf}
\caption{
(a) Spin correlations $\langle\vc S_i\cdot\vc S_j\rangle$ obtained
within ED for the bonds between nearest neighbors (black line),
spin correlations of the components active in the Kitaev interaction,
$\langle S^\gamma_iS^\gamma_j\rangle$ (blue line), and complementary
spin components, $\langle S^{\bar\gamma}_i S^{\bar\gamma}_j\rangle$
(red line). Below further neighbor spin correlations
$|\langle \vc S_i\cdot \vc S_j\rangle|$ are shown:
(b) near the AF spin-liquid phase, and
(c) for the angle $\varphi$ interval including the FM spin-liquid phase.
}
\label{cor}
\end{figure}
Additional information about the ground state is given by spin--spin
correlation functions. In Fig. ~\ref{cor}(a) one can observe
isotropic stable $\langle S^{\gamma}_{i}S^{\gamma}_{j}\rangle$
correlations in almost the entire AF phase
($\langle\boldsymbol{S}_{i}\cdot\boldsymbol{S}_{j}\rangle\approx-0.36$
for $\varphi=0$), while for FM phase the anisotropy quickly develops
when moving away from FM Heisenberg point $\varphi=\pi$ (here
$\langle\boldsymbol{S}_{i}\cdot\boldsymbol{S}_{j}\rangle$ reaches the
classical value $0.25$). This again demonstrates that the AF (and
zigzag) phase is more robust and uniform than FM (and stripy) phase.
Moreover, spin-spin correlations allow us to confirm the disordered
regions around the Kitaev points as critical cases of quantum spin
liquid \cite{Tik11}. At the Kitaev points we observe the expected
undisturbed KSL pattern: non--zero values of nearest
neighbor correlations between spin components active in the Kitaev
interaction (blue curve in Fig. ~\ref{cor}(a)) and vanishing
correlations between complementary components (red curve). In contrast,
the next nearest and further neighbor correlations disappear, see Figs.
\ref{cor}(b) and \ref{cor}(c). While moving away from the Kitaev points
the absolute values of the correlations enter the regions of slow
growth --- these are signatures of the critical spin-liquid phases and
they look similar in AF and FM spin liquid cases. At some point however
proceeding further results in rapidly growing absolute values which
mark KSL/LRO boundaries.
Furthermore, Figs. \ref{cor}(b) and \ref{cor}(c) prove that there is
a qualitative difference between the two spin-liquid regimes. This is
observed in the rapid growth of spin correlations at the onset of LRO:
step-like jump visible in Fig. \ref{cor}(b) contrasts with smoother
crossover seen in Fig. \ref{cor}(c). Below we investigate this distinct
behavior by analyzing the dynamical spin susceptibility for various
available phases. After Fourier transformation of the $z$--component
correlations, we obtain the spin structure factor to be discussed in
the context of the spin susceptibility also in Sec.~\ref{sec:chi}.
\section{Spin susceptibility and excitations
in the vicinity of the Kitaev points}
\label{sec:chi}
Below we study the spin dynamics within the KH model by
analyzing the dynamical spin susceptibility at $T=0$,
\begin{equation}
\chi_{\alpha\alpha}(\vc q,\omega) = i \int_0^\infty \left\langle\Phi_0|
[S^\alpha_{\vc q}(t),S^\alpha_{-\vc q}(0)]
\right|\Phi_0\rangle\, \mathrm{e}^{i\omega t} \,\mathrm{d} t,
\end{equation}
with the Fourier-transformed spin operator defined via
\begin{equation}
S^\alpha_{\vc q} = \frac1{\sqrt{N}}
\sum_{\vc R} \mathrm{e}^{-i\vc q\cdot\vc R} S^\alpha_{\vc R}
\;
\end{equation}
and $|\Phi_0\rangle$ denoting the cluster ground state. For $\omega>0$,
the imaginary part of $\chi(\vc q,\omega)$ reads as
\begin{equation}
\chi''_{\alpha\alpha}(\vc q,\omega) = -\mathrm{Im}\,
\langle\mathrm{\Phi_0}|\, S^\alpha_{\vc q} \;
\frac1{\omega+E_\mathrm{GS}-\mathcal{H}+i\delta}
\; S^\alpha_{-\vc q} \,
|\mathrm{\Phi_0}\rangle \;,
\end{equation}
which can be conveniently expressed as a sum over the excited states
$\{|\nu\rangle\}$,
\begin{equation}\label{eq:chiexcsum}
\chi''_{\alpha\alpha}(\vc q,\omega) = \pi\sum_{|\nu\rangle}
|\langle\nu|S^\alpha_{-\vc q}|\Phi_0\rangle|^2
\delta(\omega-E_\mathrm{\nu}) \;,
\end{equation}
where the excitation energy $E_\nu$ is measured relative to the ground
state energy $E_{\mathrm{GS}}$. We have evaluated
$\chi_{\alpha\alpha}(\vc q,\omega)$ by ED on a hexagonal cluster of
$N=24$ sites. In the ED approach, the exact ground state of the cluster
$|\Phi_0\rangle$ is found by Lanczos diagonalization, the operator
$S^\alpha_{-\vc q}$ is applied, and the average of the resolvent
$1/(z-\mathcal{H})$ is determined by Lanczos method using normalized
$S^\alpha_{-\vc q}|\Phi_0\rangle$ as a starting vector \cite{Ful95}.
\begin{figure}[t!]
\includegraphics[width=8.4cm]{fig6.pdf}
\caption{
(a)~Dynamical spin susceptibility $\chi''(\vc q,\omega)$ obtained by
ED near the AF KSL phase at the characteristic wavevector of the AF
order, $\vc q=\Gamma'$.
(b)~The same for the zigzag wavevector $\vc q=M$.
(c)~Brillouin zone portraits of the spin-structure factor
$\langle S^z_{-\vc q} S^z_{\vc q}\rangle$ at $\varphi=87.5^\circ$,
$90^\circ$, and $92.5^\circ$ (interpolated from the ED data). The inner
hexagon is the Brillouin zone of the honeycomb lattice, the outer one
corresponds to the triangular lattice with the missing sites filled in.
(d,e)~The same as in panels (a,b) but for the interval containing the
FM ($\vc{q}=\Gamma$) and stripy ($\vc q=X$) phase.
(f)~Brillouin zone portraits of the spin-structure factor obtained at
$\varphi=255^\circ$, $270^\circ$, and $285^\circ$.
}
\label{susc}
\end{figure}
In our case of the KH model, the calculation generally requires a
relatively large number of Lanczos steps (up to one thousand) to
achieve convergence of the dense high-energy part of the spectrum.
Having the advantage of being exact, the method is limited by the
$\vc q$ vectors accessible for a finite cluster and compatible with
the PBC, and by finite-size effects due to small $N$. These concern
mainly the low-energy part of $\chi''$ and lead e.g. to an enlarged
gap of spin excitations in LRO phases of AF nature. Nevertheless,
a qualitative understanding can still be obtained.
The evolution of numerically obtained $\chi''$ with varying $\varphi$
is presented in Figs.~\ref{susc}(a) and \ref{susc}(b) for the region
including AF spin-liquid phase, as well as in Figs.~\ref{susc}(d) and
\ref{susc}(e) for the region including the FM spin-liquid phase.
The transitions are well visible at the characteristic $\vc q$ vectors
of the individual LRO phases. The structure factor pattern, see
Figs.~\ref{susc}(c) and \ref{susc}(f), changes accordingly between the
sharply peaked one in LRO phases and a wave-like form characteristic
for nearest neighbor correlations in the spin-liquid phases.
After entering the spin-liquid phase, further changes of the spin response are
very different for the AF and FM case. In the AF case, there is a sharp
transition --- a level crossing at our cluster, so that the ground state
changes abruptly. The original intense pseudo-Goldstone mode as well as many
other excited states become inactive in the spin-liquid phase. The observed
low--energy gap in $\chi''$ varies only slightly with $\varphi$.
In contrast, when entering the FM spin-liquid phase the excitation
that used to be the gapless magnon mode is characterized by a
gradually increasing gap which culminates at the Kitaev point.
Starting from the Kitaev point, the gradual reduction of the
low--energy gap in $\chi''$ due to the Heisenberg perturbation
manifests itself by a development of spin correlations beyond nearest
neighbors (already reported in Fig.~2 of Ref.~\cite{Cha10}) and an
increase of the static susceptibility to the magnetic field
Zeeman-coupled to the order parameter of the neighboring LRO phase.
This susceptibility then diverges at the transition point
(see also Fig.~3 of Ref.~\cite{Cha10}).
\section{Summary and conclusions}
\label{sec:summa}
In the present paper we studied the phase diagram of the Kitaev-Heisenberg
model by a combination of exact diagonalization and cluster mean field theory
(CMFT), supplemented by the insights from linear spin-wave theory and the
second--order perturbation theory. Both methods allowed to stabilize
previously known ordered phases: N\'eel, zigzag, FM and stripy. Moreover, the
ordered moment analysis provided by cluster mean field approach demonstrates
N\'eel--zigzag and FM--stripy connections described before \cite{Cha13}.
Compared to the previous CMFT studies utilizing $N=6$ site cluster (see Ref.
\cite{Got15} or the Appendix), we have used a sufficiently large cluster of
$N=24$ sites preserving the lattice symmetries and improving the ratio between
internal and boundary bonds. This led to a balanced approach which allowed us
to treat both ordered and disordered (spin-liquid) states on equal footing.
As the main result, the present study uncovers a fundamental difference
between the onset of broken symmetry phases in the vicinity of Kitaev
points with antiferromagnetic or ferromagnetic interactions. While the
spin liquids obtained at $K=+1$ and $K=-1$ are strictly equivalent and
can be transformed one into the other in the absence of Heisenberg
interactions (at $J=0$), spin excitations and quantum phase transitions
emerging at finite $J$ are very different in both cases.
For antiferromagnetic Kitaev spin liquid phase ($K\simeq1$) one finds
that a gap opens abruptly in $\chi''(\vc q,\omega)$ at
$\vc q=\Gamma^{'}$ and $\vc q=M$ when the ground state changes to the
critical Kitaev quantum spin liquid. This phase transition is abrupt
and occurs by level crossing. In contrast, for ferromagnetic spin liquid
$K\simeq -1$ the gaps in $\chi''(\vc q,\omega)$ at $\vc q=\Gamma$ and
$\vc q=X$ open gradually from the points of quantum phase transition
from ordered to disordered phase. With much weaker quantum corrections
for ordered phases in the regime of ferromagnetic Kitaev interactions,
the spin liquid is more robust near $K=-1$ as a phase that contains
quantum fluctuations and survives in a broader regime than near $K=1$
when antiferromagnetic Kitaev interactions are disturbed by increasing
(antiferromagnetic or ferromagnetic) Heisenberg interactions.
This behavior is reminiscent of the ferromagnetic Kitaev model in
a weak magnetic field \cite{Tik11}.
\acknowledgments
We thank Giniyat Khaliullin for insightful discussions.
We kindly acknowledge support by Narodowe Centrum Nauki
(NCN, National Science Center) under Project No. 2012/04/A/ST3/00331.
J. R. and J. C. were supported by Czech Science Foundation (GA\v{C}R)
under Project No. GJ15-14523Y and by the project CEITEC 2020 (LQ1601)
with financial support from the Ministry of Education, Youth and Sports
of the Czech Republic under the National Sustainability Programme II.
Access to computing and storage facilities owned by parties and
projects contributing to the National Grid Infrastructure MetaCentrum,
provided under the program ``Projects of Large Research, Development,
and Innovations Infrastructures" (CESNET LM2015042), is acknowledged.
G.~J.~is supported in part by the National Science Foundation under
Grant No. NSF PHY11-25915.
|
\section{Introduction}
Incompatibility lies deeply within quantum mechanics and many of the famous and key aspect of quantum theories have been traced to Heisenberg uncertainty principle, no cloning theorem, violations of Bell inequalities and other notions making use of compatibility, see \cite{HeinosaariMiyaderaZiman-compatibility} for recent review. In light of these discoveries compatibility in the framework of general probabilistic theories has been studied \cite{WolfPerezgarciaFernandez-measIncomp, BarnumHowardBarretLeifer-noBroadcast, BuschHeinosaariSchultzStevens-compatibility} in order to show the difference between classical and non-classical probabilistic theories. Also the connection of compatibility and steering in general probabilistic theories have been studied \cite{Banik-steering, UolaMoroder-steering}.
Recently incompatibility of measurements on quantum channels and combs has been in question \cite{SedlakReitznerChiribellaZiman-compatibility} as it potentially could be used as a resource in quantum theory in a similar ways as an incompatibility of measurements on quantum states \cite{HeinosaariMiyaderaZiman-compatibility}. The degree of compatibility (also called robustness of incompatibility) has been studied for measurements on channels and combs \cite{HeinosaariKiukasReitzner-robustness, WolfPerezgarciaFernandez-measIncomp, SedlakReitznerChiribellaZiman-compatibility}.
In the present article we study the notion of compatibility of measurements in the framework of probabilistic theories and we show that every two measurements are compatible if and only if the state space is a simplex. In one way this result has clear physical interpretation - classical state space is always a simplex and the existence of incompatible measurements is often seen as one of the main aspects of quantum theories.
The paper is organized as follows: Sec. \ref{sec:prelim} contains preliminary mathematical results and references. Note that Subsec. \ref{subsec:prelim-faces} contains the definition of maximal face that (to the best knowledge of present author) was not defined elsewhere (even though it has close tie to the notion of tangent half-space and tangent hyper-plane \cite[pp. 169]{Rockafellar-convex}) and is later used in Sec. \ref{sec:inc}. In Sec. \ref{sec:meas} the measurements are defined. In Sec. \ref{sec:inc} compatibility of measurements and degree of compatibility is defined and it is shown that all measurements are compatible if and only if the state space is a simplex. Also the linear program for compatibility of two two-outcome measurements is formulated.
\section{Preliminaries} \label{sec:prelim}
We present preliminary mathematical knowledge used in the paper. In all of the paper $E$ will denote a real, finite dimensional vector space equipped with the Euclidean topology and $K$ will denote non-empty compact convex subset of $E$. We will denote the convex hull of a set $X$ as $\conv(X)$, affine hull of a set $X$ as $\aff(X)$, interior of a set of a set $X$ as $\mathit{int}(X)$ and by $\partial K$ we will denote the boundary of $K$, i.e $\partial K = K \setminus \mathit{int}(K)$ as $K$ is closed.
\subsection{Structure of $A(K)$} \label{subsec:structure}
By $A(K)$ we will denote the set of real valued affine functions on $K$ and by $A(K)^+$ we will denote the set of positive affine functions on $K$, i.e. $f \in A(K)^+$ if and only if $f(x) \geq 0$ for every $x \in K$. We will denote constant functions by the value they attain. Since $K$ is compact and the functions $A(K)$ are continuous, every function reaches its maximum and minimum over $K$ at some point of $K$ and we can introduce the supremum norm for $f \in A(K)$ as
\begin{equation*}
\Vert f \Vert_A = \sup_{x \in K} |f(x)|.
\end{equation*}
The set $A(K)^+$ is:
\begin{itemize}
\item closed
\item convex, i.e. for $\lambda \in \mathbb{R}$, $0 \leq \lambda \leq 1$,$f_1, f_2 \in A(K)^+$ we have $\lambda f_1 + (1-\lambda)f_2 \in A(K)^+$
\item cone, i.e. for $\nu \in \mathbb{R}$, $f \in A(K)^+$ we have $\nu f_1 \in A(K)^+$
\item pointed, i.e. $A(K)^+ \cap (-A(K)^+) = \{0\}$
\item generating, i.e. for every $f \in A(K)$ we have $f_+, f_- \in A(K)^+$ such that $f = f_+ - f_-$.
\end{itemize}
The closed, pointed, convex cone $A(K)^+$ defines a partial order $\geq$ on $A(K)$ given for $f_1, f_2 \in A(K)$ as
\begin{equation*}
f_1 \geq f_2 \Leftrightarrow f_1 - f_2 \in A(K)^+
\end{equation*}
or equivalently $f_1 \geq f_2 \Leftrightarrow (f_1 - f_2)(x) \geq 0, \forall x \in K$. The partial order $\geq$ will play a role in our formulation of linear program for incompatibility of two-outcome measurements.
\begin{defin}
We say that $e \in A(K)^+$ is an order unit if for every $f \in A(K)^+$ there is some $\nu \in \mathbb{R}$, $\nu > 0$ such that
\begin{equation*}
\nu e \geq f.
\end{equation*}
\end{defin}
In the current setting it is easy to see that every strictly positive function is an order unit. We will omit the simple proof of the following fact.
\begin{prop}
$e \in A(K)^+$ is an order unit if and only if $e \in \mathit{int}(A(K)^+)$.
\end{prop}
We will also use the notion of a base of a cone.
\begin{defin}
Let $Q \subset E$ be a cone, then a set $\mathcal{B} \subset Q$ is called base of $Q$ if for every $0 \neq x \in Q$ there exist unique $y \in \mathcal{B}$ and $\lambda \in \mathbb{R}$ such that $x = \lambda y$.
\end{defin}
To formulate the linear programming problem we will also have to work with the dual space of $A(K)$, we will denote it $A(K)^*$. We will denote by $A(K)^{*+}$ the cone of positive functionals dual to $A(K)^+$, that is $\psi \in A(K)^{*+}$ if and only if for every $f \in A(K)^+$ we have $\psi(f) \geq 0$.
\begin{prop}
$A(K)^{*+}$ is a closed pointed convex cone.
\end{prop}
\begin{proof}
It is straightforward to see that $A(K)^{*+}$ is closed convex cone. It is pointed because $A(K)^+$ is generating.
\end{proof}
We define the dual norm for $\psi \in A(K)^*$ as
\begin{equation*}
\Vert \psi \Vert_* = \sup_{\Vert f \Vert_A \leq 1} |\psi(f)|.
\end{equation*}
For $x \in K$ let $\Phi_x \in A(K)^*$ be given for $f \in A(K)$ as
\begin{equation*}
\Phi_x (f) = f(x).
\end{equation*}
The map $\Phi: K \to A(K)^*$ is called evaluation map and it is affine. It is easy to see that $\Phi[K] = \{ \Phi_x : x \in K \}$ contains only positive functionals with unit norm such that $\Phi_x(1) = 1$ for every $x \in K$. The converse is also true:
\begin{prop}
$\Phi[K] = \{ \psi \in A(B)^* : \Vert \psi \Vert_* = \psi(1) = 1 \}$.
\end{prop}
\begin{proof}
For proof see \cite[Theorem 4.3]{AsimowEllis}. Also note that $\Vert \psi \Vert_* = \psi(1) = 1$ implies $\psi \geq 0$.
\end{proof}
The set $\Phi[K]$ is sometimes referred to as the state space as in general applications it is often easier to work with $\Phi[K]$ rather than $K$.
\subsection{Exposed faces and maximal faces of a convex set} \label{subsec:prelim-faces}
In this subsection we will define faces, exposed faces and maximal faces and prove Prop. \ref{prop:introd-faces-pointInMaximal}.
\begin{defin}
Let $C \subset K$ be a convex set (that is $C$ is a convex set that is subset of $K$). We say that $C$ is a face of $K$ if $x \in C$, $\lambda \in \mathbb{R}$, $0 < \lambda < 1$ and $x = \lambda y + (1-\lambda) z$ implies $y, z \in C$.
\end{defin}
It is straightforward that $K$ and the empty set are a faces of $K$ and they are called the trivial faces. Apart from the trivial faces it is known that all faces lie in $\partial K$ \cite[Corollary 18.1.3]{Rockafellar-convex}. Face consisting of only single point is called extreme point of $K$.
\begin{defin}
Let $C \subset K$ be a set where some affine function $f$ reaches its maximum (or minimum) over $K$, i.e. if $max_{x \in K} f(x) = M_f$, then $C = \{ x \in K : f(x) = M_f \}$. Such $C$ is called exposed face of $K$.
\end{defin}
Every exposed face is a face \cite[pp. 162]{Rockafellar-convex}. An exposed face consisting of only single point will be called exposed point. It will be important that the set of exposed points of $K$ is dense in the set of extreme points of $K$ \cite[Theorem 18.6]{Rockafellar-convex} and that every face of a closed convex set is closed \cite[Corollary 18.1.1]{Rockafellar-convex}. Also note that not every extreme point must be an exposed point, example of this is presented in \cite[pp. 163]{Rockafellar-convex}.
We proceed by defining the notion of maximal face. Maximal faces are generalization of the $n-1$ dimensional exposed faces of polytopes (that is of convex sets that are convex hull of finite number of points).
\begin{defin} \label{def:introd-faces-maximal}
Let $C \subset K$ be a nontrivial face, such that for every $x \in K \setminus C$ we have $\conv(C \cup \{x\}) \cap \mathit{int}(K) \neq \emptyset$, then we say that $C$ is a maximal face.
\end{defin}
Note that we require maximal faces to be nontrivial, i.e. $K$ itself is not a maximal face. One can show that every maximal face is exposed, because every maximal face is an intersection of $K$ and a hyper-plane tangent to $K$. Also every intersection of $K$ and hyper-plane tangent to $K$ is a maximal faces. We present a simple example of maximal faces of triangle and circle.
\begin{exm}
Assume that $K \subset \mathbb{R}^2$ is a triangle. The vertices of the triangle are extreme and exposed points of $K$, but they are not maximal faces. In this case maximal faces are the edges of the triangle.
Now consider that $K \subset \mathbb{R}^2$ is the convex hull of the unit circle, then every extreme point of $K$ is a maximal face.
\end{exm}
Maximal faces will play a role in the notion of compatibility of measurements as the condition $\conv(C \cup \{x\}) \cap \mathit{int}(K) \neq \emptyset$ will be of great importance.
\begin{prop} \label{prop:introd-faces-pointInMaximal}
Let $K \subset \mathbb{R}^n$ be a non-empty convex compact set. Then for every point $x \in \partial K$ there are maximal faces $C_1, C_2$ such that $x \in C_1$ and $x \notin C_2$.
\end{prop}
\begin{proof}
We will prove the statement in two steps. As first we will prove that that every point of $\partial K$ belongs to some maximal face. Then we prove that maximal faces that have a point in common can not form $\partial K$.
Let $x \in \partial K$, then there exists a non-constant affine function $f$ that reaches its maximum over $K$ in $x$ \cite[Colloraly 11.6.2]{Rockafellar-convex}, let $f(x) = M_f$. The set $G_0 = \{ x' \in K : f(x') = M_f \}$ is an exposed face. If $G_0$ is maximal face then we are done, if $G_0$ is not a maximal face, then there must exist a point $y \in K \setminus G_0$ such that $\conv(G_0 \cup \{y\}) \cap \mathit{int}(K) = \emptyset$. The set $\conv(G_0 \cup \{y\})$ does not have to be face itself, but since $\conv(G_0 \cup \{y\}) \cap \mathit{int}(K) = \emptyset$ then there exists a non-trivial supporting hyperplane to $K$ containing $\conv(G_0 \cup \{y\})$, see \cite[Theorem 11.6]{Rockafellar-convex} for definition of supporting hyperplane to $K$ and proof of the statement. In other words there must exist a non-constant affine function $f_1$ such that $\max_{y \in K} f_1(y) = M_{f_1}$ and $G_1 = \{ x' \in K : f_1(x') = M_{f_1}\} \supset \conv(G_0 \cup \{y\})$, i.e. $G_1$ is an exposed face of $K$ and $x \in G_1$. Moreover for the dimensions of $\aff(G_0)$ and $\aff(G_1)$ we must have $\dim(\aff(G_1)) > \dim(\aff(G_0))$, because $y \in G_1$ and $y \notin G_0$. If $G_1$ is a maximal face then we are finished, if not then we can repeat the procedure to find exposed face $G_2 \supset G_1$.
Since the affine span of every maximal face can be at most $n-1$ dimensional and the dimension of affine span of the exposed faces $G_i$ is strictly growing with $i$ it is clear that we can repeat this procedure at most $n-1$ times to obtain a maximal face, hence in this way to every $x \in \partial K$ we can find a maximal face that contains it.
Now we will proceed with the second part of the proof. Take $x \in \partial K$, denote $\{ C_i \}$ the set of all maximal faces of $K$ and assume $x \in \cap_{i} C_i$. Since every point of $\partial K$ belongs to some maximal face we must have $\cup_i C_i = \partial K$. Let us define positive affine functions $f_i$, such that $C_i = \{ y \in K : f_i (y) = 0 \}$ then since a finite dimensional convex compact set is an intersection of closed half-spaces tangent to it \cite[Theorem 18.8]{Rockafellar-convex} we have $K = \{ y \in \mathbb{R}^n : f_i (y) \geq 0, \; \forall i \}$. Since we have $f_i(x) = 0$, $\forall i$ then for any $\lambda \in \mathbb{R}$, $\lambda \geq 0$ and $z \in K$ we have
\begin{equation*}
f_i ( \lambda z + (1-\lambda) x) = \lambda f_i(z) \geq 0
\end{equation*}
for every $i$. This implies that $\lambda z + (1-\lambda) x \in K$ which is in contradiction with $K$ being compact.
\end{proof}
\section{Measurements on $K$} \label{sec:meas}
Let $E$ be a finite dimensional real vector space equipped with the Euclidean topology and let $K \subset E$ be a compact convex set. We will call $K$ a state space as it represents a set of all possible states of some system and the convex combination is interpreted as probabilistic mixture. Let $\Omega$ be a nonempty compact Hausdorff space and let $\mathcal{P}(\Omega)$ denote the set of Borel probability measures on $\Omega$.
\begin{defin}
Measurements (also called observables) on $K$ with sample space $\Omega$ are affine mappings $m: K \to \mathcal{P}(\Omega)$.
\end{defin}
The interpretation is that $\Omega$ represents all possible outcomes of a certain measurement and is usually referred to as sample space. For $x \in K$ the measure $m(x) \in \mathcal{P} (\Omega)$ is a generalized notion of assigning probabilities to the measurement outcomes. Our definition follows the usual definitions of measurements in probabilistic theories \cite{BuschHeinosaariSchultzStevens-compatibility, Banik-steering} but may be easily generalized to locally compact sample spaces $\Omega$. Let $\sigma \subset \Omega$ be a measurable set, then by $m(x; \sigma)$ we will denote the measure of the set $\sigma$ with respect to the measure $m(x)$.
\subsection{Finite outcome measurements}
Let the sample space $\Omega = \{ \omega_1, \ldots, \omega_k \}$ be a finite set. Every Borel probability measure $\mu \in \mathcal{P} (\Omega)$ is of the form
\begin{equation*}
\mu = \sum_{i=1}^k \lambda_i \delta_{\omega_i}
\end{equation*}
where $\delta_{\omega_i}$ is the Dirac measure centered at $\omega_i$ and $\lambda_i \in \mathbb{R}$, $0 \leq \lambda_i \leq 1$, $\sum_{i=1}^k \lambda_i = 1$. It follows that if $m$ is a measurement on $K$ with finite sample space $\Omega$ then there always are functions $f_j \in A(K)^+$, $0 \leq f_j \leq 1$ for $j \in \{1, \ldots, k\}$, $\sum_{j=1}^k f_j = 1$ such that
\begin{equation*}
m = \sum_{i=1}^k f_i \delta_{\omega_i}.
\end{equation*}
\begin{rem}
In the standard literature \cite{HeinosaariZiman-MLQT, Holevo-QT} usually it is instead of writing $m = \sum_{i=1}^k f_i \delta_{\omega_i}$ simply said that the function $f_j$ represents the probability of the outcome $\omega_j$. To simplify the notation we will use the formulation presented above.
\end{rem}
\section{Compatibility of measurements} \label{sec:inc}
Assume that we wish to perform two distinct measurements $m_1, m_2$ with two separate sample spaces $\Omega_1, \Omega_2$. We would like to know whether there exists a measurement that performs both $m_1$ and $m_2$ at the same time. To ask this question properly we will introduce the concept of marginal measurement. When working with the Cartesian product $\Omega_1 \times \Omega_2$ we will always consider the product topology on it given by the topologies of $\Omega_1, \Omega_2$.
\begin{defin}
Let $m : K \to \mathcal{P} (\Omega_1 \times \Omega_2)$ be a measurement on $K$ with sample space $\Omega_1 \times \Omega_2$. We say that $m_1 : K \to \mathcal{P}(\Omega_1)$ is a marginal measurement of $m$ if for every measurable set $\sigma \subset \Omega_1$ and $x \in K$ we have
\begin{equation*}
m_1(x; \sigma) = m(x; \sigma \times \Omega_2).
\end{equation*}
\end{defin}
This definition can be formally understood as
\begin{equation*}
m_1(x; \sigma) = \int_{\Omega_2} m(x; \sigma \times d \omega_2)
\end{equation*}
for every measurable set $\sigma \subset \Omega_1$. For the finite outcome measurements the integral is replaced by a sum over the outcomes, i.e. for $m = \sum_{i,j=1}^k f_{ij} \delta_{(\omega_i, \omega_j)}$, where $f_{ij} \in A(K)^+$ and $\delta_{(\omega_i, \omega_j)}$ is the Dirac measure centered at $(\omega_i, \omega_j)$, we have
\begin{equation*}
m_1 = \sum_{i,j=1}^k f_{ij} \delta_{\omega_i}.
\end{equation*}
It is straightforward to see that $m_1$ is a measurement on $K$ with sample space $\Omega_1$ as the positivity and normalization to 1 follow from the properties of $m$. Now we are ready for the definition of compatibility.
\begin{defin}
We will say that measurements $m_1 : K \to \mathcal{P}(\Omega_1)$ and $m_2 : K \to \mathcal{P}(\Omega_2)$ are compatible if there exists a measurement $m: K \to \mathcal{P}(\Omega_1 \times \Omega_2)$ such that $m_1, m_2$ are marginal measurements of $m$.
\end{defin}
This definition is the standard definition used for compatibility of measurements.
A natural question is: are there any incompatible measurements? It is of course long known that incompatible measurements in quantum mechanics exist, but mathematically it is interesting to ask what properties of $K$ imply that all measurements are compatible.
\begin{prop} \label{prop:inc-simplex}
Let $K$ be a simplex, that is let $\{x_1, \ldots, x_n\}$ be the set of extreme points of $K$ such that the points $x_1, \ldots, x_n$ are affinely independent. Then every measurement on $K$ is compatible with every other measurement on $K$.
\end{prop}
\begin{proof}
Let $K$ be a simplex then there exists affine functions $b_j: K \to \mathbb{R}$, $j \in \{1, \ldots, n\}$ defined by $b_j(x_i) = \delta_{ij}$. These functions are positive, because for every $y \in K$ we have $y = \sum_{i=1}^n \lambda_i x_i$ with $\sum_{i=1}^n \lambda_i = 1$ and $0 \leq \lambda_i \leq 1$ for every $i$.
Let $m$ be a measurement on $K$ with a sample space $\Omega$, then for $y \in K$, $y = \sum_{i=1}^n \lambda_i x_i$ we have
\begin{equation*}
m(y) = \sum_{i=1}^n \lambda_i m(x_i) = \sum_{i=1}^n b_i(y) m(x_i),
\end{equation*}
i.e. a measurement $m$ on simplex is uniquely described by the measures $m(x_i) \in \mathcal{P}(\Omega)$.
Now let $m_1, m_2$ be measurements on $K$ with the sample spaces $\Omega_1, \Omega_2$ respectively, then for $y \in K$ we have as above
\begin{equation*}
m_j(y) = \sum_{i=1}^n b_i(y) m_j(x_i),
\end{equation*}
for $j \in \{1, 2\}$. Let $(m_1 \times m_2)(x_i)$ denote the product measure obtained form the measures $m_1(x_i)$ and $m_2(x_i)$, that is for measurable sets $\sigma_i \subset \Omega_i$, $i \in \{1, 2\}$ we have
\begin{equation*}
(m_1 \times m_2)(x_i; \sigma_1 \times \sigma_2) = m_1(x_i; \sigma_1) m_1(x_i; \sigma_2).
\end{equation*}
Let the measurement $m: K \to \mathcal{P}(\Omega_1 \times \Omega_2)$ be given as
\begin{equation*}
m(y) = \sum_{i=1}^n b_i(y) (m_1 \times m_2)(x_i)
\end{equation*}
then it is easy to verify that $m_1$ and $m_2$ are marginal measurements of $m$.
\end{proof}
Note that positivity of functions $b_j$ plays a crucial role in the proof and these functions are positive only if $K$ is a simplex. Next we introduce the concept of a coin-toss (also called trivial) measurement.
\begin{defin}
Let $\mu$ be some fixed Borel probability measure on sample space $\Omega$, then by coin-toss we will refer to the measurement given as
\begin{equation*}
m(y) = \mu
\end{equation*}
for every $y \in K$.
\end{defin}
Coin-toss measurements usually represent noise, that is some random factor that affects the measurement outcomes. It can be also interpreted as the most simple measurement when we ignore any information about the state and simply "toss a coin" and return whatever value we obtain. It is straightforward that any coin-toss measurement is compatible with any other measurement.
In the following we state the usual definition of the degree of compatibility.
\begin{defin}
Let $i \in \{1, 2\}$ and let $m_i :K \to \mathcal{P}(\Omega_i)$ be a measurement on $K$ with sample space $\Omega_i$. Let $\tau_i : K \to \mathcal{P}(\Omega_i)$ be some coin-toss measurements, then we define degree of compatibility of measurements $m_1, m_2$ as
\begin{align*}
\degcom (m_1, m_2) =& \sup_{\substack{0 \leq \lambda \leq 1 \\ \tau_1, \tau_2}} \{ \lambda : \lambda m_1 + (1-\lambda) \tau_1, \\ & \lambda m_2 + (1-\lambda) \tau_2 \; \text{are compatible} \}.
\end{align*}
\end{defin}
The reason for considering different trivial measurements $\tau_1, \tau_2$ is that the sample spaces may be different and even if they would be the same due to our definitions we can not pick some preferred measure as for example properly normed Lebesgue measure on a compact subset of $\mathbb{R}^k$. Note that the supremum is taken also over the coin-toss measurements $\tau_1, \tau_2$.
Based on the analysis of compatibility presented in \cite{HeinosaariMiyaderaZiman-compatibility} we obtain the following:
\begin{prop} \label{prop:inc-minimal-deg}
For any two measurements $m_i :K \to \mathcal{P}(\Omega_i)$, $i \in \{1, 2\}$, we have $\degcom(m_1, m_2) \geq \frac{1}{2}$.
\end{prop}
\begin{proof}
The idea is that we can always toss a fair two sided coin, based on the result implement one of the measurements and substitute the other by the respective coin-toss observable. In other words let $\mu_1, \mu_2$ be any Borel probability measures on $\Omega_1, \Omega_2$ respectively that give rise to coin-toss measurements $\tau_i$ given as $\tau_i(y) = \mu_i$, $i \in \{1, 2\}$. Consider the measurement $m: K \to \mathcal{P}(\Omega_1 \times \Omega_2)$ given for $y \in K$ as
\begin{equation*}
m(y) = \dfrac{1}{2} \left( \mu_1 \times m_2(y) + m_1(y) \times \mu_2 \right).
\end{equation*}
It is straightforward to verify that the measurements $\frac{1}{2}(m_1 + \tau_1)$ and $\frac{1}{2}(m_2 + \tau_2)$ are marginal measurements of $m$.
\end{proof}
Similar result has been observed even for compatibility of quantum channels \cite{HeinosaariMiyadera-compOfChan}.
\subsection{Compatibility of two-outcome measurements} \label{subsec:inc-two-outcome}
In general it may be hard to decide whether measurements $m_1$ and $m_2$ are compatible but in the case of two-outcome measurements, that is in the case when $\Omega_1, \Omega_2$ contain only two points, we will formulate necessary and sufficient conditions for the measurements $m_1, m_2$ to be compatible. These conditions may be generalized in the same manner to general finite outcome measurements.
Let $\Omega_1 = \Omega_2 = \Omega = \{ \omega_1, \omega_2\}$ be the sample space of the measurements $m_1, m_2$, then they are of the form
\begin{align*}
m_i &= f_i \delta_{\omega_1} + (1-f_i)\delta_{\omega_2}
\end{align*}
for $i \in \{1, 2\}$. Also every measurement $m$ on $K$ with sample space $\Omega \times \Omega$ is of the form
\begin{equation*}
m = g_{11} \delta_{(\omega_1, \omega_1)} + g_{12} \delta_{(\omega_1, \omega_2)} + g_{21} \delta_{(\omega_2, \omega_1)} + g_{22} \delta_{(\omega_2, \omega_2)},
\end{equation*}
where $g_{11}, g_{12}, g_{21}, g_{22} \in A(K)^+$ and $\delta_{(\omega_j, \omega_k)}$ is a Dirac measure on $\Omega \times \Omega$ centered at $(\omega_j, \omega_k) \in \Omega \times \Omega$. Assume that $m_1$ and $m_2$ are marginal measurements of $m$, then we obtain
\begin{align}
g_{11} + g_{12} &= f_1, \label{eq:inc-two-outcome-sigmas-1} \\
g_{21} + g_{22} &= 1-f_1, \label{eq:inc-two-outcome-sigmas-2} \\
g_{11} + g_{21} &= f_2, \label{eq:inc-two-outcome-sigmas-3} \\
g_{12} + g_{22} &= 1-f_2. \label{eq:inc-two-outcome-sigmas-4}
\end{align}
These equations imply $g_{11} + g_{12} + g_{21} + g_{22} = 1$, but not $g_{jk} \geq 0$, $j,k \in \{1,2\}$ and they in general don't have a unique solution. Let $g_{11} = p$, $0 \leq p \leq 1$, then a general solution to Eq. \eqref{eq:inc-two-outcome-sigmas-1} - \eqref{eq:inc-two-outcome-sigmas-4} is
\begin{align*}
g_{12} &= f_1 - p, \\
g_{21} &= f_2 - p, \\
g_{22} &= 1 - f_1 - f_2 + p,
\end{align*}
which imply the inequalities
\begin{align}
f_1 &\geq p, \label{eq:inc-two-outcome-ineq-1} \\
f_2 &\geq p, \label{eq:inc-two-outcome-ineq-2} \\
1 + p &\geq f_1 + f_2, \label{eq:inc-two-outcome-ineq-3}
\end{align}
that come from $g_{jk} \geq 0$ for all $j,k \in \{1,2\}$. In general there may not exist such $p$ satisfying Ineq. \eqref{eq:inc-two-outcome-ineq-1} - \eqref{eq:inc-two-outcome-ineq-3}, in that case the measurements are incompatible. But if $m$ is a joint measurement of $m_1, m_2$ then the Ineq. \eqref{eq:inc-two-outcome-ineq-1} - \eqref{eq:inc-two-outcome-ineq-3} must be satisfied and Eq. \eqref{eq:inc-two-outcome-sigmas-1} - \eqref{eq:inc-two-outcome-sigmas-4} are satisfied simply because $m_1$ and $m_2$ are marginals of $m$. We have proved the following:
\begin{prop} \label{prop:inc-two-outcome-iff}
Let $m_1, m_2$ be two-outcome measurements on $K$ given as
\begin{align*}
m_i &= f_i \delta_{\omega_1} + (1-f_i)\delta_{\omega_2}
\end{align*}
for $i \in \{1, 2\}$, then they are compatible if and only if there is a function $p \in A(K)^+$, such that $0 \leq p \leq 1$ and Ineq. \eqref{eq:inc-two-outcome-ineq-1} - \eqref{eq:inc-two-outcome-ineq-3} are satisfied.
\end{prop}
Similar results in terms of operators in case of measurements on states were obtained in \cite{HeinosaariKiukasReitzner-robustness, WolfPerezgarciaFernandez-measIncomp}.
Now we will proceed by deriving some conditions on the incompatibility of two-outcome measurements based on the results of Prop. \ref{prop:inc-two-outcome-iff} that will help us prove that there exist incompatible measurements if and only if $K$ is not a simplex.
The main idea is that we will construct two functions $f_1, f_2 \in A(K)^+$ that reach both $0$ and $1$ on $K$ and for the exposed faces
\begin{equation*}
F_i = \{ x \in K : f_i(x) = 0 \},
\end{equation*}
$i \in \{1, 2\}$, it holds that $\conv( F_1 \cup F_2 ) \cap \mathit{int}(K) \neq \emptyset$. Then by the Ineq. \eqref{eq:inc-two-outcome-ineq-1} and \eqref{eq:inc-two-outcome-ineq-2} we have that $p(x) = 0$ for every $x \in \conv(F_1 \cup F_2)$. Since $\conv( F_1 \cup F_2 ) \cap \mathit{int}(K) \neq \emptyset$ and $p \geq 0$ we get $p = 0$. Then by Ineq. \eqref{eq:inc-two-outcome-ineq-3} we must have $f_1 + f_2 \leq 1$ if the measurements are compatible so we will show that we can construct functions $f_1, f_2$ with the mentioned properties such that $f_1(y) + f_2(y) > 1$ for some $y \in K$ whenever $K$ is not a simplex.
The ideas presented above were inspired by an example of incompatible measurements on a square presented in \cite{BuschHeinosaariSchultzStevens-compatibility}.
\begin{prop} \label{prop:inc-two-outcome-maximal-face}
Let $x \in K$ be an extreme point and let $F$ be a maximal face disjoint from $\{ x \}$, then there exist incompatible two-outcome measurements on $K$ if $F$ does not contain all other extreme points of $K$ except for $x$.
\end{prop}
\begin{proof}
For the definition of maximal face see Def. \ref{def:introd-faces-maximal} and remember that according to the definition $K$ itself is not a maximal face. Note that closedness of $K$ will play a role as it implies closedness of every face of $K$ \cite[Corollary 18.1.1]{Rockafellar-convex}.
Assume that there is one maximal face $F$ disjoint from $x$, but $F$ does not contain all extreme points of $K$ except for $x$, i.e. there is an extreme point $y \in K$, such that $y \notin F$ and $y \neq x$. Since $F$, $\{x\}$, $\{y\}$ are closed sets and $\{y\}$ is disjoint from both $F$ and $\{x\}$, then there exists some open neighborhood $N_\varepsilon$ containing $y$, such that $x \notin N_\varepsilon$ and $F \cap N_\varepsilon = \emptyset$. There is an exposed point $z \in N_\varepsilon$ as the set of exposed points is dense in the set of extreme points of $K$ \cite[Theorem 18.6]{Rockafellar-convex}. For the same reason we will consider $x$ an exposed point as well. Now let us construct positive affine function $f_1, f_x, f_z$ such that
\begin{align*}
F_1 &= \{ w \in K : f_1(w) = 0 \}, \\
\{x\} &= \{ w \in K : f_x(w) = 0 \}, \\
\{z\} &= \{ w \in K : f_z(w) = 0 \}, \\
\end{align*}
and
\begin{equation*}
\max_{w \in K} f_1(w) = \max_{w \in K} f_x(w) = \max_{w \in K} f_z(w) = 1.
\end{equation*}
The functions $f_1, f_x, f_z$ give rise to two-outcome measurements $m_1, m_x, m_z$ given as
\begin{align*}
m_1 &= f_1 \delta_{\omega_1} + (1 - f_1) \delta_{\omega_2}, \\
m_x &= f_x \delta_{\omega_1} + (1 - f_x) \delta_{\omega_2}, \\
m_z &= f_z \delta_{\omega_1} + (1 - f_z) \delta_{\omega_2}.
\end{align*}
Since we have
\begin{align*}
\conv( F_1 \cup \{x\}) \cap \mathit{int}(K) &\neq \emptyset, \\
\conv( F_1 \cup \{z\}) \cap \mathit{int}(K) &\neq \emptyset,
\end{align*}
we must have by Prop. \ref{prop:inc-two-outcome-iff}
\begin{align*}
&f_1 + f_x \leq 1,
&f_1 + f_z \leq 1,
\end{align*}
for the measurements $m_1, m_x$ and $m_1, m_z$ to be compatible. From $f_1 + f_x \leq 1$ we get $\{ w \in K : f_1(w) = 1 \} = \{x\}$ and from $f_1 + f_z \leq 1$ we get $\{ w \in K : f_1(w) = 1 \} = \{z\}$, which is a contradiction with $x \neq z$ implied by $x \notin N_\varepsilon$ and $z \in N_\varepsilon$.
\end{proof}
\begin{prop} \label{prop:inc-two-outcome-simplex}
Let $K \subset \mathbb{R}^n$ be a compact convex set then there exist incompatible measurements on $K$ whenever $K$ is not a simplex.
\end{prop}
\begin{proof}
We will rely on the results of Prop. \ref{prop:inc-two-outcome-maximal-face}. Assume that $x \in K$ is an extreme point that is affinely dependent on other extreme points, i.e. there are extreme points $\{y_1, \ldots, y_n\} \subset K$ such that $x = \sum_{i=1}^n \alpha_i y_i$ with $\sum_{i=1}^n \alpha_i = 1$ and let $F$ denote the maximal face disjoint from $\{x\}$. Now let us construct a non-constant positive affine function $f \in A(K)^+$ such that
\begin{equation*}
F = \{ z \in K : f(z) = 0 \}.
\end{equation*}
Again the function $f$ exists as $F$ is an exposed face. Since $x = \sum_{i=1}^n \alpha_i y_i$, $\{y_1, \ldots, y_n\} \in F$ and $f$ is affine, we have
\begin{equation*}
f(x) = \sum_{i=1}^n \alpha_i f(y_i) = 0
\end{equation*}
and we must have $x \in F$, which is a contradiction. Hence the set of exposed points must be affinely independent, finite and $K$ must be a simplex.
\end{proof}
It is an open question whether it can be in an easier fashion showed that the compactness and convexity of $K$ together with compatibility of every two-outcome measurement implies the Riesz decomposition property \cite[pp. 84]{Alfsen-convSets} as it is known that it is equivalent to $K$ being a simplex \cite[Corollary II.3.11]{Alfsen-convSets}. It is also known that in more general settings of effect algebras the result does not hold, i.e. there are effect algebras that are compatible but that do no satisfy Riesz Decomposition property, see \cite[Example 3.6]{Jenca-effAlg} for an example.
\subsection{Linear programming problem for compatibility of two-outcome measurements} \label{subsec:inc-linProg}
We will formulate the problem of compatibility of two two-outcome measurements as a problem of linear programming \cite{Barvinok-linProg} similar to the one obtained in \cite{WolfPerezgarciaFernandez-measIncomp}. We will start with the results of Prop. \ref{prop:inc-two-outcome-iff} and we will construct the linear programming problem from there.
Let $m_1, m_2$ be two-outcome measurements with sample space $\Omega = \{ \omega_1, \omega_2 \}$ given as
\begin{align*}
m_i &= f_i \delta_{\omega_1} + (1 - f_i) \delta_{\omega_2}
\end{align*}
for $i \in \{1, 2\}$ and let $\tau$ represent a coin-toss measurement given as
\begin{equation*}
\tau = \dfrac{1}{2} \left( \delta_{\omega_1} + \delta_{\omega_2} \right)
\end{equation*}
In the following calculations we will restrict ourselves only to this special coin-toss observable as it is sufficent to determine whether the measurements $m_1, m_2$ are compatible.
We want to know what is the highest possible $\lambda \in [\frac{1}{2}, 1]$, such that the measurements $\lambda m_1 + (1-\lambda) \tau, \lambda m_2 + (1-\lambda) \tau$ are compatible. In terms of Prop. \ref{prop:inc-two-outcome-iff} we want to know what is the highest value of $\lambda$ such that there exists $\tilde{p} \in A(K)^+$ such that the conditions
\begin{align*}
\lambda f_1 + \dfrac{1-\lambda}{2} &\geq \tilde{p}, \\
\lambda f_2 + \dfrac{1-\lambda}{2} &\geq \tilde{p}, \\
1 + \tilde{p} &\geq \lambda ( f_1 + f_2) + (1-\lambda)
\end{align*}
are satisfied. Denoting $p = \frac{\tilde{p}}{\lambda}$ and $\mu = \frac{1-\lambda}{\lambda}$ we obtain
\begin{align}
\dfrac{\mu}{2} - p &\geq - f_1, \label{eq:inc-linProg-ineq1} \\
\dfrac{\mu}{2} - p &\geq - f_2, \label{eq:inc-linProg-ineq2} \\
p &\geq f_1 + f_2 - 1. \label{eq:inc-linProg-ineq3}
\end{align}
Now it is important to realize that maximizing $\lambda$ is equivalent to minimizing $\mu$. In the following we will introduce new partially ordered vector spaces and a linear map as the problem of linear programming will be formulated in their terms.
Let $x \in \mathbb{R} \times A(K)$, then $x = ( \alpha, g)$ for $\alpha \in \mathbb{R}$ and $g \in A(K)$. We introduce partial ordering on $\mathbb{R} \times A(K)$ by the relation
\begin{equation*}
( \alpha, g) = x \geq 0 \quad \Leftrightarrow \quad \alpha \geq 0, \, g \in A(K)^+.
\end{equation*}
The topological dual to $\mathbb{R} \times A(K)$ is $\mathbb{R} \times A(K)^*$, for $x = ( \alpha, g)$, $\tilde{c} \in \mathbb{R} \times A(K)^*$, $\tilde{c} = (\beta, \psi)$, $\beta \in \mathbb{R}$, $\psi \in A(K)^*$ we have
\begin{equation*}
\< \tilde{c}, x \> = \alpha \beta + \psi(g).
\end{equation*}
We will also use $A(K) \times A(K) \times A(K)$ equipped with the following partial order: let $(g_1, g_2, g_3) \in A(K) \times A(K) \times A(K)$, then $(g_1, g_2, g_3) \geq 0$ if and only if $g_i \geq 0$ for every $i \in \{ 1, 2, 3 \}$.
Let $T: \mathbb{R} \times A(K) \to A(K) \times A(K) \times A(K)$ be a linear map given as
\begin{equation*}
T (\alpha, g) = ( -g + \dfrac{\alpha}{2}, -g + \dfrac{\alpha}{2}, g),
\end{equation*}
where $\frac{\alpha}{2}$ stands for the constant function attaining the value $\frac{\alpha}{2}$. It is straightforward to see that $T$ is linear.
\begin{prop} \label{prop:inc-linProg-primal}
Let $c \in \mathbb{R} \times A(K)^*$, $c = (1, 0)$, $F \in A(K) \times A(K) \times A(K)$, $F = (-f_1, -f_2, f_1 + f_2 - 1)$ and $x \in \mathbb{R} \times A(K)$, $x = (\mu, p)$ then
\begin{align*}
\inf &\< c , x \> \\
x &\geq 0 \\
Tx &\geq F
\end{align*}
is a primal linear programming problem. When the reached minimum is $0$ then the measurement $m_1, m_2$ are compatible. Moreover, there always exists primal feasible plan.
\end{prop}
\begin{proof}
The proof is straightforward. We have $\<c, x \> = \mu$ for the given $c$, $x \geq 0$ translates to $\mu \geq 0$ and $p \geq 0$. Note that $\mu \geq 0$ corresponds to $\lambda \leq 1$. $Tx \geq F$ is the same as
\begin{equation*}
( -p + \dfrac{\mu}{2}, -p + \dfrac{\mu}{2}, p ) \geq (-f_1, -f_2, f_1 + f_2 - 1)
\end{equation*}
which is in turn equivalent to conditions \eqref{eq:inc-linProg-ineq1} - \eqref{eq:inc-linProg-ineq3}.
Since $\mu = \frac{1-\lambda}{\lambda}$ then $\mu = 0$ implies $\lambda = 1$. There always exists a primal feasible plan as we know that for $\lambda = \frac{1}{2}$ the measurements are always compatible, see Prop. \ref{prop:inc-minimal-deg}.
\end{proof}
Now that we have the primal problem we will find the dual problem to obtain another condition on the compatibility of measurements $m_1, m_2$.
\begin{prop} \label{prop:inc-linProg-dual}
The dual problem to problem introduced in Prop. \ref{prop:inc-linProg-primal} is given as
\begin{align*}
\sup &\< F , l \> \\
T^* l &\leq c \\
l &\geq 0
\end{align*}
where $l \in A(K)^* \times A(K)^* \times A(K)^*$ and $T^*$ is given by the relation $\< \tilde{l}, T \tilde{x} \> = \< T^* \tilde{l}, \tilde{x} \>$ for every $\tilde{l} \in A(K)^* \times A(K)^* \times A(K)^*$ and $\tilde{x} \in \mathbb{R} \times A(K)$, i.e. $T^*: A(K)^* \times A(K)^* \times A(K)^* \to \mathbb{R} \times A(K)^*$, such that for $(\psi_1, \psi_2, \psi_3) \in A(K)^* \times A(K)^* \times A(K)^*$ we have
\begin{equation*}
T^* ( \psi_1, \psi_2, \psi_3) = \left( \dfrac{1}{2}(\psi_1 + \psi_2)(1), - \psi_1 - \psi_2 + \psi_3 \right)
\end{equation*}
where $1$ stands for the constant function on $K$ and $\psi_i(1)$ is the value of functional $\psi_1$ on this function, that is for some $z_{11}, z_{12} \in K$ and $a_1, a_2 \in \mathbb{R}$, $a_1 \geq 0$, $a_2 \geq 0$ we have $\psi_1 = a_1 \phi_{z_1} - a_2 \phi_{z_2}$ and $\psi_1 (1) = a_1 - a_2$.
\end{prop}
\begin{proof}
The only thing we need to do is to find $T^*$, the rest follows from the relation between primal and dual problems \cite[pp. 163]{Barvinok-linProg}.
From the relation $\< \tilde{l}, T \tilde{x} \> = \< T^* \tilde{l}, \tilde{x} \>$ for $\tilde{l} = (\psi_1, \psi_2, \psi_3) \in A(K)^* \times A(K)^* \times A(K)^*$ and $\tilde{x} = (\alpha, g) \in \mathbb{R} \times A(K)$ we get
\begin{align*}
\< \tilde{l}, T \tilde{x} \> &= \left\langle (\psi_1, \psi_2, \psi_3), \left( -g + \dfrac{\alpha}{2}, -g + \dfrac{\alpha}{2}, g \right) \right\rangle \\
&= \dfrac{\alpha}{2} (\psi_1 + \psi_2)(1) + (-\psi_1 - \psi_2 + \psi_3)(g) \\
&= \left\langle \left( \dfrac{1}{2}(\psi_1 + \psi_2)(1), - \psi_1 - \psi_2 + \psi_3 \right), ( \alpha, g) \right\rangle \\
&= \< T^* \tilde{l}, \tilde{x} \>.
\end{align*}
\end{proof}
\begin{prop} \label{prop:inc-linProg-zeroGap}
The duality gap between the primal problem given by Prop. \ref{prop:inc-linProg-primal} and the dual problem given by Prop. \ref{prop:inc-linProg-dual} is zero.
\end{prop}
\begin{proof}
The duality gap is zero if there is a primal feasible plan and the cone
\begin{align*}
Q &= \{ (T \tilde{x}, \<c, \tilde{x} \>) : \tilde{x} \in \mathbb{R} \times A(K), \tilde{x} \geq 0 \}, \\
Q &\subset A(K) \times A(K) \times A(K) \times \mathbb{R},
\end{align*}
where $c = (1, 0)$ as in Prop. \ref{prop:inc-linProg-primal}, is closed \cite[Theorem 7.2]{Barvinok-linProg}. To show that $Q$ is closed we will use the fact that if $V, W$ are topological vector spaces, $Q_V \subset V$ is a cone with compact convex base and $T_V : V \to W$ is a continuous linear transformation, such that $\ker(T_V) \cap Q_V = \{ 0 \}$, then the cone $T_V(Q_V)$ is closed \cite[Lemma 7.3]{Barvinok-linProg}.
Because the cone $A(K)^+$ is generating there exists a base of positive functions $h_1, \ldots h_n$ such that for every $\tilde{h} \in A(K)^+$ we have $\tilde{h} = \sum_{i=1}^n \lambda_i h_i$ for $\lambda_i \geq 0$. We introduce the $L^1$ norm on $A(K)$: for $h' \in A(K)$, $h' = \sum_{i=1}^n \nu_i h_i$ we have $\Vert h' \Vert_{L1} = \sum_{i=1}^n \vert \nu_i \vert$. Note that this norm is an affine function on $A(K)^+$.
We can introduce a norm on $\mathbb{R} \times A(K)$ as follows: let $\tilde{x} = (\alpha, g) \in \mathbb{R} \times A(K)$, then
\begin{equation*}
\Vert \tilde{x} \Vert_{\mathbb{R} \times A(K)} = |\alpha| + \Vert g \Vert_{L1}.
\end{equation*}
The base of the positive cone in $\mathbb{R} \times A(K)$ is the set
\begin{equation*}
\mathcal{K} = \{ \tilde{x} \in \mathbb{R} \times A(K) : \Vert \tilde{x} \Vert_{\mathbb{R} \times A(K)} = 1 \}.
\end{equation*}
$\mathcal{K}$ is compact and convex, because the norm $\Vert \cdot \Vert_{\mathbb{R} \times A(K)}$ is continuous and for $\alpha \geq 0$ and $g \in A(K)^+$ it is affine.
The map $T': \mathbb{R} \times A(K) \to A(K) \times A(K) \times A(K) \times \mathbb{R}$ given as
\begin{equation*}
T' \tilde{x} = ( T\tilde{x}, \<c, \tilde{x} \> )
\end{equation*}
is linear and continuous. If for $(\alpha, g) = \tilde{x} \in \mathbb{R} \times A(K)$ hold that $T' \tilde{x} = 0$ then we have to have $\tilde{x} = (0, 0)$ as $\<c, \tilde{x} \> = 0$ implies $\alpha = 0$ and $T \tilde{x} = (0, 0, 0)$ implies $g = 0$. In conclusion we have $\ker(T') = \{ (0, 0) \}$.
This shows that the cone $Q$ is closed and since we have already showed in Prop. \ref{prop:inc-linProg-primal} that primal feasible plan exists, the duality gap is zero.
\end{proof}
We will proceed with rewriting the dual problem from Prop. \ref{prop:inc-linProg-dual} into a more usable form to obtain necessary and sufficient condition for two two-outcome measurements to be incompatible. We will start from the dual problem stated in Prop. \ref{prop:inc-linProg-dual}. Since $l \in A(K)^* \times A(K)^* \times A(K)^*$ and $l \geq 0$ we must have some $z_1, z_2, z_3 \in K$ and $a_1, a_2, a_3 \in \mathbb{R}$, $a_i \geq 0$, $i \in \{1, 2, 3\}$, such that $l = (a_1 \phi_{z_1}, a_2 \phi_{z_2}, a_3 \phi_{z_3})$ in the formalism of Subsec. \ref{subsec:structure}. From $T^* l \leq c$ we obtain the conditions
\begin{align}
\dfrac{1}{2} ( a_1 + a_2 ) &\leq 1, \label{eq:inc-linProg-dual-ineq1} \\
a_3 \phi_{z_3} &\leq a_1 \phi_{z_1} + a_2 \phi_{z_2}. \label{eq:inc-linProg-dual-ineq2}
\end{align}
Moreover we have
\begin{equation*}
\< F, l \> = -a_1 f_1(z_1) - a_2 f_2(z_2) + a_3 ( f_1 (z_3) + f_2(z_3) - 1 ).
\end{equation*}
Thus we have proved:
\begin{prop}
The two-outcome measurements $m_1, m_2$ corresponding to the functions $f_1, f_2$ are incompatible if and only if there exists positive numbers $a_1, a_2, a_3 \in \mathbb{R}$ and $z_1, z_2, z_3 \in K$ such that Ineq. \eqref{eq:inc-linProg-dual-ineq1}, \eqref{eq:inc-linProg-dual-ineq2} are satisfied and $\< F, l \> > 0$.
\end{prop}
If one would wish to have $\< F, l \> > 0$ it would be a possible first idea to have $f_1(z_1) = f_2(z_2) = 0$ as then only $f_1 (z_3) + f_2(z_3) > 1$ would be required. In this case it would be easy to satisfy the Ineq. \eqref{eq:inc-linProg-dual-ineq2} by suitable choice of $a_3$ whenever $\conv( \{z_1, z_2 \} ) \cap \mathit{int}(K) \neq \emptyset$ as then for some $\nu \in [0,1]$ we would have $\nu z_1 + (1-\nu) z_2 \in \mathit{int}(K)$ and $\phi_{\nu z_1 + (1-\nu) z_2}$ would be an order unit in $A(K)^*$. Matter of fact, this is exactly the idea we used to prove Prop. \ref{prop:inc-two-outcome-simplex}.
It is worth mentioning that by similar methods of semidefinite programming it was shown that in case of measurements on states that the value of $\< F, l \>$ corresponds to maximal violation of CHSH Bell inequality \cite{WolfPerezgarciaFernandez-measIncomp}.
\section{Conclusions}
Incompatibility of measurements is one of the key aspects of quantum theories and as our results have shown, in finite dimensional cases it only differentiates classical probabilistic theories from general probabilistic theories. The quest for finding some essentially quantum restriction on probabilistic theories also considered in \cite{BarnumHowardBarretLeifer-noBroadcast} is not over as such restriction would probably help us understand quantum theories better and deeper.
It is of course an open question whether such aspect of quantum theories that would differentiate it from other non-classical probabilistic theories exists. It is also an open question whether our results hold also in the infinite dimensional case. Possible approach to generalize our results to infinite dimensional case would be to prove it using Riesz decomposition property and to observe whether the proof may be generalized for infinite-dimensional state space.
\section*{Acknowledgments}
The author is thankful to Anna Jen\v{c}ov\'{a} for helpful and stimulating conversations and to M\'{a}rio Ziman, Teiko Heinosaari and Takayuki Miyadera for sharing their research notes. This research was supported by grant VEGA 2/0069/16.
|
\section{INTRODUCTION}
Molecular clouds (MCs), as observed in CO surveys of the
Galactic plane, play an important role in studying star formation
and the structure of the Galaxy \citep[e.g.,][]{1985ApJ...295..422C,1987ApJ...322..706D,
2001ApJ...547..792D,1998ApJ...502..265H,2006ApJS..163..145J,2013PASA...30...44B}.
Several works were concentrated on the large-scale structure traced by
MCs in the first Galactic quadrant
\citep[e.g.,][]{1985ApJ...297..751D,1986ApJS...60..297C,
1988ApJ...327..139C,1986ApJ...305..892D,1987ApJS...63..821S,1988A&A...195...93J,1989ApJ...339..919S,
2010ApJ...723..492R}.
However, the CO emission lying beyond the solar circle in the first quadrant
of the Galaxy is less studied \citep[e.g., Section 3.1.5 and
Figure 3 in][]{2001ApJ...547..792D}.
The only cases are the study of the MCs
in the Outer Arm of the Galaxy for the region of $l>55^{\circ}$
in the first quadrant \citep[using the NRAO 11 m telescope,]
[]{1981ApJ...249L..15K,1988ApJ...330..399M}
and for the region between $l=65^{\circ}$ and $71^{\circ}$
\citep[using the CfA 1.2 m and the Bell 7 m telescopes,][]{1990ApJ...357L..29D}.
The Milky Way Imaging Scroll Painting (MWISP) project
\footnote{http://www.radioast.nsdc.cn/mwisp.php} is
a high resolution ($50''$) \twCO\ ($J$=1--0), \thCO\ ($J$=1--0), and C$^{18}$O ($J$=1--0)
survey of the northern Galactic plane, performed with the
Purple Mountain Observatory Delingha 13.7 m telescope.
The survey started in 2011, and will cover Galactic longitudes from
$l=-10^{\circ}$ to $250^{\circ}$ and latitudes from
$b=-5^{\circ}$ to $5^{\circ}$ over a period of $\sim$10 years.
The Galactic plane will be covered by full-sampling observations
with the spectral line On-The-Fly (OTF) mode. The survey has equal
sensitivity over all regions of the sky mapped and thus it is unbiased.
One of the goals of this project is to study the physical properties of MCs
along the northern Galactic plane.
As of this writing the survey has completed about half of its planned
area of coverage.
Benefiting from the large unbiased
survey of the MWISP, we can systematically investigate the
characteristics of the structure of the Galaxy using the molecular gas.
Recently, a CO spiral arm lying beyond the Outer Arm in the first
Galactic quadrant was identified by \cite{2011ApJ...734L..24D}
using the CfA 1.2 m telescope.
Based on our new CO data of the MWISP, the extension of the above arm,
which is probably the Scutum-Centaurus Arm into the outer second quadrant,
has been revealed very recently by \cite{2015ApJ...798L..27S}.
In this paper, we investigate the Outer Arm
\citep[the Norma-Cygnus Arm, e.g., see the recent review in][]{2014ApJS..215....1V}
in the first quadrant of the Galaxy according to the distribution
of the CO gas between $l=35^{\circ}$ to $45^{\circ}$.
The distant MCs beyond the solar circle traced by CO emission have a negative velocity
in the direction that we are interested in. The velocity range
of the CO gas within the arm is about $-60$ to $-10\km\ps$
in the direction (Section 3.3).
The Outer Arm lies beyond the solar circle in the Galactic longitude
range of $l\gsim30^{\circ}$ in the first quadrant of the Milky Way,
thus the molecular gas traced by CO emission does not suffer the
kinematic distance ambiguity encountered in the region within the Solar circle.
Generally speaking, in this quadrant, the more negative
the value of the velocity is, the larger the distance of the MC.
Moreover, the large sample of the distant MCs provides a good opportunity
to study the global cloud parameters (Section 3.6) since the data
have a high signal-to-noise ratio and there is little overlap
between the clouds (Section 3.1).
We present observations and data reductions of the survey in Section 2.
In Section 3.1 and 3.2, we describe the MCs identification and
the distribution of the distant MCs, respectively.
We find that the physical mid-plane traced by the distant CO
gas is slightly displaced from the IAU-defined plane of $b = 0^{\circ}$
(Section 3.3). We also discuss the relationship between
the two planes in Section 3.4. In Section 3.5, we derive the Sun's
offset above the physical mid-plane because of the apparent displacement
between the two planes for regions near and within the solar circle.
The statistical properties of the distant MCs are
discussed in Section 3.6.
Finally, the summary is given in Section 4.
\section{CO OBSERVATIONS AND DATA REDUCTIONS}
We mapped $\sim$110 square degrees (439 cells) in the
region between Galactic longitudes 34\fdg75--45\fdg25 and
latitudes $-$5\fdg25 to 5\fdg25 during 2011 November to
2015 March using the 13.7 m millimeter-wavelength telescope located
at Delingha in China.
The nine-beam Superconducting Spectroscopic Array Receiver (SSAR)
system was used at the front end and each Fast Fourier transform
spectrometer (FFTS) with a bandwidth of 1 GHz provides 16,384 channels
and a spectral resolution of 61 kHz \citep[see the details in][]{Shan}.
The molecular lines of \twCO\ ($J$=1--0), \thCO\ ($J$=1--0) and
C$^{18}$O ($J$=1--0) were observed simultaneously with the OTF method.
Each cell with dimension 30$'\times30'$
was scanned at least in two orthogonal directions, along the Galactic
longitude and the Galactic latitude, in order to reduce the fluctuation
of noise. The half-power beamwidth (HPBW) of the telescope
was about $50''$ at a local oscillating frequency of 112.6 GHz
and the pointing accuracy was better than $5''$ in all observing epochs.
All cells were reduced using the GILDAS/CLASS
package.\footnote{http://www.iram.fr/IRAMFR/GILDAS}
After removing the bad channels in the spectra and correcting the
first order (linear) baseline fitting, the antenna temperature
($T_{\rm A}^{\star}$) was converted to the main beam temperature
($T_{\rm mb}$) with the relation
$T_{\rm mb}=T_{\rm A}^{\star}/(f_{\rm b}\times\eta_{\rm mb})$.
In the above conversion, the beam-filling factor of $f_{\rm b}$ is
assumed to be one and the beam efficiency $\eta_{\rm mb}$ is from the
status report\footnote{http://www.radioast.nsdc.cn/mwisp.php}
of the 13.7 m telescope of the Purple Mountain Observatory.
The typical sensitivity (rms) is about 0.5 K for \twCO\ ($J$=1--0)
at the channel width of 0.16$\km\ps$ and 0.3 K for \thCO\ ($J$=1--0) and
C$^{18}$O ($J$=1--0) at 0.17$\km\ps$
with a spatial resolution of $<1'$.
Finally, all spectra of CO in each cell were converted to the
three-dimensional (3D) cube mosaic data together with a grid
spacing of $30''$ and the channel width of 0.16$\km\ps$ for
\twCO\ ($J$=1--0) and 0.17$\km\ps$ for \thCO\ ($J$=1--0)
and C$^{18}$O ($J$=1--0) for subsequent analysis.
\section{RESULTS AND DISCUSSION}
\subsection{MCs Identification}
We carefully checked the 3D cube data channel by channel
before investigating
the characters of the MCs that we are interested in. We find
that the distant MCs in the velocity interval of
$-60$ to $-10\km\ps$
are isolated in the $l$--$b$--$v$ space.
That is, most of them show the distinct structure in the space of
$l$--$b$ and there is, therefore, little overlapping in the
velocity axis over the large survey's area. Moreover, it is
not surprising that their \twCO\ ($J$=1--0) emission
is stronger than that of the \thCO\ ($J$=1--0).
No $>3\times$rms C$^{18}$O ($J$=1--0) emission
($T_{\rm peak18}$) is detected in the whole region
in the velocity range.
Only insignificant C$^{18}$O emission ($T_{\rm peak18}\sim2\times$rms)
can be found toward the MCs' region with some detectable
\thCO\ emission (e.g., $T_{\rm peak13}>5\times$rms). We do not discuss these
weak C$^{18}$O emission features in this paper.
According to the discrete feature of the distant MCs,
we used the FINDCLUMPS tool in the CUPID package
(part of the STARLINK package) to identify MCs in the
\twCO\ and \thCO\ FITS cube. The CLUMPFIND algorithm
\citep{1994ApJ...428..693W} is applied in the process of
identification. We set the parameter Tlow=3$\times$rms
for \twCO\ and Tlow=2.7$\times$rms for \thCO,
which determines the lowest level to contour an MC, in
order to obtain as much of the emission as possible
and to avoid the contamination of the noise based on
our large survey data. The parameter of DeltaT=9$\times$rms,
which shows the contour increment, is large since we do not
care about the properties of the structure of the bright
parts embedded in the extended MCs.
Other input parameters of the CLUMPFIND algorithm are
typical to identify those distant MCs for the 3D cube data.
Considering the irregular shape of the MCs, we use
a polygon to describe the MC's boundary in the $l$--$b$ space.
Some objects are rejected if their sizes
are less than the criteria: FwhmBeam=1.5 pixel,
VeloRes=2 channel, and MinPix=16.
It is worthwhile to note that sometimes an individual MC is
decomposed to several MCs because of the small velocity
separation of $\sim1-2\km\ps$ within roughly the same $l$--$b$
region. However, the above case is a rare population compared
to that of all detected MCs (Figures 1 and 2).
On the other hand, some point-like sources ($\lsim$ 3--4 pixels)
cannot be picked out because of their smaller sizes. Some faint
diffuse structures where \twCO\ ($J$=1--0) emission is near
the noise level of the survey cannot be identified
due to their poor signal-to-noise ratio. Further observations
are needed to confirm these MCs with small size or faint emission.
Finally, we identified 575 and 131 MCs from the \twCO\
and \thCO\ emission based on the CLUMPFIND algorithm, respectively.
The parameters of each MC,
such as position, LSR velocity, one-dimensional velocity dispersion,
the peak value, the size, and the luminosity, were obtained
directly from the automated detection routine. The parameters
of the resolved MCs are summarized in Tables 1 and 2:
(1) the ID of the identified MCs, arranged from the low
Galactic longitude;
(2) and (3) the MCs' Galactic coordinates ($l$ and $b$);
(4) the MCs' LSR velocity ($V_{\rm LSR}$);
(5) the MCs' full width at half maximum (FWHM, $\Delta V_{\rm FWHM}$),
defined as 2.355 times of the velocity dispersion
$\sigma_{v}$ for a Gaussian line;
(6) the MCs' peak emission ($T_{\rm peak}$);
(7) the MCs' area;
(8) the MCs' integrated intensity ($W_{\rm CO}$) within a defined
area;
(9) the MCs' total luminosity ($L_{\rm CO}=W_{\rm CO}\times \rm {Area}$).
Finally, we use MWISP Glll.lll$\pm$bb.bbb to name an
MC detected from the survey data.
We made the integrated intensity map of the \twCO\
($J$=1--0) emission in the velocity interval of
$-1.6\times l$+$13.2\km\ps$ to $-10\km\ps$ (Section 3.3)
in Figure 1.
After overlapping the 575 resolved MCs on the map, we find that
the distribution of the CO gas is well traced by the result from
the automated cloud-finding routine. It shows that the CLUMPFIND
algorithm is thus reliable to pick out the distant MCs.
In addition, the sample of the MCs can be used to
investigate the large structures of the spiral arm (Section 3.3)
and the properties of the distant MCs (Section 3.6).
Figure 2 displays the intensity-weighted
velocity (the first moment) map of the \twCO\ ($J$=1--0) emission.
The distribution of the 131 \thCO\ MCs is also consistent with
that of the bright \twCO\ ($J$=1--0) emission.
The blue stripes at $l=$41\fdg5 and $b=-$1\fdg5 are from the
bad channels,
which show regular oscillation along the velocity axis
with intensity of $-2$ to 3 K. The abnormal feature can be
easily distinguished from the CO emission of MCs.
We also present the typical spectra of four resolved MCs
(\twCO\ with the black line and \thCO $\times2$ with blue) in Figure 3.
All spectra in Figure 3 show good baselines.
The CO emission of the MCs can be easily discerned from the
spectra and the peak of \thCO\ ($J$=1--0) emission is roughly
corresponding to that of the \twCO\ ($J$=1--0) line (Figure 3).
\subsection{Spatial Distribution of the Distant Molecular Gas}
We show the 575 MCs in the Galactic coordinate system in Figure 4,
in which the filled circle indicates the
position and the size (scaled with CO luminosity) of MCs and the color
indicates the LSR velocity of MCs.
All identified MCs are in the region of the Galactic latitude
$b=-$1\fdg8 to $b=$2\fdg2, which indicates that the molecular gas is
roughly concentrated in and around the Galactic plane
(mainly in the region of $b\sim -$0\fdg5 to 1\fdg5).
The map also shows that the distribution of the MCs is not
uniform with Galactic longitude.
Figure 5 is the normalized histogram of the MCs with the Galactic
longitude (left panel) and the Galactic latitude (right panel).
According to the left panel of Figure 5, we find that the spatial
distribution
of the MCs is not uniform again. Several distribution peaks of the
MCs are clearly discerned at $l=$35\fdg25, 39$^{\circ}$, 42$^{\circ}$,
and 44$^{\circ}$, respectively. The concentrated groups of these MCs
can also be seen in the same region of Figure 4.
The map in the right panel of Figure 5 shows that the MCs
are mainly located within a limited range of the Galactic latitude,
which indicates that the molecular gas in the velocity interval
is actually from the Galactic plane with a distribution peak
of $b=$0\fdg42.
We also made the position$-$velocity (PV) map of the MCs along
the Galactic longitude (Figure 6),
in which the distribution of the MCs is similar
to that in Figure 5.
The velocity gradient of the MCs can be seen
on a large scale (see green dashed line in Figure 6),
which shows the velocity trend of the Outer Arm
along the Galactic longitude (Section 3.3).
The distribution of the MCs in the longitude$-$velocity
diagram seems to display some asymmetrical ripples,
which is a feature that is probably from the substructures
of the Outer Arm in this direction.
We note that some MCs are concentrated within groups.
These MC groups seem to display partial shell structures,
which probably relate to the star forming activity within
the Outer Arm.
We searched for the MCs associated with the massive star
formation regions based on our unbiased CO survey.
The information of the associations is listed in Table 3.
We find that the distribution of the massive star-forming
regions in the Outer Arm, which are mainly located at
$l = 39^{\circ}$, $42^{\circ}$, and $44^{\circ}$, is associated
with that of the CO peaks in the $l$--$b$ space
(see the 21 \HII\ regions marked with black circles and
five 6.7 GHz masers marked with black triangles in Figure 2).
This result shows that star formation activity is very common
in the distant MCs of the Outer Arm.
We will study the interesting structures and the massive
star-forming activities in the future.
\subsection{Outer Gas Arm Traced by CO emission}
In the direction of $l = 35^{\circ}$ to $45^{\circ}$, we note that
the Outer Arm LSR velocity locates between the Perseus Arm
and the Scutum-Centaurus Arm.
That is, the Perseus Arm, the Outer Arm, and the Scutum-Centaurus
Arm are encountered in this direction, each with different
and decreased LSR velocity
\citep[see Figures 9 and 10 in][]{2016ApJ...823...77R}.
Part of the molecular gas' emission in the velocity interval of
0 to $-10\km\ps$ is from the Perseus Arm.
The molecular gas of the Local Arm
\citep{2013ApJ...769...15X} probably also contributes to some emission
in the velocity range of 0 to $-10\km\ps$. And
the molecular gas with the LSR velocity $< -1.6\times l$+$13.2\km\ps$
is probably from a segment of a spiral arm at
Galactocentric radii of R$_{\rm GC} \sim$13--14 kpc
\citep[the value of -1.6 $\km\ps$Degree$^{-1}$ is from][]{2011ApJ...734L..24D}.
In order to isolate the emission from the three spiral arms,
we use the molecular gas in the velocity cutoff of
$-1.6\times l$+$13.2\km\ps$ to $-10\km\ps$
to trace the Outer Arm (see the yellow dashed lines in Figure 6).
Actually, the large-scale segment of the Outer Arm can be
discerned from the longitude$-$velocity diagram of CO emission, in which
the CO emission of the Outer Arm's MCs is mainly in the
above velocity range and their distribution can be
described as $V_{\rm LSR}=-2.78\times l+75.9\km\ps$
(see the green dashed line in Figure 6) based on
\thCO\ emission (excluding MCs near $l\sim44^{\circ}$ and
$V_{\rm LSR}\sim-20 \km\ps$).
We use a Gaussian function, $f(b)\sim e^{-\frac{(b-b_{0})^{2}}{2\sigma ^{2}}}$,
to fit the distribution of the MCs with the Galactic latitude
(the right panel in Figure 5).
The dashed blue line shows the best fit: $b_{0}$=0\fdg42 and
$\sigma$=0\fdg29. It indicates that the distant MCs in the velocity
interval are mainly distributed within
a narrow range of the Galactic latitudes of
FWHM=2.355$\sigma \sim$0\fdg7.
We thus suggest that the large amount of the CO emission actually
traces the Outer Arm of the Milky Way.
The best Gaussian fit also shows that the distribution peak of
the MCs in the segment is not around $b = 0^{\circ}$ but
$b$=0\fdg42, which probably indicates that
the Outer Arm traced by the molecular gas is slightly displaced
with respect to the IAU-defined plane (see the details in Section 3.4).
We also use a linear relation, $b$=0\fdg0377$\times l -$1\fdg0893 (the
thick blue line in Figure 4), to describe the distribution of the MCs
in the $l$--$b$ space.
The fitted blue line represents the physical mid-plane with respect to
the IAU-defined plane of $b = 0^{\circ}$.
It indicates that the Outer Arm traced by the CO gas passes through
the IAU-defined plane of $b = 0^{\circ}$ at $l = $28\fdg9.
The molecular gas of the Outer Arm in the range of $l\lsim$28\fdg9 is
mostly within the solar circle and the distribution of
the MCs with latitude is mainly $b< 0^{\circ}$
\citep[see the Norma-Cygnus arm in Figure 3 of][]{2014ApJS..215....1V}.
The longitude$-$velocity relation from Figure 6 also indicates
that the LSR velocity
of the Outer Arm is roughly $\gsim0 \km\ps$ at $l\lsim$28\fdg9.
The feature is probably related to the
Sun's offset from the physical mid-plane (see Sections 3.4 and 3.5).
Using the 21 cm \mbox{H\,\textsc{i}}
data from the VLA Galactic plane survey
\citep[VGPS;][]{2006AJ....132.1158S},
\cite{2011ApJ...734L..24D} found that the Outer Arm
traced by the \mbox{H\,\textsc{i}} gas is mainly
below the plane of $b = 0^{\circ}$ in the region
of $l = 25^{\circ}-30^{\circ}$ (see Figure 2 in their paper).
In other words, at $l < 30^{\circ}$ the Outer Arm seems to be
below the IAU-defined plane in the first Galactic quadrant.
According to our linear fit of the CO emission, in the region of
$l = 25^{\circ}-30^{\circ}$ the Galactic latitude distribution of
the Outer Arm should peak at $b \sim-$0\fdg1, which is
roughly consistent with the result of \mbox{H\,\textsc{i}} gas
\citep[see Figure 2 in][]{2011ApJ...734L..24D}.
Based on the Galactic parameters of Model A5 of \cite{2014ApJ...783..130R},
we suggest that the heliocentric distance of the 10$^{\circ}$ segment of the
Outer Arm is about 15.1--15.6 kpc, which is a value that is
quite consistent with the result of other
models \citep[e.g.,][]{2007ApJ...671..427M,2010ASPC..438...16F}.
The Galactocentric radii of the arm, R$_{\rm GC}$, in the direction is
about 9.5--11.4 kpc accordingly.
After subtracting the effect of the tilted structure with the
Galactic longitude (the thick blue line in Figure 4), we derived
that the scale height of the MCs in the Outer Arm is about
0\fdg6 (or $\sim$160 pc at a heliocentric distance of 15 kpc).
It is roughly consistent with the FWHM of the molecular disk of the
Milky Way in such Galactocentric radii
\citep[see Table 1 and Figure 3 in][]{2006PASJ...58..847N}.
The total \twCO\ luminosity of the segment of the spiral arm
is about 2.1$\times10^5$ K$\km\ps$pc$^2$ at a heliocentric distance of 15 kpc.
Therefore, the total molecular gas within the segment is
about 9$\times10^{5}\Msun$ by adopting the mean CO-to-H$_2$ mass
conversion factor $X_{\rm CO}$=$2\E{20}$~cm$^{-2}$K$^{-1}$km$^{-1}$s
\citep{2013ARA&A..51..207B} and a mean molecular weight per
H$_2$ molecule of 2.76.
We mention that the total mass of the segment estimated above
is probably the lower limit because some weak \twCO\
emission is not accounted for in Table 1 (Section 3.1)
and the value of $X_{\rm CO}$ is probably underestimated for the outer
Milky Way \citep{2013ARA&A..51..207B}.
Based on Table 1, the \twCO\ luminosity of the segment
of the spiral arm is about 2.1$\times10^5$ K$\km\ps$pc$^2$ at a
heliocentric distance of 15 kpc.
The molecular gas within the segment is
about 0.9$\times10^{6}\Msun$ by adopting the mean CO-to-H$_2$ mass
conversion factor $X_{\rm CO}$=$2\E{20}$~cm$^{-2}$K$^{-1}$km$^{-1}$s
\citep{2013ARA&A..51..207B} and a mean molecular weight per
H$_2$ molecule of 2.76.
For comparison, the total \twCO\ luminosity of the segment
of $\sim2.6\times10^4$ K$\km\ps$arcmin$^2$ can be obtained by
integrating velocity channels from $-1.6\times l$+$13.2\km\ps$
to $-10\km\ps$ for the covered map. Therefore, our estimate for
the total molecular mass of the segment of the Outer Arm
is 2.2$\times10^{6}\Msun$. The result indicates that
over half of the total mass appears to be missed due to the relatively
high criterion of Tlow=3$\times$rms in the CLUMPFIND method.
The difference between the above estimations probably indicates
that about 60\% of the molecular gas probably resides within small, cold,
and$\//$or faint, diffuse clouds that cannot be included in our catalog
\citep[see also, e.g., Section 5.1.3 of][]{2015ARA&A..53..583H}.
We mention that the total mass of the segment estimated above
is probably the lower limit because some weak \twCO\
emission is not accounted for and the value of $X_{\rm CO}$ is
probably underestimated for the outer
Milky Way \citep{2013ARA&A..51..207B}.
Assuming a constant value of $X_{\rm CO}$=
$2\E{20}$~cm$^{-2}$K$^{-1}$km$^{-1}$s and a heliocentric distance of 15 kpc,
the mass surface density of
the distant MCs is about $22\Msun$~pc$^{-2}$, which is roughly consistent with
the character of MCs in the outer Galaxy
\citep[see, e.g., Figure 8 in][]{2015ARA&A..53..583H}.
\subsection{Physical Mid-plane Traced by the Outer Arm}
In this section, we investigate the relationship between
the physical mid-plane and the IAU-defined plane
of $b = 0^{\circ}$ in the view
of the molecular gas on a large scale based on the new CO survey.
Generally, warps are common in spiral disks of galaxies
\citep[e.g.][]{1998A&A...337....9R,2001A&A...373..402S,2002A&A...382..513R}. For example,
NGC 4013 shows a prodigious warp in its outer parts,
in which the gas layer curves away from the plane of the inner
disk \citep{1996A&A...306..345B}. In the Milky Way, the H\textsc{i}
layer was also found to be warped from the Solar radius outwards
\citep[e.g.,][]{1957AJ.....62...90B,1958MNRAS.118..379O,1982ApJ...263..116H}.
Furthermore, the stellar warp of the Milky Way is similar to the gaseous
warp, but smaller in amplitude
\citep[e.g.,][]{1994ApJ...429L..69F,1998ApJ...492..495F,2001ApJ...556..181D}.
On the other hand, the inner disk of the Milky Way is probably
observed to be
tilting because of the offset of the Sun \citep{1995MNRAS.273..206H}.
That is, the physical mid-plane projected on the sky displays an apparent
change in latitude with respect to longitude across some regions of the sky.
Previously, astronomers often used the star-counts
method to derive the position of the Sun
\citep[e.g.,][]{1995MNRAS.273..206H,1995AJ....110.2183H,2001ApJ...553..184C,
2001AJ....121.2737M,2007MNRAS.378..768J,2009MNRAS.398..263M}.
These studies found
that the Sun's vertical displacement from the physical mid-plane is
$\sim$10--30 pc \citep[refer to Table 1 in][]{2006JRASC.100..146R}.
By studying the infrared dark cloud (IRDC) ``Nessie,''
\cite{2014ApJ...797...53G} recently suggested that the very long
and dense filamentary IRDC represents a spine-like bone of the
major spiral arm, the Scutum-Centaurus Arm, in the fourth quadrant
of the Milky Way. They also suggested that the latitude
of the true Galactic mid-plane traced by the spiral arm is
$b\sim -$0\fdg4 (but not $b=0^{\circ}$)
because of the Sun's height off of the physical mid-plane
of the Milky Way
\citep[see Figure 2 and Section 3.1 in][]{2014ApJ...797...53G}.
Here we discuss the Galactic structure traced by CO emission of the
Outer Arm.
Both of the effects of the warp and the tilted plane (because of the Sun's
offset from the physical mid-plane) were considered.
We explored three typical models for comparison with our observations:
the gaseous warp model from the Galactic H\textsc{i} data
\citep[$m=1$ mode for a cutoff at R$_{\rm GC}$=10 kpc,][]{2006ApJ...643..881L};
the stellar warp model from 2MASS data
\citep[a $\gamma_{\rm warp}$ of 0.09 and a cutoff
at R$_{\rm GC}$=8.4 kpc,][]{2009A&A...495..819R};
and the tilted-plane model for the inner Galaxy
\citep[R$_{\rm GC}<7$ kpc, and see Equation 5 in][]{1995MNRAS.273..206H}.
The $\gamma_{\rm warp}$ in the stellar warp model is defined as
the ratio between the displacement (z$_{\rm mid-plane}$)
of the physical mid-plane from the IAU-defined plane of $b = 0^{\circ}$
and the Galactocentric radii of R$_{\rm GC}$
\citep[refer to Section 2.3 and Figure 9 in][]{2009A&A...495..819R}.
For two warp models, we assume that the Sun roughly lies on the line of
nodes of the warp
\citep[i.e., $\phi_{\rm warp}=0^{\circ}$,][]{1988gera.book..295B,1998gaas.book.....B}.
Based on the best fit of the longitude$-$velocity diagram
($V_{\rm LSR}=-2.78\times l+75.9\km\ps$; Figure 6), we can
kinematically calculate the Galactocentric radii and the height
of the physical mid-plane traced by the Outer Arm using the A5
rotation curve model of \cite{2014ApJ...783..130R}.
Thus, when the R$_{\rm GC}$--z$_{\rm mid-plane}$ relation is obtained,
we can directly compare the slope and the amplitude of the different
models (the dotted lines) with those of the observations (the blue line)
in the $l-b$ map (Figure 4).
Here, the slope is defined as the ratio between
z$_{\rm mid-plane}$ and $\Delta l$.
We find that the slopes of the two
warp models are larger than that of the observations.
The amplitude of the stellar warp model is larger than the
observed displacement across the present longitude interval.
Furthermore, the amplitude of the gaseous warp model is also larger
than the observations in the region of $l\gsim$40\fdg7 or R$_{\rm GC}\gsim10.6$ kpc.
It indicates that the effect of the warp
traced by the CO gas of the Outer Arm is $not$ so large in
this segment with R$_{\rm GC}\sim9.5-11.4$ kpc.
This result is roughly consistent with the interpretation that
the amplitude of the physical mid-plane displacement is small inside a radius of 10 kpc,
and it steeply increases beyond the radius \citep{2006PASJ...58..847N}.
Moreover, \cite{2002A&A...394..883L} also suggested that the effect
of the stellar warp is large at the place of the larger R$_{\rm GC}$,
but is small in the vicinity of the solar circle
(R$_{\rm GC}\sim$ 8--10 kpc, see Figure 18 in their paper).
Our CO survey shows that the slope of the observations is intermediate
between the warp models and the tilted flat plane model
across the region.
Therefore, the warp probably affects the observed slope of the MCs
in such a region, but not to the extent predicted by the current models.
On the contrary, the slope of the tilted-plane model
seems very close to that of the observations
in the region of $l\lsim$37\fdg5 (or R$_{\rm GC}\lsim10$ kpc).
The displacement between the physical mid-plane and the $b=0^{\circ}$ plane
is also comparable to that of the tilted-plane model in the above region,
in which the offset between the two planes is
about 0\fdg2--0\fdg3 or 50--80 pc at a heliocentric distance of 15 kpc (Figure 4).
Furthermore, the amplitude of the tilted-plane model will be closing to
the displacement between the two planes with decreased R$_{\rm GC}$
(or $l$ for the mid-plane traced by the Outer Arm).
Actually, the inner H\textsc{i} disk was found to be tilted against
the $b = 0^{\circ}$ plane \citep{2003PASJ...55..191N}. Also,
the amplitude of the displacement traced by our CO gas seems comparable to
that of \cite{2006PASJ...58..847N} (see Figure 12 in
their paper) in the region with R$_{\rm GC}\sim$9--11 kpc.
Based on the above analysis, we suggest that the Galactic warp plays a role
in the region with large Galactocentric radii (e.g., R$_{\rm GC}\gsim10$ kpc)
while it could be neglected in regions near and within the solar circle.
By contrast, the tilted-plane model
suggests that the Sun's offset from the
physical mid-plane seems to become the dominant for regions
near and within the solar circle.
Roughly speaking, to first order approximation, the observed tilt of
the MCs is due to the Sun's offset from the physical mid-plane
for regions with R$_{\rm GC}\lsim10$ kpc.
\subsection{Position of the Sun}
In Section 3.4, we show that the Galactic warp is not strong
and the apparent displacement between the two planes is probably dominated by
the tilted-plane effect in the region of R$_{\rm GC}\lsim10$ kpc.
We thus do not take into account the effect of the gas warp in the region
near and within the solar circle.
Indeed, the Galactic warp is very important for the distant gas arm with large
R$_{\rm GC}$, \citep[e.g., see that the structure of the
extension of the Scutum-Centaurus Arm with R$_{\rm GC} >14$ kpc
exhibits warps along the Galactic longitude, Figure 2b in]
[]{2015ApJ...798L..27S}.
The warp of the Outer arm is also obvious between
$l=100^{\circ}$ and $150^{\circ}$ \citep[Figure 7 in][]{2016ApJS..224....7D},
in which the Galactocentric radii ($\gsim 13$ kpc) of the
Outer Arm is larger than that of the MCs discussed by us here.
According to Section 3.4, to first order, the displacement
between the physical mid-plane and the $b = 0^{\circ}$ plane is
dominated by the
tilted effect because of the Sun's height off of the physical mid-plane
for regions of R$_{\rm GC}\lsim10$ kpc.
If the tilt of the Galactic plane is the cause of the
observations here, then we can derive the offset of the Sun
by extrapolating the fitted Outer Arm into the lower longitude
(e.g., $l\sim25^{\circ}-35^{\circ}$).
In Figure 7, we construct a schematic diagram to show the
presumable relationship between the physical mid-plane
and the IAU-defined plane within the solar circle.
In Figure 7, $\alpha$=$-$0\fdg046 is the Galactic latitude
of Sgr A$^{\star}$ \citep[$(\RA{17}{45}{40}.0409,\Dec{-29}{00}{28}.118)$ in
J2000 for Sgr A$^{\star}$,][]{2004ApJ...616..872R} and
$\beta$=28\fdg9 is the Galactic longitude of the Outer Arm
at $b=0^{\circ}$ (Section 3.3).
The tilted angle, $\theta$, is to be determined from the
geometric relation shown in the schematic diagram. That is, the
Outer Arm passes through the IAU-defined plane
at the direction of $l = \beta$ and $b = 0^{\circ}$.
The value for the fitted parameter of
$\beta =l = $28\fdg9 at $b=0^{\circ}$
can be derived from the linear fit (the blue solid line in Figure 4).
Assuming the distance of $d_{\rm Sgr A^{\star}}=$8.34 kpc
\citep{2014ApJ...783..130R}, the value of $d_{\rm Crossing}$
is about 13.69 kpc (see the lower left corner of Figure 7).
Accordingly, the tilted angle $\theta$ and the Sun's offset
above the physical mid-plane $Z_{\rm Sun}$ is about
0\fdg072 and 17.1 pc, respectively.
In our geometrical model (Figure 7), the values of $\theta$
and $Z_{\rm Sun}$ that we are interested in are dependent on
$\beta$ and the distance to the Galactic center,
$d_{\rm Sgr A^{\star}}$. We note that the tilted angle
($\theta$) may slightly vary from $\sim$0.067 to $\sim$0.078 when
$\beta$ is from 27$^{\circ}$
to 31$^{\circ}$. The offset of the Sun ($Z_{\rm Sun}$)
may also vary from 15.8 pc to 18.4 pc when
$d_{\rm Sgr A^{\star}}$ is changed from 8.0 kpc to 8.5 kpc and
the $\beta$ range of 27$^{\circ}$--31$^{\circ}$.
Comparing the value of $Z_{\rm Sun}$ with other investigations
\citep[e.g., Table 1 in][]{2006JRASC.100..146R}, our
estimation of $Z_{\rm Sun}$=15.8--18.4 pc
based on the distant large-scale molecular gas
agrees well with those from the method of star counts.
Moreover, the above estimation from the \twCO\
emission (Table 1) is similar to that from the \thCO\ data
(Table 2; $b$=0\fdg0335$\times l -$0\fdg9640 and $l = \beta$=28\fdg8
at $b = 0^{\circ}$).
Finally, since the warp was neglected in regions
near and within the Solar circle (R$_{\rm GC}\lsim10$ kpc), the Sun's
$z$-height based on the tilted appearance of the distant MCs
represents an upper limit.
It should be noted that our result is slightly larger
than that of \cite{1993A&A...275...67B}, in which the authors
found $Z_{\rm Sun}$= $13\pm7$ pc according to the MCs in the range
of heliocentric distances 0.7--2 kpc. It is probably due to the
uncertainty of the MC samples located nearby the Sun in their study.
On the other hand, our independent estimate is in good agreement
with the recent result of \cite{2016AstL...42....1B}, in which
they found that the mean value of $Z_{\rm Sun}$ is $16\pm2$ pc
based on samples of various objects.
In the paper, we only use the limited survey data ($\sim$ 110
square degrees) to fit the Outer Arm in CO emission.
The MWISP project will cover the remaining region of the inner
Galaxy in the four or five coming years. We hope that the accumulated
data of the new CO survey will be more helpful in the studying
of the Outer Arm's structure
in the first Galactic quadrant.
\subsection{Correlations of the Distant MCs}
In Section 3.3, we show that the distant MCs in the velocity
range of $-1.6\times l$+$13.2\km\ps$ to $-10\km\ps$ trace the Outer Arm
in the Galactic longitude $l=35^{\circ}$ to $l=45^{\circ}$.
In this section, we investigate correlations of the
distant MCs between the observed parameters, which
are directly obtained from the CLUMPFIND algorithm (Table 1)
to avoid the uncertainty from other factors
(e.g., the kinematic
distance from the rotation curve and the estimated mass
from the CO-to-H$_2$ mass conversion factor $X_{\rm CO}$).
We only consider the correlation of the physical parameters from
\twCO\ because the sample from \twCO\ ($J$=1--0) emission of the
distant MCs is at least four times larger than that
from \thCO\ ($J$=1--0). Moreover, the
signal-to-noise ratio of the \twCO\ ($J$=1--0) emission is better
than that of \thCO\ ($J$=1--0).
The thick blue line in Figure 8 indicates the relationship between
the size (2R) and the velocity dispersion ($\sigma_{v}$) of the MCs
(the so-called size-linewidth relation):
$\sigma_{v} (\km\ps) \sim \rm {size}^{0.42} (\rm {arcmin})$.
We find that the correlation is not strong. The scattering of the
velocity dispersion is large ($\sim$0.2$\km\ps$--2$\km\ps$)
in the small size range ($\sim$1--8 arcmin).
The small dynamical range, the limited observation sensitivity,
and the non-thermal motions are probably responsible for the poor
fitting \citep[Section 4.3 in][]{2014AJ....147...46Z}.
Although the scattering is large, we also note that the
power-law index of 0.42 is between other works
(\citealp[e.g., 0.33 from][]{1981MNRAS.194..809L}; and
\citealp[0.5 from][]{1987ApJ...319..730S}).
On the other hand, it probably indicates that the velocity dispersion
of MCs is not a simple power-law function of its size
\citep{2001ApJ...551..852H,2009ApJ...699.1092H}.
The left panel of Figure 9 shows a tight correlation
between the CO luminosity ($L_{\rm CO}$) and the virial
mass ($M_{\rm virial}$):
$M_{\rm virial} \sim L_{\rm CO}^{0.87}$.
The power-law index of 0.87 is in good agreement with that of
0.81 from \cite{1987ApJ...319..730S}.
On the other hand, a tight relation between the CO integrated
intensity ($W_{\rm CO}$) and the virial mass is also found:
$M_{\rm virial} \sim W_{\rm CO}^{2.11}$ with a good correlation
coefficient of 0.98. The good correlation between the virial mass
and the column density (right panel in Figure 9) is probably related to
the dependence of $\sigma_{v}/R^{0.5}$ on the column density of MCs
\citep[see Figures 7 and 8 in][]{2009ApJ...699.1092H}.
The thick blue line in Figure 10 displays the power-law relation
between the size ($2R$) and the CO luminosity ($L_{\rm CO}$)
of the MCs: $L_{\rm CO}\sim \rm {size}^{2.74}$.
We find that the fit of this power-law relation is
excellent with a correlation coefficient of 0.92.
If we use the weighted CO luminosity to fit it,
a power-law relation of $L_{\rm CO}\sim \rm {size}^{2.41}$
also describes the data.
We emphasize that the above power-law index 2.41 is consistent
well with the result from the optically thin GRS \thCO\ ($J$=1--0) data
(\citealp[a power-law index of 2.36 in ][]{2010ApJ...723..492R},
also see Figure 1 in their paper).
It probably indicates that the distant MCs are also the unresolved
parts of a pervasive fractal structure, which is similar to
the result of \cite{1996ApJ...471..816E}.
The result also shows that the mass of the $large$ MCs traced
by the \twCO\ ($J$=1--0) luminosity is similar to that traced by
the optically thin \thCO\ ($J$=1--0) emission
as analyzed by \cite{2010ApJ...723..492R}.
Furthermore, the CO luminosity from the best fit is overestimated for the
true $L_{\rm CO}$ of MCs with smaller size or lower luminosity.
\section{SUMMARY}
The MWISP project is a new large-scale survey of molecular
gas in \twCO\ ($J$=1--0), \thCO\ ($J$=1--0), and C$^{18}$O
($J$=1--0) emission.
Comparing the data with those of other CO surveys
(\citealp[e.g., GRS \thCO\ ($J$=1--0) survey in][]{2006ApJS..163..145J}
and \citealp[COHRS \twCO\ ($J$=3--2) survey in][]{2013ApJS..209....8D}),
our new CO survey has larger spatial and velocity coverage. In the
paper, we have presented the result of 110 square degree CO emission
between $l=$ 34\fdg75 to $l=$ 45\fdg25 and $b=-$5\fdg25 to $b= $5\fdg25
in the velocity interval of
$-1.6\times l$+$13.2\km\ps$ to $-10\km\ps$
to study the distant MCs of the Milky Way. Based on the new
unbiased CO survey, the main results are summarized as follows.
1. We have identified over five hundred distant MCs
according to the \twCO\ ($J$=1--0) emission in the velocity
range in the 110 square degree
region. 131 MCs in the \thCO\ ($J$=1--0) emission are also
identified among these \twCO\ ($J$=1--0) MCs.
The parameters (e.g., position, LSR velocity, peak temperature,
size, and luminosity; see Tables 1 and 2) of the distant MCs are
presented for the first time.
2, The distribution of the distant MCs in the velocity range is not
uniform with the Galactic longitude. Four distribution peaks
of the MCs can be seen at $l=$35\fdg25, 39$^{\circ}$,
42$^{\circ}$, and 44$^{\circ}$, respectively.
The interesting implication is that the massive star formation regions
seem to be concentrated in the last three peaks.
The associations between the MCs and the massive star
formation regions can be seen in Table 3.
It also shows that the star formation activity is
very common in the distant MCs.
3. The distant MCs seem to be concentrated within a narrow
Galactic latitude range along the Galactic longitude.
Most of them are within the region of $b=-$0\fdg5 and $b=$1\fdg5.
We find that the distribution of the distant MCs can be
described by a Gaussian model with Galactic latitude.
Based on the PV map of the MCs along the Galactic longitude, we thus
suggest that the CO emission of the MCs is from the Outer Arm
in the first Galactic quadrant.
4. According to the unbiased CO survey and the Galactic
rotation curve of \cite{2014ApJ...783..130R}, the heliocentric distance,
the Galactocentric radii, the scale height, and the lower limit to
the total mass of the Outer Arm in the segment
are about 15.1--15.6 kpc, 9.5--11.4 kpc, 0\fdg6 (or 160 pc at a
heliocentric distance of 15 kpc),
and 2.2$\times10^{6}\Msun$, respectively.
Assuming a constant value of $X_{\rm CO}$=
$2\E{20}$~cm$^{-2}$K$^{-1}$km$^{-1}$s and a heliocentric distance of 15 kpc,
the mass surface density of
the distant MCs is about $22\Msun$~pc$^{-2}$.
5. We note that the physical mid-plane traced by the distant
CO arm is slightly displaced with respect to the IAU-defined
plane, $b=0^{\circ}$.
We find that the Galactic warp plays a role in the region of
R$_{\rm GC}$ ($\gsim 10$ kpc) and the tilted-plane model is probably
a good approximation in the region near and within the Solar circle.
6. If the mid-plane within the Solar circle is flat, the tilted angle
between the two planes is about 0\fdg072.
And the distance from the Sun to where the
two planes cross is $d_{\rm Crossing}\sim$ 13.69 kpc assuming
$d_{\rm Sgr A^{\star}}=$8.34 kpc.
The location of the Sun, as an upper limit, is thus about 17.1 pc above the
physical mid-plane from the estimate of the tilted angle.
The offset of the Sun, which is determined independently from the
view of the large-scale structure of the distant molecular gas,
is also in agreement with results via the star-counts
method.
7. The $L_{\rm CO}$--$M_{\rm virial}$ and the $W_{\rm CO}$--$M_{\rm virial}$
relations of the distant MCs,
as well as the size--$L_{\rm CO}$ relation, display
a convincing power-law relationship,
which is consistent with the results of other studies
\citep[e.g.,][]{1987ApJ...319..730S,2010ApJ...723..492R}.
\acknowledgments
The authors acknowledge the staff members of the Qinghai Radio
Observing Station at Delingha for their support of the observations.
We thank the anonymous referee for critical
comments and suggestions that helped to improve the paper.
The TOPCAT tool \citep{2005ASPC..347...29T} was used while
preparing the paper.
This work is supported by NSFC grant 11233007.
The work is a part of the Multi-Line Galactic Plane Survey in CO and its
Isotopic Transitions, also called the Milky Way Imaging Scroll
Painting, which is supported by the Strategic Priority Research Program,
the Emergence of Cosmological Structures of the Chinese Academy of
Sciences, grant No. XDB09000000.
\bibliographystyle{apj}
|
\section{Introduction}
Cross-lingual word embedding algorithms try to represent the vocabularies of two or more languages in one common continuous vector space. These vectors can be used to improve monolingual word similarity \cite{Faruqui2014} or support cross-lingual transfer \cite{Gouws2015NAACL}. In this work, we focus on the second (cross-lingual) aspect of these embeddings, and try to determine what makes some embedding approaches better than others {\em on a set of translation-oriented benchmarks}. While cross-lingual word embeddings have been used for a variety of cross-lingual transfer tasks, we prefer evaluating on translation-oriented benchmarks, rather than across specific NLP tasks, since the translation setting allows for a cleaner examination of cross-lingual similarity. Another important delineation of this work is that we focus on algorithms that rely on {\em sentence-aligned data}; in part, because these algorithms are particularly interesting for low-resource languages, but also to make our analysis and comparison with alignment algorithms more focused.
We observe that the top performing embedding algorithms share the same underlying feature space -- sentence IDs -- while their different algorithmic approaches seem to have a negligible impact on performance. We also notice that several statistical alignment algorithms, such as IBM Model-1 \cite{Brown1993}, operate under the same data assumptions. Specifically, we find that using the translation probabilities learnt by Model-1 as the cross-lingual similarity function (in place of the commonly-used cosine similarity between word embeddings) performs on-par with state-of-the-art cross-lingual embeddings on word alignment and bilingual dictionary induction tasks. In other words, as long as the similarity function is based on the sentence ID feature space and the embedding/alignment algorithm itself is not too na\"{i}ve, the actual difference in performance between different approaches is marginal.
This leads us to revisit another statistical alignment algorithm from the literature that uses the same sentence-based signal -- the Dice aligner \cite{Och2003}. We first observe that the vanilla Dice aligner is significantly outperformed by the Model-1 aligner. We then recast Dice as the dot-product between two word vectors (based on the sentence ID feature space), which allows us to generalize it, resulting in an embedding model that is as effective as Model-1 and other sophisticated state-of-the-art embedding methods, but takes a fraction of the time to train.
Existing approaches for creating cross-lingual word embeddings are typically restricted to training bilingual embeddings, mapping exactly two languages onto a common space. We show that our generalization of the Dice coefficient can be augmented to jointly train \emph{multi}-lingual embeddings for any number of languages. We do this by leveraging the fact that the space of sentence IDs is shared among all languages in the parallel corpus; the verses of the Bible, for example, are identical across all translations. Introducing this multi-lingual signal shows a significant performance boost, which eclipses the variance in performance among pre-existing embedding algorithms.
\paragraph{Contributions} We first establish the importance of the sentence ID feature space for cross-lingual word embedding algorithms through experiments across several translation-oriented benchmarks. We then compare cross-lingual word embedding algorithms to traditional word alignment algorithms that also rely on sentence ID signals. We show that a generalization of one of these, the Dice aligner, is a very strong baseline for cross-lingual word embedding algorithms, performing better than several state-of-the-art algorithms, especially when exploiting a multi-lingual signal. Our code and data are publicly available.\footnote{\url{bitbucket.org/omerlevy/xling_embeddings}}
\section{Background: Cross-lingual Embeddings}
\label{sec:background}
Previous approaches to cross-lingual word embeddings can be divided into three categories, according to assumptions on the training data. The first category assumes \emph{word-level alignments}, in the form of bilingual dictionaries \cite{Mikolov2013Crosslingual,Xiao2014} or automatically produced word alignments \cite{Klementiev2012,Zou2013,Faruqui2014}. Sizable bilingual dictionaries are not available for many language pairs, and the quality of automatic word alignment greatly affects the quality of the embeddings. It is also unclear whether the embedding process provides significant added value beyond the initial word alignments \cite{Zou2013}. We therefore exclude these algorithms for this study, also in order to focus our analysis and make the comparison with traditional word alignment algorithms more straightforward.
The second category makes a much weaker assumption, \emph{document-level alignments}, and uses comparable texts in different languages (not necessarily translations) such as Wikipedia articles or news reports of the same event. Algorithms in this category try to leverage massive amounts of data to make up for the lack of lower-level alignments \cite{Soegaard2015,Vulic2016}.
Algorithms in the third category take the middle ground; they use \emph{sentence-level alignments}, common in legal translations and religious texts. Also known as ``parallel corpora'', sentence-aligned data maps each sentence (as a whole) to its translation. We focus on this third category, because it does not require the strict assumption of word-aligned data (which is difficult to obtain), while still providing a cleaner and more accurate signal than document-level alignments (which have been shown, in monolingual data, to capture mainly syntagmatic relations \cite{Sahlgren2006}). In \S6, we provide evidence to the hypothesis that sentence-aligned data is indeed far more informative than document-aligned data.
Algorithms that rely on sentence-aligned data typically create intermediate sentence representations from each sentence's constituent words. \newcite{Hermann2014} proposed a deep neural model, BiCVM, which compared the two sentence representations at the final layer, while \newcite{Chandar2014} proposed a shallower autoencoder-based model, representing both source and target language sentences as the same intermediate sentence vector. Recently, a simpler model, BilBOWA \cite{Gouws2015Bilbowa}, showed similar performance \emph{without} using a hidden sentence-representation layer, giving it a dramatic speed advantage over its predecessors. BilBOWA is essentially an extension of skip-grams with negative sampling (SGNS) \cite{Mikolov2013Distributed}, which simultaneously optimizes each word's similarity to its inter-lingual context (words that appeared in the aligned target language sentence) and its intra-lingual context (as in the original monolingual model). \newcite{Luong2015NAACL} proposed a similar SGNS-based model over the same features.
We study which factors determine the success of cross-lingual word embedding algorithms that use sentence-aligned data, and evaluate them against baselines from the statistical machine translation literature that incorporate the same data assumptions. We go on to generalize one of these, the Dice aligner, showing that one variant is a much stronger baseline for cross-lingual word embedding algorithms than standard baselines.
Finally, we would like to point out the work of \newcite{Upadhyay2016}, who studied how different data assumptions affect embedding quality in both monolingual and cross-lingual tasks. Our work focuses on one specific data assumption (sentence-level alignments) and only on cross-lingual usage. This more restricted setting allows us to: (a) compare embeddings to alignment algorithms, (b) decouple the feature space from the algorithm, and make a more specific observation about the contribution of each component to the end result. In that sense, our findings complement those of \newcite{Upadhyay2016}.
\section{Which Features Make Better Cross-lingual Embeddings?}
We group state-of-the-art cross-lingual embedding algorithms according to their feature sets, and compare their performance on two cross-lingual benchmarks: word alignment and bilingual dictionary induction. In doing so, we hope to learn which features are more informative.
\subsection{Features of Sentence-aligned Data}
\label{sec:features}
We observe that cross-lingual embeddings typically use parallel corpora in one of two ways:
\paragraph{Source + Target Language Words}
Each word $w$ is represented using all the other words that appeared with it in the same sentence (source language words) \emph{and} all the words that appeared in target language sentences that were aligned to sentences in which the word $w$ appeared (target language words). This representation also stores the number of times each pair of word $w$ and feature (context) word $f$ co-occurred.
These features are analogous to the ones used by \newcite{Vulic2016} for document-aligned data, and can be built in a similar manner: create a pseudo-bilingual sentence from each aligned sentence, and for each word in question, consider all the other words in this sentence as its features. BilBOWA \cite{Gouws2015Bilbowa} also uses a similar set of features, but restricts the source language words to those that appeared within a certain distance from the word in question, and defines a slightly different interaction with target language words.
\paragraph{Sentence IDs}
Here, each word is represented by the set of sentences in which it appeared, indifferent to the number of times it appeared in each one. This feature set is also indifferent to the word ordering within each sentence. This approach is implicitly used by \newcite{Chandar2014}, who encode the bag-of-words representations of parallel sentences into the same vector. Thus, each word is not matched directly to another word, but rather used to create the sentence's language-independent representation. \newcite{Soegaard2015} use similar features, document IDs, for leveraging comparable Wikipedia articles in different languages. In \S\ref{sec:data-paradigms} we show that when using sentence IDs, even a small amount of sentence-aligned data is more powerful than a huge amount of comparable documents.
\subsection{Experiment Setup}
\paragraph{Algorithms}
We use the four algorithms mentioned in \S\ref{sec:features}: BilBOWA \cite{Gouws2015Bilbowa}, BWE-SkipGram \cite{Vulic2016}, Bilingual Autoencoders \cite{Chandar2014}, and Inverted Index \cite{Soegaard2015}. While both BWE-SkipGram and Inverted Index were originally trained on document-aligned data, in this work, we apply them to sentence-aligned data.
\paragraph{Data}
\newcite{Christodouloupoulos2015} collected translations of the Bible (or parts of it) in over 100 languages, naturally aligned by book, chapter, and verse (31,102 verses in total).\footnote{\url{homepages.inf.ed.ac.uk/s0787820/bible/}} This corpus allows us to evaluate methods across many different languages, while controlling for the training set's size. The corpus was decapitalized and tokenized using white spaces after splitting at punctuation.
\paragraph{Benchmarks}
We measure the quality of each embedding using both manually annotated word alignment datasets and bilingual dictionaries. We use 16 manually annotated word alignment datasets -- Hansards\footnote{\url{www.isi.edu/natural-language/download/hansard/}} and data from four other sources \cite{Graca2008,Lambert:ea:05,Mihalcea:Pedersen:03,Holmqvist2011,Cakmak:ea:12} -- as well as 16 bilingual dictionaries from Wiktionary.
In the word alignment benchmark, each word in a given source language sentence is aligned with the most similar target language word from the target language sentence -- this is exactly the same greedy decoding algorithm that is implemented in IBM Model-1 \cite{Brown1993}. If a source language word is out of vocabulary, it is not aligned with anything, whereas target language out-of-vocabulary words are given a default minimal similarity score, and never aligned to any candidate source language word in practice. We use the inverse of alignment error rate (1-AER) as described in \newcite{Koehn2010} to measure performance, where higher scores mean better alignments.
High quality, freely available, manually annotated word alignment datasets are rare, especially for non-European languages. We therefore include experiments on bilingual dictionary induction. We obtain bilingual dictionaries from Wiktionary for five non-Indo-European languages, namely: Arabic, Finnish, Hebrew, Hungarian, and Turkish (all represented in the Edinburgh Bible Corpus). We emphasize that unlike most previous work, we experiment with finding translation equivalents of all words and do not filter the source and target language words by part of speech. We use precision-at-one (P@1), essentially selecting the closest target-language word to the given source-language word as the translation of choice. This often means that 100\% precision is unattainable, since many words have multiple translations.
\paragraph{Hyperparameters}
\newcite{Levy2015} exposed a collection of hyperparameters that affect the performance of monolingual embeddings. We assume that the same is true for cross-lingual embeddings, and use their recommended settings across all algorithms (where applicable). Specifically, we used 500 dimensions for every algorithm, context distribution smoothing with $\alpha = 0.75$ (applicable to BilBOWA and BWE-SkipGram), the symmetric version of SVD (applicable to Inverted Index), and run iterative algorithms for 100 epochs to ensure convergence (applicable to all algorithms except Inverted Index). For BilBOWA's monolingual context window, we used the default of 5. Similarity is always measured by the vectors' cosine. Most importantly, we use a shared vocabulary, consisting of every word that appeared at least twice in the corpus (tagged with language ID). While hyperparameter tuning could admittedly affect results, we rarely have data for reliably tuning hyperparameters for truly low-resource languages.
\begin{table*}[t]
\begin{center}
\scalebox{0.8}{
\begin{tabular}{| c | c | c c || c c | c c | }
\hline
& & & & \multicolumn{2}{|c|}{\multirow{2}{*}{\bf Source+Target Words}} & \multicolumn{2}{c|}{\multirow{2}{*}{\bf Sentence IDs}} \\
& & & & & & & \\
\cline{5-8}
& & & & \multirow{2}{*}{BilBOWA} & BWE & Bilingual & Inverted \\
& & & & & SkipGram & Autoencoders & Index \\
\hline
\hline
\multirow{16}{*}{\rotatebox[origin=c]{90}{\textbf{\large{Word Alignment (1-AER)}}}} & \multirow{6}{*}{\small {\sc Gra\c{c}a}} & en & fr & .3653 & .3538 & \textbf{.4376} & .3499 \\
& & fr & en & .3264 & .3676 & \textbf{.4488} & .3995 \\
& & en & es & .2723 & .3156 & \textbf{.5000} & .3443 \\
& & es & en & .2953 & .3740 & \textbf{.5076} & .4545 \\
& & en & pt & .3716 & .3983 & \textbf{.4449} & .3263 \\
& & pt & en & .3949 & .4272 & \textbf{.4474} & .3902 \\
\cline{2-8}
& \multirow{2}{*}{\small {\sc Hansards}} & en & fr & .3189 & .3109 & \textbf{.4083} & .3336 \\
& & fr & en & .3206 & .3314 & \textbf{.4218} & .3749 \\
\cline{2-8}
& \multirow{2}{*}{\small {\sc Lambert}} & en & es & .1576 & .1897 & \textbf{.2960} & .2268 \\
& & es & en & .1617 & .2073 & \textbf{.2905} & .2696 \\
\cline{2-8}
& \multirow{2}{*}{\small {\sc Mihalcea}} & en & ro & .1621 & .1848 & \textbf{.2366} & .1951 \\
& & ro & en & .1598 & .2042 & \textbf{.2545} & .2133 \\
\cline{2-8}
& \multirow{2}{*}{\small {\sc Holmqvist}} & en & sv & .2092 & .2373 & \textbf{.2746} & .2357 \\
& & sv & en & .2121 & .2853 & \textbf{.2994} & .2881 \\
\cline{2-8}
& \multirow{2}{*}{\small {\sc Cakmak}} & en & tr & .1302 & .1547 & \textbf{.2256} & .1731 \\
& & tr & en & .1479 & .1571 & .2661 & \textbf{.2665} \\
\hline
\multirow{16}{*}{\rotatebox[origin=c]{90}{\textbf{\large{Dictionary Induction (P@1)}}}} & \multirow{16}{*}{\small {\textsc{Wiktionary}}} & en & fr & .1096 & .2176 & .2475 & \textbf{.3125} \\
& & fr & en & .1305 & .2358 & .2762 & \textbf{.3466} \\
& & en & es & .0630 & .1246 & .2738 & \textbf{.3135} \\
& & es & en & .0650 & .1399 & .3012 & \textbf{.3574} \\
& & en & pt & .1384 & \textbf{.3869} & .3281 & .3866 \\
& & pt & en & .1573 & .4119 & .3661 & \textbf{.4190} \\
& & en & ar & .0385 & \textbf{.1364} & .0995 & \textbf{.1364} \\
& & ar & en & .0722 & .2408 & .1958 & \textbf{.2825} \\
& & en & fi & .0213 & .1280 & .0887 & \textbf{.1367} \\
& & fi & en & .0527 & .1877 & .1597 & \textbf{.2477} \\
& & en & he & .0418 & \textbf{.1403} & .0985 & .1284 \\
& & he & en & .0761 & .1791 & .1701 & \textbf{.2179} \\
& & en & hu & .0533 & \textbf{.2299} & .1679 & .2182 \\
& & hu & en & .0810 & .2759 & .2234 & \textbf{.3204} \\
& & en & tr & .0567 & .2207 & .1770 & \textbf{.2245} \\
& & tr & en & .0851 & .2598 & .2069 & \textbf{.2835} \\
\hline
\hline
\multicolumn{4}{|c||}{\textbf{Average}*} & .1640 & .2505 & .2856 & \textbf{.2867} \\
\multicolumn{4}{|c||}{\textbf{Top 1}} & 0 & 3.5 & \textbf{15} & 13.5 \\
\hline
\end{tabular}
}
\end{center}
\caption{The performance of four state-of-the-art cross-lingual embedding methods. * Averages across two different metrics.}
\label{tab:results_sota}
\vspace{-5pt}
\end{table*}
\subsection{Results}
Table~\ref{tab:results_sota} shows that the two algorithms based on the sentence-ID feature space perform consistently better than those using source+target words. We suspect that the source+target feature set might be capturing more information than is actually needed for translation, such as syntagmatic or topical similarity between words (e.g. ``dog'' $\sim$ ``kennel''). This might be distracting for cross-lingual tasks such as word alignment and bilingual dictionary induction. Sentence ID features, on the other hand, are simpler, and might therefore contain a cleaner translation-oriented signal.
It is important to state that, in absolute terms, these results are quite poor. The fact that the best inverse AER is around 50\% calls into question the ability to actually utilize these embeddings in a real-life scenario. While one may suggest that this is a result of the small training dataset (Edinburgh Bible Corpus), previous work (e.g. \cite{Chandar2014}) used an even smaller dataset (the first 10K sentences in Europarl \cite{Europarl}). To ensure that our results are not an artifact of the Edinburgh Bible Corpus, we repeated our experiments on the full Europarl corpus (180K sentences) for a subset of languages (English, French, and Spanish), and observed similar trends. As this is a comparative study focused on analyzing the qualitative differences between algorithms, we place the issue of low absolute performance aside for the moment, and reopen it in \S\ref{sec:polyglot}.
\begin{table*}[t]
\begin{center}
\scalebox{0.8}{
\begin{tabular}{| c | c | c c || c c | c c | }
\hline
& & & & \multicolumn{2}{|c|}{\multirow{2}{*}{\bf Embeddings}} & \multicolumn{2}{c|}{\bf Alignment} \\
& & & & & & \multicolumn{2}{c|}{\bf Algorithms} \\
\cline{5-8}
& & & & Bilingual & Inverted & \multirow{2}{*}{Dice} & IBM \\
& & & & Autoencoders & Index & & Model-1 \\
\hline
\hline
\multirow{16}{*}{\rotatebox[origin=c]{90}{\textbf{\large{Word Alignment (1-AER)}}}} & \multirow{6}{*}{\small {\sc Gra\c{c}a}} & en & fr & \textbf{.4376} & .3499 & .3355 & .4263 \\
& & fr & en & \textbf{.4488} & .3995 & .3470 & .4248 \\
& & en & es & \textbf{.5000} & .3443 & .3919 & .4251 \\
& & es & en & \textbf{.5076} & .4545 & .3120 & .4243 \\
& & en & pt & .4449 & .3263 & .3569 & \textbf{.4729} \\
& & pt & en & .4474 & .3902 & .3598 & \textbf{.4712} \\
\cline{2-8}
& \multirow{2}{*}{\small {\sc Hansards}} & en & fr & .4083 & .3336 & .3614 & \textbf{.4360} \\
& & fr & en & .4218 & .3749 & .3663 & \textbf{.4499} \\
\cline{2-8}
& \multirow{2}{*}{\small {\sc Lambert}} & en & es & \textbf{.2960} & .2268 & .2057 & .2400 \\
& & es & en & \textbf{.2905} & .2696 & .1947 & .2443 \\
\cline{2-8}
& \multirow{2}{*}{\small {\sc Mihalcea}} & en & ro & \textbf{.2366} & .1951 & .2030 & .2335 \\
& & ro & en & \textbf{.2545} & .2133 & .1720 & .2214 \\
\cline{2-8}
& \multirow{2}{*}{\small {\sc Holmqvist}} & en & sv & .2746 & .2357 & .2435 & \textbf{.3405} \\
& & sv & en & .2994 & .2881 & .2541 & \textbf{.3559} \\
\cline{2-8}
& \multirow{2}{*}{\small {\sc Cakmak}} & en & tr & .2256 & .1731 & .2285 & \textbf{.3154} \\
& & tr & en & .2661 & .2665 & .2458 & \textbf{.3494} \\
\hline
\multirow{16}{*}{\rotatebox[origin=c]{90}{\textbf{\large{Dictionary Induction (P@1)}}}} & \multirow{16}{*}{{\small \textsc{Wiktionary}}} & en & fr & .2475 & \textbf{.3125} & .1104 & .1791 \\
& & fr & en & .2762 & \textbf{.3466} & .1330 & .1816 \\
& & en & es & .2738 & \textbf{.3135} & .1072 & .0903 \\
& & es & en & .3012 & \textbf{.3574} & .1417 & .1131 \\
& & en & pt & .3281 & \textbf{.3866} & .1384 & .3779 \\
& & pt & en & .3661 & .4190 & .1719 & \textbf{.4358} \\
& & en & ar & .0995 & \textbf{.1364} & .0449 & .1316 \\
& & ar & en & .1958 & .2825 & .0610 & \textbf{.2873} \\
& & en & fi & .0887 & \textbf{.1367} & .0423 & .1340 \\
& & fi & en & .1597 & \textbf{.2477} & .0463 & .2394 \\
& & en & he & .0985 & \textbf{.1284} & .0358 & .1224 \\
& & he & en & .1701 & \textbf{.2179} & .0328 & .2000 \\
& & en & hu & .1679 & .2182 & .0569 & \textbf{.2219} \\
& & hu & en & .2234 & \textbf{.3204} & .0737 & .2985 \\
& & en & tr & .1770 & \textbf{.2245} & .0406 & .1985 \\
& & tr & en & .2069 & .2835 & .0820 & \textbf{.3073} \\
\hline
\hline
\multicolumn{4}{|c||}{\textbf{Average}} & 0.2856 & 0.2867 & 0.1843 & \textbf{0.2922} \\
\multicolumn{4}{|c||}{\textbf{Top 1}} & 8 & \textbf{12} & 0 & \textbf{12} \\
\hline
\end{tabular}
}
\end{center}
\caption{The performance of embedding and alignment methods based on the sentence ID feature set.}
\label{tab:results_alignment}
\end{table*}
\section{Comparing Cross-lingual Embeddings with Traditional Alignment Methods}
Sentence IDs are not unique to modern embedding methods, and have been used by statistical machine translation from the very beginning. In particular, the Dice coefficient \cite{Och2003}, which is often used as a baseline for more sophisticated alignment methods, measures the cross-lingual similarity of words according to the number of aligned sentences in which they appeared. IBM Model-1 \cite{Brown1993} also makes exactly the same data assumptions as other sentence-ID methods. It therefore makes sense to use Dice similarity and the translation probabilities derived from IBM Model-1 as baselines for cross-lingual embeddings that use sentence IDs.
From Table~\ref{tab:results_alignment} we learn that the existing embedding methods are not really better than IBM Model-1. In fact, their average performance is even slightly lower than Model-1's. Although Bilingual Autoencoders, Inverted Index, and Model-1 reflect entirely different algorithmic approaches (respectively: neural networks, matrix factorization, and EM), the overall difference in performance seems to be rather marginal. This suggests that the main performance factor is not the algorithm, but the feature space: sentence IDs.
However, Dice also relies on sentence IDs, yet its performance is significantly worse. We suggest that Dice uses the sentence-ID feature set na\"{i}vely, resulting in degenerate performance with respect to the other methods. In the following section, we analyze this shortcoming and show that generalizations of Dice actually do yield similar performance Model-1 and other sentence-ID methods.
\section{Generalized Dice}
In this section, we show that the Dice coefficient \cite{Och2003} can be seen as the dot-product between two word vectors represented over the sentence-ID feature set. After providing some background, we demonstrate the mathematical connection between Dice and word-feature matrices. We then introduce a new variant of Dice, SID-SGNS, which performs on-par with Model-1 and the other embedding algorithms. This variant is able to seamlessly leverage the multi-lingual nature of sentence IDs, giving it a small but significant edge over Model-1.
\subsection{Word-Feature Matrices}
\label{sec:matrices}
In the word similarity literature, it is common to represent words as real-valued vectors and compute their ``semantic'' similarity with vector similarity metrics, such as the cosine of two vectors. These word vectors are traditionally derived from sparse word-feature matrices, either by using the matrix's rows as-is (also known as ``explicit'' representation) or by inducing a lower-dimensional representation via matrix factorization \cite{Turney2010}. Many modern methods, such as those in word2vec \cite{Mikolov2013Distributed}, also create vectors by factorizing word-feature matrices, only without representing these matrices explicitly.
Formally, we are given a vocabulary of words $V_W$ and a feature space (``vocabulary of features'') $V_F$. These features can be, for instance, the set of sentences comprising the corpus. We then define a matrix $M$ of $|V_W|$ rows and $|V_F|$ columns. Each entry in $M$ represents some statistic pertaining to that combination of word and feature. For example, $M_{w,f}$ could be the number of times the word $w$ appeared in the document $f$.
The matrix $M$ is typically processed into a ``smarter'' matrix that reflects the strength of association between each given word $w$ and feature $f$. We present three common association metrics: $L_1$ row normalization (Equation~(\ref{eq:l1})), Inverse Document Frequency (IDF, Equation~(\ref{eq:idf})), and Pointwise Mutual Information (PMI, Equation~(\ref{eq:ppmi})). The following equations show how to compute their respective matrices:
\begin{equation}
\scalebox{1.0}{$
M^{L_1}_{w,f} = \frac{I(w,f)}{I(w,*)}
$}
\label{eq:l1}
\end{equation}
\begin{equation}
\scalebox{1.0}{$
M^{IDF}_{w,f} = \log\frac{|V_F|}{I(w,*)}
$}
\label{eq:idf}
\end{equation}
\begin{equation}
\scalebox{1.0}{$
M^{PMI}_{w,f} = \log\frac{\#(w,f) \cdot \#(*, *)}{\#(w, *) \cdot \#(*, f)}
$}
\label{eq:ppmi}
\end{equation}
where $\#(\cdot, \cdot)$ is the co-occurrence count function, $I(\cdot, \cdot)$ is the co-occurrence indicator function, and $*$ is a wildcard.\footnote{A function with a wildcard should be interpreted as the sum of all possible instantiations, e.g. $I(w, *) = \sum_x I(w, x)$.}
To obtain word vectors of lower dimensionality ($V_F$ may be huge), the processed matrix is then decomposed, typically with SVD. An alternative way to create low-dimensional word vectors without explicitly constructing $M$ is to use the negative sampling algorithm (SGNS) \cite{Mikolov2013Distributed}.\footnote{For consistency with prior art, we refer to this algorithm as SGNS (skip-grams with negative sampling), even when it is applied without the skip-gram feature model.} This algorithm factorizes $M^{PMI}$ using a weighted non-linear objective \cite{Levy2014NIPS}.
\subsection{Reinterpreting the Dice Coefficient}
\label{sec:dice}
\noindent In statistical machine translation, the Dice coefficient is commonly used as a baseline for word alignment \cite{Och2003}. Given sentence-aligned data, it provides a numerical measure of how likely two words -- a source-language word $w_s$ and a target-language word $w_t$ -- are each other's translation:
\begin{equation}
\scalebox{1.0}{$
Dice(w_s, w_t) = \frac{2 \cdot S(w_s, w_t)}{S(w_s, *) \cdot S(*, w_t)}
$}
\label{eq:dice}
\end{equation}
where $S(\cdot, \cdot)$ is the number of aligned sentences in the data where both arguments occurred.
We claim that this metric is mathematically equivalent to the dot-product of two $L_1$-normalized sentence-ID word-vectors, multiplied by 2. In other words, if we use the combination of sentence-ID features and $L_1$-normalization to create our word vectors, then for any $w_s$ and $w_t$:
\begin{equation}
\scalebox{1.0}{$
w_s \cdot w_t = \frac{Dice(w_s, w_t)}{2}
$}
\end{equation}
To demonstrate this claim, let us look at the dot-product of $w_s$ and $w_t$:
\begin{equation}
\scalebox{1.0}{$
w_s \cdot w_t = \sum_i \left( \frac{I(w_s,i)}{I(w_s,*)} \cdot \frac{I(w_t,i)}{I(w_t,*)} \right)
$}
\label{eq:proof1}
\end{equation}
where $i$ is the index of an aligned sentence. Since $I(w_s,*) = S(w_s,*)$ and $I(w_t, *) = S(*,w_t)$, and both are independent of $i$, we can rewrite the equation as follows:
\begin{equation}
\scalebox{1.0}{$
w_s \cdot w_t = \frac{\sum_i I(w_s,i) \cdot I(w_t,i)}{S(w_s,*) \cdot S(*,w_t)}
$}
\label{eq:proof2}
\end{equation}
Since $I(w, i)$ is an indicator function of whether the word $w$ appeared in sentence $i$, it stands to reason that the product $I(w_s, i) \cdot I(w_t, i)$ is an indicator of whether both $w_s$ and $w_t$ appeared in $i$. Ergo, the numerator of Equation~(\ref{eq:proof2}) is exactly the number of aligned sentences in which both $w_s$ and $w_t$ occurred: $S(w_s, w_t)$. Therefore:
\begin{equation}
\scalebox{1.0}{$
w_s \cdot w_t = \frac{S(w_s, w_t)}{S(w_s, *) \cdot S(*, w_t)} = \frac{Dice(w_s, w_t)}{2}
$}
\label{eq:proof3}
\end{equation}
This theoretical result implies that the cross-lingual similarity function derived from embeddings based on sentence IDs is essentially a generalization of the Dice coefficient.
\subsection{SGNS with Sentence IDs}
\begin{table*}[t]
\begin{center}
\scalebox{0.8}{
\begin{tabular}{| c | c | c c || c c c | c c | }
\hline
& & & & \multicolumn{3}{c|}{\multirow{2}{*}{\bf Prior Art}} & \multicolumn{2}{c|}{\multirow{2}{*}{\bf This Work}} \\
& & & & & & & & \\
\cline{5-9}
& & & & Bilingual & Inverted & IBM & Bilingual & Multilingual \\
& & & & Autoencoders & Index & Model-1 & SID-SGNS & SID-SGNS \\
\hline
\hline
\multirow{16}{*}{\rotatebox[origin=c]{90}{\textbf{\large{Word Alignment (1-AER)}}}} & \multirow{6}{*}{\small {\sc Gra\c{c}a}} & en & fr & .4376 & .3499 & .4263 & .4167 & \textbf{.4433} \\
& & fr & en & .4488 & .3995 & .4248 & .4300 & \textbf{.4632} \\
& & en & es & \textbf{.5000} & .3443 & .4251 & .4200 & .4893 \\
& & es & en & \textbf{.5076} & .4545 & .4243 & .3610 & .5015 \\
& & en & pt & .4449 & .3263 & \textbf{.4729} & .3983 & .4047 \\
& & pt & en & .4474 & .3902 & \textbf{.4712} & .4272 & .4151 \\
\cline{2-9}
& \multirow{2}{*}{\small {\sc Hansards}} & en & fr & .4083 & .3336 & \textbf{.4360} & .3810 & .4091 \\
& & fr & en & .4218 & .3749 & \textbf{.4499} & .3806 & .4302 \\
\cline{2-9}
& \multirow{2}{*}{\small {\sc Lambert}} & en & es & .2960 & .2268 & .2400 & .2471 & \textbf{.2989} \\
& & es & en & .2905 & .2696 & .2443 & .2415 & \textbf{.3049} \\
\cline{2-9}
& \multirow{2}{*}{\small {\sc Mihalcea}} & en & ro & .2366 & .1951 & .2335 & .1986 & \textbf{.2514} \\
& & ro & en & .2545 & .2133 & .2214 & .1914 & \textbf{.2753} \\
\cline{2-9}
& \multirow{2}{*}{\small {\sc Holmqvist}} & en & sv & .2746 & .2357 & \textbf{.3405} & .2373 & .2737 \\
& & sv & en & .2994 & .2881 & \textbf{.3559} & .2853 & .3195 \\
\cline{2-9}
& \multirow{2}{*}{\small {\sc Cakmak}} & en & tr & .2256 & .1731 & \textbf{.3154} & .1547 & .2404 \\
& & tr & en & .2661 & .2665 & \textbf{.3494} & .1571 & .2945 \\
\hline
\multirow{16}{*}{\rotatebox[origin=c]{90}{\textbf{\large{Dictionary Induction (P@1)}}}} & \multirow{16}{*}{{\small \textsc{Wiktionary}}} & en & fr & .2475 & .3125 & .1791 & .3182 & \textbf{.3304} \\
& & fr & en & .2762 & .3466 & .1816 & .3379 & \textbf{.3893} \\
& & en & es & .2738 & .3135 & .0903 & .3268 & \textbf{.3509} \\
& & es & en & .3012 & .3574 & .1131 & .3483 & \textbf{.3868} \\
& & en & pt & .3281 & .3866 & .3779 & .3869 & \textbf{.4058} \\
& & pt & en & .3661 & .4190 & .4358 & .4119 & \textbf{.4376} \\
& & en & ar & .0995 & .1364 & .1316 & .1364 & \textbf{.1605} \\
& & ar & en & .1958 & .2825 & .2873 & .2408 & \textbf{.3082} \\
& & en & fi & .0887 & .1367 & .1340 & .1280 & \textbf{.1591} \\
& & fi & en & .1597 & .2477 & .2394 & .1877 & \textbf{.2584} \\
& & en & he & .0985 & .1284 & .1224 & .1403 & \textbf{.1448} \\
& & he & en & .1701 & .2179 & .2000 & .1791 & \textbf{.2403} \\
& & en & hu & .1679 & .2182 & .2219 & .2299 & \textbf{.2482} \\
& & hu & en & .2234 & .3204 & .2985 & .2759 & \textbf{.3372} \\
& & en & tr & .1770 & .2245 & .1985 & .2207 & \textbf{.2437} \\
& & tr & en & .2069 & .2835 & .3073 & .2598 & \textbf{.3080} \\
\hline
\hline
\multicolumn{4}{|c||}{\textbf{Average}} & 0.2856 & 0.2867 & 0.2922 & 0.2830 & \textbf{0.3289} \\
\multicolumn{4}{|c||}{\textbf{Top 1}} & 2 & 0 & 8 & 0 & \textbf{22} \\
\hline
\end{tabular}}
\end{center}
\caption{The performance of SID-SGNS compared to state-of-the-art cross-lingual embedding methods and traditional alignment methods.}
\label{tab:results_sgns}
\end{table*}
The Dice coefficient appears to be a particularly na\"{i}ve variant of matrix-based methods that use sentence IDs. For example, Inverted Index \cite{Soegaard2015}), which uses SVD over IDF followed by $L_2$ normalization (instead of $L_1$ normalization), shows significantly better performance. We propose using a third variant, sentence-ID SGNS (SID-SGNS), which simply applies SGNS \cite{Mikolov2013Distributed} to the word/sentence-ID matrix (see \S5.1).
Table~\ref{tab:results_sgns} compares its performance (Bilingual SID-SGNS) to the other methods, and shows that indeed, this algorithm behaves similarly to other sentence-ID-based methods. We observe similar results for other variants as well, such as SVD over positive PMI (not shown).
\subsection{Embedding Multiple Languages}
\label{sec:polyglot}
Up until now, we used bilingual data to train cross-lingual embeddings, even though our parallel corpus (the Bible) is in fact multi-lingual. Can we make better use of this fact?
An elegant property of the sentence-ID feature set is that it is a truly inter-lingual representation. This means that multiple languages can be represented together in the same matrix before factorizing it. This raises a question: does dimensionality reduction over a \emph{multi}-lingual matrix produce better cross-lingual vectors than doing so over a bilingual matrix?
We test our hypothesis by comparing the performance of embeddings trained with SID-SGNS over all 57 languages of the Bible corpus to that of the bilingual embeddings we used earlier. This consistently improves performance across all the development benchmarks, providing a 4.69\% average increase in performance (Table~\ref{tab:results_sgns}). With this advantage, SID-SGNS performs significantly better than the other methods combined.\footnote{We observed a similar increase in performance when applying the multi-lingual signal to S{\o}gaard et al.'s \shortcite{Soegaard2015} IDF-based method and to SVD over positive PMI.} This result is similar in vein to recent findings in the parsing literature \cite{Ammar2016,Guo2016}, where multi-lingual transfer was shown to improve upon bilingual transfer.
In absolute terms, Multilingual SID-SGNS's performance is still very low. However, this experiment demonstrates that one way of making significant improvement in cross-lingual embeddings is by considering additional sources of information, such as the multi-lingual signal demonstrated here. We hypothesize that, regardless of the algorithmic approach, relying solely on sentence IDs from bilingual parallel corpora will probably not be able to improve much beyond IBM Model-1.
\section{Data Paradigms}
\label{sec:data-paradigms}
In \S\ref{sec:background}, we assumed that using sentence-aligned data is a better approach than utilizing document-aligned data. Is this the case?
To compare the data paradigms, we run the same algorithm, SID-SGNS, also on document IDs from Wikipedia.\footnote{We use the word-document matrix mined by \newcite{Soegaard2015}, which contains only a subset of our target languages: English, French, and Spanish.} We use the bilingual (not multilingual) version for both data types to control for external effects. During evaluation, we use a common vocabulary for both sentence-aligned and document-aligned embeddings.
Table~\ref{tab:data_paradigms} shows that using sentence IDs from the Bible usually outperforms Wikipedia. This remarkable result, where a small amount of parallel sentences is enough to outperform one of the largest collections of multi-lingual texts in existence, indicates that document-aligned data is an inferior paradigm for translation-related tasks such as word alignment and dictionary induction.
\begin{table}
\begin{center}
\scalebox{0.8}{
\begin{tabular}{| c | c c || c | c |}
\hline
& & & \multirow{2}{*}{\textbf{The Bible}} & \multirow{2}{*}{\textbf{Wikipedia}} \\
& & & & \\
\hline
\hline
\multirow{4}{*}{\small {\textsc{Gra\c{c}a}}} & en & fr & \textbf{.3169} & .2602 \\
& fr & en & \textbf{.3089} & .2440 \\
& en & es & \textbf{.3225} & .2429 \\
& es & en & \textbf{.3207} & .2504 \\
\hline
\multirow{2}{*}{\small {\textsc{Hansards}}} & en & fr & \textbf{.3661} & .2365 \\
& fr & en & \textbf{.3345} & .1723 \\
\hline
\multirow{2}{*}{\small {\textsc{Lambert}}} & en & es & \textbf{.2161} & .1215 \\
& es & en & \textbf{.2123} & .1027 \\
\hline
\multirow{4}{*}{\small {\textsc{Wiktionary}}} & en & fr & .3232 & \textbf{.3889} \\
& fr & en & .3418 & \textbf{.4135} \\
& en & es & \textbf{.3307} & .3262 \\
& es & en & \textbf{.3509} & .3310 \\
\hline
\hline
\multicolumn{3}{|c||}{\textbf{Average}} & \textbf{.3121} & .2575 \\
\multicolumn{3}{|c||}{\textbf{Top 1}} & \textbf{10} & 2 \\
\hline
\end{tabular}}
\end{center}
\caption{The performance of SID-SGNS with sentence-aligned data from the Bible (31,102 verses) vs document-aligned data from Wikipedia (195,000 documents).}
\label{tab:data_paradigms}
\end{table}
\section{Conclusions}
In this paper, we draw both empirical and theoretical parallels between modern cross-lingual word embeddings based on sentence alignments and traditional word alignment algorithms. We show the importance of sentence ID features and present a new, strong baseline for cross-lingual word embeddings, inspired by the Dice aligner. Our results suggest that apart from faster algorithms and more compact representations, recent cross-lingual word embedding algorithms are still unable to outperform the traditional methods by a significant margin. However, introducing our new multi-lingual signal considerably improves performance. Therefore, we hypothesize that the information in bilingual sentence-aligned data has been thoroughly mined by existing methods, and suggest that future work explore additional sources of information in order to make substantial progress.
\section*{Acknowledgments}
The work was supported in part by The European Research Council (grant number 313695) and The Israeli Science Foundation (grant number 1555/15). We would like to thank Sarath Chandar for helping us run Bilingual Autoencoders on large datasets.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.